⑂ main

AGENTS.md #28

Merged@k33g wants to merge feature/agents-file into main

Merged as 3c42ab80f861.

39 files changed

renamed .tickets/issues/0080-structured-output-from-go-type-declarations-in-t.yaml +5 -5 Viewed

similarity index 95%
rename from .tickets/issues/0080-structured-output-from-go-type-declarations.yaml
rename to .tickets/issues/0080-structured-output-from-go-type-declarations-in-t.yaml
@@ -1,12 +1,12 @@
1	1	# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2	2	id: 80
3	3	title: Structured output from Go type declarations in the config
4		-state: open
	4	+state: closed
5	5	author:
6	6	name: k33g
7	7	email: ph.charriere@gmail.com
8	8	createdAt: 2026-10-08T10:00:00.000Z
9		-updatedAt: 2026-10-08T16:00:00.000Z
	9	+updatedAt: 2026-10-09T03:16:38.703Z
10	10	labels:
11	11	- enhancement
12	12	epic: Models
@@ -88,7 +88,7 @@ tasks:
88	88	updatedAt: 2026-10-08T16:00:00.000Z
89	89	closedAt: 2026-10-08T16:00:00.000Z
90	90	- id: 4
91		- title: "Refuse structuredOutput when the agent declares any tool (config error listing the tools)"
	91	+ title: Refuse structuredOutput when the agent declares any tool (config error listing the tools)
92	92	state: closed
93	93	priority: none
94	94	author:
@@ -108,7 +108,7 @@ tasks:
108	108	updatedAt: 2026-10-08T16:00:00.000Z
109	109	closedAt: 2026-10-08T16:00:00.000Z
110	110	- id: 6
111		- title: "Quality gate (quality skill)"
	111	+ title: Quality gate (quality skill)
112	112	state: open
113	113	priority: none
114	114	author:
@@ -116,7 +116,7 @@ tasks:
116	116	email: ph.charriere@gmail.com
117	117	createdAt: 2026-10-08T10:00:00.000Z
118	118	- id: 7
119		- title: "Docs EN+FR (how-to + reference/configuration) and .memory update"
	119	+ title: Docs EN+FR (how-to + reference/configuration) and .memory update
120	120	state: closed
121	121	priority: none
122	122	author:

  # Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
  id: 80
  title: Structured output from Go type declarations in the config
--state: open
++state: closed
  author:
    name: k33g
    email: ph.charriere@gmail.com
  createdAt: 2026-10-08T10:00:00.000Z
--updatedAt: 2026-10-08T16:00:00.000Z
++updatedAt: 2026-10-09T03:16:38.703Z
  labels:
    - enhancement
  epic: Models
      updatedAt: 2026-10-08T16:00:00.000Z
      closedAt: 2026-10-08T16:00:00.000Z
    - id: 4
--    title: "Refuse structuredOutput when the agent declares any tool (config error listing the tools)"
++    title: Refuse structuredOutput when the agent declares any tool (config error listing the tools)
      state: closed
      priority: none
      author:
      updatedAt: 2026-10-08T16:00:00.000Z
      closedAt: 2026-10-08T16:00:00.000Z
    - id: 6
--    title: "Quality gate (quality skill)"
++    title: Quality gate (quality skill)
      state: open
      priority: none
      author:
        email: ph.charriere@gmail.com
      createdAt: 2026-10-08T10:00:00.000Z
    - id: 7
--    title: "Docs EN+FR (how-to + reference/configuration) and .memory update"
++    title: Docs EN+FR (how-to + reference/configuration) and .memory update
      state: closed
      priority: none
      author:

added .tickets/issues/0081-support-agents-md-project-instructions-for-the-a.yaml +155 -0 Viewed

new file mode 100644
@@ -0,0 +1,155 @@
	1	+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
	2	+id: 81
	3	+title: Support AGENTS.md (project instructions for the agent)
	4	+state: open
	5	+author:
	6	+ name: k33g
	7	+ email: ph.charriere@gmail.com
	8	+createdAt: 2026-10-09T08:00:00.000Z
	9	+updatedAt: 2026-10-09T15:00:00.000Z
	10	+labels:
	11	+ - enhancement
	12	+epic: Agent Memory
	13	+related:
	14	+ - 79
	15	+body: \|
	16	+ ## Goal
	17	+
	18	+ When mm works in a project that has an `AGENTS.md` file, read it and give its content to the model as project instructions: build/test commands, conventions, what not to touch. Today the only instructions are the config's `system` prompt, the same for every project.
	19	+
	20	+ `AGENTS.md` (https://agents.md) is a shared convention among coding agents: a plain Markdown file at the repository root, no required fields. In a monorepo, nested `AGENTS.md` files can sit in sub-directories; the one closest to the file being worked on takes precedence, and an explicit instruction from the user in the chat wins over the file.
	21	+
	22	+ ## Where it plugs in (2026-10-09)
	23	+
	24	+ - The system message is built in one place: `session.Fresh(system)` (`internal/session/session.go`), called by the line REPL (`internal/agent/agent.go`), `-p` (`RunPrompt`, `RunStructuredPrompt`), the window (`internal/tui/turn.go`, `commands.go` for `/new`), and the ACP façade.
	25	+ - The working directory differs per front end: the process cwd in the terminal, the session's `cwd` in ACP (`ui.Resolve` / the sink's `WorkDir`, as for `sessions.dir` and `trace.dir`).
	26	+ - A `/new`, a `/resume` and a context compression must keep the instructions (they live in the system message, which compression already keeps).
	27	+
	28	+ ## Decided (2026-10-09, with the user)
	29	+
	30	+ How the file joins the prompt:
	31	+
	32	+ - Appended to `system`, after the config's own prompt, under a heading naming its origin — `PROJECT INSTRUCTIONS (from <absolute path>)` — never in place of it: `system` says what the agent is, AGENTS.md what the project is. Not a first user message: small models weigh the system message more, and Gemma/Mistral templates refuse two consecutive user turns. Same mechanism as `structuredOutput`'s `OUTPUT FORMAT` paragraph (ticket 80).
	33	+ - Read at start-up, then again at every `/new` and every new ACP session — never in the middle of a conversation: a system message that changes mid-way breaks the server's prompt cache (llama.cpp re-reads the whole history; see ticket 53). An edited file is picked up by the next conversation.
	34	+ - Still to settle in the implementation: `/resume` — keep the stored system message, or replace it with the current file (ACP already has `history.WithSystem` for a reloaded session). Compression already keeps the system message as is.
	35	+ - Size: a warning, not a cut. The banner says the file's estimated size and its share of the context window (`AGENTS.md: ~6200 tokens, 40 % of the 16k window`) when it is large. An optional cap (e.g. `agentsMd.maxChars`, off by default) may be added; beyond it the file is not truncated silently — truncation can drop the one rule that matters, at the end of the file.
	36	+
	37	+ Answered by the user in comment 1 (2026-10-09):
	38	+
	39	+ - Only `AGENTS.md` — no `CLAUDE.md` / `.github/copilot-instructions.md` fallback.
	40	+ - The working directory only — no walk up to the git root, no nested files concatenated. (Working directory = the process cwd in the terminal and `-p`, the session's `cwd` over ACP.)
	41	+ - A switch, on by default, with a path override.
	42	+ - An asked agent (`ask_agent`, `/ask`, `mm --ask`) reads it only if ITS OWN config has the switch on — so, the switch being on by default, an asked agent reads the AGENTS.md of the working directory it is started in unless its file says `false`.
	43	+ - Skipped for structured agents (`structuredOutput` set, ticket 80): prose instructions would fight the JSON constraint.
	44	+
	45	+ Settled with the user on 2026-10-09 (the last open points):
	46	+
	47	+ - Config shape: a block, like `trace` and `sessions` —
	48	+ `agentsMd: {enabled: true, path: AGENTS.md}` (enabled by default). A relative `path` is resolved against the WORKING DIRECTORY (the session's cwd over ACP), like `sessions.dir` and `trace.dir` — not against the config file: one `agent.yaml` serves many projects. The optional cap, if added, goes in the same block (`agentsMd.maxChars`, 0 = none).
	49	+ - `/resume` re-reads the file and the resumed conversation continues under the current one — the rule already in force: `session.WithSystem` (`internal/session/session.go`) replaces a stored system message with the agent's current prompt, in `store/keeper.go` (terminal, window) and `acp/persist.go` (session/load). No cache cost: a resumed session is re-read whole by the server anyway.
	50	+ - Beyond the optional cap: the file is SKIPPED, with a banner warning (`AGENTS.md skipped: <n> characters > agentsMd.maxChars <cap>`) — never truncated: a missing file is a visible state, a half-read one hides the rule that was cut.
	51	+ - A structured agent that skips an existing AGENTS.md says so neutrally in the banner (`agents.md: ignored (structured output)`, only when a file exists), not as a warning — the same spirit as `tools: none (structured output: JSON answers)`. No banner in `-p` / `--ask`, so nothing there.
	52	+
	53	+ No question left open.
	54	+
	55	+ ## Implemented (2026-10-09, branch `feature/agents-file`)
	56	+
	57	+ - `internal/agentsmd`: `Load(workDir)` → Disabled / Missing / Loaded / Skipped / Ignored / Unreadable; `System(base, workDir)`.
	58	+ - Wired wherever a conversation starts: line mode (start, /new, /resume, -p), window (start, /new, /resume), ACP (session/new with its cwd, /new, session/load).
	59	+ - Banner column `agents.md: …` + warnings (≥ 20 % of a known window, skipped, unreadable) — `agentsmd_banner.go`.
	60	+ - Live on llmman (Qwen3.5-4B): the model followed a test AGENTS.md (`make check`, a signature); an agent asked with `--ask` read it with its switch on and not with `enabled: false`; a structured agent ignored it and its JSON stayed clean.
	61	+ - Task 6 (quality gate) open: qlty is not installed in the sandbox.
	62	+tasks:
	63	+ - id: 1
	64	+ title: Decide the open questions below with the user (which files, where to look, how it joins the system prompt, a switch to turn it off)
	65	+ state: closed
	66	+ priority: none
	67	+ author:
	68	+ name: k33g
	69	+ email: ph.charriere@gmail.com
	70	+ createdAt: 2026-10-09T08:00:00.000Z
	71	+ updatedAt: 2026-10-09T15:00:00.000Z
	72	+ closedAt: 2026-10-09T15:00:00.000Z
	73	+ - id: 2
	74	+ title: "Failing tests first: an AGENTS.md in the working directory reaches the system message; none = system prompt unchanged"
	75	+ state: closed
	76	+ priority: none
	77	+ author:
	78	+ name: k33g
	79	+ email: ph.charriere@gmail.com
	80	+ createdAt: 2026-10-09T08:00:00.000Z
	81	+ updatedAt: 2026-10-09T15:00:00.000Z
	82	+ closedAt: 2026-10-09T15:00:00.000Z
	83	+ - id: 3
	84	+ title: Read AGENTS.md from the session's working directory in every front end (terminal, window, -p, ACP session cwd, an agent asked through ask_agent / --ask)
	85	+ state: closed
	86	+ priority: none
	87	+ author:
	88	+ name: k33g
	89	+ email: ph.charriere@gmail.com
	90	+ createdAt: 2026-10-09T08:00:00.000Z
	91	+ updatedAt: 2026-10-09T15:00:00.000Z
	92	+ closedAt: 2026-10-09T15:00:00.000Z
	93	+ - id: 4
	94	+ title: "Banner / ACP: say which instruction file(s) were loaded, and warn when one is too big"
	95	+ state: closed
	96	+ priority: none
	97	+ author:
	98	+ name: k33g
	99	+ email: ph.charriere@gmail.com
	100	+ createdAt: 2026-10-09T08:00:00.000Z
	101	+ updatedAt: 2026-10-09T15:00:00.000Z
	102	+ closedAt: 2026-10-09T15:00:00.000Z
	103	+ - id: 5
	104	+ title: Live check on llmman with a demo project holding an AGENTS.md
	105	+ state: closed
	106	+ priority: none
	107	+ author:
	108	+ name: k33g
	109	+ email: ph.charriere@gmail.com
	110	+ createdAt: 2026-10-09T08:00:00.000Z
	111	+ updatedAt: 2026-10-09T15:00:00.000Z
	112	+ closedAt: 2026-10-09T15:00:00.000Z
	113	+ - id: 6
	114	+ title: Quality gate (quality skill)
	115	+ state: open
	116	+ priority: none
	117	+ author:
	118	+ name: k33g
	119	+ email: ph.charriere@gmail.com
	120	+ createdAt: 2026-10-09T08:00:00.000Z
	121	+ - id: 7
	122	+ title: Docs EN+FR (how-to + reference/configuration + explanation) and .memory update
	123	+ state: closed
	124	+ priority: none
	125	+ author:
	126	+ name: k33g
	127	+ email: ph.charriere@gmail.com
	128	+ createdAt: 2026-10-09T08:00:00.000Z
	129	+ updatedAt: 2026-10-09T15:00:00.000Z
	130	+ closedAt: 2026-10-09T15:00:00.000Z
	131	+comments:
	132	+ - id: 1
	133	+ author:
	134	+ name: k33g
	135	+ email: ph.charriere@gmail.com
	136	+ createdAt: 2026-10-09T03:27:47.320Z
	137	+ body: \|+
	138	+ ## Answers to open questions:
	139	+
	140	+ - Which files: `AGENTS.md` only, or also `CLAUDE.md` / `.github/copilot-instructions.md` as fallbacks?
	141	+ - Only `AGENTS.md`
	142	+
	143	+ - Where to look: the working directory only, or walk up to the git root and concatenate root → cwd (nested files)?
	144	+ - the working directory only
	145	+
	146	+ - A switch: `agentsMd: true\|false` in the config (on by default?), and a path override?
	147	+ - yes and yes and yes
	148	+
	149	+ - An agent asked through `ask_agent` / `mm --ask`: does it read the AGENTS.md of its asker's project (same cwd), or none?
	150	+ - Only if agentsMd: true in the asked agent config file
	151	+
	152	+ - Interaction with `structuredOutput` (ticket 80): an instruction file asking for prose would fight the JSON constraint — skip it for structured agents?
	153	+ - skip it for structured agents
	154	+
	155	+

++# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
++id: 81
++title: Support AGENTS.md (project instructions for the agent)
++state: open
++author:
++  name: k33g
++  email: ph.charriere@gmail.com
++createdAt: 2026-10-09T08:00:00.000Z
++updatedAt: 2026-10-09T15:00:00.000Z
++labels:
++  - enhancement
++epic: Agent Memory
++related:
++  - 79
++body: |
++  ## Goal
++
++  When mm works in a project that has an `AGENTS.md` file, read it and give its content to the model as project instructions: build/test commands, conventions, what not to touch. Today the only instructions are the config's `system` prompt, the same for every project.
++
++  `AGENTS.md` (https://agents.md) is a shared convention among coding agents: a plain Markdown file at the repository root, no required fields. In a monorepo, nested `AGENTS.md` files can sit in sub-directories; the one closest to the file being worked on takes precedence, and an explicit instruction from the user in the chat wins over the file.
++
++  ## Where it plugs in (2026-10-09)
++
++  - The system message is built in one place: `session.Fresh(system)` (`internal/session/session.go`), called by the line REPL (`internal/agent/agent.go`), `-p` (`RunPrompt`, `RunStructuredPrompt`), the window (`internal/tui/turn.go`, `commands.go` for `/new`), and the ACP façade.
++  - The working directory differs per front end: the process cwd in the terminal, the session's `cwd` in ACP (`ui.Resolve` / the sink's `WorkDir`, as for `sessions.dir` and `trace.dir`).
++  - A `/new`, a `/resume` and a context compression must keep the instructions (they live in the system message, which compression already keeps).
++
++  ## Decided (2026-10-09, with the user)
++
++  How the file joins the prompt:
++
++  - **Appended to `system`**, after the config's own prompt, under a heading naming its origin — `PROJECT INSTRUCTIONS (from <absolute path>)` — never in place of it: `system` says what the agent is, AGENTS.md what the project is. Not a first user message: small models weigh the system message more, and Gemma/Mistral templates refuse two consecutive user turns. Same mechanism as `structuredOutput`'s `OUTPUT FORMAT` paragraph (ticket 80).
++  - **Read at start-up, then again at every `/new` and every new ACP session** — never in the middle of a conversation: a system message that changes mid-way breaks the server's prompt cache (llama.cpp re-reads the whole history; see ticket 53). An edited file is picked up by the next conversation.
++  - Still to settle in the implementation: `/resume` — keep the stored system message, or replace it with the current file (ACP already has `history.WithSystem` for a reloaded session). Compression already keeps the system message as is.
++  - **Size: a warning, not a cut.** The banner says the file's estimated size and its share of the context window (`AGENTS.md: ~6200 tokens, 40 % of the 16k window`) when it is large. An optional cap (e.g. `agentsMd.maxChars`, off by default) may be added; beyond it the file is not truncated silently — truncation can drop the one rule that matters, at the end of the file.
++
++  Answered by the user in comment 1 (2026-10-09):
++
++  - **Only `AGENTS.md`** — no `CLAUDE.md` / `.github/copilot-instructions.md` fallback.
++  - **The working directory only** — no walk up to the git root, no nested files concatenated. (Working directory = the process cwd in the terminal and `-p`, the session's `cwd` over ACP.)
++  - **A switch, on by default, with a path override.**
++  - **An asked agent** (`ask_agent`, `/ask`, `mm --ask`) reads it only if ITS OWN config has the switch on — so, the switch being on by default, an asked agent reads the AGENTS.md of the working directory it is started in unless its file says `false`.
++  - **Skipped for structured agents** (`structuredOutput` set, ticket 80): prose instructions would fight the JSON constraint.
++
++  Settled with the user on 2026-10-09 (the last open points):
++
++  - **Config shape: a block**, like `trace` and `sessions` —
++    `agentsMd: {enabled: true, path: AGENTS.md}` (enabled by default). A relative `path` is resolved against the WORKING DIRECTORY (the session's cwd over ACP), like `sessions.dir` and `trace.dir` — not against the config file: one `agent.yaml` serves many projects. The optional cap, if added, goes in the same block (`agentsMd.maxChars`, 0 = none).
++  - **`/resume` re-reads the file** and the resumed conversation continues under the current one — the rule already in force: `session.WithSystem` (`internal/session/session.go`) replaces a stored system message with the agent's current prompt, in `store/keeper.go` (terminal, window) and `acp/persist.go` (session/load). No cache cost: a resumed session is re-read whole by the server anyway.
++  - **Beyond the optional cap: the file is SKIPPED**, with a banner warning (`AGENTS.md skipped: <n> characters > agentsMd.maxChars <cap>`) — never truncated: a missing file is a visible state, a half-read one hides the rule that was cut.
++  - **A structured agent that skips an existing AGENTS.md says so neutrally in the banner** (`agents.md: ignored (structured output)`, only when a file exists), not as a warning — the same spirit as `tools: none (structured output: JSON answers)`. No banner in `-p` / `--ask`, so nothing there.
++
++  No question left open.
++
++  ## Implemented (2026-10-09, branch `feature/agents-file`)
++
++  - `internal/agentsmd`: `Load(workDir)` → Disabled / Missing / Loaded / Skipped / Ignored / Unreadable; `System(base, workDir)`.
++  - Wired wherever a conversation starts: line mode (start, /new, /resume, -p), window (start, /new, /resume), ACP (session/new with its cwd, /new, session/load).
++  - Banner column `agents.md: …` + warnings (≥ 20 % of a known window, skipped, unreadable) — `agentsmd_banner.go`.
++  - Live on llmman (Qwen3.5-4B): the model followed a test AGENTS.md (`make check`, a signature); an agent asked with `--ask` read it with its switch on and not with `enabled: false`; a structured agent ignored it and its JSON stayed clean.
++  - Task 6 (quality gate) open: qlty is not installed in the sandbox.
++tasks:
++  - id: 1
++    title: Decide the open questions below with the user (which files, where to look, how it joins the system prompt, a switch to turn it off)
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-09T08:00:00.000Z
++    updatedAt: 2026-10-09T15:00:00.000Z
++    closedAt: 2026-10-09T15:00:00.000Z
++  - id: 2
++    title: "Failing tests first: an AGENTS.md in the working directory reaches the system message; none = system prompt unchanged"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-09T08:00:00.000Z
++    updatedAt: 2026-10-09T15:00:00.000Z
++    closedAt: 2026-10-09T15:00:00.000Z
++  - id: 3
++    title: Read AGENTS.md from the session's working directory in every front end (terminal, window, -p, ACP session cwd, an agent asked through ask_agent / --ask)
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-09T08:00:00.000Z
++    updatedAt: 2026-10-09T15:00:00.000Z
++    closedAt: 2026-10-09T15:00:00.000Z
++  - id: 4
++    title: "Banner / ACP: say which instruction file(s) were loaded, and warn when one is too big"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-09T08:00:00.000Z
++    updatedAt: 2026-10-09T15:00:00.000Z
++    closedAt: 2026-10-09T15:00:00.000Z
++  - id: 5
++    title: Live check on llmman with a demo project holding an AGENTS.md
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-09T08:00:00.000Z
++    updatedAt: 2026-10-09T15:00:00.000Z
++    closedAt: 2026-10-09T15:00:00.000Z
++  - id: 6
++    title: Quality gate (quality skill)
++    state: open
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-09T08:00:00.000Z
++  - id: 7
++    title: Docs EN+FR (how-to + reference/configuration + explanation) and .memory update
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-09T08:00:00.000Z
++    updatedAt: 2026-10-09T15:00:00.000Z
++    closedAt: 2026-10-09T15:00:00.000Z
++comments:
++  - id: 1
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-09T03:27:47.320Z
++    body: |+
++      ## Answers to open questions:
++
++      - Which files: `AGENTS.md` only, or also `CLAUDE.md` / `.github/copilot-instructions.md` as fallbacks?
++        - Only `AGENTS.md`
++
++      - Where to look: the working directory only, or walk up to the git root and concatenate root → cwd (nested files)?
++        - the working directory only
++
++      - A switch: `agentsMd: true|false` in the config (on by default?), and a path override?
++        - yes and yes and yes
++
++      - An agent asked through `ask_agent` / `mm --ask`: does it read the AGENTS.md of its asker's project (same cwd), or none?
++        - Only if agentsMd: true in the asked agent config file
++
++      - Interaction with `structuredOutput` (ticket 80): an instruction file asking for prose would fight the JSON constraint — skip it for structured agents?
++        - skip it for structured agents
++
++

added agentsmd_banner.go +57 -0 Viewed

new file mode 100644
@@ -0,0 +1,57 @@
	1	+package main
	2	+
	3	+import (
	4	+ "fmt"
	5	+
	6	+ "mm/internal/agentsmd"
	7	+ "mm/internal/config"
	8	+)
	9	+
	10	+// largeAgentsMdShare is the share of the context window (in percent) from
	11	+// which the banner warns about the size of AGENTS.md: the file rides in the
	12	+// system message of every request, for the whole conversation.
	13	+const largeAgentsMdShare = 20
	14	+
	15	+// agentsMdBanner returns the banner's `agents.md:` column (with its leading
	16	+// " \| ", or "" when there is nothing to say) and the warnings about the file.
	17	+// window is the context window in tokens, 0 when unknown. In ACP mode each
	18	+// session reads the file of its own cwd, which no session has given yet: the
	19	+// column only says where it will be looked for.
	20	+//
	21	+// col, warnings := agentsMdBanner(false, 16384)
	22	+// // " \| agents.md: /home/me/project/AGENTS.md (~350 tokens)", nil
	23	+func agentsMdBanner(acpMode bool, window int) (string, []string) {
	24	+ settings := config.Cfg.AgentsMd
	25	+ if !settings.Enabled {
	26	+ return "", nil
	27	+ }
	28	+ if acpMode {
	29	+ return " \| agents.md: <session cwd>/" + settings.Path, nil
	30	+ }
	31	+ f := agentsmd.Load("")
	32	+ switch f.Status {
	33	+ case agentsmd.Loaded:
	34	+ return fmt.Sprintf(" \| agents.md: %s (~%d tokens)", f.Path, f.Tokens()), largeAgentsMdWarning(f, window)
	35	+ case agentsmd.Ignored:
	36	+ // Neutral, not a warning: a structured agent skips it on purpose.
	37	+ return " \| agents.md: ignored (structured output)", nil
	38	+ case agentsmd.Skipped:
	39	+ return " \| agents.md: skipped (too long)", []string{fmt.Sprintf(
	40	+ "%s skipped: %d characters > agentsMd.maxChars %d — it is not read at all (shorten it, or raise the limit)",
	41	+ f.Path, f.Chars(), settings.MaxChars)}
	42	+ case agentsmd.Unreadable:
	43	+ return "", []string{fmt.Sprintf("%s cannot be read: %v — no project instructions", f.Path, f.Err)}
	44	+ }
	45	+ return "", nil
	46	+}
	47	+
	48	+// largeAgentsMdWarning says when the file takes a large share of a known
	49	+// context window. Nothing is cut: the user decides (agentsMd.maxChars).
	50	+func largeAgentsMdWarning(f agentsmd.File, window int) []string {
	51	+ if window <= 0 \|\| f.Tokens()100 < windowlargeAgentsMdShare {
	52	+ return nil
	53	+ }
	54	+ return []string{fmt.Sprintf(
	55	+ "%s is ~%d tokens, %d %% of the %d-token context window — every request carries it (shorten it, or set agentsMd.maxChars)",
	56	+ f.Path, f.Tokens(), f.Tokens()*100/window, window)}
	57	+}

++package main
++
++import (
++	"fmt"
++
++	"mm/internal/agentsmd"
++	"mm/internal/config"
++)
++
++// largeAgentsMdShare is the share of the context window (in percent) from
++// which the banner warns about the size of AGENTS.md: the file rides in the
++// system message of every request, for the whole conversation.
++const largeAgentsMdShare = 20
++
++// agentsMdBanner returns the banner's `agents.md:` column (with its leading
++// " | ", or "" when there is nothing to say) and the warnings about the file.
++// window is the context window in tokens, 0 when unknown. In ACP mode each
++// session reads the file of its own cwd, which no session has given yet: the
++// column only says where it will be looked for.
++//
++//	col, warnings := agentsMdBanner(false, 16384)
++//	// " | agents.md: /home/me/project/AGENTS.md (~350 tokens)", nil
++func agentsMdBanner(acpMode bool, window int) (string, []string) {
++	settings := config.Cfg.AgentsMd
++	if !settings.Enabled {
++		return "", nil
++	}
++	if acpMode {
++		return " | agents.md: <session cwd>/" + settings.Path, nil
++	}
++	f := agentsmd.Load("")
++	switch f.Status {
++	case agentsmd.Loaded:
++		return fmt.Sprintf(" | agents.md: %s (~%d tokens)", f.Path, f.Tokens()), largeAgentsMdWarning(f, window)
++	case agentsmd.Ignored:
++		// Neutral, not a warning: a structured agent skips it on purpose.
++		return " | agents.md: ignored (structured output)", nil
++	case agentsmd.Skipped:
++		return " | agents.md: skipped (too long)", []string{fmt.Sprintf(
++			"%s skipped: %d characters > agentsMd.maxChars %d — it is not read at all (shorten it, or raise the limit)",
++			f.Path, f.Chars(), settings.MaxChars)}
++	case agentsmd.Unreadable:
++		return "", []string{fmt.Sprintf("%s cannot be read: %v — no project instructions", f.Path, f.Err)}
++	}
++	return "", nil
++}
++
++// largeAgentsMdWarning says when the file takes a large share of a known
++// context window. Nothing is cut: the user decides (agentsMd.maxChars).
++func largeAgentsMdWarning(f agentsmd.File, window int) []string {
++	if window <= 0 || f.Tokens()*100 < window*largeAgentsMdShare {
++		return nil
++	}
++	return []string{fmt.Sprintf(
++		"%s is ~%d tokens, %d %% of the %d-token context window — every request carries it (shorten it, or set agentsMd.maxChars)",
++		f.Path, f.Tokens(), f.Tokens()*100/window, window)}
++}

added agentsmd_banner_test.go +89 -0 Viewed

new file mode 100644
@@ -0,0 +1,89 @@
	1	+package main
	2	+
	3	+import (
	4	+ "os"
	5	+ "path/filepath"
	6	+ "strings"
	7	+ "testing"
	8	+
	9	+ "mm/internal/config"
	10	+)
	11	+
	12	+// inProject runs the test in a fresh directory holding an AGENTS.md of text
	13	+// ("" = no file), with the given settings.
	14	+func inProject(t *testing.T, text string, settings config.AgentsMdConfig) string {
	15	+ t.Helper()
	16	+ saved, savedSchema := config.Cfg.AgentsMd, config.Cfg.OutputSchema
	17	+ config.Cfg.AgentsMd, config.Cfg.OutputSchema = settings, nil
	18	+ t.Cleanup(func() { config.Cfg.AgentsMd, config.Cfg.OutputSchema = saved, savedSchema })
	19	+ dir := t.TempDir()
	20	+ t.Chdir(dir)
	21	+ if text != "" {
	22	+ if err := os.WriteFile("AGENTS.md", []byte(text), 0o644); err != nil {
	23	+ t.Fatal(err)
	24	+ }
	25	+ }
	26	+ abs, _ := filepath.Abs(filepath.Join(dir, "AGENTS.md"))
	27	+ return abs
	28	+}
	29	+
	30	+var agentsMdOn = config.AgentsMdConfig{Enabled: true, Path: "AGENTS.md"}
	31	+
	32	+func TestBannerNamesTheLoadedFile(t *testing.T) {
	33	+ path := inProject(t, strings.Repeat("a", 700), agentsMdOn)
	34	+ col, warnings := agentsMdBanner(false, 100000)
	35	+ if col != " \| agents.md: "+path+" (~200 tokens)" \|\| len(warnings) != 0 {
	36	+ t.Errorf("column %q, warnings %v", col, warnings)
	37	+ }
	38	+}
	39	+
	40	+func TestBannerWarnsWhenTheFileFillsTheWindow(t *testing.T) {
	41	+ path := inProject(t, strings.Repeat("a", 700), agentsMdOn) // ~200 tokens
	42	+ _, warnings := agentsMdBanner(false, 1000) // 20 %
	43	+ if len(warnings) != 1 \|\| !strings.Contains(warnings[0], path+" is ~200 tokens, 20 % of the 1000-token context window") {
	44	+ t.Errorf("warnings %v", warnings)
	45	+ }
	46	+ if _, warnings := agentsMdBanner(false, 1001); len(warnings) != 0 {
	47	+ t.Errorf("under 20 %%: %v", warnings)
	48	+ }
	49	+ if _, warnings := agentsMdBanner(false, 0); len(warnings) != 0 {
	50	+ t.Errorf("unknown window: %v", warnings)
	51	+ }
	52	+}
	53	+
	54	+func TestBannerSaysASkippedFileIsNotRead(t *testing.T) {
	55	+ inProject(t, "too long rules", config.AgentsMdConfig{Enabled: true, Path: "AGENTS.md", MaxChars: 5})
	56	+ col, warnings := agentsMdBanner(false, 0)
	57	+ if col != " \| agents.md: skipped (too long)" \|\| len(warnings) != 1 \|\| !strings.Contains(warnings[0], "14 characters > agentsMd.maxChars 5") {
	58	+ t.Errorf("column %q, warnings %v", col, warnings)
	59	+ }
	60	+}
	61	+
	62	+// A structured agent: a neutral mention, no warning — and only when a file
	63	+// exists.
	64	+func TestBannerMentionsAFileAStructuredAgentIgnores(t *testing.T) {
	65	+ inProject(t, "prose rules", agentsMdOn)
	66	+ config.Cfg.OutputSchema = map[string]any{"type": "object"}
	67	+ col, warnings := agentsMdBanner(false, 0)
	68	+ if col != " \| agents.md: ignored (structured output)" \|\| len(warnings) != 0 {
	69	+ t.Errorf("column %q, warnings %v", col, warnings)
	70	+ }
	71	+}
	72	+
	73	+func TestBannerSaysNothingWithoutAFileOrWhenOff(t *testing.T) {
	74	+ inProject(t, "", agentsMdOn)
	75	+ if col, warnings := agentsMdBanner(false, 0); col != "" \|\| warnings != nil {
	76	+ t.Errorf("no file: column %q, warnings %v", col, warnings)
	77	+ }
	78	+ inProject(t, "rules", config.AgentsMdConfig{Enabled: false, Path: "AGENTS.md"})
	79	+ if col, warnings := agentsMdBanner(false, 0); col != "" \|\| warnings != nil {
	80	+ t.Errorf("disabled: column %q, warnings %v", col, warnings)
	81	+ }
	82	+}
	83	+
	84	+func TestBannerInACPModeSaysWhereEachSessionLooks(t *testing.T) {
	85	+ inProject(t, "", agentsMdOn)
	86	+ if col, _ := agentsMdBanner(true, 0); col != " \| agents.md: <session cwd>/AGENTS.md" {
	87	+ t.Errorf("column %q", col)
	88	+ }
	89	+}

++package main
++
++import (
++	"os"
++	"path/filepath"
++	"strings"
++	"testing"
++
++	"mm/internal/config"
++)
++
++// inProject runs the test in a fresh directory holding an AGENTS.md of text
++// ("" = no file), with the given settings.
++func inProject(t *testing.T, text string, settings config.AgentsMdConfig) string {
++	t.Helper()
++	saved, savedSchema := config.Cfg.AgentsMd, config.Cfg.OutputSchema
++	config.Cfg.AgentsMd, config.Cfg.OutputSchema = settings, nil
++	t.Cleanup(func() { config.Cfg.AgentsMd, config.Cfg.OutputSchema = saved, savedSchema })
++	dir := t.TempDir()
++	t.Chdir(dir)
++	if text != "" {
++		if err := os.WriteFile("AGENTS.md", []byte(text), 0o644); err != nil {
++			t.Fatal(err)
++		}
++	}
++	abs, _ := filepath.Abs(filepath.Join(dir, "AGENTS.md"))
++	return abs
++}
++
++var agentsMdOn = config.AgentsMdConfig{Enabled: true, Path: "AGENTS.md"}
++
++func TestBannerNamesTheLoadedFile(t *testing.T) {
++	path := inProject(t, strings.Repeat("a", 700), agentsMdOn)
++	col, warnings := agentsMdBanner(false, 100000)
++	if col != " | agents.md: "+path+" (~200 tokens)" || len(warnings) != 0 {
++		t.Errorf("column %q, warnings %v", col, warnings)
++	}
++}
++
++func TestBannerWarnsWhenTheFileFillsTheWindow(t *testing.T) {
++	path := inProject(t, strings.Repeat("a", 700), agentsMdOn) // ~200 tokens
++	_, warnings := agentsMdBanner(false, 1000)                 // 20 %
++	if len(warnings) != 1 || !strings.Contains(warnings[0], path+" is ~200 tokens, 20 % of the 1000-token context window") {
++		t.Errorf("warnings %v", warnings)
++	}
++	if _, warnings := agentsMdBanner(false, 1001); len(warnings) != 0 {
++		t.Errorf("under 20 %%: %v", warnings)
++	}
++	if _, warnings := agentsMdBanner(false, 0); len(warnings) != 0 {
++		t.Errorf("unknown window: %v", warnings)
++	}
++}
++
++func TestBannerSaysASkippedFileIsNotRead(t *testing.T) {
++	inProject(t, "too long rules", config.AgentsMdConfig{Enabled: true, Path: "AGENTS.md", MaxChars: 5})
++	col, warnings := agentsMdBanner(false, 0)
++	if col != " | agents.md: skipped (too long)" || len(warnings) != 1 || !strings.Contains(warnings[0], "14 characters > agentsMd.maxChars 5") {
++		t.Errorf("column %q, warnings %v", col, warnings)
++	}
++}
++
++// A structured agent: a neutral mention, no warning — and only when a file
++// exists.
++func TestBannerMentionsAFileAStructuredAgentIgnores(t *testing.T) {
++	inProject(t, "prose rules", agentsMdOn)
++	config.Cfg.OutputSchema = map[string]any{"type": "object"}
++	col, warnings := agentsMdBanner(false, 0)
++	if col != " | agents.md: ignored (structured output)" || len(warnings) != 0 {
++		t.Errorf("column %q, warnings %v", col, warnings)
++	}
++}
++
++func TestBannerSaysNothingWithoutAFileOrWhenOff(t *testing.T) {
++	inProject(t, "", agentsMdOn)
++	if col, warnings := agentsMdBanner(false, 0); col != "" || warnings != nil {
++		t.Errorf("no file: column %q, warnings %v", col, warnings)
++	}
++	inProject(t, "rules", config.AgentsMdConfig{Enabled: false, Path: "AGENTS.md"})
++	if col, warnings := agentsMdBanner(false, 0); col != "" || warnings != nil {
++		t.Errorf("disabled: column %q, warnings %v", col, warnings)
++	}
++}
++
++func TestBannerInACPModeSaysWhereEachSessionLooks(t *testing.T) {
++	inProject(t, "", agentsMdOn)
++	if col, _ := agentsMdBanner(true, 0); col != " | agents.md: <session cwd>/AGENTS.md" {
++		t.Errorf("column %q", col)
++	}
++}

added demo-agents-file/README.en.md +159 -0 Viewed

new file mode 100644
@@ -0,0 +1,159 @@
	1	+# mini-me demo: the project's instructions (`AGENTS.md`)
	2	+
	3	+This directory shows how mini-me reads the `AGENTS.md` file of the project it works in and adds it to its system prompt: the agent then follows the project's rules (which command runs the tests, what not to touch…) without any change to its configuration. The models are served by [llmman](https://github.com/llmmanorg/llmman). Version française : [README.md](README.md).
	4	+
	5	+## What is here
	6	+
	7	+\| File \| Role \|
	8	+\|---\|---\|
	9	+\| `project/` \| A small "project", Tangerine: the directory mm is started in. \|
	10	+\| `project/AGENTS.md` \| The project's instructions: tests run with `make check` (never `go test`), do not touch `vendor/` or `generated/`, sign every answer `-- Tangerine bot`. \|
	11	+\| `project/Makefile`, `project/greet.sh` \| What `make check` runs. \|
	12	+\| `agent.yaml` \| Bob, a coding agent (`bash`, file tools, `ask_agent`). He reads `AGENTS.md` (the `agentsMd` setting is written out in full, with its default values). \|
	13	+\| `agent.helper.yaml` \| helper, asked by Bob: no tools, reads `AGENTS.md` (the default). \|
	14	+\| `agent.reviewer.yaml` \| reviewer, asked by Bob: no tools, does not read `AGENTS.md` (`agentsMd.enabled: false`). \|
	15	+
	16	+All three agents use `hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M`.
	17	+
	18	+## Prerequisites
	19	+
	20	+1. llmman running (`llmman serve`, on `127.0.0.1:17434`; from a Docker Sandbox, mm falls back to `host.docker.internal:17434` by itself), and the model pulled: `llmman pull hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M`.
	21	+2. `mm` built at the repository root (`go build -o mm .`).
	22	+3. Start mm from `project/`: `AGENTS.md` is looked for in the working directory, not next to the configuration file.
	23	+
	24	+ ```bash
	25	+ cd demo-agents-file/project
	26	+ ```
	27	+
	28	+## The setting
	29	+
	30	+In `agent.yaml` (these are the defaults: a config that never mentions `agentsMd` reads `./AGENTS.md` all the same):
	31	+
	32	+```yaml
	33	+agentsMd:
	34	+ enabled: true # read at the start of every conversation: start-up, /new, /resume
	35	+ path: AGENTS.md # relative to the working directory (the project), not to this file
	36	+ maxChars: 0 # 0 = no limit; above it the file is skipped whole, never cut
	37	+```
	38	+
	39	+The model receives its `system` prompt, then:
	40	+
	41	+```text
	42	+PROJECT INSTRUCTIONS (from /…/demo-agents-file/project/AGENTS.md)
	43	+The instructions of the project you work in. Follow them; when the user asks otherwise in the conversation, the user wins.
	44	+
	45	+# Project "Tangerine" — instructions for coding agents
	46	+…
	47	+```
	48	+
	49	+## 1. The banner says which file is read
	50	+
	51	+```bash
	52	+../../mm ../agent.yaml
	53	+```
	54	+
	55	+```
	56	+mm … \| tools: bash, read_file, write_file, edit_file, ask_agent \| skills: 0 \| yolo: on \| agents.md: /…/demo-agents-file/project/AGENTS.md (~125 tokens)
	57	+```
	58	+
	59	+(The `no skills found` warning is expected: this demo has no skills.)
	60	+
	61	+## 2. Bob follows the project's instructions
	62	+
	63	+```bash
	64	+../../mm ../agent.yaml -p "Run the tests."
	65	+```
	66	+
	67	+```
	68	+🛠️ bash: make check
	69	+ │ running the Tangerine checks…
	70	+ │ ✅ all checks passed
	71	+🤖 ✅ all checks passed
	72	+```
	73	+
	74	+Bob runs `make check`, the command `AGENTS.md` sets, not a guessed one.
	75	+
	76	+To compare, move the file aside and ask the same thing. `AGENT_YOLO=false` keeps the commands from running: you see what Bob wanted to run.
	77	+
	78	+```bash
	79	+mv AGENTS.md AGENTS.md.off
	80	+AGENT_YOLO=false ../../mm ../agent.yaml -p "Run the tests."
	81	+mv AGENTS.md.off AGENTS.md
	82	+```
	83	+
	84	+```
	85	+🛠️ bash: python -m pytest
	86	+ │ [command rejected by the user]
	87	+🛠️ bash: pytest
	88	+ │ [command rejected by the user]
	89	+…
	90	+```
	91	+
	92	+Without instructions the model guesses (here `pytest`, in a project without a line of Python; in a run with `yolo: true` it went as far as `pip3 install pytest`).
	93	+
	94	+## 3. An asked agent decides with ITS OWN config
	95	+
	96	+When Bob hands a question to another agent (`ask_agent`, `/ask`, `mm --ask`), the asked agent works in the same directory, but its own `agentsMd` setting decides whether it reads the file.
	97	+
	98	+```bash
	99	+../../mm ../agent.yaml --ask helper -p "How do I run the tests in this project?"
	100	+```
	101	+
	102	+```
	103	+To run the tests, execute `make check` as instructed by the project's build and test guidelines. Always quote the last line of the command output in your response.
	104	+-- Tangerine bot
	105	+```
	106	+
	107	+```bash
	108	+../../mm ../agent.yaml --ask reviewer -p "How do I run the tests in this project?"
	109	+```
	110	+
	111	+```
	112	+Run the tests by executing the command specified in your project's documentation, typically `npm test` for Node.js projects or `pytest` for Python projects. …
	113	+```
	114	+
	115	+helper knows `make check` and signs; reviewer (`agentsMd.enabled: false`) gives a generic answer: it never saw the file.
	116	+
	117	+## 4. Editing `AGENTS.md` during a session
	118	+
	119	+The file is read at the start of every conversation, never in the middle of one (changing the system message mid-way would make the server re-read the whole history). To take an edit into account:
	120	+
	121	+```bash
	122	+../../mm ../agent.yaml
	123	+> Run the tests. # make check
	124	+```
	125	+
	126	+Edit `AGENTS.md` in another terminal (for instance: the command becomes `sh greet.sh Tangerine`), then:
	127	+
	128	+```text
	129	+> /new
	130	+🆕 new session: 4 message(s) forgotten
	131	+> Run the tests.
	132	+🛠️ bash: sh greet.sh Tangerine
	133	+```
	134	+
	135	+The new instruction applies from the new conversation on. (In our run Bob then also ran `make check`, despite "Never run make": a 4B model does not honour every prohibition.)
	136	+
	137	+`/resume` also reads the current file.
	138	+
	139	+## 5. Check what the model received
	140	+
	141	+```bash
	142	+AGENT_TRACE=true ../../mm ../agent.yaml -p "Run the tests."
	143	+jq -r '.messages[0].content' .mm/traces/*-0001.request.json
	144	+```
	145	+
	146	+The system message ends with the `PROJECT INSTRUCTIONS (from …)` section.
	147	+
	148	+## Good to know
	149	+
	150	+- A small model does not follow every rule every time: with Qwen3.5-4B, Bob ran `make check` on every try but left out the `-- Tangerine bot` signature; helper, with no tools, added it. Short, concrete instructions work best.
	151	+- Size: the file rides with every request. The banner warns when it takes 20 % or more of the context window; `agentsMd.maxChars` sets a hard limit (above it the file is skipped whole, and the banner says `skipped (too long)`).
	152	+- A JSON agent (`structuredOutput`) never reads `AGENTS.md`; its banner says `agents.md: ignored (structured output)`. See [demo-structured-output](../demo-structured-output/README.en.md).
	153	+- In an editor (Zed…, ACP mode), each session reads the `AGENTS.md` of the project it opened.
	154	+
	155	+## Further reading
	156	+
	157	+- [Give the agent your project's instructions](../docs/en/how-to/give-project-instructions.md)
	158	+- [`agentsMd` reference](../docs/en/reference/configuration.md#agentsmd-keys)
	159	+- [Project instructions — explanation](../docs/en/explanation/project-instructions.md)

++# mini-me demo: the project's instructions (`AGENTS.md`)
++
++This directory shows how mini-me reads the `AGENTS.md` file of the project it works in and adds it to its system prompt: the agent then follows the project's rules (which command runs the tests, what not to touch…) without any change to its configuration. The models are served by [llmman](https://github.com/llmmanorg/llmman). Version française : [README.md](README.md).
++
++## What is here
++
++| File | Role |
++|---|---|
++| `project/` | A small "project", **Tangerine**: the directory mm is started in. |
++| `project/AGENTS.md` | The project's instructions: tests run with `make check` (never `go test`), do not touch `vendor/` or `generated/`, sign every answer `-- Tangerine bot`. |
++| `project/Makefile`, `project/greet.sh` | What `make check` runs. |
++| `agent.yaml` | **Bob**, a coding agent (`bash`, file tools, `ask_agent`). He reads `AGENTS.md` (the `agentsMd` setting is written out in full, with its default values). |
++| `agent.helper.yaml` | **helper**, asked by Bob: no tools, **reads** `AGENTS.md` (the default). |
++| `agent.reviewer.yaml` | **reviewer**, asked by Bob: no tools, does **not** read `AGENTS.md` (`agentsMd.enabled: false`). |
++
++All three agents use `hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M`.
++
++## Prerequisites
++
++1. llmman running (`llmman serve`, on `127.0.0.1:17434`; from a Docker Sandbox, mm falls back to `host.docker.internal:17434` by itself), and the model pulled: `llmman pull hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M`.
++2. `mm` built at the repository root (`go build -o mm .`).
++3. **Start mm from `project/`**: `AGENTS.md` is looked for in the working directory, not next to the configuration file.
++
++   ```bash
++   cd demo-agents-file/project
++   ```
++
++## The setting
++
++In `agent.yaml` (these are the defaults: a config that never mentions `agentsMd` reads `./AGENTS.md` all the same):
++
++```yaml
++agentsMd:
++  enabled: true      # read at the start of every conversation: start-up, /new, /resume
++  path: AGENTS.md    # relative to the working directory (the project), not to this file
++  maxChars: 0        # 0 = no limit; above it the file is skipped whole, never cut
++```
++
++The model receives its `system` prompt, then:
++
++```text
++PROJECT INSTRUCTIONS (from /…/demo-agents-file/project/AGENTS.md)
++The instructions of the project you work in. Follow them; when the user asks otherwise in the conversation, the user wins.
++
++# Project "Tangerine" — instructions for coding agents
++…
++```
++
++## 1. The banner says which file is read
++
++```bash
++../../mm ../agent.yaml
++```
++
++```
++mm … | tools: bash, read_file, write_file, edit_file, ask_agent | skills: 0 | yolo: on | agents.md: /…/demo-agents-file/project/AGENTS.md (~125 tokens)
++```
++
++(The `no skills found` warning is expected: this demo has no skills.)
++
++## 2. Bob follows the project's instructions
++
++```bash
++../../mm ../agent.yaml -p "Run the tests."
++```
++
++```
++🛠️  bash: make check
++   │ running the Tangerine checks…
++   │ ✅ all checks passed
++🤖 ✅ all checks passed
++```
++
++Bob runs `make check`, the command `AGENTS.md` sets, not a guessed one.
++
++**To compare**, move the file aside and ask the same thing. `AGENT_YOLO=false` keeps the commands from running: you see what Bob *wanted* to run.
++
++```bash
++mv AGENTS.md AGENTS.md.off
++AGENT_YOLO=false ../../mm ../agent.yaml -p "Run the tests."
++mv AGENTS.md.off AGENTS.md
++```
++
++```
++🛠️  bash: python -m pytest
++   │ [command rejected by the user]
++🛠️  bash: pytest
++   │ [command rejected by the user]
++…
++```
++
++Without instructions the model guesses (here `pytest`, in a project without a line of Python; in a run with `yolo: true` it went as far as `pip3 install pytest`).
++
++## 3. An asked agent decides with ITS OWN config
++
++When Bob hands a question to another agent (`ask_agent`, `/ask`, `mm --ask`), the asked agent works in the same directory, but **its own** `agentsMd` setting decides whether it reads the file.
++
++```bash
++../../mm ../agent.yaml --ask helper -p "How do I run the tests in this project?"
++```
++
++```
++To run the tests, execute `make check` as instructed by the project's build and test guidelines. Always quote the last line of the command output in your response.
++-- Tangerine bot
++```
++
++```bash
++../../mm ../agent.yaml --ask reviewer -p "How do I run the tests in this project?"
++```
++
++```
++Run the tests by executing the command specified in your project's documentation, typically `npm test` for Node.js projects or `pytest` for Python projects. …
++```
++
++helper knows `make check` and signs; reviewer (`agentsMd.enabled: false`) gives a generic answer: it never saw the file.
++
++## 4. Editing `AGENTS.md` during a session
++
++The file is read **at the start of every conversation**, never in the middle of one (changing the system message mid-way would make the server re-read the whole history). To take an edit into account:
++
++```bash
++../../mm ../agent.yaml
++> Run the tests.            # make check
++```
++
++Edit `AGENTS.md` in another terminal (for instance: the command becomes `sh greet.sh Tangerine`), then:
++
++```text
++> /new
++🆕 new session: 4 message(s) forgotten
++> Run the tests.
++🛠️  bash: sh greet.sh Tangerine
++```
++
++The new instruction applies from the new conversation on. (In our run Bob then also ran `make check`, despite "Never run make": a 4B model does not honour every prohibition.)
++
++`/resume` also reads the current file.
++
++## 5. Check what the model received
++
++```bash
++AGENT_TRACE=true ../../mm ../agent.yaml -p "Run the tests."
++jq -r '.messages[0].content' .mm/traces/*-0001.request.json
++```
++
++The system message ends with the `PROJECT INSTRUCTIONS (from …)` section.
++
++## Good to know
++
++- **A small model does not follow every rule every time**: with Qwen3.5-4B, Bob ran `make check` on every try but left out the `-- Tangerine bot` signature; helper, with no tools, added it. Short, concrete instructions work best.
++- **Size**: the file rides with every request. The banner warns when it takes 20 % or more of the context window; `agentsMd.maxChars` sets a hard limit (above it the file is skipped whole, and the banner says `skipped (too long)`).
++- **A JSON agent** (`structuredOutput`) never reads `AGENTS.md`; its banner says `agents.md: ignored (structured output)`. See [demo-structured-output](../demo-structured-output/README.en.md).
++- **In an editor** (Zed…, ACP mode), each session reads the `AGENTS.md` of the project it opened.
++
++## Further reading
++
++- [Give the agent your project's instructions](../docs/en/how-to/give-project-instructions.md)
++- [`agentsMd` reference](../docs/en/reference/configuration.md#agentsmd-keys)
++- [Project instructions — explanation](../docs/en/explanation/project-instructions.md)

added demo-agents-file/README.md +159 -0 Viewed

new file mode 100644
@@ -0,0 +1,159 @@
	1	+# Démo mini-me : les consignes du projet (`AGENTS.md`)
	2	+
	3	+Ce dossier montre comment mini-me lit le fichier `AGENTS.md` du projet dans lequel il travaille et l'ajoute à son prompt système : l'agent suit alors les règles du projet (quelle commande lance les tests, ce qu'il ne faut pas toucher…) sans qu'on modifie sa configuration. Les modèles sont servis par [llmman](https://github.com/llmmanorg/llmman). English version: [README.en.md](README.en.md).
	4	+
	5	+## Ce qu'il y a ici
	6	+
	7	+\| Fichier \| Rôle \|
	8	+\|---\|---\|
	9	+\| `project/` \| Un petit « projet », Tangerine : c'est le dossier dans lequel on lance mm. \|
	10	+\| `project/AGENTS.md` \| Les consignes du projet : les tests se lancent avec `make check` (jamais `go test`), ne pas toucher `vendor/` ni `generated/`, signer chaque réponse `-- Tangerine bot`. \|
	11	+\| `project/Makefile`, `project/greet.sh` \| De quoi faire tourner `make check`. \|
	12	+\| `agent.yaml` \| Bob, un agent de code (`bash`, outils de fichiers, `ask_agent`). Il lit `AGENTS.md` (le réglage `agentsMd` y est écrit en entier, avec ses valeurs par défaut). \|
	13	+\| `agent.helper.yaml` \| helper, appelé par Bob : sans outils, il lit `AGENTS.md` (réglage par défaut). \|
	14	+\| `agent.reviewer.yaml` \| reviewer, appelé par Bob : sans outils, il ne lit pas `AGENTS.md` (`agentsMd.enabled: false`). \|
	15	+
	16	+Les trois agents utilisent `hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M`.
	17	+
	18	+## Prérequis
	19	+
	20	+1. llmman qui tourne (`llmman serve`, sur `127.0.0.1:17434` ; depuis un Docker Sandbox, mm se replie tout seul sur `host.docker.internal:17434`), et le modèle téléchargé : `llmman pull hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M`.
	21	+2. `mm` compilé à la racine du dépôt (`go build -o mm .`).
	22	+3. Lancez mm depuis `project/` : `AGENTS.md` est cherché dans le répertoire de travail, pas à côté du fichier de configuration.
	23	+
	24	+ ```bash
	25	+ cd demo-agents-file/project
	26	+ ```
	27	+
	28	+## Le réglage
	29	+
	30	+Dans `agent.yaml` (ce sont les valeurs par défaut : une config qui ne mentionne pas `agentsMd` lit `./AGENTS.md` de la même façon) :
	31	+
	32	+```yaml
	33	+agentsMd:
	34	+ enabled: true # lu au début de chaque conversation : démarrage, /new, /resume
	35	+ path: AGENTS.md # relatif au répertoire de travail (le projet), pas à ce fichier
	36	+ maxChars: 0 # 0 = pas de limite ; au-delà, le fichier est ignoré en entier, jamais coupé
	37	+```
	38	+
	39	+Le modèle reçoit son prompt `system`, puis :
	40	+
	41	+```text
	42	+PROJECT INSTRUCTIONS (from /…/demo-agents-file/project/AGENTS.md)
	43	+The instructions of the project you work in. Follow them; when the user asks otherwise in the conversation, the user wins.
	44	+
	45	+# Project "Tangerine" — instructions for coding agents
	46	+…
	47	+```
	48	+
	49	+## 1. La bannière dit quel fichier est lu
	50	+
	51	+```bash
	52	+mm ../agent.yaml
	53	+```
	54	+
	55	+```
	56	+mm … \| tools: bash, read_file, write_file, edit_file, ask_agent \| skills: 0 \| yolo: on \| agents.md: /…/demo-agents-file/project/AGENTS.md (~125 tokens)
	57	+```
	58	+
	59	+(L'avertissement `no skills found` est normal : cette démo n'a pas de skills.)
	60	+
	61	+## 2. Bob suit les consignes du projet
	62	+
	63	+```bash
	64	+mm ../agent.yaml -p "Run the tests."
	65	+```
	66	+
	67	+```
	68	+🛠️ bash: make check
	69	+ │ running the Tangerine checks…
	70	+ │ ✅ all checks passed
	71	+🤖 ✅ all checks passed
	72	+```
	73	+
	74	+Bob lance `make check`, la commande que fixe `AGENTS.md`, et pas une commande devinée.
	75	+
	76	+Pour comparer, mettez le fichier de côté et reposez la même question. `AGENT_YOLO=false` empêche les commandes de s'exécuter : on voit ce que Bob voulait lancer.
	77	+
	78	+```bash
	79	+mv AGENTS.md AGENTS.md.off
	80	+AGENT_YOLO=false ../../mm ../agent.yaml -p "Run the tests."
	81	+mv AGENTS.md.off AGENTS.md
	82	+```
	83	+
	84	+```
	85	+🛠️ bash: python -m pytest
	86	+ │ [command rejected by the user]
	87	+🛠️ bash: pytest
	88	+ │ [command rejected by the user]
	89	+…
	90	+```
	91	+
	92	+Sans consigne, le modèle devine (ici, `pytest`, alors que le projet n'a pas une ligne de Python ; lors d'un essai avec `yolo: true`, il est allé jusqu'à lancer `pip3 install pytest`).
	93	+
	94	+## 3. Un agent appelé décide avec SA propre config
	95	+
	96	+Quand Bob confie une question à un autre agent (`ask_agent`, `/ask`, `mm --ask`), l'agent appelé travaille dans le même répertoire, mais c'est son réglage `agentsMd` qui décide s'il lit le fichier.
	97	+
	98	+```bash
	99	+mm ../agent.yaml --ask helper -p "How do I run the tests in this project?"
	100	+```
	101	+
	102	+```
	103	+To run the tests, execute `make check` as instructed by the project's build and test guidelines. Always quote the last line of the command output in your response.
	104	+-- Tangerine bot
	105	+```
	106	+
	107	+```bash
	108	+mm ../agent.yaml --ask reviewer -p "How do I run the tests in this project?"
	109	+```
	110	+
	111	+```
	112	+Run the tests by executing the command specified in your project's documentation, typically `npm test` for Node.js projects or `pytest` for Python projects. …
	113	+```
	114	+
	115	+helper connaît `make check` et signe ; reviewer (`agentsMd.enabled: false`) donne une réponse générique : il n'a pas vu le fichier.
	116	+
	117	+## 4. Modifier `AGENTS.md` pendant une session
	118	+
	119	+Le fichier est lu au début de chaque conversation, jamais au milieu (changer le message système en cours de route obligerait le serveur à relire tout l'historique). Pour prendre une modification en compte :
	120	+
	121	+```bash
	122	+../../mm ../agent.yaml
	123	+> Run the tests. # make check
	124	+```
	125	+
	126	+Modifiez `AGENTS.md` dans un autre terminal (par exemple : la commande devient `sh greet.sh Tangerine`), puis :
	127	+
	128	+```text
	129	+> /new
	130	+🆕 new session: 4 message(s) forgotten
	131	+> Run the tests.
	132	+🛠️ bash: sh greet.sh Tangerine
	133	+```
	134	+
	135	+La nouvelle consigne s'applique dès la nouvelle conversation. (Lors de notre essai, Bob a ensuite lancé aussi `make check`, malgré « Never run make » : un modèle 4B ne respecte pas chaque interdiction.)
	136	+
	137	+`/resume` relit aussi le fichier actuel.
	138	+
	139	+## 5. Vérifier ce que le modèle a reçu
	140	+
	141	+```bash
	142	+AGENT_TRACE=true ../../mm ../agent.yaml -p "Run the tests."
	143	+jq -r '.messages[0].content' .mm/traces/*-0001.request.json
	144	+```
	145	+
	146	+Le message système se termine par la section `PROJECT INSTRUCTIONS (from …)`.
	147	+
	148	+## Ce qu'il faut savoir
	149	+
	150	+- Un petit modèle ne suit pas toutes les règles à chaque fois : avec Qwen3.5-4B, Bob a lancé `make check` à chaque essai, mais a oublié la signature `-- Tangerine bot` ; helper, sans outils, l'a mise. Des consignes courtes et concrètes passent mieux.
	151	+- Taille : le fichier part avec chaque requête. La bannière avertit quand il occupe 20 % ou plus de la fenêtre de contexte ; `agentsMd.maxChars` fixe une limite ferme (au-delà, il est ignoré en entier, et la bannière l'indique `skipped (too long)`).
	152	+- Un agent JSON (`structuredOutput`) ne lit jamais `AGENTS.md` ; sa bannière indique `agents.md: ignored (structured output)`. Voir [demo-structured-output](../demo-structured-output/README.md).
	153	+- Dans un éditeur (Zed…, mode ACP), chaque session lit le `AGENTS.md` du projet ouvert.
	154	+
	155	+## Pour aller plus loin
	156	+
	157	+- [Donner à l'agent les consignes de votre projet](../docs/fr/how-to/give-project-instructions.md)
	158	+- [Référence de `agentsMd`](../docs/fr/reference/configuration.md#clés-de-agentsmd)
	159	+- [Consignes du projet — explication](../docs/fr/explanation/project-instructions.md)

++# Démo mini-me : les consignes du projet (`AGENTS.md`)
++
++Ce dossier montre comment mini-me lit le fichier `AGENTS.md` du projet dans lequel il travaille et l'ajoute à son prompt système : l'agent suit alors les règles du projet (quelle commande lance les tests, ce qu'il ne faut pas toucher…) sans qu'on modifie sa configuration. Les modèles sont servis par [llmman](https://github.com/llmmanorg/llmman). English version: [README.en.md](README.en.md).
++
++## Ce qu'il y a ici
++
++| Fichier | Rôle |
++|---|---|
++| `project/` | Un petit « projet », **Tangerine** : c'est le dossier dans lequel on lance mm. |
++| `project/AGENTS.md` | Les consignes du projet : les tests se lancent avec `make check` (jamais `go test`), ne pas toucher `vendor/` ni `generated/`, signer chaque réponse `-- Tangerine bot`. |
++| `project/Makefile`, `project/greet.sh` | De quoi faire tourner `make check`. |
++| `agent.yaml` | **Bob**, un agent de code (`bash`, outils de fichiers, `ask_agent`). Il lit `AGENTS.md` (le réglage `agentsMd` y est écrit en entier, avec ses valeurs par défaut). |
++| `agent.helper.yaml` | **helper**, appelé par Bob : sans outils, il **lit** `AGENTS.md` (réglage par défaut). |
++| `agent.reviewer.yaml` | **reviewer**, appelé par Bob : sans outils, il **ne lit pas** `AGENTS.md` (`agentsMd.enabled: false`). |
++
++Les trois agents utilisent `hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M`.
++
++## Prérequis
++
++1. llmman qui tourne (`llmman serve`, sur `127.0.0.1:17434` ; depuis un Docker Sandbox, mm se replie tout seul sur `host.docker.internal:17434`), et le modèle téléchargé : `llmman pull hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M`.
++2. `mm` compilé à la racine du dépôt (`go build -o mm .`).
++3. **Lancez mm depuis `project/`** : `AGENTS.md` est cherché dans le répertoire de travail, pas à côté du fichier de configuration.
++
++   ```bash
++   cd demo-agents-file/project
++   ```
++
++## Le réglage
++
++Dans `agent.yaml` (ce sont les valeurs par défaut : une config qui ne mentionne pas `agentsMd` lit `./AGENTS.md` de la même façon) :
++
++```yaml
++agentsMd:
++  enabled: true      # lu au début de chaque conversation : démarrage, /new, /resume
++  path: AGENTS.md    # relatif au répertoire de travail (le projet), pas à ce fichier
++  maxChars: 0        # 0 = pas de limite ; au-delà, le fichier est ignoré en entier, jamais coupé
++```
++
++Le modèle reçoit son prompt `system`, puis :
++
++```text
++PROJECT INSTRUCTIONS (from /…/demo-agents-file/project/AGENTS.md)
++The instructions of the project you work in. Follow them; when the user asks otherwise in the conversation, the user wins.
++
++# Project "Tangerine" — instructions for coding agents
++…
++```
++
++## 1. La bannière dit quel fichier est lu
++
++```bash
++mm ../agent.yaml
++```
++
++```
++mm … | tools: bash, read_file, write_file, edit_file, ask_agent | skills: 0 | yolo: on | agents.md: /…/demo-agents-file/project/AGENTS.md (~125 tokens)
++```
++
++(L'avertissement `no skills found` est normal : cette démo n'a pas de skills.)
++
++## 2. Bob suit les consignes du projet
++
++```bash
++mm ../agent.yaml -p "Run the tests."
++```
++
++```
++🛠️  bash: make check
++   │ running the Tangerine checks…
++   │ ✅ all checks passed
++🤖 ✅ all checks passed
++```
++
++Bob lance `make check`, la commande que fixe `AGENTS.md`, et pas une commande devinée.
++
++**Pour comparer**, mettez le fichier de côté et reposez la même question. `AGENT_YOLO=false` empêche les commandes de s'exécuter : on voit ce que Bob *voulait* lancer.
++
++```bash
++mv AGENTS.md AGENTS.md.off
++AGENT_YOLO=false ../../mm ../agent.yaml -p "Run the tests."
++mv AGENTS.md.off AGENTS.md
++```
++
++```
++🛠️  bash: python -m pytest
++   │ [command rejected by the user]
++🛠️  bash: pytest
++   │ [command rejected by the user]
++…
++```
++
++Sans consigne, le modèle devine (ici, `pytest`, alors que le projet n'a pas une ligne de Python ; lors d'un essai avec `yolo: true`, il est allé jusqu'à lancer `pip3 install pytest`).
++
++## 3. Un agent appelé décide avec SA propre config
++
++Quand Bob confie une question à un autre agent (`ask_agent`, `/ask`, `mm --ask`), l'agent appelé travaille dans le même répertoire, mais c'est **son** réglage `agentsMd` qui décide s'il lit le fichier.
++
++```bash
++mm ../agent.yaml --ask helper -p "How do I run the tests in this project?"
++```
++
++```
++To run the tests, execute `make check` as instructed by the project's build and test guidelines. Always quote the last line of the command output in your response.
++-- Tangerine bot
++```
++
++```bash
++mm ../agent.yaml --ask reviewer -p "How do I run the tests in this project?"
++```
++
++```
++Run the tests by executing the command specified in your project's documentation, typically `npm test` for Node.js projects or `pytest` for Python projects. …
++```
++
++helper connaît `make check` et signe ; reviewer (`agentsMd.enabled: false`) donne une réponse générique : il n'a pas vu le fichier.
++
++## 4. Modifier `AGENTS.md` pendant une session
++
++Le fichier est lu **au début de chaque conversation**, jamais au milieu (changer le message système en cours de route obligerait le serveur à relire tout l'historique). Pour prendre une modification en compte :
++
++```bash
++../../mm ../agent.yaml
++> Run the tests.            # make check
++```
++
++Modifiez `AGENTS.md` dans un autre terminal (par exemple : la commande devient `sh greet.sh Tangerine`), puis :
++
++```text
++> /new
++🆕 new session: 4 message(s) forgotten
++> Run the tests.
++🛠️  bash: sh greet.sh Tangerine
++```
++
++La nouvelle consigne s'applique dès la nouvelle conversation. (Lors de notre essai, Bob a ensuite lancé aussi `make check`, malgré « Never run make » : un modèle 4B ne respecte pas chaque interdiction.)
++
++`/resume` relit aussi le fichier actuel.
++
++## 5. Vérifier ce que le modèle a reçu
++
++```bash
++AGENT_TRACE=true ../../mm ../agent.yaml -p "Run the tests."
++jq -r '.messages[0].content' .mm/traces/*-0001.request.json
++```
++
++Le message système se termine par la section `PROJECT INSTRUCTIONS (from …)`.
++
++## Ce qu'il faut savoir
++
++- **Un petit modèle ne suit pas toutes les règles à chaque fois** : avec Qwen3.5-4B, Bob a lancé `make check` à chaque essai, mais a oublié la signature `-- Tangerine bot` ; helper, sans outils, l'a mise. Des consignes courtes et concrètes passent mieux.
++- **Taille** : le fichier part avec chaque requête. La bannière avertit quand il occupe 20 % ou plus de la fenêtre de contexte ; `agentsMd.maxChars` fixe une limite ferme (au-delà, il est ignoré en entier, et la bannière l'indique `skipped (too long)`).
++- **Un agent JSON** (`structuredOutput`) ne lit jamais `AGENTS.md` ; sa bannière indique `agents.md: ignored (structured output)`. Voir [demo-structured-output](../demo-structured-output/README.md).
++- **Dans un éditeur** (Zed…, mode ACP), chaque session lit le `AGENTS.md` du projet ouvert.
++
++## Pour aller plus loin
++
++- [Donner à l'agent les consignes de votre projet](../docs/fr/how-to/give-project-instructions.md)
++- [Référence de `agentsMd`](../docs/fr/reference/configuration.md#clés-de-agentsmd)
++- [Consignes du projet — explication](../docs/fr/explanation/project-instructions.md)

added demo-agents-file/agent.helper.yaml +26 -0 Viewed

new file mode 100644
@@ -0,0 +1,26 @@
	1	+# Agent settings — helper: no tools, and it READS the project's AGENTS.md
	2	+# (agentsMd is on by default). Asked by Bob, it runs in Bob's working
	3	+# directory, so it gets the same project rules.
	4	+provider: llmman
	5	+model: hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
	6	+baseUrl: http://127.0.0.1:17434/v1
	7	+fallback: http://host.docker.internal:17434/v1
	8	+
	9	+bashTool: false
	10	+editTools: false
	11	+planTool: false
	12	+skillsDir: no-skills
	13	+
	14	+acp:
	15	+ name: helper
	16	+ title: Helper (reads AGENTS.md)
	17	+
	18	+system: \|
	19	+ You are Helper. You answer questions about how to work in a project, in
	20	+ two sentences at most. You have no tools.
	21	+
	22	+sampling:
	23	+ temperature: 0.0
	24	+ max_tokens: 1024
	25	+ chat_template_kwargs:
	26	+ enable_thinking: false

++# Agent settings — helper: no tools, and it READS the project's AGENTS.md
++# (agentsMd is on by default). Asked by Bob, it runs in Bob's working
++# directory, so it gets the same project rules.
++provider: llmman
++model: hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
++baseUrl: http://127.0.0.1:17434/v1
++fallback: http://host.docker.internal:17434/v1
++
++bashTool: false
++editTools: false
++planTool: false
++skillsDir: no-skills
++
++acp:
++  name: helper
++  title: Helper (reads AGENTS.md)
++
++system: |
++  You are Helper. You answer questions about how to work in a project, in
++  two sentences at most. You have no tools.
++
++sampling:
++  temperature: 0.0
++  max_tokens: 1024
++  chat_template_kwargs:
++    enable_thinking: false

added demo-agents-file/agent.reviewer.yaml +30 -0 Viewed

new file mode 100644
@@ -0,0 +1,30 @@
	1	+# Agent settings — reviewer: no tools, and it does NOT read the project's AGENTS.md
	2	+# (agentsMd.enabled: false below): its answers are an outside opinion,
	3	+# free of the project's rules — even though it runs in the project directory.
	4	+provider: llmman
	5	+model: hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
	6	+baseUrl: http://127.0.0.1:17434/v1
	7	+fallback: http://host.docker.internal:17434/v1
	8	+
	9	+bashTool: false
	10	+editTools: false
	11	+planTool: false
	12	+skillsDir: no-skills
	13	+
	14	+# The one difference with agent.helper.yaml.
	15	+agentsMd:
	16	+ enabled: false
	17	+
	18	+acp:
	19	+ name: reviewer
	20	+ title: Reviewer (ignores AGENTS.md)
	21	+
	22	+system: \|
	23	+ You are Reviewer. You answer questions about how to work in a project, in
	24	+ two sentences at most. You have no tools.
	25	+
	26	+sampling:
	27	+ temperature: 0.0
	28	+ max_tokens: 1024
	29	+ chat_template_kwargs:
	30	+ enable_thinking: false

++# Agent settings — reviewer: no tools, and it does NOT read the project's AGENTS.md
++# (agentsMd.enabled: false below): its answers are an outside opinion,
++# free of the project's rules — even though it runs in the project directory.
++provider: llmman
++model: hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
++baseUrl: http://127.0.0.1:17434/v1
++fallback: http://host.docker.internal:17434/v1
++
++bashTool: false
++editTools: false
++planTool: false
++skillsDir: no-skills
++
++# The one difference with agent.helper.yaml.
++agentsMd:
++  enabled: false
++
++acp:
++  name: reviewer
++  title: Reviewer (ignores AGENTS.md)
++
++system: |
++  You are Reviewer. You answer questions about how to work in a project, in
++  two sentences at most. You have no tools.
++
++sampling:
++  temperature: 0.0
++  max_tokens: 1024
++  chat_template_kwargs:
++    enable_thinking: false

added demo-agents-file/agent.yaml +62 -0 Viewed

new file mode 100644
@@ -0,0 +1,62 @@
	1	+# Agent settings — Bob, a coding agent that follows the AGENTS.md of the
	2	+# project it works in (ticket 81). Run him FROM THE PROJECT directory:
	3	+#
	4	+# cd demo-agents-file/project
	5	+# ../../mm ../agent.yaml
	6	+#
	7	+# The banner then shows: agents.md: …/project/AGENTS.md (~N tokens)
	8	+provider: llmman
	9	+model: hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
	10	+baseUrl: http://127.0.0.1:17434/v1
	11	+fallback: http://host.docker.internal:17434/v1
	12	+
	13	+# The project's instructions. These are the defaults — written out for the
	14	+# demo; a config that never mentions agentsMd reads ./AGENTS.md all the same.
	15	+# enabled read it at the start of every conversation (start-up, /new, /resume)
	16	+# path relative to the WORKING directory (the project), not to this file
	17	+# maxChars 0 = no limit; above it the file is skipped whole, never cut
	18	+agentsMd:
	19	+ enabled: true
	20	+ path: AGENTS.md
	21	+ maxChars: 0
	22	+
	23	+# Two agents Bob can hand work to with `ask_agent`. Each one decides with
	24	+# ITS OWN agentsMd setting whether it reads the project's AGENTS.md: helper
	25	+# does (default), reviewer does not (agentsMd.enabled: false).
	26	+mmConfigs:
	27	+ helper:
	28	+ path: agent.helper.yaml
	29	+ description: A second coding agent with no tools — answers questions about how to work in this project.
	30	+ reviewer:
	31	+ path: agent.reviewer.yaml
	32	+ description: A neutral reviewer with no tools — gives an outside opinion, without the project's rules.
	33	+
	34	+bashTool: true
	35	+editTools: true
	36	+planTool: false
	37	+skillsDir: no-skills
	38	+# Bob's commands and edits run without asking, so that `-p` runs to the end.
	39	+yolo: true
	40	+bashTimeout: 1m
	41	+
	42	+acp:
	43	+ name: bob
	44	+ title: Bob (coding agent)
	45	+
	46	+system: \|
	47	+ Your name is Bob. You are a coding agent working in a terminal, with a
	48	+ `bash` tool and file tools (`read_file`, `write_file`, `edit_file`).
	49	+ Never state the output of a command you did not run in this answer.
	50	+ `ask_agent` hands a question to another agent; call it only when the user
	51	+ names that agent.
	52	+
	53	+sampling:
	54	+ temperature: 0.0
	55	+ parallel_tool_calls: false
	56	+ max_tokens: 4096
	57	+ chat_template_kwargs:
	58	+ enable_thinking: false
	59	+
	60	+maxTurns: 20
	61	+watchdogTimeout: 2m
	62	+firstTokenTimeout: 3m

++# Agent settings — Bob, a coding agent that follows the AGENTS.md of the
++# project it works in (ticket 81). Run him FROM THE PROJECT directory:
++#
++#   cd demo-agents-file/project
++#   ../../mm ../agent.yaml
++#
++# The banner then shows: agents.md: …/project/AGENTS.md (~N tokens)
++provider: llmman
++model: hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
++baseUrl: http://127.0.0.1:17434/v1
++fallback: http://host.docker.internal:17434/v1
++
++# The project's instructions. These are the defaults — written out for the
++# demo; a config that never mentions agentsMd reads ./AGENTS.md all the same.
++#   enabled   read it at the start of every conversation (start-up, /new, /resume)
++#   path      relative to the WORKING directory (the project), not to this file
++#   maxChars  0 = no limit; above it the file is skipped whole, never cut
++agentsMd:
++  enabled: true
++  path: AGENTS.md
++  maxChars: 0
++
++# Two agents Bob can hand work to with `ask_agent`. Each one decides with
++# ITS OWN agentsMd setting whether it reads the project's AGENTS.md: helper
++# does (default), reviewer does not (agentsMd.enabled: false).
++mmConfigs:
++  helper:
++    path: agent.helper.yaml
++    description: A second coding agent with no tools — answers questions about how to work in this project.
++  reviewer:
++    path: agent.reviewer.yaml
++    description: A neutral reviewer with no tools — gives an outside opinion, without the project's rules.
++
++bashTool: true
++editTools: true
++planTool: false
++skillsDir: no-skills
++# Bob's commands and edits run without asking, so that `-p` runs to the end.
++yolo: true
++bashTimeout: 1m
++
++acp:
++  name: bob
++  title: Bob (coding agent)
++
++system: |
++  Your name is Bob. You are a coding agent working in a terminal, with a
++  `bash` tool and file tools (`read_file`, `write_file`, `edit_file`).
++  Never state the output of a command you did not run in this answer.
++  `ask_agent` hands a question to another agent; call it only when the user
++  names that agent.
++
++sampling:
++  temperature: 0.0
++  parallel_tool_calls: false
++  max_tokens: 4096
++  chat_template_kwargs:
++    enable_thinking: false
++
++maxTurns: 20
++watchdogTimeout: 2m
++firstTokenTimeout: 3m

added demo-agents-file/project/AGENTS.md +12 -0 Viewed