spacehub

surshar / claude-code

Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.

Built by Anthropic · github.com/anthropics

claude-code / mods / types / claude-code.d.ts
13,165 lines · 494.5 KBRaw
1// Written by Claude Code 2.1.277.
2// Claude Code function hooks: the plugin API's TypeScript declarations.
3//
4// EARLY ACCESS: this surface may change between releases without notice.
5// Written by `/plugin-types`; regenerate with that command after an update
6// rather than editing. The first line names the Claude Code version that
7// wrote it. TypeScript 5.4 or newer reads it. `claude plugin validate <dir>`
8// is the other half: it reads a plugin's manifest and its hooks module's
9// source the way the engine will and reports what the module hooks and
10// calls and everything the engine would refuse, before any session loads it.
11//
12// What is here: the module a hooks module may import types from,
13// import type { Register, On, EngineInterface } from 'claude-code'
14// (at run time the import is empty), and the globals a hooks module has:
15// `h` and `Fragment` (what JSX compiles against), the JSX namespace, and
16// the environment's web APIs (URL, TextEncoder, AbortController,
17// crypto.subtle, ...). A hooks module runs in an environment of its own:
18// no DOM, no Node. The module, and every file it imports from the plugin,
19// is named .ts, .tsx, .jsx, .js, .mjs, .cjs, .mts or .cts (a file named
20// otherwise is not loaded) and is an ES module whatever its suffix: there
21// is no `require`. The elements a render hook draws with (`Box`, `Text`,
22// `Button`, ...) are not globals: they come from the surface's table,
23// const { Box, Text } = $.ui.resolve(e)
24//
25// Also here: 'claude-code/testing', the kit a plugin's *.test.ts and
26// *.test.tsx files import under `claude plugin test <dir>`, which runs each
27// in an environment like the one a plugin's hooks run in (no fs, network or
28// process), the plugin loaded from the folder by the engine's own host:
29// `test(name, async ($, on) => { ... })`, where `$` is the engine's own
30// and the hooks `on` registers sit beneath every plugin; with describe,
31// expect, tier, and `mock`, whose clock, store and env answer those nouns
32// beneath the plugins from memory. `$.ui.mount` draws a component through
33// the plugin on the surface the test names (terminal, desktop, vscode or
34// mobile: never assumed) and hands back the drawing to read and act on by
35// key: find an element, press a Button, type into an Input, drive a
36// `Client`'s module (keys, pointer, posts, its frame clock), each act typed
37// by that surface's element table, so one test body run over several
38// surfaces covers the mod's hooks and description on each. It exercises
39// the mod (its hooks, the tree they return under each surface's rules, its
40// `Client` modules), never a surface's paint (Ink's, the desktop page's):
41//
42// test('notes add up on every surface that takes input', async $ => {
43// for (const surface of ['terminal', 'desktop'] as const) {
44// const ui = await $.ui.mount({ plugin: 'notes', surface, ...BAND })
45// await ui.input({ key: 'new', text: 'milk' })
46// expect((await ui.find({ key: 'count' }))?.text).toBe('1 note')
47// await ui.key({ key: 'down', in: 'list' })
48// await ui.unmount()
49// }
50// for (const surface of ['terminal', 'desktop', 'vscode', 'mobile'] as const) {
51// const ui = await $.ui.mount({ plugin: 'notes', surface, ...HINT })
52// expect(await ui.find({ type: 'Text', text: /notes/ })).toBeDefined()
53// await ui.press({ key: 'dismiss' })
54// }
55// })
56//
57// Typing a plugin against it:
58// export const register: Register = (on, options) => { ... }
59// or, in a .js module,
60// /** @type {import('claude-code').Register} */
61// export const register = (on, options) => { ... }
62//
63// A tsconfig.json (or jsconfig.json) that fits a hooks module:
64// {
65// "compilerOptions": {
66// "target": "es2023", "lib": ["es2023"], "types": [],
67// "module": "esnext", "moduleResolution": "bundler",
68// "strict": true, "noUncheckedIndexedAccess": true,
69// "noEmit": true, "skipLibCheck": true,
70// "jsx": "react", "jsxFactory": "h", "jsxFragmentFactory": "Fragment"
71// },
72// "include": [".claude/types", "hooks", "tests"]
73// }
74// ".claude/types" is where /plugin-types writes this file and, beside it,
75// claude-code-mcp.d.ts and claude-code-plugins.d.ts, the index of the
76// enabled plugins' type contracts, each copied to claude-code-plugins/
77// <plugin>.d.ts: what a plugin adds to `$` in engine.create, so a plugin
78// you depend on is typed with nothing copied (the include above takes the
79// whole folder). "hooks" is the plugin's hooks/ folder and "tests" its test
80// files. `lib` names no DOM: the environment has none, and its `Text`
81// would shadow the element.
82//
83// A plugin that adds a noun to `$` ships its own contract: a .d.ts its
84// plugin.json names as "types", exporting the noun's types at its top level
85// and declaring the noun on the engine's interface,
86// export type Topo = { ... }
87// declare module 'claude-code' {
88// interface EngineInterface { topo: Topo }
89// }
90// with no import or reference, its exported names led by the noun's
91// PascalCase name; the plugin's own hooks module imports them from it.
92
93declare module 'claude-code' {
94 /**
95 * What an Agent call a plugin raised (`$.tool.call({ tool: "Agent", ... })`)
96 * answers as `result` once the agent settles; `text` is its final answer.
97 *
98 * A plugin's spawn always runs in the background, so the call resolves with
99 * this record and not the Agent tool's own (BuiltinToolResults), which is
100 * what a model's Agent call carries at `tool.call`.
101 */
102 export type AgentCallRecord = {
103 /**
104 * The spawned agent's id, as `$.agent.list()` names it.
105 */
106 agentId: string;
107 /**
108 * The model the agent resolved to, when one is known.
109 */
110 resolvedModel?: string;
111 };
112
113 /**
114 * One agent loop of this session as `$.agent.list()` returns it: a subagent
115 * or an in-process teammate.
116 */
117 export type AgentInfo = {
118 /**
119 * The agent's id: the same string its loop's `tool.call` events carry as
120 * `agentId`, and a spawn inside it as `parentAgentId`.
121 */
122 id: string;
123 /**
124 * Its row's label.
125 */
126 description: string;
127 /**
128 * The agent definition it runs as (`general-purpose`, `Explore`, ...), or
129 * `teammate` for an in-process teammate.
130 */
131 type: string;
132 /**
133 * `running`, `completed`, `failed`, `killed`, or another of the engine's task
134 * statuses.
135 */
136 status: string;
137 /**
138 * The id of the subagent whose loop spawned it; absent when the main loop
139 * did.
140 */
141 parentId?: string;
142 /**
143 * The plugin whose `$.agent.spawn` (or Agent `$.tool.call`) started it;
144 * absent when the model or the person did.
145 *
146 * The origin axis: which plugin caused it, whichever loop that hook ran in.
147 */
148 spawnedBy?: string;
149 /**
150 * What SendMessage addresses it by (`Agent({ name })`, or the engine's own
151 * for a background agent), when it has one; not `description` or `type`.
152 */
153 name?: string;
154 };
155
156 /**
157 * Which model loop an event happened in: the loop's agent id inside a
158 * subagent's or a teammate's loop, absent on the main loop.
159 *
160 * The agent axis, not the origin axis: `next.origin` names the plugin whose
161 * hook frame caused the dispatch (the recursion skip), whichever loop it ran
162 * in; `agentId` names the loop, whoever caused the call.
163 */
164 export type AgentLoop = {
165 /**
166 * The id of the loop this call runs in: for a subagent or a teammate, the
167 * `id` `$.agent.list()` gives it; absent on the main loop.
168 *
169 * A workflow's agents and the engine's own forks (compaction, memory) carry
170 * ids no list names. Pinned: a different value is refused, one left out is
171 * kept. Which loop, not which plugin caused it (that is `next.origin`).
172 */
173 agentId?: string;
174 };
175
176 /**
177 * The input of `agent.offer`: one agent type, at the moment the engine
178 * offers it to the model.
179 *
180 * Listed for the model (the agent listing) or named by it at dispatch
181 * (`subagent_type`): the same event at each. Model-facing only: a plugin's
182 * own `$.agent.spawn` of a type is no offer, and `agent.spawn` governs it.
183 */
184 export type AgentOfferInput = {
185 /**
186 * Which type (`Explore`, `Plan`, a plugin's agent); the key a matcher
187 * narrows on.
188 */
189 agent: string;
190 /**
191 * Its listing line's text as the definition states it (`whenToUse`).
192 */
193 description: string;
194 /**
195 * Where the definition came from (`built-in`, `plugin`, a settings
196 * source), so a matcher tells a built-in from a user's agent of its name.
197 */
198 source: string;
199 /**
200 * Who provides this agent: the plugin and its tier; `{ plugin: "engine",
201 * tier: "core" }` for a built-in. Pinned: a rewrite is refused.
202 */
203 provider: Origin;
204 };
205
206 /**
207 * What an `agent.offer` hook returns: whether the model is offered the agent
208 * type, in its listing and at dispatch.
209 */
210 export type AgentOfferResult = {
211 isOffered: boolean;
212 };
213
214 /**
215 * `agent.spawn`'s input as the call takes it: what the Agent tool's caller
216 * says; the engine fills the rest.
217 *
218 * A call's spawn runs in the background: it resolves once the subagent
219 * started, and nothing holds it in the foreground.
220 */
221 export type AgentSpawnArgs = Pick<AgentSpawnInput, 'prompt'> & Partial<Pick<AgentSpawnInput, 'description' | 'subagentType' | 'model' | 'name' | 'cwd'>>;
222
223 /**
224 * The input of `agent.spawn`: what the Agent tool decided about the
225 * subagent it is about to start, before its model is resolved.
226 *
227 * A hook rewrites content (prompt, description, subagentType, model,
228 * background, cwd), read back as the tool's parameters; tool_use_id, name,
229 * fork, parentModel, permissionMode, parentAgentId and provider are pinned.
230 */
231 export type AgentSpawnInput = {
232 /**
233 * The Agent tool call this spawn belongs to (for `$.ui.notice`). Pinned:
234 * the spawn's identity.
235 */
236 tool_use_id: string;
237 /**
238 * The task the subagent is given: the Agent tool's `prompt` parameter. A
239 * rewrite is the prompt the subagent runs with.
240 */
241 prompt: string;
242 /**
243 * The Agent tool's short `description` of the task (a few words). A rewrite
244 * is what the task shows as.
245 */
246 description: string;
247 /**
248 * The resolved agent type (`general-purpose`, `Explore`, a plugin's agent,
249 * `fork`). A rewrite names another agent this call can dispatch, exactly.
250 *
251 * That definition is the one spawned; a name matching none refuses the
252 * spawn, and a fork dispatches no other.
253 */
254 subagentType: string;
255 /**
256 * Who provides this agent: the plugin and its tier; `{ plugin: "engine",
257 * tier: "core" }` for a built-in. Pinned with the spawn's identity.
258 */
259 provider: Origin;
260 /**
261 * The Agent tool's `model` parameter as given, an alias (`haiku`) or a
262 * full id; undefined lets the agent's own model, then the parent's, decide.
263 *
264 * Ignored for forks, which always inherit. A hook sets this to pick the
265 * subagent's model.
266 */
267 model?: string;
268 /**
269 * The parent's effective model, what `inherit` resolves to. Pinned: a fact
270 * of the parent (set `model` to change what the subagent runs on).
271 */
272 parentModel: string;
273 /**
274 * The id of the loop the spawn happens in (its `tool.call` `agentId`; for a
275 * subagent or teammate its `$.agent.list()` id); absent from main.
276 *
277 * Pinned: a different value is refused, one left out is kept. The agent
278 * axis, not the origin: `next.origin` names the plugin whose hook frame
279 * caused the spawn, this names the model loop it happened in.
280 */
281 parentAgentId?: string;
282 /**
283 * The parent's permission mode (`default`, `acceptEdits`, `plan`, ...), which
284 * the subagent inherits. Pinned: a fact of the parent.
285 */
286 permissionMode?: string;
287 /**
288 * True when the subagent will run in the background (or remotely). A
289 * rewrite is read back as the call's `run_in_background`.
290 *
291 * The agent's own definition and remote isolation can still force it on,
292 * and disabled background tasks force it off.
293 */
294 background: boolean;
295 /**
296 * True for a fork of the parent: it inherits the parent's context and model,
297 * and `model` is ignored. Pinned: the spawn's identity.
298 */
299 fork: boolean;
300 /**
301 * Given by the call (`Agent({ name })`, addressable by SendMessage);
302 * undefined when unnamed. Pinned: the address the parent routes by.
303 */
304 name?: string;
305 /**
306 * The directory the subagent runs in when the call set one (`cwd`); undefined
307 * means the parent's. A rewrite is where the subagent runs.
308 */
309 cwd?: string;
310 };
311
312 /**
313 * What an `agent.spawn` hook returns and what `next(e)` resolves to: the
314 * started subagent's `{ model, agentId }`, or `{ deny: reason }`, refusing it.
315 *
316 * It and `$.agent.spawn(input)` resolve once the subagent started; its answer
317 * is its own `turn.complete`, carrying this `agentId`. Keep the id and set
318 * `got[agentId]` to what waits; a `turn.complete` hook resolves it:
319 *
320 * @example
321 * on("turn.complete", ($, e, next) => (got[e.agentId]?.(e.answer), next(e)))
322 */
323 export type AgentSpawnResult = {
324 /**
325 * What the subagent runs on: from core the resolved id; from a hook an
326 * alias (`haiku`) or an id.
327 */
328 model: string;
329 /**
330 * The started subagent's id: the same string its loop's `tool.call`
331 * events carry as `agentId` and `$.agent.list()` lists it by.
332 *
333 * Set by core; a hook that answers without `next` started none.
334 */
335 agentId?: string;
336 deny?: undefined;
337 } | {
338 /**
339 * Refuses the spawn, so nothing starts; the model sees the text as the
340 * Agent tool's error.
341 */
342 deny: string;
343 model?: undefined;
344 agentId?: undefined;
345 };
346
347 /**
348 * What `$.agent.register` takes: an agent type this plugin defines, spelled
349 * as an agent definition in settings JSON is, plus its `name`.
350 *
351 * Every field an agent file or `--agents` entry may carry is here and takes
352 * effect as it does there; the engine validates it with the same schema.
353 */
354 type AgentSpec = {
355 /**
356 * The type's short name (letters, digits, `_`, `-`; up to 64); the type is
357 * `<plugin>:<name>`, as a plugin's agent file `agents/<name>.md` names one.
358 */
359 name: string;
360 /**
361 * When to delegate to this agent: the line the model reads in its listing
362 * of agent types, and `agent.offer`'s `description`.
363 */
364 description: string;
365 /**
366 * The agent's system prompt, whole: it replaces the session's.
367 */
368 prompt: string;
369 /**
370 * What the agent may call, by tool name (`Read`, `Bash`, `mcp__x__y`); a
371 * call to any other is refused. Left out: every tool the parent has.
372 */
373 tools?: readonly string[];
374 /**
375 * Tools withheld from the agent, by name, out of whatever `tools` allows.
376 */
377 disallowedTools?: readonly string[];
378 /**
379 * The agent's model: an alias (`haiku`, `sonnet`, `opus`), a full id, or
380 * `inherit` for the parent's. Left out: the session's subagent default.
381 */
382 model?: string;
383 /**
384 * The agent's effort: a level (`low`, `medium`, `high`, `xhigh`, `max`)
385 * or an integer budget.
386 */
387 effort?: string | number;
388 /**
389 * How the agent's permission asks are decided, a mode: `default`,
390 * `acceptEdits`, `plan`, `auto`, `dontAsk` or `bypassPermissions`.
391 *
392 * Left out: the parent's.
393 */
394 permissionMode?: string;
395 /**
396 * Servers connected over MCP for the agent's run: a configured server's
397 * name, or `{ "<name>": { command, args } }`, one server's config inline.
398 */
399 mcpServers?: readonly (string | Record<string, unknown>)[];
400 /**
401 * Settings hooks in force while the agent runs, in settings.json's `hooks`
402 * shape (`{ PreToolUse: [{ matcher, hooks: [...] }] }`).
403 */
404 hooks?: Record<string, unknown>;
405 /**
406 * The most turns the agent takes before it stops.
407 */
408 maxTurns?: number;
409 /**
410 * Preloaded into the agent's context before its first turn: skill names.
411 */
412 skills?: readonly string[];
413 /**
414 * The agent's first user turn; `{{intent}}` in it takes the spawn's prompt.
415 * Left out: the spawn's prompt is the first turn.
416 */
417 initialPrompt?: string;
418 /**
419 * A persistent memory the agent keeps between runs, and where: `user`,
420 * `project` or `local`.
421 */
422 memory?: 'user' | 'project' | 'local';
423 /**
424 * Set so that every spawn of the agent runs in the background.
425 */
426 background?: true;
427 /**
428 * Set so that the agent's context carries no CLAUDE.md block (the person's,
429 * the project's, the rules): it takes what it needs from its prompt.
430 */
431 omitClaudeMd?: true;
432 /**
433 * Where the agent runs apart from the session: `worktree`, a git worktree
434 * of its own; `remote`, a cloud session where the build allows one.
435 */
436 isolation?: 'worktree' | 'remote';
437 };
438
439 /**
440 * The hook `on("*", hook)` takes: it runs on every event, plugin nouns no
441 * declaration names included, so `e` is `unknown` and `next` is StarNext.
442 *
443 * Until `next.is(pattern, e)` narrows `e`, all a hook can do with it is pass
444 * it on, time it, log it, or fail it; once narrowed it is an ordinary hook on
445 * those events.
446 *
447 * @param $ the engine interface; at `engine.create` the empty table, so a hook
448 * that reads `$` tests `next.is("engine.create", e)` first
449 */
450 export type AnyEventHook = ($: EngineInterface, e: unknown, next: StarNext) => unknown;
451
452 /**
453 * Every key of every variant, index signatures included.
454 */
455 type AnyKeyOf<I> = I extends unknown ? keyof I : never;
456
457 /**
458 * The argument of event `N`: `e` in its hooks, and what its call takes. For a
459 * union of names, the union of their arguments.
460 */
461 export type Args<N extends EventName = EventName> = EventOf[N];
462
463 /**
464 * Options of `$.ui.ask`.
465 */
466 export type AskOptions = {
467 /**
468 * 2-4 option labels; fewer than two are padded with Yes/No; free text is the
469 * dialog's Other.
470 */
471 options?: readonly string[];
472 /**
473 * A short chip beside the question (`Approach`, 12 characters at most).
474 */
475 header?: string;
476 /**
477 * Allow several options; the answer comes back comma-joined.
478 */
479 multiSelect?: true;
480 };
481
482 /**
483 * The input of `attribution.text`: one text the engine asks the model to
484 * write into a commit or a pull request, at the moment it is composed.
485 *
486 * Composed for the Bash description, the commit skills, a PR's body, the
487 * pre-ship mandate and the commit gate's deny: the same event at each.
488 */
489 export type AttributionTextInput = {
490 /**
491 * Which text (`commit`, `pr`, `exemption`, `remedy`); the key a matcher
492 * narrows on.
493 */
494 kind: AttributionTextKind;
495 /**
496 * As the engine composed it, the settings applied.
497 */
498 text: string;
499 };
500
501 /**
502 * Which git text `attribution.text` carries: the commit trailer, the PR
503 * footer, the mandate's or the commit gate's sentence naming the exemption.
504 */
505 export type AttributionTextKind = 'commit' | 'pr' | 'exemption' | 'remedy';
506
507 /**
508 * What an `attribution.text` hook returns: the text the model reads in
509 * that place.
510 */
511 export type AttributionTextResult = {
512 text: string;
513 };
514
515 /**
516 * What `$.audio.play` plays: a URL the engine fetches, or the bytes.
517 */
518 export type AudioClip = {
519 /**
520 * A file of the calling plugin's own, relative to the plugin's
521 * directory (`fx/open.wav`); no `.`, no leading slash.
522 *
523 * The engine resolves it and loads the file from disk.
524 */
525 asset: string;
526 url?: undefined;
527 base64?: undefined;
528 mime?: undefined;
529 } | {
530 /**
531 * The clip's URL; the engine fetches it (never the plugin).
532 */
533 url: string;
534 asset?: undefined;
535 base64?: undefined;
536 mime?: undefined;
537 } | {
538 /**
539 * The clip's bytes, base64.
540 */
541 base64: string;
542 /**
543 * What the bytes are (`audio/mpeg`, `audio/wav`).
544 */
545 mime: string;
546 asset?: undefined;
547 url?: undefined;
548 };
549
550 type BackgroundTaskSummary = {
551 id: string;
552 /**
553 * Friendly task-type label (e.g. 'shell', 'subagent', 'monitor', 'workflow'). Falls back to the raw discriminant for unknown types.
554 */
555 type: string;
556 status: string;
557 /**
558 * Free-text description. Capped at 1000 chars; clipped values append an in-string "... [+N chars]" marker.
559 */
560 description: string;
561 /**
562 * Shell command line. Only present for 'shell' tasks. Capped at 1000 chars with the same "... [+N chars]" marker.
563 */
564 command?: string;
565 /**
566 * Subagent type name. Only present for 'subagent' tasks.
567 */
568 agent_type?: string;
569 /**
570 * MCP server name. Only present for 'monitor' / 'MCP task' tasks.
571 */
572 server?: string;
573 /**
574 * MCP tool name. Only present for 'monitor' / 'MCP task' tasks.
575 */
576 tool?: string;
577 /**
578 * Workflow name. Only present for 'workflow' tasks.
579 */
580 name?: string;
581 };
582
583 type BaseHookInput = {
584 session_id: string;
585 transcript_path: string;
586 cwd: string;
587 /**
588 * UUID correlating a user prompt with all subsequent events until the next prompt. Same value emitted on OpenTelemetry events as the `prompt.id` attribute, so hook output can be joined to OTel events at prompt grain. Absent until the first user input of the process lifetime.
589 */
590 prompt_id?: string;
591 permission_mode?: string;
592 /**
593 * Subagent identifier. Present only when the hook fires from within a subagent (e.g., a tool called by an AgentTool worker). Absent for the main thread, even in --agent sessions. Use this field (not agent_type) to distinguish subagent calls from main-thread calls.
594 */
595 agent_id?: string;
596 /**
597 * Agent type name (e.g., "general-purpose", "code-reviewer"). Present when the hook fires from within a subagent (alongside agent_id), or on the main thread of a session started with --agent (without agent_id).
598 */
599 agent_type?: string;
600 /**
601 * Reasoning effort applied to the current turn. Same shape as StatusLineCommandInput.effort. Present for hooks that fire within a tool-use context (PreToolUse, PostToolUse, Stop, SubagentStop, etc.) on a model that supports the effort parameter; absent for session-lifecycle hooks and models without effort support.
602 */
603 effort?: {
604 /**
605 * Active effort level for the current turn (e.g., "low", "medium", "high", "xhigh", "max"), after any silent downgrade for the selected model. Also exposed to hook commands and Bash as the CLAUDE_EFFORT env var.
606 */
607 level: string;
608 };
609 };
610
611 /**
612 * The `Box` props a `hover` may override, none of which moves the Box's
613 * siblings, and `scope`, which names the hover group the Box joins.
614 *
615 * `display` is `"flex"` alone, on a Box drawn `display: "none"`: with a
616 * `scope`, every member of the lit group is revealed, in whichever site it
617 * sits, so a pointer on a glyph in the transcript can swap an entry into a
618 * fixed row of the band, or reveal an absolutely positioned card.
619 * `borderStyle` only restyles a border the Box has; the offsets move a Box
620 * drawn `position: "absolute"`, none in the flow. The surface applies a
621 * hover as the pointer moves, so a reveal that moves what the pointer rests
622 * on is drawn once, not in a loop.
623 */
624 export type BoxHoverProps = {
625 /**
626 * Names a hover group of this plugin's: every element it draws with the
627 * same `scope`, in any site, lights while any is hovered, reveals included.
628 *
629 * A Pane row and a mark on a transcript message can share one. Another
630 * plugin's elements under the same string are a different group. One to
631 * 64 characters, no control characters; no keyed Box needed; no hook runs.
632 */
633 scope?: string;
634 borderStyle?: string;
635 borderColor?: string;
636 borderDimColor?: boolean;
637 backgroundColor?: string;
638 display?: 'flex';
639 top?: number;
640 left?: number;
641 right?: number;
642 bottom?: number;
643 };
644
645 /**
646 * The props of `Box`: the layout, position, margin, padding and border props
647 * of Ink's Box a tree may set, and the two of hover.
648 */
649 export type BoxProps = {
650 /**
651 * Makes the Box a hover scope: while the pointer is anywhere over it, its
652 * own `hover` and that of every element beneath it apply.
653 *
654 * A nested Box with a `key` of its own scopes what is beneath it, a placed
655 * card outside its rows included. A key a sibling Box already took, or one
656 * that is no plain string, names no scope: hovers beneath it stay inert.
657 */
658 key?: string;
659 /**
660 * Style overrides applied by the surface while the pointer is over the
661 * nearest keyed `Box`, this one included, or, given a `scope`, its group.
662 *
663 * No hook runs and nothing crosses to the plugin. To reveal on hover, draw
664 * the Box `display: "none"` with `hover: { display: "flex" }` in a visible
665 * keyed Box or under a `scope`; on a `position: "absolute"` Box, no reflow.
666 */
667 hover?: BoxHoverProps;
668 /**
669 * `"absolute"` leaves the flow, as in CSS: placed against its parent by
670 * the offsets, no room among its siblings, painted over those before it.
671 *
672 * So showing or moving it (a hover may) moves nothing; the pointer on it is
673 * on its parent. Clipped, pointer and paint, by the region its site is in
674 * (viewport, pane, band), by the site itself on the main screen or export.
675 *
676 * @example <Box key="k"><Text>glyph</Text><Box position="absolute" top={-2}
677 * left={2} display="none" hover={{ display: 'flex' }}>the card</Box></Box>
678 */
679 position?: 'relative' | 'absolute';
680 /**
681 * Rows from the parent's top edge, in character cells: an integer,
682 * negative above the edge.
683 *
684 * With `bottom` and no `height` the Box spans the two.
685 */
686 top?: number;
687 /**
688 * Columns from the parent's left edge, in character cells: an integer,
689 * negative left of the edge.
690 *
691 * With `right` and no `width` the Box spans the two.
692 */
693 left?: number;
694 /**
695 * Columns from the parent's right edge, in character cells: an integer,
696 * negative right of the edge.
697 */
698 right?: number;
699 /**
700 * Rows from the parent's bottom edge, in character cells: an integer,
701 * negative below the edge.
702 */
703 bottom?: number;
704 flexDirection?: 'row' | 'column' | 'row-reverse' | 'column-reverse';
705 flexGrow?: number;
706 flexShrink?: number;
707 flexWrap?: 'nowrap' | 'wrap' | 'wrap-reverse';
708 alignItems?: 'flex-start' | 'center' | 'flex-end' | 'stretch';
709 alignSelf?: 'flex-start' | 'center' | 'flex-end' | 'auto';
710 justifyContent?: 'flex-start' | 'center' | 'flex-end' | 'space-between' | 'space-around' | 'space-evenly';
711 gap?: number;
712 columnGap?: number;
713 rowGap?: number;
714 width?: number | string;
715 height?: number | string;
716 minWidth?: number | string;
717 minHeight?: number | string;
718 margin?: number;
719 marginX?: number;
720 marginY?: number;
721 marginTop?: number;
722 marginBottom?: number;
723 marginLeft?: number;
724 marginRight?: number;
725 padding?: number;
726 paddingX?: number;
727 paddingY?: number;
728 paddingTop?: number;
729 paddingBottom?: number;
730 paddingLeft?: number;
731 paddingRight?: number;
732 borderStyle?: string;
733 borderColor?: string;
734 borderDimColor?: boolean;
735 backgroundColor?: string;
736 overflow?: 'visible' | 'hidden';
737 display?: 'flex' | 'none';
738 };
739
740 /**
741 * One variant per built-in tool; with none in the table (a plugin author's
742 * project before `/plugin-types` ran), one loose variant over every name.
743 */
744 export type BuiltinToolCallInput = [BuiltinToolName] extends [never] ? BuiltinToolCallInputFallback : {
745 [N in BuiltinToolName]: ToolInputOf<N, BuiltinToolInputs[N]>;
746 }[BuiltinToolName];
747
748 /**
749 * The built-in branch's answer when no built-in tool is declared: every
750 * name, its args unconstrained.
751 */
752 type BuiltinToolCallInputFallback = {
753 /**
754 * The name of the tool being called (`Bash`); comparing it narrows `e` once
755 * the table has entries. Reserved: a rewrite of it is ignored by core.
756 */
757 tool: string;
758 /**
759 * The tool_use block's id: the same at every event of the call and in
760 * `$.ui.notice`. Reserved: a rewrite of it is ignored by core.
761 */
762 tool_use_id: string;
763 [argument: string]: unknown;
764 };
765
766 /**
767 * The arguments of each built-in tool by name, for declaration merging;
768 * empty until a declaration file adds entries, then `e.tool === "Bash"`
769 * narrows `e` to Bash's arguments.
770 *
771 * `/plugin-types` writes this build's set beneath the engine's declarations
772 * (claude-code.d.ts), from each tool's input schema.
773 *
774 * @example
775 * interface BuiltinToolInputs { Bash: { command: string; timeout?: number } }
776 */
777 export interface BuiltinToolInputs {
778 }
779
780 /**
781 * The names of the built-in tools.
782 */
783 export type BuiltinToolName = keyof BuiltinToolInputs & string;
784
785 /**
786 * The structured result of each built-in tool by name, for declaration
787 * merging; empty until a declaration file adds entries, then after
788 * `e.tool === "Bash"` the `result` of `next(e)` is Bash's record.
789 *
790 * `/plugin-types` writes this build's set beneath the engine's declarations
791 * (claude-code.d.ts), from each tool's output schema; a tool without one is
792 * `unknown`.
793 *
794 * @example
795 * interface BuiltinToolResults { Bash: { stdout: string; stderr: string } }
796 */
797 export interface BuiltinToolResults {
798 }
799
800 /**
801 * The props of `Button`, every surface's pressable leaf: an address, a
802 * label, the closure a press runs, and the label styles a hover overrides.
803 *
804 * The terminal draws `[ label ]` (when `plain`, `1: label` or the label
805 * alone), a desktop a native button; a click, a `hotkey`, the chord for its
806 * `action`, or Enter under the focus raises `ui.press`, its bottom `onPress`.
807 */
808 export type ButtonProps = {
809 /**
810 * The element's address: `e.element` at `ui.press`, what a matcher names.
811 * Defaults to the label.
812 */
813 key?: string;
814 /**
815 * The text drawn on the button; or the one string child.
816 */
817 label?: string;
818 /**
819 * One digit (`"1"`) or one lowercase letter (`"w"`) that presses it while
820 * the plugin's site holds the focus; anything else is refused.
821 *
822 * The band after ctrl+x tab or a click, a focused `Pane`; never from the
823 * composer, save a bare digit in an empty one pressing a band Button, as a
824 * survey is answered. Shift+w matches `"w"`; two on one key, the later wins.
825 */
826 hotkey?: string;
827 /**
828 * An engine keybinding action (`"app:cycleDiffBase"`) whose chord, as the
829 * person bound it, presses this from the prompt; unknown names refused.
830 *
831 * Chords, or a modified key Global or an active context binds, on the
832 * terminal while mounted, no dialog up and no engine handler of the action
833 * mounted; a pane's over the band's over another's, then the last drawn.
834 */
835 action?: string;
836 /**
837 * Drawn without chrome: the hotkey in the accent color, a colon, the
838 * label (`1: Yes`), as a survey's row reads; no `hotkey`, the label alone.
839 *
840 * A one-glyph label (`'\u{1F50A}'`, a speaker) is then a control by
841 * itself: the focus and the pointer still invert it, `dimColor` and `hover`
842 * still apply. A desktop draws its native button either way.
843 */
844 plain?: true;
845 /**
846 * The label drawn dim at rest, as `Text`'s `dimColor`, and at full strength
847 * under the pointer or the focus: a secondary control, a path in a list.
848 */
849 dimColor?: boolean;
850 /**
851 * The site's focus ring starts here when the site takes the keyboard,
852 * instead of on nothing, as the DOM's `autofocus`: Enter acts on it at once.
853 *
854 * A pane opened with `focus`, or the person's focus chord or click, is the
855 * take. Of several in one site the first drawn wins; it raises `ui.focus`,
856 * origin this plugin. A ring the person has moved stays where it was put.
857 */
858 autoFocus?: true;
859 /**
860 * Label style overrides (the `Text` set) applied by the surface while the
861 * nearest enclosing keyed `Box`, or given a `scope` its group, is hovered.
862 *
863 * No hook runs and nothing crosses to the plugin; under the pointer itself
864 * the button inverts as it always has. Refused outside a keyed Box unless
865 * it names a `scope`.
866 */
867 hover?: TextHoverProps;
868 /**
869 * What the press runs, in the plugin's own environment: the bottom of the
870 * `ui.press` chain. The host holds only a handle, for the drawing's life.
871 */
872 onPress: () => void;
873 };
874
875 /**
876 * The handler `on(...).catch(handler)` takes for a hook of type `F`: the
877 * hook's `($, e, next)`, run afresh when it throws, misreturns or overruns.
878 *
879 * `next` carries `error` and `called` (Caught) and is replay-safe; a return
880 * within the grace is the hook's result, `undefined` the hook absent. On a
881 * streaming event the handler is a generator too, continuing the stream.
882 */
883 export type CatchHandler<F> = F extends ($: infer D, e: infer E, next: infer N) => infer R ? [R] extends [AsyncGenerator<unknown, unknown, unknown>] ? ($: D, e: E, next: N & Caught) => R : ($: D, e: E, next: N & Caught) => R | undefined | Promise<Awaited<R> | undefined> : never;
884
885 /**
886 * What `next` carries into a `.catch` handler and nowhere else: why the
887 * hook failed, and whether it had called `next` before it did.
888 *
889 * There `next` is replay-safe: when `called`, `next(e)` resolves to what the
890 * hook's last call settled to, nothing beneath running again, the argument
891 * unread; when not, it runs the hooks beneath once and a later call replays.
892 */
893 export type Caught = {
894 /**
895 * Why the hook failed (HookFailure); undefined on an ordinary hook's
896 * `next`, so its presence says a handler is running.
897 */
898 readonly error: HookFailure;
899 /**
900 * True when the failed hook had called `next()` or `next.to()` at least
901 * once, settled or in flight; the handler runs once that call settled.
902 */
903 readonly called: boolean;
904 };
905
906 /**
907 * What a hook on streaming event `N` yields, and what its `next(e)` yields
908 * to it (ChunkOf by name).
909 */
910 export type Chunk<N extends StreamingEventName = StreamingEventName> = ChunkOf[N];
911
912 /**
913 * The chunk type of each streaming event, by name: what its stream yields.
914 */
915 export type ChunkOf = {
916 /**
917 * One piece of the model's response (TurnStepChunk).
918 */
919 'turn.step': TurnStepChunk;
920 };
921
922 /**
923 * What every `turn.step` chunk may carry: the engine's handle on the item
924 * of its own stream the chunk was read off, absent on a chunk a hook made.
925 *
926 * A hook that passes a chunk on, or rewrites it by spreading it, keeps the
927 * handle, and whatever it left unchanged reaches the engine as the engine
928 * streamed it; a chunk built afresh has none and is taken at its word.
929 */
930 type ChunkRef = {
931 /**
932 * The engine's handle on its own streamed item; opaque, this step's only.
933 */
934 ref?: number;
935 };
936
937 /**
938 * The name of a classic hook event as a function-hooks event: the settings
939 * hook's own name under `classic` (`classic.Stop`, `classic.PreToolUse`).
940 */
941 export type ClassicEventName = `classic.${ClassicHookEvent}`;
942
943 /**
944 * The classic (settings) hook events, one per classic event name: `e` is
945 * what the classic hook receives on stdin for that event.
946 *
947 * The chain is [managed settings hooks, ...hooks modules, the other settings
948 * hooks as core], so a managed block ends it above every module. In shape
949 * `classic.PreToolUse` alone differs: its `e` is ToolCallEnvelope, no more.
950 */
951 export type ClassicEventOf = {
952 [E in ClassicHookEvent as `classic.${E}`]: E extends 'PreToolUse' ? ToolCallEnvelope : ClassicHookInputs[E];
953 };
954
955 /**
956 * The name of a classic hook event: `PreToolUse`, `Stop`, and the rest.
957 */
958 export type ClassicHookEvent = HookInput['hook_event_name'];
959
960 /**
961 * What a classic hook receives on stdin for each event, by event name: the
962 * Agent SDK's `<Event>HookInput`.
963 */
964 export type ClassicHookInputs = {
965 [I in HookInput as I['hook_event_name']]: I;
966 };
967
968 /**
969 * Everything a classic hook event's answer can carry, named as the classic
970 * hook's JSON output names it; each event reads its subset (ClassicResultOf).
971 *
972 * The settings hooks below fold into one of these (last write wins, contexts
973 * concatenate); a hooks module returns `next(e)`, a copy with fields changed,
974 * or its own. A field of the wrong shape fails the hook, which is skipped.
975 */
976 export type ClassicResult = {
977 /**
978 * `decision: "block"` with this text as `reason` (a command hook's exit
979 * code 2): the event's block, veto or re-prompt.
980 */
981 block?: string;
982 /**
983 * `continue: false`: the session stops after this event.
984 */
985 preventContinuation?: true;
986 /**
987 * Shown when `preventContinuation` stops the session (`stopReason`).
988 */
989 stopReason?: string;
990 /**
991 * `hookSpecificOutput.additionalContext`, one entry per hook: text handed
992 * to the model with the event.
993 */
994 additionalContext?: string[];
995 /**
996 * `hookSpecificOutput.sessionTitle` (UserPromptSubmit, SessionStart).
997 */
998 sessionTitle?: string;
999 /**
1000 * `hookSpecificOutput.suppressOriginalPrompt` (UserPromptSubmit,
1001 * UserPromptExpansion).
1002 */
1003 suppressOriginalPrompt?: true;
1004 /**
1005 * `hookSpecificOutput.initialUserMessage` (SessionStart).
1006 */
1007 initialUserMessage?: string;
1008 /**
1009 * `hookSpecificOutput.watchPaths` (SessionStart).
1010 */
1011 watchPaths?: string[];
1012 /**
1013 * `hookSpecificOutput.reloadSkills` (SessionStart).
1014 */
1015 reloadSkills?: true;
1016 /**
1017 * `hookSpecificOutput.permissionDecision` (PreModelSwitch).
1018 */
1019 permissionDecision?: 'allow' | 'deny' | 'ask';
1020 /**
1021 * `hookSpecificOutput.permissionDecisionReason` (PreModelSwitch).
1022 */
1023 permissionDecisionReason?: string;
1024 /**
1025 * `hookSpecificOutput.decision` (PermissionRequest).
1026 */
1027 decision?: PermissionRequestDecision;
1028 /**
1029 * `hookSpecificOutput.updatedToolOutput` (PostToolUse): replaces what the
1030 * model sees of the tool's result.
1031 */
1032 updatedToolOutput?: unknown;
1033 /**
1034 * `hookSpecificOutput.updatedMCPToolOutput` (PostToolUse, MCP tools only).
1035 */
1036 updatedMCPToolOutput?: unknown;
1037 /**
1038 * `hookSpecificOutput.retry` (PermissionDenied).
1039 */
1040 retry?: true;
1041 /**
1042 * `hookSpecificOutput.displayContent` (MessageDisplay).
1043 */
1044 displayContent?: string;
1045 /**
1046 * `hookSpecificOutput.worktreePath` (WorktreeCreate; a command hook prints
1047 * it): the worktree the hook created, absolute or relative to its cwd.
1048 *
1049 * Left unset, the session creates its git worktree as it would unhooked.
1050 */
1051 worktreePath?: string;
1052 };
1053
1054 /**
1055 * The event-specific fields of ClassicResult each classic event reads (its
1056 * `hookSpecificOutput`), by event; an event absent here reads none of them.
1057 *
1058 * `block`, `preventContinuation` and `stopReason` are every event's.
1059 */
1060 export type ClassicResultFields = {
1061 UserPromptSubmit: 'additionalContext' | 'sessionTitle' | 'suppressOriginalPrompt';
1062 UserPromptExpansion: 'additionalContext' | 'suppressOriginalPrompt';
1063 SessionStart: 'additionalContext' | 'initialUserMessage' | 'sessionTitle' | 'watchPaths' | 'reloadSkills';
1064 Setup: 'additionalContext';
1065 PreModelSwitch: 'permissionDecision' | 'permissionDecisionReason';
1066 PostModelSwitch: 'additionalContext';
1067 SubagentStart: 'additionalContext';
1068 PostToolUse: 'additionalContext' | 'updatedToolOutput' | 'updatedMCPToolOutput';
1069 PostToolUseFailure: 'additionalContext';
1070 PostToolBatch: 'additionalContext';
1071 Stop: 'additionalContext';
1072 SubagentStop: 'additionalContext';
1073 PermissionDenied: 'retry';
1074 PermissionRequest: 'decision';
1075 MessageDisplay: 'displayContent';
1076 WorktreeCreate: 'worktreePath';
1077 };
1078
1079 /**
1080 * What each classic hook event's hook returns and its `next(e)` resolves to:
1081 * the event's own subset of ClassicResult.
1082 *
1083 * `classic.PreToolUse` keeps its `allow` / `ask` / `deny` result
1084 * (PreToolUseResult).
1085 */
1086 export type ClassicResultOf = {
1087 [E in ClassicHookEvent as `classic.${E}`]: E extends 'PreToolUse' ? PreToolUseResult : Pick<ClassicResult, 'block' | 'preventContinuation' | 'stopReason' | (E extends keyof ClassicResultFields ? ClassicResultFields[E] : never)>;
1088 };
1089
1090 /**
1091 * Options of `$.model.classify`.
1092 */
1093 export type ClassifyOptions = {
1094 /**
1095 * An alias (`haiku`) or a full model id; default the engine's small fast
1096 * model.
1097 */
1098 model?: string;
1099 };
1100
1101 /**
1102 * The element table a surface module draws with, `surface.elements`: the
1103 * terminal's (Elements) less `Client` (none nests), `Raster` and `Image`.
1104 *
1105 * What `$.ui.resolve(e)` is to a hooks module; a Button, Input or Select
1106 * keeps its handler here. A tree of them keeps every tree's bounds (20,000
1107 * nodes, 32 deep, 100,000 characters serialized) or the instance unmounts.
1108 */
1109 export type ClientElements = Omit<Elements['terminal'], 'Client' | 'Raster' | 'Image'>;
1110
1111 /**
1112 * One key the person pressed while a `Client` had the focus, as
1113 * `surface.onKey` hands it. Escape never arrives: it returns the focus.
1114 */
1115 export type ClientKeyEvent = {
1116 /**
1117 * A special key's name (`up`, `down`, `left`, `right`, `return`, `tab`,
1118 * `backspace`, `delete`, `pageup`, `home`, ...) or the character typed.
1119 */
1120 key: string;
1121 /**
1122 * Modifier keys held with it, each present only when true.
1123 */
1124 ctrl?: true;
1125 shift?: true;
1126 meta?: true;
1127 };
1128
1129 /**
1130 * The component a surface module exports (default, or its one PascalCase
1131 * export): from props and surface to the tree drawn (no nested `Client`).
1132 *
1133 * Runs on the drawing thread, a second per call at most; `h` and `Fragment`
1134 * are the environment's, so nothing of the module's takes those names. A
1135 * throw, an overrun or a tree past the bounds (ClientElements) unmounts it.
1136 *
1137 * @example
1138 * const Tag: ClientModule<string> = (t, s) => s.elements.Text({ children: t })
1139 */
1140 export type ClientModule<P extends JsonValue = JsonValue, S = unknown> = (props: P, surface: ClientSurface<S>) => RenderElement;
1141
1142 /**
1143 * One pointer event over a `Client`'s region, as `surface.onPointer` hands
1144 * it: region-relative cells, and the sub-cell position where known.
1145 *
1146 * After a `down` in the region the instance holds the pointer until the
1147 * `up`: every `move` reaches it, past the edges too (negative, or beyond
1148 * `columns`/`rows`), and the transcript neither selects nor scrolls.
1149 */
1150 export type ClientPointerEvent = {
1151 type: ClientPointerType;
1152 /**
1153 * The column under the pointer, 0 at the region's left edge; on `enter`
1154 * and `leave`, the last column a move reported.
1155 */
1156 x: number;
1157 /**
1158 * The row under the pointer, 0 at the region's top edge.
1159 */
1160 y: number;
1161 /**
1162 * The same position with the fraction of a cell kept (`x: 12`, `fine.x:
1163 * 12.375`), where the surface knows it: a terminal that reports pixels.
1164 *
1165 * kitty, Ghostty, iTerm2, WezTerm and foot do; tmux passes none through.
1166 * Absent elsewhere and on `enter`/`leave`; present, a `move` comes within
1167 * a cell too. What a module showing another program's pixels forwards.
1168 *
1169 * @example
1170 * surface.onPointer(e => click((e.fine?.x ?? e.x + 0.5) * pixelsPerCell))
1171 */
1172 fine?: {
1173 x: number;
1174 y: number;
1175 };
1176 /**
1177 * Which button is down (`down`, `move` while held) or came up (`up`);
1178 * absent on a hover `move`, `enter` and `leave`.
1179 */
1180 button?: 'left' | 'middle' | 'right';
1181 /**
1182 * Modifier keys the terminal reported held, each present only when true.
1183 */
1184 shift?: true;
1185 alt?: true;
1186 ctrl?: true;
1187 };
1188
1189 /**
1190 * What the pointer did over a `Client`'s region: a button went down, the
1191 * pointer moved, the button came up, or the pointer crossed the region's edge.
1192 */
1193 export type ClientPointerType = 'down' | 'move' | 'up' | 'enter' | 'leave';
1194
1195 /**
1196 * The props of `Client`: which of the plugin's surface modules draws here,
1197 * under what key, with what data, in how much room.
1198 */
1199 export type ClientProps = {
1200 /**
1201 * The instance's address within the drawing: two `Client`s of one plugin in
1202 * one tree take two keys. What `e.element` carries at `ui.message`.
1203 *
1204 * The engine keeps the instance (its local state, its timers) across the
1205 * plugin's redraws while a `Client` under this key stays in the tree.
1206 */
1207 key: string;
1208 /**
1209 * The surface module's path, a string literal relative to this file:
1210 * `module: "./<name>.tsx"` (or `.jsx`, `.ts`, `.js`, `.mjs`).
1211 *
1212 * Read off the source: a variable there is refused at load, as is a path
1213 * outside the plugin or naming no file. Its default export draws, else
1214 * its one PascalCase export; loaded, the tree carries the plugin's path.
1215 */
1216 module: string;
1217 /**
1218 * Plain data (JsonValue) handed to the module function; a new value on a
1219 * redraw reaches the running instance, its state kept.
1220 *
1221 * Bounded as a tree's text is; not a channel for closures.
1222 * Typed `unknown` so a matcher over a tree stays shallow.
1223 */
1224 props?: unknown;
1225 /**
1226 * Columns the instance's region takes: a count, or a percentage of the
1227 * parent. Absent, the region is as wide as what the module draws.
1228 */
1229 width?: number | string;
1230 /**
1231 * Rows the instance's region takes: a count, or a percentage of the parent.
1232 * Absent, the region is as tall as what the module draws.
1233 */
1234 height?: number | string;
1235 /**
1236 * How the region grows into free room along the parent's direction, as a
1237 * Box's `flexGrow`.
1238 */
1239 flexGrow?: number;
1240 };
1241
1242 /**
1243 * What a surface module's function receives as its second argument: its
1244 * elements, the instance's local state, its region, input, clock and port.
1245 *
1246 * Called again (same `surface`, same `state`) on new props, after
1247 * `setState`, and on a resize; what it returns is drawn in the region. No
1248 * `$` here: the hooks module has it, and `post` is the way to reach it.
1249 */
1250 export type ClientSurface<S = unknown> = {
1251 /**
1252 * The surface's element table (ClientElements), the tags the module draws
1253 * with: `const { Box, Text } = surface.elements`. No `Client` in it.
1254 */
1255 readonly elements: ClientElements;
1256 /**
1257 * The instance's local state: `undefined` until the first `setState`.
1258 * Kept across the plugin's redraws; dropped with the instance.
1259 */
1260 readonly state: S | undefined;
1261 /**
1262 * Replaces the local state and schedules one more call of the function on
1263 * the next frame; several calls before it coalesce into one redraw.
1264 *
1265 * A `setState` on each of three renders in a row with no key, pointer,
1266 * tick, press or props between is a render loop: the instance unmounts.
1267 */
1268 setState: (next: S) => void;
1269 /**
1270 * The region's width in cells, as last laid out (0 before the first
1271 * layout).
1272 */
1273 readonly columns: number;
1274 /**
1275 * The region's height in cells, as last laid out (0 before the first
1276 * layout).
1277 */
1278 readonly rows: number;
1279 /**
1280 * Calls `fn` every `ms` milliseconds on the surface's frame clock until
1281 * the returned function is called or the instance unmounts.
1282 *
1283 * Start it once (while `state` is still undefined), not on every call.
1284 */
1285 every: (ms: number, fn: () => void) => () => void;
1286 /**
1287 * Sets the instance's pointer listener (one; a later call replaces it)
1288 * and returns what clears it. See ClientPointerEvent for capture.
1289 */
1290 onPointer: (fn: (event: ClientPointerEvent) => void) => () => void;
1291 /**
1292 * Sets the instance's key listener (one; a later call replaces it),
1293 * reached while a click has given it the focus; Escape returns that.
1294 */
1295 onKey: (fn: (event: ClientKeyEvent) => void) => () => void;
1296 /**
1297 * Sends plain data to the plugin's hooks module: `e.data` of a
1298 * `ui.message` only that plugin's hooks see, one per frame at most.
1299 *
1300 * A later post in the same frame replaces an undelivered one; a hook
1301 * answering `{ props }` hands this instance its next props. At most
1302 * 20,000 values, 32 deep, 100,000 characters, or the post is not sent.
1303 */
1304 post: (data: JsonValue) => void;
1305 };
1306
1307 /**
1308 * The argument of the `$.clock` waits (`sleep`, `after`, `every`): how long,
1309 * in milliseconds, before the dispatch resolves.
1310 */
1311 type ClockWait = {
1312 /**
1313 * The wait, a non-negative number of milliseconds.
1314 */
1315 ms: number;
1316 };
1317
1318 /**
1319 * The props of `Code`, source text every surface draws with the engine's own
1320 * highlighter: coloured tokens, a line gutter on request, or a unified diff.
1321 *
1322 * A leaf: no children. `source` is the element's data as a string is a
1323 * Text's, bounded and free of control characters the same way; the colour
1324 * on screen is the engine's, never the plugin's.
1325 */
1326 export type CodeProps = {
1327 /**
1328 * The text drawn: source code, or under `format: 'diff'` one or more
1329 * unified-diff hunks.
1330 *
1331 * At most 10000 characters; tab and newline are the only control
1332 * characters it may hold.
1333 */
1334 source: string;
1335 /**
1336 * A highlighter language id or alias (`typescript`, `ts`, `py`), a
1337 * plugin-contributed grammar's included.
1338 *
1339 * Absent, the language is inferred from `path`; when neither resolves,
1340 * the text is drawn plain.
1341 */
1342 language?: string;
1343 /**
1344 * A file path the language is inferred from when `language` is absent:
1345 * its extension or name, else a shebang on the first line.
1346 *
1347 * Drawn nowhere and never read: nothing touches the disk.
1348 */
1349 path?: string;
1350 /**
1351 * The 1-based number of the first line of `source`: present, a dim
1352 * right-aligned gutter numbers the lines from it; absent, no gutter.
1353 *
1354 * Ignored under `format: 'diff'`, whose hunks carry their own numbers.
1355 */
1356 startLine?: number;
1357 /**
1358 * `'source'` (the default) draws `source` as code; `'diff'` reads it as
1359 * unified-diff hunks and draws gutters, markers, add and remove colouring.
1360 *
1361 * A hunk is `@@ -a,b +c,d @@` then lines starting ` `, `+` or `-`; a
1362 * leading `---`/`+++` pair and `\ No newline at end of file` are read
1363 * past. A source that does not parse as hunks is refused.
1364 */
1365 format?: 'source' | 'diff';
1366 /**
1367 * What a line wider than the room does, in Text's `wrap` vocabulary:
1368 * `'wrap'` (the default) continues it on rows under the gutter.
1369 *
1370 * `'truncate-end'` cuts it at the edge with an ellipsis; Text's other
1371 * spellings are refused, since a gutter leaves them no sense here.
1372 */
1373 wrap?: 'wrap' | 'truncate-end';
1374 };
1375
1376 /**
1377 * The input of `command.describe`: how one slash command presents in the
1378 * typeahead and `/help`, at the moment the engine lists it.
1379 */
1380 export type CommandDescribeInput = {
1381 /**
1382 * Names the command without its slash; the key a matcher narrows on. A
1383 * rewrite is refused.
1384 */
1385 command: string;
1386 /**
1387 * The one-line description the menu shows, as the command declares it.
1388 */
1389 description: string;
1390 /**
1391 * The hint drawn dim after the name (`[name]`), when the command has one.
1392 */
1393 argumentHint?: string;
1394 /**
1395 * Whether the typeahead and `/help` leave the command out; a hidden
1396 * command still runs when typed in full.
1397 */
1398 isHidden: boolean;
1399 /**
1400 * Whether the command declares it runs at once when typed mid-turn,
1401 * instead of waiting for the turn to end.
1402 *
1403 * One that decides per invocation reads false. Read only: not part of
1404 * what a hook answers, and a rewrite is refused.
1405 */
1406 immediate: boolean;
1407 /**
1408 * Who provides this command: the plugin and its tier; `{ plugin: "engine",
1409 * tier: "core" }` for a built-in. Pinned: a rewrite is refused.
1410 */
1411 provider: Origin;
1412 };
1413
1414 /**
1415 * What a `command.describe` hook returns: the description, hint and hidden
1416 * flag the menu uses; the name, `immediate` and `provider` stay as they were.
1417 */
1418 export type CommandDescribeResult = Omit<CommandDescribeInput, 'command' | 'immediate' | 'provider'>;
1419
1420 /**
1421 * One slash command as `$.command.list()` returns it.
1422 */
1423 export type CommandInfo = {
1424 /**
1425 * What the person runs it by, without the slash.
1426 */
1427 name: string;
1428 /**
1429 * The one line the typeahead shows for it.
1430 */
1431 description: string;
1432 /**
1433 * Where it comes from (CommandSource).
1434 */
1435 source: CommandSource;
1436 /**
1437 * Which plugin added it, when `source` is `plugin` and the engine knows:
1438 * the one that registered it, or whose manifest carries it.
1439 */
1440 plugin?: string;
1441 };
1442
1443 /**
1444 * Where a command's answer will show: which of the terminal's two layouts
1445 * the surface renders, and how wide it is when the command runs.
1446 *
1447 * A fact the engine stamps on `command.run`, so a command that draws (opens
1448 * a pane, prints a wide table) can suit the room it has: the fullscreen
1449 * layout docks a pane beside the transcript from 110 columns, the main
1450 * screen places it inline above the prompt at any width.
1451 */
1452 export type CommandPresentation = {
1453 /**
1454 * True under the fullscreen (alternate-screen) layout; false on the main
1455 * screen (`CLAUDE_CODE_NO_FLICKER=0`, tmux by default) and headless.
1456 *
1457 * The fact `RenderViewport`'s `isFullscreen` carries on every terminal
1458 * drawing, from the same source: the two agree.
1459 */
1460 isFullscreen: boolean;
1461 /**
1462 * The terminal's width in cells as the command runs; 80 where no terminal
1463 * has measured (headless with no tty).
1464 */
1465 columns: number;
1466 };
1467
1468 /**
1469 * `command.run`'s input as a plugin's `$.command.run` takes it: `args` may
1470 * be left out (`/command`, bare); `origin` and `presentation` the engine sets.
1471 */
1472 export type CommandRunArgs = Omit<CommandRunInput, 'origin' | 'args' | 'presentation'> & {
1473 /**
1474 * Everything after the name, as the person would type it; left out, `""`.
1475 */
1476 args?: string;
1477 };
1478
1479 /**
1480 * The input of `command.run`: one slash command about to run, the way the
1481 * person typed it (`/name args`), and where the run came from.
1482 */
1483 export type CommandRunInput = {
1484 /**
1485 * Names the command without its slash (`compact`, `hello`), aliases and
1486 * folds resolved; the key a matcher narrows on. A rewrite is refused.
1487 */
1488 command: string;
1489 /**
1490 * Everything after the name, as typed (`""` when nothing was); a hook
1491 * rewrites it with `next({ ...e, args })`.
1492 */
1493 args: string;
1494 /**
1495 * Where the run came from, in `prompt.submit`'s words (PromptOrigin):
1496 * the person's Enter (`composer`), the bridge, the SDK, or a plugin.
1497 *
1498 * A plugin's `$.command.run` reads `{ kind: 'plugin', name }`. `next(e)`
1499 * passes it on as received.
1500 */
1501 origin: PromptOrigin;
1502 /**
1503 * Where the command's answer will show (CommandPresentation): the
1504 * fullscreen layout or the main screen, and the terminal's width.
1505 *
1506 * Pinned: the engine stamps it, `next(e)` passes it on, a rewrite that
1507 * leaves it out keeps it and one that changes it is refused.
1508 */
1509 presentation: CommandPresentation;
1510 };
1511
1512 /**
1513 * What a `command.run` hook returns and what `next(e)` and `$.command.run`
1514 * resolve to: the command's output text and the notes it leaves the model.
1515 *
1516 * From core, `text` is what the command printed (a `local` command's
1517 * returned text; a panel command may print nothing), `context` what it
1518 * recorded for the model beside that, and `ref` names the run.
1519 *
1520 * A hook's own answer without `next` runs no command: its `text` is shown
1521 * as the command's output, under the names of the plugins hooking the
1522 * command unless each is bundled with Claude Code, whose answer reads as
1523 * the built-in command's own, and its `context` is recorded after it.
1524 *
1525 * Both are the plugin's to size, as a core command's output is; what bounds
1526 * them is what bounds any transcript row where a surface draws it.
1527 */
1528 export type CommandRunResult = {
1529 /**
1530 * The command's output as a transcript line, or undefined when the
1531 * command showed nothing as text (a panel, a prompt for the model).
1532 */
1533 text?: string;
1534 /**
1535 * What the model reads after the command's output and the person never
1536 * sees, each entry one hidden user message recorded after the output row.
1537 *
1538 * From core, the notes the command left the model, absent when none. Kept
1539 * whole from `next`: left out after `next`, the last answer's notes ride
1540 * along; written, it keeps every entry that answer had (none empty).
1541 */
1542 context?: readonly string[];
1543 /**
1544 * Set by core on what `next(e)` resolves to: names the engine's run of
1545 * the command (its result stays on the host side).
1546 *
1547 * A hook that returns the object it got makes the engine use that run
1548 * verbatim. Absent on a hook's own `{ text }` and on `$.command.run`'s.
1549 */
1550 ref?: number;
1551 };
1552
1553 /**
1554 * Where a slash command comes from, as `$.command.list()` tells them apart.
1555 *
1556 * `builtin` ships with Claude Code; `plugin` is a plugin's markdown command,
1557 * skill or `$.command.register`; `user` is the user's or project's own
1558 * file; `mcp` an MCP server's prompt.
1559 */
1560 export type CommandSource = 'builtin' | 'plugin' | 'user' | 'mcp';
1561
1562 /**
1563 * What `$.command.register` takes: the slash command this plugin serves.
1564 */
1565 export type CommandSpec = {
1566 /**
1567 * The command's name without the slash (letters, digits, `_`, `-`; up to
1568 * 64); the person runs it as `/<name>`.
1569 */
1570 name: string;
1571 /**
1572 * The one line the typeahead and `/help` show for it.
1573 */
1574 description: string;
1575 /**
1576 * The hint drawn dim after the name (`[name]`), when it takes arguments.
1577 */
1578 argumentHint?: string;
1579 /**
1580 * Set so that `/<name>` typed while a turn is in flight runs at once
1581 * instead of waiting for the turn to end, as it does when left out.
1582 *
1583 * Its `command.run` hook then runs while a turn may still be streaming and
1584 * must not assume the turn's state (what the transcript holds, whether a
1585 * tool is mid-call); its `{ text }` prints as an idle run's does.
1586 */
1587 immediate?: true;
1588 };
1589
1590 type ConfigChangeHookInput = BaseHookInput & {
1591 hook_event_name: 'ConfigChange';
1592 source: 'user_settings' | 'project_settings' | 'local_settings' | 'policy_settings' | 'skills';
1593 file_path?: string;
1594 };
1595
1596 /**
1597 * The input of `config.describe`: how one `/config` row presents, at the
1598 * moment the menu lists it (and for `$.config.list`).
1599 */
1600 export type ConfigDescribeInput = {
1601 /**
1602 * Names the row, as `config.set`'s `key` does; the key a matcher narrows
1603 * on. Pinned: a rewrite is refused.
1604 */
1605 key: string;
1606 /**
1607 * What the menu draws for the row, before its value.
1608 */
1609 label: string;
1610 /**
1611 * The help text beneath the label, when the row has one (a plugin
1612 * field's `description`); absent for the panel's own rows.
1613 */
1614 description?: string;
1615 /**
1616 * Whether the menu leaves the row out; a hidden row still answers
1617 * `$.config.set` and `/config key=value`.
1618 */
1619 isHidden: boolean;
1620 /**
1621 * Who owns the row, as `config.set`'s `provider`. Pinned.
1622 */
1623 provider: Origin;
1624 };
1625
1626 /**
1627 * What a `config.describe` hook returns: the label, help text and hidden
1628 * flag the menu uses; the key and provider stay as they were.
1629 */
1630 export type ConfigDescribeResult = Omit<ConfigDescribeInput, 'key' | 'provider'>;
1631
1632 /**
1633 * How a `/config` row takes its value: `boolean` toggles, `choice` picks one
1634 * of its `options`, `text` takes a string, `number` a number.
1635 */
1636 export type ConfigKind = 'boolean' | 'choice' | 'text' | 'number';
1637
1638 /**
1639 * Where a `config.set` came from, in `prompt.submit`'s words: the person
1640 * in the `/config` menu (`composer`), the bridge, or a plugin, named.
1641 *
1642 * Set by the engine; `next(e)` passes it on as received and none sets it.
1643 */
1644 export type ConfigOrigin = {
1645 /**
1646 * The person's own change in the `/config` menu (a toggle, a pick, a
1647 * typed value), or their `/config key=value`.
1648 */
1649 kind: 'composer';
1650 } | {
1651 /**
1652 * A `/config key=value` that arrived over the Remote Control bridge (a
1653 * phone or web client, or a relay): not attestably the owner's hand.
1654 */
1655 kind: 'bridge';
1656 } | {
1657 /**
1658 * A plugin's `$.config.set`; that plugin's own hooks do not see it.
1659 */
1660 kind: 'plugin';
1661 /**
1662 * The calling plugin's name.
1663 */
1664 name: string;
1665 };
1666
1667 /**
1668 * One `/config` row as `$.config.list()` returns it: what the menu would
1669 * draw now, after every `config.describe` hook, a hidden row left out.
1670 */
1671 export type ConfigRow = {
1672 /**
1673 * Names the row: the panel's id for a built-in, `<plugin>.<field>` for a
1674 * plugin's `userConfig` field; what `$.config.set` takes.
1675 */
1676 key: string;
1677 /**
1678 * What the menu draws for the row, before its value.
1679 */
1680 label: string;
1681 /**
1682 * The help text under the label, when the row has one.
1683 */
1684 description?: string;
1685 /**
1686 * How the row takes its value (ConfigKind).
1687 */
1688 kind: ConfigKind;
1689 /**
1690 * What the row holds now.
1691 */
1692 value: ConfigValue;
1693 /**
1694 * The values a `choice` row takes, in order (a plugin string field's
1695 * declared `options` for its row); absent otherwise.
1696 */
1697 options?: readonly string[];
1698 /**
1699 * Who owns the row: the engine for the panel's own, else the plugin.
1700 */
1701 provider: Origin;
1702 /**
1703 * Whether a trusted source (managed settings, the organization's policy)
1704 * owns the value, so the menu shows it and refuses a change.
1705 */
1706 isLocked: boolean;
1707 };
1708
1709 /**
1710 * `config.set`'s input as a plugin's `$.config.set(args)` takes it: the
1711 * row's key and the value; the engine fills the rest.
1712 */
1713 export type ConfigSetArgs = Pick<ConfigSetInput, 'key' | 'value'>;
1714
1715 /**
1716 * The input of `config.set`: one `/config` row about to change, from the
1717 * menu or a plugin's `$.config.set`, with what it holds now and who owns it.
1718 */
1719 export type ConfigSetInput = {
1720 /**
1721 * Names the row: the panel's own id for a built-in (`theme`, `verbose`,
1722 * `autoCompact`), `<plugin>.<field>` for a plugin's `userConfig` field.
1723 *
1724 * The key a matcher narrows on; pinned: a rewrite is refused.
1725 */
1726 key: string;
1727 /**
1728 * What the row is being set to; `next({ ...e, value })` clamps or
1729 * replaces it, held to the row's kind (a toggle takes a boolean).
1730 */
1731 value: ConfigValue;
1732 /**
1733 * What the row holds before the change, as the menu shows it. Pinned.
1734 */
1735 previous: ConfigValue;
1736 /**
1737 * Who owns the row: `{ plugin: "engine", tier: "core" }` for the panel's
1738 * own, the plugin and its tier for a `userConfig` field. Pinned.
1739 */
1740 provider: Origin;
1741 /**
1742 * Where the change came from (ConfigOrigin): the person in the menu
1743 * (`composer`) or a plugin's `$.config.set`. Pinned.
1744 */
1745 origin: ConfigOrigin;
1746 };
1747
1748 /**
1749 * What a `config.set` hook returns and what `next(e)` resolves to:
1750 * `{ value }` once written, or `{ deny: reason }`, the row left as it was.
1751 *
1752 * The menu shows a deny's reason beside the row; a plugin's `$.config.set`
1753 * resolves with it.
1754 */
1755 export type ConfigSetResult = {
1756 value: ConfigValue;
1757 deny?: undefined;
1758 } | {
1759 deny: string;
1760 value?: undefined;
1761 };
1762
1763 /**
1764 * A `/config` row's value as a hook and `$.config` see it: a toggle's
1765 * boolean, a choice's or a text's string, a number, or a list of strings.
1766 */
1767 export type ConfigValue = boolean | string | number | readonly string[];
1768
1769 /**
1770 * One custom agent whose description the Agent tool's prompt carries;
1771 * built-in agents are left out.
1772 */
1773 export type ContextAgent = {
1774 /**
1775 * The agent's type, as the Agent tool names it.
1776 */
1777 agentType: string;
1778 /**
1779 * Where it was defined, by the engine's word (`projectSettings`,
1780 * `userSettings`, `plugin`); the display label is the renderer's.
1781 */
1782 source: string;
1783 /**
1784 * The description's estimated tokens.
1785 */
1786 tokens: number;
1787 };
1788
1789 /**
1790 * The token counts the last API response of the live window reported, as
1791 * the API spells them; the breakdown's `Messages` row is reconciled to it.
1792 */
1793 export type ContextApiUsage = {
1794 /**
1795 * Uncached input tokens the response was answered over.
1796 */
1797 input_tokens: number;
1798 /**
1799 * Tokens the response generated.
1800 */
1801 output_tokens: number;
1802 /**
1803 * Input tokens written to the prompt cache by the request.
1804 */
1805 cache_creation_input_tokens: number;
1806 /**
1807 * Input tokens read from the prompt cache by the request.
1808 */
1809 cache_read_input_tokens: number;
1810 };
1811
1812 /**
1813 * How a context breakdown is counted: `full` with the token-count API per
1814 * category, `summary` from the last response's usage and local estimates.
1815 *
1816 * The SDK's `get_context_usage` takes the same two words as its `detail`.
1817 */
1818 export type ContextBreakdownDetail = 'summary' | 'full';
1819
1820 /**
1821 * One row of the breakdown, as /context lists it beside the grid (`System
1822 * prompt`, `Messages`, `Free space`, `Autocompact buffer`).
1823 */
1824 export type ContextCategory = {
1825 /**
1826 * The row's label as /context prints it.
1827 */
1828 name: string;
1829 /**
1830 * The row's estimated tokens; a `deferred` row's do not count toward the
1831 * total.
1832 */
1833 tokens: number;
1834 /**
1835 * The theme colour /context draws the row and its squares in, by its key
1836 * in the theme (`promptBorder`, `inactive`, `permission`).
1837 */
1838 color: string;
1839 /**
1840 * Whether the row is tool schemas loaded on demand, which the grid leaves
1841 * out; the same fact as `kind` `deferred`.
1842 */
1843 isDeferred: boolean;
1844 /**
1845 * What the row is (ContextCategoryKind), stamped by the engine.
1846 */
1847 kind: ContextCategoryKind;
1848 };
1849
1850 /**
1851 * What a breakdown row is; branch on this, never on the row's `name`.
1852 *
1853 * `used` content occupies the window, `free` is the window left, `buffer`
1854 * the compaction reserve, `deferred` tool schemas loaded on demand and
1855 * outside the window.
1856 */
1857 export type ContextCategoryKind = 'used' | 'free' | 'buffer' | 'deferred';
1858
1859 /**
1860 * One square of the grid /context draws: which row it belongs to and how
1861 * full it is.
1862 */
1863 export type ContextGridSquare = {
1864 /**
1865 * The theme colour of the square's row, by its key in the theme.
1866 */
1867 color: string;
1868 /**
1869 * Whether the square holds any of its row's tokens.
1870 */
1871 isFilled: boolean;
1872 /**
1873 * The `name` of the row the square belongs to (`Free space` for the
1874 * window left).
1875 */
1876 categoryName: string;
1877 /**
1878 * The row's tokens, repeated on each of its squares.
1879 */
1880 tokens: number;
1881 /**
1882 * The row's share of the window as a whole percentage, repeated likewise.
1883 */
1884 percentage: number;
1885 /**
1886 * How full this one square is, 0 to 1: a row's last square is the partial
1887 * one (/context draws it hollow under 0.7).
1888 */
1889 squareFullness: number;
1890 };
1891
1892 /**
1893 * One MCP tool's schema as the context carries it.
1894 */
1895 export type ContextMcpTool = {
1896 /**
1897 * The tool's wire name (`mcp__linear__create_issue`).
1898 */
1899 name: string;
1900 /**
1901 * The server it belongs to, as /mcp lists it.
1902 */
1903 serverName: string;
1904 /**
1905 * The schema's estimated tokens.
1906 */
1907 tokens: number;
1908 /**
1909 * Whether the schema is inside the window now: always, unless tool
1910 * schemas load on demand and this one has not been searched for yet.
1911 */
1912 isLoaded: boolean;
1913 };
1914
1915 /**
1916 * One memory file the context carries (a CLAUDE.md, a rules file, an
1917 * auto-memory entry).
1918 */
1919 export type ContextMemoryFile = {
1920 /**
1921 * The file's path, absolute.
1922 */
1923 path: string;
1924 /**
1925 * The display label of where it was loaded from (`Project`, `User`,
1926 * `Local`, `Managed`, `AutoMem`).
1927 */
1928 type: string;
1929 /**
1930 * The file's estimated tokens.
1931 */
1932 tokens: number;
1933 };
1934
1935 /**
1936 * One skill whose listing the context carries.
1937 */
1938 export type ContextSkill = {
1939 /**
1940 * The skill's name as `/skills` lists it.
1941 */
1942 name: string;
1943 /**
1944 * Where it came from, by the engine's word (`userSettings`, `plugin`,
1945 * `built-in`, `mcp`, `syncedSkills`); the display label is the renderer's.
1946 */
1947 source: string;
1948 /**
1949 * The providing plugin's name, when the skill comes from one.
1950 */
1951 pluginName?: string;
1952 /**
1953 * The listing's estimated tokens.
1954 */
1955 tokens: number;
1956 };
1957
1958 /**
1959 * The skills the context lists for the model: how many there are, how many
1960 * fit the listing's budget, and each one's share.
1961 */
1962 export type ContextSkills = {
1963 /**
1964 * How many skills the session has.
1965 */
1966 totalSkills: number;
1967 /**
1968 * How many the listing included within its token budget.
1969 */
1970 includedSkills: number;
1971 /**
1972 * The listing's tokens in all.
1973 */
1974 tokens: number;
1975 /**
1976 * One entry per listed skill.
1977 */
1978 skillFrontmatter: ContextSkill[];
1979 };
1980
1981 /**
1982 * The slash commands the Skill tool's prompt lists, counted.
1983 */
1984 export type ContextSlashCommands = {
1985 /**
1986 * How many commands the session has.
1987 */
1988 totalCommands: number;
1989 /**
1990 * How many the listing included.
1991 */
1992 includedCommands: number;
1993 /**
1994 * The listing's tokens in all.
1995 */
1996 tokens: number;
1997 };
1998
1999 /**
2000 * How the window the breakdown measures against was settled; /context
2001 * prints its `Auto-compact window` line off this.
2002 *
2003 * `auto` is the model's own limit; the rest name a compaction window by who
2004 * set it: the `CLAUDE_CODE_AUTO_COMPACT_WINDOW` variable, the settings, the
2005 * account, an experiment, the model's default, or an unrecognised model's.
2006 */
2007 export type ContextWindowSource = 'env' | 'settings' | 'clientdata' | 'experiment' | 'model-default' | 'unknown-model' | 'auto';
2008
2009 /**
2010 * The plugin's identity (`plugin`) and the nouns core contributes to `$` as
2011 * the innermost step of the `engine.create` fold.
2012 *
2013 * A plugin's finished `$` (EngineInterface) has every core noun an outer
2014 * step did not withhold, and every noun the plugins' steps added.
2015 */
2016 export interface CoreEngineInterface {
2017 /**
2018 * This plugin, as loaded: its manifest name and its directory.
2019 */
2020 plugin: {
2021 /**
2022 * From plugin.json; debug-log and `$.ui.log` lines carry it.
2023 */
2024 name: string;
2025 /**
2026 * The plugin's directory (the one holding plugin.json), absolute.
2027 */
2028 root: string;
2029 };
2030 /**
2031 * Display: a line under an open dialog, a redraw or a repaint, a
2032 * transcript or debug line, panes the surface places, a window or ring.
2033 */
2034 ui: {
2035 /**
2036 * Shows `text` as one line under the dialog open for `tool_use_id`, or
2037 * removes the line when `text` is undefined.
2038 *
2039 * Core removes the line when the call resolves. A call that is not open
2040 * is refused, as is another plugin's `$.tool.call` run.
2041 *
2042 * @param tool_use_id the call whose open dialog gets the line
2043 * @param text the line to show; undefined removes it
2044 * @example
2045 * $.ui.notice(e.tool_use_id, "checked by my-plugin"); return next(e)
2046 */
2047 notice: (tool_use_id: string, text: string | undefined) => void;
2048 /**
2049 * Re-runs an event whose results the engine caches: `ui.render` draws the
2050 * instances this plugin may draw again; the others drop the cached answers.
2051 *
2052 * A render hook whose state changed (a countdown) calls it for a redraw, at
2053 * most ten a second, thirty for the shown pane and the band (calls sooner
2054 * fold); a prompt section, context or attachment hook: dropped next turn.
2055 *
2056 * @param event `ui.render`, or a cached-answer event: `prompt.section`,
2057 * `prompt.context`, `prompt.attachment`, `tool/command/config.describe`
2058 */
2059 invalidate: (event: InvalidatableEventName) => void;
2060 /**
2061 * Repaints a mounted `Raster` this plugin's own render hook drew with new
2062 * cells, or swaps a keyed `Image` it drew to a new source; no redraw.
2063 *
2064 * The surface paints the cells or source at its next frame, so blits
2065 * between frames fold into one: up to 120 a second taken, some sixty
2066 * shown. An Image swap sends one small command; a resize is a redraw.
2067 *
2068 * @param args `requestId` (the site), `key`, and `cells` (RasterProps) or
2069 * `source` (ImageSource); `columns`, `rows` (the mounted size)
2070 * @returns `{}` once the cells or source are its next frame, or `{ deny }`
2071 * (not mounted, not this plugin's, another size, cells that do
2072 * not decode, a bad source, or an Image drawing its alt there:
2073 * no placeholder images, or a file this terminal cannot read)
2074 * @example
2075 * $.clock.every(33, () => $.ui.blit({ requestId, key, cells: frame() }))
2076 * @example
2077 * await $.ui.blit({ requestId: 'browser', key: 'view',
2078 * source: { shm: name, format: 'rgb', width, height } })
2079 */
2080 blit: (args: UiBlitArgs) => Promise<UiBlitResult>;
2081 /**
2082 * The elements of the surface `e` is drawn on (Elements[e.surface]): a
2083 * frozen table of constructors, the JSX tags a render hook draws with.
2084 *
2085 * A read, not a dispatch: the engine ran `ui.resolve` (the other hooks,
2086 * then core) at load, per surface and component. Narrowed `e.surface`:
2087 * that table exactly; unnarrowed: the union, so shared names type-check.
2088 *
2089 * @param e this hook's own `ui.render` argument (its surface and
2090 * component pick the table)
2091 * @returns the surface's frozen element table (`Elements[e.surface]`)
2092 * @example
2093 * const { Box, Text } = $.ui.resolve(e)
2094 * return <Box>{await next(e)}<Text dimColor>done</Text></Box>
2095 */
2096 resolve: <E extends ResolveInput>(e: E) => Elements[E['surface']];
2097 /**
2098 * Appends one line to the transcript, drawn like a system notice (dim;
2099 * not sent to the model), or with `{ to: "debug" }` to the debug log alone.
2100 *
2101 * A row of its own at the next frame, in logging order; a `-p` or SDK
2102 * host receives it as `ui_log`; the debug log has every line under this
2103 * plugin's name. Raised as `ui.log`: a hook above may rewrite `e.to`.
2104 *
2105 * @param text the line's text
2106 * @param options `to`: `transcript` (the default) or `debug`
2107 * @example
2108 * $.ui.log(`prompt from ${e.origin.kind}: ${e.text.length} chars`)
2109 * @example
2110 * $.ui.log(`cache miss for ${e.tool_use_id}`, { to: "debug" })
2111 */
2112 log: (text: string, options?: UiLogOptions) => void;
2113 /**
2114 * Asks the user `question` in the engine's own AskUserQuestion dialog and
2115 * resolves to the label they chose, or the text typed under "Other".
2116 *
2117 * A `tool.call` of `AskUserQuestion` through every hook but the calling
2118 * one, drawn by `ui.render` on `AskUserQuestion`; a multi-select answer is
2119 * comma-joined. Rejects when dismissed, and in a `-p` run (no one to ask).
2120 *
2121 * @param question the question, ending in a question mark
2122 * @param options 2-4 option labels, or `{ options, header, multiSelect }`
2123 * @returns the label chosen, the chosen labels comma-joined, or free text
2124 * typed under "Other" (so compare it with the labels exactly)
2125 * @example
2126 * const mood = await $.ui.ask("How careful?", ["Bold", "Careful"])
2127 */
2128 ask: (question: string, options?: readonly string[] | AskOptions) => Promise<string>;
2129 /**
2130 * Shows `text` on the notification bar under the prompt for a few
2131 * seconds, the way the engine's own "context left" notice appears.
2132 *
2133 * It leaves the transcript and the model untouched.
2134 *
2135 * @param text the line to show; an unpaired surrogate half in it is drawn
2136 * as U+FFFD
2137 * @param options `timeoutMs`: how long it stays (default 4000)
2138 * @example
2139 * $.ui.toast(`turn took ${Math.round(e.durationMs / 1000)} s`)
2140 */
2141 toast: (text: string, options?: ToastOptions) => void;
2142 /**
2143 * Pins `text` as this plugin's status line under the prompt, beside the
2144 * engine's own pinned notices, until the next call replaces it.
2145 *
2146 * One per plugin; `undefined` removes it.
2147 *
2148 * @param text the line to keep on screen (an unpaired surrogate half is
2149 * drawn as U+FFFD); undefined clears it
2150 * @example
2151 * $.ui.status("thinking..."); return next(e)
2152 */
2153 status: (text: string | undefined) => void;
2154 /**
2155 * Opens a pane: a framed region the surface places, whose body this
2156 * plugin draws by hooking `ui.render` for `{ component: "Pane" }`.
2157 *
2158 * One instance per id (`requestId`): opening an open id retitles it; a hook
2159 * on `ui.open` may refuse it. The keyboard is the person's; unasked, it
2160 * waits undrawn below 144 columns (110 once asked), judged at each open.
2161 *
2162 * @param pane `id` (1-64 of letters, digits, `_`, `-`), `title`, `focus`,
2163 * `closeOnEscape` and `holdToasts` (a dialog), `rows` / `columns` wanted
2164 * @returns settles once the pane is open (or retitled)
2165 * @example
2166 * await $.ui.open({ id: "clock", title: "Clock" })
2167 * @example
2168 * await $.ui.open({ id: "ask", focus: true, closeOnEscape: true, rows: 9 })
2169 */
2170 open: (pane: PaneOpenArgs) => Promise<void>;
2171 /**
2172 * Closes one of the open panes; an id that is not open is left alone.
2173 *
2174 * Every close raises `ui.close`, `e.origin` naming whose it is: this call
2175 * (`plugin`), the person's mark or key (`person`), an unload (`unload`).
2176 * A hook answering without `next` keeps the pane open, save on an unload.
2177 *
2178 * @param pane `id`: the id the pane was opened under
2179 * @returns settles once the pane is gone or a hook answered for it;
2180 * rejects on a hook's `{ deny }`
2181 * @example
2182 * onPress: () => $.ui.close({ id: "clock" })
2183 */
2184 close: (pane: PaneCloseArgs) => Promise<void>;
2185 /**
2186 * Lists this plugin's own open panes (UiPane): each one's id and title,
2187 * and whether it is shown, holds the keyboard, and is placed.
2188 *
2189 * The engine's record, not the module's: a module reloaded while its
2190 * pane stayed up finds it here. Another plugin's panes are not listed.
2191 *
2192 * @returns the panes in open order, placed ones first; empty with none
2193 * @example
2194 * const isUp = (await $.ui.panes()).some(pane => pane.id === "clock")
2195 * if (!isUp) await $.ui.open({ id: "clock", title: "Clock" })
2196 */
2197 panes: () => Promise<readonly UiPane[]>;
2198 /**
2199 * Scrolls something into view as the DOM's `scrollIntoView` would: a
2200 * render instance by `requestId`, an element by `key`, a site's edge.
2201 *
2202 * A site of this plugin's (its pane, the band it draws into) moves under
2203 * the event `ui.scroll`, origin `plugin`. A transcript row moves only
2204 * while this call answers the person's own input, where one scrolls.
2205 *
2206 * @param args `to` (what), `in` (which site, required for `start` and
2207 * `end`), `block` (where it lands; `nearest` by default)
2208 * @returns `{}` once it moved, or `{ deny }` saying why not
2209 * @example
2210 * onPress: () => $.ui.scroll({ in: "log", to: "end" })
2211 */
2212 scroll: (args: UiScrollArgs) => Promise<UiScrollResult>;
2213 /**
2214 * Moves the focus ring of one of this plugin's sites onto an element it
2215 * drew there, as the DOM's `element.focus()`, while it holds the keys.
2216 *
2217 * Raised as the event `ui.focus`, origin `plugin`; the engine's inverse
2218 * marks the element. The keyboard is the person's to give: a site not
2219 * holding it, or holding it on another plugin's element, is `{ deny }`.
2220 *
2221 * @param args `requestId` (which site: a pane's id, the band's) and `key`
2222 * (the element's, as drawn)
2223 * @returns `{}` once it moved, or `{ deny }` saying why not
2224 * @example
2225 * onPress: () => $.ui.focus({ requestId: "files", key: "row:0" })
2226 */
2227 focus: (args: UiFocusArgs) => Promise<UiFocusResult>;
2228 };
2229 /**
2230 * Completions through the session's own client and credentials.
2231 */
2232 model: {
2233 /**
2234 * Runs one text completion through the session's own API client and
2235 * resolves to the reply's text.
2236 *
2237 * No tools, no history, no system prompt beyond the CLI's identity block
2238 * and `request.system`.
2239 *
2240 * @param request the model (an alias such as `haiku`, or a full id,
2241 * resolved like a `--model` value), the prompt, and a token cap
2242 * @returns the reply's text
2243 * @example
2244 * const reply = await $.model.complete({ model: "haiku", prompt: "Hi." })
2245 */
2246 complete: (request: ModelCompleteRequest) => Promise<string>;
2247 /**
2248 * Runs one tool-less completion over the session's OWN transcript, sharing
2249 * the main thread's prompt cache; the reply's text and the fork's usage.
2250 *
2251 * Not `complete`: sharing the cache prefix means the model and system
2252 * prompt are the session's, read from its last turn's cache-safe snapshot,
2253 * and tools are denied. Null on a cold snapshot or an API error.
2254 *
2255 * @param request the one user message the fork answers
2256 * @returns the reply's text with the fork's usage, or null on a cold
2257 * snapshot or an API error
2258 * @example
2259 * const reply = await $.model.fork({ prompt: "One line to learn?" })
2260 * if (reply !== null) $.ui.log(`${reply.usage.output_tokens} tokens`)
2261 */
2262 fork: (request: ModelForkRequest) => Promise<ModelForkResult | null>;
2263 /**
2264 * Picks one of `labels` for `text` with one completion over
2265 * `$.model.complete` and a fixed classifier prompt.
2266 *
2267 * `text` is data; the model answers with a label alone. It resolves
2268 * `undefined` when the answer named none of `labels`. A failed request,
2269 * an abort or a reply with no text rejects, naming the cause.
2270 *
2271 * @param text what to classify
2272 * @param labels the labels to choose from (2 or more)
2273 * @param options `model`: an alias or id; default the engine's small
2274 * fast model
2275 * @returns the label the model named, or undefined when it named none;
2276 * rejects when the request fails or the reply has no text
2277 * @example
2278 * const kind = await $.model.classify(e.text, ["bug", "feature"])
2279 * if (kind === undefined) return next(e)
2280 */
2281 classify: (text: string, labels: readonly string[], options?: ClassifyOptions) => Promise<string | undefined>;
2282 };
2283 /**
2284 * Sound: clip playback and platform speech.
2285 */
2286 audio: {
2287 /**
2288 * Plays one audio clip, starting now; clips are not queued, so two calls
2289 * play together (a bed under speech).
2290 *
2291 * `{ asset }` is the plugin's own file, loaded by the engine and played
2292 * through the platform's player (`afplay` on macOS). Resolves when
2293 * playback ends; rejects, naming the cause, when the clip cannot play.
2294 *
2295 * @param clip the plugin's own file (`{ asset }`), a URL the engine
2296 * fetches, or the bytes as base64 with their MIME type
2297 * @param options `shouldLoop`, `gain`, and an AbortSignal that stops the
2298 * clip
2299 */
2300 play: (clip: AudioClip, options?: PlayOptions) => Promise<void>;
2301 /**
2302 * Speaks `text` with the platform's own synthesizer (`say` on macOS).
2303 * Plain text.
2304 *
2305 * Utterances are queued among themselves; clips are not. Resolves when
2306 * the utterance has ended; rejects, naming the cause, when there is no
2307 * synthesizer, the voice is not installed, or the utterance failed.
2308 *
2309 * @param text what to say, as plain text
2310 * @param options `voice`: the system voice's exact name (`Samantha`);
2311 * absent, the synthesizer's default
2312 * @returns which synthesizer spoke, once the utterance has ended
2313 */
2314 speak: (text: string, options?: SpeakOptions) => Promise<SpeakResult>;
2315 };
2316 /**
2317 * The engine's connected MCP servers.
2318 */
2319 mcp: {
2320 /**
2321 * Calls `tool` on one of the engine's connected MCP servers with the
2322 * engine's own connection and credentials.
2323 *
2324 * A `cached` server is dialed on first use. No permission prompt: the
2325 * plugin's call, seen by the hooks above it, is the grant. Positional, not
2326 * the `{ tool: "mcp__server__tool", ... }` shape a `tool.call` hook sees.
2327 *
2328 * @param server the server's name as /mcp lists it (`claude.ai Gmail`;
2329 * the tool-name spelling `claude_ai_Gmail` is accepted too)
2330 * @param tool the tool's name on that server (`create_draft`)
2331 * @param args the tool's arguments; none when absent
2332 * @returns the tool's result as MCP returns it: `content` blocks and
2333 * `isError`
2334 * @example
2335 * const { content } = await $.mcp.call("claude.ai Gmail", "create_draft", {
2336 * to: "team@example.com",
2337 * subject: "Release notes",
2338 * })
2339 */
2340 call: (server: string, tool: string, args?: Record<string, unknown>) => Promise<McpToolResult>;
2341 };
2342 /**
2343 * The running session, read as plain data, and compacting it.
2344 */
2345 session: {
2346 /**
2347 * Returns the transcript so far, one entry per user or assistant
2348 * message; progress rows, `$.ui.log` lines and notices are not messages.
2349 *
2350 * Each entry is a SessionMessage, `{ role, text, toolUses }` (a user
2351 * message may add `toolResults`; a `toolUses` entry adds its `result` and
2352 * `text` once answered). A long transcript answers its newest 4096.
2353 *
2354 * @example
2355 * const last = (await $.session.messages()).at(-1)
2356 */
2357 messages: () => Promise<SessionMessage[]>;
2358 /**
2359 * Returns the directory the session runs in, absolute.
2360 */
2361 cwd: () => Promise<string>;
2362 /**
2363 * Returns the session's project root, absolute: where it started, or
2364 * where `/cd`, a host's directory change or a worktree move took it.
2365 *
2366 * A shell `cd` during the session does not move it; nested instruction
2367 * files are read only beneath it.
2368 */
2369 root: () => Promise<string>;
2370 /**
2371 * Returns the main loop's model, as `/model` shows it.
2372 */
2373 model: () => Promise<string>;
2374 /**
2375 * Returns how many prompts the user has sent this session (user turns in
2376 * the transcript).
2377 */
2378 turns: () => Promise<number>;
2379 /**
2380 * Returns the session's id (the transcript file's name).
2381 */
2382 id: () => Promise<string>;
2383 /**
2384 * Returns the git repository the session runs in, read from the working
2385 * copy on each call; null when the directory is not inside one.
2386 *
2387 * @example
2388 * const repo = await $.session.repo(); const publicRepo = !repo?.internal
2389 */
2390 repo: () => Promise<SessionRepo | null>;
2391 /**
2392 * Returns every surface the session draws on, each once: `terminal` under
2393 * the REPL first, then the remote ones in the order they attached.
2394 *
2395 * A session may draw on several at once (a terminal and two phones):
2396 * clients attach (`session.attach`) and detach, and a render hook still
2397 * reads `e.surface` per ask. Empty in a plain -p run; never rejects.
2398 *
2399 * @example
2400 * const inApp = (await $.session.surfaces()).some(s => s !== "terminal")
2401 */
2402 surfaces: () => Promise<readonly RenderSurface[]>;
2403 /**
2404 * Returns the first of `$.session.surfaces()`, or null where nothing
2405 * draws.
2406 *
2407 * @deprecated use `surfaces()`; a session may draw on several surfaces at
2408 * once
2409 */
2410 surface: () => Promise<RenderSurface | null>;
2411 /**
2412 * Returns the context window's fill, the rate-limit windows and the
2413 * cost, as the status line has them; with `breakdown`, by category too.
2414 *
2415 * The plain call costs nothing; `"full"` counts each category with the
2416 * token-count API as /context does, `"summary"` estimates locally, and
2417 * `context.breakdown` comes back in the SDK's `get_context_usage` shape.
2418 *
2419 * @param args `{ breakdown, columns }`: how the breakdown is counted and
2420 * the width its grid is drawn in; nothing for the status line's figures
2421 * @returns `{ context, rateLimits, cost }` as the status line has them
2422 * @example
2423 * const { context } = await $.session.usage()
2424 * if ((context.percent ?? 0) >= 85) await $.session.compact()
2425 * @example
2426 * const usage = await $.session.usage({ breakdown: "full", columns })
2427 * for (const row of usage.context.breakdown?.gridRows ?? []) draw(row)
2428 */
2429 usage: (args?: SessionUsageArgs) => Promise<SessionUsage>;
2430 /**
2431 * Returns the version of the engine the session runs on, the release it
2432 * is built from, and when it was built.
2433 *
2434 * The same three values the engine's own analytics rows carry, answered
2435 * in every mode and build; `base` is absent when the version is not
2436 * spelled as a release, `builtAt` in a run from source that stamps none.
2437 *
2438 * @returns `{ version, base, builtAt }`: the full version, its release
2439 * (`2.1.280`, or `2.1.280-dev` for a development build) and an
2440 * ISO 8601 build time
2441 * @example
2442 * const { version, base, builtAt } = await $.session.version()
2443 * row.env = { version, version_base: base, build_time: builtAt }
2444 * @example
2445 * const isDevelopmentBuild = (await $.session.version()).base?.endsWith("-dev")
2446 */
2447 version: () => Promise<SessionVersion>;
2448 /**
2449 * Compacts the conversation: the event `session.compact` with `trigger`
2450 * `plugin`, the same call `/compact` makes, between turns.
2451 *
2452 * It runs through every hook but the calling one, then core: a summary
2453 * and the kept messages in the transcript's place. Resolves `{ skip }`
2454 * when a hook vetoed it; rejects while a turn runs.
2455 *
2456 * @example
2457 * const { skip } = await $.session.compact({ instructions: "the plan" })
2458 */
2459 compact: EventCalls['session']['compact'];
2460 /**
2461 * Holds the session's Anthropic credential on the host and answers an
2462 * opaque handle and its kind; the secret never reaches the plugin.
2463 *
2464 * The handle is spent through `$.http.fetch(url, { auth: handle })`,
2465 * which sets the credential header, only for a first-party host. Null
2466 * where the build or the provider has no first-party credential to hold.
2467 *
2468 * @example
2469 * await $.http.fetch(url, { auth: (await $.session.authorize())?.handle })
2470 */
2471 authorize: () => Promise<SessionAuthorization>;
2472 };
2473 /**
2474 * The running model turn: ending it.
2475 */
2476 turn: {
2477 /**
2478 * Cancels the running model turn: the one whose id `turn.start` handed
2479 * this plugin, its running tools stopped, no interruption marker.
2480 *
2481 * The event `turn.abort`, seen by the hooks above; the prompt this
2482 * plugin submits next is the context. Rejects, naming both ids, when
2483 * `turnId` is not the running turn's; a hook may end its own turn.
2484 *
2485 * @param input `turnId`: the id `turn.start` carried
2486 * @example
2487 * on("turn.start", ($, e, next) => { held = e.turnId; return next(e) })
2488 */
2489 abort: (input: OpEventOf['turn.abort']) => Promise<void>;
2490 };
2491 /**
2492 * Submitting a prompt the model reads as a user turn, and the person's
2493 * prompt box: read as it stands, written, or proposed into.
2494 */
2495 prompt: {
2496 /**
2497 * Submits a prompt: the event `prompt.submit`, the same call the engine
2498 * makes for a typed prompt; `input.text` runs when the session is idle.
2499 *
2500 * It goes through every hook but the calling one (the plugin's others
2501 * see it) with `e.origin` `{ kind: 'plugin', name }`, the name the
2502 * model reads it under unless a hook leaves it out of its answer.
2503 *
2504 * @example
2505 * void $.prompt.submit({ text: "List the TODOs you just mentioned." })
2506 */
2507 submit: EventCalls['prompt']['submit'];
2508 /**
2509 * Returns the prompt box as it stands, the draft typed so far and the
2510 * cursor's offset into it, so a `fill` can keep what the person typed.
2511 *
2512 * Never rejects: `{ text: '', cursor: 0 }` where the session draws no box
2513 * (a -p run, an SDK host) or none is mounted yet.
2514 *
2515 * @example
2516 * const { text, cursor } = await $.prompt.read()
2517 */
2518 read: () => Promise<PromptBox>;
2519 /**
2520 * Puts `input.text` in the prompt box as the draft, by `mode`: `replace`
2521 * (the default) over it, `append` after it, `insert` at the cursor.
2522 *
2523 * The event `prompt.fill` through the other plugins' hooks; `isFilled:
2524 * false` under a dialog or headless. To hand the model text WITH the next
2525 * prompt instead, a `prompt.submit` hook adds `context` (second example).
2526 *
2527 * @example
2528 * await $.prompt.fill({ text: `> ${quote}\n`, mode: "insert" })
2529 * @example
2530 * ($, e, next) => next({ ...e, context: [...(e.context ?? []), hunk] })
2531 */
2532 fill: (input: PromptFillArgs) => Promise<PromptFilled>;
2533 /**
2534 * Proposes `input.text` as the prompt box's dim suggestion, Tab to take:
2535 * the event `prompt.suggest`, as the engine's own guess after a turn.
2536 *
2537 * It goes through every other plugin's hook with `e.origin` `{ kind:
2538 * 'plugin', name }`, the engine's own suggestions on or off; `{ isShown:
2539 * false }` while the box holds text, a turn runs, or headless (no box).
2540 *
2541 * @example
2542 * void $.prompt.suggest({ text: "run the tests you just wrote" })
2543 */
2544 suggest: EventCalls['prompt']['suggest'];
2545 };
2546 /**
2547 * The tools the model has in this session, and running one.
2548 */
2549 tool: {
2550 /**
2551 * Returns the tools the model can call now, built-in and MCP alike, in
2552 * the order the model sees them.
2553 *
2554 * @example
2555 * const names = (await $.tool.list()).map(t => t.name)
2556 */
2557 list: () => Promise<ToolInfo[]>;
2558 /**
2559 * Calls a tool: the event `tool.call`, the same call the engine makes for
2560 * the model's tool calls, under a `tool_use_id` of its own.
2561 *
2562 * It runs through every hook but the calling one (the plugin's others
2563 * see it), the permission check and its dialog, then the tool. Rejects
2564 * when no tool has that name or the call is aborted.
2565 *
2566 * @example
2567 * const { text } = await $.tool.call({ tool: "Read", file_path: "a.md" })
2568 */
2569 call: EventCalls['tool']['call'];
2570 /**
2571 * Asks the engine's permission decision for a tool call now: the event
2572 * `tool.check`, resolved to `{ decision, reason?, rule? }`.
2573 *
2574 * The hooks run (the calling hook's own frame skipped, `next.origin` this
2575 * plugin, no `tool_use_id`); nothing runs, no dialog opens, no PreToolUse
2576 * hook or classifier is asked.
2577 *
2578 * @example
2579 * const { decision } = await $.tool.check({ tool: "Read", input })
2580 */
2581 check: EventCalls['tool']['check'];
2582 /**
2583 * Declares a tool the model can call from the next prompt on: the name,
2584 * description and input schema of `mcp__<plugin>__<name>`.
2585 *
2586 * Serve it with a `tool.call` hook on `{ tool: "mcp__<plugin>__<name>" }`
2587 * that returns the result (a call no hook answers fails); a name registered
2588 * again is replaced. Rejects until the session binds, at `session.start`.
2589 *
2590 * @param tool `name`, `description` (what the model reads), `inputSchema`
2591 * (a JSON schema object; default `{ type: "object" }`)
2592 * @returns `{ tool }`, the registered tool's full name
2593 * `mcp__<plugin>__<name>`
2594 * @example
2595 * await $.tool.register({ name: "weather", description: "Weather." })
2596 */
2597 register: (tool: ToolSpec) => Promise<OpValueOf['tool.register']>;
2598 };
2599 /**
2600 * The slash commands the person can run in this session, and running one.
2601 */
2602 command: {
2603 /**
2604 * Returns the slash commands the person can run now, built-in, plugin
2605 * and MCP alike, in the order the typeahead lists them.
2606 *
2607 * @example
2608 * const names = (await $.command.list()).map(c => c.name)
2609 */
2610 list: () => Promise<CommandInfo[]>;
2611 /**
2612 * Runs a slash command as if the person typed `/command args`: the
2613 * event `command.run`, queued and run once the session is idle.
2614 *
2615 * It runs through every hook but the calling one with `e.origin`
2616 * `{ kind: 'plugin', name }`, its lines in the transcript. Rejects an
2617 * unknown name, and inside a hook the turn is waiting on.
2618 *
2619 * @example
2620 * const { text } = await $.command.run({ command: "status" })
2621 */
2622 run: EventCalls['command']['run'];
2623 /**
2624 * Declares the slash command `/<name>` for this session, listed in the
2625 * typeahead from the next keystroke on.
2626 *
2627 * Serve it with a `command.run` hook on `{ command: "<name>" }` that
2628 * returns `{ text }`; a run no hook answers says so as its output.
2629 * Registering a name again replaces it; a built-in's name is refused.
2630 *
2631 * @param command `name`, `description` (what the menu shows),
2632 * `argumentHint` (dim after the name), `immediate` (runs mid-turn)
2633 * @returns `{ command }`, the registered name
2634 * @example
2635 * await $.command.register({ name: "hello", description: "Says hi." })
2636 */
2637 register: (command: CommandSpec) => Promise<OpValueOf['command.register']>;
2638 };
2639 /**
2640 * Every row of the settings menu (`/config`), the panel's own and each
2641 * enabled plugin's `userConfig` fields alike: listing and changing them.
2642 */
2643 config: {
2644 /**
2645 * Returns the rows the `/config` menu would draw now, in its order,
2646 * each with its current value, its kind, its owner and its lock.
2647 *
2648 * After every `config.describe` hook: a hidden row is left out, a
2649 * relabelled one carries the new label.
2650 *
2651 * @example
2652 * const theme = (await $.config.list()).find(row => row.key === "theme")
2653 */
2654 list: () => Promise<ConfigRow[]>;
2655 /**
2656 * Changes one row as if the person did in the menu: the event
2657 * `config.set` with `origin` `{ kind: 'plugin', name }`, then the writer.
2658 *
2659 * Through the other plugins' hooks, this plugin's own skipped; `{ deny }`
2660 * when a hook refused, the value does not fit, a trusted source owns the
2661 * row or only its dialog changes it. Rejects a key no row has.
2662 *
2663 * @param args `key` (as `list` names it) and `value` (the row's kind)
2664 * @returns `{ value }` once written, or `{ deny }`
2665 * @example
2666 * const { deny } = await $.config.set({ key: "verbose", value: true })
2667 */
2668 set: EventCalls['config']['set'];
2669 };
2670 /**
2671 * Subagents.
2672 */
2673 agent: {
2674 /**
2675 * Spawns a subagent: the event `agent.spawn`, the same call the engine
2676 * makes when the Agent tool starts one; the engine fills the rest.
2677 *
2678 * It runs every hook but the calling one, then the Agent tool in the
2679 * background under this call's origin: `{ model, agentId }` once the
2680 * subagent started (its answer is its `turn.complete`), or `{ deny }`.
2681 *
2682 * @example
2683 * const { agentId } = await $.agent.spawn({ prompt: "Read README.md." })
2684 */
2685 spawn: EventCalls['agent']['spawn'];
2686 /**
2687 * Returns the session's subagents so far, the ones the model spawned and
2688 * the ones plugins did alike.
2689 */
2690 list: () => Promise<AgentInfo[]>;
2691 /**
2692 * Defines an agent type the Agent tool dispatches from the next turn on,
2693 * named `<plugin>:<name>`: the event `agent.register`.
2694 *
2695 * Every field takes effect as in an agent file; re-registered, a name is
2696 * replaced; unloaded, a plugin's types go. `agent.offer` hides it from
2697 * the model alone; any plugin's `$.agent.spawn` answers to `agent.spawn`.
2698 *
2699 * @param spec `name`, `description` (when to delegate), `prompt` (its
2700 * system prompt), and any other field of an agent definition
2701 * @returns `{ agent }`, the full name; rejects until the session binds,
2702 * on a spec the schema refuses (its reason), on a hook's deny
2703 * @example
2704 * await $.agent.register({ name: "runner", description: "Runs a spec",
2705 * prompt: RUNNER_PROMPT, tools: ["Read", "Bash"], omitClaudeMd: true })
2706 * @example
2707 * // runner-only: hidden from the model, spawned by this plugin's tool,
2708 * // answered by the subagent's turn.complete (matched by agentId)
2709 * on("agent.offer", { agent: "lab:runner" }, () => ({ isOffered: false }))
2710 * on("tool.call", { tool: "mcp__lab__run" }, async ($, e) => {
2711 * const { agentId, deny } = await $.agent.spawn({
2712 * subagentType: "lab:runner", prompt: e.spec, description: "run" })
2713 * return { result: deny ?? (await answerOf(agentId)) }
2714 * })
2715 */
2716 register: (spec: AgentSpec) => Promise<OpValueOf['agent.register']>;
2717 };
2718 /**
2719 * The file system as the engine's own process reaches it; a relative path
2720 * is under the session's working directory, and text is UTF-8.
2721 *
2722 * An absolute path is used as given; where one may go is an `fs.*` hook's
2723 * to say. A read or write over 4 MiB rejects, a foreign network location
2724 * as spelled rejects untouched, and an OS refusal rejects with its errno.
2725 */
2726 fs: {
2727 /**
2728 * Reads a file and returns its text, or with `{ as: "bytes" }` its bytes
2729 * as `{ base64 }`.
2730 *
2731 * Rejects when missing, or over 4 MiB, which bounds what one read copies
2732 * into the plugin's environment. A file the plugin ships is under
2733 * `$.plugin.root`.
2734 *
2735 * @param path relative to the working directory, or absolute
2736 * @param options `as`: `"text"` (the default) or `"bytes"`
2737 * @returns the file's text, or `{ base64 }`
2738 * @example
2739 * const readme = await $.fs.read("README.md")
2740 * @example
2741 * const { base64 } = await $.fs.read(
2742 * `${$.plugin.root}/hooks/weights.bin`, { as: "bytes" })
2743 * const weights = Uint8Array.fromBase64(base64)
2744 */
2745 read: FsReadCall;
2746 /**
2747 * Writes `text` to a file, creating it and its directories as needed.
2748 *
2749 * @param path relative to the working directory, or absolute
2750 * @param text the whole new content
2751 */
2752 write: (path: string, text: string) => Promise<void>;
2753 /**
2754 * Lists a directory: `{ name, kind, size, isLink }` per entry, by name,
2755 * each entry as it stands (a symbolic link is `other` with `isLink`).
2756 *
2757 * @param path the directory's path; absent, the working directory
2758 * @returns the entries, `{ name, kind, size, isLink }` each
2759 */
2760 list: (path?: string) => Promise<FsEntry[]>;
2761 /**
2762 * Returns whether the path exists; rejects only a network location as
2763 * spelled, which no `fs` call reaches.
2764 */
2765 exists: (path: string) => Promise<boolean>;
2766 /**
2767 * Returns `{ kind, size, mtimeMs, isLink }` of the path: what it leads to,
2768 * and whether it is itself a symbolic link. Rejects when missing.
2769 *
2770 * With `{ resolve: true }` also `realPath`, where the path lands, absent
2771 * when it leads nowhere; a guard matches it and denies without it, since a
2772 * spelling it cannot resolve (`~`, a file not there yet) the tool may open.
2773 *
2774 * @param path relative to the working directory, or absolute
2775 * @param options `resolve`: also answer `realPath` (one more file system
2776 * call)
2777 * @returns the stat, `realPath` with it when asked and resolvable; rejects
2778 * `ENOENT` for a missing path
2779 * @example
2780 * // ROOT was resolved the same way and SEP is its separator; an
2781 * // allow-list under it is the robust guard, a deny-list on spellings
2782 * // only best effort (a hard link or a case alias keeps its own)
2783 * on("tool.call", { tool: "Read" }, async ($, e, next) => {
2784 * const stat = await $.fs.stat(e.file_path, { resolve: true })
2785 * .catch(() => undefined)
2786 * const real = stat?.realPath
2787 * const isInside = real !== undefined && real.startsWith(ROOT + SEP)
2788 * return isInside ? next(e) : { deny: "outside the project" }
2789 * })
2790 * @example
2791 * // a Write may name a file not there yet: `placed` answers where the
2792 * // path lands or undefined, and the guard denies on undefined or
2793 * // outside ROOT. Unplaceable by spelling first, with no file system
2794 * // call (drive-relative `D:x`, a `\\` or `//` network or device path,
2795 * // a name that is empty, ".", ".." or itself `C:...`); then the file
2796 * // if it stats; else its folder, cut after the last separator and
2797 * // keeping it so a drive or share root stays that root, plus the name
2798 * const placed = async (path) => {
2799 * const cut = Math.max(path.lastIndexOf("/"), path.lastIndexOf("\\"))
2800 * const name = path.slice(cut + 1)
2801 * const isPlaceable = !/^[A-Za-z]:(?![\\/])/.test(path) &&
2802 * !/^[\\/][\\/]/.test(path) && !/^[A-Za-z]:/.test(name) &&
2803 * name !== "" && name !== "." && name !== ".."
2804 * if (!isPlaceable) return undefined
2805 * const own = await $.fs.stat(path, { resolve: true })
2806 * .catch(() => undefined)
2807 * if (own) return own.realPath
2808 * const folder = cut < 0 ? "." : path.slice(0, cut + 1)
2809 * const dir = await $.fs.stat(folder, { resolve: true })
2810 * .catch(() => undefined)
2811 * return dir?.realPath === undefined ? undefined
2812 * : `${dir.realPath.replace(/[\\/]$/, "")}${SEP}${name}`
2813 * }
2814 * const real = await placed(e.file_path)
2815 * const isInside = real !== undefined && real.startsWith(ROOT + SEP)
2816 * return isInside ? next(e) : { deny: "cannot place it, or outside" }
2817 */
2818 stat: (path: string, options?: FsStatOptions) => Promise<FsStat>;
2819 /**
2820 * Reads the named instruction files in every directory above the
2821 * session's original working directory, the way the engine reads CLAUDE.md.
2822 *
2823 * Root first, each `{ dir, name, content }` that exists, the content
2824 * with its `@include`s after it; `of` and `below` together are the
2825 * engine's own walk for a nested CLAUDE.md between root and read file.
2826 *
2827 * @param request `names`, relative `.md` file names (no `..`); `of`, the
2828 * file walked down to; `below`, the directory the walk stays inside
2829 * @returns the files found, root first
2830 * @example
2831 * const found = await $.fs.ancestors({ names: ["AGENTS.md"] })
2832 * const stack = await $.fs.ancestors({ names: ["AGENTS.md"], of: path })
2833 * const nested = await $.fs.ancestors({
2834 * names: ["AGENTS.md"],
2835 * of: e.file_path,
2836 * below: await $.session.root(),
2837 * })
2838 */
2839 ancestors: (request: FsAncestorsRequest) => Promise<readonly FsAncestor[]>;
2840 };
2841 /**
2842 * This plugin's own key-value store, kept between sessions and hot
2843 * reloads; values are JSON data.
2844 *
2845 * A JSON file of the plugin's own under the user's Claude Code
2846 * configuration directory.
2847 */
2848 store: {
2849 /**
2850 * Returns the value under `key`, or `undefined` when unset.
2851 *
2852 * @example
2853 * const count = Number((await $.store.get("count")) ?? 0) + 1
2854 */
2855 get: (key: string) => Promise<unknown>;
2856 /**
2857 * Sets `key` to `value`, which must be JSON data.
2858 *
2859 * `get` reads back `JSON.parse(JSON.stringify(value))`: a Date is its ISO
2860 * string, an `undefined` field is dropped, a Map or Set is `{}`. Rejects
2861 * a function, a cycle, or a store over 4 MiB of JSON text in all.
2862 */
2863 set: (key: string, value: unknown) => Promise<void>;
2864 /**
2865 * Removes `key` from the store.
2866 */
2867 delete: (key: string) => Promise<void>;
2868 /**
2869 * Returns every key set, in insertion order.
2870 */
2871 keys: () => Promise<string[]>;
2872 };
2873 /**
2874 * The time and timers, each an event through the host: `clock.now` reads
2875 * the time; `clock.sleep`, `after` and `every` wait until it has passed.
2876 *
2877 * A timer's callback is the plugin's own function, kept in its environment
2878 * and run there when the wait resolves; a hot reload of the plugin cancels
2879 * its pending waits with the old environment.
2880 */
2881 clock: {
2882 /**
2883 * Resolves milliseconds since the epoch, now.
2884 *
2885 * @example
2886 * const startedAt = await $.clock.now()
2887 */
2888 now: () => Promise<number>;
2889 /**
2890 * Resolves after `ms` milliseconds; rejects at once when `signal` aborts.
2891 *
2892 * @param ms how long, in milliseconds
2893 * @param options `signal`: ends the wait early with a rejection (pass
2894 * `next.signal` so a hook's wait ends with its dispatch)
2895 * @example
2896 * await $.clock.sleep(500, { signal: next.signal })
2897 */
2898 sleep: (ms: number, options?: SleepOptions) => Promise<void>;
2899 /**
2900 * Calls `fn` once after `ms` milliseconds; `cancel()` before then stops it.
2901 *
2902 * One `clock.after` dispatch: `fn` runs when it resolves, and never when
2903 * a hook refuses it.
2904 */
2905 after: TimerCall;
2906 /**
2907 * Calls `fn` every `ms` milliseconds (at least 1) until `cancel()`.
2908 *
2909 * One `clock.every` dispatch per period: `fn` runs when it resolves and
2910 * the next period is asked; a refused period ends the interval.
2911 *
2912 * @example
2913 * const tick = $.clock.every(1000, () => $.ui.status("polling"))
2914 */
2915 every: TimerCall;
2916 };
2917 /**
2918 * The network, through the host.
2919 */
2920 http: {
2921 /**
2922 * Fetches `url` through the host (never the plugin's own network) and
2923 * resolves `{ status, ok, headers, text }` once the body is read.
2924 *
2925 * http or https, to whatever the host process can reach, unless the
2926 * administrator's policy switches refuse it; an `auth` handle from
2927 * `$.session.authorize()` rides https only, to a first-party host.
2928 *
2929 * @param url the URL (http or https)
2930 * @param init `{ method, headers, body, auth, socketPath }` (body a
2931 * string; socketPath a Unix socket to go over instead of TCP)
2932 * @returns `{ status, ok, headers, text }` once the body is read
2933 * @example
2934 * const { ok, text } = await $.http.fetch("https://example.com/status")
2935 * @example
2936 * await $.http.fetch("http://bridge/reload", {
2937 * method: "POST",
2938 * socketPath: `${runDirectory}/bridge.sock`,
2939 * })
2940 */
2941 fetch: (url: string, init?: HttpInit) => Promise<HttpResponse>;
2942 };
2943 /**
2944 * Commands on the host, run as the user the session runs as. CLI only.
2945 *
2946 * Local execution, not a network path: what a command of its own reaches
2947 * is its own, as for the Bash tool and a settings `command` hook.
2948 */
2949 process: {
2950 /**
2951 * Runs a command on the host by its argument vector (no shell) and
2952 * resolves `{ exitCode, stdout, stderr }` once it exits, any exit code.
2953 *
2954 * One shot: the whole output is read, so a background process left
2955 * writing holds the call until the timeout. Rejects when the command
2956 * cannot start or is still running then. Git runs with repo hooks off.
2957 *
2958 * @param argv the command and its arguments, `argv[0]` the executable
2959 * @param init `{ cwd, env, stdin, timeoutMs }` (cwd the session's by
2960 * default; timeout 30 s by default, ten minutes at most)
2961 * @returns `{ exitCode, stdout, stderr }` once the child exits
2962 * @example
2963 * const { exitCode, stdout } = await $.process.run(["git", "status"])
2964 */
2965 run: (argv: readonly string[], init?: ProcessRunInit) => Promise<ProcessRunResult>;
2966 };
2967 /**
2968 * What the settings files, `--settings` and managed policy hold, as the
2969 * engine runs under it; read only.
2970 *
2971 * Every key crosses as the source holds it, `env` and the helper commands
2972 * included: nothing is filtered. The OAuth session and the global config
2973 * (~/.claude.json) are not settings and are never read.
2974 */
2975 settings: {
2976 /**
2977 * Resolves with the settings merged over every source, as the engine
2978 * reads them, or with one source's settings as loaded (`{ source }`).
2979 *
2980 * A snapshot in plain data each call; a source with no file answers
2981 * `{}`. The sources (SettingsSource) rise in precedence from `user` to
2982 * `policy`: the merge takes a key from the last source that has it.
2983 *
2984 * @param args `{ source }` to read one source; nothing for the merge
2985 * @returns the settings object, keyed as a settings.json is
2986 * @example
2987 * const { permissions } = await $.settings.read()
2988 * const policy = await $.settings.read({ source: "policy" })
2989 */
2990 read: (args?: SettingsReadArgs) => Promise<Settings>;
2991 };
2992 /**
2993 * The environment of this process, the one every Bash child, MCP server
2994 * and `$.process.run` command started after inherits.
2995 *
2996 * `get` and `set` take the variable's name as a string literal, so what a
2997 * module reads and writes is read off its source: `claude plugin validate`
2998 * lists the names, and a name the module does not spell is refused.
2999 */
3000 env: {
3001 /**
3002 * Resolves with the variable's value, or `undefined` when it is unset.
3003 *
3004 * `name` must be a string literal; `claude plugin validate` lists the
3005 * names your module reads and writes.
3006 *
3007 * @example
3008 * const home = await $.env.get("HOME")
3009 */
3010 get: (name: string) => Promise<string | undefined>;
3011 /**
3012 * Sets the variable for this process and everything it starts after, or
3013 * unsets it when `value` is `undefined`.
3014 *
3015 * `name` must be a string literal; `claude plugin validate` lists the
3016 * names your module reads and writes.
3017 *
3018 * @example
3019 * await $.env.set("GIT_PAGER", "cat")
3020 */
3021 set: (name: string, value: string | undefined) => Promise<void>;
3022 };
3023 }
3024
3025 /**
3026 * The name of an event the engine defines itself (a key of CoreEventOf);
3027 * EventName adds the declared plugin nouns' events.
3028 */
3029 type CoreEventName = keyof CoreEventOf;
3030
3031 /**
3032 * The argument of each event the engine defines itself: its call sites'
3033 * (EngineEventOf), the classic hooks' (ClassicEventOf), the calls on `$`.
3034 */
3035 type CoreEventOf = EngineEventOf & ClassicEventOf & OpEventOf;
3036
3037 type CwdChangedHookInput = BaseHookInput & {
3038 hook_event_name: 'CwdChanged';
3039 old_cwd: string;
3040 new_cwd: string;
3041 };
3042
3043 type DirectoryAddedHookInput = BaseHookInput & {
3044 hook_event_name: 'DirectoryAdded';
3045 /**
3046 * Absolute path of the directory that was added.
3047 */
3048 directory: string;
3049 /**
3050 * How the directory was added: "slash_command" for /add-dir, "register_repo_root" for the SDK control_request.
3051 */
3052 source: 'slash_command' | 'register_repo_root';
3053 };
3054
3055 /**
3056 * The `children` field every element constructor's props carry, appended
3057 * beside its own props type: one child, or a list that may nest.
3058 *
3059 * JSX types a lone child as the child itself (`<Text dimColor>done</Text>`
3060 * passes the string, `<Text>{count}</Text>` the number), and a mapped list
3061 * beside a sibling as a nested list (`<Box>{rows.map(row)}<Text>ok</Text>`).
3062 */
3063 export type ElementChildren = {
3064 children?: RenderChildren;
3065 };
3066
3067 /**
3068 * An element as `$.ui.resolve(e)` hands it out: a constructor from props to
3069 * the frozen plain-data element, `children` among the props as JSX passes.
3070 *
3071 * `const { Box } = $.ui.resolve(e)` then `<Box gap={1}>...</Box>` compiles to
3072 * `h(Box, { gap: 1 }, ...children)`, and `h` calls a function tag with its
3073 * props, so the table's constructors are the JSX tags.
3074 */
3075 export type ElementConstructor<P> = (props: P & ElementChildren) => RenderElement;
3076
3077 /**
3078 * Every element name of every surface: what a table handed out is completed to
3079 * (an omitted one draws a fragment; see `ui.resolve`).
3080 */
3081 export type ElementName = {
3082 [P in RenderSurface]: keyof Elements[P];
3083 }[RenderSurface];
3084
3085 /**
3086 * The element constructors each surface draws, by `e.surface`: what
3087 * `$.ui.resolve(e)` returns and a `ui.resolve` hook passes on; no globals.
3088 *
3089 * All carry `Box`, `Text`, `Button`, `Link`, `Code`, `Markdown`; every remote
3090 * surface `Svg`; all but mobile `Input` and `Select`; terminal and desktop
3091 * `Client`; terminal `Raster` and `Image`. Narrowed on `e.surface`, that table.
3092 */
3093 export type Elements = {
3094 terminal: {
3095 Box: ElementConstructor<BoxProps>;
3096 Text: ElementConstructor<TextProps>;
3097 Button: ElementConstructor<ButtonProps>;
3098 Input: ElementConstructor<InputProps>;
3099 Select: ElementConstructor<SelectProps>;
3100 Link: ElementConstructor<LinkProps>;
3101 Code: ElementConstructor<CodeProps>;
3102 Markdown: ElementConstructor<MarkdownProps>;
3103 /**
3104 * A region one of the plugin's surface modules draws (ClientModule), named
3105 * by `module`: a string literal, the module's path relative to this file.
3106 *
3107 * The engine reads the surface module off the hooks module's source
3108 * before anything runs, so a variable, a template with a substitution or
3109 * a computed path as `module` is refused at load, naming the line.
3110 */
3111 Client: ElementConstructor<ClientProps>;
3112 Raster: ElementConstructor<RasterProps>;
3113 Image: ElementConstructor<ImageProps>;
3114 };
3115 desktop: {
3116 Box: ElementConstructor<BoxProps>;
3117 Text: ElementConstructor<TextProps>;
3118 Button: ElementConstructor<ButtonProps>;
3119 Input: ElementConstructor<InputProps>;
3120 Select: ElementConstructor<SelectProps>;
3121 Svg: ElementConstructor<SvgProps>;
3122 Link: ElementConstructor<LinkProps>;
3123 Code: ElementConstructor<CodeProps>;
3124 Markdown: ElementConstructor<MarkdownProps>;
3125 Client: ElementConstructor<ClientProps>;
3126 };
3127 /**
3128 * No `Input` or `Select`: the control protocol carries presses (ui_press)
3129 * but no ui_input or ui_select yet; not a limit of the device.
3130 *
3131 * The table grows when those messages exist.
3132 */
3133 mobile: {
3134 Box: ElementConstructor<BoxProps>;
3135 Text: ElementConstructor<TextProps>;
3136 Button: ElementConstructor<ButtonProps>;
3137 Svg: ElementConstructor<SvgProps>;
3138 Link: ElementConstructor<LinkProps>;
3139 Code: ElementConstructor<CodeProps>;
3140 Markdown: ElementConstructor<MarkdownProps>;
3141 };
3142 /**
3143 * The desktop's table without `Client`: a remote `Client`'s module, presses
3144 * and posts (ui_client_module, ui_client_press, ui_message) name no surface.
3145 *
3146 * They are the desktop's alone today, not a limit of the editor's webview:
3147 * the table gains `Client` when those asks name a surface.
3148 */
3149 vscode: {
3150 Box: ElementConstructor<BoxProps>;
3151 Text: ElementConstructor<TextProps>;
3152 Button: ElementConstructor<ButtonProps>;
3153 Input: ElementConstructor<InputProps>;
3154 Select: ElementConstructor<SelectProps>;
3155 Svg: ElementConstructor<SvgProps>;
3156 Link: ElementConstructor<LinkProps>;
3157 Code: ElementConstructor<CodeProps>;
3158 Markdown: ElementConstructor<MarkdownProps>;
3159 };
3160 };
3161
3162 /**
3163 * The table `ui.resolve` answers for an argument of surface `P`.
3164 */
3165 export type ElementTable<P extends RenderSurface = RenderSurface> = Elements[P];
3166
3167 /**
3168 * Hook input for the Elicitation event. Fired when an MCP server requests user input. Hooks can auto-respond (accept/decline) instead of showing the dialog.
3169 */
3170 type ElicitationHookInput = BaseHookInput & {
3171 hook_event_name: 'Elicitation';
3172 mcp_server_name: string;
3173 message: string;
3174 mode?: 'form' | 'url';
3175 url?: string;
3176 elicitation_id?: string;
3177 requested_schema?: Record<string, unknown>;
3178 };
3179
3180 /**
3181 * Hook input for the ElicitationResult event. Fired after the user responds to an MCP elicitation. Hooks can observe or override the response before it is sent to the server.
3182 */
3183 type ElicitationResultHookInput = BaseHookInput & {
3184 hook_event_name: 'ElicitationResult';
3185 mcp_server_name: string;
3186 elicitation_id?: string;
3187 mode?: 'form' | 'url';
3188 action: 'accept' | 'decline' | 'cancel';
3189 content?: Record<string, unknown>;
3190 };
3191
3192 /**
3193 * The input of `engine.create`: the fold that builds `$`, once per load,
3194 * core innermost.
3195 *
3196 * A hook is written in post-order: `const built = await next(e)` is `$` as
3197 * built so far; `return { ...built, voice: { say } }` adds this plugin's noun.
3198 * See EngineEventOf's `engine.create` for what a step may and may not do.
3199 */
3200 export type EngineCreateInput = {
3201 /**
3202 * The modules this fold builds, in list order, first outermost (managed
3203 * plugins first, so an org plugin's withholding wins).
3204 *
3205 * The whole set at a load, this plugin alone at its reload, the ones added
3206 * or changed at a refresh that keeps the rest: a module added later that
3207 * does not hook this event is built against the table as it stands.
3208 */
3209 plugins: readonly string[];
3210 };
3211
3212 /**
3213 * What an `engine.create` hook returns: `$` as built so far with this
3214 * plugin's nouns added, less any it withheld.
3215 *
3216 * Every declared noun is optional here. Between hooks it crosses the chain
3217 * as interface descriptors; a hook sees objects (EngineInterfaceBuilt).
3218 */
3219 export type EngineCreateResult = Partial<EngineInterface> & {
3220 readonly [noun: string]: unknown;
3221 };
3222
3223 /**
3224 * The events the engine raises at its call sites, and `engine.create`; the
3225 * classic settings hooks' events are ClassicEventOf.
3226 *
3227 * At every one, a hook that fails (throws, overruns its budget: HookBudget,
3228 * answers a wrong shape) is skipped: the hooks beneath and core run in its
3229 * place, or its last `next` result stands; the failure is reported by name.
3230 */
3231 export type EngineEventOf = {
3232 /**
3233 * Fires when the engine is about to run a tool. `next(e)` runs the hooks
3234 * beneath, then core (the permission prompt, the tool itself).
3235 *
3236 * Return `{ deny: reason }` to refuse or `{ result }` to answer yourself; a
3237 * hook that returns while its `next` is pending aborts what runs beneath.
3238 * The managed-settings hooks run first: their deny is the call's result.
3239 */
3240 'tool.call': ToolCallInput;
3241 /**
3242 * Fires when the engine decides whether a tool call may run, after the
3243 * `tool.call` and PreToolUse hooks and before the mode settles an ask.
3244 *
3245 * `next(e)` resolves to the engine's verdict (rules, mode, the tool's own
3246 * check, PreToolUse's decision); return any `{ decision }`. `$.tool.check`
3247 * runs the same chain and executes nothing.
3248 *
3249 * @example
3250 * on("tool.check", { tool: "Read" }, () => ({ decision: "allow" }))
3251 */
3252 'tool.check': ToolCheckInput;
3253 /**
3254 * Fires when the engine is about to draw a component: once per input value
3255 * (props, viewport width), plugin load or `$.ui.invalidate("ui.render")`.
3256 *
3257 * A repaint reuses the answer; a clock invalidates. `next(e)` resolves to the
3258 * drawing: return it, wrap it, draw your own, or rewrite `props`. A tree that
3259 * does not validate draws the engine's own; `--plugin-dir` is told why.
3260 */
3261 'ui.render': RenderInput;
3262 /**
3263 * Fires when the plugins load (not per draw), once per surface, component
3264 * and plugin: `e` names the surface and component, never the props.
3265 *
3266 * `next(e)` resolves to the surface's table, which `$.ui.resolve(e)` then
3267 * reads. Return it, one with an element restyled for every other plugin,
3268 * or one with a key left out (a fragment there); own table: hook skipped.
3269 */
3270 'ui.resolve': ResolveInput;
3271 /**
3272 * Fires when a `Button` a render hook drew is pressed on a surface; `e` is
3273 * `{ plugin, element, component, surface }`, `element` the button's `key`.
3274 *
3275 * `next(e)` runs the hooks beneath, then core: the element's own `onPress`
3276 * closure, in its plugin's environment, resolving to `{ element }`. Return
3277 * `next(e)` to let the press through, or `{ element }` to take it.
3278 */
3279 'ui.press': UiPressArgument;
3280 /**
3281 * Fires when an `Input` a render hook drew changes or is submitted; `e` is
3282 * `{ plugin, element, component, surface, kind, value }`.
3283 *
3284 * `next(e)` runs the hooks beneath, then the element's own `onInput` or
3285 * `onSubmit` with `e.value` as the chain left it, resolving to `{ element,
3286 * value }`; `next({ ...e, value })` rewrites the typing, an answer takes it.
3287 */
3288 'ui.input': UiInputArgument;
3289 /**
3290 * Fires when a `Select` a render hook drew is picked from; `e` is
3291 * `{ plugin, element, component, surface, value }`.
3292 *
3293 * `next(e)` runs the hooks beneath, then the element's own `onSelect` with
3294 * `e.value` as the chain left it, resolving to `{ element, value }`;
3295 * `next({ ...e, value })` rewrites the pick, an answer takes it.
3296 */
3297 'ui.select': UiSelectArgument;
3298 /**
3299 * Fires when a `Client` THIS plugin drew posts from its surface module
3300 * (`surface.post(data)`); only this plugin's hooks see it.
3301 *
3302 * `next.origin` names `client`: `data` came from code. Core answers `{}`;
3303 * `next({ ...e, data })` rewrites the data; `{ props }` hands the posting
3304 * instance its next props without a redraw. One per instance per frame.
3305 */
3306 'ui.message': UiMessageArgument;
3307 /**
3308 * Fires before a site's window moves: the person's wheel or scroll keys on
3309 * a `Pane` body or the `AbovePrompt` band, at its edges too; `$.ui.scroll`.
3310 *
3311 * `next(e)` moves it to `e.offset` and draws: `{}`; `next({ ...e, offset
3312 * })` elsewhere; no `next` (`{}` or `{ deny }`) leaves it undrawn, so a hook
3313 * drawing its own rows under a header moves them by `e.by` and invalidates.
3314 *
3315 * @example
3316 * on("ui.scroll", { requestId: "log" }, ($, e) => (scrollOwnRows(e.by), {}))
3317 */
3318 'ui.scroll': UiScrollInput;
3319 /**
3320 * Fires before a site's focus ring moves: the person's Tab, arrows or click
3321 * in a `Pane` or the band; an `autoFocus` element taking it; `$.ui.focus`.
3322 *
3323 * `next(e)` lands it on `e.element` (absent: one of the engine's stops) and
3324 * draws: `{}`; `next({ ...e, element })` on another of `e.plugin`'s; no
3325 * `next` (`{}` or `{ deny }`) keeps it where it was, drawn as it was.
3326 *
3327 * @example
3328 * on("ui.focus", { requestId: "list" }, ($, e, next) => (mark(e), next(e)))
3329 */
3330 'ui.focus': UiFocusInput;
3331 /**
3332 * Fires when the engine offers an agent type to the model, in the agent
3333 * listing and again at dispatch; `next(e)` resolves to `{ isOffered: true }`.
3334 *
3335 * Return `{ isOffered: false }` to keep the type out of the listing and
3336 * refuse its dispatch. A hook that fails passes it through.
3337 *
3338 * @example
3339 * on("agent.offer", { agent: "Plan" }, () => ({ isOffered: false }))
3340 */
3341 'agent.offer': AgentOfferInput;
3342 /**
3343 * Fires when the Agent tool is about to start a subagent, everything
3344 * decided and its model not yet resolved.
3345 *
3346 * `next(e)` resolves to `{ model }`. Return it, `next({ ...e, model })`,
3347 * `{ model }` of your own (an alias resolves like the tool's parameter), or
3348 * `{ deny: reason }`.
3349 */
3350 'agent.spawn': AgentSpawnInput;
3351 /**
3352 * Fires when a prompt is submitted, before the turn starts. `next(e)` runs
3353 * the hooks beneath and the UserPromptSubmit settings hooks.
3354 *
3355 * Rewrite with `next({ ...e, text })` (the user message on screen follows)
3356 * or stop it with `{ drop: reason }`; a broken plugin never blocks a prompt.
3357 * A prompt typed while a turn ran fires at Enter, with that turn's id.
3358 */
3359 'prompt.submit': PromptSubmitInput;
3360 /**
3361 * Fires when a text is about to be put in the prompt box as the person's
3362 * draft (a plugin's `$.prompt.fill`); `next(e)` writes it by `e.mode`.
3363 *
3364 * `replace` over the draft, `append` after it, `insert` at the cursor;
3365 * rewrite `text` or `mode` going down, or answer `{ isFilled: false }`
3366 * without `next` to keep it out, as core does under a dialog or headless.
3367 *
3368 * @example
3369 * on("prompt.fill", ($, e, next) => next({ ...e, text: e.text.trim() }))
3370 */
3371 'prompt.fill': PromptFillInput;
3372 /**
3373 * Fires when a text is proposed as the prompt box's dim suggestion, Tab to
3374 * take: the engine's guess after a turn, or a plugin's `$.prompt.suggest`.
3375 *
3376 * `next(e)` shows it: `{ isShown }`. Rewrite with `next({ ...e, text })`,
3377 * or answer `{ isShown: false }` without `next` to drop it; core answers
3378 * that too while the box holds text or a turn runs.
3379 *
3380 * @example
3381 * on("prompt.suggest", { origin: { kind: "suggestion" } }, hide)
3382 */
3383 'prompt.suggest': PromptSuggestInput;
3384 /**
3385 * Fires when the person edits the main prompt box: a key the editor took as
3386 * an edit, or a paste; `next(e)` resolves the box the editor shows.
3387 *
3388 * `e` is the draft before and the splice (`start`, `end`, `inputText`); a
3389 * burst of keys is one edit. Rewrite `inputText` going down or the box
3390 * coming up; `{ text: e.text, cursor: e.cursor }` without `next` consumes.
3391 *
3392 * @example
3393 * on("prompt.edit", ($, e, next) => next({ ...e, inputText: up(e) }))
3394 */
3395 'prompt.edit': PromptEditInput;
3396 /**
3397 * Fires once per named section of the system prompt, when the engine
3398 * assembles it; `next(e)` resolves to `{ text }` as core computed it.
3399 *
3400 * Sections are cached by name for the session until
3401 * `$.ui.invalidate("prompt.section")`: an unstable answer spends the
3402 * model's prompt cache on every call. A hook that fails passes it through.
3403 *
3404 * @example
3405 * on("prompt.section", { name: "memory" }, () => ({ text: null }))
3406 */
3407 'prompt.section': PromptSectionInput;
3408 /**
3409 * Fires once per conversation, when the engine computes the context blocks
3410 * its first user message carries; `next(e)` resolves to `{ blocks }`.
3411 *
3412 * Append, drop, reorder or rewrite with `next({ ...e, blocks })`; the
3413 * engine renders what comes back, in order, until
3414 * `$.ui.invalidate("prompt.context")` or a re-read (compaction, `/clear`).
3415 *
3416 * @example
3417 * on("prompt.context", () => ({ blocks: [] }))
3418 */
3419 'prompt.context': PromptContextInput;
3420 /**
3421 * Fires when the engine renders a system prompt; `next(e)` resolves to
3422 * `{ sections }`, each `{ id, text, scope }`, in the order they are sent.
3423 *
3424 * The engine answers no sections: the plugins hooked here are the prompt.
3425 * Append, replace by id, reorder or drop what `next(e)` answered; `e` is
3426 * the facts to compose from. The engine joins, and places every cache mark.
3427 *
3428 * @example
3429 * on("prompt.compose", () => ({ sections: [INTRO, TOOLS, PLACE] }))
3430 */
3431 'prompt.compose': PromptComposeInput;
3432 /**
3433 * Fires once per message the engine injects for the model on its own (a
3434 * reminder, a mode transition, a mentioned file), as a request carries it.
3435 *
3436 * `next(e)` resolves to `{ text }`; `{ text: null }` leaves it out. The
3437 * answer holds per attachment for the process (asked again on resume or
3438 * `$.ui.invalidate`); the transcript keeps the engine's record.
3439 *
3440 * @example
3441 * on("prompt.attachment", { type: "todo_reminder" }, () => ({ text: null }))
3442 */
3443 'prompt.attachment': PromptAttachmentInput;
3444 /**
3445 * Fires once per tool, when the engine first renders the tool's schema in
3446 * a session; `next(e)` resolves to `{ description, isDeferred? }`.
3447 *
3448 * Cached for the session until `$.ui.invalidate("tool.describe")`: an
3449 * unstable answer spends the model's prompt cache. An explicit `isDeferred`
3450 * moves the tool behind ToolSearch (true) or into the prompt's list (false).
3451 *
3452 * @example
3453 * on("tool.describe", { tool: "Bash" }, ($, e) => ({ ...e, description }))
3454 * @example
3455 * on("tool.describe", { tool: "Monitor" }, pin) // {...e, isDeferred: false}
3456 */
3457 'tool.describe': ToolDescribeInput;
3458 /**
3459 * Fires when a slash command is about to run (`/name args` typed, or a
3460 * plugin's `$.command.run`); `next(e)` resolves to `{ text }`, its output.
3461 *
3462 * Core is the engine's command (a registered one has none). Rewrite `args`
3463 * with `next`, or return `{ text }` without it to answer in its place; one
3464 * after `next` replaces a printed output, not a panel or prompt it opened.
3465 *
3466 * @example
3467 * on("command.run", { command: "hello" }, () => ({ text: "hello" }))
3468 */
3469 'command.run': CommandRunInput;
3470 /**
3471 * Fires once per command, when the engine lists it for the typeahead and
3472 * `/help`; `next(e)` resolves to `{ description, argumentHint, isHidden }`.
3473 *
3474 * Listed answers are cached for the session until
3475 * `$.ui.invalidate("command.describe")`. A hook that fails passes it
3476 * through.
3477 *
3478 * @example
3479 * on("command.describe", ($, e, next) => next({ ...e, isHidden: true }))
3480 */
3481 'command.describe': CommandDescribeInput;
3482 /**
3483 * Fires when a `/config` row is about to change, from the menu or a
3484 * plugin's `$.config.set`; `next(e)` resolves to `{ value }` once written.
3485 *
3486 * Return `{ deny: reason }` to leave the row as it is (the menu says why),
3487 * or `next({ ...e, value })` to clamp it; a value of the wrong kind for
3488 * the row is refused. A row a trusted source owns is core's to refuse.
3489 *
3490 * @example
3491 * on("config.set", { key: "theme" }, () => ({ deny: "the theme stays" }))
3492 */
3493 'config.set': ConfigSetInput;
3494 /**
3495 * Fires once per `/config` row, when the menu lists it and for
3496 * `$.config.list`; `next(e)` resolves to `{ label, description, isHidden }`.
3497 *
3498 * Relabel, re-describe or hide with `next({ ...e, isHidden: true })`; the
3499 * answers are cached until `$.ui.invalidate("config.describe")` or the
3500 * loaded plugins change. A hook that fails passes the row through.
3501 *
3502 * @example
3503 * on("config.describe", { key: "tips" }, hide) // answers isHidden: true
3504 */
3505 'config.describe': ConfigDescribeInput;
3506 /**
3507 * Fires when the engine expands a skill's prompt for the model (`/name`,
3508 * the Skill tool, a preload); `next(e)` resolves to `{ text }` as computed.
3509 *
3510 * Return `{ text }` with the text the model reads instead. A hook that
3511 * fails passes it through.
3512 *
3513 * @example
3514 * on("skill.prompt", { skill: "commit" }, () => ({ text: "A haiku." }))
3515 */
3516 'skill.prompt': SkillPromptInput;
3517 /**
3518 * Fires when the engine composes a git text the model is to write (`kind`:
3519 * `commit`, `pr`, `exemption`, `remedy`); `next(e)` resolves to `{ text }`.
3520 *
3521 * Return `{ text }` with the text the model reads instead. A hook that
3522 * fails passes it through.
3523 *
3524 * @example
3525 * on("attribution.text", { kind: "commit" }, () => ({ text: "" }))
3526 */
3527 'attribution.text': AttributionTextInput;
3528 /**
3529 * Fires once per process for each loaded plugin, before the first prompt,
3530 * then once per fresh load of one (never `/clear`); `next(e)` is `{ cwd }`.
3531 *
3532 * Observe. The first is awaited: a `$.tool.register` is listed by turn one.
3533 * A later one runs its hooks alone: an enable, a worker respawn, or a reload
3534 * (changed modules only; all if one hooks `engine.create`/`plugin.register`).
3535 *
3536 * @example
3537 * on("session.start", ($, e, next) => $.tool.register(t).then(() => next(e)))
3538 */
3539 'session.start': SessionStartInput;
3540 /**
3541 * Fires when a delivery reaches the session (a relay's event, a peer's
3542 * message, a Remote Control prompt), before it is queued; `{ text }`.
3543 *
3544 * Rewrite with `next({ ...e, text })`, or return `{ consumed: reason }` to
3545 * take it: nothing is queued, shown or read by the model. `origin` and
3546 * `event` pass on as received; `session.send`, its dual, is reserved.
3547 *
3548 * @example
3549 * on("session.receive", { origin: "peer" }, () => ({ consumed: "muted" }))
3550 */
3551 'session.receive': SessionReceiveInput;
3552 /**
3553 * Fires when the conversation is about to be compacted (`/compact`, the
3554 * threshold, a plugin, or ahead of time); `next(e)` resolves `{ messages }`.
3555 *
3556 * Rewrite `instructions` or `messages` on the way down, the messages on
3557 * the way up, or answer `{ messages }` of your own; `{ skip: reason }`
3558 * leaves the conversation as it is. `trigger` passes on as received.
3559 *
3560 * @example
3561 * on("session.compact", { trigger: "precompute" }, () => ({ skip: "off" }))
3562 */
3563 'session.compact': SessionCompactInput;
3564 /**
3565 * Fires when a remote client joins the session's roster of attached
3566 * surfaces: it said so (ui_attach), or it first asked to draw.
3567 *
3568 * Observe (a phone joined: draw the lobby); `next(e)` resolves to
3569 * `{ clientId }`, a different return changes nothing. `$.session.surfaces()`
3570 * reads the roster; a render hook still reads `e.surface` per ask.
3571 *
3572 * @example
3573 * on("session.attach", { surface: "mobile" }, ($, e, next) => next(e))
3574 */
3575 'session.attach': SessionAttachInput;
3576 /**
3577 * Fires when a client leaves the roster: it detached, or the session ended
3578 * with it attached (`e.reason`). Observe; `next(e)` echoes `{ clientId }`.
3579 */
3580 'session.detach': SessionDetachInput;
3581 /**
3582 * Fires when the engine measures the session and a unit moved: after each
3583 * main-thread turn, and when a rate-limit window moves a whole point.
3584 *
3585 * Observe; `next(e)` echoes `{ changed }`. `$.session.usage()`'s figures,
3586 * pushed, not polled: compare them with your own threshold here, call the
3587 * op for the breakdown. One at a time, a burst folding into one more.
3588 *
3589 * @example
3590 * on("session.measure", ($, e, next) => (toastPast90(e.rateLimits), next(e)))
3591 */
3592 'session.measure': SessionMeasureInput;
3593 /**
3594 * Fires once when the session ends (exit, /clear, resume, logout, signal, a
3595 * `-p` run done), after its SessionEnd settings hooks; `e.reason` says which.
3596 *
3597 * `e.resume.id` is `--resume`'s id; `next(e)` runs the engine's end step for
3598 * the plugins (at an exit the attached clients leave): `{ sessionId }`. The
3599 * SessionEnd bound afresh, 1.5 s by default; a `kill -9` raises nothing.
3600 *
3601 * @example
3602 * on("session.end", async ($, e, next) => (await keep(e.resume), next(e)))
3603 */
3604 'session.end': SessionEndInput;
3605 /**
3606 * Fires once per hooks module about to join the chain, at load (the set
3607 * folded and built, nothing swapped in) and at reload; core allows.
3608 *
3609 * Return `{ refuse: reason }` and it never joins: no hook, no noun, no tool
3610 * of it; the transcript names who refused. Its judges, `$` whole, are the
3611 * plugins admitted before it and the binary's; judge by `tier` and `uses`.
3612 *
3613 * @example
3614 * on("plugin.register", { tier: "user" }, () => ({ refuse: "managed only" }))
3615 */
3616 'plugin.register': PluginRegisterInput;
3617 /**
3618 * Fires when a model turn begins, before its first model call; `next(e)`
3619 * resolves to `{ turnId }`. Observe: a different return changes nothing.
3620 */
3621 'turn.start': TurnStartInput;
3622 /**
3623 * Fires when the engine is about to send a model request of a turn, main's
3624 * or a subagent's (`e.agentId`); `next(e)` resolves to the whole response.
3625 *
3626 * `next({ ...e, model })` or `effort` sends another; the turn, the index and
3627 * the message count are pinned. An answer without `next` sends no request.
3628 * It streams (StreamNext): the hook's budget counts its own code alone.
3629 */
3630 'turn.step': TurnStepInput;
3631 /**
3632 * Fires when a model turn has ended, at the point its duration is reported;
3633 * `next(e)` resolves to `{ text }`, the answer. `e.reason` says why.
3634 *
3635 * Return `{ text }` with a different text to show it beneath the answer (a
3636 * synopsis, a TL;DR line); the transcript's record is never rewritten. A
3637 * hook that fails leaves the answer as it was.
3638 */
3639 'turn.complete': TurnCompleteInput;
3640 /**
3641 * Runs while `$` is being built, once per load or reload of this plugin
3642 * and before any other hook of it; `next(e)` resolves to `$` built so far.
3643 *
3644 * A step may ADD nouns and WITHHOLD nouns (leave one out, or return without
3645 * `next`); it may NOT REPLACE one another step added: the step fails, named
3646 * with both plugins. A step that fails unloads its plugin; `$` is rebuilt.
3647 */
3648 'engine.create': EngineCreateInput;
3649 };
3650
3651 /**
3652 * `$`, the first parameter of every hook. Frozen; core's interface plus
3653 * every noun the plugins' `engine.create` steps added.
3654 *
3655 * Flat, `<noun>.<event>`; it does not carry `on`, since registration happens
3656 * before `$` exists. An interface so a plugin types the noun it provides by
3657 * declaration merging, the way a jQuery plugin types `$.fn`.
3658 *
3659 * @example
3660 * declare module "claude-code" { interface EngineInterface { voice: Voice } }
3661 */
3662 export interface EngineInterface extends CoreEngineInterface {
3663 }
3664
3665 /**
3666 * What `next(e)` resolves to at `engine.create`: `$` as the steps beneath
3667 * built it, typed as `$` is, open to nouns no declaration names yet.
3668 *
3669 * A withheld noun is on it as a stub (a step inside withheld it, or the
3670 * last fold did and this is a reload), and the host refuses an op on one a
3671 * step outside withholds later; a typed module bootstraps at load on it.
3672 */
3673 export type EngineInterfaceBuilt = EngineInterface & {
3674 readonly [noun: string]: unknown;
3675 };
3676
3677 /**
3678 * The engine's events' results.
3679 */
3680 export type EngineResultOf = {
3681 /**
3682 * `{ result, context? }`, `{ deny }`, or core's `{ ref, result }`.
3683 */
3684 'tool.call': ToolCallResult;
3685 /**
3686 * `{ decision, reason?, rule? }`.
3687 */
3688 'tool.check': ToolCheckResult;
3689 /**
3690 * The tree to draw; `{ type: "engine", ref }` is core's own drawing.
3691 */
3692 'ui.render': RenderElement;
3693 /**
3694 * The surface's element table (Elements[e.surface]): constructors from props
3695 * to a RenderElement.
3696 */
3697 'ui.resolve': ElementTable;
3698 /**
3699 * `{ element }`: the element whose handler the press reached.
3700 */
3701 'ui.press': UiPressResult;
3702 /**
3703 * `{ element, value }`: the field whose handler the input reached, and
3704 * the text it received.
3705 */
3706 'ui.input': UiInputResult;
3707 /**
3708 * `{ element, value }`: the picker whose handler the pick reached, and
3709 * the value it received.
3710 */
3711 'ui.select': UiSelectResult;
3712 /**
3713 * `{ props? }`: the posting instance's next props, when a hook hands some.
3714 */
3715 'ui.message': UiMessageResult;
3716 /**
3717 * `{}` once the window moved, or `{ deny }`.
3718 */
3719 'ui.scroll': UiScrollResult;
3720 /**
3721 * `{}` once the ring moved, or `{ deny }`.
3722 */
3723 'ui.focus': UiFocusResult;
3724 /**
3725 * `{ isOffered }`.
3726 */
3727 'agent.offer': AgentOfferResult;
3728 /**
3729 * `{ model }` or `{ deny }`.
3730 */
3731 'agent.spawn': AgentSpawnResult;
3732 /**
3733 * `{ text, context? }` or `{ drop }`.
3734 */
3735 'prompt.submit': PromptSubmitResult;
3736 /**
3737 * `{ isFilled }`.
3738 */
3739 'prompt.fill': PromptFillResult;
3740 /**
3741 * `{ isShown }`.
3742 */
3743 'prompt.suggest': PromptSuggestResult;
3744 /**
3745 * `{ text, cursor }`, the box the editor shows next.
3746 */
3747 'prompt.edit': PromptEditResult;
3748 /**
3749 * `{ text }` (null leaves the section out).
3750 */
3751 'prompt.section': PromptSectionResult;
3752 /**
3753 * `{ blocks }` (a block left out is not sent).
3754 */
3755 'prompt.context': PromptContextResult;
3756 /**
3757 * `{ sections }`, every `shared` one ahead of every `session` one (a
3758 * section left out is not sent).
3759 */
3760 'prompt.compose': PromptComposeResult;
3761 /**
3762 * `{ text }` (null leaves the attachment out).
3763 */
3764 'prompt.attachment': PromptAttachmentResult;
3765 /**
3766 * `{ description, isDeferred? }`.
3767 */
3768 'tool.describe': ToolDescribeResult;
3769 /**
3770 * `{ text }` (the command's output, when it printed one).
3771 */
3772 'command.run': CommandRunResult;
3773 /**
3774 * `{ description, argumentHint, isHidden }`.
3775 */
3776 'command.describe': CommandDescribeResult;
3777 /**
3778 * `{ value }` once written, or `{ deny }`.
3779 */
3780 'config.set': ConfigSetResult;
3781 /**
3782 * `{ label, description, isHidden }`.
3783 */
3784 'config.describe': ConfigDescribeResult;
3785 /**
3786 * `{ text }`.
3787 */
3788 'skill.prompt': SkillPromptResult;
3789 /**
3790 * `{ text }`.
3791 */
3792 'attribution.text': AttributionTextResult;
3793 /**
3794 * `{ cwd }`.
3795 */
3796 'session.start': SessionStartResult;
3797 /**
3798 * `{ text }`, or `{ consumed }`.
3799 */
3800 'session.receive': SessionReceiveResult;
3801 /**
3802 * `{ messages, tokensBefore?, tokensAfter? }`, or `{ skip }`.
3803 */
3804 'session.compact': SessionCompactResult;
3805 /**
3806 * `{ clientId }`.
3807 */
3808 'session.attach': SessionAttachResult;
3809 /**
3810 * `{ clientId }`.
3811 */
3812 'session.detach': SessionDetachResult;
3813 /**
3814 * `{ changed }`.
3815 */
3816 'session.measure': SessionMeasureResult;
3817 /**
3818 * `{ sessionId }`.
3819 */
3820 'session.end': SessionEndResult;
3821 /**
3822 * `{ allow: true }`, or `{ refuse }`.
3823 */
3824 'plugin.register': PluginRegisterResult;
3825 /**
3826 * `{ turnId }`.
3827 */
3828 'turn.start': TurnStartResult;
3829 /**
3830 * The response: `{ turnId, index, answer, toolUses, stopReason, usage }`.
3831 */
3832 'turn.step': TurnStepResult;
3833 /**
3834 * `{ text }`.
3835 */
3836 'turn.complete': TurnCompleteResult;
3837 /**
3838 * `$` as built so far, with this plugin's interface added and any it
3839 * withheld left out; `next(e)` resolves to EngineInterfaceBuilt.
3840 */
3841 'engine.create': EngineCreateResult;
3842 };
3843
3844 /**
3845 * The engine's own events as calls on `$`, one signature each:
3846 * `$.<noun>.<event>(input)` resolves to its result, or to its stream.
3847 *
3848 * The engine raises its events through these same calls; a plugin's call
3849 * runs the same chain with the calling hook alone skipped. `input` may leave
3850 * out what the engine fills (`tool_use_id`, the parent agent).
3851 */
3852 export type EventCalls = {
3853 tool: {
3854 call: ToolCallOverloads;
3855 check: (input: ToolCheckArgs) => Promise<ToolCheckResult>;
3856 describe: (input: ToolDescribeInput) => Promise<ToolDescribeResult>;
3857 };
3858 command: {
3859 run: (input: CommandRunArgs) => Promise<CommandRunResult>;
3860 describe: (input: CommandDescribeInput) => Promise<CommandDescribeResult>;
3861 };
3862 config: {
3863 set: (input: ConfigSetArgs) => Promise<ConfigSetResult>;
3864 describe: (input: ConfigDescribeInput) => Promise<ConfigDescribeResult>;
3865 };
3866 prompt: {
3867 submit: (input: PromptSubmitArgs) => Promise<PromptSubmitResult>;
3868 fill: (input: PromptFillArgs) => Promise<PromptFillResult>;
3869 suggest: (input: PromptSuggestArgs) => Promise<PromptSuggestResult>;
3870 section: (input: PromptSectionInput) => Promise<PromptSectionResult>;
3871 context: (input: PromptContextInput) => Promise<PromptContextResult>;
3872 attachment: (input: PromptAttachmentInput) => Promise<PromptAttachmentResult>;
3873 compose: (input: PromptComposeInput) => Promise<PromptComposeResult>;
3874 };
3875 skill: {
3876 prompt: (input: SkillPromptInput) => Promise<SkillPromptResult>;
3877 };
3878 attribution: {
3879 text: (input: AttributionTextInput) => Promise<AttributionTextResult>;
3880 };
3881 agent: {
3882 offer: (input: AgentOfferInput) => Promise<AgentOfferResult>;
3883 spawn: (input: AgentSpawnArgs) => Promise<AgentSpawnResult>;
3884 };
3885 session: {
3886 start: (input: SessionStartInput) => Promise<SessionStartResult>;
3887 receive: (input: SessionReceiveInput) => Promise<SessionReceiveResult>;
3888 compact: (input?: SessionCompactArgs) => Promise<SessionCompactResult>;
3889 attach: (input: SessionAttachInput) => Promise<SessionAttachResult>;
3890 detach: (input: SessionDetachInput) => Promise<SessionDetachResult>;
3891 measure: (input: SessionMeasureInput) => Promise<SessionMeasureResult>;
3892 end: (input: SessionEndInput) => Promise<SessionEndResult>;
3893 };
3894 turn: {
3895 start: (input: TurnStartInput) => Promise<TurnStartResult>;
3896 step: (input: TurnStepInput) => HookStream<TurnStepChunk, TurnStepResult>;
3897 complete: (input: TurnCompleteInput) => Promise<TurnCompleteResult>;
3898 };
3899 ui: {
3900 render: <C extends RenderComponent>(input: RenderInput<C>) => Promise<RenderElement>;
3901 resolve: <E extends ResolveInput>(e: E) => Elements[E['surface']];
3902 scroll: (input: UiScrollArgs) => Promise<UiScrollResult>;
3903 focus: (input: UiFocusArgs) => Promise<UiFocusResult>;
3904 };
3905 };
3906
3907 /**
3908 * The name of an event: a key of EventOf, the engine's own (CoreEventName)
3909 * and the declared plugin nouns' (NounEventName).
3910 */
3911 export type EventName = keyof EventOf;
3912
3913 /**
3914 * The argument of each event, by event name: what a hook receives as `e`
3915 * and what the call on `$` takes. Plain data, frozen to every depth.
3916 *
3917 * The engine's own events (CoreEventOf: a plugin's `$.fs.write(...)` is
3918 * a dispatch the hooks above it see) and the methods of the plugin nouns
3919 * declared on EngineInterface (NounEventOf).
3920 */
3921 export type EventOf = CoreEventOf & NounEventOf;
3922
3923 /**
3924 * The result of event `N`: what its hooks return and what their `next(e)`
3925 * resolves to.
3926 */
3927 export type EventResult<N extends EventName = EventName> = ResultOf[N];
3928
3929 /**
3930 * The hook signature of each event, `($, e, next)`, as one mapped type over
3931 * EventOf; a streaming event's is the generator form (StreamHook).
3932 *
3933 * With one handler type per event, `Events[E]` for a generic E would be a
3934 * union; as one mapped type it stays a single function type the engine
3935 * calls without a cast (TS 4.6 correlated unions).
3936 *
3937 * @param $ the engine interface, frozen, the same object at every invocation;
3938 * at `engine.create` the empty table, since `$` exists after the fold
3939 * @param e the event's argument, frozen to every depth (`e.command = "ls"`
3940 * is a type error and throws); a rewrite is a copy passed to `next`
3941 * @param next the rest of the chain; a chain a hook raises through `$` skips
3942 * this hook and runs every other, its plugin's others too
3943 */
3944 export type Events = {
3945 [E in keyof EventOf]: E extends StreamingEventName ? StreamHook<E> : ($: E extends 'engine.create' ? NoEngineInterface : EngineInterface, e: Frozen<Args<E>>, next: Next<E>) => EventResult<E> | Promise<EventResult<E>>;
3946 };
3947
3948 type ExitReason = 'clear' | 'resume' | 'logout' | 'prompt_input_exit' | 'other';
3949
3950 type FileChangedHookInput = BaseHookInput & {
3951 hook_event_name: 'FileChanged';
3952 file_path: string;
3953 event: 'change' | 'add' | 'unlink';
3954 };
3955
3956 /**
3957 * `T` with every property read-only to every depth, arrays and tuples kept
3958 * as declared: how a hook's `e` is typed.
3959 *
3960 * `e.command = 'x'` is a type error; `next({ ...e, command: 'x' })` compiles.
3961 */
3962 export type Frozen<T> = T extends (...args: never[]) => unknown ? T : T extends readonly unknown[] ? {
3963 [K in keyof T]: Frozen<T[K]>;
3964 } : T extends object ? {
3965 readonly [K in keyof T]: Frozen<T[K]>;
3966 } : T;
3967
3968 /**
3969 * One file `$.fs.ancestors` found: the directory it stands in, the name it
3970 * was asked for by, and its text as the engine's memory loader reads it.
3971 */
3972 export type FsAncestor = {
3973 /**
3974 * The directory the file stands in, absolute.
3975 */
3976 dir: string;
3977 /**
3978 * The spelling the caller asked for it by.
3979 */
3980 name: string;
3981 /**
3982 * The file's text, with what its `@include`s bring after it.
3983 */
3984 content: string;
3985 /**
3986 * The file and then each file its `@` imports brought, in load order,
3987 * path and text apiece; `content` is these texts joined.
3988 */
3989 parts: readonly FsAncestorPart[];
3990 };
3991
3992 /**
3993 * One file of an ancestor entry: the file itself or one it imported.
3994 */
3995 export type FsAncestorPart = {
3996 /**
3997 * The file's path, absolute.
3998 */
3999 path: string;
4000 /**
4001 * Its text as the engine's memory loader reads it.
4002 */
4003 content: string;
4004 };
4005
4006 /**
4007 * The argument of `$.fs.ancestors`: the file names to look for in each
4008 * directory, the file to walk down to, and the directory to walk beneath.
4009 */
4010 export type FsAncestorsRequest = {
4011 /**
4012 * Relative `.md` file names, each looked for in every directory.
4013 */
4014 names: readonly string[];
4015 /**
4016 * The file the walk goes on down to the directory of, relative to the
4017 * working directory or absolute; absent, it ends at the working directory.
4018 */
4019 of?: string;
4020 /**
4021 * The directory the walk starts beneath, relative to the working directory
4022 * or absolute; absent, the walk starts at the filesystem root.
4023 *
4024 * Only directories strictly inside it are read, so with `of` a file under
4025 * the project root the walk reads the directories between the two, as the
4026 * engine reads a nested CLAUDE.md; a file not inside it finds nothing.
4027 */
4028 below?: string;
4029 };
4030
4031 /**
4032 * What `$.fs.read(path, { as: "bytes" })` resolves with: the file's bytes,
4033 * base64, since only plain data crosses into a plugin's environment.
4034 */
4035 export type FsBytes = {
4036 /**
4037 * The file's content, standard padded base64;
4038 * `Uint8Array.fromBase64(base64)` gives the bytes back.
4039 */
4040 base64: string;
4041 };
4042
4043 /**
4044 * One entry of `$.fs.list`: the entry itself, a link not followed.
4045 */
4046 export type FsEntry = {
4047 /**
4048 * The entry's name (no directory part).
4049 */
4050 name: string;
4051 /**
4052 * `file`, `dir`, or `other`, of the entry itself: a symbolic link is
4053 * `other` (`$.fs.stat` says what it leads to, and with `resolve` where).
4054 */
4055 kind: 'file' | 'dir' | 'other';
4056 /**
4057 * Bytes, for a file.
4058 */
4059 size: number;
4060 /**
4061 * True when the entry is a symbolic link.
4062 */
4063 isLink: boolean;
4064 };
4065
4066 /**
4067 * How `$.fs.read` answers: `text` (UTF-8, the default) or `bytes` (base64).
4068 */
4069 export type FsReadAs = 'text' | 'bytes';
4070
4071 /**
4072 * The options of `$.fs.read` that ask for the bytes: the call answers
4073 * `{ base64 }`.
4074 */
4075 export type FsReadBytesOptions = {
4076 /**
4077 * `bytes`.
4078 */
4079 as: 'bytes';
4080 };
4081
4082 /**
4083 * `$.fs.read`: the file's text, or with `{ as: "bytes" }` its bytes as
4084 * `{ base64 }`.
4085 */
4086 export type FsReadCall = {
4087 (path: string): Promise<string>;
4088 (path: string, options: FsReadBytesOptions): Promise<FsBytes>;
4089 (path: string, options: FsReadOptions): Promise<string | FsBytes>;
4090 };
4091
4092 /**
4093 * The options of `$.fs.read`.
4094 */
4095 export type FsReadOptions = {
4096 /**
4097 * `bytes` answers `{ base64 }` for a binary file; `text` (the default) the
4098 * file's text.
4099 */
4100 as: FsReadAs;
4101 };
4102
4103 /**
4104 * What `$.fs.stat` resolves with: what the path leads to, whether the path
4105 * itself is a symbolic link, and where it lands when asked.
4106 */
4107 export type FsStat = {
4108 /**
4109 * `file`, `dir`, or `other`, of what the path leads to: a link is
4110 * followed, and one that leads nowhere is `other`.
4111 */
4112 kind: 'file' | 'dir' | 'other';
4113 /**
4114 * Bytes, for a file.
4115 */
4116 size: number;
4117 /**
4118 * Last modification, milliseconds since the epoch.
4119 */
4120 mtimeMs: number;
4121 /**
4122 * True when the path itself is a symbolic link; `kind`, `size` and
4123 * `mtimeMs` then describe what it points at, or the link when that is gone.
4124 */
4125 isLink: boolean;
4126 /**
4127 * Where the path landed when asked with `{ resolve: true }`: absolute,
4128 * every symbolic link followed, `.` and `..` folded; else absent.
4129 *
4130 * Absent too when the path leads nowhere or a hook above withheld it, so
4131 * a guard denies without it; and a hard link, a volume or file-id spelling
4132 * (macOS `/.vol/`) or a case alias keeps its own spelling, `isLink` false.
4133 *
4134 * @remarks a deny-list on spellings is thus best effort; an allow-list on
4135 * `realPath` under a root resolved the same way is the robust guard.
4136 */
4137 realPath?: string;
4138 };
4139
4140 /**
4141 * The options of `$.fs.stat`.
4142 */
4143 export type FsStatOptions = {
4144 /**
4145 * True to resolve where the path lands as well.
4146 *
4147 * The answer then carries `realPath` when the path leads somewhere, at
4148 * the cost of one more file system call: where it landed then, not a hold
4149 * on what a tool opens afterwards.
4150 */
4151 resolve: boolean;
4152 };
4153
4154 /**
4155 * Every event (`*`), or every event under a namespace (`classic.*`: each
4156 * one whose name starts with `classic.`).
4157 */
4158 export type Glob = '*' | `${Namespace}.*`;
4159
4160 /**
4161 * The hook `on(pattern, hook)` takes for a glob or a negation: one function
4162 * placed on every selected event, `e` and the result typed as their union.
4163 *
4164 * `next.event` says which event a run is; `next.is(pattern, e)` narrows `e`
4165 * to one of them. At `engine.create` (a negation may select it) `$` is the
4166 * empty table and the hook observes the fold, as a `*` hook does.
4167 */
4168 export type GlobHook<P extends Pattern, N extends EventName = Selected<P>> = ($: EngineInterface, e: Frozen<Args<N>>, next: GlobNext<P>) => EventResult<N> | Promise<EventResult<N>>;
4169
4170 /**
4171 * `next` in a hook on a glob or a negation: an overload per selected event,
4172 * then one over their union for an `e` not yet narrowed.
4173 *
4174 * `is` narrows `e` to the selected events its pattern names; `event` is one
4175 * of the selected names.
4176 */
4177 export type GlobNext<P extends Pattern, N extends EventName = Selected<P>> = OrderedOverloads<N> & {
4178 (e: Args<N>): Promise<GlobNextResult<N>>;
4179 /**
4180 * Continues this dispatch at a tier, as Next's `to` (a managed hook's),
4181 * over the selected events' union for an `e` not yet narrowed.
4182 */
4183 readonly to: (e: Args<N>, tier: TargetTier) => Promise<GlobNextResult<N>>;
4184 readonly signal: AbortSignal;
4185 readonly is: <M extends PatternOver<N>>(pattern: M, e: unknown) => e is Frozen<Args<Extract<N, Selected<M>>>>;
4186 readonly event: N;
4187 readonly origin: Origin;
4188 readonly trace: readonly TraceEntry<N, Args<N>, GlobNextResult<N>>[];
4189 readonly budget: NextBudget;
4190 };
4191
4192 /**
4193 * What `next(e)` resolves to in a glob hook before `e` is narrowed: the
4194 * NextResult of each selected event, as a union.
4195 */
4196 type GlobNextResult<N extends EventName> = {
4197 [K in N]: NextResult<K>;
4198 }[N];
4199
4200 /**
4201 * One hook, `($, e, next)`, on event `E`.
4202 */
4203 export type Hook<E extends EventName = EventName> = Events[E];
4204
4205 /**
4206 * The time bounds every hook runs under, in milliseconds: the engine's own
4207 * constants are typed by these members, and `next.budget` reads the live one.
4208 *
4209 * Each bounds the hook's OWN time: the clock stops while a `next(e)` call or
4210 * any `$` call of the hook's is in flight (a `$.clock` wait excepted), so a
4211 * slow chain beneath or a minute-long `$.model.complete` costs it nothing.
4212 *
4213 * @example
4214 * await $.model.complete(ask) // a minute; next.budget.remainingMs unmoved
4215 */
4216 export type HookBudget = {
4217 /**
4218 * A hook's budget per dispatch, from its call to its return; past it the
4219 * hook is absent (its `.catch` asked, else `next(e)` run on its behalf).
4220 *
4221 * A streaming hook's (`turn.step`, an async generator) spans its whole
4222 * run and counts only while its own code runs: never at a `yield`, never
4223 * while it reads the stream beneath. Not per chunk: the sum of its work.
4224 */
4225 readonly ms: 10_000;
4226 /**
4227 * A `.catch` handler's grace: a fresh budget from the moment it is called,
4228 * on the same clock (its `next` replay and its `$` calls are free).
4229 *
4230 * Past it the hook is absent as if it had no handler; `next.error.budget`
4231 * and `next.budget.ms` both read it there. `engine.create` has no budget.
4232 */
4233 readonly catchMs: 1_000;
4234 /**
4235 * How long a hook may keep running after `next.signal` aborted (the person
4236 * interrupted, a hook above settled first, its own budget ran out).
4237 *
4238 * Past it the hook is reported as lingering; the dispatch had already gone
4239 * on without it when the signal aborted.
4240 */
4241 readonly lingerMs: 5_000;
4242 };
4243
4244 /**
4245 * Why a hook failed, as its `.catch` handler reads it on `next.error`: plain
4246 * frozen data.
4247 *
4248 * `throw`: the hook threw, or returned what the site refuses, `message`
4249 * saying what; `timeout`: it outran its budget, `message` then what its last
4250 * `next()` rejected with, if it did. `budget` is the handler's own grace.
4251 */
4252 export type HookFailure = {
4253 readonly kind: 'throw' | 'timeout';
4254 /**
4255 * The thrown error's message, or for a timeout what the hook's last
4256 * `next()` rejected with; absent for a timeout with nothing rejected.
4257 */
4258 readonly message?: string;
4259 /**
4260 * The grace the handler runs under, in milliseconds; past it, the hook is
4261 * absent as if it had no handler.
4262 */
4263 readonly budget: number;
4264 };
4265
4266 /**
4267 * The hook type per pattern: an event's own (Events), `*`'s (AnyEventHook),
4268 * or a glob's over the events it selects (GlobHook), as one conditional type.
4269 *
4270 * One type, so the two-argument `on` stays ONE generic signature: as an
4271 * overload set the language service offers no tool-name completions inside
4272 * `e.tool === "`; as an index into a table TS intersects every argument.
4273 */
4274 export type HookFor<P extends Pattern> = P extends '*' ? AnyEventHook : P extends EventName ? Events[P] : GlobHook<P>;
4275
4276 type HookInput = PreToolUseHookInput | PostToolUseHookInput | PostToolUseFailureHookInput | PostToolBatchHookInput | PermissionDeniedHookInput | NotificationHookInput | UserPromptSubmitHookInput | UserPromptExpansionHookInput | SessionStartHookInput | SessionEndHookInput | StopHookInput | StopFailureHookInput | SubagentStartHookInput | SubagentStopHookInput | PreCompactHookInput | PostCompactHookInput | PreModelSwitchHookInput | PostModelSwitchHookInput | PermissionRequestHookInput | SetupHookInput | TeammateIdleHookInput | TaskCreatedHookInput | TaskCompletedHookInput | ElicitationHookInput | ElicitationResultHookInput | ConfigChangeHookInput | InstructionsLoadedHookInput | WorktreeCreateHookInput | WorktreeRemoveHookInput | CwdChangedHookInput | FileChangedHookInput | DirectoryAddedHookInput | MessageDisplayHookInput;
4277
4278 /**
4279 * The hook event `E` takes: an async generator over its chunks for a
4280 * streaming event (`turn.step`), `($, e, next) => result` for every other.
4281 */
4282 export type HookOf<E extends EventName> = Events[E];
4283
4284 /**
4285 * What a hooks module exports: `register`, and nothing the loader reads
4286 * besides.
4287 */
4288 export type HooksModule = {
4289 register: Register;
4290 };
4291
4292 /**
4293 * What `next(e)` returns on a streaming event: the stream of everything
4294 * beneath, chunk by chunk, whose return value is the result from beneath.
4295 *
4296 * `yield* next(e)` forwards the chunks and evaluates to that result; a
4297 * transforming hook reads `for await (const chunk of stream)` and then
4298 * `await stream.result`. Each call runs what is beneath afresh.
4299 */
4300 export type HookStream<C, R> = AsyncGenerator<C, R> & {
4301 /**
4302 * Settles with what beneath returned once the stream has been read to
4303 * its end; rejects if the stream is closed before that.
4304 */
4305 readonly result: Promise<R>;
4306 };
4307
4308 /**
4309 * Options of `$.http.fetch`.
4310 */
4311 export type HttpInit = {
4312 /**
4313 * `GET` (default), `POST`, ...
4314 */
4315 method?: string;
4316 /**
4317 * Request headers.
4318 */
4319 headers?: Record<string, string>;
4320 /**
4321 * The request body, as text.
4322 */
4323 body?: string;
4324 /**
4325 * The handle `$.session.authorize()` answered: the engine sets the
4326 * session's credential header itself, only for a first-party host.
4327 */
4328 auth?: string;
4329 /**
4330 * The absolute path of a Unix domain socket the request goes over instead
4331 * of TCP (near 100 B at most): one per session, in a private directory.
4332 *
4333 * The URL still gives path, query and Host; redirects stay on it; no proxy;
4334 * `https:` is TLS verified against the URL's host (always, once `auth`
4335 * rides), so a credential reaches only its host. No relative path, no NUL.
4336 *
4337 * @example
4338 * await $.http.fetch(url, { socketPath: `${runDirectory}/bridge.sock` })
4339 */
4340 socketPath?: string;
4341 };
4342
4343 /**
4344 * What `$.http.fetch` resolves with.
4345 */
4346 export type HttpResponse = {
4347 /**
4348 * The HTTP status code.
4349 */
4350 status: number;
4351 /**
4352 * True for a 2xx status.
4353 */
4354 ok: boolean;
4355 /**
4356 * Response headers, lower-cased names.
4357 */
4358 headers: Record<string, string>;
4359 /**
4360 * The body, as text.
4361 */
4362 text: string;
4363 };
4364
4365 /**
4366 * A `$.ui.blit` argument swapping one of the caller's mounted keyed Images
4367 * to its next picture; every call sends it, so a stream needs no generation.
4368 *
4369 * The surface writes them with its frames, some sixty a second: byte and
4370 * `file` sources fold to the last per frame; every `shm` source reaches the
4371 * terminal (it unlinks each), and is denied while frames are not written.
4372 *
4373 * @example await $.ui.blit({ requestId: 'browser', key: 'view',
4374 * source: { shm: '/tb-4', format: 'rgb', width: 1280, height: 720 } })
4375 */
4376 export type ImageBlitArgs = {
4377 /**
4378 * The site the Image is drawn in, by the `requestId` this plugin draws it
4379 * under.
4380 */
4381 requestId: string;
4382 /**
4383 * The Image's `key` in that drawing.
4384 */
4385 key: string;
4386 /**
4387 * The next picture (ImageSource): bytes, or a name the terminal reads.
4388 */
4389 source: ImageSource;
4390 /**
4391 * Refused unless it is the mounted Image's width in cells. Absent, that.
4392 */
4393 columns?: number;
4394 /**
4395 * Refused unless it is the mounted Image's height in cells. Absent, that.
4396 */
4397 rows?: number;
4398 };
4399
4400 /**
4401 * The props of `Image`, the terminal surface's picture leaf: pixels over a
4402 * box of cells where the terminal can (kitty, Ghostty), the `alt` elsewhere.
4403 *
4404 * A leaf: no children, `hover` or `onPress`; the cells are text, so the
4405 * picture scrolls and clips as a word does. Drawn again with another `source`
4406 * it is replaced in place; keyed, `$.ui.blit` swaps it at the frame rate.
4407 *
4408 * @example const { bytes } = await $.fs.read('chart.png', { as: 'bytes' })
4409 * <Image source={{ png: bytes.toBase64() }} columns={40} rows={12} alt="p95" />
4410 * @example <Image key="view" source={{ shm: '/tb-7', format: 'rgb', width: 960,
4411 * height: 600 }} columns={80} rows={25} alt="the page" />
4412 */
4413 export type ImageProps = {
4414 /**
4415 * The element's address within the drawing: what `$.ui.blit` names to swap.
4416 *
4417 * Unique among the Images of one tree; absent, only a redraw changes it.
4418 */
4419 key?: string;
4420 /**
4421 * The picture (ImageSource): `{ png }` or `{ rgba, width, height }` bytes,
4422 * or `{ file, format }` / `{ shm, format, width, height }` read here.
4423 */
4424 source: ImageSource;
4425 /** How many terminal columns wide, 1 to 255; the picture is scaled to fill
4426 * the box and the site clips what its body cannot show. */
4427 columns: number;
4428 /**
4429 * How many terminal rows tall, 1 to 255.
4430 */
4431 rows: number;
4432 /**
4433 * What the picture says, drawn dim in its place where the picture cannot
4434 * be (and read by a screen reader); required, may be a single space.
4435 */
4436 alt: string;
4437 };
4438
4439 /**
4440 * The picture an `Image` shows: base64 bytes the plugin holds (at most 2 MiB
4441 * decoded), or the name of a file or POSIX shared-memory object it does not.
4442 *
4443 * A name is one another process on this machine wrote; the terminal, running
4444 * as the person, opens, reads and decodes it itself, so no pixel crosses `$`
4445 * and the engine never touches it. One it will not read (not a regular file,
4446 * under `/proc`, `/sys` or `/dev`, gone, across ssh) draws a blank box, its
4447 * words in the debug log; it unlinks a shared-memory object once read, so
4448 * each frame is a fresh one. A source equal to the last drawn sends nothing;
4449 * `generation` makes new content under an unchanged name a new source.
4450 *
4451 * @example { rgba: pixels.toBase64(), width: 64, height: 32 }
4452 * @example { file: '/dev/shm/frame-3.rgba', format: 'rgba', width: 640,
4453 * height: 400 }
4454 * @example { shm: '/tb-view-2', format: 'rgb', width: 1280, height: 720 }
4455 */
4456 export type ImageSource = {
4457 /**
4458 * A whole PNG file, base64.
4459 */
4460 png: string;
4461 } | {
4462 /**
4463 * `width * height` RGBA pixels, 4 bytes each, base64.
4464 */
4465 rgba: string;
4466 /**
4467 * Pixels per row, 1 to 2048.
4468 */
4469 width: number;
4470 /**
4471 * Rows of pixels, 1 to 2048.
4472 */
4473 height: number;
4474 } | {
4475 /**
4476 * The absolute path of a regular file holding a whole PNG, at most
4477 * 3072 bytes of path; read by the terminal, left in place.
4478 */
4479 file: string;
4480 /**
4481 * The file is a whole PNG; the terminal decodes it and sizes it itself.
4482 */
4483 format: 'png';
4484 /**
4485 * A whole number that changes when the file's content does under the
4486 * same path, so a redraw reads it again; absent, the path alone tells.
4487 */
4488 generation?: number;
4489 } | {
4490 /**
4491 * The absolute path of a regular file of `width * height` raw pixels, at
4492 * most 3072 bytes of path; read by the terminal, left in place.
4493 *
4494 * On Linux a file under `/dev/shm` is memory.
4495 */
4496 file: string;
4497 /**
4498 * `rgba`, 4 bytes a pixel, or `rgb`, 3.
4499 */
4500 format: 'rgba' | 'rgb';
4501 /**
4502 * Pixels per row, 1 to 4096.
4503 */
4504 width: number;
4505 /**
4506 * Rows of pixels, 1 to 4096.
4507 */
4508 height: number;
4509 /**
4510 * A whole number that changes when the file's content does under the
4511 * same path, so a redraw reads it again; absent, the path alone tells.
4512 */
4513 generation?: number;
4514 } | {
4515 /**
4516 * The name of a POSIX shared-memory object of `width * height` raw
4517 * pixels: `/` then up to 254 of `A-Z a-z 0-9 . _ -` (macOS takes 30).
4518 *
4519 * The terminal unlinks it after reading, so a name feeds one Image
4520 * drawn once and is never sent again on a redraw; POSIX terminals only.
4521 */
4522 shm: string;
4523 /**
4524 * `rgba`, 4 bytes a pixel, or `rgb`, 3.
4525 */
4526 format: 'rgba' | 'rgb';
4527 /**
4528 * Pixels per row, 1 to 4096.
4529 */
4530 width: number;
4531 /**
4532 * Rows of pixels, 1 to 4096.
4533 */
4534 height: number;
4535 /**
4536 * A whole number that changes when a fresh object reuses a name, so a
4537 * redraw reads it again; absent, the name alone tells.
4538 */
4539 generation?: number;
4540 };
4541
4542 /**
4543 * The keys of object pattern `P` that object member `E` cannot satisfy, `D`
4544 * levels down; `never` when there is none, which is what keeps the member.
4545 *
4546 * A key `E` does not have (an open record has every string key), or one
4547 * whose value `P` narrows to nothing.
4548 */
4549 type ImpossibleKeys<E, P, D extends readonly unknown[]> = {
4550 [K in keyof P]-?: K extends keyof E ? [NarrowedValue<Exclude<E[K], undefined>, P[K], D>] extends [never] ? K : never : K;
4551 }[keyof P];
4552
4553 /**
4554 * The props of `Input`, every surface's one-line text field: an address,
4555 * optional texts, and the closures a change and a submit run. A leaf.
4556 *
4557 * Focused through the same ring as `Button` (`abovePrompt:focus`); while it
4558 * has focus every printable key reaches it alone and Esc returns them; a
4559 * change and Enter raise `ui.input`, whose bottom is `onInput` / `onSubmit`.
4560 */
4561 export type InputProps = {
4562 /**
4563 * The element's address: `e.element` at `ui.input`, what a matcher names.
4564 */
4565 key: string;
4566 /**
4567 * Text drawn before the field.
4568 */
4569 label?: string;
4570 /**
4571 * Text drawn dim in an empty field.
4572 */
4573 placeholder?: string;
4574 /**
4575 * The text the field holds when drawn; the person's typing replaces it
4576 * until the hook draws another.
4577 */
4578 value?: string;
4579 /**
4580 * What Enter does, in a word or two, drawn beside the field while it has
4581 * focus (`send`). Defaults to `submit`.
4582 */
4583 submitLabel?: string;
4584 /**
4585 * The site's focus ring starts here when the site takes the keyboard,
4586 * instead of on nothing, as the DOM's `autofocus`: Enter acts on it at once.
4587 *
4588 * A pane opened with `focus`, or the person's focus chord or click, is the
4589 * take. Of several in one site the first drawn wins; it raises `ui.focus`,
4590 * origin this plugin. A ring the person has moved stays where it was put.
4591 */
4592 autoFocus?: true;
4593 /**
4594 * Runs on every change of the text, in the plugin's own environment: the
4595 * bottom of a `ui.input` chain of kind `change`.
4596 */
4597 onInput?: (value: string, e: UiInputArgument) => void;
4598 /**
4599 * Runs on Enter with the text, in the plugin's own environment: the bottom
4600 * of a `ui.input` chain of kind `submit`. No model turn unless it asks one.
4601 */
4602 onSubmit: (value: string, e: UiInputArgument) => void;
4603 };
4604
4605 /**
4606 * Every tier an instruction file can belong to, for checking a hook's
4607 * answer; the kind type is derived from this list.
4608 */
4609 const INSTRUCTION_FILE_KINDS: readonly ["managed", "user", "project", "local", "memory"];
4610
4611 /**
4612 * One instruction file behind the `claudeMd` block: where it was read, its
4613 * tier, its text as loaded, and the file that `@`-imported it if one did.
4614 */
4615 export type InstructionFile = {
4616 /**
4617 * The file's path, absolute.
4618 */
4619 path: string;
4620 /**
4621 * Its tier.
4622 */
4623 kind: InstructionFileKind;
4624 /**
4625 * The text as loaded (comments and frontmatter already stripped).
4626 */
4627 content: string;
4628 /**
4629 * The path of the file whose `@` import brought this one, when one did.
4630 */
4631 parent?: string;
4632 };
4633
4634 /**
4635 * What tier an instruction file belongs to: the organization's managed
4636 * policy, the person's own, the project's checked-in or private ones, memory.
4637 */
4638 export type InstructionFileKind = (typeof INSTRUCTION_FILE_KINDS)[number];
4639
4640 type InstructionsLoadedHookInput = BaseHookInput & {
4641 hook_event_name: 'InstructionsLoaded';
4642 file_path: string;
4643 memory_type: 'User' | 'Project' | 'Local' | 'Managed';
4644 load_reason: 'session_start' | 'nested_traversal' | 'path_glob_match' | 'include' | 'compact';
4645 globs?: string[];
4646 trigger_file_path?: string;
4647 parent_file_path?: string;
4648 };
4649
4650 /**
4651 * What `$.ui.invalidate` takes: a render event, or one of the six events
4652 * whose answers the engine caches for the session.
4653 */
4654 export type InvalidatableEventName = RenderEventName | 'prompt.section' | 'prompt.context' | 'prompt.attachment' | 'tool.describe' | 'command.describe' | 'config.describe';
4655
4656 /**
4657 * Whether tag key `K` selects members of `I`: it does when each member gives
4658 * it ONE literal (`component: "ToolUse"` on the ToolUse variant).
4659 *
4660 * A key that is the same union on every member is a filter at runtime; it
4661 * is left out of the selection so that it cannot defeat the narrowing the
4662 * other keys give.
4663 */
4664 type IsDiscriminant<I, K> = I extends unknown ? K extends KnownKeys<I> ? IsSingleLiteral<I[K]> : true : never;
4665
4666 /**
4667 * Whether `V` is made of literals only: `"a" | "b"` is, `string` is not.
4668 */
4669 type IsLiteralValued<V> = string extends V ? false : number extends V ? false : boolean extends V ? false : [V] extends [string | number | boolean] ? true : false;
4670
4671 /**
4672 * Whether `V` is exactly one string, number or boolean literal.
4673 */
4674 type IsSingleLiteral<V> = [V] extends [string | number | boolean] ? IsUnion<V> extends true ? false : true : false;
4675
4676 /**
4677 * Whether `T` is a union of two or more members.
4678 */
4679 type IsUnion<T, U = T> = T extends unknown ? [U] extends [T] ? false : true : never;
4680
4681 /**
4682 * Plain data: what JSON holds, and what crosses between a plugin's hooks
4683 * module and its surface module whole (a `Client`'s props, a post's data).
4684 *
4685 * A function, a class instance, `undefined` or a cycle is not plain data;
4686 * the engine refuses one where it checks, and drops it where it clones.
4687 */
4688 export type JsonValue = string | number | boolean | null | readonly JsonValue[] | {
4689 readonly [key: string]: JsonValue;
4690 };
4691
4692 /**
4693 * What `next` takes in a matched hook: the variants of `e` the matcher can
4694 * match (KeptMembers), as declared, so a rewrite of a pinned field passes.
4695 */
4696 type KeptEvent<P extends Pattern, M> = MatchedNames<P, M> extends infer N extends EventName ? N extends unknown ? KeptMembers<Args<N>, M> : never : never;
4697
4698 /**
4699 * The members of the argument union `E` matcher `P` can match, as declared;
4700 * what `next` takes, so a rewrite may change a field the matcher pinned.
4701 *
4702 * A member is dropped when `P` names a key it lacks, or gives a key, at any
4703 * depth, a value none of that key's values can equal (ImpossibleKeys).
4704 */
4705 type KeptMembers<E, P> = E extends unknown ? [ImpossibleKeys<E, P, []>] extends [never] ? E : never : never;
4706
4707 /**
4708 * The declared keys of `T`, the string and number index signatures left out.
4709 */
4710 type KnownKeys<T> = keyof {
4711 [K in keyof T as string extends K ? never : number extends K ? never : K]: 0;
4712 };
4713
4714 /**
4715 * The events whose overload must come after the rest, lest it shadow them.
4716 *
4717 * `classic.PreToolUse` shares `tool.call`'s envelope, `turn.abort` every
4718 * turn event's `turnId`, and an object of any shape is assignable to NoArgs.
4719 */
4720 type LateOverload = 'classic.PreToolUse' | 'turn.abort' | NoArgsEvent;
4721
4722 /**
4723 * The props of `Link`, a hyperlink every surface draws: an OSC 8 span on the
4724 * terminal (else its text then the URL in dim), an anchor on desktop.
4725 *
4726 * An inline element: its children are the text, strings and inline
4727 * elements; absent children the `label`, absent both the URL. The engine
4728 * bounds `href` before the tree crosses.
4729 */
4730 export type LinkProps = {
4731 /**
4732 * Where the link goes: an `https:` URL (or `http://localhost`), at most
4733 * 2048 characters of printable ASCII, spelled as `new URL(href).href`.
4734 *
4735 * No `user@host` part, no raw `@`, space or non-ASCII letter (encode them);
4736 * anything else refuses the tree the Link is in.
4737 */
4738 href: string;
4739 /**
4740 * The text drawn when the element has no children; absent both, the URL
4741 * itself is the text.
4742 */
4743 label?: string;
4744 };
4745
4746 /**
4747 * What a matcher value selects by: itself, or `unknown` for a RegExp, which
4748 * selects nothing.
4749 */
4750 type Literal<X> = X extends RegExp ? unknown : X;
4751
4752 /**
4753 * What a `Markdown` carries across the boundary: its address, text, dimness
4754 * and which links it answers; `onLinkPress` stays behind, a `press` instead.
4755 */
4756 export type MarkdownLeafProps = {
4757 /**
4758 * The element's address: what `e.element` carries and what a matcher
4759 * names; present whenever the element carries a `press`.
4760 */
4761 key?: string;
4762 /**
4763 * The markdown drawn, bounded as a Text's string is.
4764 */
4765 text: string;
4766 /**
4767 * The whole block dim, as `Text`'s `dimColor`; absent draws as false.
4768 */
4769 dimColor?: boolean;
4770 /**
4771 * Which links a press belongs to, by `href` as written; absent, with a
4772 * `press`, every link drawn. Compared with the pressed target only.
4773 */
4774 pressableLinks?: readonly string[];
4775 };
4776
4777 /**
4778 * The props of `Markdown`, a block of markdown every surface draws as it
4779 * draws an assistant reply's text: its own renderer, links, tables, fences.
4780 *
4781 * A leaf: no children. `text` is the element's data as a `Text`'s string is,
4782 * bounded the same way. With `onLinkPress` the links it draws are the
4783 * plugin's to answer: a press on one raises `ui.press` addressed to `key`.
4784 */
4785 export type MarkdownProps = {
4786 /**
4787 * The element's address: `e.element` at `ui.press`, what a matcher names.
4788 * Required with `onLinkPress`, since a press needs one; else optional.
4789 */
4790 key?: string;
4791 /**
4792 * The markdown drawn, as an assistant reply would write it.
4793 *
4794 * At most 10000 characters, tab and newline its only control characters;
4795 * a link whose scheme is not `https:`, `http:` or `file:` draws as text,
4796 * never clickable. Not drawn around the approval dialog.
4797 */
4798 text: string;
4799 /**
4800 * The whole block dim, as `Text`'s `dimColor`: a thought, an aside.
4801 */
4802 dimColor?: boolean;
4803 /**
4804 * What a press on a link in `text` runs, in the plugin's own environment:
4805 * the bottom of a `ui.press` chain whose `e.link` names the link.
4806 *
4807 * A press is a plain single click where the surface reports clicks (the
4808 * fullscreen terminal): it opens nothing and lands once no double-click
4809 * followed; a ctrl, alt or terminal-kept cmd click opens it as before.
4810 */
4811 onLinkPress?: (link: PressedLink, e: UiPressArgument) => void;
4812 /**
4813 * Which links in `text` a press belongs to, by `href` as the markdown
4814 * writes it; absent, every link drawn. Only with `onLinkPress`.
4815 *
4816 * A link left out keeps the surface's own behaviour. At most 256 entries
4817 * of at most 2048 characters; compared with the pressed link's target,
4818 * never opened.
4819 */
4820 pressableLinks?: readonly string[];
4821 };
4822
4823 /**
4824 * The argument a matched hook receives: `e` narrowed by `M` (Narrowed), per
4825 * event the registration covers.
4826 */
4827 export type MatchedEvent<P extends Pattern, M> = MatchedNames<P, M> extends infer N extends EventName ? N extends unknown ? Narrowed<Args<N>, M> : never : never;
4828
4829 /**
4830 * The hook `on(pattern, matcher, hook)` takes: `($, e, next)` with `e`
4831 * narrowed by the matcher (MatchedEvent), and a tagged result the same way.
4832 *
4833 * On a streaming event it is the generator form (MatchedStreamHook). `next`
4834 * takes the variants the matcher keeps, as declared (KeptEvent); `next.is`
4835 * names the events the registration covers and narrows as the matcher does.
4836 */
4837 export type MatchedHook<P extends Pattern, M> = P extends StreamingEventName ? MatchedStreamHook<P, M> : ($: EngineInterface, e: Frozen<MatchedEvent<P, M>>, next: Next<MatchedNames<P, M>, KeptEvent<P, M>, MatchedResult<P, M>, {
4838 [K in MatchedNames<P, M>]: Narrowed<Args<K>, M>;
4839 }>) => MatchedResult<P, M> | Promise<MatchedResult<P, M>>;
4840
4841 /**
4842 * The events a matched registration on `P` covers: the event named, or for a
4843 * glob every selected event whose input has each key the matcher names.
4844 */
4845 type MatchedNames<P, M = never> = P extends EventName ? P : {
4846 [N in Selected<P & string>]: [M] extends [never] ? N : keyof M extends AnyKeyOf<Args<N>> ? N : never;
4847 }[Selected<P & string>];
4848
4849 /**
4850 * What a matched hook returns: the event's result, narrowed by `M` where the
4851 * result is a union tagged by the matcher's tag keys.
4852 */
4853 export type MatchedResult<P extends Pattern, M> = MatchedNames<P, M> extends infer N extends EventName ? N extends unknown ? Select<EventResult<N>, Selection<Args<N>, M>> : never : never;
4854
4855 /**
4856 * The hook `on(event, matcher, hook)` takes on a streaming event: the
4857 * generator form, `e` narrowed by the matcher, `next(e)` the stream beneath.
4858 */
4859 export type MatchedStreamHook<P extends StreamingEventName, M> = ($: EngineInterface, e: Frozen<MatchedEvent<P, M>>, next: MatchedStreamNext<P, M>) => StreamHookBody<Chunk<P>, MatchedResult<P, M>>;
4860
4861 /**
4862 * What a matched streaming hook's `next.is(pattern, e)` narrows `e` to:
4863 * the streaming event's argument narrowed by the matcher.
4864 */
4865 type MatchedStreamNarrowings<P extends StreamingEventName, M> = {
4866 [K in P]: Narrowed<Args<K>, M>;
4867 };
4868
4869 /**
4870 * A matched streaming hook's `next`: the variants the matcher keeps, the
4871 * result tagged the same way, `next.is` narrowing as the matcher does.
4872 */
4873 type MatchedStreamNext<P extends StreamingEventName, M> = StreamNext<P, KeptEvent<P, M>, MatchedResult<P, M>, MatchedStreamNarrowings<P, M>>;
4874
4875 /**
4876 * What `on(event, matcher, hook)` takes for an argument of type `I`: the
4877 * shape of the `e` the hook wants, a partial of it at any depth.
4878 *
4879 * A leaf is `===` or a RegExp (`{ command: /^p4 / }`); an array is any-of;
4880 * an object is a partial of an OBJECT, so `{ command: { startsWith } }` is
4881 * a type error where `e` is typed and free where it is `unknown`.
4882 */
4883 export type Matcher<I, All = I> = I extends unknown ? {
4884 readonly [K in KnownKeys<I>]?: MatcherValue<I[K], MatcherValueOf<All, K>>;
4885 } & (string extends keyof I ? OpenMatcher<I, All> : unknown) : never;
4886
4887 /**
4888 * Any matcher at all, for a field typed `unknown` (a tool's input, a
4889 * result's output): the kinds the engine accepts, unchecked there.
4890 *
4891 * The four kinds: a scalar is `===`; a RegExp tests the value as a string
4892 * (`{ command: /^p4 / }`); an array matches if any element does; an object
4893 * is a partial of an object. `{ startsWith: 'p4' }` never matches a string.
4894 */
4895 type MatcherData = string | number | boolean | null | RegExp | readonly MatcherData[] | {
4896 readonly [key: string]: MatcherData;
4897 };
4898
4899 /**
4900 * The declared keys of every variant of `I` (index signatures aside).
4901 */
4902 type MatcherKeys<I> = I extends unknown ? KnownKeys<I> : never;
4903
4904 /**
4905 * What matches one value of type `V`, by the runtime's kinds:
4906 * a scalar leaf takes the value or a RegExp; an object, a partial of it.
4907 *
4908 * For an array, what matches one ELEMENT of it, since a pattern against an
4909 * array value holds when some element matches; for `unknown`, any matcher.
4910 * A scalar matches by `===`, a RegExp tests the value as a string.
4911 */
4912 type MatcherOne<V> = unknown extends V ? MatcherData : V extends readonly (infer Item)[] ? MatcherOne<Item> : V extends string | number | boolean | null ? V | RegExp : V extends object ? Matcher<V> : V extends undefined ? never : unknown;
4913
4914 /**
4915 * What a matcher gives a key whose value is `V` on this variant and `Across`
4916 * over every variant: one MatcherOne, or an array of them matched as one-of.
4917 *
4918 * The one-of is typed over every variant, so `{ tool: ['Bash', 'Read'] }`
4919 * types on the Bash variant; with a nested pattern beside it, the pattern is
4920 * checked against a variant the one-of names, not against each of them.
4921 */
4922 type MatcherValue<V, Across = V> = MatcherOne<V> | readonly MatcherOne<Across>[];
4923
4924 /**
4925 * The type of key `K` across the variants of `I` that declare it.
4926 */
4927 type MatcherValueOf<I, K> = I extends unknown ? K extends KnownKeys<I> ? I[K] : never : never;
4928
4929 /**
4930 * One block of an MCP result: `type` and the fields that kind of block carries.
4931 */
4932 export type McpContentBlock = {
4933 /**
4934 * The block's kind: `text`, `image`, `audio`, `resource`, `resource_link`.
4935 */
4936 type: string;
4937 /**
4938 * Set on a `text` block.
4939 */
4940 text?: string;
4941 /**
4942 * Set on a `resource_link` (or embedded `resource`) block.
4943 */
4944 uri?: string;
4945 /**
4946 * Declared by an image, audio or resource block.
4947 */
4948 mimeType?: string;
4949 [field: string]: unknown;
4950 };
4951
4952 /**
4953 * The MCP server serving this tool, for `mcp__*` tools: `name` is the server's config key (for `source: "sdk"`, exactly the name the SDK host registered in `sdkMcpServers` / `mcp_set_servers`; for any other source, the key as authored in that configuration - untrusted text, the same value `mcp_status` and system/init report, to be escaped before display), `source` is where its definition came from - `sdk` (an in-process server the SDK host runs; only the host can register one, so a configured server of the same name never reads `sdk`), `plugin` (a server a plugin ships or registers at runtime), or a config scope (`user`, `project`, `local`, `dynamic` for --mcp-config / `mcp_set_servers` process servers, `managed`, `enterprise`, `claudeai`, `agent`). Key trust on `source`, not on the name or the tool-name prefix. Absent for non-MCP tools.
4954 */
4955 type McpServerProvenance = {
4956 name: string;
4957 /**
4958 * sdk | plugin | user | project | local | dynamic | managed | enterprise | claudeai | agent - an open set; treat unknown values as an unrecognized configured source, never as sdk.
4959 */
4960 source: string;
4961 };
4962
4963 /**
4964 * The MCP branch of a `tool.call` hook's `e` (and of `$.tool.call`'s input):
4965 * one variant per declared tool, else the loose McpToolCallInputFallback.
4966 *
4967 * `$.mcp.call(server, tool, args)` takes none of these: its parameters are
4968 * positional, the server and tool by name, then the arguments object.
4969 */
4970 export type McpToolCallInput = [keyof McpToolInputs] extends [never] ? McpToolCallInputFallback : {
4971 [N in keyof McpToolInputs & string]: ToolInputOf<N, McpToolInputs[N] & Record<string, unknown>>;
4972 }[keyof McpToolInputs & string];
4973
4974 /**
4975 * The `e` a `tool.call` (or `classic.PreToolUse`) hook receives for an MCP
4976 * tool while no MCP tool is declared: every `mcp__*` name, loose arguments.
4977 *
4978 * McpToolInputs has no entries until `/plugin-types` writes the connected
4979 * tools' declarations; also the `input` of `$.tool.call({ tool:
4980 * "mcp__<server>__<tool>", ... })`. Not `$.mcp.call`'s, which is positional:
4981 *
4982 * @example
4983 * on("tool.call", { tool: "mcp__gh__issue" }, ($, e, next) => audit(e.title))
4984 */
4985 type McpToolCallInputFallback = {
4986 /**
4987 * The name of the tool being called (`mcp__<server>__<tool>`); comparing
4988 * it narrows `e`. Reserved: a rewrite of it is ignored by core.
4989 */
4990 tool: McpToolName;
4991 /**
4992 * The tool_use block's id: the same at every event of the call and in
4993 * `$.ui.notice`. Reserved: a rewrite of it is ignored by core.
4994 */
4995 tool_use_id: string;
4996 [argument: string]: unknown;
4997 };
4998
4999 /**
5000 * The inputs of the MCP tools this project knows, keyed by full tool name,
5001 * for declaration merging; empty by default, then every MCP tool is loose.
5002 *
5003 * A `.d.ts` in the plugin author's project (written by `/plugin-types <dir>`
5004 * from the connected servers' JSON Schemas, or by hand) adds entries under
5005 * `declare module "claude-code"`; `e.tool === <name>` then narrows to them.
5006 *
5007 * @example
5008 * interface McpToolInputs { "mcp__my_server__send": { to: string } }
5009 */
5010 export interface McpToolInputs {
5011 }
5012
5013 /**
5014 * The name of an MCP tool as the engine spells it: `mcp__<server>__<tool>`.
5015 */
5016 export type McpToolName = `mcp__${string}__${string}`;
5017
5018 /**
5019 * An MCP tools/call result as the SDK returns it, plain data.
5020 */
5021 export type McpToolResult = {
5022 /**
5023 * The result's content blocks, in order (text, image, resource,
5024 * resource_link, ...).
5025 */
5026 content: McpContentBlock[];
5027 /**
5028 * True when the server reported the call as failed; the blocks then describe
5029 * the error.
5030 */
5031 isError: boolean;
5032 /**
5033 * The server's structured result, when its tool declares an output schema.
5034 */
5035 structuredContent?: unknown;
5036 };
5037
5038 /**
5039 * Hook input for the MessageDisplay event. Fired with each batch of newly completed lines while an assistant message streams. Display-only: the stored message and what the model sees are untouched.
5040 */
5041 type MessageDisplayHookInput = BaseHookInput & {
5042 hook_event_name: 'MessageDisplay';
5043 /**
5044 * UUID of the current turn.
5045 */
5046 turn_id: string;
5047 /**
5048 * UUID of the assistant message being displayed. Stable across every flush of the same message. Not the API msg_... id.
5049 */
5050 message_id: string;
5051 /**
5052 * Zero-based index of this delta within the message. Increments by one per flush.
5053 */
5054 index: number;
5055 /**
5056 * True on the message's last flush. Exactly one flush per message has it.
5057 */
5058 final: boolean;
5059 /**
5060 * The newly completed lines since the prior flush. Always whole lines, except on the final flush which may end mid-line. The delta of the final flush is empty when the message ends on a newline; treat final as the end-of-message signal regardless.
5061 */
5062 delta: string;
5063 };
5064
5065 /**
5066 * What `$.model.complete` takes.
5067 */
5068 export type ModelCompleteRequest = {
5069 /**
5070 * An alias (`haiku`) or a full model id; resolved and allowlist-checked like
5071 * a `--model` value.
5072 */
5073 model: string;
5074 /**
5075 * The one user message; the reply's text comes back.
5076 */
5077 prompt: string;
5078 /**
5079 * Precedes the completion as its system prompt, after the CLI's identity
5080 * block. Default none.
5081 */
5082 system?: string;
5083 /**
5084 * The reply's token cap: any positive integer up to what one reply can
5085 * hold, the model's own output limit or 64000, whichever is lower.
5086 *
5087 * Default 1024. The reply comes back whole, not streamed, and a provider
5088 * ends such a request at ten minutes, hence the 64000. One past the limit
5089 * is refused, naming it.
5090 */
5091 maxTokens?: number;
5092 };
5093
5094 /**
5095 * What `$.model.fork` takes.
5096 */
5097 export type ModelForkRequest = {
5098 /**
5099 * The one user message, appended to the session's own transcript.
5100 *
5101 * The reply's text and usage come back, or null on an API error or a cold
5102 * transcript.
5103 */
5104 prompt: string;
5105 };
5106
5107 /**
5108 * What `$.model.fork` resolves to when the fork answered: the reply's text
5109 * and what the fork cost.
5110 */
5111 export type ModelForkResult = {
5112 /**
5113 * The non-error replies' text, joined by newlines.
5114 */
5115 text: string;
5116 /**
5117 * The fork's token counts, so a plugin can account for what it spent.
5118 */
5119 usage: ModelForkUsage;
5120 };
5121
5122 /**
5123 * What one fork cost, as the API counted it: the four token counts of the
5124 * fork's completions summed.
5125 */
5126 export type ModelForkUsage = {
5127 input_tokens: number;
5128 output_tokens: number;
5129 cache_read_input_tokens: number;
5130 cache_creation_input_tokens: number;
5131 };
5132
5133 /**
5134 * The prefixes a glob may name: one or more whole leading segments of an
5135 * event name (`tool` of `tool.call`), derived, plugin nouns included.
5136 */
5137 export type Namespace<N extends string = EventName> = N extends `${infer Head}.${infer Rest}` ? Head | `${Head}.${Namespace<Rest>}` : never;
5138
5139 /**
5140 * How many object or array levels a matcher narrows `e` through, counted as
5141 * a tuple's length: the runtime's own limit, which refuses a deeper matcher.
5142 *
5143 * Past it a field keeps its declared type; nothing becomes `any`.
5144 */
5145 type NarrowDepth = 8;
5146
5147 /**
5148 * `e` in a matched hook: the members of the argument union `E` matcher `P`
5149 * can match, each with the keys `P` names narrowed to what a match implies.
5150 *
5151 * A scalar narrows to the literal, a one-of to what its alternatives give,
5152 * an object key recursively, an array some element of which must match to
5153 * a non-empty tuple; other keys, `unknown` and RegExp-matched fields keep.
5154 */
5155 export type Narrowed<E, P> = E extends unknown ? NarrowedMember<E, P, []> : never;
5156
5157 /**
5158 * A value of declared type `V` under a one-of, folded over the tuple into
5159 * `Found`: the union of what each alternative narrows `V` to (NarrowedByOne).
5160 *
5161 * An alternative that cannot match adds nothing, so a one-of none of whose
5162 * alternatives can is `never`; a one-of typed as a plain array rather than
5163 * a tuple narrows by the union of its elements at once.
5164 */
5165 type NarrowedByAny<V, Alternatives extends readonly unknown[], D extends readonly unknown[], Found = never> = Alternatives extends readonly [infer First, ...infer Rest] ? NarrowedByAny<V, Rest, D, Found | NarrowedByOne<V, First, D>> : Alternatives extends readonly [] ? Found : Found | NarrowedByOne<V, Alternatives[number], D>;
5166
5167 /**
5168 * A value of declared type `V` under one matcher node `Q` that is not a
5169 * one-of, member of `V` by member; a member that cannot match is `never`.
5170 *
5171 * An array some element of which must match becomes NonEmpty; a RegExp
5172 * keeps the member; an object pattern recurses into an object member, one
5173 * level down; a scalar keeps a member as narrow, and replaces a wider one.
5174 */
5175 type NarrowedByOne<V, Q, D extends readonly unknown[]> = V extends readonly (infer Item)[] ? [NarrowedValue<Item, Q, [...D, unknown]>] extends [never] ? never : NonEmpty<V, Item> : Q extends RegExp ? V : Q extends object ? V extends object ? NarrowedMember<V, Q, [...D, unknown]> : never : V extends Q ? V : Q extends V ? Q : never;
5176
5177 /**
5178 * One object member `E` under object pattern `P`, `D` levels down: `never`
5179 * when a key of `P` is impossible on it (ImpossibleKeys), else `E` narrowed.
5180 *
5181 * Each key `P` names is narrowed (NarrowedValue); every other key, and each
5182 * key's optionality, stays as declared.
5183 */
5184 type NarrowedMember<E, P, D extends readonly unknown[]> = [
5185 ImpossibleKeys<E, P, D>
5186 ] extends [never] ? {
5187 [K in keyof E]: K extends keyof P ? NarrowedValue<E[K], P[K], D> : E[K];
5188 } : never;
5189
5190 /**
5191 * A value of declared type `V` where the matcher gives `Q`, `D` levels
5192 * down: NarrowedByAny under a one-of (an array), NarrowedByOne otherwise.
5193 *
5194 * As declared once `D` reaches NarrowDepth, or for a field typed `unknown`
5195 * (a tool's input), which no pattern narrows.
5196 */
5197 type NarrowedValue<V, Q, D extends readonly unknown[]> = D['length'] extends NarrowDepth ? V : unknown extends V ? V : Q extends readonly unknown[] ? NarrowedByAny<V, Q, D> : NarrowedByOne<V, Q, D>;
5198
5199 /**
5200 * `!` before a name or a glob: every event except the ones it selects. `!*`
5201 * would select none, so it is no pattern.
5202 */
5203 type Negation = `!${Exclude<EventName | Glob, '*'>}`;
5204
5205 /**
5206 * The rest of the chain, as one hook receives it: made once per dispatch per
5207 * hook, frozen; `next(e)` resolves to the downstream result.
5208 *
5209 * Each call runs the hooks below again; core is the last, and below it `next`
5210 * rejects. Called with no argument it rejects, naming the hook. A hook that
5211 * returns without calling it ends the chain; returning nothing is a failure.
5212 *
5213 * @template S what `next.is(pattern, e)` narrows `e` to, per event: the event's
5214 * argument, or, under a matcher, that argument narrowed by it (MatchedHook)
5215 * @template T the tool `e` names, when it names one: on `tool.call` it types
5216 * the result (NextResultFor); an `e` without `tool` takes the line beneath
5217 */
5218 export type Next<N extends EventName = EventName, E = Args<N>, O = NextResult<N>, S extends {
5219 [K in N]?: unknown;
5220 } = {
5221 [K in N]: Args<K>;
5222 }> = {
5223 <T extends string>(e: E & ToolNamed<T>): Promise<NextResultFor<N, O, T>>;
5224 (e: E): Promise<O>;
5225 /**
5226 * Continues this dispatch at a tier: `next(e)` with every link between
5227 * this hook's own tier and that one skipped, by tier and narrowing only.
5228 *
5229 * A managed hook's: prepend may name append, builtin or core, append core;
5230 * it never skips a tier with more authority, so no user hook skips the
5231 * org's. A literal on the hook's own `next`; skipped links are traced.
5232 */
5233 readonly to: {
5234 <T extends string>(e: E & ToolNamed<T>, tier: TargetTier): Promise<NextResultFor<N, O, T>>;
5235 (e: E, tier: TargetTier): Promise<O>;
5236 };
5237 /**
5238 * Aborts when the call this dispatch belongs to is abandoned: the user
5239 * interrupted, a hook above settled first, or this hook ran out of budget.
5240 *
5241 * Anything the hook started (timers, requests) should stop on it. It is an
5242 * AbortSignal of the plugin's own environment, driven by the chain's.
5243 */
5244 readonly signal: AbortSignal;
5245 /**
5246 * Whether this dispatch's event is selected by `pattern` (a name, a glob,
5247 * a negation), as a type predicate on `e`: `next.is("tool.call", e)`.
5248 *
5249 * Under a matcher the narrowing includes it. `pattern` names events this
5250 * hook covers (PatternOver); a name outside them is a compile error.
5251 */
5252 readonly is: <M extends PatternOver<N>>(pattern: M, e: unknown) => e is Frozen<S[Extract<N, Selected<M>>]>;
5253 /**
5254 * The name of this dispatch's event, as a value, for a glob hook to log or
5255 * switch on.
5256 */
5257 readonly event: N;
5258 /**
5259 * Who raised this dispatch: the calling plugin's name and the tier it sits
5260 * in (Origin); the engine reads `{ plugin: "engine", tier: "core" }`.
5261 *
5262 * Set by the host alone, from the environment the call came from (its own
5263 * MessagePort) and that plugin's seat; nothing a plugin writes reaches it.
5264 * Every hook of one dispatch sees the same origin, and `next.to` keeps it.
5265 *
5266 * @example
5267 * return next.origin.tier === "prepend" ? next.to(e, "append") : next(e)
5268 */
5269 readonly origin: Origin;
5270 /**
5271 * What settled beneath this hook on its latest `next()` call, the one
5272 * started last: an entry per link beneath, nearest first, the engine's last.
5273 *
5274 * Empty before `next` is called; filled even when `next` rejected; a link
5275 * still running joins in place later, nothing listed leaves; it ends short
5276 * of the engine at a link that answered its last call itself. Data, frozen.
5277 */
5278 readonly trace: readonly TraceEntry<N, E, O>[];
5279 /**
5280 * How much time this hook runs under and how much is left (NextBudget),
5281 * read fresh on each access; HookBudget says what the clock counts.
5282 *
5283 * It stands still while a `next` or `$` call of the hook's is in flight;
5284 * in a `.catch` handler it is the grace, `ms` being `next.error.budget`.
5285 *
5286 * @example
5287 * if (next.budget.remainingMs < 2_000) return next(e) // skip the polish
5288 */
5289 readonly budget: NextBudget;
5290 };
5291
5292 /**
5293 * The budget the code reading `next.budget` runs under: the whole allowance
5294 * and what is left of it now, plain data read fresh on each access.
5295 *
5296 * In a hook `ms` is HookBudget's `ms`; in a `.catch` handler its `catchMs`
5297 * (equal to `next.error.budget`); on a `next` nothing meters (the engine's
5298 * own, a test's root one) `ms` is 0 and `remainingMs` Infinity.
5299 *
5300 * @example
5301 * if (next.budget.remainingMs < 2_000) return next(e) // skip the slow path
5302 */
5303 export type NextBudget = {
5304 /**
5305 * The whole budget in milliseconds; 0 where nothing meters this `next`.
5306 */
5307 readonly ms: number;
5308 /**
5309 * What is left of it now, in milliseconds, never below 0; Infinity where
5310 * nothing meters. It stands still while a `next` or `$` call is in flight.
5311 */
5312 readonly remainingMs: number;
5313 };
5314
5315 /**
5316 * What `next(e)` resolves to for event `N`: the event's result, except at
5317 * `engine.create`, where the steps beneath return `$` as built so far.
5318 *
5319 * A withheld noun is on that `$` as a stub, so the built table is typed
5320 * whole where what a hook returns (EngineCreateResult) is partial.
5321 */
5322 export type NextResult<N extends EventName> = N extends 'engine.create' ? EngineInterfaceBuilt : EventResult<N>;
5323
5324 /**
5325 * What `next(e)` resolves to once `e.tool` is the literal `T`: on `tool.call`
5326 * the result typed for that tool; on every other event, `O` as declared.
5327 *
5328 * `result` is Bash's record after `e.tool === "Bash"`; an un-narrowed `e`
5329 * names every tool, and `result` stays `unknown`.
5330 */
5331 export type NextResultFor<N extends EventName, O, T extends string> = [
5332 N
5333 ] extends ['tool.call'] ? ToolCallResult<T> : O;
5334
5335 /**
5336 * The argument of a call on `$` that takes nothing (`$.session.cwd()`): an
5337 * object with no keys.
5338 */
5339 type NoArgs = Record<never, never>;
5340
5341 /**
5342 * The events whose argument is exactly NoArgs (`session.cwd`, a declared
5343 * plugin noun's `() => ...`); their overloads come last (LateOverload).
5344 */
5345 type NoArgsEvent = {
5346 [N in EventName]: Args<N> extends NoArgs ? NoArgs extends Args<N> ? N : never : never;
5347 }[EventName];
5348
5349 /**
5350 * What an `engine.create` hook receives as `$`: nothing. Every property
5351 * reads as `never`, so `$.model` inside the hook is a compile error.
5352 */
5353 export type NoEngineInterface = {
5354 readonly [noun: string]: never;
5355 };
5356
5357 /**
5358 * Array type `V`, of element `Item`, once some element of it is known to
5359 * match: a tuple of at least one `Item`, readonly when `V` is.
5360 *
5361 * So `e` still passes wherever the declared array is taken; a `V` that is
5362 * already a non-empty tuple is kept as it is.
5363 */
5364 type NonEmpty<V, Item> = V extends readonly [unknown, ...unknown[]] ? V : V extends Item[] ? [Item, ...Item[]] : readonly [Item, ...Item[]];
5365
5366 type NotificationHookInput = BaseHookInput & {
5367 hook_event_name: 'Notification';
5368 message: string;
5369 title?: string;
5370 notification_type: string;
5371 };
5372
5373 /**
5374 * The declared plugin nouns' methods as event rows (NounEventRow), one per
5375 * `<noun>.<method>` that is a function; a member that is not is no event.
5376 */
5377 type NounEvent = {
5378 [K in PluginNoun]: {
5379 [M in keyof EngineInterface[K] & string]: EngineInterface[K][M] extends (...args: infer Parameters) => infer Result ? NounEventRow<`${K}.${M}`, Parameters extends readonly [] ? NoArgs : Parameters[0], Awaited<Result>> : never;
5380 }[keyof EngineInterface[K] & string];
5381 }[PluginNoun];
5382
5383 /**
5384 * The name of a declared plugin noun's event (`voice.speak`).
5385 */
5386 type NounEventName = keyof NounEventOf & string;
5387
5388 /**
5389 * The events of the plugin nouns declared on EngineInterface, by name: the
5390 * argument of each `<noun>.<method>`. Empty until a plugin declares a noun.
5391 *
5392 * @example
5393 * declare module "claude-code" { interface EngineInterface { voice: Voice } }
5394 */
5395 export type NounEventOf = {
5396 [E in NounEvent as E['name']]: E['args'];
5397 };
5398
5399 /**
5400 * The result of a declared plugin noun's event as its hooks see it:
5401 * `{ value }` (the method's answer) or `{ deny }`.
5402 */
5403 type NounEventResult<N extends NounEventName> = ValueOrDeny<NounValueOf[N]>;
5404
5405 /**
5406 * One method of a declared plugin noun as an event row: its event's name,
5407 * argument (the method's first parameter) and value (its awaited result).
5408 */
5409 type NounEventRow<Name extends string, Args, Value> = {
5410 name: Name;
5411 args: Args;
5412 value: Value;
5413 };
5414
5415 /**
5416 * What each declared plugin noun's method answers (the `value` of its
5417 * event's result), by event name.
5418 */
5419 type NounValueOf = {
5420 [E in NounEvent as E['name']]: E['value'];
5421 };
5422
5423 /**
5424 * Registers `hook` on the events `pattern` selects: one by name, every one
5425 * under a namespace (`classic.*`), all (`*`), or all but some (`!tool.*`).
5426 *
5427 * One function stands on every selected event (`next.event` says which),
5428 * under a matcher for the inputs it matches; a plugin's registrations nest
5429 * in order, first outermost; a repeat throws. Returns the Registration.
5430 *
5431 * @see HookBudget the time each hook has per dispatch (its own time: waits
5432 * on `next` and `$` are free), read live from `next.budget`
5433 */
5434 export type On = {
5435 <P extends Pattern>(pattern: P, hook: NoInfer<HookFor<P>>): Registration<HookFor<P>>;
5436 <P extends Pattern, const M extends Matcher<Args<MatchedNames<P>>>>(pattern: P, matcher: M, hook: NoInfer<MatchedHook<P, M>>): Registration<MatchedHook<P, M>>;
5437 };
5438
5439 /**
5440 * The part of a transcript message the surface that drew it has on screen:
5441 * units `first` to `last` of the message's `of`, counted from its start.
5442 *
5443 * The terminal counts the rows the site laid out, from its first (the blank
5444 * row the engine draws above a message is its row 0; a hook's tree starts at
5445 * its own); other surfaces count in their unit, so compare `first / of` there.
5446 *
5447 * @remarks While `null` the terminal holds the site at the rows it last laid
5448 * out (a scroller's geometry needs real rows); on screen, draw anything.
5449 * @example if (e.props.onScreen) shown.set(e.requestId, e.props.onScreen)
5450 * else shown.delete(e.requestId) // the band's legend lists `shown`
5451 */
5452 type OnScreen = {
5453 /**
5454 * The unit inside the viewport, from 0: on the terminal, `7` when the
5455 * message's top seven rows are scrolled away.
5456 */
5457 first: number;
5458 /**
5459 * The unit inside the viewport, inclusive; at least `first`.
5460 */
5461 last: number;
5462 /**
5463 * The message's whole extent in the same unit, as laid out (a hook's tree
5464 * taller than the engine's counts its own rows); more than `last`.
5465 *
5466 * One surface's numbers say nothing of another's: what the phone shows the
5467 * terminal may not, and each `ui.render` carries its own surface's.
5468 */
5469 of: number;
5470 };
5471
5472 /**
5473 * The keys a variant with a string index signature (an MCP tool's input)
5474 * takes beyond its own: another variant's key as typed there; others free.
5475 *
5476 * So a misspelt value for a key some variant declares (`command: 5`) is
5477 * refused on every variant, not admitted by the open one.
5478 */
5479 type OpenMatcher<I, All> = {
5480 readonly [K in Exclude<MatcherKeys<All>, KnownKeys<I>>]?: MatcherValue<MatcherValueOf<All, K>>;
5481 } & Readonly<Record<string, unknown>>;
5482
5483 /**
5484 * The name of a call on `$` the host serves, as an event.
5485 */
5486 export type OpEventName = keyof OpEventOf;
5487
5488 /**
5489 * The calls on `$` the host serves, as events: `e` is the call's argument as
5490 * it crosses to the host, and every one is hookable by name and by `on("*")`.
5491 *
5492 * A hook above the caller passes it on, rewrites it, refuses it with
5493 * `{ deny }` or answers with `{ value }`; core is the host's implementation.
5494 * The calling hook alone is skipped, and `next.origin` names the caller.
5495 */
5496 export type OpEventOf = {
5497 /**
5498 * The argument of `$.model.complete(request)`.
5499 */
5500 'model.complete': ModelCompleteRequest;
5501 /**
5502 * The argument of `$.model.classify(text, labels, options)`.
5503 */
5504 'model.classify': {
5505 text: string;
5506 labels: readonly string[];
5507 options?: ClassifyOptions;
5508 };
5509 /**
5510 * The argument of `$.model.fork(request)`.
5511 */
5512 'model.fork': ModelForkRequest;
5513 /**
5514 * The clip and how to play it (`shouldLoop`, `gain`); the signal does not
5515 * cross.
5516 */
5517 'audio.play': {
5518 clip: AudioClip;
5519 shouldLoop: boolean;
5520 gain?: number;
5521 };
5522 /**
5523 * The argument of `$.audio.speak(text, { voice })`.
5524 */
5525 'audio.speak': SpeakRequest;
5526 /**
5527 * The argument of `$.mcp.call(server, tool, args)`.
5528 */
5529 'mcp.call': {
5530 server: string;
5531 tool: string;
5532 args: Record<string, unknown>;
5533 };
5534 /**
5535 * The argument of `$.session.cwd()`.
5536 */
5537 'session.cwd': NoArgs;
5538 /**
5539 * The argument of `$.session.root()`.
5540 */
5541 'session.root': NoArgs;
5542 /**
5543 * The argument of `$.session.model()`.
5544 */
5545 'session.model': NoArgs;
5546 /**
5547 * The argument of `$.session.turns()`.
5548 */
5549 'session.turns': NoArgs;
5550 /**
5551 * The argument of `$.session.id()`.
5552 */
5553 'session.id': NoArgs;
5554 /**
5555 * The argument of `$.session.messages()`.
5556 */
5557 'session.messages': NoArgs;
5558 /**
5559 * The argument of `$.session.repo()`.
5560 */
5561 'session.repo': NoArgs;
5562 /**
5563 * The argument of `$.session.surface()`.
5564 *
5565 * @deprecated with `$.session.surface()`; hook `session.surfaces`
5566 */
5567 'session.surface': NoArgs;
5568 /**
5569 * The argument of `$.session.surfaces()`.
5570 */
5571 'session.surfaces': NoArgs;
5572 /**
5573 * The argument of `$.session.authorize()`.
5574 */
5575 'session.authorize': NoArgs;
5576 /**
5577 * The argument of `$.session.usage({ breakdown, columns })`.
5578 */
5579 'session.usage': SessionUsageArgs;
5580 /**
5581 * The argument of `$.session.version()`.
5582 */
5583 'session.version': NoArgs;
5584 /**
5585 * The argument of `$.turn.abort({ turnId })`.
5586 */
5587 'turn.abort': {
5588 turnId: string;
5589 };
5590 /**
5591 * The argument of `$.prompt.read()`.
5592 */
5593 'prompt.read': NoArgs;
5594 /**
5595 * The argument of `$.tool.list()`.
5596 */
5597 'tool.list': NoArgs;
5598 /**
5599 * The argument of `$.tool.register(spec)`.
5600 */
5601 'tool.register': Required<ToolSpec>;
5602 /**
5603 * The argument of `$.command.list()`.
5604 */
5605 'command.list': NoArgs;
5606 /**
5607 * The argument of `$.command.register(spec)`.
5608 */
5609 'command.register': CommandSpec;
5610 /**
5611 * The argument of `$.config.list()`.
5612 */
5613 'config.list': NoArgs;
5614 /**
5615 * The argument of `$.agent.list()`.
5616 */
5617 'agent.list': NoArgs;
5618 /**
5619 * The argument of `$.agent.register(spec)`: the agent type as the plugin
5620 * defined it. A hook above rewrites any of it; the type stays the caller's.
5621 */
5622 'agent.register': AgentSpec;
5623 /**
5624 * The argument of `$.ui.toast(text, { timeoutMs })`.
5625 */
5626 'ui.toast': {
5627 text: string;
5628 timeoutMs?: number;
5629 };
5630 /**
5631 * The argument of `$.ui.status(text)`; `text` undefined clears the line.
5632 */
5633 'ui.status': {
5634 text: string | undefined;
5635 };
5636 /**
5637 * The argument of `$.ui.log(text, { to })`; `to` is always present
5638 * (UiLogSink), and `next({ ...e, to: "debug" })` keeps a line off screen.
5639 */
5640 'ui.log': {
5641 text: string;
5642 to: UiLogSink;
5643 };
5644 /**
5645 * The argument of `$.ui.notice(tool_use_id, text)`.
5646 */
5647 'ui.notice': {
5648 tool_use_id: string;
5649 text: string | undefined;
5650 };
5651 /**
5652 * The argument of `$.ui.invalidate(event)`.
5653 */
5654 'ui.invalidate': {
5655 event: InvalidatableEventName;
5656 };
5657 /**
5658 * The argument of `$.ui.open({ id, title, focus })`; a hook above the
5659 * opener may retitle it or refuse it with `{ deny }`, never rename it.
5660 */
5661 'ui.open': PaneOpenArgs;
5662 /**
5663 * The argument of `$.ui.close({ id })` with `origin` `plugin`; the engine
5664 * raises it too, for the person (`person`) and an unload (`unload`).
5665 */
5666 'ui.close': PaneCloseInput;
5667 /**
5668 * The argument of `$.ui.panes()`.
5669 */
5670 'ui.panes': NoArgs;
5671 /**
5672 * The argument of `$.ui.blit(...)`: a Raster's `cells` or a keyed Image's
5673 * `source`; a hook above may rewrite either with `next`, or `{ deny }`.
5674 */
5675 'ui.blit': UiBlitArgs;
5676 /**
5677 * The argument of `$.fs.read(path, { as })`: `as` is `text` unless the
5678 * caller asked for `bytes`.
5679 */
5680 'fs.read': {
5681 path: string;
5682 as: FsReadAs;
5683 };
5684 /**
5685 * The argument of `$.fs.write(path, text)`.
5686 */
5687 'fs.write': {
5688 path: string;
5689 text: string;
5690 };
5691 /**
5692 * The argument of `$.fs.list(path)`.
5693 */
5694 'fs.list': {
5695 path: string;
5696 };
5697 /**
5698 * The argument of `$.fs.exists(path)`.
5699 */
5700 'fs.exists': {
5701 path: string;
5702 };
5703 /**
5704 * The argument of `$.fs.stat(path, { resolve })`: `resolve` is false
5705 * unless the caller asked where the path lands.
5706 */
5707 'fs.stat': {
5708 path: string;
5709 resolve: boolean;
5710 };
5711 /**
5712 * The argument of `$.fs.ancestors({ names, of, below })`.
5713 */
5714 'fs.ancestors': FsAncestorsRequest;
5715 /**
5716 * The argument of `$.store.get(key)`.
5717 */
5718 'store.get': {
5719 key: string;
5720 };
5721 /**
5722 * The argument of `$.store.set(key, value)`.
5723 */
5724 'store.set': {
5725 key: string;
5726 value: unknown;
5727 };
5728 /**
5729 * The argument of `$.store.delete(key)`.
5730 */
5731 'store.delete': {
5732 key: string;
5733 };
5734 /**
5735 * The argument of `$.store.keys()`.
5736 */
5737 'store.keys': NoArgs;
5738 /**
5739 * The argument of `$.clock.now()`.
5740 */
5741 'clock.now': NoArgs;
5742 /**
5743 * The argument of `$.clock.sleep(ms, { signal })`; the signal does not
5744 * cross, it aborts the dispatch.
5745 */
5746 'clock.sleep': ClockWait;
5747 /**
5748 * The argument of `$.clock.after(ms, fn)`: the wait before `fn`, which
5749 * stays in the plugin's environment and runs once the dispatch resolves.
5750 */
5751 'clock.after': ClockWait;
5752 /**
5753 * The argument of `$.clock.every(ms, fn)`, dispatched once per period:
5754 * `fn` runs each time a dispatch resolves, and the next period is asked.
5755 */
5756 'clock.every': ClockWait;
5757 /**
5758 * The argument of `$.http.fetch(url, init)`.
5759 */
5760 'http.fetch': {
5761 url: string;
5762 init?: HttpInit;
5763 };
5764 /**
5765 * The argument of `$.process.run(argv, init)`.
5766 */
5767 'process.run': {
5768 argv: readonly string[];
5769 init?: ProcessRunInit;
5770 };
5771 /**
5772 * The argument of `$.settings.read({ source })`.
5773 */
5774 'settings.read': SettingsReadArgs;
5775 /**
5776 * The argument of `$.env.get(name)`; `name` is identity, pinned.
5777 */
5778 'env.get': {
5779 name: string;
5780 };
5781 /**
5782 * The argument of `$.env.set(name, value)`; `name` is identity, pinned,
5783 * and no `value` unsets.
5784 */
5785 'env.set': {
5786 name: string;
5787 value?: string;
5788 };
5789 };
5790
5791 /**
5792 * The result of a call on `$` as its event's hooks see it: `{ value }` (the
5793 * call's answer) or `{ deny }`.
5794 */
5795 export type OpEventResult<N extends OpEventName = OpEventName> = ValueOrDeny<OpValueOf[N]>;
5796
5797 /**
5798 * What each call on `$` answers (the `value` of its event's result), by event
5799 * name.
5800 */
5801 export type OpValueOf = {
5802 'model.complete': string;
5803 'model.classify': string | undefined;
5804 'model.fork': ModelForkResult | null;
5805 'audio.play': void;
5806 'audio.speak': SpeakResult;
5807 'mcp.call': McpToolResult;
5808 'session.cwd': string;
5809 'session.root': string;
5810 'session.model': string;
5811 'session.turns': number;
5812 'session.id': string;
5813 'session.messages': SessionMessage[];
5814 'session.repo': SessionRepo | null;
5815 'session.surface': RenderSurface | null;
5816 'session.surfaces': readonly RenderSurface[];
5817 /**
5818 * The credential handle, or null where the build or the provider has no
5819 * first-party credential to hold.
5820 */
5821 'session.authorize': SessionAuthorization;
5822 'session.usage': SessionUsage;
5823 /**
5824 * The engine's version, its release, and its build time when stamped.
5825 */
5826 'session.version': SessionVersion;
5827 'turn.abort': void;
5828 /**
5829 * The box as it stands; the empty box where the session draws none.
5830 */
5831 'prompt.read': PromptBox;
5832 'tool.list': ToolInfo[];
5833 'tool.register': {
5834 tool: string;
5835 };
5836 'command.list': CommandInfo[];
5837 'command.register': {
5838 command: string;
5839 };
5840 'config.list': ConfigRow[];
5841 'agent.list': AgentInfo[];
5842 'agent.register': {
5843 agent: string;
5844 };
5845 'ui.toast': void;
5846 'ui.status': void;
5847 'ui.log': void;
5848 'ui.notice': void;
5849 'ui.invalidate': void;
5850 'ui.open': void;
5851 'ui.close': void;
5852 /**
5853 * The calling plugin's open panes, placed then unplaced, in open order.
5854 */
5855 'ui.panes': readonly UiPane[];
5856 'ui.blit': UiBlitResult;
5857 /**
5858 * The text; `{ base64 }` when asked for bytes.
5859 */
5860 'fs.read': string | FsBytes;
5861 'fs.write': void;
5862 'fs.list': FsEntry[];
5863 'fs.exists': boolean;
5864 'fs.stat': FsStat;
5865 'fs.ancestors': readonly FsAncestor[];
5866 'store.get': unknown;
5867 'store.set': void;
5868 'store.delete': void;
5869 'store.keys': string[];
5870 /**
5871 * Milliseconds since the epoch.
5872 */
5873 'clock.now': number;
5874 'clock.sleep': void;
5875 'clock.after': void;
5876 'clock.every': void;
5877 'http.fetch': HttpResponse;
5878 'process.run': ProcessRunResult;
5879 'settings.read': Settings;
5880 'env.get': string | undefined;
5881 'env.set': void;
5882 };
5883
5884 /**
5885 * One call signature per event in `Names`, intersected, the ambiguous ones
5886 * (LateOverload) after the rest: `next` for a hook covering several events.
5887 *
5888 * An empty group contributes nothing (Overloads<never> is unknown).
5889 */
5890 type OrderedOverloads<Names extends EventName> = Overloads<Exclude<Names, LateOverload>> & Overloads<Extract<Names, 'classic.PreToolUse'>> & Overloads<Extract<Names, 'turn.abort'>> & Overloads<Extract<Names, NoArgsEvent>>;
5891
5892 /**
5893 * Who raised a dispatch, as `next.origin` holds it: the calling plugin's
5894 * name and the tier it sits in; the engine reads `engine` in `core`.
5895 *
5896 * The same pair a `next.trace` entry names its link by, and an event's
5897 * `provider` its subject's definer by. Set by the host from where the call
5898 * came from and where that plugin was seated; nothing a plugin writes.
5899 */
5900 export type Origin = {
5901 /**
5902 * Whose hook made the `$` call, by name; `"engine"` for a call site,
5903 * `"client"` for a `Client` surface module's `ui.message` post.
5904 */
5905 readonly plugin: string;
5906 /**
5907 * Where that plugin sits among the chain's tiers (Tier); `"core"` for the
5908 * engine, the owning plugin's for a `client` post.
5909 */
5910 readonly tier: Tier;
5911 };
5912
5913 /**
5914 * One call signature per event in `Names`, intersected into an overload set.
5915 */
5916 type Overloads<Names extends EventName> = UnionToIntersection<{
5917 [N in Names]: (e: Args<N>) => Promise<NextResult<N>>;
5918 }[Names]>;
5919
5920 /**
5921 * The argument of `$.ui.close`: the pane to close (`{ id }`). `origin` is
5922 * the engine's to set: a plugin's call reads `plugin` at the hooks.
5923 */
5924 export type PaneCloseArgs = Omit<PaneCloseInput, 'origin'>;
5925
5926 /**
5927 * The input of `ui.close`: the pane closing and why (PaneCloseOrigin).
5928 * Closing an id that is not open does nothing.
5929 */
5930 export type PaneCloseInput = {
5931 /**
5932 * What `$.ui.open` named the pane; pinned: `next(e)` passes it on.
5933 */
5934 id: string;
5935 /**
5936 * Who closes it, set by the engine: a hook that answers without `next`
5937 * keeps the pane open on `plugin` and `person`, never on `unload`.
5938 */
5939 origin: PaneCloseOrigin;
5940 };
5941
5942 /**
5943 * Why a pane closes, as the engine stamped it at `ui.close`.
5944 *
5945 * `unload` drops a pane nothing draws any more (its plugin unloaded, or its
5946 * drawing threw): it is gone before the hooks hear of it, and its opener's
5947 * hooks do not run. `next(e)` passes the origin on as received; none sets it.
5948 */
5949 export type PaneCloseOrigin = {
5950 /**
5951 * `plugin`, a plugin's `$.ui.close`; `person`, the person's close mark or
5952 * close key; `unload`, the engine's own.
5953 */
5954 kind: 'plugin' | 'person' | 'unload';
5955 };
5956
5957 /**
5958 * The argument of `$.ui.open`: which pane, its title, whether it asks the
5959 * person's keyboard, its dialog manners, its size: rows inline, columns docked.
5960 *
5961 * An open answering the person's input (a command or prompt they entered, a
5962 * press) is placed at any width; one the plugin makes on its own waits
5963 * undrawn below 144 terminal columns, 110 once they asked for that id.
5964 */
5965 export type PaneOpenArgs = {
5966 /**
5967 * Names the pane: 1-64 of letters, digits, `_` and `-`. One pane per id:
5968 * opening an open id delivers the new title, never a second instance.
5969 *
5970 * Pinned at the hooks: `next(e)` passes it on; the title and focus rewrite.
5971 */
5972 id: string;
5973 /**
5974 * The pane's tab while more than one pane is open (with one, no title is
5975 * drawn), and the `title` its hook sees; the id when omitted.
5976 *
5977 * An unpaired surrogate half (a `.slice()` through an emoji) is drawn as
5978 * U+FFFD; a control character is refused.
5979 */
5980 title?: string;
5981 /**
5982 * A request, not a grant: the surface focuses (and raises) the pane only
5983 * while the prompt has the keys over an empty composer.
5984 *
5985 * An element of the band or a pane the person holds, text in the
5986 * composer, a dialog or a survey each refuse it: the pane opens without
5987 * the keyboard.
5988 */
5989 focus?: true;
5990 /**
5991 * While the pane holds the keyboard, the key that hands it back (Escape)
5992 * also closes it as the person's close does: `ui.close`, origin `person`.
5993 *
5994 * So does Escape at an idle, empty prompt once the prompt has the keys
5995 * again (the person typed, or `focus` was refused); over text, a running
5996 * turn, a dialog, a footer selection, a viewed agent's transcript or a
5997 * prompt mode of its own the key stays theirs.
5998 *
5999 * A hook may refuse that close and keep it open. Left out, Escape returns
6000 * the keys to the prompt and the pane stays. Each open sets it anew, as it
6001 * sets the title.
6002 */
6003 closeOnEscape?: true;
6004 /**
6005 * While the pane is on screen the surface holds its transient toasts (the
6006 * notification line under the prompt) and shows them once it closes.
6007 *
6008 * As it does behind the engine's own side panel; pinned warnings still
6009 * show. Left out, toasts show as they come. Each open sets it anew.
6010 */
6011 holdToasts?: true;
6012 /**
6013 * The body rows the pane's content wants while seated inline above the
6014 * prompt: it opens that tall, up to what the layout spares, not a third.
6015 *
6016 * A request, not a grant: a size the person dragged or keyed the block
6017 * to wins, this session's or a kept one, and the dock ignores it. A
6018 * positive whole number; left out, a third. Each open sets it anew.
6019 */
6020 rows?: number;
6021 /**
6022 * The body columns the pane's content wants while docked beside a
6023 * fullscreen transcript: the dock opens that wide, floor to ceiling.
6024 *
6025 * A request, not a grant: a width the person dragged or keyed the dock
6026 * to wins, this session's or a kept one, and the inline block ignores
6027 * it. A positive whole number; left out, the share. Each open sets it anew.
6028 *
6029 * @example
6030 * await $.ui.open({ id: "browser", focus: true, rows: 24, columns: 100 })
6031 */
6032 columns?: number;
6033 };
6034
6035 /**
6036 * What `on(pattern, hook)` and `next.is(pattern, e)` take: an event's name,
6037 * a glob (`*`, `classic.*`), or a negation of either (`!tool.describe`).
6038 */
6039 export type Pattern = EventName | Glob | Negation;
6040
6041 /**
6042 * The patterns `next.is` takes in a hook covering the events `N`: their
6043 * names, `*`, a glob over one of their namespaces, or a negation.
6044 *
6045 * A name none of them has is a compile error, as is a glob over a namespace
6046 * none is under; a negation that selects none of them narrows `e` to never.
6047 */
6048 type PatternOver<N extends EventName> = N | '*' | `${Namespace<N>}.*` | Negation;
6049
6050 type PermissionBehavior = 'allow' | 'deny' | 'ask';
6051
6052 type PermissionDeniedHookInput = BaseHookInput & {
6053 hook_event_name: 'PermissionDenied';
6054 tool_name: string;
6055 tool_input: unknown;
6056 tool_use_id: string;
6057 reason: string;
6058 mcp_server?: McpServerProvenance;
6059 };
6060
6061 /**
6062 * Permission mode for controlling how tool executions are handled. 'default' - Standard behavior, prompts for dangerous operations. 'acceptEdits' - Auto-accept file edit operations. 'bypassPermissions' - Bypass all permission checks (requires allowDangerouslySkipPermissions). 'plan' - Planning mode, no actual tool execution. 'dontAsk' - Don't prompt for permissions, deny if not pre-approved. 'auto' - Use a model classifier to approve/deny permission prompts.
6063 */
6064 type PermissionMode = 'default' | 'acceptEdits' | 'bypassPermissions' | 'plan' | 'dontAsk' | 'auto';
6065
6066 /**
6067 * A `classic.PermissionRequest` answer's `decision`, as the classic hook's
6068 * `hookSpecificOutput.decision`: allow (with a rewrite or rules) or deny.
6069 */
6070 export type PermissionRequestDecision = {
6071 behavior: 'allow';
6072 updatedInput?: Record<string, unknown>;
6073 updatedPermissions?: PermissionUpdates;
6074 } | {
6075 behavior: 'deny';
6076 message?: string;
6077 interrupt?: true;
6078 };
6079
6080 type PermissionRequestHookInput = BaseHookInput & {
6081 hook_event_name: 'PermissionRequest';
6082 tool_name: string;
6083 tool_input: unknown;
6084 permission_suggestions?: PermissionUpdate[];
6085 mcp_server?: McpServerProvenance;
6086 };
6087
6088 type PermissionRuleValue = {
6089 toolName: string;
6090 ruleContent?: string;
6091 };
6092
6093 type PermissionUpdate = {
6094 type: 'addRules';
6095 rules: PermissionRuleValue[];
6096 behavior: PermissionBehavior;
6097 destination: PermissionUpdateDestination;
6098 } | {
6099 type: 'replaceRules';
6100 rules: PermissionRuleValue[];
6101 behavior: PermissionBehavior;
6102 destination: PermissionUpdateDestination;
6103 } | {
6104 type: 'removeRules';
6105 rules: PermissionRuleValue[];
6106 behavior: PermissionBehavior;
6107 destination: PermissionUpdateDestination;
6108 } | {
6109 type: 'setMode';
6110 mode: PermissionMode;
6111 destination: PermissionUpdateDestination;
6112 } | {
6113 type: 'addDirectories';
6114 directories: string[];
6115 destination: PermissionUpdateDestination;
6116 } | {
6117 type: 'removeDirectories';
6118 directories: string[];
6119 destination: PermissionUpdateDestination;
6120 };
6121
6122 type PermissionUpdateDestination = 'userSettings' | 'projectSettings' | 'localSettings' | 'session' | 'cliArg';
6123
6124 /**
6125 * The permission rules a PermissionRequest allow may add: the shape of the
6126 * request's own `permission_suggestions` (the SDK's PermissionUpdate list).
6127 */
6128 type PermissionUpdates = NonNullable<ClassicHookInputs['PermissionRequest']['permission_suggestions']>;
6129
6130 /**
6131 * How `$.audio.play` plays a clip: looped until `signal` aborts, or once.
6132 *
6133 * A loop needs the signal that ends it; a single play takes one as an option.
6134 */
6135 export type PlayOptions = {
6136 /**
6137 * Repeat the clip until `signal` aborts (the promise then resolves).
6138 */
6139 shouldLoop: true;
6140 /**
6141 * Linear gain, from 0 to 4; default 1.
6142 */
6143 gain?: number;
6144 /**
6145 * Stops the clip: playback ends at once and the promise resolves.
6146 */
6147 signal: AbortSignal;
6148 } | {
6149 /**
6150 * Play once.
6151 */
6152 shouldLoop?: false;
6153 /**
6154 * Linear gain, from 0 to 4; default 1.
6155 */
6156 gain?: number;
6157 /**
6158 * Stops the clip early, as above.
6159 */
6160 signal?: AbortSignal;
6161 };
6162
6163 /**
6164 * The nouns a plugin declared on `$` by merging into EngineInterface; never
6165 * one the engine's own events are under (`tool`, `session`, `ui`).
6166 */
6167 type PluginNoun = Exclude<keyof EngineInterface & string, keyof CoreEngineInterface | Namespace<CoreEventName>>;
6168
6169 /**
6170 * A plugin's options as `register(on, options)` receives them: the values of
6171 * the fields its manifest's `userConfig` declares, defaults filled in.
6172 *
6173 * Stored in settings.json `pluginConfigs[<plugin>].options` (sensitive ones
6174 * in secure storage), validated against the declared `type` before the module
6175 * loads; a required field with no value fails the load, naming the field. A
6176 * string field that declares `options` holds one of them: `/config` draws it
6177 * as a picker over them, and a stored value outside them counts as unset, so
6178 * its default applies. A `--plugin-dir` plugin's key is its plugin.json
6179 * `<name>` (or `<name>@inline`).
6180 */
6181 export type PluginOptions = Readonly<Record<string, string | number | boolean | readonly string[]>>;
6182
6183 /**
6184 * The input of `plugin.register`: one hooks module the engine is about to
6185 * load, read off its manifest and its scanned source. Every field is pinned.
6186 *
6187 * `tier` and `uses` are the host's reading of the module, what a gate holds
6188 * it to; `name`, `root`, `version` and `provenance` are its own word (its
6189 * manifest, its place), so a rule keyed on them is one a rename walks past.
6190 */
6191 export type PluginRegisterInput = {
6192 /**
6193 * The plugin's name, as its own plugin.json declares it.
6194 */
6195 name: string;
6196 /**
6197 * Where its hooks would stand among the chain's tiers (Tier); never
6198 * `core`, the engine's own.
6199 */
6200 tier: Exclude<Tier, 'core'>;
6201 /**
6202 * The plugin's directory (the one holding plugin.json), absolute.
6203 */
6204 root: string;
6205 /**
6206 * Its manifest's `version`, when it states one.
6207 */
6208 version?: string;
6209 /**
6210 * Where the plugin came from, as the loader keys it: `<name>@<marketplace>`
6211 * installed, `<name>@inline` by `--plugin-dir`, `<name>@builtin` bundled.
6212 */
6213 provenance: string;
6214 /**
6215 * What the module hooks and calls, as scanned (PluginRegisterUses).
6216 */
6217 uses: PluginRegisterUses;
6218 };
6219
6220 /**
6221 * What a `plugin.register` hook returns and what `next(e)` resolves to:
6222 * `{ allow: true }` from core, or `{ refuse: reason }`.
6223 */
6224 export type PluginRegisterResult = {
6225 /**
6226 * The module loads: its `engine.create` step runs and its hooks join.
6227 */
6228 allow: true;
6229 refuse?: undefined;
6230 } | {
6231 /**
6232 * The module does not load: no step, no hooks, no tools or commands;
6233 * the transcript names the plugin that refused and this reason.
6234 */
6235 refuse: string;
6236 allow?: undefined;
6237 };
6238
6239 /**
6240 * What a hooks module uses, as the host scanned its source before loading it:
6241 * the same lists `claude plugin validate` prints and the host's rule reads.
6242 *
6243 * Exact, since a module that spells `on`, `$` or `$.env` other than literally
6244 * does not load.
6245 */
6246 export type PluginRegisterUses = {
6247 /**
6248 * The patterns its `on(...)` registrations name, as written (`tool.call`,
6249 * `*`, `classic.*`, `!tool.describe`), in registration order, each once.
6250 */
6251 events: readonly string[];
6252 /**
6253 * What it calls on `$`, spelled `noun.method` (`fs.write`, `http.fetch`),
6254 * sorted, each once.
6255 */
6256 calls: readonly string[];
6257 /**
6258 * The environment variables its `$.env.get` and `$.env.set` calls name;
6259 * absent when it calls neither.
6260 */
6261 env?: {
6262 /**
6263 * The names its `$.env.get` calls spell, sorted, each once.
6264 */
6265 reads: readonly string[];
6266 /**
6267 * The names its `$.env.set` calls spell, sorted, each once.
6268 */
6269 writes: readonly string[];
6270 };
6271 };
6272
6273 /**
6274 * Whose element: the plugin whose hook drew it, stamped by the runtime as the
6275 * tree leaves it; a Box's or Text's `group`, a Client's or Raster's own.
6276 *
6277 * A Button's `press` names its plugin the same way, beside its handle. Two
6278 * plugins under one `hover.scope` string never share a group.
6279 */
6280 type PluginStamp = {
6281 plugin: string;
6282 };
6283
6284 type PostCompactHookInput = BaseHookInput & {
6285 hook_event_name: 'PostCompact';
6286 trigger: 'manual' | 'auto';
6287 /**
6288 * The conversation summary produced by compaction
6289 */
6290 compact_summary: string;
6291 };
6292
6293 type PostModelSwitchHookInput = (BaseHookInput & {
6294 hook_event_name: 'PostModelSwitch';
6295 }) & {
6296 /**
6297 * Resolved model id the session was running before the switch
6298 */
6299 from_model: string;
6300 /**
6301 * Resolved model id the session runs after the switch
6302 */
6303 to_model: string;
6304 /**
6305 * What was asked for (alias such as "opus", a full id, or null for "default")
6306 */
6307 requested_model: string | null;
6308 /**
6309 * command: /model <name>, the /config Model row, or enabling fast mode when that promotes the model; picker: an interactive model picker; sdk: headless set_model (SDK, Remote Control, IDE); auto: automatic fallback or other programmatic change; resume: model restored while resuming a session
6310 */
6311 source: 'command' | 'picker' | 'sdk' | 'auto' | 'resume';
6312 /**
6313 * Prompt tokens the next request re-sends: the last main-thread response's input + cache_read + cache_creation + output tokens (0 before the first response; for a server-side tool loop, its last iteration's window, not the summed totals)
6314 */
6315 context_tokens: number;
6316 /**
6317 * Whether the current model's prompt cache is likely still warm (a switch then forfeits it)
6318 */
6319 prompt_cache_warm: boolean;
6320 cache_ttl: '5m' | '1h';
6321 /**
6322 * Estimated cost of re-caching context_tokens on to_model at its cache-write rate - the managed modelPricing when set, otherwise list price; excludes the response
6323 */
6324 estimated_cache_write_usd: number;
6325 /**
6326 * configured: priced at the managed modelPricing setting; catalog: list price; default: to_model unknown, the default tier was assumed
6327 */
6328 pricing: 'configured' | 'catalog' | 'default';
6329 };
6330
6331 /**
6332 * Hook input for the PostToolBatch event. Fired once after every tool call in a batch has resolved, before the next model request. PostToolUse fires per-tool and may run concurrently for parallel tool calls; PostToolBatch fires exactly once with the full batch.
6333 */
6334 type PostToolBatchHookInput = BaseHookInput & {
6335 hook_event_name: 'PostToolBatch';
6336 tool_calls: PostToolBatchToolCall[];
6337 };
6338
6339 type PostToolBatchToolCall = {
6340 tool_name: string;
6341 tool_input: unknown;
6342 tool_use_id: string;
6343 tool_response?: unknown;
6344 };
6345
6346 type PostToolUseFailureHookInput = BaseHookInput & {
6347 hook_event_name: 'PostToolUseFailure';
6348 tool_name: string;
6349 tool_input: unknown;
6350 tool_use_id: string;
6351 error: string;
6352 is_interrupt?: boolean;
6353 /**
6354 * Tool execution time in milliseconds. Excludes permission-prompt and hook time.
6355 */
6356 duration_ms?: number;
6357 mcp_server?: McpServerProvenance;
6358 };
6359
6360 type PostToolUseHookInput = BaseHookInput & {
6361 hook_event_name: 'PostToolUse';
6362 tool_name: string;
6363 tool_input: unknown;
6364 tool_response: unknown;
6365 tool_use_id: string;
6366 /**
6367 * Tool execution time in milliseconds. Excludes permission-prompt and hook time.
6368 */
6369 duration_ms?: number;
6370 mcp_server?: McpServerProvenance;
6371 };
6372
6373 type PreCompactHookInput = BaseHookInput & {
6374 hook_event_name: 'PreCompact';
6375 trigger: 'manual' | 'auto';
6376 custom_instructions: string | null;
6377 };
6378
6379 type PreModelSwitchHookInput = (BaseHookInput & {
6380 hook_event_name: 'PreModelSwitch';
6381 }) & {
6382 /**
6383 * Resolved model id the session was running before the switch
6384 */
6385 from_model: string;
6386 /**
6387 * Resolved model id the session runs after the switch
6388 */
6389 to_model: string;
6390 /**
6391 * What was asked for (alias such as "opus", a full id, or null for "default")
6392 */
6393 requested_model: string | null;
6394 /**
6395 * command: /model <name>, the /config Model row, or enabling fast mode when that promotes the model; picker: an interactive model picker; sdk: headless set_model (SDK, Remote Control, IDE)
6396 */
6397 source: 'command' | 'picker' | 'sdk';
6398 /**
6399 * Prompt tokens the next request re-sends: the last main-thread response's input + cache_read + cache_creation + output tokens (0 before the first response; for a server-side tool loop, its last iteration's window, not the summed totals)
6400 */
6401 context_tokens: number;
6402 /**
6403 * Whether the current model's prompt cache is likely still warm (a switch then forfeits it)
6404 */
6405 prompt_cache_warm: boolean;
6406 cache_ttl: '5m' | '1h';
6407 /**
6408 * Estimated cost of re-caching context_tokens on to_model at its cache-write rate - the managed modelPricing when set, otherwise list price; excludes the response
6409 */
6410 estimated_cache_write_usd: number;
6411 /**
6412 * configured: priced at the managed modelPricing setting; catalog: list price; default: to_model unknown, the default tier was assumed
6413 */
6414 pricing: 'configured' | 'catalog' | 'default';
6415 };
6416
6417 /**
6418 * The link a press landed on: one a `Markdown` drew, pressed where the
6419 * surface reports presses (a plain click in the fullscreen terminal).
6420 *
6421 * What the surface knows of the cell pressed, not which occurrence: two
6422 * links written with one target read the same here.
6423 */
6424 export type PressedLink = {
6425 /**
6426 * The link's target as the surface drew it: for an `https:` link, the
6427 * href as the markdown wrote it.
6428 */
6429 href: string;
6430 };
6431
6432 /**
6433 * The decision of a `classic.PreToolUse` result: `allow`, `ask`, `deny`, or
6434 * none.
6435 */
6436 export type PreToolUseDecision = {
6437 /**
6438 * Lets the call run without a permission prompt (the managed-settings
6439 * hooks ran first; a deny from them ended the chain above).
6440 */
6441 allow: true;
6442 ask?: undefined;
6443 deny?: undefined;
6444 } | {
6445 /**
6446 * Asks the user before the call runs; the text is shown as the reason.
6447 */
6448 ask: string;
6449 allow?: undefined;
6450 deny?: undefined;
6451 } | {
6452 /**
6453 * Refuses the call; the model receives the text as the reason.
6454 */
6455 deny: string;
6456 allow?: undefined;
6457 ask?: undefined;
6458 } | {
6459 allow?: undefined;
6460 ask?: undefined;
6461 deny?: undefined;
6462 };
6463
6464 type PreToolUseHookInput = BaseHookInput & {
6465 hook_event_name: 'PreToolUse';
6466 tool_name: string;
6467 tool_input: unknown;
6468 tool_use_id: string;
6469 mcp_server?: McpServerProvenance;
6470 };
6471
6472 /**
6473 * What a `classic.PreToolUse` hook returns: one of `allow`, `ask`, `deny`,
6474 * or none of them, which passes the call on to the normal permission flow.
6475 */
6476 export type PreToolUseResult = PreToolUseDecision & {
6477 /**
6478 * Replaces the tool's arguments; validated against the tool's schema before
6479 * the tool runs.
6480 */
6481 updatedInput?: Record<string, unknown>;
6482 /**
6483 * Extra context handed to the model with the call, one entry per note.
6484 */
6485 additionalContext?: string[];
6486 };
6487
6488 /**
6489 * Options of `$.process.run`.
6490 */
6491 export type ProcessRunInit = {
6492 /**
6493 * The child's working directory, relative to the session's or absolute;
6494 * absent, the session's working directory.
6495 */
6496 cwd?: string;
6497 /**
6498 * Variables set over the host process's own environment.
6499 */
6500 env?: Record<string, string>;
6501 /**
6502 * Text written to the child's standard input, then closed.
6503 */
6504 stdin?: string;
6505 /**
6506 * How long the child may run before it is killed and the call rejects,
6507 * in milliseconds; 30 seconds when absent, ten minutes at most.
6508 */
6509 timeoutMs?: number;
6510 };
6511
6512 /**
6513 * What `$.process.run` resolves with once the child has exited.
6514 */
6515 export type ProcessRunResult = {
6516 /**
6517 * The child's exit status; a child ended by a signal reads as 1.
6518 */
6519 exitCode: number;
6520 /**
6521 * What the child wrote to standard output, as text, cut at the output
6522 * limit.
6523 */
6524 stdout: string;
6525 /**
6526 * What the child wrote to standard error, as text, cut at the output
6527 * limit.
6528 */
6529 stderr: string;
6530 };
6531
6532 /**
6533 * The input of `prompt.attachment`: one message the engine injects into the
6534 * conversation for the model on its own, as a request is about to carry it.
6535 *
6536 * A reminder, a mode transition, a listing, a mentioned file, a hook's
6537 * context: the person never typed it and mostly never sees it. Only an
6538 * attachment that carries text for the model is raised.
6539 */
6540 export type PromptAttachmentInput = {
6541 /**
6542 * As the engine names the attachment's kind; the key a matcher narrows on.
6543 * Pinned. Builds add and retire kinds: match by name.
6544 *
6545 * Among them `todo_reminder`, `plan_mode`, `plan_mode_exit`, `auto_mode`,
6546 * `auto_mode_exit`, `instructions`, `nested_memory`, `skill_listing`,
6547 * `deferred_tools_delta`, `edited_text_file`, `file`, `queued_command`.
6548 */
6549 type: string;
6550 /**
6551 * What the model reads for this attachment, inside the engine's framing;
6552 * rewritable with `next({ ...e, text })`.
6553 *
6554 * The `<system-reminder>` wrapper (or the system channel that replaces it)
6555 * goes around what the chain answers, never inside it. An attachment
6556 * rendered as several text blocks hands them joined by newlines.
6557 */
6558 text: string;
6559 /**
6560 * Who authored the text (PromptAttachmentOrigin): the engine, a settings
6561 * hook, or a plugin's chain context.
6562 *
6563 * Pinned: a different value is refused, one left out is kept.
6564 */
6565 origin: PromptAttachmentOrigin;
6566 /**
6567 * The loop whose request carries the attachment: a subagent's id, the `id`
6568 * `$.agent.list()` gives it and its `tool.call`s carry; absent on main.
6569 *
6570 * Pinned: a different value is refused, one left out is kept. A subagent
6571 * a hook spawned through `$.agent.spawn` is resolved past that hook.
6572 */
6573 agentId?: string;
6574 };
6575
6576 /**
6577 * Who authored the text an injected attachment carries, as the engine knows
6578 * it from the attachment itself; a closed set, pinned on the event.
6579 *
6580 * A hooks module reads `e.origin.kind` to tell the engine's own prose from a
6581 * settings hook's output or another plugin's context. `next(e)` passes it on
6582 * as received; one left out is put back; no hook sets one.
6583 */
6584 export type PromptAttachmentOrigin = {
6585 /**
6586 * The engine's own prose or framing: a reminder, a mode transition, a
6587 * listing, a notice, an announced context block.
6588 *
6589 * A file the person mentioned, as the engine presents it, and a prompt
6590 * or notification delivered into a running turn are the engine's too.
6591 */
6592 kind: 'engine';
6593 } | {
6594 /**
6595 * A settings hook's output the engine injects for the model: its
6596 * additional context, a blocking error's note, a stopped continuation.
6597 */
6598 kind: 'hook';
6599 /**
6600 * The settings hook event that produced it (`SessionStart`,
6601 * `UserPromptSubmit`, `PostToolUse`, ...).
6602 */
6603 event: string;
6604 } | {
6605 /**
6606 * Text a plugin's hook attached through a chain's `context`
6607 * (`prompt.submit`, `tool.call`), as the model reads it.
6608 */
6609 kind: 'plugin';
6610 /**
6611 * The chain's event that attached it.
6612 */
6613 event: string;
6614 };
6615
6616 /**
6617 * What a `prompt.attachment` hook returns: the text the model reads for that
6618 * attachment, or null to leave the attachment out of the request.
6619 */
6620 export type PromptAttachmentResult = {
6621 text: string | null;
6622 };
6623
6624 /**
6625 * The person's prompt box as it stands: the draft and where the cursor is in
6626 * it; what `$.prompt.read()` resolves and `$.prompt.fill` hands back.
6627 *
6628 * No selection: the terminal's box has none of its own, and a surface that
6629 * binds one adds it here.
6630 */
6631 export type PromptBox = {
6632 /**
6633 * The draft as typed so far; `''` where the session draws no box.
6634 */
6635 text: string;
6636 /**
6637 * Where the next typed character lands: an offset into `text` in UTF-16
6638 * code units, 0 at the start, `text.length` at the end.
6639 */
6640 cursor: number;
6641 };
6642
6643 /**
6644 * The input of `prompt.compose`: the facts a system prompt is composed from,
6645 * each already resolved by the engine, at the moment it renders one.
6646 */
6647 export type PromptComposeInput = {
6648 /**
6649 * The id of the model the request is for; pinned, the field a matcher
6650 * narrows on.
6651 */
6652 model: string;
6653 /**
6654 * The model whose prompt is rendered: `model`, unless the engine renders
6655 * another model's prompt for it (a model it holds no prompt of its own for).
6656 */
6657 promptModel: string;
6658 /**
6659 * Where the session draws at this render, as `$.session.surfaces()`
6660 * answers: `terminal` first under the REPL; empty where nothing draws.
6661 */
6662 surfaces: readonly RenderSurface[];
6663 /**
6664 * The names of the tools the request offers the model.
6665 */
6666 tools: readonly string[];
6667 /**
6668 * What the person chose in place of the default way of answering, and
6669 * whether it keeps the coding instructions; null for the default style.
6670 */
6671 outputStyle: {
6672 name: string;
6673 isKeepingCodingInstructions: boolean;
6674 } | null;
6675 traits: readonly PromptComposeTrait[];
6676 };
6677
6678 /**
6679 * What a `prompt.compose` hook returns: the sections of the system prompt,
6680 * in order, every `shared` one ahead of every `session` one.
6681 *
6682 * A section left out is not sent. The engine joins each side, places the
6683 * cache boundary between them and every cache marker itself.
6684 */
6685 export type PromptComposeResult = {
6686 sections: readonly PromptComposeSection[];
6687 };
6688
6689 /**
6690 * Which side of the prompt cache's boundary a section of the system prompt
6691 * sits on: `shared` before it, `session` after it.
6692 *
6693 * `shared` is text that reads the same for every person on this build and
6694 * model: it is sent in the block the API may cache across organizations.
6695 * `session` is text that varies with the person, the machine or the session.
6696 *
6697 * The engine places the one boundary and every cache marker itself,
6698 * whatever a list says; `shared` text that varies hits that cache for nobody.
6699 */
6700 export type PromptComposeScope = 'shared' | 'session';
6701
6702 /**
6703 * One section of the system prompt as `prompt.compose` answers it: a stable
6704 * id, the text the model reads, and the side of the cache boundary it is on.
6705 *
6706 * @example
6707 * const POLICY = { id: "acme:policy", text: "# Policy\n...", scope: "session" }
6708 */
6709 export type PromptComposeSection = {
6710 /**
6711 * What a hook above finds the section by, to replace, move or drop it;
6712 * never empty, and unique in one list.
6713 *
6714 * A section a plugin adds is named `<plugin>:<name>`; the bare names are
6715 * the ones the plugin that defines the default prompt gives its own.
6716 */
6717 id: string;
6718 /**
6719 * The section's text, sent as written; sections on one side of the
6720 * boundary are joined by a blank line, in the list's order.
6721 */
6722 text: string;
6723 scope: PromptComposeScope;
6724 };
6725
6726 /**
6727 * One branch the engine's own composition of the system prompt takes on the
6728 * request or the session before it computes any section: a closed set.
6729 *
6730 * `bare`: the session runs with the one-line prompt (`--bare`). `lean`: the
6731 * prompt model takes the short body. `sdk-preset`: the SDK's `claude_code`
6732 * preset, whose per-person sections ride the first user message instead.
6733 *
6734 * `teammate`: an in-process teammate's render of its lead's prompt.
6735 * `analysis`: a render that measures the prompt (`/context`) and sends
6736 * nothing. `print`: a session with no terminal behind it (`-p`, the SDK).
6737 *
6738 * `skills`: the Skill tool has commands to list. `send-user-message`: the
6739 * session speaks to the person through a message tool.
6740 *
6741 * What one section's own text turns on (a flag, a setting, a model family)
6742 * is not here: it becomes a member when that text is a plugin's to write.
6743 */
6744 export type PromptComposeTrait = 'bare' | 'lean' | 'sdk-preset' | 'teammate' | 'analysis' | 'print' | 'skills' | 'send-user-message';
6745
6746 /**
6747 * One block of the context the first user message carries: a name the
6748 * engine keys it by and the text under it.
6749 */
6750 export type PromptContextBlock = {
6751 /**
6752 * The key the block renders under (`# name`): `claudeMd`, `userEmail`,
6753 * `attachedProject`, `currentDate`, or a plugin's own.
6754 *
6755 * The field a matcher narrows on; unique among one context's blocks.
6756 */
6757 name: string;
6758 /**
6759 * The block's text; `claudeMd`'s is the instruction files framed as the
6760 * engine frames them, empty when it announces none.
6761 */
6762 text: string;
6763 };
6764
6765 /**
6766 * The context blocks of a conversation's first user message, in the order
6767 * the engine renders them: what `prompt.context` takes and answers alike.
6768 */
6769 export type PromptContextBlocks = {
6770 /**
6771 * From core: `claudeMd` (when instruction files are loaded), `userEmail`,
6772 * `attachedProject`, `currentDate`, each only when present.
6773 */
6774 blocks: readonly PromptContextBlock[];
6775 };
6776
6777 /**
6778 * The input of `prompt.context`: the context blocks the engine prepends to a
6779 * conversation's first user message, and the files behind `claudeMd`.
6780 */
6781 export type PromptContextInput = {
6782 /**
6783 * From core: `claudeMd` (when instruction files are loaded), `userEmail`,
6784 * `attachedProject`, `currentDate`, each only when present.
6785 */
6786 blocks: readonly PromptContextBlock[];
6787 /**
6788 * The files behind `claudeMd`, in the order it renders them, `@` imports
6789 * included; empty when it renders none.
6790 *
6791 * Undefined when a hook above rewrote the `claudeMd` text: the files
6792 * behind that text are then unknown, and a hook adds none of its own.
6793 */
6794 instructionFiles?: readonly InstructionFile[];
6795 };
6796
6797 /**
6798 * What a `prompt.context` hook returns: the blocks the conversation
6799 * carries, in order; one left out is not sent.
6800 */
6801 export type PromptContextResult = {
6802 /**
6803 * What the conversation carries, in order.
6804 */
6805 blocks: readonly PromptContextBlock[];
6806 /**
6807 * The files now behind `claudeMd`; left out, the ones from below stand.
6808 *
6809 * Kept in step with the `claudeMd` text at every link: a changed list
6810 * renders the text the next reader gets, a rewritten text makes the files
6811 * unknown from there on. Every kind is the hook's to add, drop or rewrite.
6812 */
6813 instructionFiles?: readonly InstructionFile[];
6814 };
6815
6816 /**
6817 * The input of `prompt.edit` (prompt-edit/): one edit the person makes in the
6818 * prompt box, as the draft before it and the splice the editor made of it.
6819 */
6820 export type PromptEditInput = {
6821 /**
6822 * Who edits (PromptEditOrigin): the person at the composer. Pinned:
6823 * `next(e)` passes it on as received.
6824 */
6825 origin: PromptEditOrigin;
6826 /**
6827 * The one key that made the edit, in `Client` `onKey`'s shape, when one
6828 * did; absent for a paste and for a burst of keys folded into one edit.
6829 *
6830 * Pinned: which key the person pressed is a fact, not the hook's to change.
6831 */
6832 key?: ClientKeyEvent;
6833 /**
6834 * The draft before the edit. `next({ ...e, text })` applies the edit to
6835 * another draft instead.
6836 */
6837 text: string;
6838 /**
6839 * Where the person's caret stood in `text` before the edit, 0 to
6840 * `text.length`.
6841 */
6842 cursor: number;
6843 /**
6844 * Where in `text` the edit begins: the start of the span it replaces, or
6845 * where what was typed goes in; a bare cursor move begins where it lands.
6846 */
6847 start: number;
6848 /**
6849 * Where the replaced span of `text` ends: `start` for an insertion or a
6850 * move, past it for a deletion (Backspace, a kill).
6851 */
6852 end: number;
6853 /**
6854 * What goes in between `start` and `end`: the typed or pasted text, `''` for
6855 * a deletion or a move.
6856 *
6857 * `next({ ...e, inputText })` puts in another; the cursor lands after it.
6858 */
6859 inputText: string;
6860 };
6861
6862 /**
6863 * Who edits the prompt box at `prompt.edit`, as the engine stamps it where
6864 * the edit starts; a closed set a matcher narrows on.
6865 *
6866 * `next(e)` passes it on as received; no hook sets one.
6867 */
6868 export type PromptEditOrigin = {
6869 /**
6870 * The person, typing or pasting into the main prompt box.
6871 */
6872 kind: 'composer';
6873 };
6874
6875 /**
6876 * What a `prompt.edit` hook returns and what `next(e)` resolves to: the box
6877 * after the edit (PromptBox), which the editor then shows.
6878 *
6879 * From core, `e.text` with the splice applied and the cursor after what went
6880 * in. Rewrite it (`{ ...r, text, cursor }`) to change what lands; answer
6881 * `{ text: e.text, cursor: e.cursor }` without `next` to consume the key.
6882 */
6883 export type PromptEditResult = PromptBox;
6884
6885 /**
6886 * `prompt.fill`'s input as a plugin's `$.prompt.fill(args)` takes it: no
6887 * `origin` (the engine sets the calling plugin's), `mode` optional.
6888 */
6889 export type PromptFillArgs = {
6890 /**
6891 * What the box is to take (PromptFillInput `text`).
6892 */
6893 text: string;
6894 /**
6895 * Where it goes (PromptFillMode); `replace` when left out.
6896 */
6897 mode?: PromptFillMode;
6898 };
6899
6900 /**
6901 * What `$.prompt.fill` resolves to: whether the box took the text, and the
6902 * box afterwards (PromptBox) as the caller's own `$.prompt.read()` reads it.
6903 *
6904 * The box goes to a plugin whose module calls `$.prompt.read`, through the
6905 * hooks on it; one that never does, or is refused there, gets the empty box.
6906 */
6907 export type PromptFilled = {
6908 /**
6909 * True once the box holds the text; false where no box could take it or a
6910 * hook kept it out (PromptFillResult).
6911 */
6912 isFilled: boolean;
6913 /**
6914 * The draft after the fill; unchanged when `isFilled` is false; `''` where
6915 * no box is drawn or the caller may not read it (see above).
6916 */
6917 text: string;
6918 /**
6919 * Where the person types next: an offset into `text`, past the fill's text
6920 * for `insert`, at the end for `replace` and `append`.
6921 */
6922 cursor: number;
6923 };
6924
6925 /**
6926 * The input of `prompt.fill` (prompt-fill/): a text about to be put in the
6927 * prompt box as the person's draft, over it, after it, or at the cursor.
6928 */
6929 export type PromptFillInput = {
6930 /**
6931 * What the box takes; the person edits it or presses Enter.
6932 * `next({ ...e, text })` writes another.
6933 */
6934 text: string;
6935 /**
6936 * Where the text goes (PromptFillMode): over the draft, after it, or in at
6937 * the cursor. `next({ ...e, mode })` moves it; left out of a rewrite, kept.
6938 */
6939 mode: PromptFillMode;
6940 /**
6941 * Who writes (PromptFillOrigin), set by the engine where the write
6942 * starts. Pinned: `next(e)` passes it on as received.
6943 */
6944 origin: PromptFillOrigin;
6945 };
6946
6947 /**
6948 * Where a `prompt.fill` puts its text: over the whole draft, after it, or
6949 * into it at the cursor.
6950 *
6951 * `replace` empties the box first and leaves the cursor at the text's end;
6952 * `append` keeps the draft and adds the text after it, cursor at the end;
6953 * `insert` splices the text in at the cursor and moves the cursor past it,
6954 * so what the person had typed stays on either side.
6955 */
6956 export type PromptFillMode = 'replace' | 'append' | 'insert';
6957
6958 /**
6959 * Who writes the prompt box at `prompt.fill`, as the engine stamps it where
6960 * the write starts; a closed set a matcher narrows on.
6961 *
6962 * `next(e)` passes it on as received; no hook sets one.
6963 */
6964 export type PromptFillOrigin = {
6965 /**
6966 * The engine writing the box on its own account; reserved for its own
6967 * sites, none of which raises `prompt.fill` as shipped.
6968 */
6969 kind: 'engine';
6970 } | {
6971 /**
6972 * A plugin's `$.prompt.fill`.
6973 */
6974 kind: 'plugin';
6975 /**
6976 * The filling plugin's name.
6977 */
6978 name: string;
6979 };
6980
6981 /**
6982 * What a `prompt.fill` hook returns and what `next(e)` resolves to: whether
6983 * the text went into the prompt box.
6984 */
6985 export type PromptFillResult = {
6986 /**
6987 * True once the box holds the text; false where no box can take it (a
6988 * dialog holds the keys, a headless session has none).
6989 *
6990 * A hook answering `{ isFilled: false }` without `next` keeps the text
6991 * out.
6992 */
6993 isFilled: boolean;
6994 };
6995
6996 /**
6997 * Where a `prompt.submit` submission came from, as the engine knows it at
6998 * the site it was queued from; a closed set, never a text prefix.
6999 *
7000 * A hooks module reads `e.origin.kind` to tell the user's own Enter from a
7001 * notification, a peer session, a schedule or another plugin. `next(e)`
7002 * passes it on as received; an answer may leave it out; no hook sets one.
7003 */
7004 export type PromptOrigin = {
7005 /**
7006 * The user's own gesture at the terminal, as the engine stamped it
7007 * (never presumed from an unstamped command).
7008 *
7009 * Enter at the prompt, typed or queued, or a click on a transcript
7010 * link; a channel the engine cannot attest (a same-user socket) is
7011 * never stamped, and arrives as `unclassified`.
7012 */
7013 kind: 'composer';
7014 } | {
7015 /**
7016 * The user's message through the Remote Control bridge (a phone or
7017 * web client).
7018 */
7019 kind: 'bridge';
7020 } | {
7021 /**
7022 * The SDK host's own turn (`claude -p`, the Agent SDK), not typed at
7023 * a terminal.
7024 */
7025 kind: 'sdk';
7026 } | {
7027 /**
7028 * A background task's notification, dequeued when the session went
7029 * idle or delivered into a running turn (`turnId` set).
7030 */
7031 kind: 'task-notification';
7032 } | {
7033 /**
7034 * A scheduled task, routine or /loop firing its stored prompt.
7035 */
7036 kind: 'scheduled-trigger';
7037 } | {
7038 /**
7039 * Another Claude session's message ("Another Claude session sent a
7040 * message"), as a turn of its own or delivered into a running one.
7041 */
7042 kind: 'peer';
7043 } | {
7044 /**
7045 * Another session's SendMessage delivery, model-authored and framed
7046 * as a notification.
7047 */
7048 kind: 'peer-send-message';
7049 } | {
7050 /**
7051 * A delivery a coordinating session composed for one of its threads.
7052 */
7053 kind: 'projects-relay';
7054 } | {
7055 /**
7056 * A message from a channel an MCP server relays (Slack, Telegram).
7057 */
7058 kind: 'channel';
7059 /**
7060 * The channel server's name.
7061 */
7062 server: string;
7063 } | {
7064 /**
7065 * A coordinating session's hand-off to a worker session.
7066 */
7067 kind: 'coordinator';
7068 } | {
7069 /**
7070 * A background observer agent's report to the agent it observes.
7071 */
7072 kind: 'observer';
7073 } | {
7074 /**
7075 * An activity digest delivered to an observer agent.
7076 */
7077 kind: 'observer-activity';
7078 } | {
7079 /**
7080 * A programmatic follow-up to a user's UI action, user-initiated but
7081 * not typed this turn.
7082 */
7083 kind: 'auto-continuation';
7084 } | {
7085 /**
7086 * A turn with no provenance the engine can name: one the ingress
7087 * could not classify, or a command queued with no stamp at all.
7088 *
7089 * An idle notice or a delivery receipt the engine queued isMeta with
7090 * no stamp is one too; the engine frames that shape as a non-user
7091 * source.
7092 */
7093 kind: 'unclassified';
7094 } | {
7095 /**
7096 * The session's owner pinging it from Slack.
7097 */
7098 kind: 'slack-ping';
7099 } | {
7100 /**
7101 * A plugin's `$.prompt.submit`; the model reads the prompt under the
7102 * plugin's name unless a hook leaves the origin out of its answer.
7103 */
7104 kind: 'plugin';
7105 /**
7106 * The submitting plugin's name.
7107 */
7108 name: string;
7109 };
7110
7111 /**
7112 * The input of `prompt.section`: one named section of the system prompt, at
7113 * the moment the engine assembles it.
7114 */
7115 export type PromptSectionInput = {
7116 /**
7117 * As the engine names the section (`env_info_simple`, `memory`, ...); the
7118 * key a matcher narrows on.
7119 */
7120 name: string;
7121 /**
7122 * The section's text as core computed it, or null when core omits it.
7123 */
7124 text: string | null;
7125 };
7126
7127 /**
7128 * What a `prompt.section` hook returns: the text the prompt carries for that
7129 * section, or null to leave it out.
7130 */
7131 export type PromptSectionResult = {
7132 text: string | null;
7133 };
7134
7135 /**
7136 * `prompt.submit`'s input as a plugin's call takes it: `origin`, `turnId`
7137 * and `wait` are the engine's to set, `context` the hooks' to attach.
7138 *
7139 * `origin` is the calling plugin's name; `turnId` is the turn a prompt typed
7140 * mid-turn ran over; `wait` is false, as a plugin's prompt runs once idle.
7141 */
7142 export type PromptSubmitArgs = Omit<PromptSubmitInput, 'origin' | 'turnId' | 'wait' | 'context'>;
7143
7144 /**
7145 * A pasted or attached non-text item of a submitted prompt; its kind, never
7146 * its bytes.
7147 */
7148 export type PromptSubmitAttachment = {
7149 /**
7150 * The item's kind.
7151 */
7152 type: 'image' | 'audio' | 'document';
7153 /**
7154 * The item's MIME type (`image/png`), when known.
7155 */
7156 mediaType?: string;
7157 /**
7158 * The pasted file's name, when it had one.
7159 */
7160 filename?: string;
7161 };
7162
7163 /**
7164 * The input of `prompt.submit`: the prompt as typed, after the input became
7165 * a user message and before it enters the session.
7166 */
7167 export type PromptSubmitInput = {
7168 /**
7169 * The prompt's text as it will reach the model (pastes already expanded).
7170 */
7171 text: string;
7172 /**
7173 * Present only when the submission carried images or other non-text items.
7174 */
7175 attachments?: readonly PromptSubmitAttachment[];
7176 /**
7177 * What the model reads beside the prompt and the user never sees, each
7178 * entry one block after the prompt as typed; absent as the engine raises it.
7179 *
7180 * A hook attaches on the way down: `next({ ...e, context: [...(e.context
7181 * ?? []), mine] })`, keeping which it likes; none empty, any length: past
7182 * 100,000 characters (200,000 together) the model reads a head and path.
7183 */
7184 context?: readonly string[];
7185 /**
7186 * The id of the model turn that was running when the prompt was submitted
7187 * (`turn.start`'s `turnId`): typed over that turn, or delivered into it.
7188 *
7189 * A queued delivery (a peer session's message) reaches the model inside a
7190 * running turn. Absent for a prompt submitted while the session was idle,
7191 * and for a plugin's own (`$.prompt.submit`), which runs once it is idle.
7192 */
7193 turnId?: string;
7194 /**
7195 * Whether the user asked the prompt to wait its turn (`chat:queueSubmit`,
7196 * `ctrl+x enter` by default): true for that submission, false otherwise.
7197 *
7198 * The engine queues every prompt typed mid-turn either way; the flag is
7199 * for hooks, so one that cancels the running turn on a plain Enter can
7200 * leave a waiting prompt alone. False for a prompt a plugin submitted.
7201 */
7202 wait: boolean;
7203 /**
7204 * Where the submission came from (PromptOrigin), set by the engine where
7205 * it was queued: the user's Enter, a notification, a peer, a plugin.
7206 *
7207 * `next(e)` passes it on as received; no hook may set one.
7208 */
7209 origin: PromptOrigin;
7210 };
7211
7212 /**
7213 * What a `prompt.submit` hook returns and what `next(e)` resolves to: the
7214 * prompt that entered, `{ text, context?, origin? }`, or `{ drop: reason }`.
7215 *
7216 * `next(e)` resolves once the prompt entered the session and its turn
7217 * started, or it was queued behind the running one; not when the turn ends,
7218 * which is `turn.complete`. A hook answering without `next` enters nothing.
7219 */
7220 export type PromptSubmitResult = {
7221 /**
7222 * The prompt that entered; from core, the text that arrived at the
7223 * bottom. A rewrite passes it down, `next({ ...e, text })`.
7224 */
7225 text: string;
7226 /**
7227 * What entered beside the prompt for the model, never shown the user:
7228 * from core, the context that arrived (`e.context`).
7229 *
7230 * Each entry is one block after the prompt as typed. A hook attaches
7231 * context on the way down; one put here after `next` resolved is not
7232 * attached (the prompt had entered), and is logged.
7233 */
7234 context?: readonly string[];
7235 /**
7236 * Where the prompt entered from: from core, `e.origin` as received;
7237 * absent, the prompt is the user's own.
7238 *
7239 * A hook may put back the origin it received; it may not set another.
7240 */
7241 origin?: PromptOrigin;
7242 drop?: undefined;
7243 } | {
7244 /**
7245 * The prompt did not enter: a hook's refusal, answered without `next`,
7246 * or a settings hook's block beneath.
7247 *
7248 * The text is shown to the user as the reason.
7249 */
7250 drop: string;
7251 text?: undefined;
7252 context?: undefined;
7253 origin?: undefined;
7254 };
7255
7256 /**
7257 * `prompt.suggest`'s input as a plugin's `$.prompt.suggest(args)` takes it:
7258 * `origin` is the engine's to set (the calling plugin's name).
7259 */
7260 export type PromptSuggestArgs = Omit<PromptSuggestInput, 'origin'>;
7261
7262 /**
7263 * The input of `prompt.suggest` (prompt-suggest/): a text about to be shown
7264 * dim in the empty prompt box, for Tab (or the right arrow) to take.
7265 */
7266 export type PromptSuggestInput = {
7267 /**
7268 * The proposed prompt: shown, not written; taking it puts it in the box
7269 * for editing. `next({ ...e, text })` proposes another.
7270 */
7271 text: string;
7272 /**
7273 * Who proposes (PromptSuggestOrigin), set by the engine where the
7274 * proposal starts. Pinned: `next(e)` passes it on as received.
7275 */
7276 origin: PromptSuggestOrigin;
7277 };
7278
7279 /**
7280 * Who proposes the text at `prompt.suggest`, as the engine stamps it where
7281 * the proposal starts; a closed set a matcher narrows on.
7282 *
7283 * `next(e)` passes it on as received; no hook sets one.
7284 */
7285 export type PromptSuggestOrigin = {
7286 /**
7287 * The engine's own guess at the person's next prompt, generated after
7288 * a turn (the prompt-suggestion service).
7289 */
7290 kind: 'suggestion';
7291 } | {
7292 /**
7293 * A plugin's `$.prompt.suggest`.
7294 */
7295 kind: 'plugin';
7296 /**
7297 * The proposing plugin's name.
7298 */
7299 name: string;
7300 };
7301
7302 /**
7303 * What a `prompt.suggest` hook returns and what `next(e)` resolves to:
7304 * whether the text is now the box's dim suggestion.
7305 */
7306 export type PromptSuggestResult = {
7307 /**
7308 * True once the box has the suggestion to show, at once or as soon as a
7309 * dialog gives the box back; false where it cannot show.
7310 *
7311 * It cannot while the box holds text, a turn runs, the text is blank, or
7312 * the session is headless. A hook answering `{ isShown: false }` without
7313 * `next` keeps it from showing.
7314 */
7315 isShown: boolean;
7316 };
7317
7318 /**
7319 * A `$.ui.blit` argument repainting one of the caller's mounted Rasters.
7320 *
7321 * `columns` and `rows`, when given, must be the mounted size (a resize is a
7322 * redraw, `$.ui.invalidate("ui.render")`, not a blit).
7323 */
7324 export type RasterBlitArgs = {
7325 /**
7326 * The site the Raster is drawn in, by the `requestId` this plugin draws
7327 * it under: one of its panes' ids, a tool row's `tool_use_id`, the band's.
7328 */
7329 requestId: string;
7330 /**
7331 * The Raster's `key` in that drawing.
7332 */
7333 key: string;
7334 /**
7335 * The new cells, encoded as the element's `cells` are (RasterProps), for
7336 * the mounted `columns * rows`.
7337 */
7338 cells: string;
7339 /**
7340 * The width the cells are laid out for; refused unless it is the mounted
7341 * Raster's. Absent, the mounted width.
7342 */
7343 columns?: number;
7344 /**
7345 * The height the cells are laid out for; refused unless it is the mounted
7346 * Raster's. Absent, the mounted height.
7347 */
7348 rows?: number;
7349 };
7350
7351 /**
7352 * The props of `Raster`, the terminal surface's cell-grid leaf: a fixed box
7353 * of cells, each a glyph, a foreground and a background, packed in `cells`.
7354 *
7355 * A leaf: no children, `hover` or `onPress` yet; repainted in place by
7356 * `$.ui.blit`. Terminal only for now (elsewhere a fragment); its palette
7357 * paints 1024 distinct color pairs at once and the rest as their nearest.
7358 */
7359 export type RasterProps = {
7360 /**
7361 * The element's address within the drawing: what `$.ui.blit` names to
7362 * repaint it, unique among the Rasters of one tree.
7363 */
7364 key: string;
7365 /**
7366 * How many terminal columns wide, 1 to 512; the site clips what its body
7367 * cannot show.
7368 */
7369 columns: number;
7370 /**
7371 * How many terminal rows tall, 1 to 256.
7372 */
7373 rows: number;
7374 /**
7375 * Every cell, row-major: standard padded base64 of `columns * rows`
7376 * little-endian u32 triplets `[codePoint, foreground, background]`.
7377 *
7378 * A code point is one printable width-1 BMP character (blocks, box drawing,
7379 * braille too), or the tree is refused naming the cell's index; a color is
7380 * `0x00RRGGBB`, or `0x01000000` (bit 24 alone) for the terminal's default.
7381 *
7382 * @example const words = Uint32Array.of(0x2588, 0xff8800, 0x01000000)
7383 * const cells = new Uint8Array(words.buffer).toBase64() // one orange cell
7384 */
7385 cells: string;
7386 };
7387
7388 /**
7389 * The hooks module's entry: `export function register(on, options)`. `on`
7390 * registers hooks; `options` is the plugin's configuration (PluginOptions).
7391 *
7392 * The options are fixed for this activation: a change to them reloads the
7393 * plugin and `register` runs again with the new object. Hooks close over it.
7394 * Its return is dropped, a promise awaited: `on => on(...)` is a module.
7395 *
7396 * @example
7397 * on("tool.call", ($, e, next) => e.tool === "Bash" ? { deny: "no" } : next(e))
7398 */
7399 export type Register = (on: On, options: PluginOptions) => unknown;
7400
7401 /**
7402 * What `on(...)` returns for a hook of type `F`: the registration, which
7403 * takes one `.catch` (CatchHandler); without it a failed hook is absent.
7404 *
7405 * A second `.catch` on one registration throws, as does one after
7406 * register() returned and one on `engine.create`, whose hook has no budget
7407 * and whose failure is the load's.
7408 */
7409 export type Registration<F> = {
7410 /**
7411 * Sets the handler run when the hook throws or overruns its budget; its
7412 * answer within the grace stands as the hook's result for the dispatch.
7413 *
7414 * The budget is HookBudget's `ms` and the grace its `catchMs`, both on
7415 * the clock that stops while the code waits on `next` or `$`.
7416 */
7417 readonly catch: (handler: CatchHandler<F>) => void;
7418 };
7419
7420 /**
7421 * What an element takes as `children`, as JSX passes them: a node, a number
7422 * (drawn as its string), a value the factory drops, or a list that may nest.
7423 *
7424 * `false`, `null` and `undefined` are dropped, so `{ok && <Text>hi</Text>}`
7425 * and `{n > 0 ? <Text>{n}</Text> : null}` type; a mapped list beside a
7426 * sibling nests. The element holds the flat, normalized list of RenderNode.
7427 */
7428 export type RenderChildren = RenderNode | number | boolean | null | undefined | readonly RenderChildren[];
7429
7430 /**
7431 * Everything `ui.render` can draw: one name per component that has a render
7432 * site; a matcher narrows on it.
7433 *
7434 * The permission dialog is drawn by the engine alone, since its answer
7435 * authorises an action; a plugin adds context with `$.ui.notice`. `Pane` is
7436 * the one component whose instances a plugin opens (`$.ui.open`).
7437 */
7438 export type RenderComponent = 'AskUserQuestion' | 'UserMessage' | 'AssistantMessage' | 'ToolUse' | 'ToolResult' | 'ToolGroup' | 'ToolProgress' | 'CommandOutput' | 'Spinner' | 'TurnDuration' | 'InfoNotice' | 'SessionMode' | 'PromptHint' | 'AbovePrompt' | 'Pane';
7439
7440 /**
7441 * What a render hook returns, and what `next(e)` resolves to: a plain-data
7442 * tree of elements (a Box or Text is a StyledElement), strings as children.
7443 *
7444 * Props are an allowlisted subset of Ink's Box/Text props (the ones
7445 * ElementProps declares); a tree with any other prop fails validation as a
7446 * whole and the engine's own component is drawn with the original props.
7447 */
7448 export type RenderElement = StyledElement<'Box', BoxHoverProps> | StyledElement<'Text', TextHoverProps> | {
7449 /**
7450 * A button, on every surface: `[ label ]` on the terminal, a native
7451 * button on a desktop; a press raises `ui.press` (`e.element` the key).
7452 *
7453 * Built by `<Button>` or the table's `t.Button`. The `onPress` closure
7454 * stays in the plugin's own environment under `press.handle`; the host
7455 * holds the handle for the lifetime of the drawing. A leaf: no children.
7456 */
7457 type: 'Button';
7458 props: {
7459 /**
7460 * The element's address: what `e.element` carries and what a matcher
7461 * names (`{ element: "explain" }`).
7462 */
7463 key: string;
7464 /**
7465 * The text drawn on the button.
7466 */
7467 label: string;
7468 /**
7469 * One digit (`"1"`) or one lowercase letter (`"w"`) that presses it
7470 * while the plugin's site holds the focus; anything else is refused.
7471 *
7472 * The band after ctrl+x tab or a click, a focused `Pane`: on keydown,
7473 * lowercased; never from the composer, save a bare digit in an empty
7474 * one pressing a band Button. Two on one hotkey: the later wins.
7475 */
7476 hotkey?: string;
7477 /**
7478 * An engine keybinding action (`"app:cycleDiffBase"`) whose chord
7479 * presses the Button from the prompt; an unknown name is refused.
7480 *
7481 * Chords, or a modified key Global or an active context binds, on
7482 * the terminal while mounted, no dialog up, no engine handler of it
7483 * mounted; a pane's over the band's over another's, the last drawn.
7484 */
7485 action?: string;
7486 /**
7487 * Drawn without chrome: the hotkey in the accent color, a colon,
7488 * then the label (`1: Yes`); without a `hotkey`, the label alone.
7489 *
7490 * The focus still inverts it. In JSX the label may be the one string
7491 * child (`<Button hotkey="1" plain onPress={...}>Yes</Button>`); the
7492 * key defaults to the label.
7493 */
7494 plain?: true;
7495 /**
7496 * The label dim at rest, as `Text`'s `dimColor`, and at full strength
7497 * under the pointer or the focus; absent draws as false.
7498 */
7499 dimColor?: TextProps['dimColor'];
7500 /**
7501 * The site's ring starts on this element when the site takes the
7502 * keyboard; the first drawn of several. Absent draws as before.
7503 */
7504 autoFocus?: true;
7505 };
7506 /**
7507 * Where the handler lives: the plugin whose hook drew the element, and
7508 * the handle its environment keeps the `onPress` closure under.
7509 *
7510 * The runtime stamps the plugin as the tree leaves that hook.
7511 */
7512 press: {
7513 plugin: string;
7514 handle: number;
7515 };
7516 /**
7517 * Label style overrides (the Text set) the surface applies while the
7518 * nearest keyed Box, or the group `scope` names, is hovered; plain data.
7519 */
7520 hover?: TextHoverProps;
7521 } | {
7522 /**
7523 * A one-line text field on every surface; a change and Enter raise
7524 * `ui.input` (`e.element` the key, `e.kind` which, `e.value` the text).
7525 *
7526 * Built by `<Input>` or the table's `t.Input`. The `onInput` and
7527 * `onSubmit` closures stay in the plugin's own environment under
7528 * `press.handle`, held as a Button's is. A leaf: no children.
7529 */
7530 type: 'Input';
7531 props: {
7532 /**
7533 * The element's address: what `e.element` carries and what a matcher
7534 * names (`{ element: "reply" }`).
7535 */
7536 key: string;
7537 /**
7538 * Text drawn before the field.
7539 */
7540 label?: string;
7541 /**
7542 * Text drawn dim in an empty field.
7543 */
7544 placeholder?: string;
7545 /**
7546 * The text the field holds when drawn.
7547 */
7548 value?: string;
7549 /**
7550 * What Enter does, drawn beside the field while it has focus.
7551 */
7552 submitLabel?: string;
7553 /**
7554 * The site's ring starts on this element when the site takes the
7555 * keyboard; the first drawn of several. Absent draws as before.
7556 */
7557 autoFocus?: true;
7558 };
7559 /**
7560 * Where the handlers live: the plugin whose hook drew the element, and
7561 * the handle its environment keeps the closures under.
7562 *
7563 * The runtime stamps the plugin as the tree leaves that hook.
7564 */
7565 press: {
7566 plugin: string;
7567 handle: number;
7568 };
7569 children?: undefined;
7570 } | {
7571 /**
7572 * A one-of-several picker on every surface; a pick raises `ui.select`
7573 * (`e.element` the key, `e.value` the option's value).
7574 *
7575 * Built by `<Select>` or the table's `t.Select`. The `onSelect` closure
7576 * stays in the plugin's own environment under `press.handle`, held as
7577 * a Button's is. A leaf: no children.
7578 */
7579 type: 'Select';
7580 props: {
7581 /**
7582 * The element's address: what `e.element` carries and what a matcher
7583 * names (`{ element: "peer" }`).
7584 */
7585 key: string;
7586 /**
7587 * Text drawn before the current value.
7588 */
7589 label?: string;
7590 /**
7591 * What can be picked, in the order drawn: each a value and the text
7592 * drawn for it.
7593 */
7594 options: readonly SelectOption[];
7595 /**
7596 * Which option is selected when drawn.
7597 */
7598 value?: string;
7599 /**
7600 * The site's ring starts on this element when the site takes the
7601 * keyboard; the first drawn of several. Absent draws as before.
7602 */
7603 autoFocus?: true;
7604 };
7605 /**
7606 * Where the handler lives: the plugin whose hook drew the element, and
7607 * the handle its environment keeps the closure under.
7608 *
7609 * The runtime stamps the plugin as the tree leaves that hook.
7610 */
7611 press: {
7612 plugin: string;
7613 handle: number;
7614 };
7615 children?: undefined;
7616 } | {
7617 /**
7618 * A hyperlink every surface draws: an OSC 8 span on the terminal (its
7619 * text then the URL in dim where unsupported), an anchor on desktop.
7620 *
7621 * Inline: its children are the text, strings and inline elements;
7622 * absent children the label, absent both the URL. `href` is `https:`
7623 * (or `http://localhost`) and bounded, or the tree is refused.
7624 */
7625 type: 'Link';
7626 props: LinkProps;
7627 children?: RenderNode[];
7628 } | {
7629 /**
7630 * Source code every surface draws with the engine's highlighter, tokens
7631 * coloured by language; under `format: 'diff'`, unified-diff hunks.
7632 *
7633 * A leaf: a dim gutter numbers the lines from `startLine`; a diff has
7634 * both gutters, markers, add and remove backgrounds. `source` is
7635 * bounded as a Text's string is, or the tree is refused.
7636 */
7637 type: 'Code';
7638 props: CodeProps;
7639 children?: undefined;
7640 } | {
7641 /**
7642 * A block of markdown every surface draws as it draws an assistant
7643 * reply's text: its own renderer, theme, hyperlinks and highlighting.
7644 *
7645 * Built by `<Markdown>` or the table's `t.Markdown`. A leaf: `text` is
7646 * bounded as a Text's string is, or the tree is refused. Without a
7647 * `press` its links are the surface's own, opened as it opens links.
7648 */
7649 type: 'Markdown';
7650 props: MarkdownLeafProps;
7651 children?: undefined;
7652 } | {
7653 /**
7654 * A `Markdown` whose plugin answers the links it drew (`onLinkPress`):
7655 * a press on one raises `ui.press` with `e.link`, the surface opens none.
7656 */
7657 type: 'Markdown';
7658 props: MarkdownLeafProps;
7659 /**
7660 * Where the `onLinkPress` handler lives: the plugin whose hook drew
7661 * the element, and the handle its environment keeps the closure under.
7662 *
7663 * The runtime stamps the plugin as the tree leaves that hook.
7664 */
7665 press: {
7666 plugin: string;
7667 handle: number;
7668 };
7669 children?: undefined;
7670 } | {
7671 /**
7672 * A region one of the plugin's SURFACE MODULES, named by path, draws and
7673 * handles input for on the drawing thread, without `$` (ClientModule).
7674 *
7675 * `Client` from the table. One instance per plugin, drawing and `key`
7676 * lives while the node stays in the tree, and talks to the plugin's
7677 * hooks through `ui.message`. The desktop carries it as data.
7678 */
7679 type: 'Client';
7680 props: ClientProps;
7681 /**
7682 * Whose surface module `props.module` names: the plugin whose hook
7683 * drew the element, stamped by the runtime as the tree leaves it.
7684 */
7685 client: {
7686 plugin: string;
7687 };
7688 children?: undefined;
7689 } | {
7690 /**
7691 * A vector drawing, the remote surfaces' alone: the SVG markup is the
7692 * element's data, drawn in an isolated box, off the page.
7693 *
7694 * A leaf: hooks above wrap or replace it whole, nothing reaches inside;
7695 * a press other plugins should see goes on an enclosing Button. On a
7696 * surface whose table lacks it the tree is refused.
7697 */
7698 type: 'Svg';
7699 props: SvgProps;
7700 children?: undefined;
7701 } | {
7702 /**
7703 * A grid of terminal cells, the terminal surface's alone: each cell a
7704 * glyph, a foreground and a background, packed in `props.cells`.
7705 *
7706 * A leaf, one node however many cells; hooks above wrap or replace it
7707 * whole. A mounted one is repainted in place by its plugin's
7708 * `$.ui.blit`. On a surface whose table lacks it the tree is refused.
7709 */
7710 type: 'Raster';
7711 props: RasterProps;
7712 /**
7713 * Whose Raster: the plugin whose hook drew the element, stamped by the
7714 * runtime as the tree leaves it; the one plugin whose blit reaches it.
7715 */
7716 raster: {
7717 plugin: string;
7718 };
7719 children?: undefined;
7720 } | {
7721 /**
7722 * A picture, the terminal surface's alone: `props.source` drawn over a
7723 * box of cells where the terminal can, `props.alt` where it cannot.
7724 *
7725 * A leaf: hooks above wrap or replace it whole; a press other plugins
7726 * should see goes on an enclosing Button; keyed, its plugin's
7727 * `$.ui.blit` swaps it. On a surface without it the tree is refused.
7728 */
7729 type: 'Image';
7730 props: ImageProps;
7731 /**
7732 * Whose Image: the plugin whose hook drew the element, stamped by the
7733 * runtime as the tree leaves it; the one plugin whose blit reaches it.
7734 */
7735 image: {
7736 plugin: string;
7737 };
7738 children?: undefined;
7739 } | {
7740 /**
7741 * The component core draws itself, with the props held under `ref`.
7742 */
7743 type: 'engine';
7744 /**
7745 * Which drawing: the number core answered from `next(e)`, under which
7746 * it holds the props it received; 0 draws the original props.
7747 */
7748 ref: number;
7749 };
7750
7751 /**
7752 * The render event: `ui.render`, one event for every component that has a
7753 * render site.
7754 */
7755 export type RenderEventName = 'ui.render';
7756
7757 /**
7758 * The input of `ui.render`: a union discriminated by `component`, one member
7759 * per RenderComponent and per RenderSurface.
7760 */
7761 export type RenderInput<C extends RenderComponent = RenderComponent, P extends RenderSurface = RenderSurface> = C extends RenderComponent ? P extends RenderSurface ? RenderInputOf<C, P> : never : never;
7762
7763 /**
7764 * One `ui.render` input, for a component narrowed to one surface.
7765 */
7766 export type RenderInputOf<C extends RenderComponent, P extends RenderSurface> = {
7767 /**
7768 * Where the tree will be drawn; one literal per member, so
7769 * `if (e.surface === "terminal")` narrows `e` and `$.ui.resolve(e)`.
7770 *
7771 * Over the wire it is what the client declared, a rendering fact and not a
7772 * trust signal: do not key policy on it.
7773 */
7774 surface: P;
7775 /**
7776 * Which component this instance is; the key a matcher narrows on.
7777 */
7778 component: C;
7779 /**
7780 * The instance: the tool_use_id for a dialog or tool row, the message id
7781 * for a message, the agent id for a spinner.
7782 *
7783 * Two drawings of one component are two instances.
7784 */
7785 requestId: string;
7786 /**
7787 * The size of what the surface draws into, in character cells, and whether
7788 * its layout docks a pane; absent where no surface has measured.
7789 *
7790 * Part of the envelope: a rewrite keeps it. On the terminal, the interactive
7791 * screen's, where a change of width re-draws every hooked site once the
7792 * resize settles; on a remote surface, what it reported.
7793 */
7794 viewport?: RenderViewport;
7795 /**
7796 * The component's plain-data props.
7797 */
7798 props: RenderPropsOf[C];
7799 };
7800
7801 /**
7802 * A node of a render tree: an element, or a string (text).
7803 */
7804 export type RenderNode = RenderElement | string;
7805
7806 /**
7807 * The plain-data props of each renderable component, as `ui.render` sees
7808 * them under `e.props`; a hook rewrites them with `next({ ...e, props })`.
7809 *
7810 * A rewrite is validated by the component and an invalid one draws the
7811 * original. This table is the plugin-facing render contract; the author's
7812 * file notes which surfaces raise each member (RENDER_SURFACES_OF).
7813 */
7814 export type RenderPropsOf = {
7815 /**
7816 * The dialog the AskUserQuestion tool opens.
7817 *
7818 * Raised on every surface.
7819 */
7820 AskUserQuestion: {
7821 /**
7822 * The name of the tool whose call opened the dialog (`AskUserQuestion`).
7823 */
7824 tool: string;
7825 /**
7826 * The tool's `questions` input, as the dialog will draw them; a rewrite
7827 * must still fit the tool's schema or the original is drawn.
7828 */
7829 questions: unknown[];
7830 /**
7831 * The call's `metadata.source` (who asked) when the model gave one.
7832 * Analytics only; never drawn.
7833 */
7834 metadataSource?: string;
7835 };
7836 /**
7837 * A user-role transcript row: the person's prompt (`> ...`), a background
7838 * task's notification, or a message another agent, teammate or session sent.
7839 *
7840 * `origin`, `task` and `from` tell them apart and are read-only; a rewrite
7841 * of `text` draws in the row alone: the stored message, and the model's
7842 * framing of another party's words as that party's, stay as they were.
7843 *
7844 * Raised on every surface.
7845 */
7846 UserMessage: {
7847 /**
7848 * What the row shows: the prompt as typed, a notification's summary, or
7849 * a message's body less the engine's framing (summary line included).
7850 *
7851 * A rewrite is printable, bounded text and draws where the engine draws
7852 * the body; a teammate block of several frames or with a summary line
7853 * keeps the engine's drawing, one string not being those parts.
7854 */
7855 text: string;
7856 /**
7857 * Where the stored message came from, as `prompt.submit` named it: the
7858 * composer's, a task notification's, a peer's, a channel's, a plugin's.
7859 *
7860 * `unclassified` for a teammate's (its drain stamps none) and for a
7861 * message stored before stamps. Read-only.
7862 */
7863 origin: PromptOrigin;
7864 /**
7865 * Whether the view draws the row in full: the ctrl+o transcript,
7866 * `--verbose`, a surface with no ctrl+o (an export). Read-only.
7867 *
7868 * False, a message row is one dim line naming its sender; a hook that
7869 * draws a compact row of its own passes when true, so ctrl+o shows all.
7870 */
7871 isExpanded: boolean;
7872 /**
7873 * What a notification row (`origin.kind` `task-notification`) reports
7874 * on: its background task; absent on every other row. Read-only.
7875 */
7876 task?: UserMessageTask;
7877 /**
7878 * Who sent the message the row carries: another agent of this session, a
7879 * teammate, another session, a channel; absent otherwise. Read-only.
7880 */
7881 from?: UserMessageFrom;
7882 /**
7883 * Which of its rows the transcript's viewport shows now: `null` while
7884 * drawn outside it, absent where the surface does not say. Read-only.
7885 *
7886 * Reported once drawn and again when it changes: on a scroll, for the
7887 * messages at the viewport's edges only. A rewrite carries it on as
7888 * received; one that changes or drops it is refused.
7889 */
7890 onScreen?: OnScreen | null;
7891 };
7892 /**
7893 * One text block of an assistant reply in the transcript; a rewrite
7894 * changes the drawing and leaves the stored message alone (ctrl+o).
7895 *
7896 * Raised on every surface.
7897 */
7898 AssistantMessage: {
7899 /**
7900 * The block's text, markdown, as the transcript will draw it.
7901 */
7902 text: string;
7903 /**
7904 * True on the block that draws the bullet opening a reply.
7905 */
7906 isFirstOfReply: boolean;
7907 /**
7908 * Which of its rows the transcript's viewport shows now: `null` while
7909 * drawn outside it, absent where the surface does not say. Read-only.
7910 *
7911 * Reported once drawn and again when it changes: on a scroll, for the
7912 * messages at the viewport's edges only. A rewrite carries it on as
7913 * received; one that changes or drops it is refused.
7914 */
7915 onScreen?: OnScreen | null;
7916 };
7917 /**
7918 * A tool call's row in the transcript (`Bash(ls -la)` and its result); the
7919 * call was decided by `tool.call`, so a rewrite here changes the row alone.
7920 *
7921 * Raised on every surface.
7922 */
7923 ToolUse: {
7924 /**
7925 * The id `tool.call` carried for this call (`e.tool_use_id` there); the
7926 * same value as the row's `requestId`, where it is looked for. Read-only.
7927 */
7928 tool_use_id: string;
7929 /**
7930 * Which one the row draws (`Bash`, `Read`, a plugin's tool), as
7931 * `tool.call` named it.
7932 */
7933 tool: string;
7934 /**
7935 * The call's input, as the model sent it.
7936 */
7937 input: unknown;
7938 /**
7939 * True while the call is still running.
7940 */
7941 isRunning: boolean;
7942 /**
7943 * True when the call ended in an error (a refusal at the dialog is one).
7944 */
7945 isErrored: boolean;
7946 /**
7947 * True when an abort ended the call: the user's Esc or a plugin's
7948 * `$.turn.abort` cut it while it ran, or dropped it before it ran.
7949 *
7950 * The row draws `Interrupted` for it, as the transcript marker does.
7951 */
7952 isInterrupted: boolean;
7953 /**
7954 * The stored result once the call has resolved (`{ stdout, stderr, ... }`
7955 * for Bash: `BuiltinToolResults[tool]`); undefined while it runs.
7956 *
7957 * For a call that errored, was refused or an abort cut, it is the text
7958 * the model read (an `isInterrupted` call: the abort's own). An expanded
7959 * group's rows draw it inline; a standalone row's is its own `ToolResult`.
7960 */
7961 output?: unknown;
7962 /**
7963 * Which of its rows the transcript's viewport shows now: `null` while
7964 * drawn outside it, absent where the surface does not say. Read-only.
7965 *
7966 * Reported once drawn and again when it changes: on a scroll, for the
7967 * messages at the viewport's edges only. A rewrite carries it on as
7968 * received; one that changes or drops it is refused.
7969 */
7970 onScreen?: OnScreen | null;
7971 };
7972 /**
7973 * The result block drawn under a standalone tool row in the transcript,
7974 * which the tool's own result renderer draws from `output`.
7975 *
7976 * A rewrite of `output` is checked against the tool's output schema (one
7977 * that does not fit draws nothing; one the renderer cannot read hits the
7978 * row's error boundary). The stored result is untouched.
7979 *
7980 * Raised on every surface.
7981 */
7982 ToolResult: {
7983 /**
7984 * The id `tool.call` carried for the call this result belongs to
7985 * (`e.tool_use_id` there); the same value as `requestId`. Read-only.
7986 */
7987 tool_use_id: string;
7988 /**
7989 * Which one the result belongs to (`Bash`, `Read`, a plugin's tool).
7990 * Read-only.
7991 */
7992 tool: string;
7993 /**
7994 * The tool's own result object (`{ stdout, stderr, interrupted, ... }` for
7995 * Bash), the same one `ToolUse.output` carries; a rewrite is drawn.
7996 *
7997 * A built-in tool's is `BuiltinToolResults[tool]`, the record
7998 * `tool.call` resolved as `result`.
7999 */
8000 output: unknown;
8001 /**
8002 * True when the call ended in an error, which draws the error text and not
8003 * `output`. Read-only.
8004 */
8005 isErrored: boolean;
8006 /**
8007 * Which of its rows the transcript's viewport shows now: `null` while
8008 * drawn outside it, absent where the surface does not say. Read-only.
8009 *
8010 * Reported once drawn and again when it changes: on a scroll, for the
8011 * messages at the viewport's edges only. A rewrite carries it on as
8012 * received; one that changes or drops it is refused.
8013 */
8014 onScreen?: OnScreen | null;
8015 };
8016 /**
8017 * A run of tool calls the transcript folds into one count line (`Read 3
8018 * files, ran 2 shell commands`): reads, searches, listings.
8019 *
8020 * A hook that sets `isExpanded` unfolds the group where it is, and each row
8021 * it unfolds into is a `ToolUse` drawing a `ToolUse` hook then sees.
8022 *
8023 * Raised on every surface.
8024 */
8025 ToolGroup: {
8026 /**
8027 * In the order the model made them.
8028 */
8029 calls: ReadonlyArray<ToolGroupCall>;
8030 /**
8031 * True while the group is the live one: a call in it may still be
8032 * running and the model's next call may join it.
8033 */
8034 isActive: boolean;
8035 /**
8036 * Whether each call draws as its own `ToolUse` row (true under
8037 * `--verbose` and in the ctrl+o transcript) or the group draws one line.
8038 *
8039 * The one prop of the three a rewrite changes on the screen.
8040 */
8041 isExpanded: boolean;
8042 /**
8043 * Which of its rows the transcript's viewport shows now: `null` while
8044 * drawn outside it, absent where the surface does not say. Read-only.
8045 *
8046 * Reported once drawn and again when it changes: on a scroll, for the
8047 * messages at the viewport's edges only. A rewrite carries it on as
8048 * received; one that changes or drops it is refused.
8049 */
8050 onScreen?: OnScreen | null;
8051 };
8052 /**
8053 * The output row a slash command printed in the transcript, under its echo
8054 * (`/cost`'s lines, the `text` a `command.run` hook answered).
8055 *
8056 * What the run resolved as text, a `local` command's or a plugin's alike. A
8057 * rewrite of `text` draws there and the stored row keeps what the model
8058 * reads; a hook's own tree draws in the row's place, the transcript's width.
8059 *
8060 * Raised on every surface.
8061 */
8062 CommandOutput: {
8063 /**
8064 * Which one printed the row, as `command.run` named it (no slash); a
8065 * hook on its own command matches by it.
8066 *
8067 * Read-only: a rewrite carries it on as received; one that changes or
8068 * drops it is refused and the engine draws its own row.
8069 */
8070 command: string;
8071 /**
8072 * The arguments the run had, as the echo above the row shows them
8073 * (`***` for a command that marks its arguments sensitive). Read-only.
8074 */
8075 args: string;
8076 /**
8077 * The row's text: what the command printed, or the `text` a hook
8078 * answered under its plugin's name; markdown, as the row draws it.
8079 */
8080 text: string;
8081 /**
8082 * True when the row is the run's error line (a command that threw),
8083 * which draws in the error colour and not dim. Read-only.
8084 */
8085 isErrored: boolean;
8086 /**
8087 * Which of its rows the transcript's viewport shows now: `null` while
8088 * drawn outside it, absent where the surface does not say. Read-only.
8089 *
8090 * Reported once drawn and again when it changes: on a scroll, for the
8091 * messages at the viewport's edges only. A rewrite carries it on as
8092 * received; one that changes or drops it is refused.
8093 */
8094 onScreen?: OnScreen | null;
8095 };
8096 /**
8097 * One row of the live tool-progress region under a running tool call; today
8098 * the run-in-background pill. One instance per row.
8099 *
8100 * A union on `kind`: other progress kinds join as their props become plain
8101 * data, each with its own fields, so a hook matches `{ props: { kind:
8102 * "background_hint" } }` or branches on `e.props.kind` and stays right.
8103 *
8104 * Raised on the terminal surface only.
8105 */
8106 ToolProgress: {
8107 /**
8108 * The tool call the row belongs to, as `tool.call` and the `ToolUse` row
8109 * carry it. Read-only: a rewrite carries it on as received.
8110 *
8111 * While several foreground calls share one pill, the first of them.
8112 */
8113 tool_use_id: string;
8114 /**
8115 * Which progress row: the run-in-background pill. Read-only.
8116 */
8117 kind: 'background_hint';
8118 /**
8119 * The pill's text as the engine draws it, dim, under the tool call:
8120 * `(ctrl+b to run in background)`, or with the person's own binding.
8121 *
8122 * A rewrite is drawn in its place, dim; `""` draws nothing. A hook that
8123 * returns a tree draws that instead.
8124 */
8125 hint: string;
8126 };
8127 /**
8128 * The line that animates while a turn runs (`Sauteing... (12s, 300
8129 * tokens)`); a remote surface draws its own.
8130 *
8131 * The row reads `word` (or `message` while one overrides it), then
8132 * `suffix`, then the dim parenthetical the engine keeps (elapsed time,
8133 * tokens, effort); a hook rewrites the first three or draws its own tree.
8134 *
8135 * Raised on the terminal surface only.
8136 */
8137 Spinner: {
8138 /**
8139 * Animated by the line (`Sauteing`), as sampled for this turn.
8140 */
8141 word: string;
8142 /**
8143 * The text drawn instead of the word while a state overrides it, else null.
8144 */
8145 message: string | null;
8146 /**
8147 * What the engine draws right after the word or message to say the turn is
8148 * still going: one ellipsis character.
8149 *
8150 * Left off a text that already ends in an ellipsis, so a rewritten
8151 * `message` ending in one shows one. A rewrite is drawn as given (`""`
8152 * draws the text bare, `" ~"` draws `Sauteing ~`); left out, the engine's.
8153 */
8154 suffix: string;
8155 /**
8156 * What the turn is doing.
8157 */
8158 mode: 'requesting' | 'responding' | 'thinking' | 'tool-input' | 'tool-use';
8159 };
8160 /**
8161 * The line that closes a turn in the transcript (`Baked for 3s`); a
8162 * remote surface draws its own footer.
8163 *
8164 * Raised on the terminal surface only.
8165 */
8166 TurnDuration: {
8167 /**
8168 * The past-tense word the line drew (`Baked`), as sampled for this line.
8169 */
8170 word: string;
8171 /**
8172 * The turn's duration in milliseconds, as the line formats it (`3s`,
8173 * `1m 4s`).
8174 */
8175 durationMs: number;
8176 /**
8177 * Which of its rows the transcript's viewport shows now: `null` while
8178 * drawn outside it, absent where the surface does not say. Read-only.
8179 *
8180 * Reported once drawn and again when it changes: on a scroll, for the
8181 * messages at the viewport's edges only. A rewrite carries it on as
8182 * received; one that changes or drops it is refused.
8183 */
8184 onScreen?: OnScreen | null;
8185 };
8186 /**
8187 * One dim status line under the logo (the model source, an experiment
8188 * enrollment, a settings hint), with a trailing `/command`.
8189 *
8190 * Raised on the terminal surface only.
8191 */
8192 InfoNotice: {
8193 /**
8194 * The notice's text, flattened to one string.
8195 */
8196 text: string;
8197 /**
8198 * The slash command appended after the text, or null when the notice has
8199 * none.
8200 */
8201 command: string | null;
8202 /**
8203 * Which of its rows the transcript's viewport shows now: `null` while
8204 * drawn outside it, absent where the surface does not say. Read-only.
8205 *
8206 * Reported once drawn and again when it changes: on a scroll, for the
8207 * messages at the viewport's edges only. A rewrite carries it on as
8208 * received; one that changes or drops it is refused.
8209 */
8210 onScreen?: OnScreen | null;
8211 };
8212 /**
8213 * The dim mode labels at the right of the prompt footer (`focus`, `memory
8214 * paused`), joined by ` & `. One instance.
8215 *
8216 * A hook adds a mode by rewriting `modes`, removes one by filtering, or
8217 * draws its own tree.
8218 *
8219 * Raised on the terminal and desktop surfaces only.
8220 */
8221 SessionMode: {
8222 /**
8223 * The labels the footer shows, in order; empty when there are none.
8224 */
8225 modes: readonly string[];
8226 };
8227 /**
8228 * The dim hint line under the prompt (`? for shortcuts`, `esc to
8229 * interrupt`, the pills beside them). One instance.
8230 *
8231 * A hook rewrites `hint` and the rewrite is drawn in the line's place, or
8232 * draws its own tree; `isDraft` and `isWorking` say what the line is for.
8233 *
8234 * Raised on the terminal surface only.
8235 */
8236 PromptHint: {
8237 /**
8238 * True while the prompt holds typed text. Read-only.
8239 */
8240 isDraft: boolean;
8241 /**
8242 * True while a model turn is running. Read-only.
8243 */
8244 isWorking: boolean;
8245 /**
8246 * The line's text as the engine draws it; one string, so a rewrite
8247 * replaces the line.
8248 *
8249 * Read from the drawn line the way the screen reader reads it, one space
8250 * between parts.
8251 */
8252 hint: string;
8253 };
8254 /**
8255 * The band directly above the prompt input, where the surveys draw; the
8256 * engine draws nothing of its own here.
8257 *
8258 * A hook draws a tree, or passes; one instance. The person collapses it
8259 * (ctrl+x ctrl+a, `[-]`) or focuses it (a click, ctrl+x tab): an Input
8260 * types, a Button arms the hotkeys, a tree taller than it scrolls.
8261 *
8262 * Raised on the terminal surface only.
8263 */
8264 AbovePrompt: {
8265 /**
8266 * True while a survey holds the band; a hook yields to it. Read-only.
8267 */
8268 hasSurvey: boolean;
8269 /**
8270 * True while a model turn is running. Read-only.
8271 */
8272 isWorking: boolean;
8273 /**
8274 * Rows the band may take: in fullscreen, what the bottom slot has left
8275 * above the prompt; otherwise the terminal's height. Read-only.
8276 *
8277 * That slot is capped at half the terminal's rows, the prompt's included.
8278 * A tree of at most `maxRows` rows shows whole; a taller one scrolls in a
8279 * window of `scroll.bodyRows`, and a bare digit arms no Button's hotkey.
8280 */
8281 maxRows: number;
8282 /**
8283 * Cells across the band: the terminal's width, or the transcript
8284 * column's while a `Pane` is docked beside it. Read-only.
8285 *
8286 * A tree wider than this wraps or truncates as its Text props say; size
8287 * a table or a rule to it rather than to `viewport.columns`.
8288 */
8289 bodyColumns: number;
8290 /**
8291 * The band's window over a tree taller than `maxRows`: engine-owned,
8292 * moved by the wheel, and by the person's keys while the band is focused.
8293 *
8294 * `bodyRows` is `maxRows` less the `n more` row. Read-only.
8295 */
8296 scroll: SiteScroll;
8297 /**
8298 * Which transcript is on screen above the band: the main conversation's
8299 * (no `agentId`) or one agent's, opened from the tasks list.
8300 *
8301 * The same band under either; a switch re-runs the hook with the new
8302 * view. Read-only: a rewrite carries it on as received; one that changes
8303 * or drops it is refused, the hook that passed it failing.
8304 */
8305 view: SiteView;
8306 };
8307 /**
8308 * The framed region a plugin opened with `$.ui.open({ id })`: one instance
8309 * per id (`requestId`), its body the hook's tree, one shown, the rest tabs.
8310 *
8311 * Placed by the surface (docked in fullscreen, else above the prompt) and
8312 * keyed by the person (ctrl+x tab, Esc), who closes it from the engine's mark
8313 * or ctrl+x x: `ui.close` with origin `person`, which a hook may refuse.
8314 *
8315 * Raised on every surface.
8316 */
8317 Pane: {
8318 /**
8319 * Its tab's label while more than one pane is open (with one, the engine
8320 * draws no title): the `title` it was opened with, or its id.
8321 *
8322 * A click on a tab, or Tab onto it and Enter, shows that pane. Read-only
8323 * here; another `$.ui.open` (or a hook on `ui.open`) retitles.
8324 */
8325 title: string;
8326 /**
8327 * True while the person has given the pane the keyboard: Tab walks its
8328 * elements, a Button's `hotkey` presses it, the arrows scroll. Read-only.
8329 */
8330 isFocused: boolean;
8331 /**
8332 * Cells across the body, inside the frame. Read-only.
8333 */
8334 bodyColumns: number;
8335 /**
8336 * Where the surface seated the pane: `dock` beside the transcript (the
8337 * terminal in fullscreen from 110 columns), or `inline` above the prompt.
8338 *
8339 * Read-only: a rewrite carries it on as received; one that changes or
8340 * drops it is refused, the hook that passed it failing.
8341 */
8342 placement: 'dock' | 'inline';
8343 /**
8344 * The body's window over the tree: engine-owned, moved by the person's
8345 * keys while the pane is focused. Read-only.
8346 */
8347 scroll: SiteScroll;
8348 /**
8349 * Which transcript is on screen beside the pane: the main conversation's
8350 * (no `agentId`) or one agent's, opened from the tasks list.
8351 *
8352 * The pane stays open across a switch, one instance re-rendered for the
8353 * view, so a hook draws for the agent in view. Read-only: a rewrite carries
8354 * it on as received; one that changes or drops it is refused.
8355 */
8356 view: SiteView;
8357 };
8358 };
8359
8360 /**
8361 * What a `ui.render` hook returns and what `next(e)` resolves to: a
8362 * RenderElement tree, the same for every component.
8363 */
8364 export type RenderResultOf = {
8365 [C in RenderComponent]: RenderElement;
8366 };
8367
8368 /**
8369 * Where a render event's component is drawn: `terminal` is Ink, which draws
8370 * the hook's whole tree; the rest are remote surfaces drawing it themselves.
8371 *
8372 * `desktop` is Claude Code Desktop, `mobile` the Claude mobile app, `vscode`
8373 * Claude Code for VS Code. A remote surface asks over the wire (ui_render),
8374 * draws with the props the hook handed core, and draws the tree where it has
8375 * a slot for it.
8376 *
8377 * Each surface's ask is its own evaluation, since a tree may hold an element
8378 * only some surfaces draw (Svg, Client).
8379 */
8380 export type RenderSurface = 'terminal' | 'desktop' | 'mobile' | 'vscode';
8381
8382 /**
8383 * The size of what a surface draws into, in character cells of the
8384 * surface's monospace metric, and whether its layout docks a pane.
8385 *
8386 * On the terminal, the conversation's columns and screen rows; on a remote
8387 * surface, the pane's width and height over the advance and line height of
8388 * its code font. Cell-based until the first element lays out in pixels.
8389 */
8390 export type RenderViewport = {
8391 /**
8392 * Cells across. A tree wider than this wraps or truncates, as its Text
8393 * props say.
8394 */
8395 columns: number;
8396 /**
8397 * Cells down the whole surface, not the room left for this component.
8398 *
8399 * Informational: a change of height alone re-draws nothing and keys no
8400 * new evaluation, so a hook reads it as of the last width or props change.
8401 */
8402 rows: number;
8403 /**
8404 * Whether this surface docks a pane beside the transcript, so a pane a
8405 * plugin opens unasked is a sidebar, not a takeover; absent is unknown.
8406 *
8407 * What `command.run`'s `presentation.isFullscreen` says. The terminal always
8408 * says: `true` fullscreen, `false` on the main screen, fixed per session; a
8409 * remote surface once its client reports it with the size, absent before.
8410 *
8411 * @remarks The desktop and VS Code report it once they place panes; the
8412 * mobile app, which has no dock, `false`. A change re-draws the sites.
8413 * @example if (e.viewport?.isFullscreen === true) void $.ui.open({ id })
8414 */
8415 isFullscreen?: boolean;
8416 };
8417
8418 /**
8419 * The input of `ui.resolve`: which surface's elements, for which component;
8420 * a union with one member per surface (ResolveInputOf).
8421 *
8422 * Resolved when the plugins load, ahead of any drawing, so it names the
8423 * surface and the component and not the props; a `ui.render` argument is
8424 * one (it carries both), which is what `$.ui.resolve(e)` takes.
8425 */
8426 export type ResolveInput<C extends RenderComponent = RenderComponent, P extends RenderSurface = RenderSurface> = P extends RenderSurface ? ResolveInputOf<C, P> : never;
8427
8428 /**
8429 * One `ui.resolve` input, for a component on one surface.
8430 */
8431 export type ResolveInputOf<C extends RenderComponent, P extends RenderSurface> = {
8432 /**
8433 * Where the table draws; one literal per member, so a hook that narrows
8434 * it narrows the table `next(e)` resolves to.
8435 */
8436 surface: P;
8437 /**
8438 * Which drawing the table is resolved for; the key a matcher narrows on.
8439 */
8440 component: C;
8441 };
8442
8443 /**
8444 * What each event's hook returns, and what its `next(e)` resolves to, by event
8445 * name.
8446 */
8447 export type ResultOf = EngineResultOf & ClassicResultOf & {
8448 [N in OpEventName]: OpEventResult<N>;
8449 } & {
8450 [N in NounEventName]: NounEventResult<N>;
8451 };
8452
8453 type SDKAssistantMessageError = 'authentication_failed' | 'oauth_org_not_allowed' | 'account_on_hold' | 'verification_required' | 'billing_error' | 'rate_limit' | 'overloaded' | 'invalid_request' | 'model_not_found' | 'server_error' | 'unknown' | 'max_output_tokens' | 'cloud_credential_error';
8454
8455 /**
8456 * The members of `T` assignable to `S`; all of `T` when none is.
8457 */
8458 type Select<T, S> = [Extract<T, S>] extends [never] ? T : Extract<T, S>;
8459
8460 /**
8461 * The events a pattern selects, as a union of names: the one named, every
8462 * one under a glob's namespace, or every one a negation does not exclude.
8463 */
8464 export type Selected<P extends string> = P extends '*' ? EventName : P extends `!${infer Negated}` ? Exclude<EventName, Selected<Negated>> : P extends `${infer Prefix}.*` ? Extract<EventName, `${Prefix}.${string}`> : Extract<EventName, P>;
8465
8466 /**
8467 * The literal each tag key of `I` is held to by matcher `M`: one-of arrays
8468 * flattened, RegExps widened to `unknown`.
8469 */
8470 type Selection<I, M> = {
8471 [K in keyof M & TagKeys<I>]: Literal<M[K] extends readonly (infer One)[] ? One : M[K]>;
8472 };
8473
8474 /**
8475 * One option of a `Select`: the value `onSelect` and `ui.select` carry, and
8476 * the text drawn for it (the value when absent).
8477 */
8478 export type SelectOption = {
8479 value: string;
8480 label?: string;
8481 };
8482
8483 /**
8484 * The props of `Select`, every surface's one-of-several picker: an address,
8485 * a label, the options, the one selected, the closure a pick runs. A leaf.
8486 *
8487 * Focused through the same ring as `Button` (`abovePrompt:focus`); keys reach
8488 * it only while it has focus (arrows move, Enter picks) and Esc always
8489 * returns to the prompt; a pick raises `ui.select`, its bottom `onSelect`.
8490 */
8491 export type SelectProps = {
8492 /**
8493 * The element's address: `e.element` at `ui.select`, what a matcher names.
8494 */
8495 key: string;
8496 /**
8497 * Text drawn before the current value.
8498 */
8499 label?: string;
8500 /**
8501 * What can be picked, in the order drawn; at least one, values unique.
8502 */
8503 options: readonly SelectOption[];
8504 /**
8505 * Which option is selected when drawn; the person's pick replaces it
8506 * until the hook draws another.
8507 */
8508 value?: string;
8509 /**
8510 * The site's focus ring starts here when the site takes the keyboard,
8511 * instead of on nothing, as the DOM's `autofocus`: Enter acts on it at once.
8512 *
8513 * A pane opened with `focus`, or the person's focus chord or click, is the
8514 * take. Of several in one site the first drawn wins; it raises `ui.focus`,
8515 * origin this plugin. A ring the person has moved stays where it was put.
8516 */
8517 autoFocus?: true;
8518 /**
8519 * Runs on a pick with the option's value, in the plugin's own environment:
8520 * the bottom of a `ui.select` chain. No model turn unless it asks one.
8521 */
8522 onSelect: (value: string, e: UiSelectArgument) => void;
8523 };
8524
8525 /**
8526 * The input of `session.attach`: a surface joined the session's roster of
8527 * attached clients (a phone opened the session; the desktop app connected).
8528 *
8529 * A remote client attaches by saying so (ui_attach) or by its first ask to
8530 * draw; the terminal's attachment is the REPL's binding and raises nothing.
8531 * Every field is the engine's: a hook observes, `next(e)` passes them on.
8532 */
8533 export type SessionAttachInput = {
8534 /**
8535 * Where the client draws: what it declared, a rendering fact.
8536 */
8537 surface: RenderSurface;
8538 /**
8539 * The client's own id, unique in the roster; two phones are two clients on
8540 * one surface. A client that never named itself is `<surface>:default`.
8541 */
8542 clientId: string;
8543 /**
8544 * The size the client draws into, in character cells, and whether its
8545 * layout docks a pane, when it said.
8546 */
8547 viewport?: RenderViewport;
8548 };
8549
8550 /**
8551 * What a `session.attach` hook returns and what `next(e)` resolves to:
8552 * `{ clientId }`, echoed by core; a hook's own value changes nothing.
8553 */
8554 export type SessionAttachResult = {
8555 clientId: string;
8556 };
8557
8558 /**
8559 * What `$.session.authorize()` answers: an opaque handle for the session's
8560 * Anthropic credential and its kind, or null when there is none to hold.
8561 *
8562 * The secret itself stays with the engine; the handle is its plugin-side
8563 * face, spent through `$.http.fetch(url, { auth: handle })`.
8564 */
8565 export type SessionAuthorization = {
8566 handle: string;
8567 kind: 'bearer' | 'api-key';
8568 } | null;
8569
8570 /**
8571 * `session.compact`'s input as a plugin's `$.session.compact(args)` takes
8572 * it: `trigger` (`plugin`), `messages` and `agentId` are the engine's.
8573 */
8574 export type SessionCompactArgs = {
8575 /**
8576 * What the summary should keep or stress, as typed after `/compact`;
8577 * absent, the engine's own compaction prompt alone.
8578 */
8579 instructions?: string;
8580 };
8581
8582 /**
8583 * A compaction that stands: the conversation as it reads afterwards, and
8584 * the token counts the engine recorded when it was the one compacting.
8585 */
8586 export type SessionCompacted = {
8587 /**
8588 * The conversation after the compaction, oldest first: from core, its
8589 * summary and the messages it kept; from a hook, whatever it hands up.
8590 *
8591 * What the transcript becomes (kept, on `precompute`, for the compaction
8592 * that comes). A message with the engine's `handle` stands as the engine
8593 * has it; one without is built from its `role`, `text` and tool blocks.
8594 */
8595 messages: readonly SessionMessage[];
8596 /**
8597 * The conversation's size before, in tokens; absent when core did not
8598 * record it.
8599 */
8600 tokensBefore?: number;
8601 /**
8602 * Its size afterwards, in tokens; absent on `precompute` and when core
8603 * did not record it.
8604 */
8605 tokensAfter?: number;
8606 skip?: undefined;
8607 };
8608
8609 /**
8610 * The input of `session.compact`: one compaction of the conversation, about
8611 * to run; `messages` is the transcript it runs over.
8612 */
8613 export type SessionCompactInput = {
8614 /**
8615 * What is compacting (SessionCompactTrigger); the key a matcher narrows
8616 * on. Pinned: `next(e)` passes it on as received.
8617 */
8618 trigger: SessionCompactTrigger;
8619 /**
8620 * The id of the loop compacting, for a subagent's or a fork's own
8621 * transcript; absent for the main conversation (tool.call's `agentId`).
8622 *
8623 * Pinned: left out of a rewrite it is put back, changed it fails the hook.
8624 * `$.session.messages()` is the main conversation whatever this names.
8625 */
8626 agentId?: string;
8627 /**
8628 * What the summary should keep or stress: the text after `/compact`, a
8629 * plugin's, or absent.
8630 *
8631 * `next({ ...e, instructions })` changes what the summarizer is told.
8632 */
8633 instructions?: string;
8634 /**
8635 * The transcript being compacted, in `$.session.messages()`'s shape, each
8636 * message carrying the engine's `handle`; a subagent's own under `agentId`.
8637 *
8638 * `next({ ...e, messages })` changes what is summarized: a message kept
8639 * with its handle is the engine's own, whole; one without is read as built.
8640 */
8641 messages: readonly SessionMessage[];
8642 };
8643
8644 /**
8645 * What a `session.compact` hook returns and what `next(e)` resolves to: the
8646 * compaction (`{ messages, tokensBefore?, tokensAfter? }`) or `{ skip }`.
8647 *
8648 * There is no summary string: the summary is a message.
8649 */
8650 export type SessionCompactResult = SessionCompacted | SessionCompactSkipped;
8651
8652 /**
8653 * A compaction vetoed: on `precompute` nothing is computed or kept; on any
8654 * other trigger the conversation stays as it is and one line says why.
8655 *
8656 * Core answers it too, when a classic PreCompact hook blocks.
8657 */
8658 export type SessionCompactSkipped = {
8659 /**
8660 * Why, as the notice reads.
8661 */
8662 skip: string;
8663 messages?: undefined;
8664 };
8665
8666 /**
8667 * Who compacts: the person's `/compact` (`manual`), the engine at its
8668 * threshold or on a prompt too long (`auto`), a plugin, or a `precompute`.
8669 *
8670 * `precompute` is the one dispatch that installs nothing: its result is kept
8671 * for the compaction that comes, if the conversation it ran over still leads.
8672 */
8673 export type SessionCompactTrigger = 'manual' | 'auto' | 'plugin' | 'precompute';
8674
8675 /**
8676 * The context window broken down as /context breaks it down: the rows, the
8677 * grid and the lists beneath it, in the SDK's `get_context_usage` shape.
8678 *
8679 * Less that reply's internal-build sections. Token counts are the engine's
8680 * estimates as numbers, never formatted: a `full` breakdown counts with the
8681 * token-count API where it can, a `summary` one estimates throughout.
8682 */
8683 export type SessionContextBreakdown = {
8684 /**
8685 * One row per category, the free space and the compaction buffer among
8686 * them; `kind` says which is which.
8687 */
8688 categories: ContextCategory[];
8689 /**
8690 * Tokens in use, unclamped: past `rawMaxTokens` when over the window.
8691 */
8692 totalTokens: number;
8693 /**
8694 * The window measured against, the same figure as `rawMaxTokens`.
8695 */
8696 maxTokens: number;
8697 /**
8698 * The window measured against, in tokens: the model's limit, or a smaller
8699 * compaction window (`autocompactSource` says which).
8700 */
8701 rawMaxTokens: number;
8702 /**
8703 * How that window was settled (ContextWindowSource); the SDK's reply
8704 * carries it under this name too, outside its typed schema.
8705 */
8706 autocompactSource: ContextWindowSource;
8707 /**
8708 * `totalTokens` over `rawMaxTokens` as a whole percentage, 0 to 100 and
8709 * past it when over.
8710 */
8711 percentage: number;
8712 /**
8713 * The grid, row by row: 10 by 10, 20 by 10 for a window of a million or
8714 * more, 5 wide when asked for under 80 `columns`.
8715 */
8716 gridRows: ContextGridSquare[][];
8717 /**
8718 * Which model the breakdown was computed for, as `/model` shows it.
8719 */
8720 model: string;
8721 /**
8722 * Each memory file in the context, with its path and tokens.
8723 */
8724 memoryFiles: ContextMemoryFile[];
8725 /**
8726 * Each MCP tool's schema, with its server and tokens.
8727 */
8728 mcpTools: ContextMcpTool[];
8729 /**
8730 * The custom agents the Agent tool describes, each with its tokens.
8731 */
8732 agents: ContextAgent[];
8733 /**
8734 * The slash-command listing, counted; absent when the session lists none.
8735 */
8736 slashCommands?: ContextSlashCommands;
8737 /**
8738 * The skill listing, counted, with one entry per skill; absent when the
8739 * session lists none.
8740 */
8741 skills?: ContextSkills;
8742 /**
8743 * The token count at which auto-compaction runs; absent when it is off.
8744 */
8745 autoCompactThreshold?: number;
8746 /**
8747 * Whether auto-compaction is on for the session.
8748 */
8749 isAutoCompactEnabled: boolean;
8750 /**
8751 * The last response's own token counts, or null before one came back.
8752 */
8753 apiUsage: ContextApiUsage | null;
8754 };
8755
8756 /**
8757 * The live context window as the status line reads it, and by category as
8758 * /context breaks it down when the call asked (`{ breakdown }`).
8759 *
8760 * `tokens` and `percent` are the last API response's input side against the
8761 * model's window, absent until the first response of the live window: a
8762 * fresh session, or one just compacted, until its next response.
8763 */
8764 export type SessionContextUsage = {
8765 /**
8766 * Input tokens the last response was answered over: uncached, cache-written
8767 * and cache-read together (the status line's `total_input_tokens`).
8768 */
8769 tokens?: number;
8770 /**
8771 * The context window of the session's model, in tokens (the status line's
8772 * `context_window_size`).
8773 */
8774 window: number;
8775 /**
8776 * `tokens` over `window` as a whole percentage, 0 to 100 (the status
8777 * line's `used_percentage`).
8778 */
8779 percent?: number;
8780 /**
8781 * The window by category, as /context breaks it down: present only when
8782 * the call passed `breakdown`, and only with a session bound.
8783 *
8784 * It measures against the compaction window (`rawMaxTokens`), which may be
8785 * smaller than `window`, and estimates every category, so its
8786 * `totalTokens` need not equal `tokens`.
8787 */
8788 breakdown?: SessionContextBreakdown;
8789 };
8790
8791 /**
8792 * What the session has cost, as `/cost` and the status line total it.
8793 */
8794 export type SessionCost = {
8795 /**
8796 * US dollars, summed over every priced API response this session.
8797 */
8798 usd: number;
8799 };
8800
8801 type SessionCronSummary = {
8802 id: string;
8803 /**
8804 * Cron expression, e.g. "0 9 * * 1-5".
8805 */
8806 schedule: string;
8807 /**
8808 * False for one-shot wakeups whose cron field encodes a single fire time; true for tasks that re-fire on every match.
8809 */
8810 recurring: boolean;
8811 /**
8812 * Prompt text submitted when the cron fires. Capped at 1000 chars; clipped values append an in-string "... [+N chars]" marker.
8813 */
8814 prompt: string;
8815 };
8816
8817 /**
8818 * The input of `session.detach`: a client left the session's roster. Every
8819 * field is the engine's: a hook observes, and `next(e)` passes them on.
8820 */
8821 export type SessionDetachInput = {
8822 /**
8823 * Where the client drew.
8824 */
8825 surface: RenderSurface;
8826 /**
8827 * The id it attached under.
8828 */
8829 clientId: string;
8830 /**
8831 * Why it left: it detached, or the session ended under it.
8832 */
8833 reason: SessionDetachReason;
8834 };
8835
8836 /**
8837 * Why a client left the roster: `detach`, it said so (ui_detach); `end`, the
8838 * session ended with it still attached.
8839 */
8840 export type SessionDetachReason = 'detach' | 'end';
8841
8842 /**
8843 * What a `session.detach` hook returns and what `next(e)` resolves to:
8844 * `{ clientId }`, echoed by core; a hook's own value changes nothing.
8845 */
8846 export type SessionDetachResult = {
8847 clientId: string;
8848 };
8849
8850 type SessionEndHookInput = BaseHookInput & {
8851 hook_event_name: 'SessionEnd';
8852 reason: ExitReason;
8853 };
8854
8855 /**
8856 * The input of `session.end`: the session is ending, why, and how to come back
8857 * to it.
8858 *
8859 * Every field is the engine's: a hook observes, and `next(e)` passes them on.
8860 */
8861 export type SessionEndInput = {
8862 /**
8863 * Why it ends (SessionEndReason), the word the classic SessionEnd hook
8864 * receives as its `reason`.
8865 */
8866 reason: SessionEndReason;
8867 /**
8868 * The ending session's id (`$.session.id()` until now); after a `/clear`
8869 * or a resume the process goes on under another.
8870 */
8871 sessionId: string;
8872 /**
8873 * What `claude --resume` takes to return to it.
8874 */
8875 resume: SessionResume;
8876 };
8877
8878 /**
8879 * Why the session ended: the classic SessionEnd hook's own `reason`, word for
8880 * word.
8881 *
8882 * `prompt_input_exit`, the person left (/exit, ctrl+c, ctrl+d); `clear`, a
8883 * /clear started a fresh one; `resume`, another took its place; `logout`;
8884 * `other`, a `-p` run finished or the process got SIGINT, SIGTERM or SIGHUP.
8885 */
8886 export type SessionEndReason = ClassicHookInputs['SessionEnd']['reason'];
8887
8888 /**
8889 * What a `session.end` hook returns and what `next(e)` resolves to:
8890 * `{ sessionId }`, echoed by core; a hook's own value changes nothing.
8891 */
8892 export type SessionEndResult = {
8893 sessionId: string;
8894 };
8895
8896 /**
8897 * The input of `session.measure`: what `$.session.usage()` answers at this
8898 * moment, and which of its units moved since the last measurement a hook saw.
8899 *
8900 * The same figures the op reads, without `context.breakdown`: call
8901 * `$.session.usage({ breakdown })` from the hook when the categories matter.
8902 * Every field is the engine's: a hook observes, and `next(e)` passes them on.
8903 */
8904 export type SessionMeasureInput = {
8905 /**
8906 * The live context window (`$.session.usage()`'s `context`): the window,
8907 * and its fill once a response of the live window reported one.
8908 *
8909 * Never carries `breakdown` here.
8910 */
8911 context: SessionContextUsage;
8912 /**
8913 * The rate-limit windows the last response reported, each with its
8914 * `percentUsed`; empty off a subscription or before the first reading.
8915 */
8916 rateLimits: SessionRateLimit[];
8917 /**
8918 * What the session has cost so far; absent where the host keeps no ledger.
8919 */
8920 cost?: SessionCost;
8921 /**
8922 * Which units differ from the last measurement raised (UsageUnit), never
8923 * empty; the first measurement names every unit it has a figure for.
8924 *
8925 * `context`: the fill moved; `rateLimits`: a window moved a whole point,
8926 * appeared or left, or the account's limit status changed; `cost`: the
8927 * total grew.
8928 */
8929 changed: UsageUnit[];
8930 };
8931
8932 /**
8933 * What a `session.measure` hook returns and what `next(e)` resolves to:
8934 * `{ changed }`, echoed by core; a hook's own value changes nothing.
8935 */
8936 export type SessionMeasureResult = {
8937 changed: UsageUnit[];
8938 };
8939
8940 /**
8941 * One message of the transcript as `$.session.messages()` returns it.
8942 */
8943 export type SessionMessage = {
8944 /**
8945 * Who wrote it.
8946 */
8947 role: 'user' | 'assistant';
8948 /**
8949 * Its text blocks joined; '' when it has none.
8950 */
8951 text: string;
8952 /**
8953 * The tool_use blocks of an assistant message, each with its outcome
8954 * (`result`, `text`, `isError`) once the transcript holds it.
8955 */
8956 toolUses: ToolUseSummary[];
8957 /**
8958 * The tool_result blocks of a user message: `{ tool_use_id, text, isError,
8959 * result }`.
8960 */
8961 toolResults?: ToolResultSummary[];
8962 /**
8963 * An opaque token the engine stamps on a message it hands a
8964 * `session.compact` hook, so a list handed back maps to its own messages.
8965 *
8966 * Absent on `$.session.messages()` and on a message a hook built.
8967 */
8968 handle?: string;
8969 };
8970
8971 /**
8972 * One rate-limit window as the rate-limit notices read it.
8973 */
8974 export type SessionRateLimit = {
8975 /**
8976 * Which window: `five_hour`, `seven_day`, or a Claude gateway's
8977 * `spend_limit`.
8978 */
8979 kind: string;
8980 /**
8981 * How much of the window is used, 0 to 100 with at most one decimal:
8982 * 23.5, or 7, never 7.000000000000001; past 100 on an exceeded spend limit.
8983 */
8984 percentUsed: number;
8985 /**
8986 * When the window resets, as an ISO 8601 timestamp.
8987 */
8988 resetsAt?: string;
8989 };
8990
8991 /**
8992 * An external-event wake the delivery's text parsed as (a GitHub relay
8993 * event, a signal's notice): the envelope's attributes and its JSON body.
8994 */
8995 export type SessionReceiveEvent = {
8996 /**
8997 * The producing service (`github`).
8998 */
8999 source: string;
9000 /**
9001 * The event's discriminator (`pull_request.closed`, `check_suite`).
9002 */
9003 kind: string;
9004 /**
9005 * The envelope's JSON body (`{ pr: "acme/app#12", outcome: "merged" }`).
9006 */
9007 data: Record<string, unknown>;
9008 /**
9009 * The `data` keys whose values are free text a third party wrote (a
9010 * commenter's `author` and `comment`), as the producer marked them.
9011 *
9012 * Empty when none; the server asserts every other key's value, not these.
9013 */
9014 untrustedKeys: readonly string[];
9015 };
9016
9017 /**
9018 * The input of `session.receive`: one inbound delivery (a relay's event, a
9019 * peer's message, a Remote Control prompt), sanitized, before it is queued.
9020 */
9021 export type SessionReceiveInput = {
9022 /**
9023 * Where the delivery came from, as the bridge classified it; the key a
9024 * matcher narrows on. `next(e)` passes it on as received.
9025 */
9026 origin: SessionReceiveOrigin;
9027 /**
9028 * The delivery's body as the model would read it (a content-block
9029 * delivery's text blocks, joined).
9030 */
9031 text: string;
9032 /**
9033 * The external-event wake the text parsed as (`source` `github` for a
9034 * subscribed pull request's event); passed on as received.
9035 *
9036 * Present only for a delivery the server itself rendered as a wake (a
9037 * webhook relay, the session inbox), by its asserted origin; anyone else's
9038 * wake-shaped text stays `text`. Its `untrustedKeys` name third-party text.
9039 */
9040 event?: SessionReceiveEvent;
9041 };
9042
9043 /**
9044 * Where an inbound delivery came from, as the bridge classified it from the
9045 * server's stamps: `prompt.submit`'s `e.origin`, less what never arrives.
9046 */
9047 export type SessionReceiveOrigin = {
9048 /**
9049 * `bridge` the user's own from a Remote Control client; `peer` another
9050 * session; `unclassified` one the bridge could not place.
9051 *
9052 * `task-notification` a relay's event (a GitHub webhook) or a trigger;
9053 * `scheduled-trigger` a routine; `peer-send-message` another session's
9054 * SendMessage; `projects-relay` and `slack-ping` as at `prompt.submit`.
9055 */
9056 kind: 'bridge' | 'task-notification' | 'scheduled-trigger' | 'peer' | 'peer-send-message' | 'projects-relay' | 'slack-ping' | 'unclassified';
9057 };
9058
9059 /**
9060 * What a `session.receive` hook returns and what `next(e)` resolves to: the
9061 * delivery that was queued, `{ text }`, or `{ consumed: reason }`.
9062 *
9063 * `next(e)` resolves once the delivery is queued; a hook answering without
9064 * `next` queues nothing.
9065 */
9066 export type SessionReceiveResult = {
9067 /**
9068 * The body the delivery was queued with; from core, `e.text` as it
9069 * arrived at the bottom.
9070 *
9071 * A rewrite passes it down, `next({ ...e, text })`.
9072 */
9073 text: string;
9074 consumed?: undefined;
9075 } | {
9076 /**
9077 * Takes the delivery, answered without `next`: it is not queued, not
9078 * written to the transcript and never reaches the model.
9079 *
9080 * The reason is logged, the plugin named.
9081 */
9082 consumed: string;
9083 text?: undefined;
9084 };
9085
9086 /**
9087 * What `$.session.repo()` answers: the repository's root and its origin remote,
9088 * when the session is in one.
9089 */
9090 export type SessionRepo = {
9091 /**
9092 * The repository's root, absolute: the main working tree's for a worktree.
9093 */
9094 root: string;
9095 /**
9096 * The `origin` remote's URL as git has it (push URL preferred), or null when
9097 * the repository has none.
9098 */
9099 remote: string | null;
9100 /**
9101 * Whether the remote is one of the repositories this build treats as its
9102 * own; false in a build that has none or when the remote is unrecognized.
9103 *
9104 * The engine matches the build's own list of repositories with its
9105 * hardened remote parser. A plugin reads this to behave differently in a
9106 * public repository; which repositories are the build's is its to say.
9107 */
9108 internal: boolean;
9109 /**
9110 * The repository the allowlist matched, as `owner/name`; null when
9111 * `internal` is false.
9112 *
9113 * A working copy without a remote that the engine still recognises is
9114 * named by its own checkout configuration; a remote's name is its path.
9115 */
9116 name: string | null;
9117 };
9118
9119 /**
9120 * How to come back to the session that ended: what `claude --resume` takes.
9121 */
9122 export type SessionResume = {
9123 /**
9124 * What `claude --resume <id>` takes, the ending session's own.
9125 *
9126 * It resumes a session that wrote a transcript: one that never ran a
9127 * prompt, or ran with persistence off, left nothing under this id.
9128 */
9129 id: string;
9130 };
9131
9132 type SessionStartHookInput = BaseHookInput & {
9133 hook_event_name: 'SessionStart';
9134 source: 'startup' | 'resume' | 'clear' | 'compact' | 'fork';
9135 agent_type?: string;
9136 model?: string;
9137 session_title?: string;
9138 /**
9139 * resume/fork: seconds since the resumed transcript's last assistant response
9140 */
9141 seconds_since_last_response?: number;
9142 /**
9143 * resume/fork: the resumed transcript's last response input + cache_read + cache_creation + output tokens (for a server-side tool loop, its last iteration's window, not the summed totals)
9144 */
9145 context_tokens?: number;
9146 /**
9147 * resume/fork: seconds_since_last_response exceeds the prompt-cache TTL, so the first request re-caches context_tokens
9148 */
9149 prompt_cache_likely_expired?: boolean;
9150 /**
9151 * resume/fork: estimated cost of re-caching context_tokens on the session model - the managed modelPricing when set, otherwise list price; excludes the response
9152 */
9153 estimated_cache_write_usd?: number;
9154 };
9155
9156 /**
9157 * The input of `session.start`: the session the process starts with, read the
9158 * way `$.session` reads it at that moment.
9159 */
9160 export type SessionStartInput = {
9161 /**
9162 * The directory the session runs in, absolute (`$.session.cwd()`).
9163 */
9164 cwd: string;
9165 /**
9166 * Where the session draws at start (`$.session.surfaces()[0]`): `terminal`
9167 * under the REPL; null for a `-p` run or the SDK, which draw nowhere yet.
9168 */
9169 surface: RenderSurface | null;
9170 /**
9171 * Whether a person is at the prompt: true under the REPL, false for a `-p`
9172 * run or the SDK.
9173 */
9174 isInteractive: boolean;
9175 };
9176
9177 /**
9178 * What a `session.start` hook returns and what `next(e)` resolves to:
9179 * `{ cwd }`, echoed by core; a hook's own value does not change the session.
9180 */
9181 export type SessionStartResult = {
9182 cwd: string;
9183 };
9184
9185 /**
9186 * What `$.session.usage()` answers: the context window's fill, the account's
9187 * rate-limit windows and the session's cost, as the status line has them.
9188 *
9189 * A figure the engine does not have is left out, never zeroed: `rateLimits`
9190 * is empty off a subscription, `cost` absent where the host keeps no ledger.
9191 */
9192 export type SessionUsage = {
9193 /**
9194 * The live context window: the status line's `context_window` figures.
9195 */
9196 context: SessionContextUsage;
9197 /**
9198 * The rate-limit windows the last API response reported (`five_hour`,
9199 * `seven_day`, a gateway's `spend_limit`); empty when none has a reading.
9200 */
9201 rateLimits: SessionRateLimit[];
9202 /**
9203 * What the session has cost so far, as /cost totals it; absent only where
9204 * the host keeps no cost ledger (the CLI always has one).
9205 */
9206 cost?: SessionCost;
9207 };
9208
9209 /**
9210 * What `$.session.usage(args)` takes: nothing for the status line's figures
9211 * alone; `breakdown` to have the window broken down as /context breaks it.
9212 */
9213 export type SessionUsageArgs = {
9214 /**
9215 * Asks for `context.breakdown` and says how it is counted
9216 * (ContextBreakdownDetail); absent, none is computed and the call is free.
9217 *
9218 * `full` sends one token-count request per tool and memory file, as
9219 * /context does; `summary` estimates locally and sends none.
9220 */
9221 breakdown?: ContextBreakdownDetail;
9222 /**
9223 * The width the breakdown's grid will be drawn in, in terminal columns;
9224 * absent, the full-width grid. Read only with `breakdown`.
9225 *
9226 * Under 80 the grid is the narrow one /context draws there, 5 wide. A thin
9227 * client's breakdown is the remote workspace's, whose grid ignores this.
9228 */
9229 columns?: number;
9230 };
9231
9232 /**
9233 * What `$.session.version()` answers: the version of the engine the session
9234 * runs on, the release that version is built from, and when it was built.
9235 *
9236 * The three are what the engine's own analytics rows carry as `version`,
9237 * `version_base` and `build_time`, so a row a plugin sends and a row the
9238 * engine sends from the same binary agree.
9239 */
9240 export type SessionVersion = {
9241 /**
9242 * The engine's full version, as `claude --version` prints it: a release's
9243 * (`2.1.280`) or a development build's, which adds the build's date, time
9244 * and commit (`2.1.280-dev.20260920.t101500.sha1a2b3c4`).
9245 */
9246 version: string;
9247 /**
9248 * The release the version is built from, its semantic core and channel:
9249 * `2.1.280` for that release, `2.1.280-dev` for a development build of it.
9250 * Absent when the version is not spelled as a release.
9251 */
9252 base?: string;
9253 /**
9254 * When the binary was built, an ISO 8601 timestamp
9255 * (`2026-09-20T10:15:00Z`). Absent in a run from source that stamps none.
9256 */
9257 builtAt?: string;
9258 };
9259
9260 /**
9261 * What `$.settings.read` answers: an object keyed as a settings.json is
9262 * (`permissions`, `env`, `hooks`, `model`, `enabledPlugins`, ...).
9263 *
9264 * Each key is as https://code.claude.com/docs/en/settings documents it and
9265 * https://json.schemastore.org/claude-code-settings.json types it. A
9266 * snapshot in plain data: a write to it changes nothing the engine reads.
9267 */
9268 export type Settings = Readonly<Record<string, unknown>>;
9269
9270 /**
9271 * What `$.settings.read(args)` takes.
9272 */
9273 export type SettingsReadArgs = {
9274 /**
9275 * The one source read, as loaded; absent, the settings merged over every
9276 * source, as the engine runs under them.
9277 */
9278 source?: SettingsSource;
9279 };
9280
9281 /**
9282 * One source of settings by the name a plugin gives it, lowest precedence
9283 * first; the engine's own name is the same word with `Settings` appended.
9284 *
9285 * `user` is ~/.claude/settings.json, `project` .claude/settings.json,
9286 * `local` .claude/settings.local.json, `flag` what `--settings` and the
9287 * SDK's inline settings carry, and `policy` the managed settings, every
9288 * managed tier merged.
9289 */
9290 export type SettingsSource = 'user' | 'project' | 'local' | 'flag' | 'policy';
9291
9292 type SetupHookInput = BaseHookInput & {
9293 hook_event_name: 'Setup';
9294 trigger: 'init' | 'maintenance';
9295 };
9296
9297 /**
9298 * Where a site's window sits over the tree a hook drew in it (a pane's body,
9299 * the band): the engine's to move (the person scrolls), the plugin's to read.
9300 */
9301 export type SiteScroll = {
9302 /**
9303 * The first row of the tree the window shows; 0 at the top. Read-only.
9304 */
9305 offset: number;
9306 /**
9307 * How many rows of the tree the window shows at once: the rows the
9308 * surface gave the body. Read-only.
9309 */
9310 bodyRows: number;
9311 };
9312
9313 /**
9314 * Which transcript the person has on screen where a site draws: the main
9315 * conversation's, or one agent's, opened from the tasks list.
9316 *
9317 * The person's to switch, the plugin's to read: a switch re-runs the site's
9318 * hooks with the new view, the site itself staying where it is.
9319 */
9320 export type SiteView = {
9321 /**
9322 * The agent whose transcript is in view: the `id` `$.agent.list()` gives
9323 * it, the `agentId` its `turn.step` and `tool.call` events carry.
9324 *
9325 * Absent while the main conversation is in view. Read-only.
9326 */
9327 agentId?: string;
9328 };
9329
9330 /**
9331 * The input of `skill.prompt`: one skill's prompt, at the moment the engine
9332 * expanded it for the model.
9333 *
9334 * Typed as `/name`, called through the Skill tool, or preloaded into a
9335 * subagent: the same event at each.
9336 */
9337 export type SkillPromptInput = {
9338 /**
9339 * Which skill (`commit`); the key a matcher narrows on.
9340 */
9341 skill: string;
9342 /**
9343 * The prompt's text as the skill computed it (its text blocks, joined).
9344 */
9345 text: string;
9346 };
9347
9348 /**
9349 * What a `skill.prompt` hook returns: the text the model reads for that
9350 * skill.
9351 */
9352 export type SkillPromptResult = {
9353 text: string;
9354 };
9355
9356 /**
9357 * Options of `$.clock.sleep`.
9358 */
9359 export type SleepOptions = {
9360 /**
9361 * Aborting it rejects the sleep at once.
9362 */
9363 signal?: AbortSignal;
9364 };
9365
9366 /**
9367 * Options of `$.audio.speak`.
9368 */
9369 export type SpeakOptions = {
9370 /**
9371 * The system voice's exact name as the platform lists it (`Samantha`, or the
9372 * name of a voice you installed). Absent: the synthesizer's default voice.
9373 */
9374 voice?: string;
9375 };
9376
9377 /**
9378 * The argument of `$.audio.speak(text, options)` as the event carries it.
9379 */
9380 type SpeakRequest = SpeakOptions & {
9381 /**
9382 * What to say, as plain text, of at most 4096 characters.
9383 */
9384 text: string;
9385 };
9386
9387 /**
9388 * What `$.audio.speak` resolves with once the utterance has ended.
9389 */
9390 export type SpeakResult = {
9391 /**
9392 * Which synthesizer spoke: `system`, the platform's own (`say` on macOS).
9393 */
9394 via: 'system';
9395 };
9396
9397 /**
9398 * `next` in a `*` hook: the set of events is open at runtime, so `e` is
9399 * `unknown` until `next.is(pattern, e)` narrows it to events it knows.
9400 *
9401 * The callable is an overload per known event (OrderedOverloads), then
9402 * `(e: unknown) => Promise<unknown>` last: `next(e)` with `e` still unknown
9403 * resolves to `unknown`.
9404 */
9405 export type StarNext = OrderedOverloads<EventName> & {
9406 (e: unknown): Promise<unknown>;
9407 /**
9408 * Continues this dispatch at a tier, as Next's `to` (a managed hook's):
9409 * `e` still unknown resolves to `unknown`.
9410 */
9411 readonly to: (e: unknown, tier: TargetTier) => Promise<unknown>;
9412 readonly signal: AbortSignal;
9413 readonly is: <M extends Pattern>(pattern: M, e: unknown) => e is Frozen<Args<Selected<M>>>;
9414 readonly event: EventName;
9415 readonly origin: Origin;
9416 readonly trace: readonly TraceEntry<EventName, unknown, unknown>[];
9417 readonly budget: NextBudget;
9418 };
9419
9420 type StopFailureHookInput = BaseHookInput & {
9421 hook_event_name: 'StopFailure';
9422 error: SDKAssistantMessageError;
9423 error_details?: string;
9424 last_assistant_message?: string;
9425 };
9426
9427 type StopHookInput = BaseHookInput & {
9428 hook_event_name: 'Stop';
9429 stop_hook_active: boolean;
9430 /**
9431 * Text content of the last assistant message before stopping. Avoids the need to read and parse the transcript file.
9432 */
9433 last_assistant_message?: string;
9434 /**
9435 * In-flight background work (running/pending + backgrounded) registered in this session. Lets hooks distinguish "session is done" from "session is paused waiting for background work to wake it". Empty array when nothing is in flight.
9436 */
9437 background_tasks?: BackgroundTaskSummary[];
9438 /**
9439 * Session-scoped cron tasks (CronCreate, ScheduleWakeup, /loop) that will wake this session later. Empty array when none are scheduled.
9440 */
9441 session_crons?: SessionCronSummary[];
9442 };
9443
9444 /**
9445 * The hook a streaming event takes: `async function* ($, e, next) {}`,
9446 * yielding the event's chunks and returning its result.
9447 *
9448 * `return yield* next(e)` passes; `for await (const c of next(e)) yield
9449 * f(c)` transforms; yielding without `next` answers alone. A chunk yielded
9450 * stays: a hook that fails mid-stream is left, the rest from beneath it.
9451 *
9452 * @example on('turn.step', async function* ($, e, next) {
9453 * const r = yield* next(e); return { ...r } })
9454 */
9455 export type StreamHook<N extends StreamingEventName> = ($: EngineInterface, e: Frozen<Args<N>>, next: StreamNext<N>) => StreamHookBody<Chunk<N>, EventResult<N>>;
9456
9457 /**
9458 * What a hook on a streaming event evaluates to: the async generator an
9459 * `async function*` makes, yielding `C` and returning `R` or nothing.
9460 *
9461 * Returning nothing lets its last `next(e)`'s result stand. A plain
9462 * function is a type error here even when it returns `next(e)`: the hook
9463 * is the generator, not a function that hands one back.
9464 */
9465 export type StreamHookBody<C, R> = AsyncGenerator<C, R | void> & {
9466 /**
9467 * Absent on a generator; present on `next(e)`, which is not a hook body.
9468 */
9469 readonly result?: never;
9470 };
9471
9472 /**
9473 * The events that stream: their hooks are async generators, `next(e)` is
9474 * the stream of everything beneath, and the result is what it returns.
9475 *
9476 * `turn.step` alone: the model's response arrives in pieces, and a hook
9477 * that saw it whole could not change what had already been shown.
9478 */
9479 export type StreamingEventName = 'turn.step';
9480
9481 /**
9482 * The rest of the chain as a hook on a streaming event receives it: Next,
9483 * except that `next(e)` is the stream beneath (HookStream), not a promise.
9484 *
9485 * Each call opens a fresh stream; for `turn.step` each is a model request,
9486 * so a hook that calls it twice makes two. Not calling it yields the hook's
9487 * own chunks and returns its own result, and nothing beneath runs.
9488 *
9489 * @template S what `next.is(pattern, e)` narrows `e` to, as Next's
9490 * @see HookBudget `ms`: the hook's budget counts its own code, never a wait
9491 * at a `yield` or on this stream, however long the response beneath takes
9492 */
9493 export type StreamNext<N extends StreamingEventName = StreamingEventName, E = Args<N>, O = NextResult<N>, S extends {
9494 [K in N]?: unknown;
9495 } = {
9496 [K in N]: Args<K>;
9497 }> = Pick<Next<N, E, O, S>, 'signal' | 'is' | 'event' | 'origin' | 'budget'> & {
9498 (e: E): HookStream<Chunk<N>, O>;
9499 /**
9500 * Continues this dispatch at a tier, as Next's `to`: the stream beneath
9501 * with the links between this hook's tier and that one skipped.
9502 */
9503 readonly to: (e: E, tier: TargetTier) => HookStream<Chunk<N>, O>;
9504 /**
9505 * What settled beneath on the latest `next()` stream, as Next's; a
9506 * streaming link's entry counts the chunks it yielded up (`chunks`).
9507 */
9508 readonly trace: readonly TraceEntry<N, E, O>[];
9509 };
9510
9511 /**
9512 * The shape a `Box` and a `Text` share in a render tree: allowlisted props,
9513 * an optional `hover`, the group stamp a `hover.scope` earns, and children.
9514 *
9515 * `Tag` is which of the two; `Hover` is that element's hover props. Every
9516 * surface draws both: Ink's Box and Text on the terminal, a flex div and a
9517 * styled span on the desktop.
9518 */
9519 type StyledElement<Tag extends 'Box' | 'Text', Hover> = {
9520 type: Tag;
9521 /**
9522 * A Box's layout, position, spacing and border props and the `key` that
9523 * makes it a hover scope; a Text's colors and styles. Others are refused.
9524 */
9525 props?: Record<string, string | number | boolean>;
9526 /**
9527 * Style overrides the surface applies while the pointer is over the
9528 * nearest keyed Box, or over any member of the group `scope` names.
9529 *
9530 * Plain data, no hook. Nothing that moves a sibling: `borderStyle`
9531 * restyles a border the Box has, `display` only reveals a Box drawn
9532 * `"none"` (under a keyed Box or in a scope), an offset moves a placed Box.
9533 */
9534 hover?: Hover;
9535 /**
9536 * Whose group `hover.scope` names; absent without a `scope`.
9537 */
9538 group?: PluginStamp;
9539 /**
9540 * In order: a Box holds elements and strings (core wraps each string in
9541 * a Text); a Text holds strings and inline elements, never an engine node.
9542 */
9543 children?: RenderNode[];
9544 };
9545
9546 type SubagentStartHookInput = BaseHookInput & {
9547 hook_event_name: 'SubagentStart';
9548 agent_id: string;
9549 agent_type: string;
9550 };
9551
9552 type SubagentStopHookInput = BaseHookInput & {
9553 hook_event_name: 'SubagentStop';
9554 stop_hook_active: boolean;
9555 agent_id: string;
9556 agent_transcript_path: string;
9557 agent_type: string;
9558 /**
9559 * Text content of the last assistant message before stopping. Avoids the need to read and parse the transcript file.
9560 */
9561 last_assistant_message?: string;
9562 /**
9563 * In-flight background work (running/pending + backgrounded) registered in this session. Lets hooks distinguish "session is done" from "session is paused waiting for background work to wake it". Empty array when nothing is in flight.
9564 */
9565 background_tasks?: BackgroundTaskSummary[];
9566 /**
9567 * Session-scoped cron tasks (CronCreate, ScheduleWakeup, /loop) that will wake this session later. Empty array when none are scheduled.
9568 */
9569 session_crons?: SessionCronSummary[];
9570 };
9571
9572 /**
9573 * The props of `Svg`, the remote surfaces' vector leaf: the markup is the
9574 * element's data, as a string is a Text's, drawn isolated.
9575 *
9576 * A leaf: no children. The surface never lets the markup reach the page
9577 * (the engine bounds it; the desktop and the editor draw it as an image, or
9578 * in a sandboxed frame when `isInteractive`; the mobile app in a web view).
9579 */
9580 export type SvgProps = {
9581 /**
9582 * The SVG document, `<svg ...>...</svg>`, at most 131072 characters.
9583 */
9584 source: string;
9585 /**
9586 * What the drawing says, for a reader that cannot see it; required, since
9587 * a surface without the element draws nothing else of it.
9588 */
9589 alt: string;
9590 /**
9591 * CSS pixels; absent, the box takes the markup's own width up to the slot.
9592 */
9593 width?: number;
9594 /**
9595 * CSS pixels; absent, the markup's own height at the drawn width.
9596 */
9597 height?: number;
9598 /**
9599 * `true` draws the SVG in a script-less sandboxed frame so hover, CSS
9600 * `:hover`, SMIL animation and `<title>` tooltips work; absent, an image.
9601 *
9602 * It never enables script or event-handler attributes (the frame has no
9603 * allow-scripts and the scrub strips them); presses that other plugins
9604 * should observe go on an enclosing element.
9605 */
9606 isInteractive?: boolean;
9607 };
9608
9609 /**
9610 * The keys of `I` a matcher may select variants by: literal-valued in every
9611 * variant, and one literal per variant (IsDiscriminant).
9612 */
9613 type TagKeys<I> = {
9614 [K in MatcherKeys<I>]: IsLiteralValued<MatcherValueOf<I, K>> extends true ? IsDiscriminant<I, K> extends true ? K : never : never;
9615 }[MatcherKeys<I>];
9616
9617 /**
9618 * A tier `next.to(e, tier)` may name: one a floor can reach past a tier of
9619 * less authority to, so never `prepend` or `user`, which nothing skips to.
9620 */
9621 export type TargetTier = Exclude<Tier, 'prepend' | 'user'>;
9622
9623 type TaskCompletedHookInput = BaseHookInput & {
9624 hook_event_name: 'TaskCompleted';
9625 task_id: string;
9626 task_subject: string;
9627 task_description?: string;
9628 teammate_name?: string;
9629 /**
9630 * @deprecated Sessions have a single implicit team; this carries the session-derived team name and will be removed in a future release.
9631 */
9632 team_name?: string;
9633 };
9634
9635 type TaskCreatedHookInput = BaseHookInput & {
9636 hook_event_name: 'TaskCreated';
9637 task_id: string;
9638 task_subject: string;
9639 task_description?: string;
9640 teammate_name?: string;
9641 /**
9642 * @deprecated Sessions have a single implicit team; this carries the session-derived team name and will be removed in a future release.
9643 */
9644 team_name?: string;
9645 };
9646
9647 type TeammateIdleHookInput = BaseHookInput & {
9648 hook_event_name: 'TeammateIdle';
9649 teammate_name: string;
9650 /**
9651 * @deprecated Sessions have a single implicit team; this carries the session-derived team name and will be removed in a future release.
9652 */
9653 team_name: string;
9654 };
9655
9656 /**
9657 * The `Text` props a `hover` may override (its colors and styles, not its
9658 * wrapping) and `scope`, the hover group it joins; a `Button`'s label too.
9659 */
9660 export type TextHoverProps = {
9661 /**
9662 * Names a hover group of this plugin's: every element it draws with the
9663 * same `scope`, in any site on the surface, lights while any is hovered.
9664 *
9665 * Another plugin's elements under the same string are a different group.
9666 * One to 64 characters, no control characters; no keyed Box needed. On the
9667 * terminal a Text nested in a Text follows its group but cannot heat it.
9668 */
9669 scope?: string;
9670 color?: string;
9671 backgroundColor?: string;
9672 dimColor?: boolean;
9673 bold?: boolean;
9674 italic?: boolean;
9675 underline?: boolean;
9676 strikethrough?: boolean;
9677 inverse?: boolean;
9678 };
9679
9680 /**
9681 * The props of `Text`: the color and style props of Ink's Text a tree may
9682 * set. Colors are a theme key or a raw color.
9683 */
9684 export type TextProps = {
9685 /**
9686 * Style overrides applied by the surface while the pointer is over the
9687 * nearest enclosing keyed `Box`, or, given a `scope`, over its group.
9688 *
9689 * No hook runs and nothing crosses to the plugin. Refused outside a keyed
9690 * Box unless it names a `scope`.
9691 */
9692 hover?: TextHoverProps;
9693 color?: string;
9694 backgroundColor?: string;
9695 dimColor?: boolean;
9696 bold?: boolean;
9697 italic?: boolean;
9698 underline?: boolean;
9699 strikethrough?: boolean;
9700 inverse?: boolean;
9701 wrap?: 'wrap' | 'end' | 'middle' | 'truncate' | 'truncate-start' | 'truncate-middle' | 'truncate-end';
9702 };
9703
9704 /**
9705 * One of the chain's five tiers (TIERS), outermost first; on every
9706 * `next.trace` entry, and what `next.to(e, tier)` names.
9707 *
9708 * `prepend` and `append` are the managed plugins an administrator lists,
9709 * `user` everything a person installs, `builtin` the plugins bundled in the
9710 * binary, `core` the engine's innermost link.
9711 */
9712 export type Tier = (typeof TIERS)[number];
9713
9714 /**
9715 * The chain's five tiers, outermost first: earlier is outer is more
9716 * authority, and same-event hooks nest in this order and no other way.
9717 *
9718 * The managed plugins an administrator prepends, everything a person
9719 * installs, the managed plugins appended, the plugins bundled in the binary,
9720 * the engine's innermost link; a built-in's `$` calls still raise everywhere.
9721 */
9722 const TIERS: readonly ["prepend", "user", "append", "builtin", "core"];
9723
9724 /**
9725 * A pending timer from `$.clock.after` / `$.clock.every`.
9726 */
9727 export type Timer = {
9728 /**
9729 * Stops it; a stopped timer never fires again.
9730 */
9731 cancel: () => void;
9732 };
9733
9734 /**
9735 * A timer on `$.clock` (`after`, `every`): `fn` runs after `ms` milliseconds,
9736 * once or until `cancel()`.
9737 */
9738 export type TimerCall = (ms: number, fn: () => void) => Timer;
9739
9740 /**
9741 * Options of `$.ui.toast`.
9742 */
9743 export type ToastOptions = {
9744 /**
9745 * How long the line stays, in milliseconds; default 4000.
9746 */
9747 timeoutMs?: number;
9748 };
9749
9750 /**
9751 * `tool.call`'s input as the call takes it: `tool_use_id` and `agentId` may
9752 * ride along (a hook passing its event's input on) and are dropped.
9753 *
9754 * The run gets its own id and runs in the session's loop.
9755 */
9756 export type ToolCallArgs = ToolCallEnvelope extends infer I ? I extends ToolCallEnvelope ? Omit<I, 'tool_use_id'> & ToolCallReserved<I['tool']> : never : never;
9757
9758 /**
9759 * The envelope the two tool events share: the tool, the id of this call, and
9760 * the tool's arguments spread beside them (`e.command` for Bash).
9761 *
9762 * A union discriminated by `tool`: after `if (e.tool === "Bash")`, `e.command`
9763 * is a string and a rewrite is checked against Bash's schema. The `e` of
9764 * `classic.PreToolUse` exactly; `tool.call`'s adds the loop (ToolCallInput).
9765 */
9766 export type ToolCallEnvelope = BuiltinToolCallInput | McpToolCallInput;
9767
9768 /**
9769 * The input of `tool.call`: the tool, the id of this call, the tool's
9770 * arguments beside them (`e.command` for Bash), and `agentId` in a subagent.
9771 *
9772 * A union discriminated by `tool`: after `if (e.tool === "Bash")`, `e.command`
9773 * is a string and a rewrite is checked against Bash's schema. `tool`,
9774 * `tool_use_id` and `agentId` are reserved: a rewrite of any is refused.
9775 */
9776 export type ToolCallInput = ToolCallEnvelope & AgentLoop;
9777
9778 /**
9779 * `$.tool.call(input)`: resolves with `result` typed for the tool `input`
9780 * names (ToolCallResult), or loosely for an input that names none literally.
9781 */
9782 type ToolCallOverloads = {
9783 <T extends string>(input: ToolCallArgs & ToolNamed<T>): Promise<ToolCallResult<T>>;
9784 (input: ToolCallArgs): Promise<ToolCallResult>;
9785 };
9786
9787 /**
9788 * The keys `tool.call`'s input carries beside the tool's own arguments, none
9789 * of which the tool sees (the engine strips them before the tool runs).
9790 *
9791 * `consent` is the person's own words for the press that raised the call
9792 * (`The user pressed "1: Yes" on ...`): the run's context carries it as a human
9793 * turn, which the permission path reads as the user's request.
9794 */
9795 export type ToolCallReserved<T> = {
9796 tool: T;
9797 tool_use_id?: string;
9798 consent?: string;
9799 };
9800
9801 /**
9802 * What a `tool.call` hook returns and what `next(e)` and `$.tool.call(input)`
9803 * resolve to: the tool's result (`{ result, context? }`) or `{ deny }`.
9804 *
9805 * From core the result is `{ ref, result, text }` or, when the tool reported
9806 * an error, `{ ref, result, text, isError }`; `ref` names core's messages.
9807 *
9808 * @template Name the tool the call went to, typing `result` per built-in
9809 * tool (BuiltinToolResults) once `e.tool` is narrowed; else `unknown`
9810 */
9811 export type ToolCallResult<Name extends string = string> = {
9812 /**
9813 * Refuses the call: the model receives the text as an error result.
9814 * Absent when the call was answered.
9815 */
9816 deny: string;
9817 result?: undefined;
9818 context?: undefined;
9819 ref?: undefined;
9820 text?: undefined;
9821 isError?: undefined;
9822 } | {
9823 /**
9824 * The tool's output: from core the tool's record, typed per built-in
9825 * tool once `e.tool` and `isError` are narrowed; from a hook, its own.
9826 *
9827 * Core validates a hook's answer against the tool's output schema when
9828 * it has one, maps it for the model with the tool's own mapper, and
9829 * records it in the transcript as the tool's result. Absent on a deny.
9830 */
9831 result: ToolResultOf<Name>;
9832 /**
9833 * What the model reads after the tool's result and the user never
9834 * sees. From core, none.
9835 *
9836 * One reminder, as a PostToolUse hook's is, after the managed tier's
9837 * review; none on a plugin's own `$.tool.call`. Kept whole from `next`,
9838 * none empty, any length: past 100,000 (200,000 together) head + path.
9839 */
9840 context?: readonly string[];
9841 /**
9842 * Set by core on what `next(e)` resolves to: names the messages core
9843 * produced for the call (they stay on the host side).
9844 *
9845 * A hook that returns the object it got makes core use them verbatim.
9846 * Absent on a hook's own `{ result }` and on a deny.
9847 */
9848 ref?: number;
9849 /**
9850 * Set by core: the result as the model reads it (text blocks joined),
9851 * present whatever the tool, where `result`'s shape varies per tool.
9852 *
9853 * Absent on a hook's own `{ result }`.
9854 */
9855 text?: string;
9856 isError?: undefined;
9857 deny?: undefined;
9858 } | {
9859 /**
9860 * Set by core, present only when the tool reported an error (it threw,
9861 * was interrupted, or answered an error): `text` is what the model read.
9862 */
9863 isError: true;
9864 /**
9865 * What the transcript stored for the errored call: the error text, or
9866 * undefined when nothing was stored; never the tool's typed record.
9867 */
9868 result: unknown;
9869 /**
9870 * The error as the model reads it.
9871 */
9872 text?: string;
9873 /**
9874 * As on an answered result: names the messages core produced.
9875 */
9876 ref?: number;
9877 /**
9878 * As on an answered result.
9879 */
9880 context?: readonly string[];
9881 deny?: undefined;
9882 };
9883
9884 /**
9885 * `tool.check`'s input as `$.tool.check` takes it: the tool and its
9886 * arguments; `tool_use_id` is the engine's to set, never a query's.
9887 */
9888 type ToolCheckArgs = Pick<ToolCheckInput, 'tool' | 'input'>;
9889
9890 /**
9891 * The verdict of `tool.check`: run the tool, put it to the mode's decider
9892 * (the dialog, the auto-mode classifier, a headless host), or refuse it.
9893 */
9894 type ToolCheckDecision = 'allow' | 'ask' | 'deny';
9895
9896 /**
9897 * The input of `tool.check`: the tool, its arguments, and the call's id when
9898 * the engine is deciding a real call.
9899 *
9900 * All three are the question's identity and are pinned: a hook decides
9901 * about this call, it does not change it (`tool.call` rewrites a call).
9902 */
9903 type ToolCheckInput = {
9904 /**
9905 * As the model names it (`Bash`, `mcp__server__tool`); the key a matcher
9906 * narrows on.
9907 */
9908 tool: string;
9909 /**
9910 * The tool's arguments as the permission decision reads them
9911 * (`{ command }` for Bash, `{ file_path, ... }` for the file tools).
9912 */
9913 input: unknown;
9914 /**
9915 * The call being decided, on a real call only; absent on a query.
9916 *
9917 * `next.origin` names who raised it: `{ plugin: 'engine', tier: 'core' }`
9918 * for the model's own call, the plugin for its `$.tool.call` or its query.
9919 */
9920 tool_use_id?: string;
9921 };
9922
9923 /**
9924 * What a `tool.check` hook returns and what `next(e)` resolves to: the
9925 * verdict, why, and the settings rule behind it when one decided.
9926 *
9927 * From core, the engine's declarative decision for the session's mode and
9928 * rules. A hook may answer any verdict in either direction; the last word up
9929 * the chain is the decision.
9930 */
9931 type ToolCheckResult = {
9932 /**
9933 * `allow` runs the tool; `ask` puts it to the mode's decider; `deny`
9934 * refuses it, the reason the model's error.
9935 */
9936 decision: ToolCheckDecision;
9937 /**
9938 * Why, in a sentence: from core the rule or check that decided; from a
9939 * hook, what the model reads on a deny and the dialog shows on an ask.
9940 */
9941 reason?: string;
9942 /**
9943 * The settings rule that decided, as written (`Bash(git push:*)`).
9944 *
9945 * Absent for a mode or a tool's own check.
9946 */
9947 rule?: string;
9948 };
9949
9950 /**
9951 * Where a `tool.describe` answer places the tool: `true` behind ToolSearch
9952 * (its schema loads when the model asks for it), `false` in the prompt's list.
9953 *
9954 * Left out of an answer, the placement beneath stands.
9955 */
9956 export type ToolDeferral = boolean;
9957
9958 /**
9959 * The input of `tool.describe`: one tool's description, at the moment the
9960 * engine first renders the tool's schema for the model.
9961 */
9962 export type ToolDescribeInput = {
9963 /**
9964 * As the model sees the name (`Bash`, `mcp__server__tool`); the key a
9965 * matcher narrows on.
9966 */
9967 tool: string;
9968 /**
9969 * The tool's description as it computed it.
9970 */
9971 description: string;
9972 /**
9973 * Present, and true, when the engine lists the tool behind ToolSearch (its
9974 * schema loads when the model asks for it by name); absent for one listed.
9975 *
9976 * By the engine's rule an MCP server's tool, or one that asks to be,
9977 * unless a rule keeps it in front.
9978 */
9979 isDeferred?: true;
9980 /**
9981 * Who provides this tool: the plugin and its tier; `{ plugin: "engine",
9982 * tier: "core" }` for a built-in. Pinned: a rewrite is refused.
9983 *
9984 * A configured MCP server's tool is `mcp:<server>`, in `prepend` when the
9985 * policy settings source configures the server, else `user`.
9986 */
9987 provider: Origin;
9988 };
9989
9990 /**
9991 * What a `tool.describe` hook returns: the description the model sees for that
9992 * tool and, when the hook moves it, where the tool waits (ToolDeferral).
9993 *
9994 * `{ description }` alone, or `{ ...(await next(e)), description }`, changes
9995 * the text and keeps the engine's placement; `isDeferred: true` puts the tool
9996 * behind ToolSearch, `isDeferred: false` puts its schema in the prompt's list.
9997 */
9998 export type ToolDescribeResult = {
9999 description: string;
10000 isDeferred?: ToolDeferral;
10001 };
10002
10003 /**
10004 * `{ tool, tool_use_id, ...args }` as one flat object type, generic over
10005 * the tool name and its parsed arguments.
10006 */
10007 type ToolEnvelope<Name, Arguments> = {
10008 /**
10009 * The name of the tool being called (`Bash`, `mcp__<server>__<tool>`);
10010 * comparing it narrows `e`. Reserved: a rewrite of it is ignored by core.
10011 */
10012 tool: Name;
10013 /**
10014 * The tool_use block's id: the same at every event of the call and in
10015 * `$.ui.notice`. Reserved: a rewrite of it is ignored by core.
10016 */
10017 tool_use_id: string;
10018 } & Arguments;
10019
10020 /**
10021 * One tool call of a ToolGroup, as `ui.render` sees it under `calls`.
10022 */
10023 export type ToolGroupCall = {
10024 /**
10025 * The id `tool.call` carried for this call (`e.tool_use_id` there), so a
10026 * hook that saw the call finds its row in the group. Read-only.
10027 *
10028 * Absent on a desktop host that predates it.
10029 */
10030 tool_use_id?: string;
10031 /**
10032 * Which one the call ran (`Bash`, `Read`, `Grep`, ...).
10033 */
10034 tool: string;
10035 /**
10036 * The call's input, as the model sent it.
10037 */
10038 input: unknown;
10039 /**
10040 * True while the call is still running.
10041 */
10042 isRunning: boolean;
10043 /**
10044 * True when the call ended in an error.
10045 */
10046 isErrored: boolean;
10047 /**
10048 * True when an abort ended the call, as on `ToolUse`.
10049 */
10050 isInterrupted: boolean;
10051 /**
10052 * As on `ToolUse`; undefined while the call runs.
10053 */
10054 output?: unknown;
10055 };
10056
10057 /**
10058 * One tool as `$.tool.list()` returns it.
10059 */
10060 export type ToolInfo = {
10061 /**
10062 * What the model calls it by.
10063 */
10064 name: string;
10065 /**
10066 * What it does, in the tool's own words (its description; a first sentence at
10067 * most for MCP tools without one).
10068 */
10069 description: string;
10070 /**
10071 * True for an MCP server's tool.
10072 */
10073 mcp: boolean;
10074 };
10075
10076 /**
10077 * `{ tool, tool_use_id, ...args }` as one flat object type.
10078 *
10079 * The docs of `tool` and `tool_use_id` live on the `keyof` operand: a mapped
10080 * type takes its properties' docs from there.
10081 */
10082 export type ToolInputOf<Name extends string, Arguments> = {
10083 [K in keyof ToolEnvelope<Name, Arguments>]: ToolEnvelope<Name, Arguments>[K];
10084 };
10085
10086 /**
10087 * An input that names its tool as the literal `T`: what `next` and
10088 * `$.tool.call` read to type the call's result per tool (NextResultFor).
10089 */
10090 type ToolNamed<T extends string> = {
10091 /**
10092 * The name of the tool being called (`Bash`).
10093 */
10094 readonly tool: T;
10095 };
10096
10097 /**
10098 * The structured result of the tool named `Name`: its BuiltinToolResults
10099 * entry for a built-in tool, else `unknown`.
10100 *
10101 * Agent's is its entry or an AgentCallRecord (what a plugin-raised call
10102 * answers). `unknown` covers an MCP tool, a name the results table lacks
10103 * (one merged into the inputs table alone too), and `Name` left at `string`.
10104 */
10105 export type ToolResultOf<Name extends string> = string extends Name ? unknown : Name extends keyof BuiltinToolResults & string ? BuiltinToolResults[Name] | (Name extends 'Agent' ? AgentCallRecord : never) : unknown;
10106
10107 /**
10108 * One tool_result block of a user message.
10109 */
10110 export type ToolResultSummary = {
10111 /**
10112 * The id of the call this result answers (`tool.call`'s `e.tool_use_id`).
10113 */
10114 tool_use_id: string;
10115 /**
10116 * The result as the model read it (text blocks joined).
10117 */
10118 text: string;
10119 /**
10120 * True when the tool reported an error.
10121 */
10122 isError: boolean;
10123 /**
10124 * What the transcript stored for the call: the tool's record on an answered
10125 * one (`tool.call`'s `result`), the error text when `isError`.
10126 *
10127 * Absent when nothing was stored. Headless (`-p`), a tool may store the
10128 * record less its bulk (Bash blanks `stdout`); `text` is what the model
10129 * read either way.
10130 */
10131 result?: unknown;
10132 };
10133
10134 /**
10135 * What `$.tool.register` takes.
10136 */
10137 export type ToolSpec = {
10138 /**
10139 * The tool's short name (letters, digits, `_`, `-`; up to 64); the model
10140 * calls it as `mcp__<plugin>__<name>`.
10141 */
10142 name: string;
10143 /**
10144 * What the tool does, for the model (and the person where tools are
10145 * listed); an unpaired surrogate half in it is drawn as U+FFFD.
10146 */
10147 description: string;
10148 /**
10149 * A JSON schema object for the input (`{ type: "object", properties,
10150 * required }`); default `{ type: "object" }`.
10151 */
10152 inputSchema?: Record<string, unknown>;
10153 };
10154
10155 /**
10156 * One tool_use block of an assistant message, with its outcome once the
10157 * transcript holds the call's tool_result (paired by `tool_use_id`).
10158 */
10159 export type ToolUseSummary = {
10160 /**
10161 * The call's id, as `tool.call` carried it (`e.tool_use_id` there).
10162 */
10163 tool_use_id: string;
10164 /**
10165 * Which one was called (`Read`, `Bash`, `mcp__server__tool`), as
10166 * `tool.call` named it (`e.tool` there).
10167 */
10168 tool: string;
10169 /**
10170 * The arguments the model gave it.
10171 */
10172 input: Record<string, unknown>;
10173 /**
10174 * What the transcript stored for the call: the tool's record on an answered
10175 * one (`tool.call`'s `result`), the error text on a refused or errored one.
10176 *
10177 * Absent while the call is in flight or when nothing was stored. Headless
10178 * (`-p`), a tool may store the record less its bulk (Bash blanks `stdout`);
10179 * `text` holds either way.
10180 */
10181 result?: unknown;
10182 /**
10183 * The result as the model read it; absent while the call is in flight.
10184 */
10185 text?: string;
10186 /**
10187 * Present only when the tool reported an error, as on `tool.call`'s result.
10188 */
10189 isError?: true;
10190 };
10191
10192 /**
10193 * One settled run of a link beneath the caller, as `next.trace` lists it:
10194 * data, not a handle.
10195 *
10196 * `received` and `returned` are the live references where the hook runs
10197 * beside the chain; in a hooks module its own copies, as `e` is.
10198 */
10199 export type TraceEntry<N extends EventName = EventName, E = Args<N>, O = NextResult<N>> = {
10200 /**
10201 * The link's place in the chain, 0 the outermost.
10202 */
10203 readonly index: number;
10204 /**
10205 * The hook's plugin; `"engine"` for the engine's own core or bottom.
10206 */
10207 readonly plugin: string;
10208 /**
10209 * The link's tier (Tier); `"core"` for the engine's own core or bottom.
10210 */
10211 readonly tier: Tier;
10212 readonly event: N;
10213 readonly outcome: TraceOutcome;
10214 /**
10215 * Why the link was skipped without running, when it was: `bypassed by
10216 * <plugin>`, whose `next.to` went beneath this tier. Absent when it ran.
10217 */
10218 readonly reason?: string;
10219 /**
10220 * Its own wall time, its `next()` calls' time in flight taken out.
10221 *
10222 * The engine entry's is everything beneath the last hook: in a hooks
10223 * module, the host round trip.
10224 */
10225 readonly ms: number;
10226 /**
10227 * How many chunks the link yielded up, on a streaming event; absent on
10228 * every other. A link left mid-stream counts what it yielded before.
10229 */
10230 readonly chunks?: number;
10231 readonly received: E;
10232 /**
10233 * What the link settled on; undefined when it was skipped or rejected.
10234 */
10235 readonly returned: O | undefined;
10236 };
10237
10238 /**
10239 * What the chain decided for one link, as `next.trace` names it.
10240 *
10241 * `returned`: its result stood; `passed`: it returned, by reference, what its
10242 * last `next()` resolved to (a hooks module's hook answers with a copy of its
10243 * own, so it reads `returned`); `skipped`: it failed before `next`, or a
10244 * `next.to` above continued beneath its tier (`reason` says which), and
10245 * beneath ran in its place; `kept`: it failed after `next`, and that run's
10246 * result stands; `expired`: its budget ran out (what stands follows
10247 * skipped/kept); `caught`: it threw or its budget ran out, and its `.catch`
10248 * handler's result stands; `rejected`: the link rejected; the deepest such
10249 * entry is where the rejection came from, and the ones above it let it pass.
10250 */
10251 export type TraceOutcome = 'caught' | 'expired' | 'kept' | 'passed' | 'rejected' | 'returned' | 'skipped';
10252
10253 /**
10254 * What every `turn.complete` carries whatever its reason: the answer, the
10255 * duration, the interrupt flag, the turn's id, its loop and what it cost.
10256 */
10257 type TurnCompleteFields = {
10258 /**
10259 * The assistant's final visible text this turn ("" if none, e.g.
10260 * thinking-only).
10261 */
10262 answer: string;
10263 /**
10264 * Wall-clock length of the turn in milliseconds.
10265 */
10266 durationMs: number;
10267 /**
10268 * True when the turn ended by interruption (`reason === 'aborted'`).
10269 */
10270 isAborted: boolean;
10271 /**
10272 * The turn's id, the same one its `turn.start` and every `turn.step`
10273 * carried; a subagent's run raises no `turn.start`, its steps carry it.
10274 */
10275 turnId: string;
10276 /**
10277 * The loop the turn ran in: a subagent's id, as `$.agent.spawn` resolves
10278 * it, each run of its loop one turn; absent on the main loop.
10279 *
10280 * Pinned: a different value is refused, one left out is kept. Every hook
10281 * sees a subagent's turn, so a hook that spawns sees its children's turns
10282 * too and bounds itself.
10283 */
10284 agentId?: string;
10285 /**
10286 * What the turn cost: its responses' token counts summed and the model of
10287 * the last, read off the API responses the engine already holds.
10288 *
10289 * Absent when the turn got no response (interrupted before one, an API
10290 * error).
10291 */
10292 usage?: TurnUsage;
10293 };
10294
10295 /**
10296 * The input of `turn.complete`: the assistant's final message of a turn, at
10297 * the moment the turn ends (where the turn's duration is reported).
10298 *
10299 * `reason` says why it ended; `refusal` exists on a refusal alone.
10300 */
10301 export type TurnCompleteInput = TurnCompleteFields & (TurnCompleteRefused | TurnCompleteUnrefused);
10302
10303 /**
10304 * Why a turn ended: the model answered, the user interrupted it, the model
10305 * refused with no fallback model to retry on, or an API error ended it.
10306 */
10307 export type TurnCompleteReason = 'answer' | 'aborted' | 'refusal' | 'error';
10308
10309 /**
10310 * The end of a turn the model refused with no fallback model to retry on:
10311 * what the API said of the refusal rides along.
10312 */
10313 type TurnCompleteRefused = {
10314 reason: 'refusal';
10315 refusal: TurnRefusal;
10316 };
10317
10318 /**
10319 * What a `turn.complete` hook returns and what `next(e)` resolves to:
10320 * `{ text }`; a text other than a main-loop answer's is shown beneath it.
10321 *
10322 * Core fills `usage` from `e.usage` when the turn had one; a hook above reads
10323 * it, and one that answers its own may leave it out.
10324 */
10325 export type TurnCompleteResult = {
10326 text: string;
10327 usage?: TurnUsage;
10328 };
10329
10330 /**
10331 * The end of a turn that was not a refusal: answered, interrupted, or dead
10332 * on an API error (retries exhausted, the context limit), nothing more.
10333 */
10334 type TurnCompleteUnrefused = {
10335 reason: Exclude<TurnCompleteReason, 'refusal'>;
10336 };
10337
10338 /**
10339 * What the API said about a refusal that ended a turn: the classifier's
10340 * category and its explanation, each null when the API sent none.
10341 */
10342 export type TurnRefusal = {
10343 category: string | null;
10344 explanation: string | null;
10345 };
10346
10347 /**
10348 * The input of `turn.start`: the prompt a model turn begins with, after
10349 * `prompt.submit` settled the text and the UserPromptSubmit settings hooks ran.
10350 */
10351 export type TurnStartInput = {
10352 /**
10353 * The user's text as the turn proceeds with it ("" for a turn started without
10354 * a typed prompt, e.g. a continuation).
10355 */
10356 text: string;
10357 /**
10358 * The turn's id, minted here; the same one every `turn.step` and the
10359 * `turn.complete` of this turn carry.
10360 */
10361 turnId: string;
10362 };
10363
10364 /**
10365 * What a `turn.start` hook returns and what `next(e)` resolves to:
10366 * `{ turnId }`, echoed by core; a hook's own value does not change the turn.
10367 */
10368 export type TurnStartResult = {
10369 turnId: string;
10370 };
10371
10372 /**
10373 * One piece of a `turn.step` response as it streams through the chain:
10374 * what a hook's `next(e)` yields and what the hook yields up in turn.
10375 *
10376 * Text, thinking, a tool call's start and arguments, and the stop are
10377 * plain data a hook reads and rewrites; `engine` is the rest, passed on
10378 * unread. What leaves the outermost hook is what is shown and recorded.
10379 */
10380 export type TurnStepChunk = TurnStepTextChunk | TurnStepThinkingChunk | TurnStepToolChunk | TurnStepInputChunk | TurnStepStopChunk | TurnStepEngineChunk;
10381
10382 /**
10383 * An item of the engine's stream the other chunk kinds do not model (the
10384 * envelope, a block's start and end, a retry marker), opaque by `ref`.
10385 *
10386 * Pass it on where it came. A hook that yields its own response has none
10387 * to yield, and the engine supplies what the response needs around the
10388 * chunks it does yield.
10389 */
10390 export type TurnStepEngineChunk = {
10391 kind: 'engine';
10392 /**
10393 * The engine's handle on the item; opaque, this step's only.
10394 */
10395 ref: number;
10396 };
10397
10398 /**
10399 * The input of `turn.step`: one model request inside a turn, at the moment
10400 * the engine is about to send it; the bottom of the chain sends it.
10401 *
10402 * The transcript is not on it: `messageCount` says how many messages the
10403 * request carries, and `$.session.messages()` reads them. A hook rewrites
10404 * `model` or `effort` going down; the rest is pinned.
10405 */
10406 export type TurnStepInput = {
10407 /**
10408 * The turn this step belongs to (`turn.start`'s id; inside a subagent's
10409 * loop, the id the run's `turn.complete` will carry). Pinned.
10410 */
10411 turnId: string;
10412 /**
10413 * The step's position in the turn, from 0. Pinned.
10414 */
10415 index: number;
10416 /**
10417 * Which model the request names, as the engine resolved it for this step
10418 * (the session's, a fallback's). `next({ ...e, model })` names another.
10419 */
10420 model: string;
10421 /**
10422 * How hard the request asks the model to think: the session's setting or
10423 * the model's default, absent for a model without effort; rewritable.
10424 */
10425 effort?: 'low' | 'medium' | 'high' | 'xhigh' | 'max' | number;
10426 /**
10427 * How many messages the request carries (the conversation so far, the
10428 * turn's tool results included). Pinned: the messages are the engine's.
10429 */
10430 messageCount: number;
10431 /**
10432 * The loop the request is made in: a subagent's id, the `id`
10433 * `$.agent.list()` gives it and its `tool.call`s carry; absent on main.
10434 *
10435 * Pinned: a different value is refused, one left out is kept. A subagent
10436 * a hook spawned through `$.agent.spawn` steps past that hook, as its tool
10437 * calls do; every other hook sees its steps.
10438 */
10439 agentId?: string;
10440 };
10441
10442 /**
10443 * A piece of a tool call's arguments as they arrive: JSON text, partial,
10444 * for the `tool` chunk of the same `index`.
10445 *
10446 * The arguments the tool runs with are the pieces concatenated and parsed
10447 * once the block ends; pieces that do not parse leave the call without
10448 * arguments it can run on, as a model that wrote broken JSON would.
10449 */
10450 export type TurnStepInputChunk = ChunkRef & {
10451 kind: 'input';
10452 index: number;
10453 /**
10454 * The next piece of the arguments' JSON text.
10455 */
10456 json: string;
10457 };
10458
10459 /**
10460 * What a `turn.step` hook returns and what `next(e)` resolves to: the
10461 * model's response to the step's request, once its blocks are all in.
10462 *
10463 * From the bottom, the response the engine streamed; a hook's own value
10464 * changes what the hooks above it read, never what the engine streamed.
10465 */
10466 export type TurnStepResult = {
10467 /**
10468 * The turn this step belongs to, as received.
10469 */
10470 turnId: string;
10471 /**
10472 * The step's position in the turn, as received.
10473 */
10474 index: number;
10475 /**
10476 * The visible text of the response ("" when it only called tools, only
10477 * thought, or no request was made).
10478 */
10479 answer: string;
10480 /**
10481 * The tool calls the response made, in order; empty for a text-only step.
10482 */
10483 toolUses: readonly TurnStepToolUse[];
10484 /**
10485 * Why the model stopped; null when no response arrived (the request
10486 * failed or was interrupted before a message, or no request was made).
10487 */
10488 stopReason: TurnStopReason;
10489 /**
10490 * What the request cost as the API reported it, and the model that
10491 * answered; null when no response arrived or it carried no usage.
10492 */
10493 usage: TurnUsage | null;
10494 };
10495
10496 /**
10497 * The response is whole: why the model stopped and what the request cost,
10498 * as the step's result will carry them. The last chunk of one response.
10499 */
10500 export type TurnStepStopChunk = ChunkRef & {
10501 kind: 'stop';
10502 stopReason: TurnStopReason;
10503 usage: TurnUsage | null;
10504 };
10505
10506 /**
10507 * A piece of the response's visible text as it arrives, in block `index`.
10508 *
10509 * The text the person watches stream and the text the transcript records
10510 * are both the concatenation of these, in order, per block: a hook that
10511 * rewrites them rewrites both, one that drops them all drops the block.
10512 */
10513 export type TurnStepTextChunk = ChunkRef & {
10514 kind: 'text';
10515 /**
10516 * The content block the piece belongs to, from 0 within one response.
10517 */
10518 index: number;
10519 text: string;
10520 };
10521
10522 /**
10523 * A piece of the model's thinking as it arrives, in block `index`: what
10524 * the person sees of it live.
10525 *
10526 * The thinking block the transcript keeps is the one the model signed, as
10527 * the engine received it; thinking a hook streams without the engine's
10528 * block behind it is shown and not recorded.
10529 */
10530 export type TurnStepThinkingChunk = ChunkRef & {
10531 kind: 'thinking';
10532 index: number;
10533 text: string;
10534 };
10535
10536 /**
10537 * The model begins a tool call in block `index`: the tool's name and the
10538 * call's id; its arguments follow as `input` chunks of the same index.
10539 *
10540 * A block whose `tool` chunk never comes out of the chain is no tool call:
10541 * the engine records none and runs none.
10542 */
10543 export type TurnStepToolChunk = ChunkRef & {
10544 kind: 'tool';
10545 index: number;
10546 /**
10547 * The call's `tool_use_id`.
10548 */
10549 id: string;
10550 /**
10551 * The tool's name (`Read`, `Bash`, `mcp__server__tool`).
10552 */
10553 name: string;
10554 };
10555
10556 /**
10557 * One tool call the model asked for in a step: the tool's name and its
10558 * arguments.
10559 */
10560 export type TurnStepToolUse = {
10561 /**
10562 * The tool's name (`Read`, `Bash`, `mcp__server__tool`).
10563 */
10564 name: string;
10565 /**
10566 * The arguments the model gave it, as the tool schema shapes them.
10567 */
10568 input: unknown;
10569 };
10570
10571 /**
10572 * Why the model stopped, as a `turn.step` result and its stop chunk carry
10573 * it: one of the API's stop reasons, or null when no response arrived.
10574 */
10575 type TurnStopReason = 'end_turn' | 'max_tokens' | 'stop_sequence' | 'tool_use' | 'pause_turn' | 'compaction' | 'refusal' | 'model_context_window_exceeded' | null;
10576
10577 /**
10578 * What a model turn, or one response inside it, cost as the API reported it:
10579 * the four token counts (`$.model.fork`'s usage shape) and the model's id.
10580 *
10581 * A turn's counts are its responses' summed; its model is the last response's.
10582 */
10583 export type TurnUsage = ModelForkUsage & {
10584 /**
10585 * Which model answered, by the id the API reports.
10586 */
10587 model: string;
10588 };
10589
10590 /**
10591 * What a plugin's `$.ui.blit(args)` takes: a Raster's next `cells`
10592 * (RasterBlitArgs) or a keyed Image's next `source` (ImageBlitArgs).
10593 *
10594 * Either names an element this plugin's own `ui.render` hook drew, still
10595 * mounted; a hook above may rewrite the cells or the source, never the
10596 * address or which kind it is.
10597 */
10598 export type UiBlitArgs = RasterBlitArgs | ImageBlitArgs;
10599
10600 /**
10601 * What `$.ui.blit` resolves to and what a `ui.blit` hook's `{ value }`
10602 * holds: `{}` once the cells or source are the next frame's, or why not.
10603 */
10604 export type UiBlitResult = {
10605 /**
10606 * Absent when the cells or source were taken; else why not.
10607 *
10608 * Nothing of this plugin's is mounted there (another plugin's reads so),
10609 * the size is not the mounted one, the cells or source are bad, or the
10610 * Image draws its `alt` there, the reason spelled out for a fallback.
10611 */
10612 deny?: string;
10613 };
10614
10615 /**
10616 * What a plugin's `$.ui.focus(args)` takes: one of its own elements, by the
10617 * `key` it drew it under, in one of its sites that holds the keyboard now.
10618 *
10619 * The engine resolves it to that site's ring and raises `ui.focus` under
10620 * the plugin's origin; the keyboard is the person's to give, so a site that
10621 * does not hold it, or holds it on another plugin's element, is `{ deny }`.
10622 */
10623 export type UiFocusArgs = {
10624 /**
10625 * The site, by the `requestId` this plugin draws it under: one of its
10626 * panes' ids, or the band's.
10627 */
10628 requestId: string;
10629 /**
10630 * The element's `key`: a `Button`, `Input` or `Select` this plugin drew in
10631 * that site; of several under one key, the first in document order.
10632 */
10633 key: string;
10634 };
10635
10636 /**
10637 * The render components whose site keeps a focus ring: a pane's body and
10638 * the band above the prompt, each a ring over the elements hooks drew there.
10639 */
10640 export type UiFocusComponent = 'Pane' | 'AbovePrompt';
10641
10642 /**
10643 * The input of `ui.focus`: a site's focus ring about to move onto one of the
10644 * elements a hook drew in it (a `Button`, `Input` or `Select`), or off them.
10645 *
10646 * Every key but `element` is the engine's word, pinned: `next(e)` passes
10647 * them on, a rewrite that leaves one out keeps it, one that changes it fails
10648 * the hook. `element` is the hook's to rewrite; the ring has not moved yet.
10649 */
10650 export type UiFocusInput = {
10651 /**
10652 * Which site: a `Pane` body or the `AbovePrompt` band.
10653 */
10654 component: UiFocusComponent;
10655 /**
10656 * The instance the `ui.render` hook drawing the site sees: the pane's id,
10657 * or the band's one id.
10658 */
10659 requestId: string;
10660 /**
10661 * Whose `ui.render` hook drew the element taking the ring; absent with
10662 * `element`. Read-only.
10663 */
10664 plugin?: string;
10665 /**
10666 * The `key` of the element taking the ring, as `ui.press` names it; absent
10667 * for one of the engine's own stops (a pane's close mark or another's tab).
10668 *
10669 * `next({ ...e, element })` lands it on another element `plugin` drew in
10670 * the site instead; one not drawn there is core's `{ deny }`.
10671 */
10672 element?: string;
10673 /**
10674 * Who moves it (UiFocusOrigin), set by the engine where the move starts.
10675 */
10676 origin: UiFocusOrigin;
10677 };
10678
10679 /**
10680 * Who moves the ring at `ui.focus`, as the engine stamps it where the move
10681 * starts; a closed set a matcher narrows on.
10682 *
10683 * `next(e)` passes it on as received; no hook sets one.
10684 */
10685 export type UiFocusOrigin = {
10686 /**
10687 * The person, by Tab, the arrows or a click while the site holds the
10688 * keyboard.
10689 */
10690 kind: 'person';
10691 } | {
10692 /**
10693 * A plugin's `$.ui.focus`, or its `autoFocus` element taking the ring
10694 * as the site takes the keyboard.
10695 */
10696 kind: 'plugin';
10697 /**
10698 * The focusing plugin's name.
10699 */
10700 name: string;
10701 };
10702
10703 /**
10704 * What a `ui.focus` hook returns, what `next(e)` resolves to, and what
10705 * `$.ui.focus` hands back: `{}` once the ring moved, or why it did not.
10706 *
10707 * As `ui.scroll` spells it.
10708 */
10709 export type UiFocusResult = {
10710 /**
10711 * Absent when the ring is where the chain left it; else why nothing
10712 * moved.
10713 *
10714 * A hook kept the ring (no `next`); the site is not this plugin's, does
10715 * not hold the keyboard, or another plugin's element holds it; no element
10716 * of `plugin` is drawn under `element`; another move landed first.
10717 */
10718 deny?: string;
10719 };
10720
10721 /**
10722 * The argument of `ui.input`: a change of, or a submit from, an `Input` a
10723 * render hook drew. Flat and frozen like every event's.
10724 *
10725 * Another plugin addresses one field by matcher:
10726 * `on("ui.input", { plugin: "roster", element: "reply" }, ...)`.
10727 */
10728 export type UiInputArgument = {
10729 /**
10730 * Whose `ui.render` hook drew the element.
10731 */
10732 plugin: string;
10733 /**
10734 * The `key` the hook gave its `Input`: its address, what a matcher names.
10735 */
10736 element: string;
10737 /**
10738 * The render component the element was drawn in (`AbovePrompt`,
10739 * `ToolUse`, ...).
10740 */
10741 component: RenderComponent;
10742 /**
10743 * The instance the element was drawn in: the `requestId` the `ui.render`
10744 * hook that drew it saw (a tool row's tool_use_id, a pane's id).
10745 */
10746 requestId: string;
10747 /**
10748 * Where the typing came from, one literal per member, so
10749 * `if (e.surface === "terminal")` narrows `e`.
10750 */
10751 surface: RenderSurface;
10752 /**
10753 * `change` after every edit of the text; `submit` on Enter.
10754 */
10755 kind: 'change' | 'submit';
10756 /**
10757 * The field's whole text at that moment; a hook above may rewrite it for
10758 * the plugins beneath and the element's own handler.
10759 */
10760 value: string;
10761 };
10762
10763 /**
10764 * What a `ui.input` hook returns and what `next(e)` resolves to.
10765 *
10766 * Beneath every hook, core runs the element's `onInput` (kind `change`) or
10767 * `onSubmit` (kind `submit`) closure in its plugin's environment with the
10768 * value as the chain left it, and answers `{ element, value }`.
10769 */
10770 export type UiInputResult = {
10771 /**
10772 * Which handler the input reached, by its Input's `key`.
10773 */
10774 element: string;
10775 /**
10776 * The text the handler received.
10777 */
10778 value: string;
10779 };
10780
10781 /**
10782 * Options of `$.ui.log`.
10783 */
10784 export type UiLogOptions = {
10785 /**
10786 * Where the line goes (UiLogSink); `transcript` when left out. A hook on
10787 * `ui.log` reads it as `e.to` and may send the line elsewhere.
10788 */
10789 to?: UiLogSink;
10790 };
10791
10792 /**
10793 * Where a `$.ui.log` line goes: `transcript`, a dim row of its own (and the
10794 * debug log, as every line); `debug`, the debug log alone, nothing on screen.
10795 *
10796 * The debug log is `claude --debug` or the `--debug-file`; either way the
10797 * line is led by the plugin's name.
10798 */
10799 export type UiLogSink = 'transcript' | 'debug';
10800
10801 /**
10802 * The argument of `ui.message`: what a `Client` instance's surface module
10803 * posted (`surface.post(data)`), addressed by where the instance is drawn.
10804 *
10805 * The origin is `client`: code sent it, on nobody's behalf, so `data` is
10806 * input to validate, not a fact. Everything but `data` is the engine's word
10807 * and a hook may not rewrite it.
10808 */
10809 export type UiMessageArgument = {
10810 /**
10811 * Where the instance is drawn: `terminal`, or `desktop` once it has them.
10812 */
10813 surface: RenderSurface;
10814 /**
10815 * The render component the `Client` was drawn in (`AbovePrompt`, `Pane`,
10816 * ...).
10817 */
10818 component: RenderComponent;
10819 /**
10820 * The engine's id for the drawing the `Client` sits in: the `requestId` the
10821 * `ui.render` hook that drew it saw.
10822 */
10823 requestId: string;
10824 /**
10825 * The `Client`'s `key`: which instance posted.
10826 */
10827 element: string;
10828 /**
10829 * The `Client`'s `module`: which surface module runs there, as its path
10830 * under the plugin's folder.
10831 */
10832 module: string;
10833 /**
10834 * What the instance posted: plain data (JsonValue), bounded as a tree's
10835 * text is; typed `unknown` since it came from code. Rewritable.
10836 */
10837 data: unknown;
10838 };
10839
10840 /**
10841 * What a `ui.message` hook returns and what `next(e)` resolves to.
10842 *
10843 * Core answers `{}`: the message was heard and nothing changes. A hook that
10844 * answers `{ props }` hands the posting instance its next props directly, its
10845 * local state kept, with no `ui.render` run; the plugin's next redraw hands
10846 * props again as usual.
10847 */
10848 export type UiMessageResult = {
10849 /**
10850 * The instance's next props, plain data (JsonValue) bounded as a `Client`'s
10851 * props are; absent leaves the instance's props as they were.
10852 */
10853 props?: unknown;
10854 };
10855
10856 /**
10857 * One of this plugin's open panes as `$.ui.panes()` lists it: the pane's
10858 * id and title, and where it stands with the person right now.
10859 *
10860 * Where a surface seated it (`dock` or `inline`) is that surface's to say, on
10861 * the `Pane` render props (`placement`); a session may draw on several.
10862 */
10863 export type UiPane = {
10864 /**
10865 * What `$.ui.open({ id })` named it, which `$.ui.close` and its
10866 * `ui.render` `requestId` name too.
10867 */
10868 id: string;
10869 /**
10870 * Its tab's label: the `title` of its latest open, or the id.
10871 */
10872 title: string;
10873 /**
10874 * True for the one pane the surface shows; the rest are tabs behind it.
10875 */
10876 isShown: boolean;
10877 /**
10878 * True while the person has given it the keyboard.
10879 */
10880 isFocused: boolean;
10881 /**
10882 * False while it waits undrawn: opened unasked on a terminal too narrow
10883 * for an unrequested pane, until an open the width or the person admits.
10884 */
10885 isPlaced: boolean;
10886 };
10887
10888 /**
10889 * The argument of `ui.press`: a press on a `Button` a render hook drew, or
10890 * on an answered `Markdown` link. Flat and frozen like every event's.
10891 *
10892 * Another plugin addresses one button by matcher:
10893 * `on("ui.press", { plugin: "explainer", element: "explain" }, ...)`.
10894 */
10895 export type UiPressArgument = {
10896 /**
10897 * Whose `ui.render` hook drew the element.
10898 */
10899 plugin: string;
10900 /**
10901 * The `key` the hook gave its `Button` or `Markdown`: its address, what a
10902 * matcher names.
10903 */
10904 element: string;
10905 /**
10906 * The render component the element was drawn in (`ToolUse`,
10907 * `AssistantMessage`, ...).
10908 */
10909 component: RenderComponent;
10910 /**
10911 * The instance the element was drawn in: the `requestId` the `ui.render`
10912 * hook that drew it saw (a tool row's tool_use_id, a pane's id).
10913 */
10914 requestId: string;
10915 /**
10916 * Where the press came from, one literal per member, so
10917 * `if (e.surface === "terminal")` narrows `e`.
10918 */
10919 surface: RenderSurface;
10920 /**
10921 * Where the press landed, when the element is a `Markdown` whose links
10922 * its plugin answers (`onLinkPress`); absent for a `Button`.
10923 *
10924 * A hook may rewrite its `href` for the hooks and the closure beneath; one
10925 * that adds or drops it fails, and the press goes on beneath it.
10926 */
10927 link?: PressedLink;
10928 };
10929
10930 /**
10931 * What a `ui.press` hook returns and what `next(e)` resolves to.
10932 *
10933 * Beneath every hook, core runs the element's `onPress` closure in its
10934 * plugin's environment with `e` as the chain left it, and answers
10935 * `{ element }`.
10936 */
10937 export type UiPressResult = {
10938 /**
10939 * Which handler the press reached, by its Button's `key`.
10940 */
10941 element: string;
10942 };
10943
10944 /**
10945 * What a plugin's `$.ui.scroll(args)` takes: what to bring into view, in
10946 * which of its sites, and where in the window it lands.
10947 *
10948 * The engine resolves it to a window and an offset and raises `ui.scroll`
10949 * under the plugin's origin; a transcript row is the person's to move, so
10950 * it is revealed only while the plugin answers the person's own input.
10951 */
10952 export type UiScrollArgs = {
10953 /**
10954 * What to reveal (UiScrollTarget): a render instance by `requestId`, one
10955 * of this plugin's elements by `key`, or the `start` or `end` of `in`.
10956 */
10957 to: UiScrollTarget;
10958 /**
10959 * The site to scroll, by the `requestId` this plugin draws it under: one
10960 * of its panes' ids, or the band's.
10961 *
10962 * Required with `start` and `end`; with `{ key }` it picks the site when
10963 * the key is drawn in several.
10964 */
10965 in?: string;
10966 /**
10967 * Where the target lands in the window (UiScrollBlock); `nearest` when
10968 * left out, so a row already showing does not move.
10969 */
10970 block?: UiScrollBlock;
10971 };
10972
10973 /**
10974 * Where in its scrollable a revealed row lands, as the DOM's
10975 * `scrollIntoView({ block })` names it.
10976 *
10977 * `nearest` moves the least that shows the row whole, and not at all when
10978 * it already shows; `start`, `center` and `end` put its top, middle or
10979 * bottom at that edge of the window. A row taller than the window shows
10980 * its top.
10981 */
10982 export type UiScrollBlock = 'start' | 'center' | 'end' | 'nearest';
10983
10984 /**
10985 * The render components whose site the engine scrolls: a pane's body and
10986 * the band above the prompt, each a window over the tree a hook drew there.
10987 */
10988 export type UiScrollComponent = 'Pane' | 'AbovePrompt';
10989
10990 /**
10991 * The input of `ui.scroll`: a site's window asked to move over the tree a
10992 * hook drew in it (a pane's body, the band above the prompt).
10993 *
10994 * Every key but `offset` is the engine's word, pinned: `next(e)` passes
10995 * them on, a rewrite that leaves one out keeps it, one that changes or adds
10996 * one fails the hook. `offset` is the hook's to rewrite; nothing moved yet.
10997 */
10998 export type UiScrollInput = {
10999 /**
11000 * Which site: a `Pane` body or the `AbovePrompt` band.
11001 */
11002 component: UiScrollComponent;
11003 /**
11004 * The instance the `ui.render` hook drawing the site sees: the pane's id,
11005 * or the band's one id.
11006 */
11007 requestId: string;
11008 /**
11009 * The first row the window is to show, 0 at the top: the row it shows
11010 * now plus `by`, clamped to the tree, or where `$.ui.scroll` resolved.
11011 *
11012 * `next({ ...e, offset })` moves it elsewhere (a clamp); one past the
11013 * tree's end lands at the end. Nothing has moved when a hook reads it.
11014 */
11015 offset: number;
11016 /**
11017 * The rows the move asks for, signed, negative toward the top; what the
11018 * DOM's `scrollBy` takes. Read-only.
11019 *
11020 * The person's: a wheel tick `-1` or `1` at rest, more as ticks crowd or
11021 * summed in flight; an arrow a row while the engine has rows to scroll, a
11022 * page key `bodyRows`, Home and End `contentRows`. A plugin's: its distance.
11023 */
11024 by: number;
11025 /**
11026 * How many rows of the tree the window shows at once, as drawn now.
11027 * Read-only.
11028 */
11029 bodyRows: number;
11030 /**
11031 * How many rows the tree has, as drawn now: the window's last offset is
11032 * `contentRows - bodyRows`, none when the tree fits. Read-only.
11033 */
11034 contentRows: number;
11035 /**
11036 * Who moves it (UiScrollOrigin), set by the engine where the move starts.
11037 */
11038 origin: UiScrollOrigin;
11039 /**
11040 * The body cell the person's wheel or trackpad was over (UiScrollPointer);
11041 * absent for the scroll keys and for `$.ui.scroll`. Read-only.
11042 *
11043 * Moves summed while a dispatch was in flight carry the latest one's. A
11044 * hook pinning a list over rows it scrolls itself tells a tick over the
11045 * list (`row` among its list's rows) from one over the body.
11046 */
11047 pointer?: UiScrollPointer;
11048 };
11049
11050 /**
11051 * Who moves the window at `ui.scroll`, as the engine stamps it where the
11052 * move starts; a closed set a matcher narrows on.
11053 *
11054 * `next(e)` passes it on as received; no hook sets one.
11055 */
11056 export type UiScrollOrigin = {
11057 /**
11058 * The person, by the wheel over the site or its scroll keys while
11059 * they hold it.
11060 */
11061 kind: 'person';
11062 } | {
11063 /**
11064 * A plugin's `$.ui.scroll`.
11065 */
11066 kind: 'plugin';
11067 /**
11068 * The scrolling plugin's name.
11069 */
11070 name: string;
11071 };
11072
11073 /**
11074 * The cell the pointer was over when the person's wheel raised `ui.scroll`,
11075 * in the site's body: the box its `ui.render` hook draws into, as painted.
11076 *
11077 * The DOM's `clientY - body.top` in cells, the window's offset not added: a
11078 * hook drawing its own window reads `row` as its tree's row, one the engine
11079 * scrolls adds the `scroll.offset` it drew with. The frame lies outside.
11080 */
11081 export type UiScrollPointer = {
11082 /**
11083 * 0 at the body's left edge, as `bodyColumns` counts them; negative, or
11084 * `bodyColumns` and past, over an inline pane's side borders.
11085 */
11086 column: number;
11087 /**
11088 * 0 at the body's first showing row, as `bodyRows` counts them.
11089 *
11090 * The tree's row is `scroll.offset + row` under the engine's window, and
11091 * `row` itself under a hook's own (offset 0); negative over a pane's top
11092 * border or tab row, `bodyRows` or more over its bottom border.
11093 */
11094 row: number;
11095 };
11096
11097 /**
11098 * What a `ui.scroll` hook returns, what `next(e)` resolves to, and what
11099 * `$.ui.scroll` hands back: `{}` once the window moved, or why it did not.
11100 */
11101 export type UiScrollResult = {
11102 /**
11103 * Absent when the window is where the chain left it; else why nothing
11104 * moved, as `tool.call` and `config.set` spell a refusal.
11105 *
11106 * A hook kept the window (no `next`); the target is not this plugin's;
11107 * another move landed first (`the window moved meanwhile`); a transcript
11108 * row is not the person's ask (`not person-initiated`) or none scrolls.
11109 */
11110 deny?: string;
11111 };
11112
11113 /**
11114 * What `$.ui.scroll` brings into view: never a row number, always a thing
11115 * drawn somewhere.
11116 *
11117 * `{ requestId }` names a render instance by the id its `ui.render` hook
11118 * saw (a transcript message's, a tool row's tool_use_id), scrolled into view
11119 * inside whatever scrolls it. `{ key }` names an element this plugin drew
11120 * with that `key`, inside the site it drew it in. `start` and `end` are the
11121 * top and bottom of the site `in` names; `end` keeps up with a tree that
11122 * grows (rows the plugin just added count) until anything next moves it.
11123 */
11124 export type UiScrollTarget = {
11125 requestId: string;
11126 } | {
11127 key: string;
11128 } | 'start' | 'end';
11129
11130 /**
11131 * The argument of `ui.select`: a pick from a `Select` a render hook drew.
11132 * Flat and frozen like every event's.
11133 *
11134 * Another plugin addresses one picker by matcher:
11135 * `on("ui.select", { plugin: "roster", element: "peer" }, ...)`.
11136 */
11137 export type UiSelectArgument = {
11138 /**
11139 * Whose `ui.render` hook drew the element.
11140 */
11141 plugin: string;
11142 /**
11143 * The `key` the hook gave its `Select`: its address, what a matcher names.
11144 */
11145 element: string;
11146 /**
11147 * The render component the element was drawn in (`AbovePrompt`,
11148 * `ToolUse`, ...).
11149 */
11150 component: RenderComponent;
11151 /**
11152 * The instance the element was drawn in: the `requestId` the `ui.render`
11153 * hook that drew it saw (a tool row's tool_use_id, a pane's id).
11154 */
11155 requestId: string;
11156 /**
11157 * Where the pick came from, one literal per member, so
11158 * `if (e.surface === "terminal")` narrows `e`.
11159 */
11160 surface: RenderSurface;
11161 /**
11162 * The picked option's value; a hook above may rewrite it for the plugins
11163 * beneath and the element's own handler.
11164 */
11165 value: string;
11166 };
11167
11168 /**
11169 * What a `ui.select` hook returns and what `next(e)` resolves to.
11170 *
11171 * Beneath every hook, core runs the element's `onSelect` closure in its
11172 * plugin's environment with the value as the chain left it, and answers
11173 * `{ element, value }`.
11174 */
11175 export type UiSelectResult = {
11176 /**
11177 * Which handler the pick reached, by its Select's `key`.
11178 */
11179 element: string;
11180 /**
11181 * What the handler received: the option's value as the chain left it.
11182 */
11183 value: string;
11184 };
11185
11186 /**
11187 * The intersection of a union's members (`A | B` to `A & B`), by inferring
11188 * one parameter type from the contravariant positions.
11189 */
11190 type UnionToIntersection<U> = (U extends unknown ? (member: U) => void : never) extends (member: infer I) => void ? I : never;
11191
11192 /**
11193 * One unit of what `$.session.usage()` answers, by its key there: the
11194 * context window's fill, the rate-limit windows, the session's cost.
11195 */
11196 export type UsageUnit = 'context' | 'rateLimits' | 'cost';
11197
11198 /**
11199 * Who sent the message a `UserMessage` row carries, when someone other than
11200 * the person did: another agent, a teammate, another session, a channel.
11201 *
11202 * Read-only. The sender's words are not the person's: the model reads them
11203 * framed as that sender's whatever a hook draws for the row.
11204 */
11205 type UserMessageFrom = {
11206 /**
11207 * The sender's name as the row shows it: a subagent's name, else its type;
11208 * a teammate's; another session's title; a channel's sender, else server.
11209 *
11210 * One printable line: control and format characters stripped, whitespace
11211 * collapsed, length bounded. Absent (no `from`) for a teammate block whose
11212 * frames name different senders or none.
11213 */
11214 name: string;
11215 };
11216
11217 /**
11218 * The background task a `UserMessage` notification row reports on: a
11219 * subagent, a background shell, a workflow, a remote agent, a monitor.
11220 *
11221 * Every field is the notification's own, so the row names the same task on
11222 * every draw, a resumed session's included; the session's record of it
11223 * (name, type, description) is `$.agent.list()`'s, by `id`. Each string is
11224 * one printable line, bounded (a forged envelope's bytes capped). Read-only.
11225 */
11226 type UserMessageTask = {
11227 /**
11228 * The task's id as the notification names it; for a subagent, the
11229 * `agentId` its `turn.complete` carried and `$.agent.list()` keys.
11230 */
11231 id?: string;
11232 /**
11233 * How the task ended: `completed`, `failed` or `killed` (stopped, by the
11234 * person or by Claude), or another word its producer wrote.
11235 */
11236 status?: string;
11237 /**
11238 * What kind of task, when the notification says (`remote_agent`, an
11239 * artifact watch's kind); a subagent's and a shell's leave it out.
11240 */
11241 type?: string;
11242 /**
11243 * Which call started the task: its `tool_use_id`, when the notification
11244 * carries it.
11245 */
11246 toolUseId?: string;
11247 /**
11248 * How long the task ran, in milliseconds, when the notification says; the
11249 * row draws it after the summary (`2m 2s`).
11250 */
11251 durationMs?: number;
11252 };
11253
11254 type UserPromptExpansionHookInput = BaseHookInput & {
11255 hook_event_name: 'UserPromptExpansion';
11256 expansion_type: 'slash_command' | 'mcp_prompt';
11257 command_name: string;
11258 command_args: string;
11259 command_source?: string;
11260 prompt: string;
11261 };
11262
11263 type UserPromptSubmitHookInput = BaseHookInput & {
11264 hook_event_name: 'UserPromptSubmit';
11265 prompt: string;
11266 /**
11267 * Who authored/injected the prompt: `user` = submitted from the interactive composer, `sdk` = non-interactive entrypoint (`-p` / Agent SDK), `loop_wakeup` = dynamic /loop wakeup, `schedule_wakeup` = scheduled-task fire (CronCreate/routine), `system` = other machine-injected turns (peer/channel messages, task notifications, auto-continuation), `poll_event` = the poll-event channel enqueue-time pass (the hook fires when the host submits an event, before its delivery ack exists - a blocking verdict rejects the event). Payloads may omit it while the field rolls out.
11268 */
11269 source?: 'user' | 'sdk' | 'system' | 'loop_wakeup' | 'schedule_wakeup' | 'poll_event';
11270 session_title?: string;
11271 };
11272
11273 /**
11274 * The result of a call on `$` as the hooks on its event see it: `{ value }`,
11275 * the call's answer, or `{ deny }`, the reason the caller's promise rejects.
11276 */
11277 type ValueOrDeny<Value> = {
11278 value: Value;
11279 deny?: undefined;
11280 } | {
11281 deny: string;
11282 value?: undefined;
11283 };
11284
11285 type WorktreeCreateHookInput = BaseHookInput & {
11286 hook_event_name: 'WorktreeCreate';
11287 name: string;
11288 };
11289
11290 type WorktreeRemoveHookInput = BaseHookInput & {
11291 hook_event_name: 'WorktreeRemove';
11292 worktree_path: string;
11293 };
11294
11295 /**
11296 * The globals of a hooks module's environment: these and no others (no DOM,
11297 * no Node). No code generation either: `eval` and `new Function` over a
11298 * string throw, and there is no `WebAssembly` (deliberately; a module that
11299 * needs compiled code runs it in a process of its own through `$.process.run`).
11300 * A surface module's environment (Client) has the same globals and a
11301 * `console`; neither has timers (a hooks module waits on `$.clock`, a
11302 * surface module on `surface.every`).
11303 */
11304 global {
11305 /**
11306 * The JSX factory (classic runtime, `@jsx h`; the engine prepends the
11307 * pragma): a plain-data element from a string tag or a component.
11308 *
11309 * JSX compiles to bare `h(...)` calls resolved by name, so a module (hooks or
11310 * surface) never declares, imports or takes as a parameter anything named
11311 * `h` or `Fragment` where it writes JSX, and carries no `@jsx` pragma of its
11312 * own: either silently points its JSX away from this factory, and a
11313 * `<Client>` tag written under another factory is not found when the
11314 * module's surface modules are read off its source.
11315 */
11316 const h: (
11317 tag: string | ((props: never) => RenderNode | null | undefined),
11318 props: Record<string, unknown> | null | undefined,
11319 ...children: unknown[]
11320 ) => RenderNode | null | undefined
11321
11322 /**
11323 * `<>...</>`: a column Box around the children.
11324 */
11325 const Fragment: (props: { children?: RenderNode[] }) => RenderElement
11326
11327 /**
11328 * JSX over the element table: every tag is a constructor from
11329 * `$.ui.resolve(e)` (`const { Box, Text } = $.ui.resolve(e)`), typed by
11330 * its props; there are no intrinsic (string) tags.
11331 */
11332 namespace JSX {
11333 type Element = RenderElement
11334 type Children = RenderChildren
11335 type ElementType = (props: never) => RenderNode | null | undefined
11336 interface IntrinsicElements {}
11337 interface ElementChildrenAttribute {
11338 children: unknown
11339 }
11340 interface IntrinsicAttributes {
11341 key?: string
11342 }
11343 }
11344
11345 interface AbortSignal {
11346 readonly aborted: boolean
11347 readonly reason: unknown
11348 throwIfAborted(): void
11349 addEventListener(
11350 type: 'abort',
11351 listener: () => void,
11352 options?: { once?: boolean },
11353 ): void
11354 removeEventListener(type: 'abort', listener: () => void): void
11355 }
11356 var AbortSignal: {
11357 prototype: AbortSignal
11358 abort(reason?: unknown): AbortSignal
11359 timeout(milliseconds: number): AbortSignal
11360 any(signals: AbortSignal[]): AbortSignal
11361 }
11362 interface AbortController {
11363 readonly signal: AbortSignal
11364 abort(reason?: unknown): void
11365 }
11366 var AbortController: {
11367 prototype: AbortController
11368 new (): AbortController
11369 }
11370 interface TextEncoder {
11371 readonly encoding: string
11372 encode(input?: string): Uint8Array
11373 }
11374 var TextEncoder: { prototype: TextEncoder; new (): TextEncoder }
11375 interface TextDecoder {
11376 readonly encoding: string
11377 decode(input?: ArrayBufferView | ArrayBuffer): string
11378 }
11379 var TextDecoder: { prototype: TextDecoder; new (label?: string): TextDecoder }
11380 interface URLSearchParams {
11381 append(name: string, value: string): void
11382 delete(name: string): void
11383 get(name: string): string | null
11384 getAll(name: string): string[]
11385 has(name: string): boolean
11386 set(name: string, value: string): void
11387 toString(): string
11388 forEach(callback: (value: string, key: string) => void): void
11389 }
11390 var URLSearchParams: {
11391 prototype: URLSearchParams
11392 new (init?: string | Record<string, string> | string[][]): URLSearchParams
11393 }
11394 interface URL {
11395 hash: string
11396 host: string
11397 hostname: string
11398 href: string
11399 readonly origin: string
11400 password: string
11401 pathname: string
11402 port: string
11403 protocol: string
11404 search: string
11405 readonly searchParams: URLSearchParams
11406 username: string
11407 toString(): string
11408 toJSON(): string
11409 }
11410 var URL: {
11411 prototype: URL
11412 new (url: string, base?: string | URL): URL
11413 canParse(url: string, base?: string): boolean
11414 }
11415 function atob(data: string): string
11416 function btoa(data: string): string
11417 function structuredClone<T>(value: T): T
11418 var crypto: {
11419 readonly subtle: {
11420 digest(
11421 algorithm: string | { name: string },
11422 data: ArrayBufferView | ArrayBuffer,
11423 ): Promise<ArrayBuffer>
11424 }
11425 randomUUID(): string
11426 getRandomValues<T extends ArrayBufferView>(array: T): T
11427 }
11428 var performance: { now(): number }
11429 }
11430}
11431
11432declare module 'claude-code/testing' {
11433 import type { Args } from 'claude-code';
11434 import type { Chunk } from 'claude-code';
11435 import type { ClientKeyEvent } from 'claude-code';
11436 import type { ClientPointerEvent } from 'claude-code';
11437 import type { Elements } from 'claude-code';
11438 import type { EventCalls } from 'claude-code';
11439 import type { EventName } from 'claude-code';
11440 import type { HookStream } from 'claude-code';
11441 import type { JsonValue } from 'claude-code';
11442 import type { On } from 'claude-code';
11443 import type { PressedLink } from 'claude-code';
11444 import type { Register } from 'claude-code';
11445 import type { RenderComponent } from 'claude-code';
11446 import type { RenderElement } from 'claude-code';
11447 import type { RenderPropsOf } from 'claude-code';
11448 import type { RenderSurface } from 'claude-code';
11449 import type { RenderViewport } from 'claude-code';
11450 import type { ResultOf } from 'claude-code';
11451 import type { StreamingEventName } from 'claude-code';
11452 import type { Tier } from 'claude-code';
11453 import type { UiInputArgument } from 'claude-code';
11454 import type { UiInputResult } from 'claude-code';
11455 import type { UiPressResult } from 'claude-code';
11456 import type { UiSelectResult } from 'claude-code';
11457
11458 /**
11459 * A value that matches by a rule inside `toEqual` and its kin
11460 * (`expect.any`, `expect.objectContaining`), known by its text in a failure.
11461 */
11462 export type AsymmetricMatcher = {
11463 readonly text: string;
11464 };
11465
11466 /**
11467 * The same checks on what a promise received settles with, each resolving
11468 * once the promise has settled and the check passed.
11469 */
11470 export type AsyncMatchers = {
11471 [K in keyof Matchers]: (...args: Parameters<Matchers[K]>) => Promise<void>;
11472 };
11473
11474 /**
11475 * Which `Client` of a mounted drawing an act or a read means, by the `key`
11476 * its element carries; optional while the drawing holds exactly one.
11477 */
11478 export type ClientScope = {
11479 in?: string;
11480 };
11481
11482 /**
11483 * A class, as `toThrow`, `toBeInstanceOf` and `expect.any` take it.
11484 */
11485 export type Constructor = abstract new (...args: never[]) => unknown;
11486
11487 /**
11488 * A group of tests: its name leads the title of each test declared inside,
11489 * and its body runs at once, while the file loads.
11490 *
11491 * @param name the group's name
11492 * @param body declares the group's tests
11493 */
11494 export const describe: (name: string, body: () => void) => void;
11495
11496 /**
11497 * The element each surface-dependent act of a mounted drawing reaches: a
11498 * handle carries the act only where the surface's table has the element.
11499 *
11500 * `input` needs `Input` and `select` needs `Select` (every surface but
11501 * mobile today); `key`, `pointer`, `post`, `advance` and `resize` reach a
11502 * `Client` (terminal and desktop today). The rest need nothing but a tree.
11503 */
11504 export type ElementOfAct = {
11505 input: 'Input';
11506 select: 'Select';
11507 key: 'Client';
11508 pointer: 'Client';
11509 post: 'Client';
11510 advance: 'Client';
11511 resize: 'Client';
11512 };
11513
11514 /**
11515 * What a mounted drawing's `find` and `findAll` match an element on, every
11516 * field given at once: its tag, its `key`, and the text it shows.
11517 *
11518 * `text` matches the element's whole shown text (string children, a Button
11519 * or Link's label, a Markdown's text, a Code's source, an Input's value):
11520 * a string by inclusion, a RegExp by test.
11521 *
11522 * @example
11523 * await ui.findAll({ type: 'Text', text: /overdue/, in: 'board' })
11524 */
11525 export type ElementQuery = {
11526 type?: string;
11527 key?: string;
11528 text?: string | RegExp;
11529 /**
11530 * Searches what this `Client`'s surface module drew (by the `Client`'s
11531 * key) instead of the plugin's own tree.
11532 */
11533 in?: string;
11534 };
11535
11536 /**
11537 * What a test holds as `$`, the engine's own: every call on it is made as
11538 * the REPL, the query loop and the render sites make theirs, over every plugin.
11539 *
11540 * `next.origin` is the engine, and the whole chain runs over the plugins
11541 * loaded. A tool call's `tool_use_id` is minted when left out; `$.ui.mount`
11542 * is a surface the test names drawing an instance, `$.ui.press` a press on it.
11543 */
11544 export type Engine = {
11545 [N in keyof EventCalls]: N extends 'ui' ? EngineNoun<N> & EnginePress & EngineMount : EngineNoun<N>;
11546 };
11547
11548 /**
11549 * One call on the engine's `$`: the event's input whole, as an engine call
11550 * site passes it, to its result, or for a streaming event to its stream.
11551 */
11552 export type EngineCall<E extends EventName> = E extends StreamingEventName ? (e: Args<E>) => HookStream<Chunk<E>, ResultOf[E]> : (e: Args<E>) => Promise<ResultOf[E]>;
11553
11554 /**
11555 * A surface drawing a component instance through the plugins, then driven
11556 * by key through a handle typed by that surface (Mounted).
11557 *
11558 * `ui.render` raised there for those props, the tree validated by that
11559 * table, its `Client`s running; the reads and `press` on every surface,
11560 * `input` where the table has `Input`, the `Client` acts where `Client`.
11561 *
11562 * @example
11563 * for (const surface of SURFACES) run(await $.ui.mount({ ...HINT, surface }))
11564 */
11565 export type EngineMount = {
11566 /**
11567 * Raises the plugins' drawing of the component on the surface and holds
11568 * it, as that surface showing the instance does.
11569 *
11570 * Resolves once the tree validated and every `Client` in it drew once.
11571 *
11572 * @param target whose elements, the surface, the component, its props,
11573 * the instance's requestId, and what the surface measured
11574 * @returns the mounted drawing; rejects when no `surface` (or an unknown
11575 * one) is named, when the surface cannot draw the tree the chain
11576 * returned (with the reason: an element its table lacks, a prop not
11577 * allowed), or when a `Client`'s module failed or drew nothing
11578 */
11579 mount: <P extends RenderSurface, C extends RenderComponent>(target: MountTarget<P, C>) => Promise<Mounted<P, C>>;
11580 };
11581
11582 /**
11583 * One noun of the engine's `$`: each of its events as the engine calls it,
11584 * `tool.call` and `ui.render` typed per tool and component, `ui.resolve` out.
11585 */
11586 export type EngineNoun<N extends keyof EventCalls> = {
11587 [V in EngineNounEvent<N>]: `${N}.${V}` extends 'tool.call' | 'ui.render' ? EventCalls[N][V] : EngineCall<`${N}.${V}` & EventName>;
11588 };
11589
11590 /**
11591 * The events of one noun a test's `$` carries: every one but `ui.resolve`,
11592 * which a render hook calls on its own `$` and the engine never raises.
11593 */
11594 export type EngineNounEvent<N extends keyof EventCalls> = Exclude<keyof EventCalls[N] & string, `${N}.resolve` extends 'ui.resolve' ? 'resolve' : never>;
11595
11596 /**
11597 * A surface pressing a Button a test rendered, or a Markdown's link: the
11598 * `ui.press` chain over every plugin hooked on it, its own closure last.
11599 */
11600 export type EnginePress = {
11601 /**
11602 * Presses the Button, as a click or its hotkey does; with `link`, the
11603 * Markdown's link, as a click on it does.
11604 *
11605 * `e.surface` is the surface of the drawing that holds it.
11606 *
11607 * @param target whose element, its key, the instance and surface when
11608 * several hold it, and for a Markdown the link
11609 * @returns what the chain settled on; rejects when no such element is
11610 * drawn, or several are and neither instance nor surface tells them apart
11611 */
11612 press: (target: PressTarget) => Promise<UiPressResult | undefined>;
11613 };
11614
11615 /**
11616 * The checks on a value (`expect(received)`), and with them the matchers
11617 * that stand inside an expected value (`expect.any(Number)`).
11618 */
11619 export type Expect = Expecting & Matching;
11620
11621 /**
11622 * Checks a value: `expect(received).toEqual(expected)` throws an
11623 * AssertionError naming both sides when it fails, a message given leading.
11624 *
11625 * `.not` negates; `.resolves` and `.rejects` check what a promise settles
11626 * with; `expect.any(Number)` and its kin stand inside an expected value.
11627 */
11628 export const expect: Expect;
11629
11630 /**
11631 * What `expect(received)` answers: the checks, their negation, and the
11632 * checks on what a promise received resolves or rejects with.
11633 */
11634 export type Expectation = Negatable<Matchers> & {
11635 /**
11636 * The checks on what the promise received resolves with; a rejection
11637 * fails them.
11638 */
11639 resolves: Negatable<AsyncMatchers>;
11640 /**
11641 * The checks on what the promise received rejects with; a resolution
11642 * fails them.
11643 */
11644 rejects: Negatable<AsyncMatchers>;
11645 };
11646
11647 /**
11648 * `expect(received, message?)`: the checks on a value, a message of the
11649 * test's own leading a failure's.
11650 */
11651 export type Expecting = (received: unknown, message?: string) => Expectation;
11652
11653 /**
11654 * One element of a mounted drawing as `find` returns it: the description
11655 * the plugin's hook (or a `Client`'s module) built, read as plain data.
11656 *
11657 * What every surface is handed to draw, not what any surface painted: the
11658 * same on the terminal, the desktop, VS Code and mobile.
11659 */
11660 export type FoundElement = {
11661 /**
11662 * The element's tag: `Box`, `Text`, `Button`, `Input`, `Client`, ...
11663 */
11664 type: string;
11665 /**
11666 * What its props carry as `key`, when they do.
11667 */
11668 key: string | undefined;
11669 /**
11670 * Its props, plain data: a Button's `label`, a Text's `color`, ...;
11671 * handlers never cross (a Button's `onPress` is not here).
11672 */
11673 props: Record<string, unknown>;
11674 /**
11675 * What it shows: its string children in document order and, for the
11676 * leaves that draw a prop as text, that prop (ElementQuery `text`).
11677 */
11678 text: string;
11679 /**
11680 * Its children as drawn: strings and element descriptions.
11681 */
11682 children: unknown[];
11683 };
11684
11685 /**
11686 * The checks `expect(received)` offers; each throws an AssertionError when
11687 * it fails, naming what was expected and what was received.
11688 */
11689 export type Matchers = {
11690 /**
11691 * Passes when received is the value, by `Object.is`.
11692 *
11693 * @param expected the value
11694 */
11695 toBe: (expected: unknown) => void;
11696 /**
11697 * Passes when received equals the value in structure, properties holding
11698 * undefined ignored; `expect.any` and its kin match inside it.
11699 *
11700 * @param expected the value
11701 */
11702 toEqual: (expected: unknown) => void;
11703 /**
11704 * As `toEqual`, but undefined properties and prototypes count.
11705 *
11706 * @param expected the value
11707 */
11708 toStrictEqual: (expected: unknown) => void;
11709 /**
11710 * Passes when received holds at least the object's properties, equal.
11711 *
11712 * @param expected the object
11713 */
11714 toMatchObject: (expected: object) => void;
11715 /**
11716 * Passes when a string received includes the item, or an iterable
11717 * received has an element that is it.
11718 *
11719 * @param item the substring or element
11720 */
11721 toContain: (item: unknown) => void;
11722 /**
11723 * Passes when an iterable received has an element equal to the item.
11724 *
11725 * @param item the element
11726 */
11727 toContainEqual: (item: unknown) => void;
11728 /**
11729 * Passes when received's `length` is the number.
11730 *
11731 * @param length the number
11732 */
11733 toHaveLength: (length: number) => void;
11734 /**
11735 * Passes when received holds a property at the path (`a.b`, or the keys
11736 * in order), equal to the value when one is given.
11737 *
11738 * @param path the property's path
11739 * @param value what it must equal, when given
11740 */
11741 toHaveProperty: (path: string | readonly string[], value?: unknown) => void;
11742 /**
11743 * Passes when received is undefined.
11744 */
11745 toBeUndefined: () => void;
11746 /**
11747 * Passes when received is anything but undefined.
11748 */
11749 toBeDefined: () => void;
11750 /**
11751 * Passes when received is null.
11752 */
11753 toBeNull: () => void;
11754 /**
11755 * Passes when received is truthy.
11756 */
11757 toBeTruthy: () => void;
11758 /**
11759 * Passes when received is falsy.
11760 */
11761 toBeFalsy: () => void;
11762 /**
11763 * Passes when received is NaN.
11764 */
11765 toBeNaN: () => void;
11766 /**
11767 * Passes when received is greater than the bound.
11768 *
11769 * @param bound the bound
11770 */
11771 toBeGreaterThan: (bound: number | bigint) => void;
11772 /**
11773 * Passes when received is the bound or greater.
11774 *
11775 * @param bound the bound
11776 */
11777 toBeGreaterThanOrEqual: (bound: number | bigint) => void;
11778 /**
11779 * Passes when received is less than the bound.
11780 *
11781 * @param bound the bound
11782 */
11783 toBeLessThan: (bound: number | bigint) => void;
11784 /**
11785 * Passes when received is the bound or less.
11786 *
11787 * @param bound the bound
11788 */
11789 toBeLessThanOrEqual: (bound: number | bigint) => void;
11790 /**
11791 * Passes when a string received includes the text, or the pattern matches
11792 * it.
11793 *
11794 * @param pattern the text or pattern
11795 */
11796 toMatch: (pattern: string | RegExp) => void;
11797 /**
11798 * Passes when a string received starts with the prefix.
11799 *
11800 * @param prefix the prefix
11801 */
11802 toStartWith: (prefix: string) => void;
11803 /**
11804 * Passes when a string received ends with the suffix.
11805 *
11806 * @param suffix the suffix
11807 */
11808 toEndWith: (suffix: string) => void;
11809 /**
11810 * Passes when received is an instance of the class.
11811 *
11812 * @param expected the class
11813 */
11814 toBeInstanceOf: (expected: Constructor) => void;
11815 /**
11816 * Passes when calling received throws (or, under `rejects`, the rejection
11817 * is) an error as described: by substring, pattern, class or message.
11818 *
11819 * @param expected the description; any throw passes without one
11820 */
11821 toThrow: (expected?: ThrowExpectation) => void;
11822 };
11823
11824 /**
11825 * The matchers that stand inside an expected value, each matching received
11826 * there by a rule instead of by equality.
11827 */
11828 export type Matching = {
11829 /**
11830 * Matches any instance of the class, a primitive by its wrapper
11831 * (`expect.any(Number)`).
11832 *
11833 * @param expected the class
11834 * @returns the matcher
11835 */
11836 any: (expected: Constructor) => AsymmetricMatcher;
11837 /**
11838 * Matches anything but null and undefined.
11839 *
11840 * @returns the matcher
11841 */
11842 anything: () => AsymmetricMatcher;
11843 /**
11844 * Matches a string that includes the text.
11845 *
11846 * @param text the substring
11847 * @returns the matcher
11848 */
11849 stringContaining: (text: string) => AsymmetricMatcher;
11850 /**
11851 * Matches a string the pattern matches.
11852 *
11853 * @param pattern the pattern, or the source of one
11854 * @returns the matcher
11855 */
11856 stringMatching: (pattern: string | RegExp) => AsymmetricMatcher;
11857 /**
11858 * Matches an object holding at least the shape's properties, equal.
11859 *
11860 * @param shape the properties
11861 * @returns the matcher
11862 */
11863 objectContaining: (shape: object) => AsymmetricMatcher;
11864 /**
11865 * Matches an array holding an element equal to each of the items.
11866 *
11867 * @param items the elements
11868 * @returns the matcher
11869 */
11870 arrayContaining: (items: readonly unknown[]) => AsymmetricMatcher;
11871 };
11872
11873 /**
11874 * The world beneath the plugins, mocked noun by noun: each member registers
11875 * hooks of the test's on `on`, visible where the test calls it.
11876 */
11877 export type Mock = {
11878 /**
11879 * Answers `$.clock` from a clock in memory that moves only when the test
11880 * moves it: `clock.now` reads it, and each wait is held.
11881 *
11882 * A held wait resolves when an advance crosses the time it is due, and is
11883 * dropped when its dispatch aborts; one held past a hook's budget (ten
11884 * seconds of real time) is let go, as a hook that overran.
11885 *
11886 * @param on the test's `on`
11887 * @param options where the clock starts (`now`, 0 when not given)
11888 * @returns the clock: its time, and the calls that move it
11889 */
11890 clock: (on: On, options?: MockClockOptions) => MockClock;
11891 /**
11892 * Answers `$.store` from a store in memory: `get`, `set`, `delete` and
11893 * `keys` over it, as the engine keeps a plugin's own.
11894 *
11895 * @param on the test's `on`
11896 * @param entries what the store holds at the start (nothing when not given)
11897 */
11898 store: (on: On, entries?: Readonly<Record<string, unknown>>) => void;
11899 /**
11900 * Answers `$.env.get` from a set of variables; one not listed is unset.
11901 *
11902 * @param on the test's `on`
11903 * @param variables the environment the plugins read
11904 */
11905 env: (on: On, variables: Readonly<Record<string, string>>) => void;
11906 };
11907
11908 /**
11909 * The world beneath the plugins, mocked noun by noun: `mock.clock`,
11910 * `mock.store` and `mock.env`.
11911 *
11912 * Each registers hooks of the test's on the `on` it is handed, visible where
11913 * the test calls it, and answers its noun from memory.
11914 */
11915 export const mock: Mock;
11916
11917 /**
11918 * The clock `mock.clock` hands back: the time its hooks answer, and the only
11919 * ways it moves.
11920 */
11921 export type MockClock = {
11922 /**
11923 * The time now, in milliseconds: what `$.clock.now()` resolves beneath the
11924 * plugins.
11925 *
11926 * @returns the time
11927 */
11928 now: () => number;
11929 /**
11930 * Moves the clock on, resolving each wait due on the way (`$.clock.sleep`,
11931 * `after`, `every`, this clock's `sleep`) in the order it comes due.
11932 *
11933 * The clock reads each wait's time as it resolves, and what one started
11934 * runs before the next resolves.
11935 *
11936 * @param ms how far to move, in milliseconds
11937 * @returns resolves once the clock is there and the event loop settled
11938 */
11939 advance: (ms: number) => Promise<void>;
11940 /**
11941 * Moves the clock to a time at or past now, as `advance` would.
11942 *
11943 * @param ms the time, in milliseconds
11944 * @returns resolves once the clock is there and the event loop settled
11945 */
11946 set: (ms: number) => Promise<void>;
11947 /**
11948 * Lets what is already under way run as far as it can without the clock
11949 * moving: every wait due now resolves and the event loop settles.
11950 *
11951 * The same as `advance(0)`; the name for the step between starting a
11952 * dispatch unawaited and looking at what it did.
11953 *
11954 * @returns resolves once the event loop settled, the time where it was
11955 */
11956 settle: () => Promise<void>;
11957 /**
11958 * Resolves once the clock has moved this far past now: how a hook of the
11959 * test's answers late.
11960 *
11961 * @param ms how far, in milliseconds
11962 * @returns resolves when an advance crosses that time
11963 */
11964 sleep: (ms: number) => Promise<void>;
11965 };
11966
11967 /**
11968 * Where a mocked clock starts: `now`, in milliseconds (0 when not given).
11969 */
11970 export type MockClockOptions = {
11971 now?: number;
11972 };
11973
11974 /**
11975 * A drawing of component `C` the test mounted on surface `P`: reads over its
11976 * description, acts on its elements by key, `Client` acts where `P` has one.
11977 *
11978 * Narrowed by the `surface` the test passed, as `e.surface` narrows
11979 * `$.ui.resolve(e)`: no `input` on `mobile`, no `key` on `vscode`, whose
11980 * tables (Elements) lack `Input` and `Client`; over a union, their meet.
11981 */
11982 export type Mounted<P extends RenderSurface = RenderSurface, C extends RenderComponent = RenderComponent> = {
11983 [K in keyof MountedMembers<P, C> as K extends keyof ElementOfAct ? ElementOfAct[K] extends keyof Elements[P] ? K : never : K]: MountedMembers<P, C>[K];
11984 };
11985
11986 /**
11987 * Every member a mounted drawing of component `C` can have; `Mounted<P, C>`
11988 * keeps the ones surface `P` has the element for (ElementOfAct).
11989 *
11990 * Each act resolves once the loop settled and what a plugin invalidated
11991 * was drawn again. The kit exercises the mod (its hooks, its description
11992 * under the surface's rules, its `Client` modules), never a surface's paint.
11993 */
11994 export type MountedMembers<P extends RenderSurface, C extends RenderComponent = RenderComponent> = {
11995 /**
11996 * Where this drawing is, as the test named it.
11997 */
11998 readonly surface: P;
11999 /**
12000 * The tree as last drawn: what the plugins' `ui.render` chain returned
12001 * for this instance, validated by the surface's table, as plain data.
12002 *
12003 * With `in`, what that `Client`'s surface module last drew instead.
12004 *
12005 * @param scope `{ in }`: a `Client`'s key, to read its module's tree
12006 * @returns the tree; rejects with the refusal when the surface could not
12007 * draw what the chain returned, or with a `Client`'s fault line
12008 * @example
12009 * expect(await ui.drawn()).toMatchObject({ type: 'Box' })
12010 */
12011 drawn: (scope?: ClientScope) => Promise<RenderElement>;
12012 /**
12013 * The first element matching the query, outermost first in document
12014 * order, or undefined.
12015 *
12016 * @param query tag, key and shown text to match; `in` to search a Client
12017 * @returns the element's description, or undefined when none matches
12018 * @example
12019 * expect((await ui.find({ key: 'count' }))?.text).toBe('3 notes')
12020 */
12021 find: (query: ElementQuery) => Promise<FoundElement | undefined>;
12022 /**
12023 * Every element matching the query, in document order.
12024 *
12025 * @param query tag, key and shown text to match; `in` to search a Client
12026 * @returns the matching elements' descriptions
12027 * @example
12028 * expect(await ui.findAll({ type: 'Button' })).toHaveLength(2)
12029 */
12030 findAll: (query: ElementQuery) => Promise<FoundElement[]>;
12031 /**
12032 * Presses the Button keyed `key`, as the person activating it on this
12033 * surface does: the `ui.press` chain, the Button's own `onPress` last.
12034 *
12035 * `e.surface` is this drawing's. With `link`, presses that link of the
12036 * Markdown keyed `key` instead. A Button a `Client`'s module drew is
12037 * pressed the same way, by its key.
12038 *
12039 * @param target the key; another plugin's name to press its Button; a link
12040 * @returns what the chain settled on
12041 * @example
12042 * await ui.press({ key: 'save' })
12043 */
12044 press: (target: MountPressTarget) => Promise<UiPressResult | undefined>;
12045 /**
12046 * Types into the Input keyed `key`: the `ui.input` chain with the text as
12047 * `e.value`, the Input's own handler last.
12048 *
12049 * `kind` `submit` (the default) is Enter with that text, `change` an edit
12050 * that leaves it in the field.
12051 *
12052 * @param target the key and the text; `kind`; another plugin's name
12053 * @returns what the chain settled on
12054 * @example
12055 * await ui.input({ key: 'title', text: 'groceries' })
12056 */
12057 input: (target: MountInputTarget) => Promise<UiInputResult | undefined>;
12058 /**
12059 * Picks `value` in the Select keyed `key`: the `ui.select` chain, the
12060 * Select's own `onSelect` last.
12061 *
12062 * @param target the key and the option's value; another plugin's name
12063 * @returns what the chain settled on
12064 * @example
12065 * await ui.select({ key: 'sort', value: 'date' })
12066 */
12067 select: (target: MountSelectTarget) => Promise<UiSelectResult | undefined>;
12068 /**
12069 * Hands a `Client`'s `onKey` listener one key, as a key pressed while
12070 * its region has the focus does.
12071 *
12072 * `in` names the `Client` by key; optional when the drawing holds one.
12073 *
12074 * @param event the key (`{ key: 'right' }`, `{ key: 'a', ctrl: true }`)
12075 * @example
12076 * await ui.key({ key: 'return', in: 'board' })
12077 */
12078 key: (event: MountKeyEvent) => Promise<void>;
12079 /**
12080 * Hands a `Client`'s `onPointer` listener one event, in the cells of its
12081 * region as every surface that draws a `Client` reports them.
12082 *
12083 * @param event the event (`{ type: 'down', x: 3, y: 0, button: 'left' }`)
12084 * @example
12085 * await ui.pointer({ type: 'down', x: 2, y: 0, button: 'left' })
12086 */
12087 pointer: (event: MountPointerEvent) => Promise<void>;
12088 /**
12089 * Posts as a `Client`'s own `surface.post(data)` does: the plugin's
12090 * `ui.message` hooks run with `e.data`, `e.surface` this drawing's.
12091 *
12092 * A `{ props }` they answer reaches the instance, which redraws with
12093 * its state kept.
12094 *
12095 * @param data plain data (JsonValue)
12096 * @param scope `{ in }`: which `Client`, when the drawing holds several
12097 * @example
12098 * await ui.post({ pick: 2 })
12099 */
12100 post: (data: JsonValue, scope?: ClientScope) => Promise<void>;
12101 /**
12102 * Moves the drawing's frame clock on: each `surface.every(ms, fn)` timer
12103 * of its `Client`s fires at every interval of its own the move crosses.
12104 *
12105 * @param ms how far, in milliseconds
12106 * @example
12107 * await ui.advance(250)
12108 */
12109 advance: (ms: number) => Promise<void>;
12110 /**
12111 * Lays a `Client`'s region out at a size, as the surface measuring it
12112 * does: what its module reads as `surface.columns` and `surface.rows`.
12113 *
12114 * @param size cells across and down; `in`: which `Client`
12115 * @example
12116 * await ui.resize({ columns: 40, rows: 3 })
12117 */
12118 resize: (size: MountResizeTarget) => Promise<void>;
12119 /**
12120 * Draws the instance again, as the engine does when its props change (a
12121 * tool row's output arrived): the `ui.render` chain with the next props.
12122 *
12123 * A plugin's own `$.ui.invalidate` needs no call: the drawing follows it
12124 * before the next read.
12125 *
12126 * @param props the component's next props; the current ones when absent
12127 * @example
12128 * await ui.redraw({ ...TOOL_ROW, isRunning: false, output: 'done' })
12129 */
12130 redraw: (props?: RenderPropsOf[C]) => Promise<void>;
12131 /**
12132 * Lets the drawing go, as the surface dropping the instance does: its
12133 * `Client`s and their timers end. Later acts on the handle reject.
12134 *
12135 * @example
12136 * await ui.unmount()
12137 */
12138 unmount: () => Promise<void>;
12139 };
12140
12141 /**
12142 * What a mounted drawing's `input` takes: the Input's key, the text, which
12143 * of the two inputs it is (`submit` when unsaid), another plugin's name.
12144 */
12145 export type MountInputTarget = {
12146 key: string;
12147 text: string;
12148 kind?: UiInputArgument['kind'];
12149 plugin?: string;
12150 };
12151
12152 /**
12153 * What a mounted drawing's `key` takes: the key as a `Client`'s `onKey`
12154 * listener receives it (ClientKeyEvent), and which `Client` when several.
12155 */
12156 export type MountKeyEvent = ClientKeyEvent & ClientScope;
12157
12158 /**
12159 * What a mounted drawing's `pointer` takes: the event as a `Client`'s
12160 * `onPointer` listener receives it, in its region's cells, and which one.
12161 */
12162 export type MountPointerEvent = ClientPointerEvent & ClientScope;
12163
12164 /**
12165 * What a mounted drawing's `press` takes: the Button's key, another
12166 * plugin's name when the Button is not the mounted plugin's, a link.
12167 */
12168 export type MountPressTarget = {
12169 key: string;
12170 plugin?: string;
12171 /**
12172 * Presses this link of the Markdown keyed `key` instead of a Button.
12173 */
12174 link?: PressedLink;
12175 };
12176
12177 /**
12178 * What a mounted drawing's `resize` takes: a `Client` region's size in
12179 * cells, as the surface measuring it hands it, and which `Client`.
12180 */
12181 export type MountResizeTarget = {
12182 columns: number;
12183 rows: number;
12184 in?: string;
12185 };
12186
12187 /**
12188 * What a mounted drawing's `select` takes: the Select's key, the picked
12189 * option's value, another plugin's name when the Select is not its own.
12190 */
12191 export type MountSelectTarget = {
12192 key: string;
12193 value: string;
12194 plugin?: string;
12195 };
12196
12197 /**
12198 * What `$.ui.mount` takes: whose elements the test will act on, the surface
12199 * that draws, and the component instance the engine asks the plugins for.
12200 *
12201 * The envelope a `ui.render` hook sees as `e` (`surface`, `component`,
12202 * `requestId`, `viewport`, `props`) plus `plugin`. `surface` is never
12203 * defaulted: one body run over several surfaces is what shows independence.
12204 *
12205 * @example
12206 * $.ui.mount({ plugin: 'notes', surface, component: 'Pane', props: PANE })
12207 */
12208 export type MountTarget<P extends RenderSurface = RenderSurface, C extends RenderComponent = RenderComponent> = {
12209 /**
12210 * Whose elements the handle's acts address by bare `key`: the plugin
12211 * under test, usually. An act names another plugin to reach its element.
12212 */
12213 plugin: string;
12214 /**
12215 * What draws: its element table validates every tree the plugins
12216 * return, and every dispatch the handle raises carries it as `e.surface`.
12217 *
12218 * One of `terminal`, `desktop`, `vscode`, `mobile`.
12219 */
12220 surface: P;
12221 /**
12222 * Which component the engine asks the plugins to draw (`Pane`,
12223 * `AbovePrompt`, `ToolUse`, ...): what a `ui.render` matcher narrows on.
12224 */
12225 component: C;
12226 /**
12227 * The component's props, what the hook reads as `e.props`.
12228 */
12229 props: RenderPropsOf[C];
12230 /**
12231 * The instance drawn (a pane's id, a tool row's tool_use_id); one is
12232 * minted when absent. Two mounts of one component are two instances.
12233 */
12234 requestId?: string;
12235 /**
12236 * What the surface measured, as the hook reads it under `e.viewport`;
12237 * absent when the surface has not measured, as on `e`.
12238 *
12239 * Cells across and down in the surface's monospace metric and whether it
12240 * docks a pane. A `Client` in the drawing starts laid out at this size
12241 * (0 by 0 when absent) until the handle's `resize` lays its region out.
12242 */
12243 viewport?: RenderViewport;
12244 };
12245
12246 /**
12247 * A set of checks and, under `not`, the same set passing where they fail.
12248 */
12249 export type Negatable<M> = M & {
12250 /**
12251 * The checks negated.
12252 */
12253 not: M;
12254 };
12255
12256 /**
12257 * Written inline in a test and loaded as a plugin folder is: its name, the
12258 * tier it loads in (`user` when not given), and its hooks module's `register`.
12259 *
12260 * `register` is written `register(on) { ... }` and is self-contained, as a
12261 * module's is: it closes over nothing of the test file.
12262 */
12263 export type Plugin = {
12264 name: string;
12265 tier?: PluginTier;
12266 register: Register;
12267 };
12268
12269 /**
12270 * A tier a plugin loads in: every tier but the engine's own.
12271 */
12272 export type PluginTier = Exclude<Tier, 'core'>;
12273
12274 /**
12275 * What `$.ui.press` takes: the plugin whose `ui.render` hook drew the element,
12276 * the `key` it gave it, the instance and surface when several, a `link`.
12277 */
12278 export type PressTarget = {
12279 plugin: string;
12280 key: string;
12281 /**
12282 * Which instance holds it, when the plugin drew the key in several.
12283 */
12284 requestId?: string;
12285 /**
12286 * Which surface's drawing, when the key is drawn on more than one; the
12287 * press carries the surface of the drawing it lands in either way.
12288 */
12289 surface?: RenderSurface;
12290 /**
12291 * For a Markdown answering its links: which of them the press lands on.
12292 */
12293 link?: PressedLink;
12294 };
12295
12296 /**
12297 * One test: it passes when its body returns or resolves, and fails when it
12298 * throws, rejects or outlasts its time (5000 ms, or `timeoutMs`).
12299 *
12300 * The body gets the engine's `$` and an `on` whose hooks sit beneath every
12301 * plugin; `plugins` load inline plugins beside the one under test. A failure
12302 * carries what the engine reported meanwhile: each hook it skipped, and why.
12303 *
12304 * @param name the test's name, led in its title by the describes around it
12305 * @param rest the body, `($, on) => ...`, or the options then the body
12306 */
12307 export const test: (name: string, ...rest: TestRest) => void;
12308
12309 /**
12310 * A test: the engine's `$`, and `on`, a plugin's registrar, whose hooks sit
12311 * beneath every plugin; beneath them the bottom hook throws, naming its event.
12312 *
12313 * The plugins load at the test's first call on `$`, so a test registers its
12314 * hooks before it, as a module registers its own in `register()`.
12315 */
12316 export type TestBody = ($: Engine, on: On) => unknown;
12317
12318 /**
12319 * What `test` takes beside its name: the inline plugins it loads beside the
12320 * one under test, and how long it may run (5000 ms when not given).
12321 */
12322 export type TestOptions = {
12323 plugins?: readonly Plugin[];
12324 timeoutMs?: number;
12325 };
12326
12327 /**
12328 * What follows a test's name: its body, or its options then its body.
12329 */
12330 export type TestRest = readonly [body: TestBody] | readonly [options: TestOptions, body: TestBody];
12331
12332 /**
12333 * What `toThrow` compares the thrown error with: a substring or pattern of
12334 * its message, its class, a value carrying the whole message, or nothing.
12335 */
12336 export type ThrowExpectation = string | RegExp | Constructor | WithMessage;
12337
12338 /**
12339 * Says which tier the plugin under test loads in, once, at the top of the
12340 * file: `prepend`, `user` (when unsaid), `append` or `builtin`.
12341 *
12342 * @param tier the tier
12343 */
12344 export const tier: (tier: PluginTier) => void;
12345
12346 /**
12347 * An error-like value `toThrow` compares by its whole message.
12348 */
12349 export type WithMessage = {
12350 message: string;
12351 };
12352}
12353
12354// The inputs of the built-in tools this build has, from each tool's
12355// input schema. Merges into ToolCallInput (BuiltinToolInputs) so
12356// `e.tool === "Bash"` narrows to the tool's arguments.
12357declare module 'claude-code' {
12358 interface BuiltinToolInputs {
12359 Agent: {
12360 /** A short (3-5 word) description of the task */
12361 description: string
12362 /** The task for the agent to perform */
12363 prompt: string
12364 /** The type of specialized agent to use for this task */
12365 subagent_type?: string
12366 /** Optional model override for this agent. Takes precedence over the agent definition's model frontmatter and the configured default subagent model. If omitted, uses the agent definition's model, else the default (inherits from the parent unless a default subagent model is configured). Ignored for subagent_type: "fork" — forks always inherit the parent model. */
12367 model?: "sonnet" | "opus" | "haiku" | "fable"
12368 /** Agents run in the background by default; you will be notified when one completes. Set to false only when your very next action depends on this agent's result and nothing else could usefully happen while it runs — otherwise leave it in the background so the user can hand you other work. */
12369 run_in_background?: boolean
12370 /** Name for the spawned agent. Makes it addressable via SendMessage({to: name}) while running. */
12371 name?: string
12372 /** Deprecated; ignored. The session has a single implicit team. */
12373 team_name?: string
12374 /** Deprecated; ignored. Subagents inherit the parent session's permission mode; agent-definition frontmatter may override it. */
12375 mode?: "acceptEdits" | "auto" | "bypassPermissions" | "default" | "dontAsk" | "plan"
12376 /** Isolation mode. "worktree" creates a temporary git worktree so the agent works on an isolated copy of the repo. "remote" launches the agent in a remote cloud environment (always runs in background; availability is gated). */
12377 isolation?: "worktree" | "remote"
12378 }
12379 Bash: {
12380 /** The command to execute */
12381 command: string
12382 /** Optional timeout in milliseconds (max 600000) */
12383 timeout?: number
12384 /** Clear, concise description of what this command does in active voice. Never use words like "complex" or "risk" in the description - just describe what it does. Say what the command does in plain words: do not echo the command's text, its flags, or file paths - the user reads this description, often without seeing the command. For simple commands (git, npm, standard CLI tools), keep it brief (5-10 words): - ls → "List files in current directory" - git status → "Show working tree status" - npm install → "Install package dependencies" For commands that are harder to parse at a glance (piped commands, obscure flags, etc.), add enough context to clarify what it does: - find . -name "*.tmp" -exec rm {} \; → "Find and delete all .tmp files recursively" - git reset --hard origin/main → "Discard all local changes and match remote main" - curl -s url | jq '.data[]' → "Fetch JSON from URL and extract data array elements" */
12385 description?: string
12386 /** Set to true to run this command in the background. */
12387 run_in_background?: boolean
12388 /** Set this to true to dangerously override sandbox mode and run commands without sandboxing. */
12389 dangerouslyDisableSandbox?: boolean
12390 }
12391 CronCreate: {
12392 /** Standard 5-field cron expression in local time: "M H DoM Mon DoW" (e.g. "* /5 * * * *" = every 5 minutes, "30 14 28 2 *" = Feb 28 at 2:30pm local once). */
12393 cron: string
12394 /** The prompt to enqueue at each fire time. */
12395 prompt: string
12396 /** true (default) = fire on every cron match until deleted or auto-expired after 7 days. false = fire once at the next match, then auto-delete. Use false for "remind me at X" one-shot requests with pinned minute/hour/dom/month. */
12397 recurring?: boolean
12398 /** true = persist to .claude/scheduled_tasks.json and survive restarts. false (default) = in-memory only, dies when this Claude session ends. Use true only when the user asks the task to survive across sessions. */
12399 durable?: boolean
12400 }
12401 CronDelete: {
12402 /** Job ID returned by CronCreate. */
12403 id: string
12404 }
12405 CronList: {}
12406 Edit: {
12407 /** The absolute path to the file to modify */
12408 file_path: string
12409 /** The text to replace */
12410 old_string: string
12411 /** The text to replace it with (must be different from old_string) */
12412 new_string: string
12413 /** Replace all occurrences of old_string (default false) */
12414 replace_all?: boolean
12415 }
12416 EnterWorktree: {
12417 /** Optional name for a new worktree. Each "/"-separated segment may contain only letters, digits, dots, underscores, and dashes; max 64 chars total. A random name is generated if not provided. Mutually exclusive with `path`. */
12418 name?: string
12419 /** Path to an existing worktree to switch into instead of creating a new one. Must appear in `git worktree list` for the current repo — or, on first entry from the launch directory, for a repo nested inside it (multi-repo workspace). Mutually exclusive with `name`. */
12420 path?: string
12421 }
12422 ExitWorktree: {
12423 /** "keep" leaves the worktree and branch on disk; "remove" deletes both. */
12424 action: "keep" | "remove"
12425 /** Required true when action is "remove" and the worktree has uncommitted files or unmerged commits. The tool will refuse and list them otherwise. */
12426 discard_changes?: boolean
12427 }
12428 ListAgents: {
12429 /** Not available in this build; leave unset. */
12430 channel?: string
12431 /** Not available in this build; leave unset. */
12432 q?: string
12433 }
12434 NotebookEdit: {
12435 /** The absolute path to the Jupyter notebook file to edit (must be absolute, not relative) */
12436 notebook_path: string
12437 /** The ID of the cell to edit. When inserting a new cell, the new cell will be inserted after the cell with this ID, or at the beginning if not specified. */
12438 cell_id?: string
12439 /** The new source for the cell */
12440 new_source: string
12441 /** The type of the cell (code or markdown). If not specified, it defaults to the current cell type. If using edit_mode=insert, this is required. */
12442 cell_type?: "code" | "markdown"
12443 /** The type of edit to make (replace, insert, delete). Defaults to replace. */
12444 edit_mode?: "replace" | "insert" | "delete"
12445 }
12446 PowerShell: {
12447 /** The PowerShell command to execute */
12448 command: string
12449 /** Optional timeout in milliseconds (max 600000) */
12450 timeout?: number
12451 /** Clear, concise description of what this command does in active voice. */
12452 description?: string
12453 /** Set to true to run this command in the background. */
12454 run_in_background?: boolean
12455 /** Set this to true to dangerously override sandbox mode and run commands without sandboxing. */
12456 dangerouslyDisableSandbox?: boolean
12457 }
12458 Read: {
12459 /** The absolute path to the file to read */
12460 file_path: string
12461 /** The line number to start reading from. Only provide if the file is too large to read at once */
12462 offset?: number
12463 /** The number of lines to read. Only provide if the file is too large to read at once. */
12464 limit?: number
12465 /** Page range for PDF files (e.g., "1-5", "3", "10-20"). Only applicable to PDF files. Maximum 20 pages per request. */
12466 pages?: string
12467 }
12468 ReportFindings: {
12469 /** Effort level the review ran at */
12470 level?: "low" | "medium" | "high" | "xhigh" | "max"
12471 /** Verified findings, most-severe first; empty if none survived */
12472 findings: Array<{
12473 /** Repo-relative path of the file the finding is in */
12474 file: string
12475 /** 1-indexed line the finding anchors to */
12476 line?: number
12477 /** One-sentence statement of the defect */
12478 summary: string
12479 /** Compressed label for compact UI (≤60 chars): the claim alone, no rationale or consequence clause */
12480 short_summary?: string
12481 /** Concrete inputs/state → wrong output/crash */
12482 failure_scenario: string
12483 /** Short kebab-case slug of the finding type, e.g. "correctness", "simplification", "efficiency", "test-coverage" */
12484 category?: string
12485 /** Set when a verify pass ran; absent on inline-only reviews */
12486 verdict?: "CONFIRMED" | "PLAUSIBLE"
12487 /** Set ONLY when re-reporting after applying fixes: what happened to this finding */
12488 outcome?: "fixed" | "skipped" | "no_change_needed"
12489 }>
12490 }
12491 ScheduleWakeup: {
12492 /** Seconds from now to wake up. Clamped to [60, 3600] by the runtime. Required unless `stop` is true. */
12493 delaySeconds?: number
12494 /** One short sentence explaining the chosen delay. Goes to telemetry and is shown to the user. Be specific. Required unless `stop` is true. */
12495 reason?: string
12496 /** The /loop input to fire on wake-up. Pass the same /loop input verbatim each turn so the next firing re-enters the skill and continues the loop. For autonomous /loop (no user prompt), pass the literal sentinel `<<autonomous-loop-dynamic>>` instead (the dynamic-pacing variant, not the CronCreate-mode `<<autonomous-loop>>`). Required unless `stop` is true. */
12497 prompt?: string
12498 /** Set to true to end the dynamic loop immediately instead of scheduling another wakeup. When true, all other fields are ignored and no further wakeups fire. */
12499 stop?: boolean
12500 /** true = nothing changed (you checked and there is nothing to report). false = something happened worth keeping (edited a file, posted a message, advanced state, surfaced a finding). Consecutive noop:true ticks are collapsed in the user's terminal view and tracked as a streak. Required unless `stop` is true. */
12501 noop?: boolean
12502 }
12503 SendMessage: {
12504 /** Recipient: a name from ListAgents (append its " [ref]" only when a listing or an error shows one), a teammate name, "main", or a background agent's agentId */
12505 to: unknown & unknown
12506 /** A 5-10 word label for your own transcript row (not transmitted — the recipient previews the first line of `message`). Truncated to 200 characters rather than rejected. */
12507 summary?: string
12508 /** Plain text message content. The recipient's human sees only the FIRST LINE as a one-line preview until they expand it, so make the first line a clear, self-contained sentence saying what this is about — not a greeting, preamble, or bare @-mention. */
12509 message: string
12510 /** Ask a session ON THIS MACHINE to send you ONE notice when it next goes idle (finishes its turn with nothing queued) or exits — opt-in, one-shot, no polling. With a message: deliver it now AND subscribe. Without a message (omit it): a pure subscription that costs the other session nothing. */
12511 notify_when_idle?: boolean
12512 }
12513 Skill: {
12514 /** The name of a skill from the available-skills list. Do not guess names. */
12515 skill: string
12516 /** Optional arguments for the skill */
12517 args?: string
12518 }
12519 TaskOutput: {
12520 /** The task ID to get output from */
12521 task_id: string
12522 /** Whether to wait for completion */
12523 block: boolean
12524 /** Max wait time in ms */
12525 timeout: number
12526 }
12527 TaskStop: {
12528 /** The ID of the background task to stop. Agent-team teammates and named background agents are also accepted by agent ID or name. */
12529 task_id?: string
12530 /** Deprecated: use task_id instead */
12531 shell_id?: string
12532 }
12533 ToolSearch: {
12534 /** Query to find deferred tools. Use "select:<tool_name>" for direct selection, or keywords to search. */
12535 query: string
12536 /** Maximum number of results to return (default: 5) */
12537 max_results: number
12538 }
12539 WebFetch: {
12540 /** The URL to fetch content from */
12541 url: string
12542 /** The prompt to run on the fetched content */
12543 prompt: string
12544 }
12545 WebSearch: {
12546 /** The search query to use */
12547 query: string
12548 /** Only include search results from these domains */
12549 allowed_domains?: string[]
12550 /** Never include search results from these domains */
12551 blocked_domains?: string[]
12552 }
12553 Workflow: {
12554 /** Self-contained workflow script. Must begin with `export const meta = { name, description, phases }` (pure literal, no computed values) followed by the script body using agent()/parallel()/pipeline()/phase(). */
12555 script?: string
12556 /** Name of a predefined workflow (built-in or from .claude/workflows/). Resolves to a self-contained script. */
12557 name?: string
12558 /** Ignored — set the workflow description in the script's `meta` block. */
12559 description?: string
12560 /** Ignored — set the workflow title in the script's `meta` block. */
12561 title?: string
12562 /** Optional input value exposed to the script as the global `args`, verbatim. Pass arrays/objects as actual JSON values, NOT as a JSON-encoded string — a stringified list breaks `args.filter`/`args.map` in the script. Use for parameterized named workflows (e.g. a research question). */
12563 args?: unknown
12564 /** Path to a workflow script file on disk. Every Workflow invocation persists its script under the session directory and returns the path in the tool result. To iterate, edit that file with Write/Edit and re-invoke Workflow with the same `scriptPath` instead of re-sending the full script. Takes precedence over `script` and `name`. */
12565 scriptPath?: string
12566 /** Run ID of a prior Workflow invocation to resume from. Completed agent() calls with unchanged (prompt, opts) return their cached results instantly; only edited or new calls re-run. Same-session only. Stop the prior run first (TaskStop) before resuming. */
12567 resumeFromRunId?: string
12568 }
12569 Write: {
12570 /** The absolute path to the file to write (must be absolute, not relative) */
12571 file_path: string
12572 /** The content to write to the file */
12573 content: string
12574 }
12575 }
12576}
12577
12578// The structured results of the same tools, from each tool's output
12579// schema. Merges into ToolCallResult (BuiltinToolResults) so after
12580// `e.tool === "Bash"` the `result` of `next(e)` is the tool's record.
12581declare module 'claude-code' {
12582 interface BuiltinToolResults {
12583 Agent: {
12584 agentId: string
12585 /** @internal Count of leading harness-authored content blocks (hand-back provenance bookkeeping; not a stable consumer field) */
12586 harnessNoteCount?: number
12587 /** @internal Count of trailing harness-authored content blocks (hand-back provenance bookkeeping; not a stable consumer field) */
12588 harnessTailCount?: number
12589 /** @internal Fingerprint binding the harness section counts to the exact content they were computed against; a hook rewrite invalidates the counts rather than misplacing rewritten bytes */
12590 harnessSectionHash?: string
12591 agentType?: string
12592 /** @internal How the report reached this result under the subagent hand-back contract: 'send' = the child's classified SubagentHandback delivery, 'flagged' = that delivery under a SECURITY WARNING, 'withheld' = nothing was delivered (not a stable consumer field) */
12593 handback?: "send" | "flagged" | "withheld"
12594 content: Array<{
12595 type: "text"
12596 text: string
12597 citations?: unknown[] | null
12598 }>
12599 resolvedModel?: string
12600 modelsUsed?: string[]
12601 totalToolUseCount: number
12602 totalDurationMs: number
12603 totalTokens: number
12604 usage: {
12605 input_tokens: number
12606 output_tokens: number
12607 cache_creation_input_tokens: number | null
12608 cache_read_input_tokens: number | null
12609 server_tool_use: {
12610 web_search_requests: number
12611 web_fetch_requests: number
12612 } | null
12613 service_tier: string | null
12614 cache_creation: {
12615 ephemeral_1h_input_tokens: number
12616 ephemeral_5m_input_tokens: number
12617 } | null
12618 inference_geo?: string | null
12619 speed?: string | null
12620 iterations?: unknown
12621 output_tokens_details?: {
12622 thinking_tokens?: number | null
12623 } | null
12624 }
12625 toolStats?: {
12626 readCount: number
12627 searchCount: number
12628 bashCount: number
12629 editFileCount: number
12630 linesAdded: number
12631 linesRemoved: number
12632 otherToolCount: number
12633 frameCount?: number
12634 }
12635 status: "completed"
12636 prompt: string
12637 worktreePath?: string
12638 worktreeBranch?: string
12639 } | {
12640 status: "async_launched"
12641 isAsync?: true
12642 /** The ID of the async agent */
12643 agentId: string
12644 /** The description of the task */
12645 description: string
12646 /** Model in use at the backgrounding transition (a pre-background swap is reflected here) */
12647 resolvedModel?: string
12648 /** Ordered distinct models used before backgrounding (length > 1 means a mid-run swap) */
12649 modelsUsed?: string[]
12650 /** The prompt for the agent */
12651 prompt: string
12652 /** Path to the output file for checking agent progress */
12653 outputFile: string
12654 /** Whether the calling agent has Read/Bash tools to check progress */
12655 canReadOutputFile?: boolean
12656 } | {
12657 status: "remote_launched"
12658 /** The ID of the remote agent task */
12659 taskId: string
12660 /** The URL of the cloud session */
12661 sessionUrl: string
12662 /** The description of the task */
12663 description: string
12664 /** The prompt for the agent */
12665 prompt: string
12666 /** Path to the output file for checking agent progress */
12667 outputFile: string
12668 }
12669 Bash: {
12670 /** The standard output of the command */
12671 stdout: string
12672 /** The standard error output of the command */
12673 stderr: string
12674 /** Path to raw output file for large MCP tool outputs */
12675 rawOutputPath?: string
12676 /** Whether the command was interrupted */
12677 interrupted: boolean
12678 /** Flag to indicate if stdout contains image data */
12679 isImage?: boolean
12680 /** ID of the background task if command is running in background */
12681 backgroundTaskId?: string
12682 /** True if the user manually backgrounded the command with Ctrl+B */
12683 backgroundedByUser?: boolean
12684 /** @internal True if a plugin's turn abort moved the running command to the background */
12685 backgroundedByTurnAbort?: boolean
12686 /** @internal True if the command was moved to the background so a message queued for the model could reach it */
12687 backgroundedToDeliverMessage?: boolean
12688 /** Set when the command hit its timeout and was auto-backgrounded; the timeout value in ms */
12689 timedOutAfterMs?: number
12690 /** Model-facing note that the session cwd was not changed by a backgrounded command containing a directory-change builtin (cd/pushd/popd/chdir) */
12691 backgroundCwdHint?: string
12692 /** True when this backgrounded command is owned by a synchronous subagent and is therefore terminated when that agent gives its final response; absent when the command survives (main loop, async subagents) */
12693 backgroundEndsWithFinalResponse?: true
12694 /** Flag to indicate if sandbox mode was overridden */
12695 dangerouslyDisableSandbox?: boolean
12696 /** Semantic interpretation for non-error exit codes with special meaning */
12697 returnCodeInterpretation?: string
12698 /** Whether the command is expected to produce no output on success */
12699 noOutputExpected?: boolean
12700 /** Structured content blocks */
12701 structuredContent?: unknown[]
12702 /** Path to the persisted full output in tool-results dir (set when output is too large for inline) */
12703 persistedOutputPath?: string
12704 /** Total size of the output in bytes (set when output is too large for inline) */
12705 persistedOutputSize?: number
12706 /** Model-facing note listing readFileState entries whose mtime bumped during this command (set when WRITE_COMMAND_MARKERS matches) */
12707 staleReadFileStateHint?: string
12708 /** Model-facing system-reminder appended when a gh command reports a GitHub API rate-limit error */
12709 ghRateLimitHint?: string
12710 /** Structured classification of git/gh operations detected in this command (commit/push/merge/rebase/PR). Client-facing — lets clients render git activity without re-parsing stdout; not surfaced to the model. */
12711 gitOperation?: {
12712 commit?: {
12713 sha: string
12714 kind: "committed" | "amended" | "cherry-picked"
12715 branch?: string
12716 }
12717 push?: {
12718 branch: string
12719 }
12720 branch?: {
12721 ref: string
12722 action: "merged" | "rebased"
12723 }
12724 pr?: {
12725 number: number
12726 url?: string
12727 action: "created" | "edited" | "merged" | "commented" | "closed" | "reopened" | "ready" | "draft" | "auto-merge-enabled" | "auto-merge-disabled"
12728 }
12729 }
12730 /** @internal Per-file diff of the working-tree changes this command made, for rendering and for PostToolUse Bash hooks (changedFiles: absolute paths of every changed file known, shown or not, at most 200; cut when shorter than files.length + moreFiles); not surfaced to the model. */
12731 bashEditDiff?: {
12732 files: {
12733 filePath: string
12734 hunks: {
12735 oldStart: number
12736 oldLines: number
12737 newStart: number
12738 newLines: number
12739 lines: string[]
12740 }[]
12741 created?: true
12742 deleted?: true
12743 }[]
12744 moreFiles: number
12745 changedFiles?: string[]
12746 unavailable?: true
12747 skipped?: true
12748 shared?: true
12749 }
12750 }
12751 CronCreate: {
12752 id: string
12753 humanSchedule: string
12754 recurring: boolean
12755 durable?: boolean
12756 }
12757 CronDelete: {
12758 id: string
12759 }
12760 CronList: {
12761 jobs: {
12762 id: string
12763 cron: string
12764 humanSchedule: string
12765 prompt: string
12766 recurring?: boolean
12767 durable?: boolean
12768 }[]
12769 }
12770 Edit: {
12771 /** The file path that was edited */
12772 filePath: string
12773 /** The original string that was replaced */
12774 oldString: string
12775 /** The new string that replaced it */
12776 newString: string
12777 /** The original file contents before editing */
12778 originalFile: string | null
12779 /** Diff patch showing the changes */
12780 structuredPatch: {
12781 oldStart: number
12782 oldLines: number
12783 newStart: number
12784 newLines: number
12785 lines: string[]
12786 }[]
12787 /** Whether the user modified the proposed changes */
12788 userModified: boolean
12789 /** Whether all occurrences were replaced */
12790 replaceAll: boolean
12791 gitDiff?: {
12792 filename: string
12793 status: "modified" | "added"
12794 additions: number
12795 deletions: number
12796 changes: number
12797 patch: string
12798 /** GitHub owner/repo when available */
12799 repository?: string | null
12800 }
12801 /** True when the edit was held for the machine owner to review instead of written; the file is unchanged */
12802 staged?: boolean
12803 }
12804 EnterWorktree: {
12805 worktreePath: string
12806 worktreeBranch?: string
12807 message: string
12808 }
12809 ExitWorktree: {
12810 action: "keep" | "remove"
12811 originalCwd: string
12812 worktreePath: string
12813 worktreeBranch?: string
12814 tmuxSessionName?: string
12815 discardedFiles?: number
12816 discardedCommits?: number
12817 message: string
12818 }
12819 ListAgents: {
12820 /** Formatted list of reachable agents */
12821 listing: string
12822 }
12823 NotebookEdit: {
12824 /** The new source code that was written to the cell */
12825 new_source: string
12826 /** The previous cell source (replace/delete only). Enables cell-relative diff rendering without re-reading the notebook. */
12827 old_source?: string
12828 /** The ID of the cell that was edited */
12829 cell_id?: string
12830 /** The type of the cell */
12831 cell_type: "code" | "markdown"
12832 /** The programming language of the notebook */
12833 language: string
12834 /** The edit mode that was used */
12835 edit_mode: string
12836 /** Error message if the operation failed */
12837 error?: string
12838 /** The path to the notebook file */
12839 notebook_path: string
12840 /** The original notebook content before modification */
12841 original_file: string
12842 /** The updated notebook content after modification */
12843 updated_file: string
12844 }
12845 PowerShell: {
12846 /** The standard output of the command */
12847 stdout: string
12848 /** The standard error output of the command */
12849 stderr: string
12850 /** Whether the command was interrupted */
12851 interrupted: boolean
12852 /** Semantic interpretation for non-error exit codes with special meaning */
12853 returnCodeInterpretation?: string
12854 /** Flag to indicate if stdout contains image data */
12855 isImage?: boolean
12856 /** Path to persisted full output when too large for inline */
12857 persistedOutputPath?: string
12858 /** Total output size in bytes when persisted */
12859 persistedOutputSize?: number
12860 /** ID of the background task if command is running in background */
12861 backgroundTaskId?: string
12862 /** True if the user manually backgrounded the command with Ctrl+B */
12863 backgroundedByUser?: boolean
12864 /** @internal True if a plugin's turn abort moved the running command to the background */
12865 backgroundedByTurnAbort?: boolean
12866 /** @internal True if the command was moved to the background so a message queued for the model could reach it */
12867 backgroundedToDeliverMessage?: boolean
12868 /** Set when the command hit its timeout and was auto-backgrounded; the timeout value in ms */
12869 timedOutAfterMs?: number
12870 /** True when this backgrounded command is owned by a synchronous subagent and is therefore terminated when that agent gives its final response; absent when the command survives (main loop, async subagents) */
12871 backgroundEndsWithFinalResponse?: true
12872 /** Structured classification of git/gh operations detected in this command (commit/push/merge/rebase/PR). Client-facing — lets clients render git activity without re-parsing stdout; not surfaced to the model. */
12873 gitOperation?: {
12874 commit?: {
12875 sha: string
12876 kind: "committed" | "amended" | "cherry-picked"
12877 branch?: string
12878 }
12879 push?: {
12880 branch: string
12881 }
12882 branch?: {
12883 ref: string
12884 action: "merged" | "rebased"
12885 }
12886 pr?: {
12887 number: number
12888 url?: string
12889 action: "created" | "edited" | "merged" | "commented" | "closed" | "reopened" | "ready" | "draft" | "auto-merge-enabled" | "auto-merge-disabled"
12890 }
12891 }
12892 }
12893 Read: {
12894 type: "text"
12895 file: {
12896 /** The path to the file that was read */
12897 filePath: string
12898 /** The content of the file */
12899 content: string
12900 /** Number of lines in the returned content */
12901 numLines: number
12902 /** The starting line number */
12903 startLine: number
12904 /** Total number of lines in the file */
12905 totalLines: number
12906 /** True when a whole-file read was auto-paginated because it exceeded the token cap (the content is a partial first page). A programmatic signal for internal consumers; survives output reconstruction (unlike the render-time banner). */
12907 truncatedByTokenCap?: boolean
12908 }
12909 /** Set when this Read completed a saved Artifact source file: the Artifact and the version of it that now counts as viewed. */
12910 artifactRead?: {
12911 slug: string
12912 ver: string
12913 }
12914 } | {
12915 type: "image"
12916 file: {
12917 /** Base64-encoded image data */
12918 base64: string
12919 /** The MIME type of the image */
12920 type: "image/jpeg" | "image/png" | "image/gif" | "image/webp"
12921 /** Original file size in bytes */
12922 originalSize: number
12923 /** Image dimension info for coordinate mapping */
12924 dimensions?: {
12925 /** Original image width in pixels */
12926 originalWidth?: number
12927 /** Original image height in pixels */
12928 originalHeight?: number
12929 /** Displayed image width in pixels (after resizing) */
12930 displayWidth?: number
12931 /** Displayed image height in pixels (after resizing) */
12932 displayHeight?: number
12933 }
12934 }
12935 } | {
12936 type: "notebook"
12937 file: {
12938 /** The path to the notebook file */
12939 filePath: string
12940 /** Array of notebook cells */
12941 cells: unknown[]
12942 }
12943 } | {
12944 type: "pdf"
12945 file: {
12946 /** The path to the PDF file */
12947 filePath: string
12948 /** Base64-encoded PDF data */
12949 base64: string
12950 /** Original file size in bytes */
12951 originalSize: number
12952 }
12953 } | {
12954 type: "parts"
12955 file: {
12956 /** The path to the PDF file */
12957 filePath: string
12958 /** Original file size in bytes */
12959 originalSize: number
12960 /** Number of pages extracted */
12961 count: number
12962 /** Directory containing extracted page images */
12963 outputDir: string
12964 }
12965 /** Document page number of the first extracted page (1 when no range was requested); labels the page images in the model-facing tool_result */
12966 firstPage?: number
12967 /** Extracted page images, in page order. Present only transiently in-process: the page image bytes are delivered solely as image blocks in the model-facing tool_result content and are not retained on the tool_use_result, so this key is absent on the emitted/persisted result */
12968 pages?: Array<{
12969 /** Base64-encoded page image; empty when the page could not be processed */
12970 base64: string
12971 /** The MIME type of the image */
12972 mediaType: "image/jpeg" | "image/png" | "image/gif" | "image/webp"
12973 /** Why the page could not be processed as an image; set only when base64 is empty */
12974 error?: string
12975 }>
12976 } | {
12977 type: "file_unchanged"
12978 file: {
12979 /** The path to the file */
12980 filePath: string
12981 }
12982 /** Set when the dedup matched a startup-seeded entry (CLAUDE.md / nested memory) rather than a prior Read tool_result */
12983 source?: "seeded"
12984 }
12985 ReportFindings: {
12986 /** Number of findings reported */
12987 count: number
12988 /** Effort level the review ran at */
12989 level?: "low" | "medium" | "high" | "xhigh" | "max"
12990 /** Echoed for the result body */
12991 findings: Array<{
12992 /** Repo-relative path of the file the finding is in */
12993 file: string
12994 /** 1-indexed line the finding anchors to */
12995 line?: number
12996 /** One-sentence statement of the defect */
12997 summary: string
12998 /** Compressed label for compact UI (≤60 chars): the claim alone, no rationale or consequence clause */
12999 short_summary?: string
13000 /** Concrete inputs/state → wrong output/crash */
13001 failure_scenario: string
13002 /** Short kebab-case slug of the finding type, e.g. "correctness", "simplification", "efficiency", "test-coverage" */
13003 category?: string
13004 /** Set when a verify pass ran; absent on inline-only reviews */
13005 verdict?: "CONFIRMED" | "PLAUSIBLE"
13006 /** Set ONLY when re-reporting after applying fixes: what happened to this finding */
13007 outcome?: "fixed" | "skipped" | "no_change_needed"
13008 }>
13009 }
13010 ScheduleWakeup: {
13011 /** Epoch ms timestamp when the next wakeup will fire */
13012 scheduledFor: number
13013 /** Actual delay used after clamping to runtime bounds */
13014 clampedDelaySeconds: number
13015 /** True if the requested delaySeconds was outside [60, 3600] */
13016 wasClamped: boolean
13017 /** True when the model ended the loop via `stop: true` */
13018 stopped?: boolean
13019 /** How many pending dynamic-loop wakeups stop:true cancelled. 0 means nothing was pending — a recurring /loop cron is not cancelled by stop:true. */
13020 cancelledWakeups?: number
13021 }
13022 SendMessage: unknown
13023 Skill: {
13024 /** Whether the skill is valid */
13025 success: boolean
13026 /** The name of the skill */
13027 commandName: string
13028 /** Tools allowed by this skill */
13029 allowedTools?: string[]
13030 /** Resolved model the skill turn runs on when a frontmatter model override took effect; omitted otherwise */
13031 model?: string
13032 /** Execution status */
13033 status?: "inline"
13034 /** True when the skill instructions were loaded read-only (nothing was executed) */
13035 readOnly?: boolean
13036 } | {
13037 /** Whether the skill completed successfully */
13038 success: boolean
13039 /** The name of the skill */
13040 commandName: string
13041 /** Execution status */
13042 status: "forked"
13043 /** The ID of the sub-agent that executed the skill */
13044 agentId: string
13045 /** The result from the forked skill execution */
13046 result: string
13047 /** True when the sub-agent was launched in the background: `result` describes the launch, and the skill outcome arrives later as a task notification. */
13048 background?: boolean
13049 }
13050 TaskOutput: unknown
13051 TaskStop: {
13052 /** Status message about the operation */
13053 message: string
13054 /** The ID of the task that was stopped */
13055 task_id: string
13056 /** The type of the task that was stopped */
13057 task_type: string
13058 /** The command or description of the stopped task */
13059 command?: string
13060 }
13061 ToolSearch: {
13062 matches: string[]
13063 query: string
13064 total_deferred_tools: number
13065 pending_mcp_servers?: string[]
13066 failed_mcp_servers?: {
13067 name: string
13068 errorCode?: string
13069 error?: string
13070 }[]
13071 }
13072 WebFetch: {
13073 /** Size of the fetched content in bytes */
13074 bytes: number
13075 /** HTTP response code */
13076 code: number
13077 /** HTTP response code text */
13078 codeText: string
13079 /** Processed result from applying the prompt to the content */
13080 result: string
13081 /** Time taken to fetch and process the content */
13082 durationMs: number
13083 /** The URL that was fetched */
13084 url: string
13085 artifactRead?: {
13086 slug: string
13087 ver?: string
13088 seeded?: false
13089 }
13090 }
13091 WebSearch: {
13092 /** The search query that was executed */
13093 query: string
13094 /** Search results and/or text commentary from the model */
13095 results: Array<{
13096 /** ID of the tool use */
13097 tool_use_id: string
13098 /** Array of search hits */
13099 content: Array<{
13100 /** The title of the search result */
13101 title: string
13102 /** The URL of the search result */
13103 url: string
13104 }>
13105 } | string>
13106 /** Time taken to complete the search operation */
13107 durationSeconds: number
13108 /** Number of web searches performed */
13109 searchCount?: number
13110 }
13111 Workflow: {
13112 status: "async_launched" | "remote_launched"
13113 taskId: string
13114 /** TaskType of the registered background task — 'local_workflow' for in-process runs, 'remote_agent' when remote:true dispatches to CCR. Set on all new writes; absent only on transcripts written before this field existed. */
13115 taskType?: "local_workflow" | "remote_agent"
13116 /** meta.name from the workflow script — same value as task_started.workflow_name. Set on all new writes; absent only on transcripts written before this field existed. */
13117 workflowName?: string
13118 /** Local workflow run identifier for resumeFromRunId. Absent for remote_launched (the CCR session URL is the resume handle there) and on transcripts written before this field existed. */
13119 runId?: string
13120 summary?: string
13121 /** Directory where subagent transcripts are written during execution */
13122 transcriptDir?: string
13123 /** Path to the persisted workflow script for this invocation. Editable via Write/Edit; pass back as `scriptPath` to re-run without resending the script. */
13124 scriptPath?: string
13125 /** CCR session URL when status is remote_launched */
13126 sessionUrl?: string
13127 /** Non-blocking heads-up (e.g. local git state diverges from the pushed branch the cloud session will clone) */
13128 warning?: string
13129 /** Set if syntax check failed */
13130 error?: string
13131 }
13132 Write: {
13133 /** Whether a new file was created or an existing file was updated */
13134 type: "create" | "update"
13135 /** The path to the file that was written */
13136 filePath: string
13137 /** The content that was written to the file */
13138 content: string
13139 /** Diff patch showing the changes (empty when nothing changed, the diff timed out, or — with originalFile null on an update — the previous content was too large to diff) */
13140 structuredPatch: {
13141 oldStart: number
13142 oldLines: number
13143 newStart: number
13144 newLines: number
13145 lines: string[]
13146 }[]
13147 /** The original file content before the write (null for new files, or when the previous content was too large to include) */
13148 originalFile: string | null
13149 gitDiff?: {
13150 filename: string
13151 status: "modified" | "added"
13152 additions: number
13153 deletions: number
13154 changes: number
13155 patch: string
13156 /** GitHub owner/repo when available */
13157 repository?: string | null
13158 }
13159 /** True when the user edited the proposed content in the permission dialog before accepting */
13160 userModified?: boolean
13161 /** True when the write was held for the machine owner to review instead of written; the file is unchanged */
13162 staged?: boolean
13163 }
13164 }
13165}