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 / agents-md / README.md

Alice T'Poteatmods/agents-md: the AGENTS.md project-instructions mod (#95409) a92ea1c · on Sep 18, 2026 History

163 lines · 11.3 KBPreview CodeRaw

1	# agents-md
2
3	`AGENTS.md` read the way Claude Code reads `CLAUDE.md`, as a plugin, under
4	one option, `instructionFiles`:
5
6	- `claude-md`: only `CLAUDE.md` is loaded, by the engine, as today. The plugin
7	adds nothing.
8	- `claude-md-or-agents-md` (the default): a project with no instruction files
9	of its own gets its `AGENTS.md` files instead, loaded exactly where and how
10	`CLAUDE.md` would be. "Of its own" is read off what the engine loaded for
11	the context: a `CLAUDE.md`, `.claude/CLAUDE.md` or `CLAUDE.local.md` in any
12	directory from the root down to the working directory leaves the whole
13	project to the engine, and the plugin stays out (the organization's managed
14	file, the person's `~/.claude/CLAUDE.md`, a `.claude/rules` file and an
15	added directory's `CLAUDE.md` do not count, as the nested walk does not see
16	them either). With none, every `AGENTS.md` and `.claude/AGENTS.md` on that
17	path joins the instruction files the engine renders, and a `Read` under a
18	subdirectory attaches that directory's `AGENTS.md` unless a `CLAUDE.md`
19	there claims it.
20	- `claude-md-and-agents-md`: every `AGENTS.md` is loaded beside `CLAUDE.md`,
21	up and down the tree; a file `CLAUDE.md` already `@`-imports, or is a link
22	to, is not loaded a second time (compared by path, then by content).
23	- `managed-only`: the project's checked-in and private instruction files and
24	the person's own are dropped from the context; the organization's managed
25	`CLAUDE.md` and the engine's memory stay. The engine's nested `CLAUDE.md`
26	attachments on `Read` are not an event yet and still arrive. (The engine's
27	`claudeMdExcludes` setting also exists, for user, project and local files,
28	and applies to the `AGENTS.md` files this plugin reads too.)
29
30	How the files reach the model is the engine's doing, not the plugin's:
31	`prompt.context` hands a hook the instruction files behind `claudeMd`
32	(`{ path, kind, content, parent? }`, kinds `managed`, `user`, `project`,
33	`local`, `memory`, in load order) and a hook answers the list changed. The
34	engine then renders `claudeMd` from the answered files with its own preamble
35	and framing, announces them by name, and keeps only the `managed` ones for an
36	agent that omits project instructions (Explore, Plan, a custom agent with
37	`omitClaudeMd`). So an `AGENTS.md` this plugin adds as a `project` file is, to
38	everything downstream, a project instruction file: same place in the context,
39	same framing, same omission rules, same announcement. An organization's
40	prepended plugin on `prompt.context` sits above this one and has the last
41	word on the files.
42
43	`hooks/register.ts` is the module; everything under `hooks/` is its parts,
44	importing `claude-code` and one another alone. `tests/` runs under
45	`claude plugin test <this folder>`.
46
47	## Setting the option
48
49	As a built-in its option is the `/config` row "Project instructions", a
50	picker over the four values, each described there. By hand it is
51
52	```json
53	{
54	"pluginConfigs": {
55	"agents-md@builtin": {
56	"options": { "instructionFiles": "claude-md-and-agents-md" }
57	}
58	}
59	}
60	```
61
62	in user settings (`~/.claude/settings.json`), `--settings`, or managed
63	settings; a project's `.claude/settings.json` is not read for plugin
64	options. Changing it reloads the module, and the next context the engine
65	builds (the next turn after the reload, a new conversation, `/clear`, a
66	compaction) carries the new mode's files. A hand-typed value outside the
67	four is told once in the transcript and reads as the default. `/plugin`
68	lists the plugin among the built-ins, where a person can turn it off; with
69	it off the engine reads `CLAUDE.md` alone.
70
71	The option was first keyed `projectInstructions`, with the values `claude`,
72	`agents-fallback`, `both` and `none`. A value still stored under that key is
73	honoured for now while `instructionFiles` reads as its default: `none` as
74	`managed-only`, `claude` as `claude-md`, `agents-fallback` as
75	`claude-md-or-agents-md`, `both` as `claude-md-and-agents-md`, any other
76	value as `claude-md` (which adds nothing, never as the default, which loads
77	`AGENTS.md`); the first `session.start` of a load says in the transcript how
78	it was read. Once `instructionFiles` is set to anything but its default, the
79	old key is not read and the transcript says to remove it.
80
81	Run from this folder instead (`claude --plugin-dir mods/agents-md`), the
82	same entry is keyed `"agents-md"`.
83
84	## What it hooks
85
86	\| event \| what the hook does \|
87	\| --- \| --- \|
88	\| `session.start` \| in every mode: passes the start straight through and floats the usage row for the configured mode, never awaited; the first start of a load logs how a stored `projectInstructions` value is read. The session's start never waits on this plugin \|
89	\| `prompt.context` \| under `claude-md-or-agents-md` and `claude-md-and-agents-md`: walks `$.fs.ancestors` for the `AGENTS.md` files above the working directory and answers them as `project` instruction files, each `@` import its own entry after its file, each placed where a project file of its directory stands (root first, before the first deeper project file, else after the last project file, before memory); files the engine already holds by path or by content are left out; under `claude-md-or-agents-md` it answers nothing when the project has a `CLAUDE.md` of its own (among the handed files, else found by a `$.fs.ancestors` walk, so a `CLAUDE.md` the engine loaded and then withheld still counts), and logs which files it loaded once, and again after a move to another project root; handed unknown files (a hook above rewrote the `claudeMd` text) it adds nothing; the first context of a load sends the load row (counts) and the feature mark; under `managed-only` (matcher: a `project`, `local` or `user` file present): answers the list without those kinds \|
90	\| `agent.spawn` on `fork: true` \| under `claude-md-or-agents-md` and `claude-md-and-agents-md`: a fork the Agent tool starts shares its parent's prompt prefix, so the parent loop's delivered nested files are copied to the fork's loop and not attached to it again (a `/fork` or `/subtask` fork does not raise `agent.spawn` yet and starts from an empty set, as every fork did before; a fork started in the same tool batch as a `Read` inherits that Read's file although its prefix holds a placeholder for it) \|
91	\| `tool.call` on `Read` \| under `claude-md-or-agents-md` and `claude-md-and-agents-md`, for a file under the session's project root (`$.session.root()`, read live, so `/cd`, a host's directory change and worktree moves are followed and a moved root starts the delivered sets and the fallback decision over; a file elsewhere gets nothing, as the engine attaches no nested `CLAUDE.md` there): walks only the directories strictly between the root and the read file (`$.fs.ancestors` with `below: root`, as the engine walks only those for a nested `CLAUDE.md`, never up to the filesystem root again) and attaches their `AGENTS.md` files not yet given to that agent loop, not already among the context's instruction files (by path or, for a project file, by text) and not claimed by a `CLAUDE.md` of the same directory (or imported by one), as `context` after the tool result, framed `Contents of <path>:` byte for byte as the engine frames a nested `CLAUDE.md`, whatever its size; each file once per loop and conversation (the context's recomputation after a compaction or `/clear` starts the count over), the context's files never; a Read that attached files sends the nested row. A `~` or `~/` path is read under the home directory as the Read tool reads it \|
92
93	## What it calls on `$`
94
95	`fs.ancestors` (with each found file's `parts`: the file and its imports
96	apart; with `below` on a Read; it finds nothing on a thin client, whose
97	workspace files are remote, as the engine's own walk does), `session.root`,
98	`session.cwd`, `env.get` (`HOME` and `USERPROFILE`, once per load, the
99	profile first on a Windows spelling of the working directory, so a `~/` path
100	the model hands a Read resolves where the Read tool reads it), `ui.log`,
101	`telemetry.log` and `telemetry.mark`.
102
103	`$.telemetry` is the [telemetry](../telemetry) plugin's noun; where that
104	plugin is not seated the calls find no noun and are dropped without a trace,
105	and nothing else changes.
106
107	## What it logs
108
109	Counts and closed choices only; no path and no file text. Each row goes
110	through `$.telemetry.log`, so it exists only where the telemetry plugin
111	does:
112
113	\| event \| when \| properties \|
114	\| --- \| --- \| --- \|
115	\| `agents_md_mode` \| once per fresh load, at `session.start` \| `mode` (`claude-md` \\| `claude-md-or-agents-md` \\| `claude-md-and-agents-md` \\| `managed-only`), `is_interactive` \|
116	\| `agents_md_load` \| the first context of a load, under `claude-md-or-agents-md` and `claude-md-and-agents-md` \| `mode`, `file_count` (`AGENTS.md` files handed to the engine), `import_count` (their `@` imports), `total_content_length`, `yielded` (`claude-md-or-agents-md` stood down for a `CLAUDE.md` of the project's own), `walk_failed`; with it one `$.telemetry.mark` for feature `agents_md`: `ok`, or `sad` with reason `walk_failed` \|
117	\| `agents_md_nested` \| a Read that attached nested files \| `mode`, `file_count` \|
118
119	## Where it still differs from CLAUDE.md
120
121	All of these apply only to the modes that load `AGENTS.md`,
122	`claude-md-or-agents-md` (the default) and `claude-md-and-agents-md`. Each
123	names a loader fact a plugin cannot reach through the events it has today.
124
125	1. Nested files attach on a text `Read` only. The engine also attaches a
126	directory's `CLAUDE.md` for a file `@`-mentioned in the prompt, for the
127	IDE's opened file or selection, and for the `Read` tool's notebook, image
128	and PDF results.
129	2. A nested file the plugin attaches is not registered in the loop's
130	read-file state, so after a compaction the engine does not restore it
131	among the recently read files (the plugin attaches it again at the next
132	`Read` under that directory instead), and a change to it mid-session is
133	not re-announced.
134	3. `/cd` carries the new tree's `CLAUDE.md` in its own notice; the plugin's
135	files for the new tree arrive in the same next request through the
136	engine's instructions announcement instead.
137	4. Paths compare by spelling; the engine resolves a symlinked alias of the
138	working directory before deciding a file is inside it.
139	5. `--add-dir` directories contribute no `AGENTS.md`, where the engine can
140	load their `CLAUDE.md`.
141	6. `/memory` and the `#` shortcut do not know `AGENTS.md` files, and the
142	engine's own initial-load row does not count them (this plugin's
143	`agents_md_load` row does).
144	7. An `@` import outside the working directory inside an `AGENTS.md` is
145	honoured only once the approval the engine asks for a `CLAUDE.md`'s
146	external imports has been given (without it the import is left out, as a
147	`CLAUDE.md`'s is); the approval dialog itself is raised for `CLAUDE.md`
148	imports alone.
149	8. A subagent that is not a fork gets a nested `AGENTS.md` at its own first
150	`Read` under that directory even when its parent's loop was already given
151	it; the engine does not hand such a subagent the nested `CLAUDE.md` again.
152	A fork matches the engine on both sides.
153
154	## Testing
155
156	claude plugin test mods/agents-md
157
158	`tests/register.test.ts` covers the default mode: a project with `AGENTS.md`
159	alone gets it as a project instruction file and one transcript line naming
160	it, a project with a `CLAUDE.md` of its own is left to the engine without a
161	walk, a failed walk leaves the context as handed, and the start hands
162	`$.telemetry` the mode row alone where a test seats a provider for that
163	noun, and goes on untouched where none is seated.