bots-garden/mini-me · 8654e3f

⑂ 8654e3f

Merge feature/talk-to-another-mm into main

rickub committed 2026-10-02T19:59:49Z Browse files

8654e3f parents: 36b4793 45e0704

added .tickets/issues/0066-talk-to-another-mm.yaml +391 -0 Viewed

new file mode 100644
@@ -0,0 +1,391 @@
	1	+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
	2	+id: 66
	3	+title: talk to another mm
	4	+state: open
	5	+author:
	6	+ name: k33g
	7	+ email: ph.charriere@gmail.com
	8	+createdAt: 2026-10-02T18:48:36.275Z
	9	+updatedAt: 2026-10-02T20:30:00.000Z
	10	+labels:
	11	+ - kind::feature
	12	+related:
	13	+ - 57
	14	+ - 59
	15	+body: \|
	16	+ Goal: from the current session, send a prompt to ANOTHER mm (another config:
	17	+ other model, provider, system prompt, tools, skills) and bring its answer
	18	+ back into the current conversation, so the current model can use it.
	19	+
	20	+ /ask bob <prompt>
	21	+ /ask riker <prompt>
	22	+
	23	+ Config (user's proposal):
	24	+
	25	+ mm-configs:
	26	+ bob: path/to/agent.bob.yaml
	27	+ riker: path/to/agent.riker.yaml
	28	+
	29	+ Feasibility: yes. Checked 2026-10-02 (no code written).
	30	+
	31	+ Recommended transport: ACP over stdio, mm talking to a child `mm -acp`.
	32	+ - The child is started with os.Executable() (same mm build, not whatever is
	33	+ on PATH) and `-config <path>`, cwd = the current working directory.
	34	+ - The parent is an ACP CLIENT: github.com/coder/acp-go-sdk (already a
	35	+ dependency, v0.13.5) has NewClientSideConnection + the Client interface.
	36	+ - mm's ACP server only ever calls two client methods (grep internal/acp):
	37	+ SessionUpdate (message/thought/tool-call/usage chunks) and
	38	+ RequestPermission. The rest of the Client interface (fs, terminal) can be
	39	+ stubs returning "not supported", and the parent advertises no fs/terminal
	40	+ capability.
	41	+ - Flow: initialize -> session/new {cwd} -> session/prompt {text} -> collect
	42	+ agent_message_chunk text until the prompt response (stopReason) ->
	43	+ insert the answer in the current history.
	44	+ - Proof it works as a protocol: tests/tool-order/session.py and the sidekick
	45	+ web client already drive `mm -acp` from outside.
	46	+
	47	+ Why ACP rather than the alternatives:
	48	+ - `mm -p` subprocess (one-shot): simpler, but stdout mixes the answer with
	49	+ thoughts, tool echoes, ANSI and the final command summary (fragile
	50	+ parsing); line mode never asks permission, so bob would run bash/write
	51	+ with NO gate; no cancel except kill; no usage; no follow-up turns.
	52	+ - In-process second Engine: no process, but config.Cfg is a global read
	53	+ everywhere (engine, tools, compact, store...). Two configs in one process
	54	+ means de-globalising the config first: a large refactor for this feature.
	55	+ The subprocess also isolates bob's tools, MCP servers and skills.
	56	+ - ACP gives, for free: structured events (tool calls can be shown as
	57	+ "bob: bash ls"), permission requests that can be FORWARDED to the user
	58	+ through the current front end's existing sink.Allow (line mode: no gate,
	59	+ window: inline prompt, editor: the editor's dialog), session/cancel for
	60	+ /abort and Ctrl+C, usage, and a session id -> multi-turn with bob
	61	+ (keep the child alive, reuse the session: "/ask bob and what about X?").
	62	+
	63	+ Proposed behaviour (to confirm):
	64	+ - One child per name, started lazily on the first /ask, kept alive for the
	65	+ parent session (start-up cost), killed on /quit and exit; a fresh ACP
	66	+ session (session/new) per /ask if 3 is confirmed.
	67	+ - The /ask exchange enters the CURRENT history as one user message, e.g.
	68	+ [bob (ollama/qwen3.5:2b) was asked: <prompt>]
	69	+ <bob's answer>
	70	+ so the current model sees it on the next question and the session file
	71	+ persists and replays it on /resume. Must keep the history valid for
	72	+ the next model call (user/model alternation: merge into the next user
	73	+ message, or a user+model pair "[noted]" — to decide in the code).
	74	+ - Bob's answer is shown while it streams, prefixed/styled as bob's
	75	+ (decided: displayed, kept for the next question).
	76	+ - Bob's thoughts: not shown (or dim), never inserted in the history.
	77	+ - Long answers: capped (maxOutput-like) before insertion — the current
	78	+ model's context pays for them.
	79	+
	80	+ Risks / things to decide:
	81	+ - Ticket 57: a permission forwarded to the user and answered slowly will
	82	+ trip the CHILD's watchdog exactly like today in an editor.
	83	+ - Ticket 59 (yolo): which permission policy for the child — forward to the
	84	+ user (default), or auto per config entry?
	85	+ - Two agents on the same working tree: the parent waits while bob works
	86	+ (/ask is blocking), so no concurrent edits; but bob may change files the
	87	+ parent model "remembers" — say so in the inserted message?
	88	+ - Recursion: /ask is a USER command, the model cannot call it, so no
	89	+ automatic loop. If it ever becomes a tool ("ask_agent"), cap the depth
	90	+ (env var MM_ASK_DEPTH) and refuse cycles.
	91	+ - Paths in mm-configs: relative to the PARENT config file's directory (the
	92	+ ACP mode already resolves skillsDir that way; a path relative to the cwd
	93	+ breaks when the config is elsewhere). The child's own relative paths
	94	+ (skillsDir) resolve against ITS config dir in -acp mode.
	95	+ - Start-up cost: Genkit init + probe + MCP servers per child (~1 s+);
	96	+ hence keep the child alive.
	97	+ - Child stderr: capture it and show it only on failure (start error,
	98	+ unknown config, provider down -> one line: "bob: nothing answers at …").
	99	+
	100	+ Demo ready (2026-10-02): demo-llmman/agent.llmman.yaml declares
	101	+ mmConfigs:
	102	+ jan: agent.jan-nano.yaml # hf.co/Menlo/Jan-nano-gguf:Q4_K_M
	103	+ lucy: agent.lucy.yaml # hf.co/Menlo/Lucy-gguf:Q4_K_M
	104	+ (the key is ignored until implemented: config loading is not strict). Both
	105	+ configs answered a bash-tool prompt live through llmman.
	106	+
	107	+ Decisions (user, 2026-10-02):
	108	+ 1. Key in camelCase: `mmConfigs:` (name -> path of a config file).
	109	+ 2. Bob's answer is DISPLAYED, and stays in the current history for the
	110	+ NEXT question; the current model does not answer right away.
	111	+ 3. Two commands:
	112	+ - /ask <name> <prompt>: bob receives the typed prompt only;
	113	+ - /ask-with-context <name> <prompt>: bob also receives the context of
	114	+ the current conversation.
	115	+ Interpretation to confirm: every /ask starts a FRESH bob session (no
	116	+ memory between two /ask); continuity comes from /ask-with-context.
	117	+ 4. Both commands work in the three front ends (Terminal, Window, Editor).
	118	+ 5. Context: only with /ask-with-context (see 3).
	119	+
	120	+ Editor visibility (checked): yes. The ACP facade already sends
	121	+ `available_commands_update` built from command.For(Editor)
	122	+ (internal/acp/acp.go availableCommands), and skills are already advertised
	123	+ with `input.unstructured.hint`. Today the built-in commands are sent
	124	+ WITHOUT an input hint: /ask and /ask-with-context need one, e.g.
	125	+ "bob\|riker <prompt>", built from mmConfigs. The editor completes the
	126	+ command name, not the argument; alternative to discuss: one command per
	127	+ agent (/ask-bob, /ask-with-context-bob) so the picker lists the agents.
	128	+ The CHILD's own available_commands_update is ignored by the parent.
	129	+
	130	+ Decided too (user, 2026-10-02):
	131	+ a. Context for /ask-with-context = a text transcript of the current
	132	+ history: user questions and assistant answers, each tool call as one
	133	+ line (name + short args), no tool outputs, no reasoning, the compaction
	134	+ summary if any; capped (chars), newest turns kept first. Sent as:
	135	+ [Conversation so far, from <parent agent name>]
	136	+ ...
	137	+ [Question]
	138	+ <prompt>
	139	+ b. Every /ask (and /ask-with-context) starts a FRESH bob session
	140	+ (session/new); no memory on bob's side between two calls.
	141	+ c. Editors: ONE COMMAND PER AGENT, generated from mmConfigs —
	142	+ /ask-bob, /ask-with-context-bob, /ask-riker, … each with an input hint
	143	+ "<prompt>". Consequences:
	144	+ - agent names become command names: validate them in config
	145	+ (lowercase letters, digits, '-'; error otherwise);
	146	+ - a skill or built-in with the same name: the existing guard in
	147	+ availableCommands (commandNamed) keeps the command and drops the skill,
	148	+ with the banner warning — extend it to the generated names;
	149	+ - terminal and window (confirmed by the user, 2026-10-02): accept the
	150	+ per-agent forms there too (same table, no drift), and keep
	151	+ `/ask <name> <prompt>` as the generic form; the window's "/" picker
	152	+ lists the per-agent commands.
	153	+
	154	+ Proposed behaviour (to confirm):
	155	+ - One child per name, started lazily on the first /ask, kept alive for the
	156	+ parent session (start-up cost), killed on /quit and exit; a fresh ACP
	157	+ session (session/new) per /ask if 3 is confirmed.
	158	+ - The /ask exchange enters the CURRENT history as one user message, e.g.
	159	+ [bob (ollama/qwen3.5:2b) was asked: <prompt>]
	160	+ <bob's answer>
	161	+ so the current model sees it on the next question and the session file
	162	+ persists and replays it on /resume. Must keep the history valid for
	163	+ the next model call (user/model alternation: merge into the next user
	164	+ message, or a user+model pair "[noted]" — to decide in the code).
	165	+ - Bob's answer is shown while it streams, prefixed/styled as bob's
	166	+ (decided: displayed, kept for the next question).
	167	+ - Bob's thoughts: not shown (or dim), never inserted in the history.
	168	+ - Long answers: capped (maxOutput-like) before insertion — the current
	169	+ model's context pays for them.
	170	+
	171	+ Risks / things to decide:
	172	+ - Ticket 57: a permission forwarded to the user and answered slowly will
	173	+ trip the CHILD's watchdog exactly like today in an editor.
	174	+ - Ticket 59 (yolo): which permission policy for the child — forward to the
	175	+ user (default), or auto per config entry?
	176	+ - Two agents on the same working tree: the parent waits while bob works
	177	+ (/ask is blocking), so no concurrent edits; but bob may change files the
	178	+ parent model "remembers" — say so in the inserted message?
	179	+ - Recursion: /ask is a USER command, the model cannot call it, so no
	180	+ automatic loop. If it ever becomes a tool ("ask_agent"), cap the depth
	181	+ (env var MM_ASK_DEPTH) and refuse cycles.
	182	+ - Paths in mm-configs: relative to the PARENT config file's directory (the
	183	+ ACP mode already resolves skillsDir that way; a path relative to the cwd
	184	+ breaks when the config is elsewhere). The child's own relative paths
	185	+ (skillsDir) resolve against ITS config dir in -acp mode.
	186	+ - Start-up cost: Genkit init + probe + MCP servers per child (~1 s+);
	187	+ hence keep the child alive.
	188	+ - Child stderr: capture it and show it only on failure (start error,
	189	+ unknown config, provider down -> one line: "bob: nothing answers at …").
	190	+
	191	+ Decisions (user, 2026-10-02):
	192	+ 1. Key in camelCase: `mmConfigs:` (name -> path of a config file).
	193	+ 2. Bob's answer is DISPLAYED, and stays in the current history for the
	194	+ NEXT question; the current model does not answer right away.
	195	+ 3. Two commands:
	196	+ - /ask <name> <prompt>: bob receives the typed prompt only;
	197	+ - /ask-with-context <name> <prompt>: bob also receives the context of
	198	+ the current conversation.
	199	+ Interpretation to confirm: every /ask starts a FRESH bob session (no
	200	+ memory between two /ask); continuity comes from /ask-with-context.
	201	+ 4. Both commands work in the three front ends (Terminal, Window, Editor).
	202	+ 5. Context: only with /ask-with-context (see 3).
	203	+
	204	+ Editor visibility (checked): yes. The ACP facade already sends
	205	+ `available_commands_update` built from command.For(Editor)
	206	+ (internal/acp/acp.go availableCommands), and skills are already advertised
	207	+ with `input.unstructured.hint`. Today the built-in commands are sent
	208	+ WITHOUT an input hint: /ask and /ask-with-context need one, e.g.
	209	+ "bob\|riker <prompt>", built from mmConfigs. The editor completes the
	210	+ command name, not the argument; alternative to discuss: one command per
	211	+ agent (/ask-bob, /ask-with-context-bob) so the picker lists the agents.
	212	+ The CHILD's own available_commands_update is ignored by the parent.
	213	+
	214	+ Still open:
	215	+ a. What "context" is for /ask-with-context. Proposal: a text transcript of
	216	+ the current history — user questions and assistant answers, each tool
	217	+ call as one line (name + short args), no tool outputs, no reasoning,
	218	+ the compaction summary if any — capped (chars) to protect bob's
	219	+ window, newest turns kept first. Sent as the first part of the prompt:
	220	+ [Conversation so far, from <parent agent name>]
	221	+ ...
	222	+ [Question]
	223	+ <prompt>
	224	+ b. Fresh bob session per /ask (interpretation of 3) — confirm.
	225	+ c. Editor: one /ask with a hint, or one command per agent?
	226	+tasks:
	227	+ - id: 1
	228	+ title: "Answer the open questions (key name, immediate answer, multi-turn, fronts, context passed)"
	229	+ state: closed
	230	+ priority: none
	231	+ author:
	232	+ name: k33g
	233	+ email: ph.charriere@gmail.com
	234	+ createdAt: 2026-10-02T19:30:00.000Z
	235	+ updatedAt: 2026-10-02T20:00:00.000Z
	236	+ - id: 2
	237	+ title: "Confirm the remaining points a/b/c (context format, fresh session per /ask, editor hint vs one command per agent)"
	238	+ state: closed
	239	+ priority: none
	240	+ author:
	241	+ name: k33g
	242	+ email: ph.charriere@gmail.com
	243	+ createdAt: 2026-10-02T19:30:00.000Z
	244	+ updatedAt: 2026-10-02T20:00:00.000Z
	245	+ - id: 3
	246	+ title: "config: mmConfigs map (validation: name = [a-z0-9-]+ since it becomes a command name, file exists; paths relative to the config file dir)"
	247	+ state: closed
	248	+ priority: none
	249	+ author:
	250	+ name: k33g
	251	+ email: ph.charriere@gmail.com
	252	+ createdAt: 2026-10-02T19:30:00.000Z
	253	+ updatedAt: 2026-10-02T20:00:00.000Z
	254	+ - id: 4
	255	+ title: "internal/askagent (name TBD): ACP client over a child `mm -acp -config <path>` (os.Executable), lazy start, one child per name, shutdown"
	256	+ state: closed
	257	+ priority: none
	258	+ author:
	259	+ name: k33g
	260	+ email: ph.charriere@gmail.com
	261	+ createdAt: 2026-10-02T19:30:00.000Z
	262	+ updatedAt: 2026-10-02T20:00:00.000Z
	263	+ - id: 5
	264	+ title: "Client side: SessionUpdate -> stream text/tool-call lines to the current sink; RequestPermission -> current sink.Allow; fs/terminal stubs"
	265	+ state: closed
	266	+ priority: none
	267	+ author:
	268	+ name: k33g
	269	+ email: ph.charriere@gmail.com
	270	+ createdAt: 2026-10-02T19:30:00.000Z
	271	+ updatedAt: 2026-10-02T20:00:00.000Z
	272	+ - id: 6
	273	+ title: "command table: /ask <name> <prompt>, /ask-with-context <name> <prompt>, and per-agent /ask-<name>, /ask-with-context-<name> generated from mmConfigs (all fronts); ACP advertises the per-agent ones with hint <prompt>; window picker; name clash guard"
	274	+ state: closed
	275	+ priority: none
	276	+ author:
	277	+ name: k33g
	278	+ email: ph.charriere@gmail.com
	279	+ createdAt: 2026-10-02T19:30:00.000Z
	280	+ updatedAt: 2026-10-02T20:00:00.000Z
	281	+ - id: 7
	282	+ title: "Context builder for /ask-with-context: transcript of the current history (text + one line per tool call), capped"
	283	+ state: closed
	284	+ priority: none
	285	+ author:
	286	+ name: k33g
	287	+ email: ph.charriere@gmail.com
	288	+ createdAt: 2026-10-02T19:30:00.000Z
	289	+ updatedAt: 2026-10-02T20:00:00.000Z
	290	+ - id: 8
	291	+ title: "Insert the exchange in the current history (format, cap, valid alternation), persisted by the store and replayed on /resume"
	292	+ state: closed
	293	+ priority: none
	294	+ author:
	295	+ name: k33g
	296	+ email: ph.charriere@gmail.com
	297	+ createdAt: 2026-10-02T19:30:00.000Z
	298	+ updatedAt: 2026-10-02T20:00:00.000Z
	299	+ - id: 9
	300	+ title: "/abort, Ctrl+C and editor cancel -> session/cancel on the child; child crash -> one-line error, history untouched"
	301	+ state: closed
	302	+ priority: none
	303	+ author:
	304	+ name: k33g
	305	+ email: ph.charriere@gmail.com
	306	+ createdAt: 2026-10-02T19:30:00.000Z
	307	+ updatedAt: 2026-10-02T20:00:00.000Z
	308	+ - id: 10
	309	+ title: "Tests: fake child agent (acp-go-sdk agent side) for the client; end-to-end with two real configs (demo-ollama + demo-llmman)"
	310	+ state: closed
	311	+ priority: none
	312	+ author:
	313	+ name: k33g
	314	+ email: ph.charriere@gmail.com
	315	+ createdAt: 2026-10-02T19:30:00.000Z
	316	+ updatedAt: 2026-10-02T20:00:00.000Z
	317	+ - id: 11
	318	+ title: "Docs EN+FR: how-to ask another agent, reference (config key, commands), explanation (why ACP, why a subprocess)"
	319	+ state: closed
	320	+ priority: none
	321	+ author:
	322	+ name: k33g
	323	+ email: ph.charriere@gmail.com
	324	+ createdAt: 2026-10-02T19:30:00.000Z
	325	+ updatedAt: 2026-10-02T20:00:00.000Z
	326	+ - id: 12
	327	+ title: "Quality gate (qlty) where available — not run: qlty absent from this sandbox"
	328	+ state: open
	329	+ priority: none
	330	+ author:
	331	+ name: k33g
	332	+ email: ph.charriere@gmail.com
	333	+ createdAt: 2026-10-02T20:30:00.000Z
	334	+ updatedAt: 2026-10-02T20:30:00.000Z
	335	+ - id: 13
	336	+ title: "Live check of the window mode (/ask-jan in mm -tui) — unit-tested only"
	337	+ state: open
	338	+ priority: none
	339	+ author:
	340	+ name: k33g
	341	+ email: ph.charriere@gmail.com
	342	+ createdAt: 2026-10-02T20:30:00.000Z
	343	+ updatedAt: 2026-10-02T20:30:00.000Z
	344	+ - id: 14
	345	+ title: "Review, commit, merge"
	346	+ state: open
	347	+ priority: none
	348	+ author:
	349	+ name: k33g
	350	+ email: ph.charriere@gmail.com
	351	+ createdAt: 2026-10-02T20:30:00.000Z
	352	+ updatedAt: 2026-10-02T20:30:00.000Z
	353	+comments:
	354	+ - id: 1
	355	+ author:
	356	+ name: claude
	357	+ email: noreply@anthropic.com
	358	+ createdAt: 2026-10-02T20:30:00.000Z
	359	+ body: \|
	360	+ Implemented on branch `feature/talk-to-another-mm` (uncommitted, 2026-10-02).
	361	+
	362	+ - config: `mmConfigs` (internal/config/agents.go) — names [a-z0-9-], not
	363	+ starting with `with-context`; relative paths resolved against the config
	364	+ file's dir at Load; missing file = start-up warning. `AGENT_SESSIONS=off`
	365	+ env override (applyEnv).
	366	+ - command: /ask family generated by SetAgents (internal/command/ask.go);
	367	+ `Hint` field; ParseAsk returns (req, isAsk, err).
	368	+ - internal/ask (new): Transcript/Prompt (context.go), Exchange as a
	369	+ user+model pair + Describe (exchange.go), ChildEnv/Depth/CanAsk
	370	+ (env.go, MaxDepth 2), ACP client (client.go), Pool with one lazy child
	371	+ per agent + fresh session per ask (pool.go), Pool.Run + Runner interface
	372	+ + Label + Missing (run.go).
	373	+ - Front ends: agent/ask.go (Ctrl+C, /abort, /quit), tui/ask.go (job
	374	+ "asking <agent>"), acp/ask.go (per-agent commands with `<prompt>` hint);
	375	+ main.go wiring; store.FailAsk.
	376	+ - Traps found: the SDK dispatches the child's requests concurrently with
	377	+ its notifications — a permission request can overtake its tool_call
	378	+ (client shows the call first, once); the SDK's default slog would print
	379	+ "connection closed" on the user's terminal (silenced); the parent's
	380	+ AGENT_* env would override the child's config (stripped).
	381	+ - Tests: config, command, ask (fake ACP agent = the test binary re-run,
	382	+ no network), agent, tui, acp. `go test ./...` green 3x; ask tests 15x.
	383	+ `-race` not available (no C compiler in the sandbox).
	384	+ - Live (host llmman): line mode `/ask-jan` → answer kept, the current
	385	+ model quoted it next turn; ACP `/ask-with-context-lucy` → per-agent
	386	+ commands announced, Lucy knew the name given earlier, her bash
	387	+ permission went to the client titled "lucy · bash: …"; no child
	388	+ session files, no leftover processes.
	389	+ - Docs EN+FR: how-to/ask-another-agent.md, explanation/asking-another-agent.md,
	390	+ reference (slash-commands, configuration, cli), architecture table,
	391	+ READMEs, docs/diagrams/packages.drawio (internal/ask + edges).

++# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
++id: 66
++title: talk to another mm
++state: open
++author:
++  name: k33g
++  email: ph.charriere@gmail.com
++createdAt: 2026-10-02T18:48:36.275Z
++updatedAt: 2026-10-02T20:30:00.000Z
++labels:
++  - kind::feature
++related:
++  - 57
++  - 59
++body: |
++  Goal: from the current session, send a prompt to ANOTHER mm (another config:
++  other model, provider, system prompt, tools, skills) and bring its answer
++  back into the current conversation, so the current model can use it.
++
++    /ask bob   <prompt>
++    /ask riker <prompt>
++
++  Config (user's proposal):
++
++    mm-configs:
++      bob:   path/to/agent.bob.yaml
++      riker: path/to/agent.riker.yaml
++
++  Feasibility: yes. Checked 2026-10-02 (no code written).
++
++  Recommended transport: ACP over stdio, mm talking to a child `mm -acp`.
++  - The child is started with os.Executable() (same mm build, not whatever is
++    on PATH) and `-config <path>`, cwd = the current working directory.
++  - The parent is an ACP CLIENT: github.com/coder/acp-go-sdk (already a
++    dependency, v0.13.5) has NewClientSideConnection + the Client interface.
++  - mm's ACP server only ever calls two client methods (grep internal/acp):
++    SessionUpdate (message/thought/tool-call/usage chunks) and
++    RequestPermission. The rest of the Client interface (fs, terminal) can be
++    stubs returning "not supported", and the parent advertises no fs/terminal
++    capability.
++  - Flow: initialize -> session/new {cwd} -> session/prompt {text} -> collect
++    agent_message_chunk text until the prompt response (stopReason) ->
++    insert the answer in the current history.
++  - Proof it works as a protocol: tests/tool-order/session.py and the sidekick
++    web client already drive `mm -acp` from outside.
++
++  Why ACP rather than the alternatives:
++  - `mm -p` subprocess (one-shot): simpler, but stdout mixes the answer with
++    thoughts, tool echoes, ANSI and the final command summary (fragile
++    parsing); line mode never asks permission, so bob would run bash/write
++    with NO gate; no cancel except kill; no usage; no follow-up turns.
++  - In-process second Engine: no process, but config.Cfg is a global read
++    everywhere (engine, tools, compact, store...). Two configs in one process
++    means de-globalising the config first: a large refactor for this feature.
++    The subprocess also isolates bob's tools, MCP servers and skills.
++  - ACP gives, for free: structured events (tool calls can be shown as
++    "bob: bash ls"), permission requests that can be FORWARDED to the user
++    through the current front end's existing sink.Allow (line mode: no gate,
++    window: inline prompt, editor: the editor's dialog), session/cancel for
++    /abort and Ctrl+C, usage, and a session id -> multi-turn with bob
++    (keep the child alive, reuse the session: "/ask bob and what about X?").
++
++  Proposed behaviour (to confirm):
++  - One child per name, started lazily on the first /ask, kept alive for the
++    parent session (start-up cost), killed on /quit and exit; a fresh ACP
++    session (session/new) per /ask if 3 is confirmed.
++  - The /ask exchange enters the CURRENT history as one user message, e.g.
++      [bob (ollama/qwen3.5:2b) was asked: <prompt>]
++      <bob's answer>
++    so the current model sees it on the next question and the session file
++    persists and replays it on /resume. Must keep the history valid for
++    the next model call (user/model alternation: merge into the next user
++    message, or a user+model pair "[noted]" — to decide in the code).
++  - Bob's answer is shown while it streams, prefixed/styled as bob's
++    (decided: displayed, kept for the next question).
++  - Bob's thoughts: not shown (or dim), never inserted in the history.
++  - Long answers: capped (maxOutput-like) before insertion — the current
++    model's context pays for them.
++
++  Risks / things to decide:
++  - Ticket 57: a permission forwarded to the user and answered slowly will
++    trip the CHILD's watchdog exactly like today in an editor.
++  - Ticket 59 (yolo): which permission policy for the child — forward to the
++    user (default), or auto per config entry?
++  - Two agents on the same working tree: the parent waits while bob works
++    (/ask is blocking), so no concurrent edits; but bob may change files the
++    parent model "remembers" — say so in the inserted message?
++  - Recursion: /ask is a USER command, the model cannot call it, so no
++    automatic loop. If it ever becomes a tool ("ask_agent"), cap the depth
++    (env var MM_ASK_DEPTH) and refuse cycles.
++  - Paths in mm-configs: relative to the PARENT config file's directory (the
++    ACP mode already resolves skillsDir that way; a path relative to the cwd
++    breaks when the config is elsewhere). The child's own relative paths
++    (skillsDir) resolve against ITS config dir in -acp mode.
++  - Start-up cost: Genkit init + probe + MCP servers per child (~1 s+);
++    hence keep the child alive.
++  - Child stderr: capture it and show it only on failure (start error,
++    unknown config, provider down -> one line: "bob: nothing answers at …").
++
++  Demo ready (2026-10-02): demo-llmman/agent.llmman.yaml declares
++    mmConfigs:
++      jan: agent.jan-nano.yaml   # hf.co/Menlo/Jan-nano-gguf:Q4_K_M
++      lucy: agent.lucy.yaml      # hf.co/Menlo/Lucy-gguf:Q4_K_M
++  (the key is ignored until implemented: config loading is not strict). Both
++  configs answered a bash-tool prompt live through llmman.
++
++  Decisions (user, 2026-10-02):
++  1. Key in camelCase: `mmConfigs:` (name -> path of a config file).
++  2. Bob's answer is DISPLAYED, and stays in the current history for the
++     NEXT question; the current model does not answer right away.
++  3. Two commands:
++     - /ask <name> <prompt>: bob receives the typed prompt only;
++     - /ask-with-context <name> <prompt>: bob also receives the context of
++       the current conversation.
++     Interpretation to confirm: every /ask starts a FRESH bob session (no
++     memory between two /ask); continuity comes from /ask-with-context.
++  4. Both commands work in the three front ends (Terminal, Window, Editor).
++  5. Context: only with /ask-with-context (see 3).
++
++  Editor visibility (checked): yes. The ACP facade already sends
++  `available_commands_update` built from command.For(Editor)
++  (internal/acp/acp.go availableCommands), and skills are already advertised
++  with `input.unstructured.hint`. Today the built-in commands are sent
++  WITHOUT an input hint: /ask and /ask-with-context need one, e.g.
++  "bob|riker <prompt>", built from mmConfigs. The editor completes the
++  command name, not the argument; alternative to discuss: one command per
++  agent (/ask-bob, /ask-with-context-bob) so the picker lists the agents.
++  The CHILD's own available_commands_update is ignored by the parent.
++
++  Decided too (user, 2026-10-02):
++  a. Context for /ask-with-context = a text transcript of the current
++     history: user questions and assistant answers, each tool call as one
++     line (name + short args), no tool outputs, no reasoning, the compaction
++     summary if any; capped (chars), newest turns kept first. Sent as:
++       [Conversation so far, from <parent agent name>]
++       ...
++       [Question]
++       <prompt>
++  b. Every /ask (and /ask-with-context) starts a FRESH bob session
++     (session/new); no memory on bob's side between two calls.
++  c. Editors: ONE COMMAND PER AGENT, generated from mmConfigs —
++     /ask-bob, /ask-with-context-bob, /ask-riker, … each with an input hint
++     "<prompt>". Consequences:
++     - agent names become command names: validate them in config
++       (lowercase letters, digits, '-'; error otherwise);
++     - a skill or built-in with the same name: the existing guard in
++       availableCommands (commandNamed) keeps the command and drops the skill,
++       with the banner warning — extend it to the generated names;
++     - terminal and window (confirmed by the user, 2026-10-02): accept the
++       per-agent forms there too (same table, no drift), and keep
++       `/ask <name> <prompt>` as the generic form; the window's "/" picker
++       lists the per-agent commands.
++
++  Proposed behaviour (to confirm):
++  - One child per name, started lazily on the first /ask, kept alive for the
++    parent session (start-up cost), killed on /quit and exit; a fresh ACP
++    session (session/new) per /ask if 3 is confirmed.
++  - The /ask exchange enters the CURRENT history as one user message, e.g.
++      [bob (ollama/qwen3.5:2b) was asked: <prompt>]
++      <bob's answer>
++    so the current model sees it on the next question and the session file
++    persists and replays it on /resume. Must keep the history valid for
++    the next model call (user/model alternation: merge into the next user
++    message, or a user+model pair "[noted]" — to decide in the code).
++  - Bob's answer is shown while it streams, prefixed/styled as bob's
++    (decided: displayed, kept for the next question).
++  - Bob's thoughts: not shown (or dim), never inserted in the history.
++  - Long answers: capped (maxOutput-like) before insertion — the current
++    model's context pays for them.
++
++  Risks / things to decide:
++  - Ticket 57: a permission forwarded to the user and answered slowly will
++    trip the CHILD's watchdog exactly like today in an editor.
++  - Ticket 59 (yolo): which permission policy for the child — forward to the
++    user (default), or auto per config entry?
++  - Two agents on the same working tree: the parent waits while bob works
++    (/ask is blocking), so no concurrent edits; but bob may change files the
++    parent model "remembers" — say so in the inserted message?
++  - Recursion: /ask is a USER command, the model cannot call it, so no
++    automatic loop. If it ever becomes a tool ("ask_agent"), cap the depth
++    (env var MM_ASK_DEPTH) and refuse cycles.
++  - Paths in mm-configs: relative to the PARENT config file's directory (the
++    ACP mode already resolves skillsDir that way; a path relative to the cwd
++    breaks when the config is elsewhere). The child's own relative paths
++    (skillsDir) resolve against ITS config dir in -acp mode.
++  - Start-up cost: Genkit init + probe + MCP servers per child (~1 s+);
++    hence keep the child alive.
++  - Child stderr: capture it and show it only on failure (start error,
++    unknown config, provider down -> one line: "bob: nothing answers at …").
++
++  Decisions (user, 2026-10-02):
++  1. Key in camelCase: `mmConfigs:` (name -> path of a config file).
++  2. Bob's answer is DISPLAYED, and stays in the current history for the
++     NEXT question; the current model does not answer right away.
++  3. Two commands:
++     - /ask <name> <prompt>: bob receives the typed prompt only;
++     - /ask-with-context <name> <prompt>: bob also receives the context of
++       the current conversation.
++     Interpretation to confirm: every /ask starts a FRESH bob session (no
++     memory between two /ask); continuity comes from /ask-with-context.
++  4. Both commands work in the three front ends (Terminal, Window, Editor).
++  5. Context: only with /ask-with-context (see 3).
++
++  Editor visibility (checked): yes. The ACP facade already sends
++  `available_commands_update` built from command.For(Editor)
++  (internal/acp/acp.go availableCommands), and skills are already advertised
++  with `input.unstructured.hint`. Today the built-in commands are sent
++  WITHOUT an input hint: /ask and /ask-with-context need one, e.g.
++  "bob|riker <prompt>", built from mmConfigs. The editor completes the
++  command name, not the argument; alternative to discuss: one command per
++  agent (/ask-bob, /ask-with-context-bob) so the picker lists the agents.
++  The CHILD's own available_commands_update is ignored by the parent.
++
++  Still open:
++  a. What "context" is for /ask-with-context. Proposal: a text transcript of
++     the current history — user questions and assistant answers, each tool
++     call as one line (name + short args), no tool outputs, no reasoning,
++     the compaction summary if any — capped (chars) to protect bob's
++     window, newest turns kept first. Sent as the first part of the prompt:
++       [Conversation so far, from <parent agent name>]
++       ...
++       [Question]
++       <prompt>
++  b. Fresh bob session per /ask (interpretation of 3) — confirm.
++  c. Editor: one /ask with a hint, or one command per agent?
++tasks:
++  - id: 1
++    title: "Answer the open questions (key name, immediate answer, multi-turn, fronts, context passed)"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-02T19:30:00.000Z
++    updatedAt: 2026-10-02T20:00:00.000Z
++  - id: 2
++    title: "Confirm the remaining points a/b/c (context format, fresh session per /ask, editor hint vs one command per agent)"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-02T19:30:00.000Z
++    updatedAt: 2026-10-02T20:00:00.000Z
++  - id: 3
++    title: "config: mmConfigs map (validation: name = [a-z0-9-]+ since it becomes a command name, file exists; paths relative to the config file dir)"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-02T19:30:00.000Z
++    updatedAt: 2026-10-02T20:00:00.000Z
++  - id: 4
++    title: "internal/askagent (name TBD): ACP client over a child `mm -acp -config <path>` (os.Executable), lazy start, one child per name, shutdown"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-02T19:30:00.000Z
++    updatedAt: 2026-10-02T20:00:00.000Z
++  - id: 5
++    title: "Client side: SessionUpdate -> stream text/tool-call lines to the current sink; RequestPermission -> current sink.Allow; fs/terminal stubs"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-02T19:30:00.000Z
++    updatedAt: 2026-10-02T20:00:00.000Z
++  - id: 6
++    title: "command table: /ask <name> <prompt>, /ask-with-context <name> <prompt>, and per-agent /ask-<name>, /ask-with-context-<name> generated from mmConfigs (all fronts); ACP advertises the per-agent ones with hint <prompt>; window picker; name clash guard"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-02T19:30:00.000Z
++    updatedAt: 2026-10-02T20:00:00.000Z
++  - id: 7
++    title: "Context builder for /ask-with-context: transcript of the current history (text + one line per tool call), capped"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-02T19:30:00.000Z
++    updatedAt: 2026-10-02T20:00:00.000Z
++  - id: 8
++    title: "Insert the exchange in the current history (format, cap, valid alternation), persisted by the store and replayed on /resume"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-02T19:30:00.000Z
++    updatedAt: 2026-10-02T20:00:00.000Z
++  - id: 9
++    title: "/abort, Ctrl+C and editor cancel -> session/cancel on the child; child crash -> one-line error, history untouched"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-02T19:30:00.000Z
++    updatedAt: 2026-10-02T20:00:00.000Z
++  - id: 10
++    title: "Tests: fake child agent (acp-go-sdk agent side) for the client; end-to-end with two real configs (demo-ollama + demo-llmman)"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-02T19:30:00.000Z
++    updatedAt: 2026-10-02T20:00:00.000Z
++  - id: 11
++    title: "Docs EN+FR: how-to ask another agent, reference (config key, commands), explanation (why ACP, why a subprocess)"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-02T19:30:00.000Z
++    updatedAt: 2026-10-02T20:00:00.000Z
++  - id: 12
++    title: "Quality gate (qlty) where available — not run: qlty absent from this sandbox"
++    state: open
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-02T20:30:00.000Z
++    updatedAt: 2026-10-02T20:30:00.000Z
++  - id: 13
++    title: "Live check of the window mode (/ask-jan in mm -tui) — unit-tested only"
++    state: open
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-02T20:30:00.000Z
++    updatedAt: 2026-10-02T20:30:00.000Z
++  - id: 14
++    title: "Review, commit, merge"
++    state: open
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-02T20:30:00.000Z
++    updatedAt: 2026-10-02T20:30:00.000Z
++comments:
++  - id: 1
++    author:
++      name: claude
++      email: noreply@anthropic.com
++    createdAt: 2026-10-02T20:30:00.000Z
++    body: |
++      **Implemented** on branch `feature/talk-to-another-mm` (uncommitted, 2026-10-02).
++
++      - config: `mmConfigs` (internal/config/agents.go) — names [a-z0-9-], not
++        starting with `with-context`; relative paths resolved against the config
++        file's dir at Load; missing file = start-up warning. `AGENT_SESSIONS=off`
++        env override (applyEnv).
++      - command: /ask family generated by SetAgents (internal/command/ask.go);
++        `Hint` field; ParseAsk returns (req, isAsk, err).
++      - internal/ask (new): Transcript/Prompt (context.go), Exchange as a
++        user+model pair + Describe (exchange.go), ChildEnv/Depth/CanAsk
++        (env.go, MaxDepth 2), ACP client (client.go), Pool with one lazy child
++        per agent + fresh session per ask (pool.go), Pool.Run + Runner interface
++        + Label + Missing (run.go).
++      - Front ends: agent/ask.go (Ctrl+C, /abort, /quit), tui/ask.go (job
++        "asking <agent>"), acp/ask.go (per-agent commands with `<prompt>` hint);
++        main.go wiring; store.FailAsk.
++      - Traps found: the SDK dispatches the child's requests concurrently with
++        its notifications — a permission request can overtake its tool_call
++        (client shows the call first, once); the SDK's default slog would print
++        "connection closed" on the user's terminal (silenced); the parent's
++        AGENT_* env would override the child's config (stripped).
++      - Tests: config, command, ask (fake ACP agent = the test binary re-run,
++        no network), agent, tui, acp. `go test ./...` green 3x; ask tests 15x.
++        `-race` not available (no C compiler in the sandbox).
++      - Live (host llmman): line mode `/ask-jan` → answer kept, the current
++        model quoted it next turn; ACP `/ask-with-context-lucy` → per-agent
++        commands announced, Lucy knew the name given earlier, her bash
++        permission went to the client titled "lucy · bash: …"; no child
++        session files, no leftover processes.
++      - Docs EN+FR: how-to/ask-another-agent.md, explanation/asking-another-agent.md,
++        reference (slash-commands, configuration, cli), architecture table,
++        READMEs, docs/diagrams/packages.drawio (internal/ask + edges).

added demo-llmman/agent.jan-nano.yaml +199 -0 Viewed

new file mode 100644
@@ -0,0 +1,199 @@
	1	+# Agent settings — Jan: mini-me on hf.co/Menlo/Jan-nano-gguf:Q4_K_M, served by llmman (https://github.com/llmmanorg/llmman)
	2	+# through its OpenAI-compatible endpoint (/v1/chat/completions).
	3	+#
	4	+# llmman pull hf.co/Menlo/Jan-nano-gguf:Q4_K_M # optional: the first question pulls it otherwise
	5	+# llmman serve # 127.0.0.1:17434
	6	+#
	7	+# llmman loads the model on the first request (in its own llama-server) and
	8	+# unloads it after five idle minutes.
	9	+provider: llmman
	10	+
	11	+# The model id exactly as `llmman list` / GET /v1/models spells it.
	12	+model: hf.co/Menlo/Jan-nano-gguf:Q4_K_M
	13	+
	14	+# `llmman serve`'s default bind; the OpenAI routes sit under /v1.
	15	+baseUrl: http://127.0.0.1:17434/v1
	16	+
	17	+# From a Docker sandbox, the daemon runs on the host.
	18	+fallback: http://host.docker.internal:17434/v1
	19	+
	20	+# Only when the daemon was started with LLMMAN_API_KEYS (mandatory when it
	21	+# listens beyond loopback); unset otherwise. The key never goes in this file.
	22	+# apiKeyEnv: LLMMAN_API_KEY
	23	+
	24	+# The `bash` tool.
	25	+# `false` removes it: the model can still read and edit files (editTools) and load skills,
	26	+# but never runs a command. If you turn it off, reword the system prompt below — it says "You have a bash tool".
	27	+bashTool: true
	28	+
	29	+# The built-in file tools: read_file, write_file, edit_file.
	30	+editTools: true
	31	+
	32	+# The built-in plan tool for managing multi-step plans.
	33	+# Set to false to disable plan creation and tracking via the `plan` tool.
	34	+planTool: true
	35	+
	36	+# Language servers for the `lsp` tool (definition, references, hover, symbols, diagnostics, rename).
	37	+# The file extension picks the server. A server missing from the PATH is skipped with a warning.
	38	+# No server listed = no `lsp` tool. Examples (uncomment what you have installed):
	39	+# lsp:
	40	+# servers:
	41	+# - name: gopls
	42	+# command: gopls
	43	+# extensions: [".go"]
	44	+# - name: typescript
	45	+# command: typescript-language-server
	46	+# args: ["--stdio"]
	47	+# extensions: [".ts", ".tsx", ".js", ".jsx"]
	48	+# - name: python
	49	+# command: pyright-langserver
	50	+# args: ["--stdio"]
	51	+# extensions: [".py"]
	52	+# - name: rust
	53	+# command: rust-analyzer
	54	+# extensions: [".rs"]
	55	+
	56	+# How many tokens the server actually serves.
	57	+# 0 = ask llmman: /api/ps tells it once the model is loaded, so the window is
	58	+# learned before the second question. Set it here if you want it from the start
	59	+# — and to the same value as LLMMAN_CONTEXT_LENGTH if you set that variable.
	60	+contextWindow: 0
	61	+
	62	+# Max number of characters a tool returns to the model.
	63	+# Context safeguard: beyond that, the output is truncated (beginning + end kept).
	64	+maxOutput: 16000
	65	+
	66	+# Max number of model <-> tools round trips for a single question.
	67	+# Raise it for tasks that need long chains of commands.
	68	+maxTurns: 40
	69	+
	70	+# Where the `read_skill` tool looks for its markdown procedures.
	71	+# The path is relative to the CURRENT DIRECTORY: run the agent from demo-llmman/.
	72	+skillsDir: ../skills
	73	+
	74	+# Persistent sessions, per project.
	75	+# When enabled, every conversation is saved as one JSON file under `dir`.
	76	+# `/sessions` lists them, `/resume` continues one; in an editor, "Reload Agent" replays the thread.
	77	+sessions:
	78	+ enabled: true
	79	+ dir: .mm/sessions
	80	+
	81	+# How many lines of a command's output are echoed to the terminal.
	82	+# 0 disables it.
	83	+previewLines: 20
	84	+
	85	+displayCommands: true
	86	+# Show what a thinking model reasons about before it answers (dim italic in the
	87	+# terminal, its own style in the window, a thought chunk in an editor).
	88	+# The reasoning is always READ — it is what keeps the watchdog alive while the
	89	+# model thinks — this only says whether to show it.
	90	+showThinking: true
	91	+
	92	+
	93	+# The system prompt: what the agent is, and what it is allowed to do.
	94	+# This is the knob to play with — it is the shortest path to changing behaviour.
	95	+system: \|
	96	+ Your name is Jan.
	97	+ You are a coding agent working in a terminal.
	98	+ You have a "bash" tool to run shell commands.
	99	+ Use it to explore files, run tests, inspect the repository, etc.
	100	+ Chain several commands if needed, then answer clearly in English.
	101	+
	102	+ A request often mixes things you answer from yourself ("say hello") with
	103	+ things only a command can answer ("list the files"). Handle every part, in
	104	+ the order asked, and run a command for each part that needs one.
	105	+ Never state the contents of a file, the output of a command, or the state of
	106	+ the repository unless a command in THIS answer returned it. What you did not
	107	+ read, you do not know: run the command instead of recalling it.
	108	+
	109	+ SKILLS
	110	+ You have a second tool, `read_skill`. Its description lists the procedures
	111	+ available for this project — one per kind of task.
	112	+
	113	+ Any request to DO something to a Go project is a skill, not a shell command
	114	+ you invent. Match the request against that list, call `read_skill` FIRST,
	115	+ before any bash command, and then follow what it says step by step.
	116	+
	117	+ FILE EDITING
	118	+ You have three tools for files: `read_file`, `edit_file` and `write_file`.
	119	+ They are how a file gets read and changed here: each change is exact,
	120	+ checked before it is written, and comes back as a diff with line numbers.
	121	+ bash is for running things — building, testing, listing, searching.
	122	+
	123	+ - Read before you write: call `read_file` on the file (numbered=true when
	124	+ you need line numbers). You cannot target text you have not seen; never
	125	+ rely on what you think you remember about a file.
	126	+ - To change an existing file, call `edit_file` with one or more {old, new}
	127	+ pairs. `old` is copied from the file character for character — same
	128	+ spaces, same indentation, same line breaks — and appears exactly once:
	129	+ add the surrounding lines until it is unique. Several pairs are applied
	130	+ together, against the original file. An empty `new` deletes the text.
	131	+ - Call `write_file` only to create a file, or to rewrite one entirely and
	132	+ on purpose. On an existing file it replaces everything, including what
	133	+ you did not intend to touch.
	134	+ - Read the diff the tool returns: it says exactly what changed and on which
	135	+ line. If `edit_file` refuses — text not found, ambiguous, overlapping
	136	+ edits — read the file again and fix `old`. Do not fall back to
	137	+ `write_file` to force the change through.
	138	+ - After editing code, run the narrowest check with bash: the formatter, the
	139	+ compiler, or the test covering that file.
	140	+
	141	+ RULES
	142	+ - Keep everything the file already does, unless the user asked to remove it.
	143	+ - Touch only the files the request is about. Do not add tests, files or
	144	+ features that were not asked for.
	145	+ - Never run a git command unless the user says git, commit or push.
	146	+ - Never move, rename or delete a file unless the user asked for it.
	147	+ - Then answer in English, in a few lines.
	148	+ - If you don't know how to use a <cli>, run `<cli> --help` (or `<cli> help`)
	149	+ to understand the options, then run the command.
	150	+
	151	+ BACKGROUND JOBS
	152	+ Never let a command block the answer. Anything that serves, watches or runs
	153	+ long goes to the background, with BOTH streams redirected and its pid kept:
	154	+
	155	+ nohup <command> > /tmp/<job>.log 2>&1 & echo $! > /tmp/<job>.pid
	156	+
	157	+ Redirecting only stdout still blocks until the process exits. Read the
	158	+ `bg-jobs` skill before you wait on, inspect or stop such a job — each has a
	159	+ rule you cannot guess. Stop every job you started before you finish, and say
	160	+ which ones you left running.
	161	+
	162	+# Generation settings (OpenAI API keys). Low temperature for a coding agent:
	163	+# we want precise and reproducible answers, not creativity.
	164	+sampling:
	165	+ temperature: 0.0
	166	+ parallel_tool_calls: false
	167	+ top_p: 0.9
	168	+ max_tokens: 4096
	169	+
	170	+# Silence allowed between two tokens once the answer has started.
	171	+watchdogTimeout: 30s
	172	+
	173	+# Silence allowed while the model reads the prompt, before the first token:
	174	+# on a laptop a 25k-token prompt can take minutes. llama-server reports its
	175	+# progress meanwhile, through llmman ("Reading the prompt 12.0k/25.3k tokens"), and each report
	176	+# restarts this clock, so it only has to cover one batch of the prompt.
	177	+firstTokenTimeout: 3m
	178	+
	179	+# Context compression.
	180	+# ON here: the agent learns the window from /api/ps, so `threshold` has a number to apply to.
	181	+context:
	182	+ enabled: true
	183	+
	184	+ # Compress when the history reaches this share of the window.
	185	+ threshold: 75
	186	+
	187	+ # Fallback on the message count, for when the window is unknown.
	188	+ # One command costs 2 messages (call + response), so 80 is roughly 30 commands of history.
	189	+ maxMessages: 80
	190	+
	191	+ # How many recent questions (with their commands and answers) are kept as-is instead of being summarised.
	192	+ # The model works on the last few turns, so summarising them would make it redo what it just did.
	193	+ keepLastTurns: 3
	194	+
	195	+ # max_tokens of the summary request. Reasoning is turned off for it on this provider
	196	+ # (a thinking model would otherwise spend the budget thinking and write nothing).
	197	+ summaryMaxTokens: 1200
	198	+
	199	+ showStats: true

++# Agent settings — Jan: mini-me on hf.co/Menlo/Jan-nano-gguf:Q4_K_M, served by llmman (https://github.com/llmmanorg/llmman)
++# through its OpenAI-compatible endpoint (/v1/chat/completions).
++#
++#   llmman pull hf.co/Menlo/Jan-nano-gguf:Q4_K_M   # optional: the first question pulls it otherwise
++#   llmman serve                                       # 127.0.0.1:17434
++#
++# llmman loads the model on the first request (in its own llama-server) and
++# unloads it after five idle minutes.
++provider: llmman
++
++# The model id exactly as `llmman list` / GET /v1/models spells it.
++model: hf.co/Menlo/Jan-nano-gguf:Q4_K_M
++
++# `llmman serve`'s default bind; the OpenAI routes sit under /v1.
++baseUrl: http://127.0.0.1:17434/v1
++
++# From a Docker sandbox, the daemon runs on the host.
++fallback: http://host.docker.internal:17434/v1
++
++# Only when the daemon was started with LLMMAN_API_KEYS (mandatory when it
++# listens beyond loopback); unset otherwise. The key never goes in this file.
++# apiKeyEnv: LLMMAN_API_KEY
++
++# The `bash` tool.
++# `false` removes it: the model can still read and edit files (editTools) and load skills,
++# but never runs a command. If you turn it off, reword the system prompt below — it says "You have a bash tool".
++bashTool: true
++
++# The built-in file tools: read_file, write_file, edit_file.
++editTools: true
++
++# The built-in plan tool for managing multi-step plans.
++# Set to false to disable plan creation and tracking via the `plan` tool.
++planTool: true
++
++# Language servers for the `lsp` tool (definition, references, hover, symbols, diagnostics, rename).
++# The file extension picks the server. A server missing from the PATH is skipped with a warning.
++# No server listed = no `lsp` tool. Examples (uncomment what you have installed):
++# lsp:
++#   servers:
++#     - name: gopls
++#       command: gopls
++#       extensions: [".go"]
++#     - name: typescript
++#       command: typescript-language-server
++#       args: ["--stdio"]
++#       extensions: [".ts", ".tsx", ".js", ".jsx"]
++#     - name: python
++#       command: pyright-langserver
++#       args: ["--stdio"]
++#       extensions: [".py"]
++#     - name: rust
++#       command: rust-analyzer
++#       extensions: [".rs"]
++
++# How many tokens the server actually serves.
++# 0 = ask llmman: /api/ps tells it once the model is loaded, so the window is
++# learned before the second question. Set it here if you want it from the start
++# — and to the same value as LLMMAN_CONTEXT_LENGTH if you set that variable.
++contextWindow: 0
++
++# Max number of characters a tool returns to the model.
++# Context safeguard: beyond that, the output is truncated (beginning + end kept).
++maxOutput: 16000
++
++# Max number of model <-> tools round trips for a single question.
++# Raise it for tasks that need long chains of commands.
++maxTurns: 40
++
++# Where the `read_skill` tool looks for its markdown procedures.
++# The path is relative to the CURRENT DIRECTORY: run the agent from demo-llmman/.
++skillsDir: ../skills
++
++# Persistent sessions, per project.
++# When enabled, every conversation is saved as one JSON file under `dir`.
++# `/sessions` lists them, `/resume` continues one; in an editor, "Reload Agent" replays the thread.
++sessions:
++  enabled: true
++  dir: .mm/sessions
++
++# How many lines of a command's output are echoed to the terminal.
++# 0 disables it.
++previewLines: 20
++
++displayCommands: true
++# Show what a thinking model reasons about before it answers (dim italic in the
++# terminal, its own style in the window, a thought chunk in an editor).
++# The reasoning is always READ — it is what keeps the watchdog alive while the
++# model thinks — this only says whether to show it.
++showThinking: true
++
++
++# The system prompt: what the agent is, and what it is allowed to do.
++# This is the knob to play with — it is the shortest path to changing behaviour.
++system: |
++  Your name is Jan.
++  You are a coding agent working in a terminal.
++  You have a "bash" tool to run shell commands.
++  Use it to explore files, run tests, inspect the repository, etc.
++  Chain several commands if needed, then answer clearly in English.
++
++  A request often mixes things you answer from yourself ("say hello") with
++  things only a command can answer ("list the files"). Handle every part, in
++  the order asked, and run a command for each part that needs one.
++  Never state the contents of a file, the output of a command, or the state of
++  the repository unless a command in THIS answer returned it. What you did not
++  read, you do not know: run the command instead of recalling it.
++
++  SKILLS
++  You have a second tool, `read_skill`. Its description lists the procedures
++  available for this project — one per kind of task.
++
++  Any request to DO something to a Go project is a skill, not a shell command
++  you invent. Match the request against that list, call `read_skill` FIRST,
++  before any bash command, and then follow what it says step by step.
++
++  FILE EDITING
++  You have three tools for files: `read_file`, `edit_file` and `write_file`.
++  They are how a file gets read and changed here: each change is exact,
++  checked before it is written, and comes back as a diff with line numbers.
++  bash is for running things — building, testing, listing, searching.
++
++  - Read before you write: call `read_file` on the file (numbered=true when
++    you need line numbers). You cannot target text you have not seen; never
++    rely on what you think you remember about a file.
++  - To change an existing file, call `edit_file` with one or more {old, new}
++    pairs. `old` is copied from the file character for character — same
++    spaces, same indentation, same line breaks — and appears exactly once:
++    add the surrounding lines until it is unique. Several pairs are applied
++    together, against the original file. An empty `new` deletes the text.
++  - Call `write_file` only to create a file, or to rewrite one entirely and
++    on purpose. On an existing file it replaces everything, including what
++    you did not intend to touch.
++  - Read the diff the tool returns: it says exactly what changed and on which
++    line. If `edit_file` refuses — text not found, ambiguous, overlapping
++    edits — read the file again and fix `old`. Do not fall back to
++    `write_file` to force the change through.
++  - After editing code, run the narrowest check with bash: the formatter, the
++    compiler, or the test covering that file.
++
++  RULES
++  - Keep everything the file already does, unless the user asked to remove it.
++  - Touch only the files the request is about. Do not add tests, files or
++    features that were not asked for.
++  - Never run a git command unless the user says git, commit or push.
++  - Never move, rename or delete a file unless the user asked for it.
++  - Then answer in English, in a few lines.
++  - If you don't know how to use a <cli>, run `<cli> --help` (or `<cli> help`)
++    to understand the options, then run the command.
++
++  BACKGROUND JOBS
++  Never let a command block the answer. Anything that serves, watches or runs
++  long goes to the background, with BOTH streams redirected and its pid kept:
++
++    nohup <command> > /tmp/<job>.log 2>&1 & echo $! > /tmp/<job>.pid
++
++  Redirecting only stdout still blocks until the process exits. Read the
++  `bg-jobs` skill before you wait on, inspect or stop such a job — each has a
++  rule you cannot guess. Stop every job you started before you finish, and say
++  which ones you left running.
++
++# Generation settings (OpenAI API keys). Low temperature for a coding agent:
++# we want precise and reproducible answers, not creativity.
++sampling:
++  temperature: 0.0
++  parallel_tool_calls: false
++  top_p: 0.9
++  max_tokens: 4096
++
++# Silence allowed between two tokens once the answer has started.
++watchdogTimeout: 30s
++
++# Silence allowed while the model reads the prompt, before the first token:
++# on a laptop a 25k-token prompt can take minutes. llama-server reports its
++# progress meanwhile, through llmman ("Reading the prompt 12.0k/25.3k tokens"), and each report
++# restarts this clock, so it only has to cover one batch of the prompt.
++firstTokenTimeout: 3m
++
++# Context compression.
++# ON here: the agent learns the window from /api/ps, so `threshold` has a number to apply to.
++context:
++  enabled: true
++
++  # Compress when the history reaches this share of the window.
++  threshold: 75
++
++  # Fallback on the message count, for when the window is unknown.
++  # One command costs 2 messages (call + response), so 80 is roughly 30 commands of history.
++  maxMessages: 80
++
++  # How many recent questions (with their commands and answers) are kept as-is instead of being summarised.
++  # The model works on the last few turns, so summarising them would make it redo what it just did.
++  keepLastTurns: 3
++
++  # max_tokens of the summary request. Reasoning is turned off for it on this provider
++  # (a thinking model would otherwise spend the budget thinking and write nothing).
++  summaryMaxTokens: 1200
++
++  showStats: true

modified demo-llmman/agent.llmman.yaml +9 -0 Viewed

@@ -17,6 +17,15 @@ baseUrl: http://127.0.0.1:17434/v1
17	17	# From a Docker sandbox, the daemon runs on the host.
18	18	fallback: http://host.docker.internal:17434/v1
19	19
	20	+# Other mm agents this one can ask. Each name becomes /ask-<name> and
	21	+# /ask-with-context-<name> (also `/ask <name> <prompt>` in the terminal and the
	22	+# window); the answer is shown and kept in the history for the next question.
	23	+# Each agent runs as a child `mm -acp` started on its first /ask. Paths are
	24	+# relative to THIS file's directory. Names: lowercase letters, digits and '-'.
	25	+mmConfigs:
	26	+ jan: agent.jan-nano.yaml
	27	+ lucy: agent.lucy.yaml
	28	+
20	29	# Only when the daemon was started with LLMMAN_API_KEYS (mandatory when it
21	30	# listens beyond loopback); unset otherwise. The key never goes in this file.
22	31	# apiKeyEnv: LLMMAN_API_KEY

  # From a Docker sandbox, the daemon runs on the host.
  fallback: http://host.docker.internal:17434/v1
++# Other mm agents this one can ask. Each name becomes /ask-<name> and
++# /ask-with-context-<name> (also `/ask <name> <prompt>` in the terminal and the
++# window); the answer is shown and kept in the history for the next question.
++# Each agent runs as a child `mm -acp` started on its first /ask. Paths are
++# relative to THIS file's directory. Names: lowercase letters, digits and '-'.
++mmConfigs:
++  jan: agent.jan-nano.yaml
++  lucy: agent.lucy.yaml
++
  # Only when the daemon was started with LLMMAN_API_KEYS (mandatory when it
  # listens beyond loopback); unset otherwise. The key never goes in this file.
  # apiKeyEnv: LLMMAN_API_KEY

added demo-llmman/agent.lucy.yaml +199 -0 Viewed

new file mode 100644
@@ -0,0 +1,199 @@
	1	+# Agent settings — Lucy: mini-me on hf.co/Menlo/Lucy-gguf:Q4_K_M, served by llmman (https://github.com/llmmanorg/llmman)
	2	+# through its OpenAI-compatible endpoint (/v1/chat/completions).
	3	+#
	4	+# llmman pull hf.co/Menlo/Lucy-gguf:Q4_K_M # optional: the first question pulls it otherwise
	5	+# llmman serve # 127.0.0.1:17434
	6	+#
	7	+# llmman loads the model on the first request (in its own llama-server) and
	8	+# unloads it after five idle minutes.
	9	+provider: llmman
	10	+
	11	+# The model id exactly as `llmman list` / GET /v1/models spells it.
	12	+model: hf.co/Menlo/Lucy-gguf:Q4_K_M
	13	+
	14	+# `llmman serve`'s default bind; the OpenAI routes sit under /v1.
	15	+baseUrl: http://127.0.0.1:17434/v1
	16	+
	17	+# From a Docker sandbox, the daemon runs on the host.
	18	+fallback: http://host.docker.internal:17434/v1
	19	+
	20	+# Only when the daemon was started with LLMMAN_API_KEYS (mandatory when it
	21	+# listens beyond loopback); unset otherwise. The key never goes in this file.
	22	+# apiKeyEnv: LLMMAN_API_KEY
	23	+
	24	+# The `bash` tool.
	25	+# `false` removes it: the model can still read and edit files (editTools) and load skills,
	26	+# but never runs a command. If you turn it off, reword the system prompt below — it says "You have a bash tool".
	27	+bashTool: true
	28	+
	29	+# The built-in file tools: read_file, write_file, edit_file.
	30	+editTools: true
	31	+
	32	+# The built-in plan tool for managing multi-step plans.
	33	+# Set to false to disable plan creation and tracking via the `plan` tool.
	34	+planTool: true
	35	+
	36	+# Language servers for the `lsp` tool (definition, references, hover, symbols, diagnostics, rename).
	37	+# The file extension picks the server. A server missing from the PATH is skipped with a warning.
	38	+# No server listed = no `lsp` tool. Examples (uncomment what you have installed):
	39	+# lsp:
	40	+# servers:
	41	+# - name: gopls
	42	+# command: gopls
	43	+# extensions: [".go"]
	44	+# - name: typescript
	45	+# command: typescript-language-server
	46	+# args: ["--stdio"]
	47	+# extensions: [".ts", ".tsx", ".js", ".jsx"]
	48	+# - name: python
	49	+# command: pyright-langserver
	50	+# args: ["--stdio"]
	51	+# extensions: [".py"]
	52	+# - name: rust
	53	+# command: rust-analyzer
	54	+# extensions: [".rs"]
	55	+
	56	+# How many tokens the server actually serves.
	57	+# 0 = ask llmman: /api/ps tells it once the model is loaded, so the window is
	58	+# learned before the second question. Set it here if you want it from the start
	59	+# — and to the same value as LLMMAN_CONTEXT_LENGTH if you set that variable.
	60	+contextWindow: 0
	61	+
	62	+# Max number of characters a tool returns to the model.
	63	+# Context safeguard: beyond that, the output is truncated (beginning + end kept).
	64	+maxOutput: 16000
	65	+
	66	+# Max number of model <-> tools round trips for a single question.
	67	+# Raise it for tasks that need long chains of commands.
	68	+maxTurns: 40
	69	+
	70	+# Where the `read_skill` tool looks for its markdown procedures.
	71	+# The path is relative to the CURRENT DIRECTORY: run the agent from demo-llmman/.
	72	+skillsDir: ../skills
	73	+
	74	+# Persistent sessions, per project.
	75	+# When enabled, every conversation is saved as one JSON file under `dir`.
	76	+# `/sessions` lists them, `/resume` continues one; in an editor, "Reload Agent" replays the thread.
	77	+sessions:
	78	+ enabled: true
	79	+ dir: .mm/sessions
	80	+
	81	+# How many lines of a command's output are echoed to the terminal.
	82	+# 0 disables it.
	83	+previewLines: 20
	84	+
	85	+displayCommands: true
	86	+# Show what a thinking model reasons about before it answers (dim italic in the
	87	+# terminal, its own style in the window, a thought chunk in an editor).
	88	+# The reasoning is always READ — it is what keeps the watchdog alive while the
	89	+# model thinks — this only says whether to show it.
	90	+showThinking: true
	91	+
	92	+
	93	+# The system prompt: what the agent is, and what it is allowed to do.
	94	+# This is the knob to play with — it is the shortest path to changing behaviour.
	95	+system: \|
	96	+ Your name is Lucy.
	97	+ You are a coding agent working in a terminal.
	98	+ You have a "bash" tool to run shell commands.
	99	+ Use it to explore files, run tests, inspect the repository, etc.
	100	+ Chain several commands if needed, then answer clearly in English.
	101	+
	102	+ A request often mixes things you answer from yourself ("say hello") with
	103	+ things only a command can answer ("list the files"). Handle every part, in
	104	+ the order asked, and run a command for each part that needs one.
	105	+ Never state the contents of a file, the output of a command, or the state of
	106	+ the repository unless a command in THIS answer returned it. What you did not
	107	+ read, you do not know: run the command instead of recalling it.
	108	+
	109	+ SKILLS
	110	+ You have a second tool, `read_skill`. Its description lists the procedures
	111	+ available for this project — one per kind of task.
	112	+
	113	+ Any request to DO something to a Go project is a skill, not a shell command
	114	+ you invent. Match the request against that list, call `read_skill` FIRST,
	115	+ before any bash command, and then follow what it says step by step.
	116	+
	117	+ FILE EDITING
	118	+ You have three tools for files: `read_file`, `edit_file` and `write_file`.
	119	+ They are how a file gets read and changed here: each change is exact,
	120	+ checked before it is written, and comes back as a diff with line numbers.
	121	+ bash is for running things — building, testing, listing, searching.
	122	+
	123	+ - Read before you write: call `read_file` on the file (numbered=true when
	124	+ you need line numbers). You cannot target text you have not seen; never
	125	+ rely on what you think you remember about a file.
	126	+ - To change an existing file, call `edit_file` with one or more {old, new}
	127	+ pairs. `old` is copied from the file character for character — same
	128	+ spaces, same indentation, same line breaks — and appears exactly once:
	129	+ add the surrounding lines until it is unique. Several pairs are applied
	130	+ together, against the original file. An empty `new` deletes the text.
	131	+ - Call `write_file` only to create a file, or to rewrite one entirely and
	132	+ on purpose. On an existing file it replaces everything, including what
	133	+ you did not intend to touch.
	134	+ - Read the diff the tool returns: it says exactly what changed and on which
	135	+ line. If `edit_file` refuses — text not found, ambiguous, overlapping
	136	+ edits — read the file again and fix `old`. Do not fall back to
	137	+ `write_file` to force the change through.
	138	+ - After editing code, run the narrowest check with bash: the formatter, the
	139	+ compiler, or the test covering that file.
	140	+
	141	+ RULES
	142	+ - Keep everything the file already does, unless the user asked to remove it.
	143	+ - Touch only the files the request is about. Do not add tests, files or
	144	+ features that were not asked for.
	145	+ - Never run a git command unless the user says git, commit or push.
	146	+ - Never move, rename or delete a file unless the user asked for it.
	147	+ - Then answer in English, in a few lines.
	148	+ - If you don't know how to use a <cli>, run `<cli> --help` (or `<cli> help`)
	149	+ to understand the options, then run the command.
	150	+
	151	+ BACKGROUND JOBS
	152	+ Never let a command block the answer. Anything that serves, watches or runs
	153	+ long goes to the background, with BOTH streams redirected and its pid kept:
	154	+
	155	+ nohup <command> > /tmp/<job>.log 2>&1 & echo $! > /tmp/<job>.pid
	156	+
	157	+ Redirecting only stdout still blocks until the process exits. Read the
	158	+ `bg-jobs` skill before you wait on, inspect or stop such a job — each has a
	159	+ rule you cannot guess. Stop every job you started before you finish, and say
	160	+ which ones you left running.
	161	+
	162	+# Generation settings (OpenAI API keys). Low temperature for a coding agent:
	163	+# we want precise and reproducible answers, not creativity.
	164	+sampling:
	165	+ temperature: 0.0
	166	+ parallel_tool_calls: false
	167	+ top_p: 0.9
	168	+ max_tokens: 4096
	169	+
	170	+# Silence allowed between two tokens once the answer has started.
	171	+watchdogTimeout: 30s
	172	+
	173	+# Silence allowed while the model reads the prompt, before the first token:
	174	+# on a laptop a 25k-token prompt can take minutes. llama-server reports its
	175	+# progress meanwhile, through llmman ("Reading the prompt 12.0k/25.3k tokens"), and each report
	176	+# restarts this clock, so it only has to cover one batch of the prompt.
	177	+firstTokenTimeout: 3m
	178	+
	179	+# Context compression.
	180	+# ON here: the agent learns the window from /api/ps, so `threshold` has a number to apply to.
	181	+context:
	182	+ enabled: true
	183	+
	184	+ # Compress when the history reaches this share of the window.
	185	+ threshold: 75
	186	+
	187	+ # Fallback on the message count, for when the window is unknown.
	188	+ # One command costs 2 messages (call + response), so 80 is roughly 30 commands of history.
	189	+ maxMessages: 80
	190	+
	191	+ # How many recent questions (with their commands and answers) are kept as-is instead of being summarised.
	192	+ # The model works on the last few turns, so summarising them would make it redo what it just did.
	193	+ keepLastTurns: 3
	194	+
	195	+ # max_tokens of the summary request. Reasoning is turned off for it on this provider
	196	+ # (a thinking model would otherwise spend the budget thinking and write nothing).
	197	+ summaryMaxTokens: 1200
	198	+
	199	+ showStats: true

++# Agent settings — Lucy: mini-me on hf.co/Menlo/Lucy-gguf:Q4_K_M, served by llmman (https://github.com/llmmanorg/llmman)
++# through its OpenAI-compatible endpoint (/v1/chat/completions).
++#
++#   llmman pull hf.co/Menlo/Lucy-gguf:Q4_K_M   # optional: the first question pulls it otherwise
++#   llmman serve                                       # 127.0.0.1:17434
++#
++# llmman loads the model on the first request (in its own llama-server) and
++# unloads it after five idle minutes.
++provider: llmman
++
++# The model id exactly as `llmman list` / GET /v1/models spells it.
++model: hf.co/Menlo/Lucy-gguf:Q4_K_M
++
++# `llmman serve`'s default bind; the OpenAI routes sit under /v1.
++baseUrl: http://127.0.0.1:17434/v1
++
++# From a Docker sandbox, the daemon runs on the host.
++fallback: http://host.docker.internal:17434/v1
++
++# Only when the daemon was started with LLMMAN_API_KEYS (mandatory when it
++# listens beyond loopback); unset otherwise. The key never goes in this file.
++# apiKeyEnv: LLMMAN_API_KEY
++
++# The `bash` tool.
++# `false` removes it: the model can still read and edit files (editTools) and load skills,
++# but never runs a command. If you turn it off, reword the system prompt below — it says "You have a bash tool".
++bashTool: true
++
++# The built-in file tools: read_file, write_file, edit_file.
++editTools: true
++
++# The built-in plan tool for managing multi-step plans.
++# Set to false to disable plan creation and tracking via the `plan` tool.
++planTool: true
++
++# Language servers for the `lsp` tool (definition, references, hover, symbols, diagnostics, rename).
++# The file extension picks the server. A server missing from the PATH is skipped with a warning.
++# No server listed = no `lsp` tool. Examples (uncomment what you have installed):
++# lsp:
++#   servers:
++#     - name: gopls
++#       command: gopls
++#       extensions: [".go"]
++#     - name: typescript
++#       command: typescript-language-server
++#       args: ["--stdio"]
++#       extensions: [".ts", ".tsx", ".js", ".jsx"]
++#     - name: python
++#       command: pyright-langserver
++#       args: ["--stdio"]
++#       extensions: [".py"]
++#     - name: rust
++#       command: rust-analyzer
++#       extensions: [".rs"]
++
++# How many tokens the server actually serves.
++# 0 = ask llmman: /api/ps tells it once the model is loaded, so the window is
++# learned before the second question. Set it here if you want it from the start
++# — and to the same value as LLMMAN_CONTEXT_LENGTH if you set that variable.
++contextWindow: 0
++
++# Max number of characters a tool returns to the model.
++# Context safeguard: beyond that, the output is truncated (beginning + end kept).
++maxOutput: 16000
++
++# Max number of model <-> tools round trips for a single question.
++# Raise it for tasks that need long chains of commands.
++maxTurns: 40
++
++# Where the `read_skill` tool looks for its markdown procedures.
++# The path is relative to the CURRENT DIRECTORY: run the agent from demo-llmman/.
++skillsDir: ../skills
++
++# Persistent sessions, per project.
++# When enabled, every conversation is saved as one JSON file under `dir`.
++# `/sessions` lists them, `/resume` continues one; in an editor, "Reload Agent" replays the thread.
++sessions:
++  enabled: true
++  dir: .mm/sessions
++
++# How many lines of a command's output are echoed to the terminal.
++# 0 disables it.
++previewLines: 20
++
++displayCommands: true
++# Show what a thinking model reasons about before it answers (dim italic in the
++# terminal, its own style in the window, a thought chunk in an editor).
++# The reasoning is always READ — it is what keeps the watchdog alive while the
++# model thinks — this only says whether to show it.
++showThinking: true
++
++
++# The system prompt: what the agent is, and what it is allowed to do.
++# This is the knob to play with — it is the shortest path to changing behaviour.
++system: |
++  Your name is Lucy.
++  You are a coding agent working in a terminal.
++  You have a "bash" tool to run shell commands.
++  Use it to explore files, run tests, inspect the repository, etc.
++  Chain several commands if needed, then answer clearly in English.
++
++  A request often mixes things you answer from yourself ("say hello") with
++  things only a command can answer ("list the files"). Handle every part, in
++  the order asked, and run a command for each part that needs one.
++  Never state the contents of a file, the output of a command, or the state of
++  the repository unless a command in THIS answer returned it. What you did not
++  read, you do not know: run the command instead of recalling it.
++
++  SKILLS
++  You have a second tool, `read_skill`. Its description lists the procedures
++  available for this project — one per kind of task.
++
++  Any request to DO something to a Go project is a skill, not a shell command
++  you invent. Match the request against that list, call `read_skill` FIRST,
++  before any bash command, and then follow what it says step by step.
++
++  FILE EDITING
++  You have three tools for files: `read_file`, `edit_file` and `write_file`.
++  They are how a file gets read and changed here: each change is exact,
++  checked before it is written, and comes back as a diff with line numbers.
++  bash is for running things — building, testing, listing, searching.
++
++  - Read before you write: call `read_file` on the file (numbered=true when
++    you need line numbers). You cannot target text you have not seen; never
++    rely on what you think you remember about a file.
++  - To change an existing file, call `edit_file` with one or more {old, new}
++    pairs. `old` is copied from the file character for character — same
++    spaces, same indentation, same line breaks — and appears exactly once:
++    add the surrounding lines until it is unique. Several pairs are applied
++    together, against the original file. An empty `new` deletes the text.
++  - Call `write_file` only to create a file, or to rewrite one entirely and
++    on purpose. On an existing file it replaces everything, including what
++    you did not intend to touch.
++  - Read the diff the tool returns: it says exactly what changed and on which
++    line. If `edit_file` refuses — text not found, ambiguous, overlapping
++    edits — read the file again and fix `old`. Do not fall back to
++    `write_file` to force the change through.
++  - After editing code, run the narrowest check with bash: the formatter, the
++    compiler, or the test covering that file.
++
++  RULES
++  - Keep everything the file already does, unless the user asked to remove it.
++  - Touch only the files the request is about. Do not add tests, files or
++    features that were not asked for.
++  - Never run a git command unless the user says git, commit or push.
++  - Never move, rename or delete a file unless the user asked for it.
++  - Then answer in English, in a few lines.
++  - If you don't know how to use a <cli>, run `<cli> --help` (or `<cli> help`)
++    to understand the options, then run the command.
++
++  BACKGROUND JOBS
++  Never let a command block the answer. Anything that serves, watches or runs
++  long goes to the background, with BOTH streams redirected and its pid kept:
++
++    nohup <command> > /tmp/<job>.log 2>&1 & echo $! > /tmp/<job>.pid
++
++  Redirecting only stdout still blocks until the process exits. Read the
++  `bg-jobs` skill before you wait on, inspect or stop such a job — each has a
++  rule you cannot guess. Stop every job you started before you finish, and say
++  which ones you left running.
++
++# Generation settings (OpenAI API keys). Low temperature for a coding agent:
++# we want precise and reproducible answers, not creativity.
++sampling:
++  temperature: 0.0
++  parallel_tool_calls: false
++  top_p: 0.9
++  max_tokens: 4096
++
++# Silence allowed between two tokens once the answer has started.
++watchdogTimeout: 30s
++
++# Silence allowed while the model reads the prompt, before the first token:
++# on a laptop a 25k-token prompt can take minutes. llama-server reports its
++# progress meanwhile, through llmman ("Reading the prompt 12.0k/25.3k tokens"), and each report
++# restarts this clock, so it only has to cover one batch of the prompt.
++firstTokenTimeout: 3m
++
++# Context compression.
++# ON here: the agent learns the window from /api/ps, so `threshold` has a number to apply to.
++context:
++  enabled: true
++
++  # Compress when the history reaches this share of the window.
++  threshold: 75
++
++  # Fallback on the message count, for when the window is unknown.
++  # One command costs 2 messages (call + response), so 80 is roughly 30 commands of history.
++  maxMessages: 80
++
++  # How many recent questions (with their commands and answers) are kept as-is instead of being summarised.
++  # The model works on the last few turns, so summarising them would make it redo what it just did.
++  keepLastTurns: 3
++
++  # max_tokens of the summary request. Reasoning is turned off for it on this provider
++  # (a thinking model would otherwise spend the budget thinking and write nothing).
++  summaryMaxTokens: 1200
++
++  showStats: true

modified docs/diagrams/packages.drawio +1 -1 Viewed

@@ -1,3 +1,3 @@
1	1	<mxfile host="app.diagrams.net" modified="2026-09-15T00:00:00.000Z" agent="mm docs" version="24.0.0" type="device"><diagram id="packages" name="packages"><mxGraphModel dx="1400" dy="900" grid="1" gridSize="10" guides="1" tooltips="1" connect="1" arrows="1" fold="1" page="1" pageScale="1" pageWidth="1400" pageHeight="900"><root><mxCell id="0"/><mxCell id="1" parent="0"/><mxCell id="main" value="<b>main (mm)</b><br/><font style="font-size:10px">wiring: config, engine, tools, then one front end</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#fff2cc;strokeColor=#d6b656;" vertex="1" parent="1"><mxGeometry x="420" y="40" width="200" height="60" as="geometry"/></mxCell><mxCell id="agent" value="<b>internal/agent</b><br/><font style="font-size:10px">terminal REPL: /quit /abort /compact /new</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1"><mxGeometry x="160" y="180" width="200" height="60" as="geometry"/></mxCell><mxCell id="acp" value="<b>internal/acp</b><br/><font style="font-size:10px">editor front end: Agent Client Protocol on stdio</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1"><mxGeometry x="700" y="180" width="200" height="60" as="geometry"/></mxCell><mxCell id="session" value="<b>internal/session</b><br/><font style="font-size:10px">what a conversation starts from; the /new reset</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="320" width="200" height="60" as="geometry"/></mxCell><mxCell id="compact" value="<b>internal/compact</b><br/><font style="font-size:10px">history compression into a model-written summary</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="40" y="320" width="200" height="60" as="geometry"/></mxCell><mxCell id="engine" value="<b>internal/engine</b><br/><font style="font-size:10px">LLM connection + generation, one Provider per server</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="700" y="320" width="200" height="60" as="geometry"/></mxCell><mxCell id="tools" value="<b>internal/tools</b><br/><font style="font-size:10px">built-in tools: bash, read_skill, file tools, lsp, fetch_content</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="460" width="200" height="60" as="geometry"/></mxCell><mxCell id="detector" value="<b>internal/detector</b><br/><font style="font-size:10px">repeated-action (loop) detection</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="40" y="600" width="200" height="60" as="geometry"/></mxCell><mxCell id="config" value="<b>internal/config</b><br/><font style="font-size:10px">settings with YAML overrides (agent.yaml)</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="260" y="600" width="200" height="60" as="geometry"/></mxCell><mxCell id="spinner" value="<b>internal/spinner</b><br/><font style="font-size:10px">terminal spinner and styling</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="480" y="600" width="200" height="60" as="geometry"/></mxCell><mxCell id="ui" value="<b>internal/ui</b><br/><font style="font-size:10px">human output destination + shared event sink</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="700" y="600" width="200" height="60" as="geometry"/></mxCell><mxCell id="fileedit" value="<b>internal/fileedit</b><br/><font style="font-size:10px">exact-replacement file editing</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="920" y="600" width="200" height="60" as="geometry"/></mxCell><mxCell id="skills" value="<b>internal/skills</b><br/><font style="font-size:10px">markdown skills catalogue (./skills)</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="1140" y="600" width="200" height="60" as="geometry"/></mxCell><mxCell id="genkit" value="<b>github.com/firebase/genkit</b><br/><font style="font-size:10px">ai.Message, tools, model calls</font>" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1"><mxGeometry x="160" y="760" width="200" height="60" as="geometry"/></mxCell><mxCell id="acpsdk" value="<b>github.com/coder/acp-go-sdk</b><br/><font style="font-size:10px">JSON-RPC transport for ACP</font>" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1"><mxGeometry x="700" y="760" width="200" height="60" as="geometry"/></mxCell><mxCell id="openai" value="<b>github.com/openai/openai-go</b><br/><font style="font-size:10px">OpenAI-compatible HTTP client</font>" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1"><mxGeometry x="920" y="760" width="200" height="60" as="geometry"/></mxCell><mxCell id="yaml" value="<b>gopkg.in/yaml.v3</b><br/><font style="font-size:10px">YAML parsing</font>" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1"><mxGeometry x="400" y="760" width="200" height="60" as="geometry"/></mxCell><mxCell id="e0" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="agent"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="version" value="<b>internal/version</b><br/><font style="font-size:10px">which mm this is: build stamp, else the commit</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="1380" y="600" width="200" height="60" as="geometry"/></mxCell>
2	2	<mxCell id="e61" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="version"><mxGeometry relative="1" as="geometry"/></mxCell>
3		- <mxCell id="e1" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="acp"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e2" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e3" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="detector"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e4" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="engine"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e5" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="skills"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e6" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="spinner"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e7" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="tools"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e8" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e9" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="engine"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e10" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="session"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e11" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e12" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="acp" target="acpsdk"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e13" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="acp" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e14" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="compact"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e15" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e16" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="detector"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e17" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="engine"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e18" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="session"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e19" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="spinner"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e20" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="agent" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e21" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="compact" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e22" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="compact" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e23" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="config" target="yaml"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e24" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="engine" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e25" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="engine" target="spinner"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e26" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="engine" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e27" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="engine" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e28" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="engine" target="openai"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e29" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e30" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="detector"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e31" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="fileedit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e32" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="skills"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e33" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="spinner"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e34" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e35" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="tools" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e36" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="session" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="mention" value="<b>internal/mention</b><br/><font style="font-size:10px">the @path notation of a question</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="400" width="200" height="60" as="geometry"/></mxCell><mxCell id="e37" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="mention"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e38" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="mention"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="store" value="<b>internal/store</b><br/><font style="font-size:10px">per-project session files (.mm/sessions/<id>.json)</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="480" width="200" height="60" as="geometry"/></mxCell><mxCell id="e39" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="store"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e40" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="store"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="tui" value="<b>internal/tui</b><br/><font style="font-size:10px">window front end: one screen on tcell (mm -tui)</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1"><mxGeometry x="430" y="160" width="200" height="60" as="geometry"/></mxCell><mxCell id="command" value="<b>internal/command</b><br/><font style="font-size:10px">the one slash-command table, per front end</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="560" width="200" height="60" as="geometry"/></mxCell><mxCell id="tcell" value="<b>github.com/gdamore/tcell/v2</b><br/><font style="font-size:10px">terminal screen, keys, mouse</font>" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1"><mxGeometry x="920" y="840" width="200" height="60" as="geometry"/></mxCell><mxCell id="e41" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="tui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e42" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e43" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="engine"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e44" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="compact"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e45" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="store"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e46" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="mention"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e47" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="session"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e48" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="detector"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e49" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e50" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="command"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e51" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="tcell"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e52" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="command"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e53" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="command"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e54" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="command" target="session"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="clipboard" value="<b>internal/clipboard</b><br/><font style="font-size:10px">system clipboard: native tool, then OSC 52</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="1140" y="680" width="200" height="60" as="geometry"/></mxCell><mxCell id="e55" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="clipboard"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e56" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="clipboard"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="lsp" value="<b>internal/lsp</b><br/><font style="font-size:10px">language-server client: definition, references, hover, symbols, diagnostics, rename</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="1160" y="600" width="220" height="60" as="geometry"/></mxCell><mxCell id="lspdev" value="<b>go.lsp.dev/protocol + jsonrpc2</b><br/><font style="font-size:10px">LSP types and JSON-RPC 2.0 over stdio</font>" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#ffffff;strokeColor=#999999;" vertex="1" parent="1"><mxGeometry x="1160" y="700" width="220" height="60" as="geometry"/></mxCell><mxCell id="e57" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="lsp"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e58" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="lsp"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e59" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="lsp" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e60" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="lsp" target="lspdev"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="legend" value="<b>mm — package dependencies</b><br/>Arrow: "imports". Yellow: entry point. Blue: front ends. Green: shared logic. Grey: leaf packages. Dashed purple: third-party modules.<br/>Regenerate the edge list from: go list -f '{{.ImportPath}}: {{join .Imports " "}}' ./..." style="text;html=1;whiteSpace=wrap;align=left;verticalAlign=top;fillColor=#ffffff;strokeColor=#999999;" vertex="1" parent="1"><mxGeometry x="900" y="40" width="440" height="90" as="geometry"/></mxCell><mxCell id="webmd" value="<b>internal/webmd</b><br/><font style="font-size:10px">web page → Markdown: fetch (robots.txt, charset → UTF-8), clean, convert</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="1400" y="460" width="220" height="60" as="geometry"/></mxCell><mxCell id="xnet" value="<b>golang.org/x/net/html (+ x/text)</b><br/><font style="font-size:10px">HTML5 parser, charset lookup</font>" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1"><mxGeometry x="1400" y="760" width="220" height="60" as="geometry"/></mxCell><mxCell id="e_main_webmd" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="webmd"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_tools_webmd" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="webmd"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_webmd_xnet" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="webmd" target="xnet"><mxGeometry relative="1" as="geometry"/></mxCell></root></mxGraphModel></diagram></mxfile>
		\ No newline at end of file
	3	+ <mxCell id="e1" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="acp"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e2" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e3" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="detector"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e4" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="engine"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e5" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="skills"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e6" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="spinner"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e7" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="tools"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e8" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e9" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="engine"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e10" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="session"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e11" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e12" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="acp" target="acpsdk"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e13" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="acp" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e14" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="compact"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e15" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e16" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="detector"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e17" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="engine"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e18" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="session"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e19" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="spinner"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e20" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="agent" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e21" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="compact" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e22" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="compact" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e23" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="config" target="yaml"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e24" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="engine" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e25" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="engine" target="spinner"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e26" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="engine" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e27" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="engine" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e28" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="engine" target="openai"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e29" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e30" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="detector"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e31" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="fileedit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e32" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="skills"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e33" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="spinner"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e34" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e35" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="tools" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e36" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="session" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="mention" value="<b>internal/mention</b><br/><font style="font-size:10px">the @path notation of a question</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="400" width="200" height="60" as="geometry"/></mxCell><mxCell id="e37" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="mention"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e38" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="mention"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="store" value="<b>internal/store</b><br/><font style="font-size:10px">per-project session files (.mm/sessions/<id>.json)</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="480" width="200" height="60" as="geometry"/></mxCell><mxCell id="e39" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="store"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e40" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="store"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="tui" value="<b>internal/tui</b><br/><font style="font-size:10px">window front end: one screen on tcell (mm -tui)</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1"><mxGeometry x="430" y="160" width="200" height="60" as="geometry"/></mxCell><mxCell id="command" value="<b>internal/command</b><br/><font style="font-size:10px">the one slash-command table, per front end</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="560" width="200" height="60" as="geometry"/></mxCell><mxCell id="tcell" value="<b>github.com/gdamore/tcell/v2</b><br/><font style="font-size:10px">terminal screen, keys, mouse</font>" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1"><mxGeometry x="920" y="840" width="200" height="60" as="geometry"/></mxCell><mxCell id="e41" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="tui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e42" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e43" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="engine"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e44" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="compact"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e45" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="store"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e46" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="mention"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e47" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="session"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e48" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="detector"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e49" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e50" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="command"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e51" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="tcell"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e52" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="command"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e53" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="command"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e54" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="command" target="session"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="clipboard" value="<b>internal/clipboard</b><br/><font style="font-size:10px">system clipboard: native tool, then OSC 52</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="1140" y="680" width="200" height="60" as="geometry"/></mxCell><mxCell id="e55" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="clipboard"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e56" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="clipboard"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="lsp" value="<b>internal/lsp</b><br/><font style="font-size:10px">language-server client: definition, references, hover, symbols, diagnostics, rename</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="1160" y="600" width="220" height="60" as="geometry"/></mxCell><mxCell id="lspdev" value="<b>go.lsp.dev/protocol + jsonrpc2</b><br/><font style="font-size:10px">LSP types and JSON-RPC 2.0 over stdio</font>" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#ffffff;strokeColor=#999999;" vertex="1" parent="1"><mxGeometry x="1160" y="700" width="220" height="60" as="geometry"/></mxCell><mxCell id="e57" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="lsp"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e58" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="lsp"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e59" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="lsp" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e60" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="lsp" target="lspdev"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="legend" value="<b>mm — package dependencies</b><br/>Arrow: "imports". Yellow: entry point. Blue: front ends. Green: shared logic. Grey: leaf packages. Dashed purple: third-party modules.<br/>Regenerate the edge list from: go list -f '{{.ImportPath}}: {{join .Imports " "}}' ./..." style="text;html=1;whiteSpace=wrap;align=left;verticalAlign=top;fillColor=#ffffff;strokeColor=#999999;" vertex="1" parent="1"><mxGeometry x="900" y="40" width="440" height="90" as="geometry"/></mxCell><mxCell id="webmd" value="<b>internal/webmd</b><br/><font style="font-size:10px">web page → Markdown: fetch (robots.txt, charset → UTF-8), clean, convert</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="1400" y="460" width="220" height="60" as="geometry"/></mxCell><mxCell id="xnet" value="<b>golang.org/x/net/html (+ x/text)</b><br/><font style="font-size:10px">HTML5 parser, charset lookup</font>" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1"><mxGeometry x="1400" y="760" width="220" height="60" as="geometry"/></mxCell><mxCell id="e_main_webmd" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="webmd"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_tools_webmd" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="webmd"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_webmd_xnet" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="webmd" target="xnet"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="ask" value="<b>internal/ask</b><br/><font style="font-size:10px">/ask another mm agent: child mm -acp, ACP client side</font>" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="640" width="200" height="60" as="geometry"/></mxCell><mxCell id="e_main_ask" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="ask"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_agent_ask" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="ask"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_acp_ask" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="ask"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_tui_ask" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="ask"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_ask_command" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="ask" target="command"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_ask_config" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="ask" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_ask_compact" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="ask" target="compact"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_ask_ui" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="ask" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_ask_acpsdk" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="ask" target="acpsdk"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_ask_genkit" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="ask" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_ask_yaml" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="ask" target="yaml"><mxGeometry relative="1" as="geometry"/></mxCell></root></mxGraphModel></diagram></mxfile>
		\ No newline at end of file

  <mxfile host="app.diagrams.net" modified="2026-09-15T00:00:00.000Z" agent="mm docs" version="24.0.0" type="device"><diagram id="packages" name="packages"><mxGraphModel dx="1400" dy="900" grid="1" gridSize="10" guides="1" tooltips="1" connect="1" arrows="1" fold="1" page="1" pageScale="1" pageWidth="1400" pageHeight="900"><root><mxCell id="0"/><mxCell id="1" parent="0"/><mxCell id="main" value="&lt;b&gt;main (mm)&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;wiring: config, engine, tools, then one front end&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#fff2cc;strokeColor=#d6b656;" vertex="1" parent="1"><mxGeometry x="420" y="40" width="200" height="60" as="geometry"/></mxCell><mxCell id="agent" value="&lt;b&gt;internal/agent&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;terminal REPL: /quit /abort /compact /new&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1"><mxGeometry x="160" y="180" width="200" height="60" as="geometry"/></mxCell><mxCell id="acp" value="&lt;b&gt;internal/acp&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;editor front end: Agent Client Protocol on stdio&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1"><mxGeometry x="700" y="180" width="200" height="60" as="geometry"/></mxCell><mxCell id="session" value="&lt;b&gt;internal/session&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;what a conversation starts from; the /new reset&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="320" width="200" height="60" as="geometry"/></mxCell><mxCell id="compact" value="&lt;b&gt;internal/compact&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;history compression into a model-written summary&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="40" y="320" width="200" height="60" as="geometry"/></mxCell><mxCell id="engine" value="&lt;b&gt;internal/engine&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;LLM connection + generation, one Provider per server&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="700" y="320" width="200" height="60" as="geometry"/></mxCell><mxCell id="tools" value="&lt;b&gt;internal/tools&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;built-in tools: bash, read_skill, file tools, lsp, fetch_content&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="460" width="200" height="60" as="geometry"/></mxCell><mxCell id="detector" value="&lt;b&gt;internal/detector&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;repeated-action (loop) detection&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="40" y="600" width="200" height="60" as="geometry"/></mxCell><mxCell id="config" value="&lt;b&gt;internal/config&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;settings with YAML overrides (agent.yaml)&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="260" y="600" width="200" height="60" as="geometry"/></mxCell><mxCell id="spinner" value="&lt;b&gt;internal/spinner&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;terminal spinner and styling&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="480" y="600" width="200" height="60" as="geometry"/></mxCell><mxCell id="ui" value="&lt;b&gt;internal/ui&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;human output destination + shared event sink&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="700" y="600" width="200" height="60" as="geometry"/></mxCell><mxCell id="fileedit" value="&lt;b&gt;internal/fileedit&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;exact-replacement file editing&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="920" y="600" width="200" height="60" as="geometry"/></mxCell><mxCell id="skills" value="&lt;b&gt;internal/skills&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;markdown skills catalogue (./skills)&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="1140" y="600" width="200" height="60" as="geometry"/></mxCell><mxCell id="genkit" value="&lt;b&gt;github.com/firebase/genkit&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;ai.Message, tools, model calls&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1"><mxGeometry x="160" y="760" width="200" height="60" as="geometry"/></mxCell><mxCell id="acpsdk" value="&lt;b&gt;github.com/coder/acp-go-sdk&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;JSON-RPC transport for ACP&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1"><mxGeometry x="700" y="760" width="200" height="60" as="geometry"/></mxCell><mxCell id="openai" value="&lt;b&gt;github.com/openai/openai-go&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;OpenAI-compatible HTTP client&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1"><mxGeometry x="920" y="760" width="200" height="60" as="geometry"/></mxCell><mxCell id="yaml" value="&lt;b&gt;gopkg.in/yaml.v3&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;YAML parsing&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1"><mxGeometry x="400" y="760" width="200" height="60" as="geometry"/></mxCell><mxCell id="e0" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="agent"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="version" value="&lt;b&gt;internal/version&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;which mm this is: build stamp, else the commit&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="1380" y="600" width="200" height="60" as="geometry"/></mxCell>
          <mxCell id="e61" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="version"><mxGeometry relative="1" as="geometry"/></mxCell>
--        <mxCell id="e1" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="acp"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e2" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e3" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="detector"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e4" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="engine"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e5" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="skills"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e6" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="spinner"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e7" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="tools"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e8" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e9" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="engine"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e10" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="session"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e11" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e12" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="acp" target="acpsdk"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e13" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="acp" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e14" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="compact"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e15" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e16" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="detector"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e17" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="engine"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e18" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="session"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e19" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="spinner"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e20" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="agent" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e21" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="compact" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e22" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="compact" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e23" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="config" target="yaml"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e24" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="engine" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e25" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="engine" target="spinner"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e26" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="engine" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e27" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="engine" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e28" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="engine" target="openai"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e29" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e30" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="detector"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e31" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="fileedit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e32" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="skills"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e33" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="spinner"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e34" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e35" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="tools" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e36" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="session" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="mention" value="&lt;b&gt;internal/mention&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;the @path notation of a question&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="400" width="200" height="60" as="geometry"/></mxCell><mxCell id="e37" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="mention"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e38" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="mention"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="store" value="&lt;b&gt;internal/store&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;per-project session files (.mm/sessions/&lt;id&gt;.json)&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="480" width="200" height="60" as="geometry"/></mxCell><mxCell id="e39" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="store"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e40" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="store"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="tui" value="&lt;b&gt;internal/tui&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;window front end: one screen on tcell (mm -tui)&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1"><mxGeometry x="430" y="160" width="200" height="60" as="geometry"/></mxCell><mxCell id="command" value="&lt;b&gt;internal/command&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;the one slash-command table, per front end&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="560" width="200" height="60" as="geometry"/></mxCell><mxCell id="tcell" value="&lt;b&gt;github.com/gdamore/tcell/v2&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;terminal screen, keys, mouse&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1"><mxGeometry x="920" y="840" width="200" height="60" as="geometry"/></mxCell><mxCell id="e41" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="tui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e42" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e43" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="engine"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e44" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="compact"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e45" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="store"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e46" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="mention"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e47" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="session"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e48" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="detector"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e49" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e50" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="command"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e51" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="tcell"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e52" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="command"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e53" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="command"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e54" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="command" target="session"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="clipboard" value="&lt;b&gt;internal/clipboard&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;system clipboard: native tool, then OSC 52&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="1140" y="680" width="200" height="60" as="geometry"/></mxCell><mxCell id="e55" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="clipboard"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e56" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="clipboard"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="lsp" value="&lt;b&gt;internal/lsp&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;language-server client: definition, references, hover, symbols, diagnostics, rename&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="1160" y="600" width="220" height="60" as="geometry"/></mxCell><mxCell id="lspdev" value="&lt;b&gt;go.lsp.dev/protocol + jsonrpc2&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;LSP types and JSON-RPC 2.0 over stdio&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#ffffff;strokeColor=#999999;" vertex="1" parent="1"><mxGeometry x="1160" y="700" width="220" height="60" as="geometry"/></mxCell><mxCell id="e57" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="lsp"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e58" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="lsp"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e59" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="lsp" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e60" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="lsp" target="lspdev"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="legend" value="&lt;b&gt;mm — package dependencies&lt;/b&gt;&lt;br/&gt;Arrow: &quot;imports&quot;. Yellow: entry point. Blue: front ends. Green: shared logic. Grey: leaf packages. Dashed purple: third-party modules.&lt;br/&gt;Regenerate the edge list from: go list -f '{{.ImportPath}}: {{join .Imports &quot; &quot;}}' ./..." style="text;html=1;whiteSpace=wrap;align=left;verticalAlign=top;fillColor=#ffffff;strokeColor=#999999;" vertex="1" parent="1"><mxGeometry x="900" y="40" width="440" height="90" as="geometry"/></mxCell><mxCell id="webmd" value="&lt;b&gt;internal/webmd&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;web page → Markdown: fetch (robots.txt, charset → UTF-8), clean, convert&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="1400" y="460" width="220" height="60" as="geometry"/></mxCell><mxCell id="xnet" value="&lt;b&gt;golang.org/x/net/html (+ x/text)&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;HTML5 parser, charset lookup&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1"><mxGeometry x="1400" y="760" width="220" height="60" as="geometry"/></mxCell><mxCell id="e_main_webmd" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="webmd"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_tools_webmd" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="webmd"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_webmd_xnet" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="webmd" target="xnet"><mxGeometry relative="1" as="geometry"/></mxCell></root></mxGraphModel></diagram></mxfile>
 \ No newline at end of file
++        <mxCell id="e1" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="acp"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e2" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e3" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="detector"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e4" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="engine"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e5" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="skills"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e6" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="spinner"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e7" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="tools"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e8" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e9" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="engine"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e10" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="session"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e11" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e12" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="acp" target="acpsdk"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e13" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="acp" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e14" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="compact"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e15" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e16" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="detector"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e17" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="engine"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e18" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="session"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e19" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="spinner"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e20" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="agent" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e21" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="compact" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e22" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="compact" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e23" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="config" target="yaml"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e24" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="engine" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e25" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="engine" target="spinner"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e26" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="engine" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e27" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="engine" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e28" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="engine" target="openai"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e29" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e30" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="detector"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e31" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="fileedit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e32" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="skills"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e33" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="spinner"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e34" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e35" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="tools" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e36" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="session" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="mention" value="&lt;b&gt;internal/mention&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;the @path notation of a question&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="400" width="200" height="60" as="geometry"/></mxCell><mxCell id="e37" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="mention"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e38" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="mention"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="store" value="&lt;b&gt;internal/store&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;per-project session files (.mm/sessions/&lt;id&gt;.json)&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="480" width="200" height="60" as="geometry"/></mxCell><mxCell id="e39" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="store"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e40" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="store"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="tui" value="&lt;b&gt;internal/tui&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;window front end: one screen on tcell (mm -tui)&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#dae8fc;strokeColor=#6c8ebf;" vertex="1" parent="1"><mxGeometry x="430" y="160" width="200" height="60" as="geometry"/></mxCell><mxCell id="command" value="&lt;b&gt;internal/command&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;the one slash-command table, per front end&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="560" width="200" height="60" as="geometry"/></mxCell><mxCell id="tcell" value="&lt;b&gt;github.com/gdamore/tcell/v2&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;terminal screen, keys, mouse&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1"><mxGeometry x="920" y="840" width="200" height="60" as="geometry"/></mxCell><mxCell id="e41" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="tui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e42" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e43" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="engine"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e44" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="compact"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e45" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="store"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e46" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="mention"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e47" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="session"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e48" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="detector"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e49" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e50" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="command"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e51" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="tcell"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e52" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="command"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e53" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="command"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e54" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="command" target="session"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="clipboard" value="&lt;b&gt;internal/clipboard&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;system clipboard: native tool, then OSC 52&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="1140" y="680" width="200" height="60" as="geometry"/></mxCell><mxCell id="e55" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="clipboard"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e56" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="clipboard"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="lsp" value="&lt;b&gt;internal/lsp&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;language-server client: definition, references, hover, symbols, diagnostics, rename&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="1160" y="600" width="220" height="60" as="geometry"/></mxCell><mxCell id="lspdev" value="&lt;b&gt;go.lsp.dev/protocol + jsonrpc2&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;LSP types and JSON-RPC 2.0 over stdio&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#ffffff;strokeColor=#999999;" vertex="1" parent="1"><mxGeometry x="1160" y="700" width="220" height="60" as="geometry"/></mxCell><mxCell id="e57" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="lsp"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e58" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="lsp"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e59" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="lsp" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e60" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="lsp" target="lspdev"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="legend" value="&lt;b&gt;mm — package dependencies&lt;/b&gt;&lt;br/&gt;Arrow: &quot;imports&quot;. Yellow: entry point. Blue: front ends. Green: shared logic. Grey: leaf packages. Dashed purple: third-party modules.&lt;br/&gt;Regenerate the edge list from: go list -f '{{.ImportPath}}: {{join .Imports &quot; &quot;}}' ./..." style="text;html=1;whiteSpace=wrap;align=left;verticalAlign=top;fillColor=#ffffff;strokeColor=#999999;" vertex="1" parent="1"><mxGeometry x="900" y="40" width="440" height="90" as="geometry"/></mxCell><mxCell id="webmd" value="&lt;b&gt;internal/webmd&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;web page → Markdown: fetch (robots.txt, charset → UTF-8), clean, convert&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#f5f5f5;strokeColor=#666666;" vertex="1" parent="1"><mxGeometry x="1400" y="460" width="220" height="60" as="geometry"/></mxCell><mxCell id="xnet" value="&lt;b&gt;golang.org/x/net/html (+ x/text)&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;HTML5 parser, charset lookup&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;dashed=1;fillColor=#e1d5e7;strokeColor=#9673a6;" vertex="1" parent="1"><mxGeometry x="1400" y="760" width="220" height="60" as="geometry"/></mxCell><mxCell id="e_main_webmd" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="webmd"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_tools_webmd" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tools" target="webmd"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_webmd_xnet" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="webmd" target="xnet"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="ask" value="&lt;b&gt;internal/ask&lt;/b&gt;&lt;br/&gt;&lt;font style=&quot;font-size:10px&quot;&gt;/ask another mm agent: child mm -acp, ACP client side&lt;/font&gt;" style="rounded=1;whiteSpace=wrap;html=1;fillColor=#d5e8d4;strokeColor=#82b366;" vertex="1" parent="1"><mxGeometry x="430" y="640" width="200" height="60" as="geometry"/></mxCell><mxCell id="e_main_ask" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="main" target="ask"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_agent_ask" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="agent" target="ask"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_acp_ask" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="acp" target="ask"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_tui_ask" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="tui" target="ask"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_ask_command" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="ask" target="command"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_ask_config" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="ask" target="config"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_ask_compact" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="ask" target="compact"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_ask_ui" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;" edge="1" parent="1" source="ask" target="ui"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_ask_acpsdk" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="ask" target="acpsdk"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_ask_genkit" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="ask" target="genkit"><mxGeometry relative="1" as="geometry"/></mxCell><mxCell id="e_ask_yaml" style="endArrow=block;html=1;rounded=0;edgeStyle=orthogonalEdgeStyle;dashed=1;" edge="1" parent="1" source="ask" target="yaml"><mxGeometry relative="1" as="geometry"/></mxCell></root></mxGraphModel></diagram></mxfile>
 \ No newline at end of file

modified docs/en/README.md +2 -0 Viewed

@@ -13,6 +13,7 @@ This documentation is organised along the four [Diátaxis](https://diataxis.fr)
13	13	- [Run mini-me on llama.cpp instead of Docker Model Runner](how-to/run-with-llama-cpp.md)
14	14	- [Run mini-me on llmman](how-to/run-with-llmman.md) — `provider: llmman`, the demo config in `demo-llmman/`
15	15	- [Run mini-me on Ollama](how-to/run-with-ollama.md) — `provider: ollama`, the demo config in `demo-ollama/`
	16	+- [Ask another mm agent from your session](how-to/ask-another-agent.md) — `mmConfigs`, `/ask-<name>`, `/ask-with-context-<name>`
16	17	- [Use mini-me from an editor (ACP)](how-to/use-from-an-editor.md)
17	18	- [Run mini-me in a Docker Sandbox](how-to/run-in-a-docker-sandbox.md) — the template and the two kits under `docker-sbx/`, window mode and editor
18	19	- [Run mini-me with the v3 sandbox kits](how-to/run-with-sandbox-kits-v3.md) — the same two kits for the kit specification v3, under `docker-sbx-v3/`
@@ -56,6 +57,7 @@ This documentation is organised along the four [Diátaxis](https://diataxis.fr)
56	57	- [The `lsp` tool](explanation/lsp-tool.md) — one tool with an `action` field, a `symbol` field instead of columns, files synced from disk, rename gated by permission
57	58	- [The `fetch_content` tool](explanation/fetch-content-tool.md) — why a built-in tool beside MCP `fetch`, a home-made converter, robots.txt, errors as numbered text, known limits
58	59	- [Thinking models](explanation/thinking-models.md) — where a model's reasoning comes from, why `mm` reads it below Genkit, and why reading it keeps the watchdog alive
	60	+- [Asking another agent](explanation/asking-another-agent.md) — why another mm agent is a child process spoken to over ACP, the history pair, fresh sessions, the depth limit
59	61	- Package diagram (draw.io source): [`docs/diagrams/packages.drawio`](../diagrams/packages.drawio)
60	62
61	63	### MCP (Model Context Protocol) Explanation

  - [Run mini-me on llama.cpp instead of Docker Model Runner](how-to/run-with-llama-cpp.md)
  - [Run mini-me on llmman](how-to/run-with-llmman.md) — `provider: llmman`, the demo config in `demo-llmman/`
  - [Run mini-me on Ollama](how-to/run-with-ollama.md) — `provider: ollama`, the demo config in `demo-ollama/`
++- [Ask another mm agent from your session](how-to/ask-another-agent.md) — `mmConfigs`, `/ask-<name>`, `/ask-with-context-<name>`
  - [Use mini-me from an editor (ACP)](how-to/use-from-an-editor.md)
  - [Run mini-me in a Docker Sandbox](how-to/run-in-a-docker-sandbox.md) — the template and the two kits under `docker-sbx/`, window mode and editor
  - [Run mini-me with the v3 sandbox kits](how-to/run-with-sandbox-kits-v3.md) — the same two kits for the kit specification v3, under `docker-sbx-v3/`
  - [The `lsp` tool](explanation/lsp-tool.md) — one tool with an `action` field, a `symbol` field instead of columns, files synced from disk, rename gated by permission
  - [The `fetch_content` tool](explanation/fetch-content-tool.md) — why a built-in tool beside MCP `fetch`, a home-made converter, robots.txt, errors as numbered text, known limits
  - [Thinking models](explanation/thinking-models.md) — where a model's reasoning comes from, why `mm` reads it below Genkit, and why reading it keeps the watchdog alive
++- [Asking another agent](explanation/asking-another-agent.md) — why another mm agent is a child process spoken to over ACP, the history pair, fresh sessions, the depth limit
  - Package diagram (draw.io source): [`docs/diagrams/packages.drawio`](../diagrams/packages.drawio)
  ### MCP (Model Context Protocol) Explanation

modified docs/en/explanation/architecture.md +1 -0 Viewed

@@ -15,6 +15,7 @@ The entry point, `main.go`, only does wiring: it loads the settings, opens the e
15	15	\| `lsp` \| The language-server client behind the `lsp` tool: starts the configured servers on first use (config `lsp.servers`), syncs files from disk, asks for definition, references, hover, symbols, diagnostics and rename, and shapes the answers as `path:line:column <line>` text. Built on `go.lsp.dev/protocol` and `go.lsp.dev/jsonrpc2`. \|
16	16	\| `skills` \| Discovery of markdown procedures and rendering of the `read_skill` catalogue. \|
17	17	\| `command` \| The one table of slash commands: name, description, which front ends offer it, whether it takes an argument. The REPL dispatches on it, the ACP façade advertises its Editor slice, the window's `/` picker lists it. \|
	18	+\| `ask` \| `/ask-<name>`: starts another mm agent (config `mmConfigs`) as a child `mm -acp`, talks to it as an ACP client, streams its answer and tool calls through the asking front end's `ui.Sink`, and returns the exchange for the history. See [Asking another agent](asking-another-agent.md). \|
18	19	\| `clipboard` \| Puts text on the system clipboard: a native tool on `PATH` first (`pbcopy`, `wl-copy`, `xclip`, `xsel`), OSC 52 as the fallback, and it reports which. Used by `/copy` in the REPL and the window. \|
19	20	\| `store` \| Per-project session files (`<dir>/<id>.json`, whole rewrite, atomic): create, save, load, list, latest. Used by both front ends when `sessions.enabled`. \|
20	21	\| `mention` \| The `@path` notation of a question: existing paths become `[attached file: …]` lines, the same shape as an editor's `resource_link`. Used by both front ends. \|

  | `lsp` | The language-server client behind the `lsp` tool: starts the configured servers on first use (config `lsp.servers`), syncs files from disk, asks for definition, references, hover, symbols, diagnostics and rename, and shapes the answers as `path:line:column  <line>` text. Built on `go.lsp.dev/protocol` and `go.lsp.dev/jsonrpc2`. |
  | `skills` | Discovery of markdown procedures and rendering of the `read_skill` catalogue. |
  | `command` | The one table of slash commands: name, description, which front ends offer it, whether it takes an argument. The REPL dispatches on it, the ACP façade advertises its Editor slice, the window's `/` picker lists it. |
++| `ask` | `/ask-<name>`: starts another mm agent (config `mmConfigs`) as a child `mm -acp`, talks to it as an ACP client, streams its answer and tool calls through the asking front end's `ui.Sink`, and returns the exchange for the history. See [Asking another agent](asking-another-agent.md). |
  | `clipboard` | Puts text on the system clipboard: a native tool on `PATH` first (`pbcopy`, `wl-copy`, `xclip`, `xsel`), OSC 52 as the fallback, and it reports which. Used by `/copy` in the REPL and the window. |
  | `store` | Per-project session files (`<dir>/<id>.json`, whole rewrite, atomic): create, save, load, list, latest. Used by both front ends when `sessions.enabled`. |
  | `mention` | The `@path` notation of a question: existing paths become `[attached file: …]` lines, the same shape as an editor's `resource_link`. Used by both front ends. |

added docs/en/explanation/asking-another-agent.md +38 -0 Viewed

new file mode 100644
@@ -0,0 +1,38 @@
	1	+# Asking another agent — explanation
	2	+
	3	+## What is this about?
	4	+
	5	+`/ask-bob <prompt>` lets the agent you are talking to consult another mm agent: another model, another system prompt, other tools and skills. Bob's answer is shown, then kept in your conversation so your own model can use it on your next question. This page explains how that is built and why.
	6	+
	7	+## Why it is designed this way
	8	+
	9	+Another agent is another process. Bob runs as a child `mm -acp -config <bob's file>`, the same binary as the agent that asks. Running him inside the same process was the obvious alternative and the wrong one: mm's configuration is one global, read by the engine, the tools, the compaction and the sessions alike, and bob's differs in all of them. A second engine in the same process would first have meant pulling that global apart everywhere. A child keeps each agent whole, with its own tools, skills and MCP servers.
	10	+
	11	+The two talk ACP. mm already speaks the Agent Client Protocol as a server, for editors. Asking bob makes the asking agent an ACP client of its child. The protocol gives, for free, what scraping a terminal never would: bob's answer as text, his tool calls as structured events, his permission requests as requests that can be answered, and a standard way to cancel. mm's ACP server only ever calls two client methods — session updates and permission requests — so the client side is small; file and terminal requests are refused as absent.
	12	+
	13	+Bob's permissions are yours to give. When bob wants to run a command, the request is forwarded to the front end you are using: in an editor its own dialog, in the window the `permissions` setting and its inline prompt, in the line mode nothing — exactly as for your own agent's tools there. Bob's tools are named `bob/<tool>`, so allowing your agent's `bash` "for this session" does not allow bob's.
	14	+
	15	+The answer is a pair of messages. It enters your history as a user message (`[asked bob (provider/model): …]`) followed by an assistant message (`[bob's answer]` and the text). One message would have been simpler, but your next question is a user message too, and several chat templates (Gemma's, Mistral's) refuse two user messages in a row. The marks say who said what, so your model does not mistake bob's answer for its own. The answer is capped like a tool's output, at `maxOutput` characters: it lands in your model's window.
	16	+
	17	+Each ask is a fresh conversation for bob. Bob keeps no memory between two asks. When he needs to know what you were doing, `/ask-with-context-bob` sends a transcript: questions, answers, one line per tool call — never the tool outputs, which would cost his window what they already cost yours, never your system prompt or reasoning. Continuity is explicit, chosen per question, rather than a second conversation drifting on its own.
	18	+
	19	+One child per agent, started once. Starting mm costs its engine set-up, its server probe and its MCP servers. The first ask starts bob; later asks reuse him, each with a new session. He stops when your agent exits.
	20	+
	21	+The child is fenced off from what is meant for its parent. Variables like `AGENT_MODEL`, `AGENT_PROVIDER` or `AGENT_BASE_URL` override a configuration; inherited by bob, they would silently turn him into a copy of you. They are removed from his environment. Bob also runs with `AGENT_SESSIONS=off`: he works in your project, and his session files would otherwise appear in your `/sessions` and `/resume`.
	22	+
	23	+Chains stop. `/ask` is a command a person types; no model can call it, so no agent loops on its own. A person could still type `/ask` into an agent that was itself asked, so each child is told its depth (`MM_ASK_DEPTH`): your agent may ask bob, bob may ask riker, riker may not ask anyone.
	24	+
	25	+## Rejected alternatives
	26	+
	27	+- `mm -p "<prompt>"` as a subprocess. One process per question and nothing to keep alive — but its output is a terminal screen (answer, reasoning, tool echoes, colours, the command recap) that would have to be parsed; it never asks permission, so bob would run commands unchecked; it cannot be cancelled except by killing it.
	28	+- A second engine in the same process. Fast, no process — at the price of turning the global configuration into a per-agent one throughout the code base.
	29	+- Letting the current model answer at once with bob's input. The user chose otherwise: bob's answer is shown and kept, and you decide what to ask next.
	30	+- Keeping bob's own conversation across asks. Two conversations evolving in parallel, one of them invisible, are harder to reason about than an explicit transcript sent when asked.
	31	+- A generic `/ask <name>` only, in editors too. An editor's `/` menu completes a command name, never its argument; one command per agent is the only way for the menu to list the agents.
	32	+
	33	+## How it relates to the rest
	34	+
	35	+- The child is an ordinary [ACP agent](../reference/acp.md); the parent reuses the protocol from the other side.
	36	+- Its tool calls and permissions go through the same event interface the three front ends implement for your own agent's tools.
	37	+- The exchange is part of the history like any other turn: it is saved with the [session](persistent-sessions.md), replayed by `/resume`, and summarised by [context compression](context-compression.md) when it gets old.
	38	+- Using it: [how to ask another agent](../how-to/ask-another-agent.md). Every command and message: [slash commands reference](../reference/slash-commands.md#asking-another-agent).

++# Asking another agent — explanation
++
++## What is this about?
++
++`/ask-bob <prompt>` lets the agent you are talking to consult another mm agent: another model, another system prompt, other tools and skills. Bob's answer is shown, then kept in your conversation so your own model can use it on your next question. This page explains how that is built and why.
++
++## Why it is designed this way
++
++**Another agent is another process.** Bob runs as a child `mm -acp -config <bob's file>`, the same binary as the agent that asks. Running him inside the same process was the obvious alternative and the wrong one: mm's configuration is one global, read by the engine, the tools, the compaction and the sessions alike, and bob's differs in all of them. A second engine in the same process would first have meant pulling that global apart everywhere. A child keeps each agent whole, with its own tools, skills and MCP servers.
++
++**The two talk ACP.** mm already speaks the Agent Client Protocol as a server, for editors. Asking bob makes the asking agent an ACP *client* of its child. The protocol gives, for free, what scraping a terminal never would: bob's answer as text, his tool calls as structured events, his permission requests as requests that can be answered, and a standard way to cancel. mm's ACP server only ever calls two client methods — session updates and permission requests — so the client side is small; file and terminal requests are refused as absent.
++
++**Bob's permissions are yours to give.** When bob wants to run a command, the request is forwarded to the front end you are using: in an editor its own dialog, in the window the `permissions` setting and its inline prompt, in the line mode nothing — exactly as for your own agent's tools there. Bob's tools are named `bob/<tool>`, so allowing your agent's `bash` "for this session" does not allow bob's.
++
++**The answer is a pair of messages.** It enters your history as a user message (`[asked bob (provider/model): …]`) followed by an assistant message (`[bob's answer]` and the text). One message would have been simpler, but your next question is a user message too, and several chat templates (Gemma's, Mistral's) refuse two user messages in a row. The marks say who said what, so your model does not mistake bob's answer for its own. The answer is capped like a tool's output, at `maxOutput` characters: it lands in your model's window.
++
++**Each ask is a fresh conversation for bob.** Bob keeps no memory between two asks. When he needs to know what you were doing, `/ask-with-context-bob` sends a transcript: questions, answers, one line per tool call — never the tool outputs, which would cost his window what they already cost yours, never your system prompt or reasoning. Continuity is explicit, chosen per question, rather than a second conversation drifting on its own.
++
++**One child per agent, started once.** Starting mm costs its engine set-up, its server probe and its MCP servers. The first ask starts bob; later asks reuse him, each with a new session. He stops when your agent exits.
++
++**The child is fenced off from what is meant for its parent.** Variables like `AGENT_MODEL`, `AGENT_PROVIDER` or `AGENT_BASE_URL` override a configuration; inherited by bob, they would silently turn him into a copy of you. They are removed from his environment. Bob also runs with `AGENT_SESSIONS=off`: he works in your project, and his session files would otherwise appear in your `/sessions` and `/resume`.
++
++**Chains stop.** `/ask` is a command a person types; no model can call it, so no agent loops on its own. A person could still type `/ask` into an agent that was itself asked, so each child is told its depth (`MM_ASK_DEPTH`): your agent may ask bob, bob may ask riker, riker may not ask anyone.
++
++## Rejected alternatives
++
++- **`mm -p "<prompt>"` as a subprocess.** One process per question and nothing to keep alive — but its output is a terminal screen (answer, reasoning, tool echoes, colours, the command recap) that would have to be parsed; it never asks permission, so bob would run commands unchecked; it cannot be cancelled except by killing it.
++- **A second engine in the same process.** Fast, no process — at the price of turning the global configuration into a per-agent one throughout the code base.
++- **Letting the current model answer at once with bob's input.** The user chose otherwise: bob's answer is shown and kept, and you decide what to ask next.
++- **Keeping bob's own conversation across asks.** Two conversations evolving in parallel, one of them invisible, are harder to reason about than an explicit transcript sent when asked.
++- **A generic `/ask <name>` only, in editors too.** An editor's `/` menu completes a command name, never its argument; one command per agent is the only way for the menu to list the agents.
++
++## How it relates to the rest
++
++- The child is an ordinary [ACP agent](../reference/acp.md); the parent reuses the protocol from the other side.
++- Its tool calls and permissions go through the same event interface the three front ends implement for your own agent's tools.
++- The exchange is part of the history like any other turn: it is saved with the [session](persistent-sessions.md), replayed by `/resume`, and summarised by [context compression](context-compression.md) when it gets old.
++- Using it: [how to ask another agent](../how-to/ask-another-agent.md). Every command and message: [slash commands reference](../reference/slash-commands.md#asking-another-agent).

added docs/en/how-to/ask-another-agent.md +61 -0 Viewed

new file mode 100644
@@ -0,0 +1,61 @@
	1	+# How to ask another mm agent from your session
	2	+
	3	+This guide shows how to put a question to another mm agent — another model, another system prompt, other tools — and keep its answer in your conversation. It assumes you already run `mm` with a working provider.
	4	+
	5	+## Steps
	6	+
	7	+1. Write (or reuse) one configuration file per agent you want to ask. Each is an ordinary mm configuration; `demo-llmman/` ships two, for llmman:
	8	+
	9	+ - `demo-llmman/agent.jan-nano.yaml` — `hf.co/Menlo/Jan-nano-gguf:Q4_K_M`, named Jan;
	10	+ - `demo-llmman/agent.lucy.yaml` — `hf.co/Menlo/Lucy-gguf:Q4_K_M`, named Lucy.
	11	+
	12	+2. Declare them in YOUR configuration, under `mmConfigs`. Paths are relative to the directory of the file you write them in:
	13	+
	14	+ ```yaml
	15	+ mmConfigs:
	16	+ jan: agent.jan-nano.yaml
	17	+ lucy: agent.lucy.yaml
	18	+ ```
	19	+
	20	+ `demo-llmman/agent.llmman.yaml` already does this.
	21	+
	22	+3. Start `mm` with that configuration:
	23	+
	24	+ ```bash
	25	+ cd demo-llmman
	26	+ ../mm agent.llmman.yaml
	27	+ ```
	28	+
	29	+ An agent whose file is missing shows up as a `[warning: mmConfigs: …]` line; the others work.
	30	+
	31	+4. Ask a question:
	32	+
	33	+ ```text
	34	+ > /ask-jan In one short sentence: what is the capital of France?
	35	+ 🗣 asking jan (llmman/hf.co/Menlo/Jan-nano-gguf:Q4_K_M)…
	36	+ The capital of France is Paris.
	37	+ 🗣 jan's answer is in the history for your next question
	38	+ ```
	39	+
	40	+ The first ask of an agent starts it, so it takes a little longer; the next ones reuse it.
	41	+
	42	+5. Use the answer: your next question goes to your own model, which now sees Jan's answer.
	43	+
	44	+ ```text
	45	+ > What did Jan answer?
	46	+ 🤖 Jan answered that the capital of France is Paris.
	47	+ ```
	48	+
	49	+## Variants
	50	+
	51	+- Let the other agent see the conversation. Use `/ask-with-context-lucy <prompt>`: Lucy receives a transcript of the conversation (questions, answers, one line per tool call — never tool outputs) before your question.
	52	+- In the terminal or the window, by name. `/ask lucy <prompt>` and `/ask-with-context lucy <prompt>` are the same commands with the agent as an argument.
	53	+- In an editor (ACP). The menu lists `/ask-jan`, `/ask-with-context-jan`, `/ask-lucy`, `/ask-with-context-lucy`; when Lucy wants to run a command, the editor's own permission dialog asks you.
	54	+- Stop a long answer. `/abort`, `Esc` or `Ctrl+C` — nothing is kept from a cancelled ask.
	55	+- Agents on other providers. Any mm configuration works: an Ollama agent next to an llmman one is just another `mmConfigs` entry.
	56	+
	57	+## See also
	58	+
	59	+- Every form, message and error: [slash commands reference](../reference/slash-commands.md#asking-another-agent)
	60	+- The `mmConfigs` key: [configuration reference](../reference/configuration.md)
	61	+- Why another agent is a separate process spoken to over ACP: [asking another agent — explanation](../explanation/asking-another-agent.md)

++# How to ask another mm agent from your session
++
++This guide shows how to put a question to another mm agent — another model, another system prompt, other tools — and keep its answer in your conversation. It assumes you already run `mm` with a working provider.
++
++## Steps
++
++1. Write (or reuse) one configuration file per agent you want to ask. Each is an ordinary mm configuration; `demo-llmman/` ships two, for llmman:
++
++   - `demo-llmman/agent.jan-nano.yaml` — `hf.co/Menlo/Jan-nano-gguf:Q4_K_M`, named Jan;
++   - `demo-llmman/agent.lucy.yaml` — `hf.co/Menlo/Lucy-gguf:Q4_K_M`, named Lucy.
++
++2. Declare them in YOUR configuration, under `mmConfigs`. Paths are relative to the directory of the file you write them in:
++
++   ```yaml
++   mmConfigs:
++     jan: agent.jan-nano.yaml
++     lucy: agent.lucy.yaml
++   ```
++
++   `demo-llmman/agent.llmman.yaml` already does this.
++
++3. Start `mm` with that configuration:
++
++   ```bash
++   cd demo-llmman
++   ../mm agent.llmman.yaml
++   ```
++
++   An agent whose file is missing shows up as a `[warning: mmConfigs: …]` line; the others work.
++
++4. Ask a question:
++
++   ```text
++   > /ask-jan In one short sentence: what is the capital of France?
++   🗣 asking jan (llmman/hf.co/Menlo/Jan-nano-gguf:Q4_K_M)…
++   The capital of France is Paris.
++   🗣 jan's answer is in the history for your next question
++   ```
++
++   The first ask of an agent starts it, so it takes a little longer; the next ones reuse it.
++
++5. Use the answer: your next question goes to your own model, which now sees Jan's answer.
++
++   ```text
++   > What did Jan answer?
++   🤖 Jan answered that the capital of France is Paris.
++   ```
++
++## Variants
++
++- **Let the other agent see the conversation.** Use `/ask-with-context-lucy <prompt>`: Lucy receives a transcript of the conversation (questions, answers, one line per tool call — never tool outputs) before your question.
++- **In the terminal or the window, by name.** `/ask lucy <prompt>` and `/ask-with-context lucy <prompt>` are the same commands with the agent as an argument.
++- **In an editor (ACP).** The menu lists `/ask-jan`, `/ask-with-context-jan`, `/ask-lucy`, `/ask-with-context-lucy`; when Lucy wants to run a command, the editor's own permission dialog asks you.
++- **Stop a long answer.** `/abort`, `Esc` or `Ctrl+C` — nothing is kept from a cancelled ask.
++- **Agents on other providers.** Any mm configuration works: an Ollama agent next to an llmman one is just another `mmConfigs` entry.
++
++## See also
++
++- Every form, message and error: [slash commands reference](../reference/slash-commands.md#asking-another-agent)
++- The `mmConfigs` key: [configuration reference](../reference/configuration.md)
++- Why another agent is a separate process spoken to over ACP: [asking another agent — explanation](../explanation/asking-another-agent.md)

modified docs/en/reference/cli.md +3 -1 Viewed

@@ -55,6 +55,8 @@ The first existing source wins.
55	55	\| `AGENT_MODEL` \| Overrides the `model` key. \|
56	56	\| `AGENT_BASE_URL` \| Overrides `baseUrl` and disables the fallback probe. \|
57	57	\| `DMR_BASE_URL` \| Same as `AGENT_BASE_URL`, for the `dmr` provider only; `AGENT_BASE_URL` wins when both are set. \|
	58	+\| `AGENT_SESSIONS` \| `off` turns session persistence off for this run, whatever `sessions.enabled` says. Set by `/ask` on the agents it starts. \|
	59	+\| `MM_ASK_DEPTH` \| How many `/ask` levels deep this agent runs (unset = 0). Set by `/ask`; at 2, `/ask` is refused. \|
58	60	\| Variable named by `apiKeyEnv` \| Holds the API key. Default name: none for `dmr`, `LLAMA_API_KEY` for `llamacpp`, `LLMMAN_API_KEY` for `llmman`, `OLLAMA_API_KEY` for `ollama`. \|
59	61
60	62	## Start-up output
@@ -63,7 +65,7 @@ Printed on stdout in terminal mode, on stderr in ACP mode. Dimmed when stdout is
63	65
64	66	\| Line \| Content \|
65	67	\|------\|---------\|
66		-\| `[warning: …]` \| Zero or more probe warnings: server unreachable, model not in the server's list, no chat template, compression that can never trigger, no tool declared at all. \|
	68	+\| `[warning: …]` \| Zero or more probe warnings: server unreachable, model not in the server's list, no chat template, compression that can never trigger, no tool declared at all, an `mmConfigs` file not found (`mmConfigs: <name> → <path> not found — /ask-<name> will fail`). \|
67	69	\| `mm <version> \\| config: … \\| provider: … \\| model: … \\| ctx: … \\| tools: … \\| skills: N` \| Version (see above), configuration path, provider name, model identifier, context window with its origin (`config`, `/props`) or `unknown`, declared tool names, number of skills found. \|
68	70	\| `Agent ready. Commands: …` \| Terminal mode only. \|
69	71	\| `Shortcut: Ctrl+C …` \| Terminal mode only. \|

  | `AGENT_MODEL` | Overrides the `model` key. |
  | `AGENT_BASE_URL` | Overrides `baseUrl` and disables the fallback probe. |
  | `DMR_BASE_URL` | Same as `AGENT_BASE_URL`, for the `dmr` provider only; `AGENT_BASE_URL` wins when both are set. |
++| `AGENT_SESSIONS` | `off` turns session persistence off for this run, whatever `sessions.enabled` says. Set by `/ask` on the agents it starts. |
++| `MM_ASK_DEPTH` | How many `/ask` levels deep this agent runs (unset = 0). Set by `/ask`; at 2, `/ask` is refused. |
  | Variable named by `apiKeyEnv` | Holds the API key. Default name: none for `dmr`, `LLAMA_API_KEY` for `llamacpp`, `LLMMAN_API_KEY` for `llmman`, `OLLAMA_API_KEY` for `ollama`. |
  ## Start-up output
  | Line | Content |
  |------|---------|
--| `[warning: …]` | Zero or more probe warnings: server unreachable, model not in the server's list, no chat template, compression that can never trigger, no tool declared at all. |
++| `[warning: …]` | Zero or more probe warnings: server unreachable, model not in the server's list, no chat template, compression that can never trigger, no tool declared at all, an `mmConfigs` file not found (`mmConfigs: <name> → <path> not found — /ask-<name> will fail`). |
  | `mm <version> \| config: … \| provider: … \| model: … \| ctx: … \| tools: … \| skills: N` | Version (see above), configuration path, provider name, model identifier, context window with its origin (`config`, `/props`) or `unknown`, declared tool names, number of skills found. |
  | `Agent ready. Commands: …` | Terminal mode only. |
  | `Shortcut: Ctrl+C …` | Terminal mode only. |

modified docs/en/reference/configuration.md +4 -0 Viewed

@@ -31,6 +31,7 @@
31	31	\| `lsp` \| map \| see below \| Language servers behind the `lsp` tool. No server listed = no tool. \|
32	32	\| `tui` \| bool \| `false` \| Start in the single-window mode without the `-tui` flag. Falls back to the line mode when stdin or stdout is not a terminal; ignored with `-acp`. \|
33	33	\| `permissions` \| string \| `auto` \| `ask`: in the window mode, `bash`, `write_file` and `edit_file` wait for the user's answer (`a`/`s`/`r`) before running. `auto`: tools run at once. The line mode and the ACP façade do not read it (ACP has its own dialog). Must be `auto` or `ask`. \|
	34	+\| `mmConfigs` \| map name → path \| none \| Other mm agents this one can ask. Each name generates `/ask-<name>` and `/ask-with-context-<name>` (see [slash commands](slash-commands.md#asking-another-agent)). A relative path is relative to the directory of the file that declares it. Names: lowercase letters, digits and `-`, starting with a letter or a digit, not starting with `with-context`. A path whose file does not exist is a start-up warning, not an error. \|
34	35
35	36	## `context` keys
36	37
@@ -117,6 +118,9 @@ Reported as `[config error: <path>: <message>]`, exit code 1.
117	118	\| `lsp.servers[N].extensions: "go" must start with a dot, like ".go"` \| an extension without its dot \|
118	119	\| `lsp.servers[N]: extension ".ts" is already served by "typescript"` \| two servers on one extension \|
119	120	\| `permissions must be "auto" or "ask"` \| any other value \|
	121	+\| `mmConfigs: "Bob" is not a valid name (…)` \| a name with uppercase letters, spaces or other characters, or starting with `-` \|
	122	+\| `mmConfigs: "with-context-x" is not a valid name (it would read as /ask-with-context-…)` \| a name starting with `with-context` \|
	123	+\| `mmConfigs.<name>: the path must not be empty` \| `<name>: ""` \|
120	124	\| `unknown provider "x" (known: dmr, llamacpp, llmman, ollama)` \| Reported by the engine, as `[engine error: …]`. \|
121	125
122	126	A start-up warning, not an error, is printed when `context.enabled` is `true`, the window is unknown and `context.maxMessages` is `0`.

  | `lsp` | map | see below | Language servers behind the `lsp` tool. No server listed = no tool. |
  | `tui` | bool | `false` | Start in the single-window mode without the `-tui` flag. Falls back to the line mode when stdin or stdout is not a terminal; ignored with `-acp`. |
  | `permissions` | string | `auto` | `ask`: in the window mode, `bash`, `write_file` and `edit_file` wait for the user's answer (`a`/`s`/`r`) before running. `auto`: tools run at once. The line mode and the ACP façade do not read it (ACP has its own dialog). Must be `auto` or `ask`. |
++| `mmConfigs` | map name → path | none | Other mm agents this one can ask. Each name generates `/ask-<name>` and `/ask-with-context-<name>` (see [slash commands](slash-commands.md#asking-another-agent)). A relative path is relative to the directory of the file that declares it. Names: lowercase letters, digits and `-`, starting with a letter or a digit, not starting with `with-context`. A path whose file does not exist is a start-up warning, not an error. |
  ## `context` keys
  | `lsp.servers[N].extensions: "go" must start with a dot, like ".go"` | an extension without its dot |
  | `lsp.servers[N]: extension ".ts" is already served by "typescript"` | two servers on one extension |
  | `permissions must be "auto" or "ask"` | any other value |
++| `mmConfigs: "Bob" is not a valid name (…)` | a name with uppercase letters, spaces or other characters, or starting with `-` |
++| `mmConfigs: "with-context-x" is not a valid name (it would read as /ask-with-context-…)` | a name starting with `with-context` |
++| `mmConfigs.<name>: the path must not be empty` | `<name>: ""` |
  | `unknown provider "x" (known: dmr, llamacpp, llmman, ollama)` | Reported by the engine, as `[engine error: …]`. |
  A start-up warning, not an error, is printed when `context.enabled` is `true`, the window is unknown and `context.maxMessages` is `0`.

modified docs/en/reference/slash-commands.md +45 -0 Viewed

@@ -17,6 +17,49 @@
17	17
18	18	With `sessions.enabled`, `/new` leaves the current session's file as it is; the next question opens a new file. An empty line at the prompt is ignored.
19	19
	20	+## Asking another agent
	21	+
	22	+Present only when `mmConfigs` declares at least one agent; `/help` and the `/` picker list them. `<name>` is a key of `mmConfigs`.
	23	+
	24	+\| Command \| Front ends \| Effect \|
	25	+\|---------\|------------\|--------\|
	26	+\| `/ask-<name> <prompt>` \| terminal, window, editor \| Sends `<prompt>` alone to `<name>`. \|
	27	+\| `/ask-with-context-<name> <prompt>` \| terminal, window, editor \| Sends a transcript of the current conversation, then `<prompt>`. \|
	28	+\| `/ask <name> <prompt>` \| terminal, window \| Same as `/ask-<name> <prompt>`. \|
	29	+\| `/ask-with-context <name> <prompt>` \| terminal, window \| Same as `/ask-with-context-<name> <prompt>`. \|
	30	+
	31	+While it runs:
	32	+
	33	+\| Item \| Behaviour \|
	34	+\|------\|-----------\|
	35	+\| First line \| `🗣 asking <name> (<provider>/<model>)…`, or `…, with the conversation…` for the context forms. \|
	36	+\| `<name>`'s answer \| Streamed where answers appear. Its reasoning is shown as reasoning when `showThinking` is on. \|
	37	+\| `<name>`'s tool calls \| Shown as tool calls titled `<name> · <tool>: <args>`. Permission requests go through this front end: none in the line mode (allowed), `permissions` in the window (`a`/`s`/`r`, "for this session" keyed on `<name>/<tool>`), the editor's dialog over ACP. \|
	38	+\| `/abort`, `Esc`, `Ctrl+C`, `session/cancel` \| Cancel the ask; `🛑 Ask aborted — nothing kept.` \|
	39	+\| `/quit` \| Cancels the ask and exits. \|
	40	+\| Any other line (line mode) \| `⚠️ <name> is still answering — /abort to stop it.`; the line is dropped. \|
	41	+
	42	+When it ends:
	43	+
	44	+\| Outcome \| History \| Message \|
	45	+\|---------\|---------\|---------\|
	46	+\| Answered \| Two messages appended: user `[asked <name> (<provider>/<model>)[, with this conversation as context]: <prompt>]`, assistant `[<name>'s answer]` + the answer (capped at `maxOutput` characters, beginning and end kept). Saved with the session. \| `🗣 <name>'s answer is in the history for your next question` \|
	47	+\| Cancelled \| Unchanged. \| `🛑 Ask aborted — nothing kept.` \|
	48	+\| Failed \| Unchanged; recorded in the session file with `where: "ask"`. \| `[error: …]` (terminal, window) or a JSON-RPC error (editor). \|
	49	+
	50	+The transcript sent by the context forms holds the user's questions, the assistant's answers, one line per tool call (`Assistant ran <tool>: <args>`, args cut at 160 characters) and the compaction summary if any; never the system prompt, tool outputs or reasoning. Capped at `maxOutput` characters, newest lines kept, `[… earlier conversation omitted …]` marking the cut.
	51	+
	52	+\| Error \| Cause \|
	53	+\|-------\|-------\|
	54	+\| `usage: /ask <agent> <prompt> — agents: …` \| `/ask` without a name. \|
	55	+\| `unknown agent "x" — agents: …` \| `/ask x …` with a name not in `mmConfigs`. \|
	56	+\| `nothing to ask: type the question after the command` \| No prompt. \|
	57	+\| `<name>: config file <path>: no such file or directory` \| The file of `mmConfigs.<name>` does not exist. \|
	58	+\| `<name> did not start: … — <last stderr line>` \| The agent's process exited or did not answer `initialize` within 60 s. \|
	59	+\| `<name> failed: …` \| The agent's turn ended with an error (its own one-line explanation). \|
	60	+\| `<name> gave no answer` \| The agent ended its turn without any text. \|
	61	+\| `this agent was itself asked through 2 level(s) of /ask; it cannot ask further (max 2)` \| `MM_ASK_DEPTH` is 2. \|
	62	+
20	63	## `@path` in a question
21	64
22	65	\| Input \| Effect \|
@@ -132,4 +175,6 @@ Two commands exist in ACP mode: `/new` and `/compact`. They are advertised to th
132	175	\| `/new` \| Same as in the terminal: history back to the system prompt, token count forgotten; the session id, `cwd` and allow-always grants are kept. \|
133	176	\| `/compact` \| Same as in the terminal, always forced: `🗜️ compressed …`, `🗜️ nothing to compact: …` or `[compact: failed, history kept: …]`. Waits for a running turn to finish first. \|
134	177
	178	+With `mmConfigs`, `/ask-<name>` and `/ask-with-context-<name>` are advertised too, each with the input hint `<prompt>`; the generic `/ask <name>` is not, since an editor's menu completes a command name but not its argument. See [Asking another agent](#asking-another-agent).
	179	+
135	180	`/quit` and `/abort` do not exist in ACP mode: the editor owns the process and ends it by closing stdin, and cancels a turn with its own stop key through `session/cancel`. Each skill is offered there as a command of its own — ACP has no picker for skills, so `/greetings Bob` in the editor's menu is the equivalent of `$greetings Bob` in the window, and it is rewritten into exactly that. See the [ACP reference](acp.md#skills-as-commands).

  With `sessions.enabled`, `/new` leaves the current session's file as it is; the next question opens a new file. An empty line at the prompt is ignored.
++## Asking another agent
++
++Present only when `mmConfigs` declares at least one agent; `/help` and the `/` picker list them. `<name>` is a key of `mmConfigs`.
++
++| Command | Front ends | Effect |
++|---------|------------|--------|
++| `/ask-<name> <prompt>` | terminal, window, editor | Sends `<prompt>` alone to `<name>`. |
++| `/ask-with-context-<name> <prompt>` | terminal, window, editor | Sends a transcript of the current conversation, then `<prompt>`. |
++| `/ask <name> <prompt>` | terminal, window | Same as `/ask-<name> <prompt>`. |
++| `/ask-with-context <name> <prompt>` | terminal, window | Same as `/ask-with-context-<name> <prompt>`. |
++
++While it runs:
++
++| Item | Behaviour |
++|------|-----------|
++| First line | `🗣 asking <name> (<provider>/<model>)…`, or `…, with the conversation…` for the context forms. |
++| `<name>`'s answer | Streamed where answers appear. Its reasoning is shown as reasoning when `showThinking` is on. |
++| `<name>`'s tool calls | Shown as tool calls titled `<name> · <tool>: <args>`. Permission requests go through this front end: none in the line mode (allowed), `permissions` in the window (`a`/`s`/`r`, "for this session" keyed on `<name>/<tool>`), the editor's dialog over ACP. |
++| `/abort`, `Esc`, `Ctrl+C`, `session/cancel` | Cancel the ask; `🛑 Ask aborted — nothing kept.` |
++| `/quit` | Cancels the ask and exits. |
++| Any other line (line mode) | `⚠️ <name> is still answering — /abort to stop it.`; the line is dropped. |
++
++When it ends:
++
++| Outcome | History | Message |
++|---------|---------|---------|
++| Answered | Two messages appended: user `[asked <name> (<provider>/<model>)[, with this conversation as context]: <prompt>]`, assistant `[<name>'s answer]` + the answer (capped at `maxOutput` characters, beginning and end kept). Saved with the session. | `🗣 <name>'s answer is in the history for your next question` |
++| Cancelled | Unchanged. | `🛑 Ask aborted — nothing kept.` |
++| Failed | Unchanged; recorded in the session file with `where: "ask"`. | `[error: …]` (terminal, window) or a JSON-RPC error (editor). |
++
++The transcript sent by the context forms holds the user's questions, the assistant's answers, one line per tool call (`Assistant ran <tool>: <args>`, args cut at 160 characters) and the compaction summary if any; never the system prompt, tool outputs or reasoning. Capped at `maxOutput` characters, newest lines kept, `[… earlier conversation omitted …]` marking the cut.
++
++| Error | Cause |
++|-------|-------|
++| `usage: /ask <agent> <prompt> — agents: …` | `/ask` without a name. |
++| `unknown agent "x" — agents: …` | `/ask x …` with a name not in `mmConfigs`. |
++| `nothing to ask: type the question after the command` | No prompt. |
++| `<name>: config file <path>: no such file or directory` | The file of `mmConfigs.<name>` does not exist. |
++| `<name> did not start: … — <last stderr line>` | The agent's process exited or did not answer `initialize` within 60 s. |
++| `<name> failed: …` | The agent's turn ended with an error (its own one-line explanation). |
++| `<name> gave no answer` | The agent ended its turn without any text. |
++| `this agent was itself asked through 2 level(s) of /ask; it cannot ask further (max 2)` | `MM_ASK_DEPTH` is 2. |
++
  ## `@path` in a question
  | Input | Effect |
  | `/new` | Same as in the terminal: history back to the system prompt, token count forgotten; the session id, `cwd` and allow-always grants are kept. |
  | `/compact` | Same as in the terminal, always forced: `🗜️ compressed …`, `🗜️ nothing to compact: …` or `[compact: failed, history kept: …]`. Waits for a running turn to finish first. |
++With `mmConfigs`, `/ask-<name>` and `/ask-with-context-<name>` are advertised too, each with the input hint `<prompt>`; the generic `/ask <name>` is not, since an editor's menu completes a command name but not its argument. See [Asking another agent](#asking-another-agent).
++
  `/quit` and `/abort` do not exist in ACP mode: the editor owns the process and ends it by closing stdin, and cancels a turn with its own stop key through `session/cancel`. Each **skill** is offered there as a command of its own — ACP has no picker for skills, so `/greetings Bob` in the editor's menu is the equivalent of `$greetings Bob` in the window, and it is rewritten into exactly that. See the [ACP reference](acp.md#skills-as-commands).

modified docs/fr/README.md +2 -0 Viewed

@@ -13,6 +13,7 @@ Cette documentation est organisée selon les quatre quadrants [Diátaxis](https:
13	13	- [Faire tourner mini-me sur llama.cpp plutôt que Docker Model Runner](how-to/run-with-llama-cpp.md)
14	14	- [Faire tourner mini-me sur llmman](how-to/run-with-llmman.md) — `provider: llmman`, la config de démo dans `demo-llmman/`
15	15	- [Faire tourner mini-me sur Ollama](how-to/run-with-ollama.md) — `provider: ollama`, la config de démo dans `demo-ollama/`
	16	+- [Interroger un autre agent mm depuis votre session](how-to/ask-another-agent.md) — `mmConfigs`, `/ask-<nom>`, `/ask-with-context-<nom>`
16	17	- [Utiliser mini-me depuis un éditeur (ACP)](how-to/use-from-an-editor.md)
17	18	- [Faire tourner mini-me dans un Docker Sandbox](how-to/run-in-a-docker-sandbox.md) — le template et les deux kits sous `docker-sbx/`, mode fenêtre et éditeur
18	19	- [Faire tourner mini-me avec les kits sandbox v3](how-to/run-with-sandbox-kits-v3.md) — les deux mêmes kits pour la spécification de kit v3, sous `docker-sbx-v3/`
@@ -56,6 +57,7 @@ Cette documentation est organisée selon les quatre quadrants [Diátaxis](https:
56	57	- [L'outil `lsp`](explanation/lsp-tool.md) — un seul outil avec un champ `action`, un champ `symbol` plutôt que des colonnes, des fichiers synchronisés depuis le disque, un renommage soumis à permission
57	58	- [L'outil `fetch_content`](explanation/fetch-content-tool.md) — pourquoi un outil intégré à côté du `fetch` MCP, un convertisseur maison, le robots.txt, des erreurs numérotées en texte, les limites connues
58	59	- [Modèles à raisonnement](explanation/thinking-models.md) — d'où vient le raisonnement d'un modèle, pourquoi `mm` le lit sous Genkit, et pourquoi le lire garde le chien de garde en vie
	60	+- [Interroger un autre agent](explanation/asking-another-agent.md) — pourquoi un autre agent mm est un processus enfant interrogé en ACP, la paire de messages, les sessions neuves, la limite de profondeur
59	61	- Diagramme des paquets (source draw.io) : [`docs/diagrams/packages.drawio`](../diagrams/packages.drawio)
60	62
61	63	### Explication MCP (Model Context Protocol)

  - [Faire tourner mini-me sur llama.cpp plutôt que Docker Model Runner](how-to/run-with-llama-cpp.md)
  - [Faire tourner mini-me sur llmman](how-to/run-with-llmman.md) — `provider: llmman`, la config de démo dans `demo-llmman/`
  - [Faire tourner mini-me sur Ollama](how-to/run-with-ollama.md) — `provider: ollama`, la config de démo dans `demo-ollama/`
++- [Interroger un autre agent mm depuis votre session](how-to/ask-another-agent.md) — `mmConfigs`, `/ask-<nom>`, `/ask-with-context-<nom>`
  - [Utiliser mini-me depuis un éditeur (ACP)](how-to/use-from-an-editor.md)
  - [Faire tourner mini-me dans un Docker Sandbox](how-to/run-in-a-docker-sandbox.md) — le template et les deux kits sous `docker-sbx/`, mode fenêtre et éditeur
  - [Faire tourner mini-me avec les kits sandbox v3](how-to/run-with-sandbox-kits-v3.md) — les deux mêmes kits pour la spécification de kit v3, sous `docker-sbx-v3/`
  - [L'outil `lsp`](explanation/lsp-tool.md) — un seul outil avec un champ `action`, un champ `symbol` plutôt que des colonnes, des fichiers synchronisés depuis le disque, un renommage soumis à permission
  - [L'outil `fetch_content`](explanation/fetch-content-tool.md) — pourquoi un outil intégré à côté du `fetch` MCP, un convertisseur maison, le robots.txt, des erreurs numérotées en texte, les limites connues
  - [Modèles à raisonnement](explanation/thinking-models.md) — d'où vient le raisonnement d'un modèle, pourquoi `mm` le lit sous Genkit, et pourquoi le lire garde le chien de garde en vie
++- [Interroger un autre agent](explanation/asking-another-agent.md) — pourquoi un autre agent mm est un processus enfant interrogé en ACP, la paire de messages, les sessions neuves, la limite de profondeur
  - Diagramme des paquets (source draw.io) : [`docs/diagrams/packages.drawio`](../diagrams/packages.drawio)
  ### Explication MCP (Model Context Protocol)

modified docs/fr/explanation/architecture.md +1 -0 Viewed

@@ -15,6 +15,7 @@ Le point d'entrée, `main.go`, ne fait que du câblage : il charge les réglages
15	15	\| `lsp` \| Le client de serveur de langage derrière l'outil `lsp` : démarre les serveurs configurés au premier usage (clé `lsp.servers`), synchronise les fichiers depuis le disque, demande définition, références, hover, symboles, diagnostics et renommage, et met les réponses en forme `chemin:ligne:colonne <ligne>`. Construit sur `go.lsp.dev/protocol` et `go.lsp.dev/jsonrpc2`. \|
16	16	\| `skills` \| La découverte des procédures markdown et le rendu du catalogue de `read_skill`. \|
17	17	\| `command` \| L'unique table des commandes slash : nom, description, façades qui l'offrent, argument éventuel. Le REPL s'y branche, la façade ACP en annonce la tranche Éditeur, le sélecteur `/` de la fenêtre la liste. \|
	18	+\| `ask` \| `/ask-<nom>` : démarre un autre agent mm (clé `mmConfigs`) comme enfant `mm -acp`, lui parle en client ACP, affiche sa réponse et ses appels d'outils via le `ui.Sink` de la façade qui pose la question, et renvoie l'échange pour l'historique. Voir [Interroger un autre agent](asking-another-agent.md). \|
18	19	\| `clipboard` \| Met du texte dans le presse-papiers système : un outil natif du `PATH` d'abord (`pbcopy`, `wl-copy`, `xclip`, `xsel`), OSC 52 en repli, et il dit lequel a servi. Utilisé par `/copy` dans le REPL et la fenêtre. \|
19	20	\| `store` \| Fichiers de session par projet (`<dir>/<id>.json`, réécriture complète, atomique) : création, sauvegarde, chargement, liste, plus récente. Utilisé par les deux façades quand `sessions.enabled`. \|
20	21	\| `mention` \| La notation `@chemin` d'une question : les chemins existants deviennent des lignes `[attached file: …]`, la même forme que le `resource_link` d'un éditeur. Utilisé par les deux façades. \|

  | `lsp` | Le client de serveur de langage derrière l'outil `lsp` : démarre les serveurs configurés au premier usage (clé `lsp.servers`), synchronise les fichiers depuis le disque, demande définition, références, hover, symboles, diagnostics et renommage, et met les réponses en forme `chemin:ligne:colonne  <ligne>`. Construit sur `go.lsp.dev/protocol` et `go.lsp.dev/jsonrpc2`. |
  | `skills` | La découverte des procédures markdown et le rendu du catalogue de `read_skill`. |
  | `command` | L'unique table des commandes slash : nom, description, façades qui l'offrent, argument éventuel. Le REPL s'y branche, la façade ACP en annonce la tranche Éditeur, le sélecteur `/` de la fenêtre la liste. |
++| `ask` | `/ask-<nom>` : démarre un autre agent mm (clé `mmConfigs`) comme enfant `mm -acp`, lui parle en client ACP, affiche sa réponse et ses appels d'outils via le `ui.Sink` de la façade qui pose la question, et renvoie l'échange pour l'historique. Voir [Interroger un autre agent](asking-another-agent.md). |
  | `clipboard` | Met du texte dans le presse-papiers système : un outil natif du `PATH` d'abord (`pbcopy`, `wl-copy`, `xclip`, `xsel`), OSC 52 en repli, et il dit lequel a servi. Utilisé par `/copy` dans le REPL et la fenêtre. |
  | `store` | Fichiers de session par projet (`<dir>/<id>.json`, réécriture complète, atomique) : création, sauvegarde, chargement, liste, plus récente. Utilisé par les deux façades quand `sessions.enabled`. |
  | `mention` | La notation `@chemin` d'une question : les chemins existants deviennent des lignes `[attached file: …]`, la même forme que le `resource_link` d'un éditeur. Utilisé par les deux façades. |

added docs/fr/explanation/asking-another-agent.md +38 -0 Viewed

new file mode 100644
@@ -0,0 +1,38 @@
	1	+# Interroger un autre agent — explication
	2	+
	3	+## De quoi s'agit-il ?
	4	+
	5	+`/ask-bob <prompt>` permet à l'agent avec qui vous parlez de consulter un autre agent mm : un autre modèle, un autre prompt système, d'autres outils et skills. La réponse de bob est affichée, puis gardée dans votre conversation pour que votre propre modèle s'en serve à votre question suivante. Cette page explique comment c'est construit et pourquoi.
	6	+
	7	+## Pourquoi c'est conçu ainsi
	8	+
	9	+Un autre agent est un autre processus. Bob tourne comme un enfant `mm -acp -config <fichier de bob>`, le même binaire que l'agent qui l'interroge. Le faire tourner dans le même processus était l'alternative évidente, et la mauvaise : la configuration de mm est une variable globale, lue par le moteur, les outils, la compaction et les sessions, et celle de bob diffère sur tous ces points. Un second moteur dans le même processus aurait d'abord demandé de démanteler cette variable globale partout. Un enfant garde chaque agent entier, avec ses propres outils, skills et serveurs MCP.
	10	+
	11	+Les deux parlent ACP. mm parle déjà l'Agent Client Protocol comme serveur, pour les éditeurs. Interroger bob fait de l'agent qui pose la question un client ACP de son enfant. Le protocole apporte gratuitement ce que lire un terminal n'aurait jamais donné : la réponse de bob sous forme de texte, ses appels d'outils comme des événements structurés, ses demandes de permission comme des requêtes auxquelles on peut répondre, et une façon standard d'annuler. Le serveur ACP de mm n'appelle que deux méthodes du client — les mises à jour de session et les demandes de permission — donc le côté client est petit ; les requêtes de fichiers et de terminal sont refusées comme absentes.
	12	+
	13	+Les permissions de bob, c'est vous qui les donnez. Quand bob veut lancer une commande, la demande est transmise à la façade que vous utilisez : dans un éditeur, sa propre boîte de dialogue ; dans la fenêtre, le réglage `permissions` et sa question dans la zone de saisie ; en mode ligne, rien — exactement comme pour les outils de votre propre agent à cet endroit. Les outils de bob s'appellent `bob/<outil>`, donc autoriser le `bash` de votre agent « pour cette session » n'autorise pas celui de bob.
	14	+
	15	+La réponse est une paire de messages. Elle entre dans votre historique comme un message utilisateur (`[asked bob (fournisseur/modèle): …]`) suivi d'un message de l'assistant (`[bob's answer]` et le texte). Un seul message aurait été plus simple, mais votre question suivante est aussi un message utilisateur, et plusieurs templates de chat (ceux de Gemma, de Mistral) refusent deux messages utilisateur à la suite. Les marques disent qui a dit quoi, pour que votre modèle ne prenne pas la réponse de bob pour la sienne. La réponse est plafonnée comme la sortie d'un outil, à `maxOutput` caractères : elle atterrit dans la fenêtre de votre modèle.
	16	+
	17	+Chaque demande est une conversation neuve pour bob. Bob ne garde aucune mémoire entre deux demandes. Quand il doit savoir ce que vous faisiez, `/ask-with-context-bob` envoie une transcription : questions, réponses, une ligne par appel d'outil — jamais les sorties d'outils, qui coûteraient à sa fenêtre ce qu'elles ont déjà coûté à la vôtre, jamais votre prompt système ni le raisonnement. La continuité est explicite, choisie à chaque question, plutôt qu'une seconde conversation qui dérive de son côté.
	18	+
	19	+Un enfant par agent, démarré une fois. Démarrer mm coûte la mise en place de son moteur, la sonde du serveur et ses serveurs MCP. La première demande démarre bob ; les suivantes le réutilisent, chacune avec une nouvelle session. Il s'arrête quand votre agent se termine.
	20	+
	21	+L'enfant est protégé de ce qui est destiné à son parent. Des variables comme `AGENT_MODEL`, `AGENT_PROVIDER` ou `AGENT_BASE_URL` surchargent une configuration ; héritées par bob, elles en feraient silencieusement une copie de vous. Elles sont retirées de son environnement. Bob tourne aussi avec `AGENT_SESSIONS=off` : il travaille dans votre projet, et ses fichiers de session apparaîtraient sinon dans vos `/sessions` et `/resume`.
	22	+
	23	+Les chaînes s'arrêtent. `/ask` est une commande tapée par une personne ; aucun modèle ne peut l'appeler, donc aucun agent ne boucle tout seul. Une personne pourrait quand même taper `/ask` dans un agent lui-même interrogé, alors chaque enfant connaît sa profondeur (`MM_ASK_DEPTH`) : votre agent peut interroger bob, bob peut interroger riker, riker ne peut interroger personne.
	24	+
	25	+## Alternatives écartées
	26	+
	27	+- `mm -p "<prompt>"` en sous-processus. Un processus par question et rien à garder en vie — mais sa sortie est un écran de terminal (réponse, raisonnement, échos d'outils, couleurs, récapitulatif des commandes) qu'il faudrait analyser ; il ne demande jamais de permission, donc bob lancerait des commandes sans contrôle ; on ne peut l'annuler qu'en le tuant.
	28	+- Un second moteur dans le même processus. Rapide, sans processus — au prix de transformer la configuration globale en configuration par agent dans tout le code.
	29	+- Faire répondre le modèle courant tout de suite avec l'apport de bob. L'utilisateur a choisi autrement : la réponse de bob est affichée et gardée, et vous décidez de la question suivante.
	30	+- Garder la conversation de bob d'une demande à l'autre. Deux conversations qui évoluent en parallèle, dont une invisible, sont plus difficiles à suivre qu'une transcription explicite envoyée quand on la demande.
	31	+- Seulement un `/ask <nom>` générique, y compris dans les éditeurs. Le menu `/` d'un éditeur complète un nom de commande, jamais son argument ; une commande par agent est la seule façon pour que le menu liste les agents.
	32	+
	33	+## Lien avec le reste
	34	+
	35	+- L'enfant est un [agent ACP](../reference/acp.md) ordinaire ; le parent réutilise le protocole de l'autre côté.
	36	+- Ses appels d'outils et ses permissions passent par la même interface d'événements que les trois façades implémentent pour les outils de votre propre agent.
	37	+- L'échange fait partie de l'historique comme n'importe quel tour : il est enregistré avec la [session](persistent-sessions.md), rejoué par `/resume`, et résumé par la [compression du contexte](context-compression.md) quand il vieillit.
	38	+- L'utiliser : [comment interroger un autre agent](../how-to/ask-another-agent.md). Toutes les commandes et messages : [référence des commandes slash](../reference/slash-commands.md#interroger-un-autre-agent).

++# Interroger un autre agent — explication
++
++## De quoi s'agit-il ?
++
++`/ask-bob <prompt>` permet à l'agent avec qui vous parlez de consulter un autre agent mm : un autre modèle, un autre prompt système, d'autres outils et skills. La réponse de bob est affichée, puis gardée dans votre conversation pour que votre propre modèle s'en serve à votre question suivante. Cette page explique comment c'est construit et pourquoi.
++
++## Pourquoi c'est conçu ainsi
++
++**Un autre agent est un autre processus.** Bob tourne comme un enfant `mm -acp -config <fichier de bob>`, le même binaire que l'agent qui l'interroge. Le faire tourner dans le même processus était l'alternative évidente, et la mauvaise : la configuration de mm est une variable globale, lue par le moteur, les outils, la compaction et les sessions, et celle de bob diffère sur tous ces points. Un second moteur dans le même processus aurait d'abord demandé de démanteler cette variable globale partout. Un enfant garde chaque agent entier, avec ses propres outils, skills et serveurs MCP.
++
++**Les deux parlent ACP.** mm parle déjà l'Agent Client Protocol comme serveur, pour les éditeurs. Interroger bob fait de l'agent qui pose la question un *client* ACP de son enfant. Le protocole apporte gratuitement ce que lire un terminal n'aurait jamais donné : la réponse de bob sous forme de texte, ses appels d'outils comme des événements structurés, ses demandes de permission comme des requêtes auxquelles on peut répondre, et une façon standard d'annuler. Le serveur ACP de mm n'appelle que deux méthodes du client — les mises à jour de session et les demandes de permission — donc le côté client est petit ; les requêtes de fichiers et de terminal sont refusées comme absentes.
++
++**Les permissions de bob, c'est vous qui les donnez.** Quand bob veut lancer une commande, la demande est transmise à la façade que vous utilisez : dans un éditeur, sa propre boîte de dialogue ; dans la fenêtre, le réglage `permissions` et sa question dans la zone de saisie ; en mode ligne, rien — exactement comme pour les outils de votre propre agent à cet endroit. Les outils de bob s'appellent `bob/<outil>`, donc autoriser le `bash` de votre agent « pour cette session » n'autorise pas celui de bob.
++
++**La réponse est une paire de messages.** Elle entre dans votre historique comme un message utilisateur (`[asked bob (fournisseur/modèle): …]`) suivi d'un message de l'assistant (`[bob's answer]` et le texte). Un seul message aurait été plus simple, mais votre question suivante est aussi un message utilisateur, et plusieurs templates de chat (ceux de Gemma, de Mistral) refusent deux messages utilisateur à la suite. Les marques disent qui a dit quoi, pour que votre modèle ne prenne pas la réponse de bob pour la sienne. La réponse est plafonnée comme la sortie d'un outil, à `maxOutput` caractères : elle atterrit dans la fenêtre de votre modèle.
++
++**Chaque demande est une conversation neuve pour bob.** Bob ne garde aucune mémoire entre deux demandes. Quand il doit savoir ce que vous faisiez, `/ask-with-context-bob` envoie une transcription : questions, réponses, une ligne par appel d'outil — jamais les sorties d'outils, qui coûteraient à sa fenêtre ce qu'elles ont déjà coûté à la vôtre, jamais votre prompt système ni le raisonnement. La continuité est explicite, choisie à chaque question, plutôt qu'une seconde conversation qui dérive de son côté.
++
++**Un enfant par agent, démarré une fois.** Démarrer mm coûte la mise en place de son moteur, la sonde du serveur et ses serveurs MCP. La première demande démarre bob ; les suivantes le réutilisent, chacune avec une nouvelle session. Il s'arrête quand votre agent se termine.
++
++**L'enfant est protégé de ce qui est destiné à son parent.** Des variables comme `AGENT_MODEL`, `AGENT_PROVIDER` ou `AGENT_BASE_URL` surchargent une configuration ; héritées par bob, elles en feraient silencieusement une copie de vous. Elles sont retirées de son environnement. Bob tourne aussi avec `AGENT_SESSIONS=off` : il travaille dans votre projet, et ses fichiers de session apparaîtraient sinon dans vos `/sessions` et `/resume`.
++
++**Les chaînes s'arrêtent.** `/ask` est une commande tapée par une personne ; aucun modèle ne peut l'appeler, donc aucun agent ne boucle tout seul. Une personne pourrait quand même taper `/ask` dans un agent lui-même interrogé, alors chaque enfant connaît sa profondeur (`MM_ASK_DEPTH`) : votre agent peut interroger bob, bob peut interroger riker, riker ne peut interroger personne.
++
++## Alternatives écartées
++
++- **`mm -p "<prompt>"` en sous-processus.** Un processus par question et rien à garder en vie — mais sa sortie est un écran de terminal (réponse, raisonnement, échos d'outils, couleurs, récapitulatif des commandes) qu'il faudrait analyser ; il ne demande jamais de permission, donc bob lancerait des commandes sans contrôle ; on ne peut l'annuler qu'en le tuant.
++- **Un second moteur dans le même processus.** Rapide, sans processus — au prix de transformer la configuration globale en configuration par agent dans tout le code.
++- **Faire répondre le modèle courant tout de suite avec l'apport de bob.** L'utilisateur a choisi autrement : la réponse de bob est affichée et gardée, et vous décidez de la question suivante.
++- **Garder la conversation de bob d'une demande à l'autre.** Deux conversations qui évoluent en parallèle, dont une invisible, sont plus difficiles à suivre qu'une transcription explicite envoyée quand on la demande.
++- **Seulement un `/ask <nom>` générique, y compris dans les éditeurs.** Le menu `/` d'un éditeur complète un nom de commande, jamais son argument ; une commande par agent est la seule façon pour que le menu liste les agents.
++
++## Lien avec le reste
++
++- L'enfant est un [agent ACP](../reference/acp.md) ordinaire ; le parent réutilise le protocole de l'autre côté.
++- Ses appels d'outils et ses permissions passent par la même interface d'événements que les trois façades implémentent pour les outils de votre propre agent.
++- L'échange fait partie de l'historique comme n'importe quel tour : il est enregistré avec la [session](persistent-sessions.md), rejoué par `/resume`, et résumé par la [compression du contexte](context-compression.md) quand il vieillit.
++- L'utiliser : [comment interroger un autre agent](../how-to/ask-another-agent.md). Toutes les commandes et messages : [référence des commandes slash](../reference/slash-commands.md#interroger-un-autre-agent).

added docs/fr/how-to/ask-another-agent.md +61 -0 Viewed

new file mode 100644
@@ -0,0 +1,61 @@
	1	+# Comment interroger un autre agent mm depuis votre session
	2	+
	3	+Ce guide montre comment poser une question à un autre agent mm — un autre modèle, un autre prompt système, d'autres outils — et garder sa réponse dans votre conversation. Il suppose que vous faites déjà tourner `mm` avec un fournisseur qui marche.
	4	+
	5	+## Étapes
	6	+
	7	+1. Écrivez (ou réutilisez) un fichier de configuration par agent à interroger. Chacun est une configuration mm ordinaire ; `demo-llmman/` en fournit deux, pour llmman :
	8	+
	9	+ - `demo-llmman/agent.jan-nano.yaml` — `hf.co/Menlo/Jan-nano-gguf:Q4_K_M`, nommé Jan ;
	10	+ - `demo-llmman/agent.lucy.yaml` — `hf.co/Menlo/Lucy-gguf:Q4_K_M`, nommée Lucy.
	11	+
	12	+2. Déclarez-les dans VOTRE configuration, sous `mmConfigs`. Les chemins sont relatifs au dossier du fichier où vous les écrivez :
	13	+
	14	+ ```yaml
	15	+ mmConfigs:
	16	+ jan: agent.jan-nano.yaml
	17	+ lucy: agent.lucy.yaml
	18	+ ```
	19	+
	20	+ `demo-llmman/agent.llmman.yaml` le fait déjà.
	21	+
	22	+3. Lancez `mm` avec cette configuration :
	23	+
	24	+ ```bash
	25	+ cd demo-llmman
	26	+ ../mm agent.llmman.yaml
	27	+ ```
	28	+
	29	+ Un agent dont le fichier manque apparaît sur une ligne `[warning: mmConfigs: …]` ; les autres fonctionnent.
	30	+
	31	+4. Posez une question :
	32	+
	33	+ ```text
	34	+ > /ask-jan In one short sentence: what is the capital of France?
	35	+ 🗣 asking jan (llmman/hf.co/Menlo/Jan-nano-gguf:Q4_K_M)…
	36	+ The capital of France is Paris.
	37	+ 🗣 jan's answer is in the history for your next question
	38	+ ```
	39	+
	40	+ La première demande à un agent le démarre, elle prend donc un peu plus de temps ; les suivantes le réutilisent.
	41	+
	42	+5. Servez-vous de la réponse : votre question suivante va à votre propre modèle, qui voit maintenant la réponse de Jan.
	43	+
	44	+ ```text
	45	+ > What did Jan answer?
	46	+ 🤖 Jan answered that the capital of France is Paris.
	47	+ ```
	48	+
	49	+## Variantes
	50	+
	51	+- Laisser l'autre agent voir la conversation. Utilisez `/ask-with-context-lucy <prompt>` : Lucy reçoit une transcription de la conversation (questions, réponses, une ligne par appel d'outil — jamais les sorties d'outils) avant votre question.
	52	+- Dans le terminal ou la fenêtre, par nom. `/ask lucy <prompt>` et `/ask-with-context lucy <prompt>` sont les mêmes commandes, avec l'agent en argument.
	53	+- Dans un éditeur (ACP). Le menu liste `/ask-jan`, `/ask-with-context-jan`, `/ask-lucy`, `/ask-with-context-lucy` ; quand Lucy veut lancer une commande, c'est la boîte de dialogue de permission de l'éditeur qui vous le demande.
	54	+- Arrêter une longue réponse. `/abort`, `Échap` ou `Ctrl+C` — rien n'est gardé d'une demande annulée.
	55	+- Des agents sur d'autres fournisseurs. N'importe quelle configuration mm convient : un agent Ollama à côté d'un agent llmman n'est qu'une entrée de plus dans `mmConfigs`.
	56	+
	57	+## Voir aussi
	58	+
	59	+- Toutes les formes, messages et erreurs : [référence des commandes slash](../reference/slash-commands.md#interroger-un-autre-agent)
	60	+- La clé `mmConfigs` : [référence de configuration](../reference/configuration.md)
	61	+- Pourquoi un autre agent est un processus à part, interrogé en ACP : [interroger un autre agent — explication](../explanation/asking-another-agent.md)

++# Comment interroger un autre agent mm depuis votre session
++
++Ce guide montre comment poser une question à un autre agent mm — un autre modèle, un autre prompt système, d'autres outils — et garder sa réponse dans votre conversation. Il suppose que vous faites déjà tourner `mm` avec un fournisseur qui marche.
++
++## Étapes
++
++1. Écrivez (ou réutilisez) un fichier de configuration par agent à interroger. Chacun est une configuration mm ordinaire ; `demo-llmman/` en fournit deux, pour llmman :
++
++   - `demo-llmman/agent.jan-nano.yaml` — `hf.co/Menlo/Jan-nano-gguf:Q4_K_M`, nommé Jan ;
++   - `demo-llmman/agent.lucy.yaml` — `hf.co/Menlo/Lucy-gguf:Q4_K_M`, nommée Lucy.
++
++2. Déclarez-les dans VOTRE configuration, sous `mmConfigs`. Les chemins sont relatifs au dossier du fichier où vous les écrivez :
++
++   ```yaml
++   mmConfigs:
++     jan: agent.jan-nano.yaml
++     lucy: agent.lucy.yaml
++   ```
++
++   `demo-llmman/agent.llmman.yaml` le fait déjà.
++
++3. Lancez `mm` avec cette configuration :
++
++   ```bash
++   cd demo-llmman
++   ../mm agent.llmman.yaml
++   ```
++
++   Un agent dont le fichier manque apparaît sur une ligne `[warning: mmConfigs: …]` ; les autres fonctionnent.
++
++4. Posez une question :
++
++   ```text
++   > /ask-jan In one short sentence: what is the capital of France?
++   🗣 asking jan (llmman/hf.co/Menlo/Jan-nano-gguf:Q4_K_M)…
++   The capital of France is Paris.
++   🗣 jan's answer is in the history for your next question
++   ```
++
++   La première demande à un agent le démarre, elle prend donc un peu plus de temps ; les suivantes le réutilisent.
++
++5. Servez-vous de la réponse : votre question suivante va à votre propre modèle, qui voit maintenant la réponse de Jan.
++
++   ```text
++   > What did Jan answer?
++   🤖 Jan answered that the capital of France is Paris.
++   ```
++
++## Variantes
++
++- **Laisser l'autre agent voir la conversation.** Utilisez `/ask-with-context-lucy <prompt>` : Lucy reçoit une transcription de la conversation (questions, réponses, une ligne par appel d'outil — jamais les sorties d'outils) avant votre question.
++- **Dans le terminal ou la fenêtre, par nom.** `/ask lucy <prompt>` et `/ask-with-context lucy <prompt>` sont les mêmes commandes, avec l'agent en argument.
++- **Dans un éditeur (ACP).** Le menu liste `/ask-jan`, `/ask-with-context-jan`, `/ask-lucy`, `/ask-with-context-lucy` ; quand Lucy veut lancer une commande, c'est la boîte de dialogue de permission de l'éditeur qui vous le demande.
++- **Arrêter une longue réponse.** `/abort`, `Échap` ou `Ctrl+C` — rien n'est gardé d'une demande annulée.
++- **Des agents sur d'autres fournisseurs.** N'importe quelle configuration mm convient : un agent Ollama à côté d'un agent llmman n'est qu'une entrée de plus dans `mmConfigs`.
++
++## Voir aussi
++
++- Toutes les formes, messages et erreurs : [référence des commandes slash](../reference/slash-commands.md#interroger-un-autre-agent)
++- La clé `mmConfigs` : [référence de configuration](../reference/configuration.md)
++- Pourquoi un autre agent est un processus à part, interrogé en ACP : [interroger un autre agent — explication](../explanation/asking-another-agent.md)

modified docs/fr/reference/cli.md +3 -1 Viewed

@@ -55,6 +55,8 @@ La première source existante gagne.
55	55	\| `AGENT_MODEL` \| Surcharge la clé `model`. \|
56	56	\| `AGENT_BASE_URL` \| Surcharge `baseUrl` et désactive la sonde de repli. \|
57	57	\| `DMR_BASE_URL` \| Comme `AGENT_BASE_URL`, pour le fournisseur `dmr` seulement ; `AGENT_BASE_URL` gagne quand les deux sont définies. \|
	58	+\| `AGENT_SESSIONS` \| `off` désactive l'enregistrement des sessions pour cette exécution, quoi que dise `sessions.enabled`. Posée par `/ask` sur les agents qu'il lance. \|
	59	+\| `MM_ASK_DEPTH` \| Nombre de niveaux de `/ask` au-dessus de cet agent (absente = 0). Posée par `/ask` ; à 2, `/ask` est refusée. \|
58	60	\| Variable nommée par `apiKeyEnv` \| Contient la clé d'API. Nom par défaut : aucun pour `dmr`, `LLAMA_API_KEY` pour `llamacpp`, `LLMMAN_API_KEY` pour `llmman`, `OLLAMA_API_KEY` pour `ollama`. \|
59	61
60	62	## Sortie de démarrage
@@ -63,7 +65,7 @@ Affichée sur stdout en mode terminal, sur stderr en mode ACP. Grisée quand std
63	65
64	66	\| Ligne \| Contenu \|
65	67	\|-------\|---------\|
66		-\| `[warning: …]` \| Zéro ou plusieurs avertissements de la sonde : serveur injoignable, modèle absent de la liste du serveur, pas de template de chat, compression qui ne peut jamais se déclencher, aucun outil déclaré. \|
	68	+\| `[warning: …]` \| Zéro ou plusieurs avertissements de la sonde : serveur injoignable, modèle absent de la liste du serveur, pas de template de chat, compression qui ne peut jamais se déclencher, aucun outil déclaré, un fichier de `mmConfigs` introuvable (`mmConfigs: <name> → <path> not found — /ask-<name> will fail`). \|
67	69	\| `mm <version> \\| config: … \\| provider: … \\| model: … \\| ctx: … \\| tools: … \\| skills: N` \| Version (voir ci-dessus), chemin de la configuration, nom du fournisseur, identifiant du modèle, fenêtre de contexte avec son origine (`config`, `/props`) ou `unknown`, noms des outils déclarés, nombre de skills trouvés. \|
68	70	\| `Agent ready. Commands: …` \| Mode terminal seulement. \|
69	71	\| `Shortcut: Ctrl+C …` \| Mode terminal seulement. \|

  | `AGENT_MODEL` | Surcharge la clé `model`. |
  | `AGENT_BASE_URL` | Surcharge `baseUrl` et désactive la sonde de repli. |
  | `DMR_BASE_URL` | Comme `AGENT_BASE_URL`, pour le fournisseur `dmr` seulement ; `AGENT_BASE_URL` gagne quand les deux sont définies. |
++| `AGENT_SESSIONS` | `off` désactive l'enregistrement des sessions pour cette exécution, quoi que dise `sessions.enabled`. Posée par `/ask` sur les agents qu'il lance. |
++| `MM_ASK_DEPTH` | Nombre de niveaux de `/ask` au-dessus de cet agent (absente = 0). Posée par `/ask` ; à 2, `/ask` est refusée. |
  | Variable nommée par `apiKeyEnv` | Contient la clé d'API. Nom par défaut : aucun pour `dmr`, `LLAMA_API_KEY` pour `llamacpp`, `LLMMAN_API_KEY` pour `llmman`, `OLLAMA_API_KEY` pour `ollama`. |
  ## Sortie de démarrage
  | Ligne | Contenu |
  |-------|---------|
--| `[warning: …]` | Zéro ou plusieurs avertissements de la sonde : serveur injoignable, modèle absent de la liste du serveur, pas de template de chat, compression qui ne peut jamais se déclencher, aucun outil déclaré. |
++| `[warning: …]` | Zéro ou plusieurs avertissements de la sonde : serveur injoignable, modèle absent de la liste du serveur, pas de template de chat, compression qui ne peut jamais se déclencher, aucun outil déclaré, un fichier de `mmConfigs` introuvable (`mmConfigs: <name> → <path> not found — /ask-<name> will fail`). |
  | `mm <version> \| config: … \| provider: … \| model: … \| ctx: … \| tools: … \| skills: N` | Version (voir ci-dessus), chemin de la configuration, nom du fournisseur, identifiant du modèle, fenêtre de contexte avec son origine (`config`, `/props`) ou `unknown`, noms des outils déclarés, nombre de skills trouvés. |
  | `Agent ready. Commands: …` | Mode terminal seulement. |
  | `Shortcut: Ctrl+C …` | Mode terminal seulement. |

modified docs/fr/reference/configuration.md +4 -0 Viewed

@@ -31,6 +31,7 @@
31	31	\| `lsp` \| dictionnaire \| voir ci-dessous \| Serveurs de langage derrière l'outil `lsp`. Aucun serveur listé = pas d'outil. \|
32	32	\| `tui` \| bool \| `false` \| Démarre dans le mode fenêtre unique sans l'option `-tui`. Retombe sur le mode ligne quand stdin ou stdout n'est pas un terminal ; ignoré avec `-acp`. \|
33	33	\| `permissions` \| chaîne \| `auto` \| `ask` : en mode fenêtre, `bash`, `write_file` et `edit_file` attendent la réponse de l'utilisateur (`a`/`s`/`r`) avant de s'exécuter. `auto` : les outils s'exécutent aussitôt. Le mode ligne et la façade ACP ne la lisent pas (ACP a son propre dialogue). Doit valoir `auto` ou `ask`. \|
	34	+\| `mmConfigs` \| dictionnaire nom → chemin \| aucun \| Les autres agents mm que celui-ci peut interroger. Chaque nom génère `/ask-<nom>` et `/ask-with-context-<nom>` (voir les [commandes slash](slash-commands.md#interroger-un-autre-agent)). Un chemin relatif l'est au dossier du fichier qui le déclare. Noms : lettres minuscules, chiffres et `-`, commençant par une lettre ou un chiffre, ne commençant pas par `with-context`. Un chemin dont le fichier n'existe pas donne un avertissement au démarrage, pas une erreur. \|
34	35
35	36	## Clés de `context`
36	37
@@ -117,6 +118,9 @@ Signalées sous la forme `[config error: <chemin>: <message>]`, code de sortie 1
117	118	\| `lsp.servers[N].extensions: "go" must start with a dot, like ".go"` \| une extension sans son point \|
118	119	\| `lsp.servers[N]: extension ".ts" is already served by "typescript"` \| deux serveurs sur une même extension \|
119	120	\| `permissions must be "auto" or "ask"` \| toute autre valeur \|
	121	+\| `mmConfigs: "Bob" is not a valid name (…)` \| un nom avec des majuscules, des espaces ou d'autres caractères, ou commençant par `-` \|
	122	+\| `mmConfigs: "with-context-x" is not a valid name (it would read as /ask-with-context-…)` \| un nom commençant par `with-context` \|
	123	+\| `mmConfigs.<nom>: the path must not be empty` \| `<nom>: ""` \|
120	124	\| `unknown provider "x" (known: dmr, llamacpp, llmman, ollama)` \| Signalée par le moteur, sous la forme `[engine error: …]`. \|
121	125
122	126	Un avertissement au démarrage, et non une erreur, est affiché quand `context.enabled` vaut `true`, que la fenêtre est inconnue et que `context.maxMessages` vaut `0`.

  | `lsp` | dictionnaire | voir ci-dessous | Serveurs de langage derrière l'outil `lsp`. Aucun serveur listé = pas d'outil. |
  | `tui` | bool | `false` | Démarre dans le mode fenêtre unique sans l'option `-tui`. Retombe sur le mode ligne quand stdin ou stdout n'est pas un terminal ; ignoré avec `-acp`. |
  | `permissions` | chaîne | `auto` | `ask` : en mode fenêtre, `bash`, `write_file` et `edit_file` attendent la réponse de l'utilisateur (`a`/`s`/`r`) avant de s'exécuter. `auto` : les outils s'exécutent aussitôt. Le mode ligne et la façade ACP ne la lisent pas (ACP a son propre dialogue). Doit valoir `auto` ou `ask`. |
++| `mmConfigs` | dictionnaire nom → chemin | aucun | Les autres agents mm que celui-ci peut interroger. Chaque nom génère `/ask-<nom>` et `/ask-with-context-<nom>` (voir les [commandes slash](slash-commands.md#interroger-un-autre-agent)). Un chemin relatif l'est au dossier du fichier qui le déclare. Noms : lettres minuscules, chiffres et `-`, commençant par une lettre ou un chiffre, ne commençant pas par `with-context`. Un chemin dont le fichier n'existe pas donne un avertissement au démarrage, pas une erreur. |
  ## Clés de `context`
  | `lsp.servers[N].extensions: "go" must start with a dot, like ".go"` | une extension sans son point |
  | `lsp.servers[N]: extension ".ts" is already served by "typescript"` | deux serveurs sur une même extension |
  | `permissions must be "auto" or "ask"` | toute autre valeur |
++| `mmConfigs: "Bob" is not a valid name (…)` | un nom avec des majuscules, des espaces ou d'autres caractères, ou commençant par `-` |
++| `mmConfigs: "with-context-x" is not a valid name (it would read as /ask-with-context-…)` | un nom commençant par `with-context` |
++| `mmConfigs.<nom>: the path must not be empty` | `<nom>: ""` |
  | `unknown provider "x" (known: dmr, llamacpp, llmman, ollama)` | Signalée par le moteur, sous la forme `[engine error: …]`. |
  Un avertissement au démarrage, et non une erreur, est affiché quand `context.enabled` vaut `true`, que la fenêtre est inconnue et que `context.maxMessages` vaut `0`.

modified docs/fr/reference/slash-commands.md +45 -0 Viewed

@@ -17,6 +17,49 @@
17	17
18	18	Avec `sessions.enabled`, `/new` laisse le fichier de la session courante tel quel ; la question suivante ouvre un nouveau fichier. Une ligne vide à l'invite est ignorée.
19	19
	20	+## Interroger un autre agent
	21	+
	22	+Présentes seulement quand `mmConfigs` déclare au moins un agent ; `/help` et le sélecteur `/` les listent. `<nom>` est une clé de `mmConfigs`.
	23	+
	24	+\| Commande \| Façades \| Effet \|
	25	+\|----------\|---------\|-------\|
	26	+\| `/ask-<nom> <prompt>` \| terminal, fenêtre, éditeur \| Envoie `<prompt>` seul à `<nom>`. \|
	27	+\| `/ask-with-context-<nom> <prompt>` \| terminal, fenêtre, éditeur \| Envoie une transcription de la conversation courante, puis `<prompt>`. \|
	28	+\| `/ask <nom> <prompt>` \| terminal, fenêtre \| Comme `/ask-<nom> <prompt>`. \|
	29	+\| `/ask-with-context <nom> <prompt>` \| terminal, fenêtre \| Comme `/ask-with-context-<nom> <prompt>`. \|
	30	+
	31	+Pendant l'exécution :
	32	+
	33	+\| Élément \| Comportement \|
	34	+\|---------\|--------------\|
	35	+\| Première ligne \| `🗣 asking <nom> (<fournisseur>/<modèle>)…`, ou `…, with the conversation…` pour les formes avec contexte. \|
	36	+\| Réponse de `<nom>` \| Affichée au fil de l'eau là où s'affichent les réponses. Son raisonnement est affiché comme un raisonnement quand `showThinking` est actif. \|
	37	+\| Appels d'outils de `<nom>` \| Affichés comme des appels d'outils intitulés `<nom> · <outil>: <arguments>`. Les demandes de permission passent par cette façade : aucune en mode ligne (autorisé), `permissions` en mode fenêtre (`a`/`s`/`r`, « pour cette session » indexé sur `<nom>/<outil>`), la boîte de dialogue de l'éditeur en ACP. \|
	38	+\| `/abort`, `Esc`, `Ctrl+C`, `session/cancel` \| Annulent la demande ; `🛑 Ask aborted — nothing kept.` \|
	39	+\| `/quit` \| Annule la demande et quitte. \|
	40	+\| Toute autre ligne (mode ligne) \| `⚠️ <nom> is still answering — /abort to stop it.` ; la ligne est ignorée. \|
	41	+
	42	+À la fin :
	43	+
	44	+\| Issue \| Historique \| Message \|
	45	+\|-------\|------------\|---------\|
	46	+\| Réponse obtenue \| Deux messages ajoutés : utilisateur `[asked <nom> (<fournisseur>/<modèle>)[, with this conversation as context]: <prompt>]`, assistant `[<nom>'s answer]` + la réponse (plafonnée à `maxOutput` caractères, début et fin conservés). Enregistrés avec la session. \| `🗣 <nom>'s answer is in the history for your next question` \|
	47	+\| Annulée \| Inchangé. \| `🛑 Ask aborted — nothing kept.` \|
	48	+\| Échec \| Inchangé ; enregistré dans le fichier de session avec `where: "ask"`. \| `[error: …]` (terminal, fenêtre) ou une erreur JSON-RPC (éditeur). \|
	49	+
	50	+La transcription envoyée par les formes avec contexte contient les questions de l'utilisateur, les réponses de l'assistant, une ligne par appel d'outil (`Assistant ran <outil>: <arguments>`, arguments coupés à 160 caractères) et le résumé de compaction s'il existe ; jamais le prompt système, les sorties d'outils ni le raisonnement. Plafonnée à `maxOutput` caractères, en gardant les lignes les plus récentes, `[… earlier conversation omitted …]` marquant la coupure.
	51	+
	52	+\| Erreur \| Cause \|
	53	+\|--------\|-------\|
	54	+\| `usage: /ask <agent> <prompt> — agents: …` \| `/ask` sans nom. \|
	55	+\| `unknown agent "x" — agents: …` \| `/ask x …` avec un nom absent de `mmConfigs`. \|
	56	+\| `nothing to ask: type the question after the command` \| Pas de prompt. \|
	57	+\| `<nom>: config file <chemin>: no such file or directory` \| Le fichier de `mmConfigs.<nom>` n'existe pas. \|
	58	+\| `<nom> did not start: … — <dernière ligne de stderr>` \| Le processus de l'agent s'est arrêté ou n'a pas répondu à `initialize` en 60 s. \|
	59	+\| `<nom> failed: …` \| Le tour de l'agent s'est terminé sur une erreur (sa propre explication en une ligne). \|
	60	+\| `<nom> gave no answer` \| L'agent a terminé son tour sans aucun texte. \|
	61	+\| `this agent was itself asked through 2 level(s) of /ask; it cannot ask further (max 2)` \| `MM_ASK_DEPTH` vaut 2. \|
	62	+
20	63	## `@chemin` dans une question
21	64
22	65	\| Saisie \| Effet \|
@@ -134,4 +177,6 @@ Deux commandes existent en mode ACP : `/new` et `/compact`. Elles sont annoncée
134	177
135	178	Chaque skill y est proposé comme une commande à part entière : ACP n'a pas de sélecteur pour les skills, donc `/greetings Bob` dans le menu de l'éditeur est l'équivalent de `$greetings Bob` dans la fenêtre, et il est réécrit en exactement ça — voir la [référence ACP](acp.md#les-skills-comme-commandes).
136	179
	180	+Avec `mmConfigs`, `/ask-<nom>` et `/ask-with-context-<nom>` sont annoncées aussi, chacune avec l'indication de saisie `<prompt>` ; le `/ask <nom>` générique ne l'est pas, car le menu d'un éditeur complète un nom de commande mais pas son argument. Voir [Interroger un autre agent](#interroger-un-autre-agent).
	181	+
137	182	`/quit` et `/abort` n'existent pas en mode ACP : l'éditeur possède le processus et le termine en fermant stdin, et il annule un tour avec sa propre touche d'arrêt via `session/cancel`. Voir la [référence ACP](acp.md).

  Avec `sessions.enabled`, `/new` laisse le fichier de la session courante tel quel ; la question suivante ouvre un nouveau fichier. Une ligne vide à l'invite est ignorée.
++## Interroger un autre agent
++
++Présentes seulement quand `mmConfigs` déclare au moins un agent ; `/help` et le sélecteur `/` les listent. `<nom>` est une clé de `mmConfigs`.
++
++| Commande | Façades | Effet |
++|----------|---------|-------|
++| `/ask-<nom> <prompt>` | terminal, fenêtre, éditeur | Envoie `<prompt>` seul à `<nom>`. |
++| `/ask-with-context-<nom> <prompt>` | terminal, fenêtre, éditeur | Envoie une transcription de la conversation courante, puis `<prompt>`. |
++| `/ask <nom> <prompt>` | terminal, fenêtre | Comme `/ask-<nom> <prompt>`. |
++| `/ask-with-context <nom> <prompt>` | terminal, fenêtre | Comme `/ask-with-context-<nom> <prompt>`. |
++
++Pendant l'exécution :
++
++| Élément | Comportement |
++|---------|--------------|
++| Première ligne | `🗣 asking <nom> (<fournisseur>/<modèle>)…`, ou `…, with the conversation…` pour les formes avec contexte. |
++| Réponse de `<nom>` | Affichée au fil de l'eau là où s'affichent les réponses. Son raisonnement est affiché comme un raisonnement quand `showThinking` est actif. |
++| Appels d'outils de `<nom>` | Affichés comme des appels d'outils intitulés `<nom> · <outil>: <arguments>`. Les demandes de permission passent par cette façade : aucune en mode ligne (autorisé), `permissions` en mode fenêtre (`a`/`s`/`r`, « pour cette session » indexé sur `<nom>/<outil>`), la boîte de dialogue de l'éditeur en ACP. |
++| `/abort`, `Esc`, `Ctrl+C`, `session/cancel` | Annulent la demande ; `🛑 Ask aborted — nothing kept.` |
++| `/quit` | Annule la demande et quitte. |
++| Toute autre ligne (mode ligne) | `⚠️ <nom> is still answering — /abort to stop it.` ; la ligne est ignorée. |
++
++À la fin :
++
++| Issue | Historique | Message |
++|-------|------------|---------|
++| Réponse obtenue | Deux messages ajoutés : utilisateur `[asked <nom> (<fournisseur>/<modèle>)[, with this conversation as context]: <prompt>]`, assistant `[<nom>'s answer]` + la réponse (plafonnée à `maxOutput` caractères, début et fin conservés). Enregistrés avec la session. | `🗣 <nom>'s answer is in the history for your next question` |
++| Annulée | Inchangé. | `🛑 Ask aborted — nothing kept.` |
++| Échec | Inchangé ; enregistré dans le fichier de session avec `where: "ask"`. | `[error: …]` (terminal, fenêtre) ou une erreur JSON-RPC (éditeur). |
++
++La transcription envoyée par les formes avec contexte contient les questions de l'utilisateur, les réponses de l'assistant, une ligne par appel d'outil (`Assistant ran <outil>: <arguments>`, arguments coupés à 160 caractères) et le résumé de compaction s'il existe ; jamais le prompt système, les sorties d'outils ni le raisonnement. Plafonnée à `maxOutput` caractères, en gardant les lignes les plus récentes, `[… earlier conversation omitted …]` marquant la coupure.
++
++| Erreur | Cause |
++|--------|-------|
++| `usage: /ask <agent> <prompt> — agents: …` | `/ask` sans nom. |
++| `unknown agent "x" — agents: …` | `/ask x …` avec un nom absent de `mmConfigs`. |
++| `nothing to ask: type the question after the command` | Pas de prompt. |
++| `<nom>: config file <chemin>: no such file or directory` | Le fichier de `mmConfigs.<nom>` n'existe pas. |
++| `<nom> did not start: … — <dernière ligne de stderr>` | Le processus de l'agent s'est arrêté ou n'a pas répondu à `initialize` en 60 s. |
++| `<nom> failed: …` | Le tour de l'agent s'est terminé sur une erreur (sa propre explication en une ligne). |
++| `<nom> gave no answer` | L'agent a terminé son tour sans aucun texte. |
++| `this agent was itself asked through 2 level(s) of /ask; it cannot ask further (max 2)` | `MM_ASK_DEPTH` vaut 2. |
++
  ## `@chemin` dans une question
  | Saisie | Effet |
  Chaque **skill** y est proposé comme une commande à part entière : ACP n'a pas de sélecteur pour les skills, donc `/greetings Bob` dans le menu de l'éditeur est l'équivalent de `$greetings Bob` dans la fenêtre, et il est réécrit en exactement ça — voir la [référence ACP](acp.md#les-skills-comme-commandes).
++Avec `mmConfigs`, `/ask-<nom>` et `/ask-with-context-<nom>` sont annoncées aussi, chacune avec l'indication de saisie `<prompt>` ; le `/ask <nom>` générique ne l'est pas, car le menu d'un éditeur complète un nom de commande mais pas son argument. Voir [Interroger un autre agent](#interroger-un-autre-agent).
++
  `/quit` et `/abort` n'existent pas en mode ACP : l'éditeur possède le processus et le termine en fermant stdin, et il annule un tour avec sa propre touche d'arrêt via `session/cancel`. Voir la [référence ACP](acp.md).

modified internal/acp/acp.go +22 -3 Viewed

@@ -29,6 +29,7 @@ import (
29	29	"strings"
30	30	"sync"
31	31
	32	+ "mm/internal/ask"
32	33	"mm/internal/command"
33	34	"mm/internal/compact"
34	35	"mm/internal/config"
@@ -76,6 +77,9 @@ type front struct {
76	77	// so one running agent keeps one store per project.
77	78	sessionsDir string
78	79
	80	+ // asker runs /ask-<agent> (internal/ask); nil when none is configured.
	81	+ asker ask.Runner
	82	+
79	83	mu sync.Mutex // guards sessions
80	84	sessions map[sdk.SessionId]*session
81	85
@@ -91,8 +95,11 @@ var _ sdk.Agent = (*front)(nil)
91	95	// Run wires the agent side of an ACP connection on stdio and blocks until the
92	96	// client closes it. Same arguments as agent.Run: main.go's wiring — engine,
93	97	// system prompt, tool list — is identical for the two front ends.
94		-func Run(ctx context.Context, e *engine.Engine, system string, tools []ai.ToolRef) {
	98	+//
	99	+// asker runs the /ask-<agent> commands; nil when no agent is configured.
	100	+func Run(ctx context.Context, e *engine.Engine, system string, tools []ai.ToolRef, asker ask.Runner) {
95	101	f := newFront(e, system, tools, os.Stdout, os.Stdin)
	102	+ f.asker = asker
96	103	select {
97	104	case <-f.conn.Done(): // the editor closed our stdin: we are done
98	105	case <-ctx.Done():
@@ -281,7 +288,13 @@ func (f *front) NewSession(ctx context.Context, p sdk.NewSessionRequest) (sdk.Ne
281	288	func availableCommands() []sdk.AvailableCommand {
282	289	var out []sdk.AvailableCommand
283	290	for _, c := range command.For(command.Editor) {
284		- out = append(out, sdk.AvailableCommand{Name: c.Name, Description: c.Description})
	291	+ cmd := sdk.AvailableCommand{Name: c.Name, Description: c.Description}
	292	+ if c.Hint != "" {
	293	+ // /ask-bob takes the question: the editor shows the hint where
	294	+ // it goes, as it does for a skill's parameters.
	295	+ cmd.Input = &sdk.AvailableCommandInput{Unstructured: &sdk.UnstructuredCommandInput{Hint: c.Hint}}
	296	+ }
	297	+ out = append(out, cmd)
285	298	}
286	299	for _, s := range skills.Declared() {
287	300	if commandNamed(s.Name) {
@@ -370,9 +383,15 @@ func (f *front) Prompt(ctx context.Context, p sdk.PromptRequest) (sdk.PromptResp
370	383
371	384	// The sink is what turns the tools' events into session/update messages;
372	385	// installing it is what switches engine+tools to "ACP mode" for this turn.
373		- ui.SetActive(&sink{f: f, ctx: turnCtx, sid: p.SessionId, s: s})
	386	+ k := &sink{f: f, ctx: turnCtx, sid: p.SessionId, s: s}
	387	+ ui.SetActive(k)
374	388	defer ui.SetActive(nil)
375	389
	390	+ // /ask-bob: another agent answers instead of our model (ask.go).
	391	+ if req, isAsk, err := command.ParseAsk(input); isAsk {
	392	+ return f.askAgent(ctx, turnCtx, p.SessionId, s, k, req, err)
	393	+ }
	394	+
376	395	// A typed "@path" — the editor's picker bypassed, or an editor without
377	396	// one — is understood like the resource_link the picker would have sent,
378	397	// against the session's cwd, which is the project the editor opened.

  	"strings"
  	"sync"
++	"mm/internal/ask"
  	"mm/internal/command"
  	"mm/internal/compact"
  	"mm/internal/config"
  	// so one running agent keeps one store per project.
  	sessionsDir string
++	// asker runs /ask-<agent> (internal/ask); nil when none is configured.
++	asker ask.Runner
++
  	mu       sync.Mutex // guards sessions
  	sessions map[sdk.SessionId]*session
  // Run wires the agent side of an ACP connection on stdio and blocks until the
  // client closes it. Same arguments as agent.Run: main.go's wiring — engine,
  // system prompt, tool list — is identical for the two front ends.
--func Run(ctx context.Context, e *engine.Engine, system string, tools []ai.ToolRef) {
++//
++// asker runs the /ask-<agent> commands; nil when no agent is configured.
++func Run(ctx context.Context, e *engine.Engine, system string, tools []ai.ToolRef, asker ask.Runner) {
  	f := newFront(e, system, tools, os.Stdout, os.Stdin)
++	f.asker = asker
  	select {
  	case <-f.conn.Done(): // the editor closed our stdin: we are done
  	case <-ctx.Done():
  func availableCommands() []sdk.AvailableCommand {
  	var out []sdk.AvailableCommand
  	for _, c := range command.For(command.Editor) {
--		out = append(out, sdk.AvailableCommand{Name: c.Name, Description: c.Description})
++		cmd := sdk.AvailableCommand{Name: c.Name, Description: c.Description}
++		if c.Hint != "" {
++			// /ask-bob takes the question: the editor shows the hint where
++			// it goes, as it does for a skill's parameters.
++			cmd.Input = &sdk.AvailableCommandInput{Unstructured: &sdk.UnstructuredCommandInput{Hint: c.Hint}}
++		}
++		out = append(out, cmd)
  	}
  	for _, s := range skills.Declared() {
  		if commandNamed(s.Name) {
  	// The sink is what turns the tools' events into session/update messages;
  	// installing it is what switches engine+tools to "ACP mode" for this turn.
--	ui.SetActive(&sink{f: f, ctx: turnCtx, sid: p.SessionId, s: s})
++	k := &sink{f: f, ctx: turnCtx, sid: p.SessionId, s: s}
++	ui.SetActive(k)
  	defer ui.SetActive(nil)
++	// /ask-bob: another agent answers instead of our model (ask.go).
++	if req, isAsk, err := command.ParseAsk(input); isAsk {
++		return f.askAgent(ctx, turnCtx, p.SessionId, s, k, req, err)
++	}
++
  	// A typed "@path" — the editor's picker bypassed, or an editor without
  	// one — is understood like the resource_link the picker would have sent,
  	// against the session's cwd, which is the project the editor opened.

added internal/acp/ask.go +42 -0 Viewed

new file mode 100644
@@ -0,0 +1,42 @@
	1	+package acp
	2	+
	3	+import (
	4	+ "context"
	5	+ "fmt"
	6	+ "os"
	7	+
	8	+ "mm/internal/ask"
	9	+ "mm/internal/command"
	10	+ "mm/internal/store"
	11	+
	12	+ sdk "github.com/coder/acp-go-sdk"
	13	+)
	14	+
	15	+// askAgent is /ask-<agent> in an editor: the other agent's answer streams
	16	+// into the thread as agent text, its tool calls as tool calls (named after
	17	+// it), its permission requests as this session's own dialog — and the
	18	+// exchange joins the history for the next question. Like a turn, it can be
	19	+// cancelled (stopReason "cancelled"), and a failure is a JSON-RPC error in the
	20	+// agent's words and a record in the session file; either way the history is
	21	+// left as it was.
	22	+func (f front) askAgent(ctx, turnCtx context.Context, sid sdk.SessionId, s session, k *sink, req command.AskRequest, parseErr error) (sdk.PromptResponse, error) {
	23	+ if parseErr != nil {
	24	+ k.Text("⚠️ " + parseErr.Error())
	25	+ return sdk.PromptResponse{StopReason: sdk.StopReasonEndTurn, Usage: s.usage.totals()}, nil
	26	+ }
	27	+ k.Text(ask.Label(req) + "\n\n")
	28	+ exchange, err := f.asker.Run(turnCtx, req, s.messages, s.cwd, k)
	29	+ switch {
	30	+ case turnCtx.Err() != nil:
	31	+ return sdk.PromptResponse{StopReason: sdk.StopReasonCancelled, Usage: s.usage.totals()}, nil
	32	+ case err != nil:
	33	+ fmt.Fprintf(os.Stderr, "[acp] session %s: /ask %s failed: %v\n", sid, req.Agent, err)
	34	+ f.fail(sid, s, store.NewFailure(s.messages, store.FailAsk, err.Error(), err))
	35	+ return sdk.PromptResponse{}, sdk.NewInternalError(map[string]any{"error": err.Error()})
	36	+ }
	37	+ s.messages = append(s.messages, exchange...)
	38	+ f.persist(sid, s)
	39	+ fmt.Fprintf(os.Stderr, "[acp] session %s: /ask %s answered\n", sid, req.Agent)
	40	+ f.sendUsage(context.WithoutCancel(ctx), sid, contextUsed(s))
	41	+ return sdk.PromptResponse{StopReason: sdk.StopReasonEndTurn, Usage: s.usage.totals()}, nil
	42	+}

++package acp
++
++import (
++	"context"
++	"fmt"
++	"os"
++
++	"mm/internal/ask"
++	"mm/internal/command"
++	"mm/internal/store"
++
++	sdk "github.com/coder/acp-go-sdk"
++)
++
++// askAgent is /ask-<agent> in an editor: the other agent's answer streams
++// into the thread as agent text, its tool calls as tool calls (named after
++// it), its permission requests as this session's own dialog — and the
++// exchange joins the history for the next question. Like a turn, it can be
++// cancelled (stopReason "cancelled"), and a failure is a JSON-RPC error in the
++// agent's words and a record in the session file; either way the history is
++// left as it was.
++func (f *front) askAgent(ctx, turnCtx context.Context, sid sdk.SessionId, s *session, k *sink, req command.AskRequest, parseErr error) (sdk.PromptResponse, error) {
++	if parseErr != nil {
++		k.Text("⚠️ " + parseErr.Error())
++		return sdk.PromptResponse{StopReason: sdk.StopReasonEndTurn, Usage: s.usage.totals()}, nil
++	}
++	k.Text(ask.Label(req) + "\n\n")
++	exchange, err := f.asker.Run(turnCtx, req, s.messages, s.cwd, k)
++	switch {
++	case turnCtx.Err() != nil:
++		return sdk.PromptResponse{StopReason: sdk.StopReasonCancelled, Usage: s.usage.totals()}, nil
++	case err != nil:
++		fmt.Fprintf(os.Stderr, "[acp] session %s: /ask %s failed: %v\n", sid, req.Agent, err)
++		f.fail(sid, s, store.NewFailure(s.messages, store.FailAsk, err.Error(), err))
++		return sdk.PromptResponse{}, sdk.NewInternalError(map[string]any{"error": err.Error()})
++	}
++	s.messages = append(s.messages, exchange...)
++	f.persist(sid, s)
++	fmt.Fprintf(os.Stderr, "[acp] session %s: /ask %s answered\n", sid, req.Agent)
++	f.sendUsage(context.WithoutCancel(ctx), sid, contextUsed(s))
++	return sdk.PromptResponse{StopReason: sdk.StopReasonEndTurn, Usage: s.usage.totals()}, nil
++}

added internal/acp/ask_test.go +89 -0 Viewed

new file mode 100644
@@ -0,0 +1,89 @@
	1	+package acp
	2	+
	3	+import (
	4	+ "context"
	5	+ "strings"
	6	+ "testing"
	7	+
	8	+ "mm/internal/command"
	9	+ "mm/internal/config"
	10	+ "mm/internal/engine"
	11	+ "mm/internal/ui"
	12	+
	13	+ sdk "github.com/coder/acp-go-sdk"
	14	+
	15	+ "github.com/firebase/genkit/go/ai"
	16	+)
	17	+
	18	+// fakeAsker answers through the sink, as the real pool streams.
	19	+type fakeAsker struct{ answer string }
	20	+
	21	+func (a fakeAsker) Run(_ context.Context, req command.AskRequest, _ []ai.Message, _ string, sink ui.Sink) ([]ai.Message, error) {
	22	+ sink.Text(a.answer)
	23	+ return []*ai.Message{ai.NewUserTextMessage("[asked " + req.Agent + ": " + req.Prompt + "]"), ai.NewModelTextMessage(a.answer)}, nil
	24	+}
	25	+
	26	+// withBob configures one agent and its generated commands for one test.
	27	+func withBob(t *testing.T) {
	28	+ t.Helper()
	29	+ saved := config.Cfg
	30	+ t.Cleanup(func() { config.Cfg = saved; command.SetAgents(nil) })
	31	+ config.Cfg.MMConfigs = map[string]string{"bob": "/nowhere/agent.bob.yaml"}
	32	+ command.SetAgents([]string{"bob"})
	33	+}
	34	+
	35	+// The editor is offered one command per agent, each with the <prompt> hint,
	36	+// and never the generic /ask, whose argument it could not complete.
	37	+func TestAskCommandsAreAnnouncedPerAgent(t *testing.T) {
	38	+ withBob(t)
	39	+ hints := map[string]string{}
	40	+ for _, c := range availableCommands() {
	41	+ if c.Input != nil && c.Input.Unstructured != nil {
	42	+ hints[c.Name] = c.Input.Unstructured.Hint
	43	+ } else {
	44	+ hints[c.Name] = ""
	45	+ }
	46	+ }
	47	+ for _, name := range []string{"ask-bob", "ask-with-context-bob"} {
	48	+ if hint, ok := hints[name]; !ok \|\| hint != "<prompt>" {
	49	+ t.Errorf("%s: announced %v with hint %q, want <prompt>", name, ok, hint)
	50	+ }
	51	+ }
	52	+ if _, ok := hints["ask"]; ok {
	53	+ t.Error("the generic /ask is announced to the editor")
	54	+ }
	55	+}
	56	+
	57	+// /ask-bob over the wire: bob's answer streams as agent text after the label,
	58	+// the turn ends normally, and the exchange is in the session's history.
	59	+func TestAskOverACP(t *testing.T) {
	60	+ withBob(t)
	61	+ w := newWire(t, "sys")
	62	+ f := frontOf(t, w)
	63	+ f.e = &engine.Engine{}
	64	+ f.asker = fakeAsker{answer: "four"}
	65	+ w.send(1, "initialize", `{"protocolVersion":1,"clientCapabilities":{}}`)
	66	+ w.until(1)
	67	+ w.send(2, "session/new", `{"cwd":"`+t.TempDir()+`","mcpServers":[]}`)
	68	+ _, resp := w.until(2)
	69	+ sid := resp["result"].(map[string]any)["sessionId"].(string)
	70	+
	71	+ w.send(3, "session/prompt", `{"sessionId":"`+sid+`","prompt":[{"type":"text","text":"/ask-bob what is 2+2?"}]}`)
	72	+ notes, resp := w.until(3)
	73	+ if stop := resp["result"].(map[string]any)["stopReason"]; stop != "end_turn" {
	74	+ t.Errorf("stopReason = %v", stop)
	75	+ }
	76	+ var text strings.Builder
	77	+ for _, n := range notes {
	78	+ if n["sessionUpdate"] == "agent_message_chunk" {
	79	+ text.WriteString(n["content"].(map[string]any)["text"].(string))
	80	+ }
	81	+ }
	82	+ if got := text.String(); !strings.Contains(got, "🗣 asking bob") \|\| !strings.HasSuffix(got, "four") {
	83	+ t.Errorf("agent text = %q", got)
	84	+ }
	85	+ msgs := f.sessions[sdk.SessionId(sid)].messages
	86	+ if len(msgs) != 3 \|\| msgs[2].Text() != "four" {
	87	+ t.Errorf("history: %d messages", len(msgs))
	88	+ }
	89	+}

++package acp
++
++import (
++	"context"
++	"strings"
++	"testing"
++
++	"mm/internal/command"
++	"mm/internal/config"
++	"mm/internal/engine"
++	"mm/internal/ui"
++
++	sdk "github.com/coder/acp-go-sdk"
++
++	"github.com/firebase/genkit/go/ai"
++)
++
++// fakeAsker answers through the sink, as the real pool streams.
++type fakeAsker struct{ answer string }
++
++func (a fakeAsker) Run(_ context.Context, req command.AskRequest, _ []*ai.Message, _ string, sink ui.Sink) ([]*ai.Message, error) {
++	sink.Text(a.answer)
++	return []*ai.Message{ai.NewUserTextMessage("[asked " + req.Agent + ": " + req.Prompt + "]"), ai.NewModelTextMessage(a.answer)}, nil
++}
++
++// withBob configures one agent and its generated commands for one test.
++func withBob(t *testing.T) {
++	t.Helper()
++	saved := config.Cfg
++	t.Cleanup(func() { config.Cfg = saved; command.SetAgents(nil) })
++	config.Cfg.MMConfigs = map[string]string{"bob": "/nowhere/agent.bob.yaml"}
++	command.SetAgents([]string{"bob"})
++}
++
++// The editor is offered one command per agent, each with the <prompt> hint,
++// and never the generic /ask, whose argument it could not complete.
++func TestAskCommandsAreAnnouncedPerAgent(t *testing.T) {
++	withBob(t)
++	hints := map[string]string{}
++	for _, c := range availableCommands() {
++		if c.Input != nil && c.Input.Unstructured != nil {
++			hints[c.Name] = c.Input.Unstructured.Hint
++		} else {
++			hints[c.Name] = ""
++		}
++	}
++	for _, name := range []string{"ask-bob", "ask-with-context-bob"} {
++		if hint, ok := hints[name]; !ok || hint != "<prompt>" {
++			t.Errorf("%s: announced %v with hint %q, want <prompt>", name, ok, hint)
++		}
++	}
++	if _, ok := hints["ask"]; ok {
++		t.Error("the generic /ask is announced to the editor")
++	}
++}
++
++// /ask-bob over the wire: bob's answer streams as agent text after the label,
++// the turn ends normally, and the exchange is in the session's history.
++func TestAskOverACP(t *testing.T) {
++	withBob(t)
++	w := newWire(t, "sys")
++	f := frontOf(t, w)
++	f.e = &engine.Engine{}
++	f.asker = fakeAsker{answer: "four"}
++	w.send(1, "initialize", `{"protocolVersion":1,"clientCapabilities":{}}`)
++	w.until(1)
++	w.send(2, "session/new", `{"cwd":"`+t.TempDir()+`","mcpServers":[]}`)
++	_, resp := w.until(2)
++	sid := resp["result"].(map[string]any)["sessionId"].(string)
++
++	w.send(3, "session/prompt", `{"sessionId":"`+sid+`","prompt":[{"type":"text","text":"/ask-bob what is 2+2?"}]}`)
++	notes, resp := w.until(3)
++	if stop := resp["result"].(map[string]any)["stopReason"]; stop != "end_turn" {
++		t.Errorf("stopReason = %v", stop)
++	}
++	var text strings.Builder
++	for _, n := range notes {
++		if n["sessionUpdate"] == "agent_message_chunk" {
++			text.WriteString(n["content"].(map[string]any)["text"].(string))
++		}
++	}
++	if got := text.String(); !strings.Contains(got, "🗣 asking bob") || !strings.HasSuffix(got, "four") {
++		t.Errorf("agent text = %q", got)
++	}
++	msgs := f.sessions[sdk.SessionId(sid)].messages
++	if len(msgs) != 3 || msgs[2].Text() != "four" {
++		t.Errorf("history: %d messages", len(msgs))
++	}
++}

modified internal/agent/agent.go +20 -1 Viewed

@@ -10,6 +10,7 @@ import (
10	10	"os/signal"
11	11	"strings"
12	12
	13	+ "mm/internal/ask"
13	14	"mm/internal/clipboard"
14	15	"mm/internal/command"
15	16	"mm/internal/compact"
@@ -38,7 +39,9 @@ type genResult struct {
38	39
39	40	// Run starts the loop: read user input → generate (with automatic tool calls) →
40	41	// print → start over, keeping the history.
41		-func Run(ctx context.Context, e engine.Engine, system string, tools []ai.ToolRef, d detector.LoopDetector) {
	42	+//
	43	+// pool runs the /ask commands (another mm agent); nil when none is configured.
	44	+func Run(ctx context.Context, e engine.Engine, system string, tools []ai.ToolRef, d detector.LoopDetector, pool ask.Runner) {
42	45	messages := session.Fresh(system)
43	46	ctx, cancelAll := context.WithCancel(ctx)
44	47	defer cancelAll()
@@ -140,6 +143,22 @@ func Run(ctx context.Context, e *engine.Engine, system string, tools []ai.ToolRe
140	143	// DURING a generation, such a command goes through the "new command
141	144	// received" branch below, which cancels the turn and re-injects it
142	145	// here — so it always applies to the finished history.
	146	+ // /ask-bob and its family: another agent answers, and its answer
	147	+ // joins the history for the next question. Between two generations
	148	+ // too, for the same reason as the commands below.
	149	+ if req, ok, err := command.ParseAsk(input); ok {
	150	+ if err != nil {
	151	+ fmt.Println("⚠️ " + err.Error())
	152	+ continue
	153	+ }
	154	+ var quit bool
	155	+ messages, quit = asking{pool: pool, cwd: cwd, keep: keep, sigCh: sigCh, inputCh: inputCh}.run(ctx, req, messages)
	156	+ if quit {
	157	+ fmt.Println("Goodbye.")
	158	+ return
	159	+ }
	160	+ continue
	161	+ }
143	162	if next, handled := (&loop{ctx: ctx, e: e, d: d, keep: keep, system: system}).historyCommand(input, messages); handled {
144	163	messages = next
145	164	continue

  	"os/signal"
  	"strings"
++	"mm/internal/ask"
  	"mm/internal/clipboard"
  	"mm/internal/command"
  	"mm/internal/compact"
  // Run starts the loop: read user input → generate (with automatic tool calls) →
  // print → start over, keeping the history.
--func Run(ctx context.Context, e *engine.Engine, system string, tools []ai.ToolRef, d *detector.LoopDetector) {
++//
++// pool runs the /ask commands (another mm agent); nil when none is configured.
++func Run(ctx context.Context, e *engine.Engine, system string, tools []ai.ToolRef, d *detector.LoopDetector, pool ask.Runner) {
  	messages := session.Fresh(system)
  	ctx, cancelAll := context.WithCancel(ctx)
  	defer cancelAll()
  		// DURING a generation, such a command goes through the "new command
  		// received" branch below, which cancels the turn and re-injects it
  		// here — so it always applies to the finished history.
++		// /ask-bob and its family: another agent answers, and its answer
++		// joins the history for the next question. Between two generations
++		// too, for the same reason as the commands below.
++		if req, ok, err := command.ParseAsk(input); ok {
++			if err != nil {
++				fmt.Println("⚠️ " + err.Error())
++				continue
++			}
++			var quit bool
++			messages, quit = asking{pool: pool, cwd: cwd, keep: keep, sigCh: sigCh, inputCh: inputCh}.run(ctx, req, messages)
++			if quit {
++				fmt.Println("Goodbye.")
++				return
++			}
++			continue
++		}
  		if next, handled := (&loop{ctx: ctx, e: e, d: d, keep: keep, system: system}).historyCommand(input, messages); handled {
  			messages = next
  			continue

added internal/agent/ask.go +125 -0 Viewed

new file mode 100644
@@ -0,0 +1,125 @@
	1	+package agent
	2	+
	3	+import (
	4	+ "context"
	5	+ "fmt"
	6	+ "os"
	7	+
	8	+ "mm/internal/ask"
	9	+ "mm/internal/command"
	10	+ "mm/internal/config"
	11	+ "mm/internal/spinner"
	12	+ "mm/internal/store"
	13	+ "mm/internal/ui"
	14	+
	15	+ "github.com/firebase/genkit/go/ai"
	16	+)
	17	+
	18	+// asking is what the REPL needs to run one /ask: the pool of agents, where
	19	+// the session lives, and the two channels a generation also listens to, so
	20	+// that Ctrl+C, /abort and /quit stop bob exactly as they stop a turn.
	21	+type asking struct {
	22	+ pool ask.Runner
	23	+ cwd string
	24	+ keep *store.Keeper
	25	+ sigCh <-chan os.Signal
	26	+ inputCh <-chan string
	27	+}
	28	+
	29	+// run asks req.Agent, streaming its answer and tool calls, and returns the
	30	+// history with the exchange appended — unchanged when the ask failed or was
	31	+// stopped. quit is true when /quit was typed meanwhile.
	32	+func (a asking) run(ctx context.Context, req command.AskRequest, messages []ai.Message) (out []ai.Message, quit bool) {
	33	+ dimln(ask.Label(req))
	34	+ askCtx, cancel := context.WithCancel(ctx)
	35	+ defer cancel()
	36	+ type result struct {
	37	+ exchange []*ai.Message
	38	+ err error
	39	+ }
	40	+ done := make(chan result, 1)
	41	+ go func() {
	42	+ exchange, err := a.pool.Run(askCtx, req, messages, a.cwd, &lineSink{cwd: a.cwd})
	43	+ done <- result{exchange, err}
	44	+ }()
	45	+
	46	+ for {
	47	+ select {
	48	+ case res := <-done:
	49	+ return a.settle(askCtx, req, messages, res.exchange, res.err), quit
	50	+ case <-a.sigCh:
	51	+ fmt.Println("\n🛑 Interrupted (Ctrl+C).")
	52	+ cancel()
	53	+ case input := <-a.inputCh:
	54	+ switch {
	55	+ case command.Is(input, command.Abort):
	56	+ fmt.Println("🛑 Aborting…")
	57	+ cancel()
	58	+ case command.Is(input, command.Quit):
	59	+ quit = true
	60	+ cancel()
	61	+ default:
	62	+ fmt.Printf("⚠️ %s is still answering — /abort to stop it.\n", req.Agent)
	63	+ }
	64	+ }
	65	+ }
	66	+}
	67	+
	68	+// settle reports how the ask ended and keeps the exchange when there is one.
	69	+func (a asking) settle(askCtx context.Context, req command.AskRequest, messages, exchange []ai.Message, err error) []ai.Message {
	70	+ fmt.Println()
	71	+ switch {
	72	+ case err != nil && askCtx.Err() != nil:
	73	+ fmt.Println("🛑 Ask aborted — nothing kept.")
	74	+ return messages
	75	+ case err != nil:
	76	+ fmt.Println("[error: " + err.Error() + "]")
	77	+ record(a.keep, messages, store.NewFailure(messages, store.FailAsk, err.Error(), err))
	78	+ return messages
	79	+ }
	80	+ messages = append(messages, exchange...)
	81	+ save(a.keep, messages)
	82	+ dimln(fmt.Sprintf("🗣 %s's answer is in the history for your next question", req.Agent))
	83	+ return messages
	84	+}
	85	+
	86	+// lineSink shows the other agent in the line mode: its answer as it streams,
	87	+// its thoughts dimmed, one line per tool call. Like the line mode's own
	88	+// tools, its calls are not gated: this front end has no permission prompt.
	89	+type lineSink struct {
	90	+ cwd string
	91	+}
	92	+
	93	+var _ ui.Sink = (*lineSink)(nil)
	94	+
	95	+func (s *lineSink) Text(text string) {
	96	+ spinner.Stop()
	97	+ fmt.Print(text)
	98	+}
	99	+
	100	+func (s *lineSink) Reasoning(text string) {
	101	+ if !config.Cfg.ShowThinking {
	102	+ return
	103	+ }
	104	+ spinner.Stop()
	105	+ if spinner.Styled() {
	106	+ fmt.Printf("\033[2;3m%s\033[0m", text)
	107	+ return
	108	+ }
	109	+ fmt.Print(text)
	110	+}
	111	+
	112	+func (s *lineSink) ToolStart(ev ui.ToolEvent) {
	113	+ spinner.Stop()
	114	+ fmt.Printf("\n🛠️ %s\n", ev.Title)
	115	+}
	116	+
	117	+func (s *lineSink) ToolEnd(_ string, res ui.ToolResult) {
	118	+ if res.Failed {
	119	+ dimln(" │ failed")
	120	+ }
	121	+}
	122	+
	123	+func (s *lineSink) Allow(context.Context, ui.ToolEvent) bool { return true }
	124	+func (s *lineSink) ToolRunning(string) {}
	125	+func (s *lineSink) WorkDir() string { return s.cwd }

++package agent
++
++import (
++	"context"
++	"fmt"
++	"os"
++
++	"mm/internal/ask"
++	"mm/internal/command"
++	"mm/internal/config"
++	"mm/internal/spinner"
++	"mm/internal/store"
++	"mm/internal/ui"
++
++	"github.com/firebase/genkit/go/ai"
++)
++
++// asking is what the REPL needs to run one /ask: the pool of agents, where
++// the session lives, and the two channels a generation also listens to, so
++// that Ctrl+C, /abort and /quit stop bob exactly as they stop a turn.
++type asking struct {
++	pool    ask.Runner
++	cwd     string
++	keep    *store.Keeper
++	sigCh   <-chan os.Signal
++	inputCh <-chan string
++}
++
++// run asks req.Agent, streaming its answer and tool calls, and returns the
++// history with the exchange appended — unchanged when the ask failed or was
++// stopped. quit is true when /quit was typed meanwhile.
++func (a asking) run(ctx context.Context, req command.AskRequest, messages []*ai.Message) (out []*ai.Message, quit bool) {
++	dimln(ask.Label(req))
++	askCtx, cancel := context.WithCancel(ctx)
++	defer cancel()
++	type result struct {
++		exchange []*ai.Message
++		err      error
++	}
++	done := make(chan result, 1)
++	go func() {
++		exchange, err := a.pool.Run(askCtx, req, messages, a.cwd, &lineSink{cwd: a.cwd})
++		done <- result{exchange, err}
++	}()
++
++	for {
++		select {
++		case res := <-done:
++			return a.settle(askCtx, req, messages, res.exchange, res.err), quit
++		case <-a.sigCh:
++			fmt.Println("\n🛑 Interrupted (Ctrl+C).")
++			cancel()
++		case input := <-a.inputCh:
++			switch {
++			case command.Is(input, command.Abort):
++				fmt.Println("🛑 Aborting…")
++				cancel()
++			case command.Is(input, command.Quit):
++				quit = true
++				cancel()
++			default:
++				fmt.Printf("⚠️ %s is still answering — /abort to stop it.\n", req.Agent)
++			}
++		}
++	}
++}
++
++// settle reports how the ask ended and keeps the exchange when there is one.
++func (a asking) settle(askCtx context.Context, req command.AskRequest, messages, exchange []*ai.Message, err error) []*ai.Message {
++	fmt.Println()
++	switch {
++	case err != nil && askCtx.Err() != nil:
++		fmt.Println("🛑 Ask aborted — nothing kept.")
++		return messages
++	case err != nil:
++		fmt.Println("[error: " + err.Error() + "]")
++		record(a.keep, messages, store.NewFailure(messages, store.FailAsk, err.Error(), err))
++		return messages
++	}
++	messages = append(messages, exchange...)
++	save(a.keep, messages)
++	dimln(fmt.Sprintf("🗣 %s's answer is in the history for your next question", req.Agent))
++	return messages
++}
++
++// lineSink shows the other agent in the line mode: its answer as it streams,
++// its thoughts dimmed, one line per tool call. Like the line mode's own
++// tools, its calls are not gated: this front end has no permission prompt.
++type lineSink struct {
++	cwd string
++}
++
++var _ ui.Sink = (*lineSink)(nil)
++
++func (s *lineSink) Text(text string) {
++	spinner.Stop()
++	fmt.Print(text)
++}
++
++func (s *lineSink) Reasoning(text string) {
++	if !config.Cfg.ShowThinking {
++		return
++	}
++	spinner.Stop()
++	if spinner.Styled() {
++		fmt.Printf("\033[2;3m%s\033[0m", text)
++		return
++	}
++	fmt.Print(text)
++}
++
++func (s *lineSink) ToolStart(ev ui.ToolEvent) {
++	spinner.Stop()
++	fmt.Printf("\n🛠️  %s\n", ev.Title)
++}
++
++func (s *lineSink) ToolEnd(_ string, res ui.ToolResult) {
++	if res.Failed {
++		dimln("   │ failed")
++	}
++}
++
++func (s *lineSink) Allow(context.Context, ui.ToolEvent) bool { return true }
++func (s *lineSink) ToolRunning(string)                       {}
++func (s *lineSink) WorkDir() string                          { return s.cwd }

added internal/agent/ask_test.go +83 -0 Viewed

new file mode 100644
@@ -0,0 +1,83 @@
	1	+package agent
	2	+
	3	+import (
	4	+ "context"
	5	+ "errors"
	6	+ "os"
	7	+ "testing"
	8	+
	9	+ "mm/internal/command"
	10	+ "mm/internal/config"
	11	+ "mm/internal/store"
	12	+ "mm/internal/ui"
	13	+
	14	+ "github.com/firebase/genkit/go/ai"
	15	+)
	16	+
	17	+// fakeRunner answers, fails, or waits to be cancelled.
	18	+type fakeRunner struct {
	19	+ answer string
	20	+ fail error
	21	+ block bool
	22	+}
	23	+
	24	+func (f fakeRunner) Run(ctx context.Context, req command.AskRequest, _ []ai.Message, _ string, sink ui.Sink) ([]ai.Message, error) {
	25	+ if f.block {
	26	+ <-ctx.Done()
	27	+ return nil, ctx.Err()
	28	+ }
	29	+ if f.fail != nil {
	30	+ return nil, f.fail
	31	+ }
	32	+ sink.Text(f.answer)
	33	+ return []*ai.Message{ai.NewUserTextMessage("[asked " + req.Agent + "]"), ai.NewModelTextMessage(f.answer)}, nil
	34	+}
	35	+
	36	+// lineAsk builds an asking with fresh channels and no session file.
	37	+func lineAsk(t *testing.T, r fakeRunner) (asking, chan os.Signal, chan string) {
	38	+ t.Helper()
	39	+ saved := config.Cfg
	40	+ t.Cleanup(func() { config.Cfg = saved })
	41	+ config.Cfg.Sessions.Enabled = false
	42	+ config.Cfg.MMConfigs = map[string]string{"bob": "/nowhere/agent.bob.yaml"}
	43	+ sig, in := make(chan os.Signal, 1), make(chan string, 1)
	44	+ return asking{pool: r, cwd: t.TempDir(), keep: store.NewKeeper(t.TempDir()), sigCh: sig, inputCh: in}, sig, in
	45	+}
	46	+
	47	+var askHistory = []*ai.Message{ai.NewSystemTextMessage("sys")}
	48	+
	49	+// An answered ask appends the exchange.
	50	+func TestLineAskKeepsTheExchange(t *testing.T) {
	51	+ a, _, _ := lineAsk(t, fakeRunner{answer: "four"})
	52	+ got, quit := a.run(context.Background(), command.AskRequest{Agent: "bob", Prompt: "2+2?"}, askHistory)
	53	+ if quit \|\| len(got) != 3 \|\| got[2].Text() != "four" {
	54	+ t.Errorf("run = %d messages, quit %v", len(got), quit)
	55	+ }
	56	+}
	57	+
	58	+// Ctrl+C and /abort stop the ask and keep nothing; /quit also says to leave;
	59	+// a failure keeps nothing.
	60	+func TestLineAskStops(t *testing.T) {
	61	+ a, sig, _ := lineAsk(t, fakeRunner{block: true})
	62	+ sig <- os.Interrupt
	63	+ if got, quit := a.run(context.Background(), command.AskRequest{Agent: "bob", Prompt: "q"}, askHistory); quit \|\| len(got) != 1 {
	64	+ t.Errorf("Ctrl+C: %d messages, quit %v", len(got), quit)
	65	+ }
	66	+
	67	+ a, _, in := lineAsk(t, fakeRunner{block: true})
	68	+ in <- "/abort"
	69	+ if got, quit := a.run(context.Background(), command.AskRequest{Agent: "bob", Prompt: "q"}, askHistory); quit \|\| len(got) != 1 {
	70	+ t.Errorf("/abort: %d messages, quit %v", len(got), quit)
	71	+ }
	72	+
	73	+ a, _, in = lineAsk(t, fakeRunner{block: true})
	74	+ in <- "/quit"
	75	+ if got, quit := a.run(context.Background(), command.AskRequest{Agent: "bob", Prompt: "q"}, askHistory); !quit \|\| len(got) != 1 {
	76	+ t.Errorf("/quit: %d messages, quit %v", len(got), quit)
	77	+ }
	78	+
	79	+ a, _, _ = lineAsk(t, fakeRunner{fail: errors.New("bob did not start")})
	80	+ if got, _ := a.run(context.Background(), command.AskRequest{Agent: "bob", Prompt: "q"}, askHistory); len(got) != 1 {
	81	+ t.Errorf("failure: %d messages", len(got))
	82	+ }
	83	+}

++package agent
++
++import (
++	"context"
++	"errors"
++	"os"
++	"testing"
++
++	"mm/internal/command"
++	"mm/internal/config"
++	"mm/internal/store"
++	"mm/internal/ui"
++
++	"github.com/firebase/genkit/go/ai"
++)
++
++// fakeRunner answers, fails, or waits to be cancelled.
++type fakeRunner struct {
++	answer string
++	fail   error
++	block  bool
++}
++
++func (f fakeRunner) Run(ctx context.Context, req command.AskRequest, _ []*ai.Message, _ string, sink ui.Sink) ([]*ai.Message, error) {
++	if f.block {
++		<-ctx.Done()
++		return nil, ctx.Err()
++	}
++	if f.fail != nil {
++		return nil, f.fail
++	}
++	sink.Text(f.answer)
++	return []*ai.Message{ai.NewUserTextMessage("[asked " + req.Agent + "]"), ai.NewModelTextMessage(f.answer)}, nil
++}
++
++// lineAsk builds an asking with fresh channels and no session file.
++func lineAsk(t *testing.T, r fakeRunner) (asking, chan os.Signal, chan string) {
++	t.Helper()
++	saved := config.Cfg
++	t.Cleanup(func() { config.Cfg = saved })
++	config.Cfg.Sessions.Enabled = false
++	config.Cfg.MMConfigs = map[string]string{"bob": "/nowhere/agent.bob.yaml"}
++	sig, in := make(chan os.Signal, 1), make(chan string, 1)
++	return asking{pool: r, cwd: t.TempDir(), keep: store.NewKeeper(t.TempDir()), sigCh: sig, inputCh: in}, sig, in
++}
++
++var askHistory = []*ai.Message{ai.NewSystemTextMessage("sys")}
++
++// An answered ask appends the exchange.
++func TestLineAskKeepsTheExchange(t *testing.T) {
++	a, _, _ := lineAsk(t, fakeRunner{answer: "four"})
++	got, quit := a.run(context.Background(), command.AskRequest{Agent: "bob", Prompt: "2+2?"}, askHistory)
++	if quit || len(got) != 3 || got[2].Text() != "four" {
++		t.Errorf("run = %d messages, quit %v", len(got), quit)
++	}
++}
++
++// Ctrl+C and /abort stop the ask and keep nothing; /quit also says to leave;
++// a failure keeps nothing.
++func TestLineAskStops(t *testing.T) {
++	a, sig, _ := lineAsk(t, fakeRunner{block: true})
++	sig <- os.Interrupt
++	if got, quit := a.run(context.Background(), command.AskRequest{Agent: "bob", Prompt: "q"}, askHistory); quit || len(got) != 1 {
++		t.Errorf("Ctrl+C: %d messages, quit %v", len(got), quit)
++	}
++
++	a, _, in := lineAsk(t, fakeRunner{block: true})
++	in <- "/abort"
++	if got, quit := a.run(context.Background(), command.AskRequest{Agent: "bob", Prompt: "q"}, askHistory); quit || len(got) != 1 {
++		t.Errorf("/abort: %d messages, quit %v", len(got), quit)
++	}
++
++	a, _, in = lineAsk(t, fakeRunner{block: true})
++	in <- "/quit"
++	if got, quit := a.run(context.Background(), command.AskRequest{Agent: "bob", Prompt: "q"}, askHistory); !quit || len(got) != 1 {
++		t.Errorf("/quit: %d messages, quit %v", len(got), quit)
++	}
++
++	a, _, _ = lineAsk(t, fakeRunner{fail: errors.New("bob did not start")})
++	if got, _ := a.run(context.Background(), command.AskRequest{Agent: "bob", Prompt: "q"}, askHistory); len(got) != 1 {
++		t.Errorf("failure: %d messages", len(got))
++	}
++}

added internal/ask/client.go +235 -0 Viewed

new file mode 100644
@@ -0,0 +1,235 @@
	1	+package ask
	2	+
	3	+import (
	4	+ "context"
	5	+ "strings"
	6	+ "sync"
	7	+
	8	+ "mm/internal/ui"
	9	+
	10	+ sdk "github.com/coder/acp-go-sdk"
	11	+)
	12	+
	13	+// client is the ACP client side of one child agent: it receives what the
	14	+// child streams and turns it into the asking front end's own events, through
	15	+// the ui.Sink that front end installed — bob's text where the answers go, his
	16	+// tool calls as tool calls, his permission requests as the asker's own
	17	+// permission prompt (none in the line mode, the inline prompt in the window,
	18	+// the editor's dialog in ACP).
	19	+//
	20	+// mm's ACP server only ever calls SessionUpdate and RequestPermission; the
	21	+// file-system and terminal methods are refused, and the capabilities sent in
	22	+// initialize say they are absent.
	23	+type client struct {
	24	+ mu sync.Mutex
	25	+ agent string // "bob": prefixes the tool calls shown
	26	+ sid sdk.SessionId // the session of the ask in progress; others are stale
	27	+ sink ui.Sink // where the events go; nil between two asks
	28	+ answer strings.Builder
	29	+ ids map[sdk.ToolCallId]string // the child's call id -> the id shown
	30	+ // started: the calls already shown. The SDK dispatches the child's
	31	+ // notifications and requests concurrently, so a permission request can
	32	+ // overtake the tool_call that announced it (seen in the tests); whichever
	33	+ // comes first shows the call, the other does not show it twice.
	34	+ started map[sdk.ToolCallId]bool
	35	+}
	36	+
	37	+var _ sdk.Client = (*client)(nil)
	38	+
	39	+// begin points the client at one ask: its session and its sink.
	40	+func (c *client) begin(agent string, sid sdk.SessionId, sink ui.Sink) {
	41	+ c.mu.Lock()
	42	+ defer c.mu.Unlock()
	43	+ c.agent, c.sid, c.sink = agent, sid, sink
	44	+ c.answer.Reset()
	45	+ c.ids = map[sdk.ToolCallId]string{}
	46	+ c.started = map[sdk.ToolCallId]bool{}
	47	+}
	48	+
	49	+// end detaches the sink and returns the answer as streamed.
	50	+func (c *client) end() string {
	51	+ c.mu.Lock()
	52	+ defer c.mu.Unlock()
	53	+ c.sink, c.sid = nil, ""
	54	+ return c.answer.String()
	55	+}
	56	+
	57	+// current returns the sink when sid is the ask in progress, nil otherwise —
	58	+// a cancelled ask's child may still be talking about its old session.
	59	+func (c *client) current(sid sdk.SessionId) ui.Sink {
	60	+ c.mu.Lock()
	61	+ defer c.mu.Unlock()
	62	+ if sid != c.sid {
	63	+ return nil
	64	+ }
	65	+ return c.sink
	66	+}
	67	+
	68	+// SessionUpdate routes one streamed update to the sink.
	69	+func (c *client) SessionUpdate(_ context.Context, n sdk.SessionNotification) error {
	70	+ sink := c.current(n.SessionId)
	71	+ if sink == nil {
	72	+ return nil
	73	+ }
	74	+ u := n.Update
	75	+ switch {
	76	+ case u.AgentMessageChunk != nil && u.AgentMessageChunk.Content.Text != nil:
	77	+ text := u.AgentMessageChunk.Content.Text.Text
	78	+ c.mu.Lock()
	79	+ c.answer.WriteString(text)
	80	+ c.mu.Unlock()
	81	+ sink.Text(text)
	82	+ case u.AgentThoughtChunk != nil && u.AgentThoughtChunk.Content.Text != nil:
	83	+ sink.Reasoning(u.AgentThoughtChunk.Content.Text.Text) // shown as thoughts, never kept
	84	+ case u.ToolCall != nil:
	85	+ c.toolStart(sink, u.ToolCall)
	86	+ case u.ToolCallUpdate != nil:
	87	+ c.toolUpdate(sink, u.ToolCallUpdate)
	88	+ }
	89	+ return nil
	90	+}
	91	+
	92	+// toolStart shows one of the child's tool calls as the asker's own, named
	93	+// after the child: "bob · bash: ls -1".
	94	+func (c client) toolStart(sink ui.Sink, call sdk.SessionUpdateToolCall) {
	95	+ ev := c.event(call.ToolCallId, call.Title, string(call.Kind), call.RawInput)
	96	+ if c.firstSight(call.ToolCallId) {
	97	+ sink.ToolStart(ev)
	98	+ }
	99	+ if call.Status == sdk.ToolCallStatusInProgress {
	100	+ sink.ToolRunning(ev.ID)
	101	+ }
	102	+}
	103	+
	104	+// firstSight records a call as shown and reports whether it was not yet.
	105	+func (c *client) firstSight(id sdk.ToolCallId) bool {
	106	+ c.mu.Lock()
	107	+ defer c.mu.Unlock()
	108	+ if c.started == nil \|\| c.started[id] {
	109	+ return false
	110	+ }
	111	+ c.started[id] = true
	112	+ return true
	113	+}
	114	+
	115	+// toolUpdate follows a call's status: running, then done or failed.
	116	+func (c client) toolUpdate(sink ui.Sink, up sdk.SessionToolCallUpdate) {
	117	+ if up.Status == nil {
	118	+ return
	119	+ }
	120	+ id := c.shownID(up.ToolCallId)
	121	+ switch *up.Status {
	122	+ case sdk.ToolCallStatusInProgress:
	123	+ sink.ToolRunning(id)
	124	+ case sdk.ToolCallStatusCompleted, sdk.ToolCallStatusFailed:
	125	+ sink.ToolEnd(id, ui.ToolResult{Output: contentText(up.Content), Failed: *up.Status == sdk.ToolCallStatusFailed})
	126	+ }
	127	+}
	128	+
	129	+// RequestPermission asks the asker's user, through the asker's front end,
	130	+// whether the child may run a call; the answer is one of the options the
	131	+// child offered (allow once / reject once), or "cancelled" when it offered
	132	+// none that fits.
	133	+func (c *client) RequestPermission(ctx context.Context, p sdk.RequestPermissionRequest) (sdk.RequestPermissionResponse, error) {
	134	+ sink := c.current(p.SessionId)
	135	+ title := ""
	136	+ if p.ToolCall.Title != nil {
	137	+ title = *p.ToolCall.Title
	138	+ }
	139	+ // A stale request — its ask was cancelled — is refused: nobody is there
	140	+ // to answer it.
	141	+ allowed := false
	142	+ if sink != nil {
	143	+ ev := c.event(p.ToolCall.ToolCallId, title, "", p.ToolCall.RawInput)
	144	+ if c.firstSight(p.ToolCall.ToolCallId) {
	145	+ sink.ToolStart(ev) // the request overtook the tool_call: show it first
	146	+ }
	147	+ allowed = sink.Allow(ctx, ev)
	148	+ }
	149	+ want := []sdk.PermissionOptionKind{sdk.PermissionOptionKindRejectOnce, sdk.PermissionOptionKindRejectAlways}
	150	+ if allowed {
	151	+ want = []sdk.PermissionOptionKind{sdk.PermissionOptionKindAllowOnce, sdk.PermissionOptionKindAllowAlways}
	152	+ }
	153	+ for _, kind := range want {
	154	+ for _, o := range p.Options {
	155	+ if o.Kind == kind {
	156	+ return sdk.RequestPermissionResponse{Outcome: sdk.RequestPermissionOutcome{Selected: &sdk.RequestPermissionOutcomeSelected{OptionId: o.OptionId}}}, nil
	157	+ }
	158	+ }
	159	+ }
	160	+ return sdk.RequestPermissionResponse{Outcome: sdk.RequestPermissionOutcome{Cancelled: &sdk.RequestPermissionOutcomeCancelled{}}}, nil
	161	+}
	162	+
	163	+// event builds the ui.ToolEvent the asker shows for one of the child's calls.
	164	+// Tool is "<agent>/<tool>" so that "allow for this session" in the window
	165	+// grants bob's bash, not the asker's.
	166	+func (c *client) event(id sdk.ToolCallId, title, kind string, raw any) ui.ToolEvent {
	167	+ tool, _, _ := strings.Cut(title, ":")
	168	+ input, _ := raw.(map[string]any)
	169	+ c.mu.Lock()
	170	+ agent := c.agent
	171	+ c.mu.Unlock()
	172	+ return ui.ToolEvent{
	173	+ ID: c.shownID(id),
	174	+ Tool: agent + "/" + strings.TrimSpace(tool),
	175	+ Title: agent + " · " + title,
	176	+ Kind: kind,
	177	+ Input: input,
	178	+ }
	179	+}
	180	+
	181	+// shownID maps the child's call id to one unique in this process: two
	182	+// children both number their calls call_1, call_2…
	183	+func (c *client) shownID(id sdk.ToolCallId) string {
	184	+ c.mu.Lock()
	185	+ defer c.mu.Unlock()
	186	+ if shown, ok := c.ids[id]; ok {
	187	+ return shown
	188	+ }
	189	+ shown := ui.NextCallID()
	190	+ if c.ids != nil {
	191	+ c.ids[id] = shown
	192	+ }
	193	+ return shown
	194	+}
	195	+
	196	+// contentText joins the text blocks of a tool call's content.
	197	+func contentText(content []sdk.ToolCallContent) string {
	198	+ var b strings.Builder
	199	+ for _, item := range content {
	200	+ if item.Content != nil && item.Content.Content.Text != nil {
	201	+ b.WriteString(item.Content.Content.Text.Text)
	202	+ }
	203	+ }
	204	+ return b.String()
	205	+}
	206	+
	207	+// --- what mm's ACP server never asks for, refused as absent -----------------
	208	+
	209	+func (c *client) ReadTextFile(context.Context, sdk.ReadTextFileRequest) (sdk.ReadTextFileResponse, error) {
	210	+ return sdk.ReadTextFileResponse{}, sdk.NewMethodNotFound(sdk.ClientMethodFsReadTextFile)
	211	+}
	212	+
	213	+func (c *client) WriteTextFile(context.Context, sdk.WriteTextFileRequest) (sdk.WriteTextFileResponse, error) {
	214	+ return sdk.WriteTextFileResponse{}, sdk.NewMethodNotFound(sdk.ClientMethodFsWriteTextFile)
	215	+}
	216	+
	217	+func (c *client) CreateTerminal(context.Context, sdk.CreateTerminalRequest) (sdk.CreateTerminalResponse, error) {
	218	+ return sdk.CreateTerminalResponse{}, sdk.NewMethodNotFound(sdk.ClientMethodTerminalCreate)
	219	+}
	220	+
	221	+func (c *client) KillTerminal(context.Context, sdk.KillTerminalRequest) (sdk.KillTerminalResponse, error) {
	222	+ return sdk.KillTerminalResponse{}, sdk.NewMethodNotFound(sdk.ClientMethodTerminalKill)
	223	+}
	224	+
	225	+func (c *client) TerminalOutput(context.Context, sdk.TerminalOutputRequest) (sdk.TerminalOutputResponse, error) {
	226	+ return sdk.TerminalOutputResponse{}, sdk.NewMethodNotFound(sdk.ClientMethodTerminalOutput)
	227	+}
	228	+
	229	+func (c *client) ReleaseTerminal(context.Context, sdk.ReleaseTerminalRequest) (sdk.ReleaseTerminalResponse, error) {
	230	+ return sdk.ReleaseTerminalResponse{}, sdk.NewMethodNotFound(sdk.ClientMethodTerminalRelease)
	231	+}
	232	+
	233	+func (c *client) WaitForTerminalExit(context.Context, sdk.WaitForTerminalExitRequest) (sdk.WaitForTerminalExitResponse, error) {
	234	+ return sdk.WaitForTerminalExitResponse{}, sdk.NewMethodNotFound(sdk.ClientMethodTerminalWaitForExit)
	235	+}

++package ask
++
++import (
++	"context"
++	"strings"
++	"sync"
++
++	"mm/internal/ui"
++
++	sdk "github.com/coder/acp-go-sdk"
++)
++
++// client is the ACP client side of one child agent: it receives what the
++// child streams and turns it into the asking front end's own events, through
++// the ui.Sink that front end installed — bob's text where the answers go, his
++// tool calls as tool calls, his permission requests as the asker's own
++// permission prompt (none in the line mode, the inline prompt in the window,
++// the editor's dialog in ACP).
++//
++// mm's ACP server only ever calls SessionUpdate and RequestPermission; the
++// file-system and terminal methods are refused, and the capabilities sent in
++// initialize say they are absent.
++type client struct {
++	mu     sync.Mutex
++	agent  string        // "bob": prefixes the tool calls shown
++	sid    sdk.SessionId // the session of the ask in progress; others are stale
++	sink   ui.Sink       // where the events go; nil between two asks
++	answer strings.Builder
++	ids    map[sdk.ToolCallId]string // the child's call id -> the id shown
++	// started: the calls already shown. The SDK dispatches the child's
++	// notifications and requests concurrently, so a permission request can
++	// overtake the tool_call that announced it (seen in the tests); whichever
++	// comes first shows the call, the other does not show it twice.
++	started map[sdk.ToolCallId]bool
++}
++
++var _ sdk.Client = (*client)(nil)
++
++// begin points the client at one ask: its session and its sink.
++func (c *client) begin(agent string, sid sdk.SessionId, sink ui.Sink) {
++	c.mu.Lock()
++	defer c.mu.Unlock()
++	c.agent, c.sid, c.sink = agent, sid, sink
++	c.answer.Reset()
++	c.ids = map[sdk.ToolCallId]string{}
++	c.started = map[sdk.ToolCallId]bool{}
++}
++
++// end detaches the sink and returns the answer as streamed.
++func (c *client) end() string {
++	c.mu.Lock()
++	defer c.mu.Unlock()
++	c.sink, c.sid = nil, ""
++	return c.answer.String()
++}
++
++// current returns the sink when sid is the ask in progress, nil otherwise —
++// a cancelled ask's child may still be talking about its old session.
++func (c *client) current(sid sdk.SessionId) ui.Sink {
++	c.mu.Lock()
++	defer c.mu.Unlock()
++	if sid != c.sid {
++		return nil
++	}
++	return c.sink
++}
++
++// SessionUpdate routes one streamed update to the sink.
++func (c *client) SessionUpdate(_ context.Context, n sdk.SessionNotification) error {
++	sink := c.current(n.SessionId)
++	if sink == nil {
++		return nil
++	}
++	u := n.Update
++	switch {
++	case u.AgentMessageChunk != nil && u.AgentMessageChunk.Content.Text != nil:
++		text := u.AgentMessageChunk.Content.Text.Text
++		c.mu.Lock()
++		c.answer.WriteString(text)
++		c.mu.Unlock()
++		sink.Text(text)
++	case u.AgentThoughtChunk != nil && u.AgentThoughtChunk.Content.Text != nil:
++		sink.Reasoning(u.AgentThoughtChunk.Content.Text.Text) // shown as thoughts, never kept
++	case u.ToolCall != nil:
++		c.toolStart(sink, u.ToolCall)
++	case u.ToolCallUpdate != nil:
++		c.toolUpdate(sink, u.ToolCallUpdate)
++	}
++	return nil
++}
++
++// toolStart shows one of the child's tool calls as the asker's own, named
++// after the child: "bob · bash: ls -1".
++func (c *client) toolStart(sink ui.Sink, call *sdk.SessionUpdateToolCall) {
++	ev := c.event(call.ToolCallId, call.Title, string(call.Kind), call.RawInput)
++	if c.firstSight(call.ToolCallId) {
++		sink.ToolStart(ev)
++	}
++	if call.Status == sdk.ToolCallStatusInProgress {
++		sink.ToolRunning(ev.ID)
++	}
++}
++
++// firstSight records a call as shown and reports whether it was not yet.
++func (c *client) firstSight(id sdk.ToolCallId) bool {
++	c.mu.Lock()
++	defer c.mu.Unlock()
++	if c.started == nil || c.started[id] {
++		return false
++	}
++	c.started[id] = true
++	return true
++}
++
++// toolUpdate follows a call's status: running, then done or failed.
++func (c *client) toolUpdate(sink ui.Sink, up *sdk.SessionToolCallUpdate) {
++	if up.Status == nil {
++		return
++	}
++	id := c.shownID(up.ToolCallId)
++	switch *up.Status {
++	case sdk.ToolCallStatusInProgress:
++		sink.ToolRunning(id)
++	case sdk.ToolCallStatusCompleted, sdk.ToolCallStatusFailed:
++		sink.ToolEnd(id, ui.ToolResult{Output: contentText(up.Content), Failed: *up.Status == sdk.ToolCallStatusFailed})
++	}
++}
++
++// RequestPermission asks the asker's user, through the asker's front end,
++// whether the child may run a call; the answer is one of the options the
++// child offered (allow once / reject once), or "cancelled" when it offered
++// none that fits.
++func (c *client) RequestPermission(ctx context.Context, p sdk.RequestPermissionRequest) (sdk.RequestPermissionResponse, error) {
++	sink := c.current(p.SessionId)
++	title := ""
++	if p.ToolCall.Title != nil {
++		title = *p.ToolCall.Title
++	}
++	// A stale request — its ask was cancelled — is refused: nobody is there
++	// to answer it.
++	allowed := false
++	if sink != nil {
++		ev := c.event(p.ToolCall.ToolCallId, title, "", p.ToolCall.RawInput)
++		if c.firstSight(p.ToolCall.ToolCallId) {
++			sink.ToolStart(ev) // the request overtook the tool_call: show it first
++		}
++		allowed = sink.Allow(ctx, ev)
++	}
++	want := []sdk.PermissionOptionKind{sdk.PermissionOptionKindRejectOnce, sdk.PermissionOptionKindRejectAlways}
++	if allowed {
++		want = []sdk.PermissionOptionKind{sdk.PermissionOptionKindAllowOnce, sdk.PermissionOptionKindAllowAlways}
++	}
++	for _, kind := range want {
++		for _, o := range p.Options {
++			if o.Kind == kind {
++				return sdk.RequestPermissionResponse{Outcome: sdk.RequestPermissionOutcome{Selected: &sdk.RequestPermissionOutcomeSelected{OptionId: o.OptionId}}}, nil
++			}
++		}
++	}
++	return sdk.RequestPermissionResponse{Outcome: sdk.RequestPermissionOutcome{Cancelled: &sdk.RequestPermissionOutcomeCancelled{}}}, nil
++}
++
++// event builds the ui.ToolEvent the asker shows for one of the child's calls.
++// Tool is "<agent>/<tool>" so that "allow for this session" in the window
++// grants bob's bash, not the asker's.
++func (c *client) event(id sdk.ToolCallId, title, kind string, raw any) ui.ToolEvent {
++	tool, _, _ := strings.Cut(title, ":")
++	input, _ := raw.(map[string]any)
++	c.mu.Lock()
++	agent := c.agent
++	c.mu.Unlock()
++	return ui.ToolEvent{
++		ID:    c.shownID(id),
++		Tool:  agent + "/" + strings.TrimSpace(tool),
++		Title: agent + " · " + title,
++		Kind:  kind,
++		Input: input,
++	}
++}
++
++// shownID maps the child's call id to one unique in this process: two
++// children both number their calls call_1, call_2…
++func (c *client) shownID(id sdk.ToolCallId) string {
++	c.mu.Lock()
++	defer c.mu.Unlock()
++	if shown, ok := c.ids[id]; ok {
++		return shown
++	}
++	shown := ui.NextCallID()
++	if c.ids != nil {
++		c.ids[id] = shown
++	}
++	return shown
++}
++
++// contentText joins the text blocks of a tool call's content.
++func contentText(content []sdk.ToolCallContent) string {
++	var b strings.Builder
++	for _, item := range content {
++		if item.Content != nil && item.Content.Content.Text != nil {
++			b.WriteString(item.Content.Content.Text.Text)
++		}
++	}
++	return b.String()
++}
++
++// --- what mm's ACP server never asks for, refused as absent -----------------
++
++func (c *client) ReadTextFile(context.Context, sdk.ReadTextFileRequest) (sdk.ReadTextFileResponse, error) {
++	return sdk.ReadTextFileResponse{}, sdk.NewMethodNotFound(sdk.ClientMethodFsReadTextFile)
++}
++
++func (c *client) WriteTextFile(context.Context, sdk.WriteTextFileRequest) (sdk.WriteTextFileResponse, error) {
++	return sdk.WriteTextFileResponse{}, sdk.NewMethodNotFound(sdk.ClientMethodFsWriteTextFile)
++}
++
++func (c *client) CreateTerminal(context.Context, sdk.CreateTerminalRequest) (sdk.CreateTerminalResponse, error) {
++	return sdk.CreateTerminalResponse{}, sdk.NewMethodNotFound(sdk.ClientMethodTerminalCreate)
++}
++
++func (c *client) KillTerminal(context.Context, sdk.KillTerminalRequest) (sdk.KillTerminalResponse, error) {
++	return sdk.KillTerminalResponse{}, sdk.NewMethodNotFound(sdk.ClientMethodTerminalKill)
++}
++
++func (c *client) TerminalOutput(context.Context, sdk.TerminalOutputRequest) (sdk.TerminalOutputResponse, error) {
++	return sdk.TerminalOutputResponse{}, sdk.NewMethodNotFound(sdk.ClientMethodTerminalOutput)
++}
++
++func (c *client) ReleaseTerminal(context.Context, sdk.ReleaseTerminalRequest) (sdk.ReleaseTerminalResponse, error) {
++	return sdk.ReleaseTerminalResponse{}, sdk.NewMethodNotFound(sdk.ClientMethodTerminalRelease)
++}
++
++func (c *client) WaitForTerminalExit(context.Context, sdk.WaitForTerminalExitRequest) (sdk.WaitForTerminalExitResponse, error) {
++	return sdk.WaitForTerminalExitResponse{}, sdk.NewMethodNotFound(sdk.ClientMethodTerminalWaitForExit)
++}

added internal/ask/context.go +101 -0 Viewed

new file mode 100644
@@ -0,0 +1,101 @@
	1	+package ask
	2	+
	3	+import (
	4	+ "fmt"
	5	+ "strings"
	6	+
	7	+ "mm/internal/compact"
	8	+
	9	+ "github.com/firebase/genkit/go/ai"
	10	+)
	11	+
	12	+// Transcript renders the conversation for /ask-with-context: what was asked
	13	+// and answered, and which tools ran — one line each, with their arguments
	14	+// shortened — but never the tools' outputs, the system prompt or any
	15	+// reasoning. The other agent needs to know what happened, not to re-read
	16	+// every file; and its window is not ours.
	17	+//
	18	+// The newest turns are kept first: when the whole history does not fit in
	19	+// maxChars, the oldest lines go, and a marker says so.
	20	+//
	21	+// text := ask.Transcript(messages, config.Cfg.MaxOutput)
	22	+func Transcript(messages []*ai.Message, maxChars int) string {
	23	+ var lines []string
	24	+ for _, m := range messages {
	25	+ lines = append(lines, transcriptLines(m)...)
	26	+ }
	27	+ return keepNewest(lines, maxChars)
	28	+}
	29	+
	30	+// transcriptLines is one message, as the transcript shows it.
	31	+func transcriptLines(m *ai.Message) []string {
	32	+ if m == nil {
	33	+ return nil
	34	+ }
	35	+ switch {
	36	+ case compact.IsSummary(m):
	37	+ return []string{"[Summary of the earlier conversation]\n" + strings.TrimSpace(m.Text())}
	38	+ case m.Role == ai.RoleUser:
	39	+ if text := strings.TrimSpace(m.Text()); text != "" {
	40	+ return []string{"User: " + text}
	41	+ }
	42	+ case m.Role == ai.RoleModel:
	43	+ return modelLines(m)
	44	+ }
	45	+ return nil // system prompt, tool outputs
	46	+}
	47	+
	48	+// modelLines is an assistant message: its text, then one line per tool call.
	49	+func modelLines(m *ai.Message) []string {
	50	+ var lines []string
	51	+ if text := strings.TrimSpace(m.Text()); text != "" {
	52	+ lines = append(lines, "Assistant: "+text)
	53	+ }
	54	+ for _, part := range m.Content {
	55	+ if part.IsToolRequest() && part.ToolRequest != nil {
	56	+ lines = append(lines, fmt.Sprintf("Assistant ran %s: %s", part.ToolRequest.Name, shorten(fmt.Sprint(part.ToolRequest.Input), 160)))
	57	+ }
	58	+ }
	59	+ return lines
	60	+}
	61	+
	62	+// keepNewest joins lines, dropping the oldest ones that do not fit.
	63	+func keepNewest(lines []string, maxChars int) string {
	64	+ const omitted = "[… earlier conversation omitted …]"
	65	+ size, first := 0, len(lines)
	66	+ for first > 0 {
	67	+ next := len([]rune(lines[first-1])) + 1
	68	+ if size+next > maxChars-len(omitted)-1 {
	69	+ break
	70	+ }
	71	+ size += next
	72	+ first--
	73	+ }
	74	+ kept := strings.Join(lines[first:], "\n")
	75	+ if first > 0 {
	76	+ return omitted + "\n" + kept
	77	+ }
	78	+ return kept
	79	+}
	80	+
	81	+// shorten cuts s to n characters on one line.
	82	+func shorten(s string, n int) string {
	83	+ s = strings.Join(strings.Fields(s), " ")
	84	+ if r := []rune(s); len(r) > n {
	85	+ return string(r[:n]) + "…"
	86	+ }
	87	+ return s
	88	+}
	89	+
	90	+// Prompt is what the other agent receives: the question alone, or — for
	91	+// /ask-with-context — the transcript first, clearly separated, so it can tell
	92	+// the conversation it is shown from the question it must answer.
	93	+//
	94	+// ask.Prompt("is this safe?", "") // "is this safe?"
	95	+// ask.Prompt("is this safe?", ask.Transcript(…)) // "[Conversation so far…]\n…\n\n[Question]\nis this safe?"
	96	+func Prompt(question, transcript string) string {
	97	+ if transcript == "" {
	98	+ return question
	99	+ }
	100	+ return "[Conversation so far, between the user and another assistant]\n" + transcript + "\n\n[Question]\n" + question
	101	+}

++package ask
++
++import (
++	"fmt"
++	"strings"
++
++	"mm/internal/compact"
++
++	"github.com/firebase/genkit/go/ai"
++)
++
++// Transcript renders the conversation for /ask-with-context: what was asked
++// and answered, and which tools ran — one line each, with their arguments
++// shortened — but never the tools' outputs, the system prompt or any
++// reasoning. The other agent needs to know what happened, not to re-read
++// every file; and its window is not ours.
++//
++// The newest turns are kept first: when the whole history does not fit in
++// maxChars, the oldest lines go, and a marker says so.
++//
++//	text := ask.Transcript(messages, config.Cfg.MaxOutput)
++func Transcript(messages []*ai.Message, maxChars int) string {
++	var lines []string
++	for _, m := range messages {
++		lines = append(lines, transcriptLines(m)...)
++	}
++	return keepNewest(lines, maxChars)
++}
++
++// transcriptLines is one message, as the transcript shows it.
++func transcriptLines(m *ai.Message) []string {
++	if m == nil {
++		return nil
++	}
++	switch {
++	case compact.IsSummary(m):
++		return []string{"[Summary of the earlier conversation]\n" + strings.TrimSpace(m.Text())}
++	case m.Role == ai.RoleUser:
++		if text := strings.TrimSpace(m.Text()); text != "" {
++			return []string{"User: " + text}
++		}
++	case m.Role == ai.RoleModel:
++		return modelLines(m)
++	}
++	return nil // system prompt, tool outputs
++}
++
++// modelLines is an assistant message: its text, then one line per tool call.
++func modelLines(m *ai.Message) []string {
++	var lines []string
++	if text := strings.TrimSpace(m.Text()); text != "" {
++		lines = append(lines, "Assistant: "+text)
++	}
++	for _, part := range m.Content {
++		if part.IsToolRequest() && part.ToolRequest != nil {
++			lines = append(lines, fmt.Sprintf("Assistant ran %s: %s", part.ToolRequest.Name, shorten(fmt.Sprint(part.ToolRequest.Input), 160)))
++		}
++	}
++	return lines
++}
++
++// keepNewest joins lines, dropping the oldest ones that do not fit.
++func keepNewest(lines []string, maxChars int) string {
++	const omitted = "[… earlier conversation omitted …]"
++	size, first := 0, len(lines)
++	for first > 0 {
++		next := len([]rune(lines[first-1])) + 1
++		if size+next > maxChars-len(omitted)-1 {
++			break
++		}
++		size += next
++		first--
++	}
++	kept := strings.Join(lines[first:], "\n")
++	if first > 0 {
++		return omitted + "\n" + kept
++	}
++	return kept
++}
++
++// shorten cuts s to n characters on one line.
++func shorten(s string, n int) string {
++	s = strings.Join(strings.Fields(s), " ")
++	if r := []rune(s); len(r) > n {
++		return string(r[:n]) + "…"
++	}
++	return s
++}
++
++// Prompt is what the other agent receives: the question alone, or — for
++// /ask-with-context — the transcript first, clearly separated, so it can tell
++// the conversation it is shown from the question it must answer.
++//
++//	ask.Prompt("is this safe?", "")                 // "is this safe?"
++//	ask.Prompt("is this safe?", ask.Transcript(…))  // "[Conversation so far…]\n…\n\n[Question]\nis this safe?"
++func Prompt(question, transcript string) string {
++	if transcript == "" {
++		return question
++	}
++	return "[Conversation so far, between the user and another assistant]\n" + transcript + "\n\n[Question]\n" + question
++}

added internal/ask/context_test.go +66 -0 Viewed

new file mode 100644
@@ -0,0 +1,66 @@
	1	+package ask
	2	+
	3	+import (
	4	+ "strings"
	5	+ "testing"
	6	+
	7	+ "github.com/firebase/genkit/go/ai"
	8	+)
	9	+
	10	+// conversation is a short history with a system prompt, a tool call and its
	11	+// output — everything the transcript must keep or drop.
	12	+func conversation() []*ai.Message {
	13	+ call := ai.NewModelMessage(ai.NewTextPart("Let me look."), ai.NewToolRequestPart(&ai.ToolRequest{Name: "bash", Input: map[string]any{"command": "ls -1"}}))
	14	+ out := ai.NewMessage(ai.RoleTool, nil, ai.NewToolResponsePart(&ai.ToolResponse{Name: "bash", Output: "SECRET-OUTPUT"}))
	15	+ return []*ai.Message{
	16	+ ai.NewSystemTextMessage("SYSTEM-PROMPT"),
	17	+ ai.NewUserTextMessage("list the files"),
	18	+ call, out,
	19	+ ai.NewModelTextMessage("There are two files."),
	20	+ }
	21	+}
	22	+
	23	+// The transcript keeps questions, answers and one line per tool call, and
	24	+// never the system prompt or a tool's output.
	25	+func TestTranscriptKeepsWhatHappenedNotTheOutputs(t *testing.T) {
	26	+ got := Transcript(conversation(), 10000)
	27	+ for _, want := range []string{"User: list the files", "Assistant: Let me look.", "Assistant ran bash: map[command:ls -1]", "Assistant: There are two files."} {
	28	+ if !strings.Contains(got, want) {
	29	+ t.Errorf("transcript lacks %q:\n%s", want, got)
	30	+ }
	31	+ }
	32	+ for _, never := range []string{"SYSTEM-PROMPT", "SECRET-OUTPUT"} {
	33	+ if strings.Contains(got, never) {
	34	+ t.Errorf("transcript leaks %q:\n%s", never, got)
	35	+ }
	36	+ }
	37	+}
	38	+
	39	+// Over the cap, the oldest lines go first and a marker says so; the newest
	40	+// line is always there.
	41	+func TestTranscriptKeepsTheNewestTurns(t *testing.T) {
	42	+ var msgs []*ai.Message
	43	+ for i := 0; i < 50; i++ {
	44	+ msgs = append(msgs, ai.NewUserTextMessage(strings.Repeat("old ", 10)), ai.NewModelTextMessage("ok"))
	45	+ }
	46	+ msgs = append(msgs, ai.NewUserTextMessage("the latest question"))
	47	+ got := Transcript(msgs, 300)
	48	+ if len([]rune(got)) > 300 {
	49	+ t.Errorf("transcript is %d chars, cap 300", len([]rune(got)))
	50	+ }
	51	+ if !strings.HasPrefix(got, "[… earlier conversation omitted …]") \|\| !strings.HasSuffix(got, "User: the latest question") {
	52	+ t.Errorf("transcript:\n%s", got)
	53	+ }
	54	+}
	55	+
	56	+// A question alone goes as typed; with a transcript, both are labelled.
	57	+func TestPrompt(t *testing.T) {
	58	+ if got := Prompt("is it safe?", ""); got != "is it safe?" {
	59	+ t.Errorf("Prompt without context = %q", got)
	60	+ }
	61	+ got := Prompt("is it safe?", "User: hi")
	62	+ want := "[Conversation so far, between the user and another assistant]\nUser: hi\n\n[Question]\nis it safe?"
	63	+ if got != want {
	64	+ t.Errorf("Prompt with context = %q, want %q", got, want)
	65	+ }
	66	+}

++package ask
++
++import (
++	"strings"
++	"testing"
++
++	"github.com/firebase/genkit/go/ai"
++)
++
++// conversation is a short history with a system prompt, a tool call and its
++// output — everything the transcript must keep or drop.
++func conversation() []*ai.Message {
++	call := ai.NewModelMessage(ai.NewTextPart("Let me look."), ai.NewToolRequestPart(&ai.ToolRequest{Name: "bash", Input: map[string]any{"command": "ls -1"}}))
++	out := ai.NewMessage(ai.RoleTool, nil, ai.NewToolResponsePart(&ai.ToolResponse{Name: "bash", Output: "SECRET-OUTPUT"}))
++	return []*ai.Message{
++		ai.NewSystemTextMessage("SYSTEM-PROMPT"),
++		ai.NewUserTextMessage("list the files"),
++		call, out,
++		ai.NewModelTextMessage("There are two files."),
++	}
++}
++
++// The transcript keeps questions, answers and one line per tool call, and
++// never the system prompt or a tool's output.
++func TestTranscriptKeepsWhatHappenedNotTheOutputs(t *testing.T) {
++	got := Transcript(conversation(), 10000)
++	for _, want := range []string{"User: list the files", "Assistant: Let me look.", "Assistant ran bash: map[command:ls -1]", "Assistant: There are two files."} {
++		if !strings.Contains(got, want) {
++			t.Errorf("transcript lacks %q:\n%s", want, got)
++		}
++	}
++	for _, never := range []string{"SYSTEM-PROMPT", "SECRET-OUTPUT"} {
++		if strings.Contains(got, never) {
++			t.Errorf("transcript leaks %q:\n%s", never, got)
++		}
++	}
++}
++
++// Over the cap, the oldest lines go first and a marker says so; the newest
++// line is always there.
++func TestTranscriptKeepsTheNewestTurns(t *testing.T) {
++	var msgs []*ai.Message
++	for i := 0; i < 50; i++ {
++		msgs = append(msgs, ai.NewUserTextMessage(strings.Repeat("old ", 10)), ai.NewModelTextMessage("ok"))
++	}
++	msgs = append(msgs, ai.NewUserTextMessage("the latest question"))
++	got := Transcript(msgs, 300)
++	if len([]rune(got)) > 300 {
++		t.Errorf("transcript is %d chars, cap 300", len([]rune(got)))
++	}
++	if !strings.HasPrefix(got, "[… earlier conversation omitted …]") || !strings.HasSuffix(got, "User: the latest question") {
++		t.Errorf("transcript:\n%s", got)
++	}
++}
++
++// A question alone goes as typed; with a transcript, both are labelled.
++func TestPrompt(t *testing.T) {
++	if got := Prompt("is it safe?", ""); got != "is it safe?" {
++		t.Errorf("Prompt without context = %q", got)
++	}
++	got := Prompt("is it safe?", "User: hi")
++	want := "[Conversation so far, between the user and another assistant]\nUser: hi\n\n[Question]\nis it safe?"
++	if got != want {
++		t.Errorf("Prompt with context = %q, want %q", got, want)
++	}
++}

added internal/ask/doc.go +22 -0 Viewed

new file mode 100644
@@ -0,0 +1,22 @@
	1	+// Package ask lets one mm agent put a question to another: `/ask-bob <prompt>`
	2	+// starts (once) a child `mm -acp -config <bob's yaml>`, talks to it as an
	3	+// Agent Client Protocol CLIENT, shows what bob does while he does it, and
	4	+// hands back his answer so the front end can keep it in the history for the
	5	+// next question.
	6	+//
	7	+// Why a child process over ACP, rather than a second engine in this process:
	8	+// the configuration is one global (config.Cfg) read by the engine, the tools
	9	+// and the compaction alike, and bob's config differs in all of them — model,
	10	+// provider, tools, skills, MCP servers. A child keeps each agent whole. Why
	11	+// ACP rather than `mm -p`: it is structured (text, tool calls and permission
	12	+// requests arrive as messages, not as a terminal screen to scrape), it can be
	13	+// cancelled, and mm already speaks it as a server.
	14	+//
	15	+// The pieces:
	16	+//
	17	+// - context.go — the transcript /ask-with-context sends, and the prompt;
	18	+// - exchange.go — the two messages the asking agent keeps in its history;
	19	+// - env.go — the child's environment and the depth limit;
	20	+// - client.go — the ACP client: events to a ui.Sink, permissions;
	21	+// - pool.go — one child per agent, started on first use.
	22	+package ask

++// Package ask lets one mm agent put a question to another: `/ask-bob <prompt>`
++// starts (once) a child `mm -acp -config <bob's yaml>`, talks to it as an
++// Agent Client Protocol CLIENT, shows what bob does while he does it, and
++// hands back his answer so the front end can keep it in the history for the
++// next question.
++//
++// Why a child process over ACP, rather than a second engine in this process:
++// the configuration is one global (config.Cfg) read by the engine, the tools
++// and the compaction alike, and bob's config differs in all of them — model,
++// provider, tools, skills, MCP servers. A child keeps each agent whole. Why
++// ACP rather than `mm -p`: it is structured (text, tool calls and permission
++// requests arrive as messages, not as a terminal screen to scrape), it can be
++// cancelled, and mm already speaks it as a server.
++//
++// The pieces:
++//
++//   - context.go  — the transcript /ask-with-context sends, and the prompt;
++//   - exchange.go — the two messages the asking agent keeps in its history;
++//   - env.go      — the child's environment and the depth limit;
++//   - client.go   — the ACP client: events to a ui.Sink, permissions;
++//   - pool.go     — one child per agent, started on first use.
++package ask

added internal/ask/env.go +55 -0 Viewed

new file mode 100644
@@ -0,0 +1,55 @@
	1	+package ask
	2	+
	3	+import (
	4	+ "fmt"
	5	+ "os"
	6	+ "slices"
	7	+ "strconv"
	8	+ "strings"
	9	+
	10	+ "mm/internal/config"
	11	+)
	12	+
	13	+// DepthEnv carries how deep in a chain of /ask an agent runs: unset for the
	14	+// one the user started, 1 for the agent it asked, and so on.
	15	+const DepthEnv = "MM_ASK_DEPTH"
	16	+
	17	+// MaxDepth: an agent at this depth may not ask another one. /ask is a USER
	18	+// command — no model can call it — so a loop needs a person typing /ask into
	19	+// an agent that was itself asked; the cap keeps even that finite.
	20	+const MaxDepth = 2
	21	+
	22	+// Depth is this process's place in the chain (0 = started by the user).
	23	+func Depth() int {
	24	+ n, _ := strconv.Atoi(os.Getenv(DepthEnv))
	25	+ return max(n, 0)
	26	+}
	27	+
	28	+// CanAsk says whether this agent may ask another one, and why not.
	29	+func CanAsk() error {
	30	+ if d := Depth(); d >= MaxDepth {
	31	+ return fmt.Errorf("this agent was itself asked through %d level(s) of /ask; it cannot ask further (max %d)", d, MaxDepth)
	32	+ }
	33	+ return nil
	34	+}
	35	+
	36	+// overrides are the variables that would make the child ignore its own
	37	+// config file: they are meant for ONE agent, the one the user started.
	38	+var overrides = []string{"AGENT_CONFIG", "AGENT_PROVIDER", "AGENT_MODEL", "AGENT_BASE_URL", config.SessionsEnv, DepthEnv}
	39	+
	40	+// ChildEnv is the environment of an agent started by /ask: the parent's,
	41	+// minus the per-agent overrides, plus no session files (it works in the
	42	+// parent's project, and its sessions would land in the parent's /sessions)
	43	+// and its depth.
	44	+//
	45	+// cmd.Env = ask.ChildEnv(os.Environ(), ask.Depth())
	46	+func ChildEnv(parent []string, depth int) []string {
	47	+ env := make([]string, 0, len(parent)+2)
	48	+ for _, kv := range parent {
	49	+ name, _, _ := strings.Cut(kv, "=")
	50	+ if !slices.Contains(overrides, name) {
	51	+ env = append(env, kv)
	52	+ }
	53	+ }
	54	+ return append(env, config.SessionsEnv+"=off", fmt.Sprintf("%s=%d", DepthEnv, depth+1))
	55	+}

++package ask
++
++import (
++	"fmt"
++	"os"
++	"slices"
++	"strconv"
++	"strings"
++
++	"mm/internal/config"
++)
++
++// DepthEnv carries how deep in a chain of /ask an agent runs: unset for the
++// one the user started, 1 for the agent it asked, and so on.
++const DepthEnv = "MM_ASK_DEPTH"
++
++// MaxDepth: an agent at this depth may not ask another one. /ask is a USER
++// command — no model can call it — so a loop needs a person typing /ask into
++// an agent that was itself asked; the cap keeps even that finite.
++const MaxDepth = 2
++
++// Depth is this process's place in the chain (0 = started by the user).
++func Depth() int {
++	n, _ := strconv.Atoi(os.Getenv(DepthEnv))
++	return max(n, 0)
++}
++
++// CanAsk says whether this agent may ask another one, and why not.
++func CanAsk() error {
++	if d := Depth(); d >= MaxDepth {
++		return fmt.Errorf("this agent was itself asked through %d level(s) of /ask; it cannot ask further (max %d)", d, MaxDepth)
++	}
++	return nil
++}
++
++// overrides are the variables that would make the child ignore its own
++// config file: they are meant for ONE agent, the one the user started.
++var overrides = []string{"AGENT_CONFIG", "AGENT_PROVIDER", "AGENT_MODEL", "AGENT_BASE_URL", config.SessionsEnv, DepthEnv}
++
++// ChildEnv is the environment of an agent started by /ask: the parent's,
++// minus the per-agent overrides, plus no session files (it works in the
++// parent's project, and its sessions would land in the parent's /sessions)
++// and its depth.
++//
++//	cmd.Env = ask.ChildEnv(os.Environ(), ask.Depth())
++func ChildEnv(parent []string, depth int) []string {
++	env := make([]string, 0, len(parent)+2)
++	for _, kv := range parent {
++		name, _, _ := strings.Cut(kv, "=")
++		if !slices.Contains(overrides, name) {
++			env = append(env, kv)
++		}
++	}
++	return append(env, config.SessionsEnv+"=off", fmt.Sprintf("%s=%d", DepthEnv, depth+1))
++}

added internal/ask/env_test.go +42 -0 Viewed

new file mode 100644
@@ -0,0 +1,42 @@
	1	+package ask
	2	+
	3	+import (
	4	+ "slices"
	5	+ "strings"
	6	+ "testing"
	7	+)
	8	+
	9	+// The child keeps the parent's environment but none of the variables that
	10	+// would override ITS config, runs without session files, one level deeper.
	11	+func TestChildEnv(t *testing.T) {
	12	+ parent := []string{"PATH=/bin", "AGENT_MODEL=parent-model", "AGENT_PROVIDER=dmr", "AGENT_BASE_URL=http://x", "AGENT_CONFIG=p.yaml", "AGENT_SESSIONS=on", "MM_ASK_DEPTH=1", "DMR_BASE_URL=http://dmr", "HOME=/h"}
	13	+ got := ChildEnv(parent, 1)
	14	+ for _, want := range []string{"PATH=/bin", "DMR_BASE_URL=http://dmr", "HOME=/h", "AGENT_SESSIONS=off", "MM_ASK_DEPTH=2"} {
	15	+ if !slices.Contains(got, want) {
	16	+ t.Errorf("child env lacks %q: %v", want, got)
	17	+ }
	18	+ }
	19	+ for _, kv := range got {
	20	+ for _, banned := range []string{"AGENT_MODEL=", "AGENT_PROVIDER=", "AGENT_BASE_URL=", "AGENT_CONFIG=", "AGENT_SESSIONS=on", "MM_ASK_DEPTH=1"} {
	21	+ if strings.HasPrefix(kv, banned) {
	22	+ t.Errorf("child env keeps %q", kv)
	23	+ }
	24	+ }
	25	+ }
	26	+}
	27	+
	28	+// The user's agent may ask; an agent asked through MaxDepth levels may not.
	29	+func TestDepthLimit(t *testing.T) {
	30	+ t.Setenv(DepthEnv, "")
	31	+ if Depth() != 0 \|\| CanAsk() != nil {
	32	+ t.Errorf("top level: Depth %d, CanAsk %v", Depth(), CanAsk())
	33	+ }
	34	+ t.Setenv(DepthEnv, "1")
	35	+ if CanAsk() != nil {
	36	+ t.Errorf("depth 1: CanAsk %v, want allowed", CanAsk())
	37	+ }
	38	+ t.Setenv(DepthEnv, "2")
	39	+ if err := CanAsk(); err == nil \|\| !strings.Contains(err.Error(), "cannot ask further") {
	40	+ t.Errorf("depth 2: CanAsk %v, want refused", err)
	41	+ }
	42	+}

++package ask
++
++import (
++	"slices"
++	"strings"
++	"testing"
++)
++
++// The child keeps the parent's environment but none of the variables that
++// would override ITS config, runs without session files, one level deeper.
++func TestChildEnv(t *testing.T) {
++	parent := []string{"PATH=/bin", "AGENT_MODEL=parent-model", "AGENT_PROVIDER=dmr", "AGENT_BASE_URL=http://x", "AGENT_CONFIG=p.yaml", "AGENT_SESSIONS=on", "MM_ASK_DEPTH=1", "DMR_BASE_URL=http://dmr", "HOME=/h"}
++	got := ChildEnv(parent, 1)
++	for _, want := range []string{"PATH=/bin", "DMR_BASE_URL=http://dmr", "HOME=/h", "AGENT_SESSIONS=off", "MM_ASK_DEPTH=2"} {
++		if !slices.Contains(got, want) {
++			t.Errorf("child env lacks %q: %v", want, got)
++		}
++	}
++	for _, kv := range got {
++		for _, banned := range []string{"AGENT_MODEL=", "AGENT_PROVIDER=", "AGENT_BASE_URL=", "AGENT_CONFIG=", "AGENT_SESSIONS=on", "MM_ASK_DEPTH=1"} {
++			if strings.HasPrefix(kv, banned) {
++				t.Errorf("child env keeps %q", kv)
++			}
++		}
++	}
++}
++
++// The user's agent may ask; an agent asked through MaxDepth levels may not.
++func TestDepthLimit(t *testing.T) {
++	t.Setenv(DepthEnv, "")
++	if Depth() != 0 || CanAsk() != nil {
++		t.Errorf("top level: Depth %d, CanAsk %v", Depth(), CanAsk())
++	}
++	t.Setenv(DepthEnv, "1")
++	if CanAsk() != nil {
++		t.Errorf("depth 1: CanAsk %v, want allowed", CanAsk())
++	}
++	t.Setenv(DepthEnv, "2")
++	if err := CanAsk(); err == nil || !strings.Contains(err.Error(), "cannot ask further") {
++		t.Errorf("depth 2: CanAsk %v, want refused", err)
++	}
++}

added internal/ask/exchange.go +63 -0 Viewed

new file mode 100644
@@ -0,0 +1,63 @@
	1	+package ask
	2	+
	3	+import (
	4	+ "fmt"
	5	+ "os"
	6	+
	7	+ "github.com/firebase/genkit/go/ai"
	8	+ "gopkg.in/yaml.v3"
	9	+)
	10	+
	11	+// Exchange is what the asking agent keeps in its history: the question as a
	12	+// user message and the other agent's answer as the assistant's, each marked
	13	+// with who was asked. A PAIR, not one message: the next question is a user
	14	+// message too, and two user messages in a row are refused by chat templates
	15	+// that require the roles to alternate (Gemma's, Mistral's).
	16	+//
	17	+// The answer is capped at maxChars, its beginning and end kept — the same
	18	+// rule as a tool's output, since it lands in the same window.
	19	+//
	20	+// history = append(history, ask.Exchange("bob", "ollama/qwen3.5:2b", false, prompt, answer, config.Cfg.MaxOutput)...)
	21	+func Exchange(agent, label string, withContext bool, question, answer string, maxChars int) []*ai.Message {
	22	+ how := ""
	23	+ if withContext {
	24	+ how = ", with this conversation as context"
	25	+ }
	26	+ asked := fmt.Sprintf("[asked %s (%s)%s: %s]", agent, label, how, question)
	27	+ answered := fmt.Sprintf("[%s's answer]\n%s", agent, clipMiddle(answer, maxChars))
	28	+ return []*ai.Message{ai.NewUserTextMessage(asked), ai.NewModelTextMessage(answered)}
	29	+}
	30	+
	31	+// clipMiddle caps s at max characters, keeping the beginning and the end.
	32	+func clipMiddle(s string, max int) string {
	33	+ r := []rune(s)
	34	+ if max <= 0 \|\| len(r) <= max {
	35	+ return s
	36	+ }
	37	+ tail := min(2000, max/2)
	38	+ head := max - tail
	39	+ return fmt.Sprintf("%s\n\n[... %d characters truncated ...]\n\n%s", string(r[:head]), len(r)-head-tail, string(r[len(r)-tail:]))
	40	+}
	41	+
	42	+// Describe reads "<provider>/<model>" from an agent's config file, for the
	43	+// marker in the history and the line the user sees. Best effort: the child
	44	+// validates its own file, and a value it cannot read here is shown as "?".
	45	+//
	46	+// ask.Describe("demo-ollama/agent.ollama.yaml") // "ollama/qwen3.5:2b"
	47	+func Describe(path string) string {
	48	+ data, err := os.ReadFile(path)
	49	+ if err != nil {
	50	+ return "?"
	51	+ }
	52	+ var head struct {
	53	+ Provider string `yaml:"provider"`
	54	+ Model string `yaml:"model"`
	55	+ }
	56	+ if yaml.Unmarshal(data, &head) != nil \|\| head.Model == "" {
	57	+ return "?"
	58	+ }
	59	+ if head.Provider == "" {
	60	+ head.Provider = "dmr" // config.Cfg's default
	61	+ }
	62	+ return head.Provider + "/" + head.Model
	63	+}

++package ask
++
++import (
++	"fmt"
++	"os"
++
++	"github.com/firebase/genkit/go/ai"
++	"gopkg.in/yaml.v3"
++)
++
++// Exchange is what the asking agent keeps in its history: the question as a
++// user message and the other agent's answer as the assistant's, each marked
++// with who was asked. A PAIR, not one message: the next question is a user
++// message too, and two user messages in a row are refused by chat templates
++// that require the roles to alternate (Gemma's, Mistral's).
++//
++// The answer is capped at maxChars, its beginning and end kept — the same
++// rule as a tool's output, since it lands in the same window.
++//
++//	history = append(history, ask.Exchange("bob", "ollama/qwen3.5:2b", false, prompt, answer, config.Cfg.MaxOutput)...)
++func Exchange(agent, label string, withContext bool, question, answer string, maxChars int) []*ai.Message {
++	how := ""
++	if withContext {
++		how = ", with this conversation as context"
++	}
++	asked := fmt.Sprintf("[asked %s (%s)%s: %s]", agent, label, how, question)
++	answered := fmt.Sprintf("[%s's answer]\n%s", agent, clipMiddle(answer, maxChars))
++	return []*ai.Message{ai.NewUserTextMessage(asked), ai.NewModelTextMessage(answered)}
++}
++
++// clipMiddle caps s at max characters, keeping the beginning and the end.
++func clipMiddle(s string, max int) string {
++	r := []rune(s)
++	if max <= 0 || len(r) <= max {
++		return s
++	}
++	tail := min(2000, max/2)
++	head := max - tail
++	return fmt.Sprintf("%s\n\n[... %d characters truncated ...]\n\n%s", string(r[:head]), len(r)-head-tail, string(r[len(r)-tail:]))
++}
++
++// Describe reads "<provider>/<model>" from an agent's config file, for the
++// marker in the history and the line the user sees. Best effort: the child
++// validates its own file, and a value it cannot read here is shown as "?".
++//
++//	ask.Describe("demo-ollama/agent.ollama.yaml") // "ollama/qwen3.5:2b"
++func Describe(path string) string {
++	data, err := os.ReadFile(path)
++	if err != nil {
++		return "?"
++	}
++	var head struct {
++		Provider string `yaml:"provider"`
++		Model    string `yaml:"model"`
++	}
++	if yaml.Unmarshal(data, &head) != nil || head.Model == "" {
++		return "?"
++	}
++	if head.Provider == "" {
++		head.Provider = "dmr" // config.Cfg's default
++	}
++	return head.Provider + "/" + head.Model
++}

added internal/ask/exchange_test.go +56 -0 Viewed

new file mode 100644
@@ -0,0 +1,56 @@
	1	+package ask
	2	+
	3	+import (
	4	+ "os"
	5	+ "path/filepath"
	6	+ "strings"
	7	+ "testing"
	8	+
	9	+ "github.com/firebase/genkit/go/ai"
	10	+)
	11	+
	12	+// The exchange is a user/model pair (roles must alternate), marked with who
	13	+// was asked and how, the answer capped with its beginning and end kept.
	14	+func TestExchangeIsAMarkedPair(t *testing.T) {
	15	+ msgs := Exchange("bob", "ollama/qwen3.5:2b", true, "is it safe?", "Yes.", 100)
	16	+ if len(msgs) != 2 \|\| msgs[0].Role != ai.RoleUser \|\| msgs[1].Role != ai.RoleModel {
	17	+ t.Fatalf("Exchange = %v, want a user then a model message", msgs)
	18	+ }
	19	+ if got := msgs[0].Text(); got != "[asked bob (ollama/qwen3.5:2b), with this conversation as context: is it safe?]" {
	20	+ t.Errorf("question = %q", got)
	21	+ }
	22	+ if got := msgs[1].Text(); got != "[bob's answer]\nYes." {
	23	+ t.Errorf("answer = %q", got)
	24	+ }
	25	+
	26	+ long := "BEGIN" + strings.Repeat("x", 500) + "END"
	27	+ got := Exchange("bob", "x/y", false, "q", long, 100)[1].Text()
	28	+ if !strings.Contains(got, "BEGIN") \|\| !strings.HasSuffix(got, "END") \|\| !strings.Contains(got, "characters truncated") {
	29	+ t.Errorf("capped answer = %q", got)
	30	+ }
	31	+ if strings.Contains(Exchange("bob", "x/y", false, "q", "a", 100)[0].Text(), "context") {
	32	+ t.Error("a plain /ask says it sent the context")
	33	+ }
	34	+}
	35	+
	36	+// Describe reads provider/model from the agent's file, defaults the provider
	37	+// like the config does, and says "?" when it cannot.
	38	+func TestDescribe(t *testing.T) {
	39	+ dir := t.TempDir()
	40	+ write := func(name, body string) string {
	41	+ p := filepath.Join(dir, name)
	42	+ if err := os.WriteFile(p, []byte(body), 0o644); err != nil {
	43	+ t.Fatal(err)
	44	+ }
	45	+ return p
	46	+ }
	47	+ if got := Describe(write("a.yaml", "provider: ollama\nmodel: qwen3.5:2b\n")); got != "ollama/qwen3.5:2b" {
	48	+ t.Errorf("Describe = %q", got)
	49	+ }
	50	+ if got := Describe(write("b.yaml", "model: ai/gemma\n")); got != "dmr/ai/gemma" {
	51	+ t.Errorf("Describe without provider = %q", got)
	52	+ }
	53	+ if got := Describe(filepath.Join(dir, "missing.yaml")); got != "?" {
	54	+ t.Errorf("Describe(missing) = %q", got)
	55	+ }
	56	+}

++package ask
++
++import (
++	"os"
++	"path/filepath"
++	"strings"
++	"testing"
++
++	"github.com/firebase/genkit/go/ai"
++)
++
++// The exchange is a user/model pair (roles must alternate), marked with who
++// was asked and how, the answer capped with its beginning and end kept.
++func TestExchangeIsAMarkedPair(t *testing.T) {
++	msgs := Exchange("bob", "ollama/qwen3.5:2b", true, "is it safe?", "Yes.", 100)
++	if len(msgs) != 2 || msgs[0].Role != ai.RoleUser || msgs[1].Role != ai.RoleModel {
++		t.Fatalf("Exchange = %v, want a user then a model message", msgs)
++	}
++	if got := msgs[0].Text(); got != "[asked bob (ollama/qwen3.5:2b), with this conversation as context: is it safe?]" {
++		t.Errorf("question = %q", got)
++	}
++	if got := msgs[1].Text(); got != "[bob's answer]\nYes." {
++		t.Errorf("answer = %q", got)
++	}
++
++	long := "BEGIN" + strings.Repeat("x", 500) + "END"
++	got := Exchange("bob", "x/y", false, "q", long, 100)[1].Text()
++	if !strings.Contains(got, "BEGIN") || !strings.HasSuffix(got, "END") || !strings.Contains(got, "characters truncated") {
++		t.Errorf("capped answer = %q", got)
++	}
++	if strings.Contains(Exchange("bob", "x/y", false, "q", "a", 100)[0].Text(), "context") {
++		t.Error("a plain /ask says it sent the context")
++	}
++}
++
++// Describe reads provider/model from the agent's file, defaults the provider
++// like the config does, and says "?" when it cannot.
++func TestDescribe(t *testing.T) {
++	dir := t.TempDir()
++	write := func(name, body string) string {
++		p := filepath.Join(dir, name)
++		if err := os.WriteFile(p, []byte(body), 0o644); err != nil {
++			t.Fatal(err)
++		}
++		return p
++	}
++	if got := Describe(write("a.yaml", "provider: ollama\nmodel: qwen3.5:2b\n")); got != "ollama/qwen3.5:2b" {
++		t.Errorf("Describe = %q", got)
++	}
++	if got := Describe(write("b.yaml", "model: ai/gemma\n")); got != "dmr/ai/gemma" {
++		t.Errorf("Describe without provider = %q", got)
++	}
++	if got := Describe(filepath.Join(dir, "missing.yaml")); got != "?" {
++		t.Errorf("Describe(missing) = %q", got)
++	}
++}

added internal/ask/pool.go +253 -0 Viewed

new file mode 100644
@@ -0,0 +1,253 @@
	1	+package ask
	2	+
	3	+import (
	4	+ "context"
	5	+ "errors"
	6	+ "fmt"
	7	+ "io"
	8	+ "log/slog"
	9	+ "os"
	10	+ "os/exec"
	11	+ "strings"
	12	+ "sync"
	13	+ "time"
	14	+
	15	+ "mm/internal/ui"
	16	+
	17	+ sdk "github.com/coder/acp-go-sdk"
	18	+)
	19	+
	20	+// Pool keeps one child agent per name, started on its first /ask and kept for
	21	+// the life of the asking agent: starting mm costs its engine set-up, its probe
	22	+// and its MCP servers, and a second question should not pay that again. Each
	23	+// ask opens a FRESH session in the child — no memory between two asks; the
	24	+// conversation travels explicitly, with /ask-with-context.
	25	+//
	26	+// pool := ask.NewPool()
	27	+// defer pool.Close()
	28	+// answer, err := pool.Ask(ctx, "bob", "/path/agent.bob.yaml", cwd, "what is 2+2?", sink)
	29	+type Pool struct {
	30	+ // command builds the child process for a config file; a test swaps it for
	31	+ // a fake agent.
	32	+ command func(configPath string) *exec.Cmd
	33	+
	34	+ mu sync.Mutex
	35	+ children map[string]*child
	36	+}
	37	+
	38	+// NewPool returns a pool whose children run this very binary, so every
	39	+// agent is the same mm build.
	40	+func NewPool() *Pool {
	41	+ return &Pool{command: childCommand, children: map[string]*child{}}
	42	+}
	43	+
	44	+// childCommand is `<this mm> -acp -config <path>`, in the asker's environment
	45	+// minus the overrides meant for the asker alone (ChildEnv).
	46	+func childCommand(configPath string) *exec.Cmd {
	47	+ self, err := os.Executable()
	48	+ if err != nil {
	49	+ self = os.Args[0]
	50	+ }
	51	+ cmd := exec.Command(self, "-acp", "-config", configPath)
	52	+ cmd.Env = ChildEnv(os.Environ(), Depth())
	53	+ return cmd
	54	+}
	55	+
	56	+// startTimeout bounds the child's start-up: process, initialize, its probe.
	57	+// Generous — a cold Genkit and a slow MCP discovery (5 s by default) — but
	58	+// finite, so a child that never speaks ACP fails with its stderr instead of
	59	+// hanging the asker.
	60	+var startTimeout = 60 * time.Second
	61	+
	62	+// Ask puts prompt to the agent `name` (config file `configPath`), working in
	63	+// cwd — the asker's project; in an editor, the session's — streaming what it
	64	+// does to sink, and returns its answer. A cancelled ctx cancels the
	65	+// child's turn (session/cancel) and returns what was answered so far with
	66	+// the context's error. A child that died is started again on the next ask.
	67	+func (p *Pool) Ask(ctx context.Context, name, configPath, cwd, prompt string, sink ui.Sink) (string, error) {
	68	+ if err := CanAsk(); err != nil {
	69	+ return "", err
	70	+ }
	71	+ if _, err := os.Stat(configPath); err != nil {
	72	+ return "", fmt.Errorf("%s: config file %s: %w", name, configPath, errors.Unwrap(err))
	73	+ }
	74	+ c, err := p.child(ctx, name, configPath)
	75	+ if err != nil {
	76	+ return "", err
	77	+ }
	78	+ return c.ask(ctx, name, cwd, prompt, sink)
	79	+}
	80	+
	81	+// child returns the running child for name, starting it when there is none
	82	+// or the previous one exited.
	83	+func (p Pool) child(ctx context.Context, name, configPath string) (child, error) {
	84	+ p.mu.Lock()
	85	+ defer p.mu.Unlock()
	86	+ if c := p.children[name]; c != nil && c.alive() {
	87	+ return c, nil
	88	+ }
	89	+ c, err := startChild(ctx, name, p.command(configPath))
	90	+ if err != nil {
	91	+ return nil, err
	92	+ }
	93	+ p.children[name] = c
	94	+ return c, nil
	95	+}
	96	+
	97	+// Close stops every child: stdin closed (mm -acp exits on EOF), then killed
	98	+// if it is still there a moment later.
	99	+func (p *Pool) Close() {
	100	+ p.mu.Lock()
	101	+ defer p.mu.Unlock()
	102	+ for name, c := range p.children {
	103	+ c.stop()
	104	+ delete(p.children, name)
	105	+ }
	106	+}
	107	+
	108	+// child is one running agent and the ACP connection to it.
	109	+type child struct {
	110	+ name string
	111	+ cmd *exec.Cmd
	112	+ stdin io.WriteCloser
	113	+ conn *sdk.ClientSideConnection
	114	+ client *client
	115	+ stderr *tail
	116	+ exited chan struct{}
	117	+
	118	+ askMu sync.Mutex // one ask at a time per child: one sink, one answer
	119	+}
	120	+
	121	+// startChild starts the process and runs the ACP handshake.
	122	+func startChild(ctx context.Context, name string, cmd exec.Cmd) (child, error) {
	123	+ stdin, err := cmd.StdinPipe()
	124	+ if err != nil {
	125	+ return nil, err
	126	+ }
	127	+ stdout, err := cmd.StdoutPipe()
	128	+ if err != nil {
	129	+ return nil, err
	130	+ }
	131	+ c := &child{name: name, cmd: cmd, stdin: stdin, client: &client{}, stderr: &tail{max: 4096}, exited: make(chan struct{})}
	132	+ cmd.Stderr = c.stderr
	133	+ if err := cmd.Start(); err != nil {
	134	+ return nil, fmt.Errorf("%s: cannot start: %w", name, err)
	135	+ }
	136	+ go func() { _ = cmd.Wait(); close(c.exited) }()
	137	+ c.conn = sdk.NewClientSideConnection(c.client, stdin, stdout)
	138	+ // The SDK logs "connection closed" and the like through slog's default,
	139	+ // which is the asker's terminal: the child's life is reported by Ask's
	140	+ // errors, nowhere else.
	141	+ c.conn.SetLogger(slog.New(slog.NewTextHandler(io.Discard, nil)))
	142	+
	143	+ initCtx, cancel := context.WithTimeout(ctx, startTimeout)
	144	+ defer cancel()
	145	+ _, err = c.conn.Initialize(initCtx, sdk.InitializeRequest{
	146	+ ProtocolVersion: sdk.ProtocolVersionNumber,
	147	+ ClientInfo: &sdk.Implementation{Name: "mm"},
	148	+ })
	149	+ if err != nil {
	150	+ c.stop()
	151	+ return nil, c.failure("did not start", err)
	152	+ }
	153	+ return c, nil
	154	+}
	155	+
	156	+// ask runs one prompt in a fresh session.
	157	+func (c *child) ask(ctx context.Context, name, cwd, prompt string, sink ui.Sink) (string, error) {
	158	+ c.askMu.Lock()
	159	+ defer c.askMu.Unlock()
	160	+ sess, err := c.conn.NewSession(ctx, sdk.NewSessionRequest{Cwd: cwd, McpServers: []sdk.McpServer{}})
	161	+ if err != nil {
	162	+ return "", c.failure("could not open a session", err)
	163	+ }
	164	+ c.client.begin(name, sess.SessionId, sink)
	165	+ _, err = c.conn.Prompt(ctx, sdk.PromptRequest{SessionId: sess.SessionId, Prompt: []sdk.ContentBlock{sdk.TextBlock(prompt)}})
	166	+ answer := c.client.end()
	167	+ switch {
	168	+ case ctx.Err() != nil:
	169	+ return answer, ctx.Err()
	170	+ case err != nil:
	171	+ return answer, c.failure("failed", err)
	172	+ }
	173	+ return answer, nil
	174	+}
	175	+
	176	+// failure is one error line: what went wrong, the protocol's words, and —
	177	+// when the child wrote anything on stderr — its last line, which is usually
	178	+// the real cause ("[config error: …]", "[engine error: …]").
	179	+func (c *child) failure(what string, err error) error {
	180	+ msg := fmt.Sprintf("%s %s: %v", c.name, what, acpError(err))
	181	+ if last := c.stderr.lastLine(); last != "" {
	182	+ msg += " — " + last
	183	+ }
	184	+ return errors.New(msg)
	185	+}
	186	+
	187	+// acpError digs the agent's own message out of a JSON-RPC error: mm's ACP
	188	+// server puts its one-line explanation under data.error.
	189	+func acpError(err error) string {
	190	+ var re *sdk.RequestError
	191	+ if errors.As(err, &re) {
	192	+ if data, ok := re.Data.(map[string]any); ok {
	193	+ if msg, ok := data["error"].(string); ok && msg != "" {
	194	+ return msg
	195	+ }
	196	+ }
	197	+ return re.Message
	198	+ }
	199	+ return err.Error()
	200	+}
	201	+
	202	+// alive reports whether the process is still running.
	203	+func (c *child) alive() bool {
	204	+ select {
	205	+ case <-c.exited:
	206	+ return false
	207	+ default:
	208	+ return true
	209	+ }
	210	+}
	211	+
	212	+// stop ends the child: EOF on its stdin first, then a kill.
	213	+func (c *child) stop() {
	214	+ _ = c.stdin.Close()
	215	+ select {
	216	+ case <-c.exited:
	217	+ case <-time.After(2 * time.Second):
	218	+ _ = c.cmd.Process.Kill()
	219	+ <-c.exited
	220	+ }
	221	+}
	222	+
	223	+// tail keeps the last max bytes written to it: the child's stderr, kept for
	224	+// the error line and nothing else.
	225	+type tail struct {
	226	+ mu sync.Mutex
	227	+ buf []byte
	228	+ max int
	229	+}
	230	+
	231	+func (t *tail) Write(p []byte) (int, error) {
	232	+ t.mu.Lock()
	233	+ defer t.mu.Unlock()
	234	+ t.buf = append(t.buf, p...)
	235	+ if len(t.buf) > t.max {
	236	+ t.buf = t.buf[len(t.buf)-t.max:]
	237	+ }
	238	+ return len(p), nil
	239	+}
	240	+
	241	+// lastLine is the last non-empty line written.
	242	+func (t *tail) lastLine() string {
	243	+ t.mu.Lock()
	244	+ defer t.mu.Unlock()
	245	+ lines := strings.Split(strings.TrimSpace(string(t.buf)), "\n")
	246	+ return strings.TrimSpace(lines[len(lines)-1])
	247	+}
	248	+
	249	+// exists reports whether path names an existing file.
	250	+func exists(path string) bool {
	251	+ _, err := os.Stat(path)
	252	+ return err == nil
	253	+}

++package ask
++
++import (
++	"context"
++	"errors"
++	"fmt"
++	"io"
++	"log/slog"
++	"os"
++	"os/exec"
++	"strings"
++	"sync"
++	"time"
++
++	"mm/internal/ui"
++
++	sdk "github.com/coder/acp-go-sdk"
++)
++
++// Pool keeps one child agent per name, started on its first /ask and kept for
++// the life of the asking agent: starting mm costs its engine set-up, its probe
++// and its MCP servers, and a second question should not pay that again. Each
++// ask opens a FRESH session in the child — no memory between two asks; the
++// conversation travels explicitly, with /ask-with-context.
++//
++//	pool := ask.NewPool()
++//	defer pool.Close()
++//	answer, err := pool.Ask(ctx, "bob", "/path/agent.bob.yaml", cwd, "what is 2+2?", sink)
++type Pool struct {
++	// command builds the child process for a config file; a test swaps it for
++	// a fake agent.
++	command func(configPath string) *exec.Cmd
++
++	mu       sync.Mutex
++	children map[string]*child
++}
++
++// NewPool returns a pool whose children run this very binary, so every
++// agent is the same mm build.
++func NewPool() *Pool {
++	return &Pool{command: childCommand, children: map[string]*child{}}
++}
++
++// childCommand is `<this mm> -acp -config <path>`, in the asker's environment
++// minus the overrides meant for the asker alone (ChildEnv).
++func childCommand(configPath string) *exec.Cmd {
++	self, err := os.Executable()
++	if err != nil {
++		self = os.Args[0]
++	}
++	cmd := exec.Command(self, "-acp", "-config", configPath)
++	cmd.Env = ChildEnv(os.Environ(), Depth())
++	return cmd
++}
++
++// startTimeout bounds the child's start-up: process, initialize, its probe.
++// Generous — a cold Genkit and a slow MCP discovery (5 s by default) — but
++// finite, so a child that never speaks ACP fails with its stderr instead of
++// hanging the asker.
++var startTimeout = 60 * time.Second
++
++// Ask puts prompt to the agent `name` (config file `configPath`), working in
++// cwd — the asker's project; in an editor, the session's — streaming what it
++// does to sink, and returns its answer. A cancelled ctx cancels the
++// child's turn (session/cancel) and returns what was answered so far with
++// the context's error. A child that died is started again on the next ask.
++func (p *Pool) Ask(ctx context.Context, name, configPath, cwd, prompt string, sink ui.Sink) (string, error) {
++	if err := CanAsk(); err != nil {
++		return "", err
++	}
++	if _, err := os.Stat(configPath); err != nil {
++		return "", fmt.Errorf("%s: config file %s: %w", name, configPath, errors.Unwrap(err))
++	}
++	c, err := p.child(ctx, name, configPath)
++	if err != nil {
++		return "", err
++	}
++	return c.ask(ctx, name, cwd, prompt, sink)
++}
++
++// child returns the running child for name, starting it when there is none
++// or the previous one exited.
++func (p *Pool) child(ctx context.Context, name, configPath string) (*child, error) {
++	p.mu.Lock()
++	defer p.mu.Unlock()
++	if c := p.children[name]; c != nil && c.alive() {
++		return c, nil
++	}
++	c, err := startChild(ctx, name, p.command(configPath))
++	if err != nil {
++		return nil, err
++	}
++	p.children[name] = c
++	return c, nil
++}
++
++// Close stops every child: stdin closed (mm -acp exits on EOF), then killed
++// if it is still there a moment later.
++func (p *Pool) Close() {
++	p.mu.Lock()
++	defer p.mu.Unlock()
++	for name, c := range p.children {
++		c.stop()
++		delete(p.children, name)
++	}
++}
++
++// child is one running agent and the ACP connection to it.
++type child struct {
++	name   string
++	cmd    *exec.Cmd
++	stdin  io.WriteCloser
++	conn   *sdk.ClientSideConnection
++	client *client
++	stderr *tail
++	exited chan struct{}
++
++	askMu sync.Mutex // one ask at a time per child: one sink, one answer
++}
++
++// startChild starts the process and runs the ACP handshake.
++func startChild(ctx context.Context, name string, cmd *exec.Cmd) (*child, error) {
++	stdin, err := cmd.StdinPipe()
++	if err != nil {
++		return nil, err
++	}
++	stdout, err := cmd.StdoutPipe()
++	if err != nil {
++		return nil, err
++	}
++	c := &child{name: name, cmd: cmd, stdin: stdin, client: &client{}, stderr: &tail{max: 4096}, exited: make(chan struct{})}
++	cmd.Stderr = c.stderr
++	if err := cmd.Start(); err != nil {
++		return nil, fmt.Errorf("%s: cannot start: %w", name, err)
++	}
++	go func() { _ = cmd.Wait(); close(c.exited) }()
++	c.conn = sdk.NewClientSideConnection(c.client, stdin, stdout)
++	// The SDK logs "connection closed" and the like through slog's default,
++	// which is the asker's terminal: the child's life is reported by Ask's
++	// errors, nowhere else.
++	c.conn.SetLogger(slog.New(slog.NewTextHandler(io.Discard, nil)))
++
++	initCtx, cancel := context.WithTimeout(ctx, startTimeout)
++	defer cancel()
++	_, err = c.conn.Initialize(initCtx, sdk.InitializeRequest{
++		ProtocolVersion: sdk.ProtocolVersionNumber,
++		ClientInfo:      &sdk.Implementation{Name: "mm"},
++	})
++	if err != nil {
++		c.stop()
++		return nil, c.failure("did not start", err)
++	}
++	return c, nil
++}
++
++// ask runs one prompt in a fresh session.
++func (c *child) ask(ctx context.Context, name, cwd, prompt string, sink ui.Sink) (string, error) {
++	c.askMu.Lock()
++	defer c.askMu.Unlock()
++	sess, err := c.conn.NewSession(ctx, sdk.NewSessionRequest{Cwd: cwd, McpServers: []sdk.McpServer{}})
++	if err != nil {
++		return "", c.failure("could not open a session", err)
++	}
++	c.client.begin(name, sess.SessionId, sink)
++	_, err = c.conn.Prompt(ctx, sdk.PromptRequest{SessionId: sess.SessionId, Prompt: []sdk.ContentBlock{sdk.TextBlock(prompt)}})
++	answer := c.client.end()
++	switch {
++	case ctx.Err() != nil:
++		return answer, ctx.Err()
++	case err != nil:
++		return answer, c.failure("failed", err)
++	}
++	return answer, nil
++}
++
++// failure is one error line: what went wrong, the protocol's words, and —
++// when the child wrote anything on stderr — its last line, which is usually
++// the real cause ("[config error: …]", "[engine error: …]").
++func (c *child) failure(what string, err error) error {
++	msg := fmt.Sprintf("%s %s: %v", c.name, what, acpError(err))
++	if last := c.stderr.lastLine(); last != "" {
++		msg += " — " + last
++	}
++	return errors.New(msg)
++}
++
++// acpError digs the agent's own message out of a JSON-RPC error: mm's ACP
++// server puts its one-line explanation under data.error.
++func acpError(err error) string {
++	var re *sdk.RequestError
++	if errors.As(err, &re) {
++		if data, ok := re.Data.(map[string]any); ok {
++			if msg, ok := data["error"].(string); ok && msg != "" {
++				return msg
++			}
++		}
++		return re.Message
++	}
++	return err.Error()
++}
++
++// alive reports whether the process is still running.
++func (c *child) alive() bool {
++	select {
++	case <-c.exited:
++		return false
++	default:
++		return true
++	}
++}
++
++// stop ends the child: EOF on its stdin first, then a kill.
++func (c *child) stop() {
++	_ = c.stdin.Close()
++	select {
++	case <-c.exited:
++	case <-time.After(2 * time.Second):
++		_ = c.cmd.Process.Kill()
++		<-c.exited
++	}
++}
++
++// tail keeps the last max bytes written to it: the child's stderr, kept for
++// the error line and nothing else.
++type tail struct {
++	mu  sync.Mutex
++	buf []byte
++	max int
++}
++
++func (t *tail) Write(p []byte) (int, error) {
++	t.mu.Lock()
++	defer t.mu.Unlock()
++	t.buf = append(t.buf, p...)
++	if len(t.buf) > t.max {
++		t.buf = t.buf[len(t.buf)-t.max:]
++	}
++	return len(p), nil
++}
++
++// lastLine is the last non-empty line written.
++func (t *tail) lastLine() string {
++	t.mu.Lock()
++	defer t.mu.Unlock()
++	lines := strings.Split(strings.TrimSpace(string(t.buf)), "\n")
++	return strings.TrimSpace(lines[len(lines)-1])
++}
++
++// exists reports whether path names an existing file.
++func exists(path string) bool {
++	_, err := os.Stat(path)
++	return err == nil
++}

added internal/ask/pool_test.go +268 -0 Viewed

new file mode 100644
@@ -0,0 +1,268 @@
	1	+package ask
	2	+
	3	+import (
	4	+ "context"
	5	+ "fmt"
	6	+ "os"
	7	+ "os/exec"
	8	+ "path/filepath"
	9	+ "strings"
	10	+ "sync"
	11	+ "testing"
	12	+ "time"
	13	+
	14	+ "mm/internal/ui"
	15	+
	16	+ sdk "github.com/coder/acp-go-sdk"
	17	+)
	18	+
	19	+// The child agent of these tests is THIS test binary, re-run with
	20	+// ASK_FAKE_AGENT set: TestFakeAgent then serves ACP on stdin/stdout like
	21	+// `mm -acp` would, with behaviour picked by the variable's value. No model,
	22	+// no network, and the real client, pool, pipes and process handling.
	23	+func TestFakeAgent(t *testing.T) {
	24	+ mode := os.Getenv("ASK_FAKE_AGENT")
	25	+ if mode == "" {
	26	+ t.Skip("helper process only")
	27	+ }
	28	+ if mode == "crash" {
	29	+ fmt.Fprintln(os.Stderr, "[config error: bad.yaml: model must not be empty ]")
	30	+ os.Exit(1)
	31	+ }
	32	+ a := &fakeAgent{mode: mode}
	33	+ a.conn = sdk.NewAgentSideConnection(a, os.Stdout, os.Stdin)
	34	+ <-a.conn.Done()
	35	+ os.Exit(0)
	36	+}
	37	+
	38	+// fakePool is a pool whose children are the fake agent in the given mode.
	39	+func fakePool(t testing.T, mode string) Pool {
	40	+ t.Helper()
	41	+ p := NewPool()
	42	+ p.command = func(string) *exec.Cmd {
	43	+ cmd := exec.Command(os.Args[0], "-test.run=^TestFakeAgent$")
	44	+ cmd.Env = append(os.Environ(), "ASK_FAKE_AGENT="+mode)
	45	+ return cmd
	46	+ }
	47	+ t.Cleanup(p.Close)
	48	+ return p
	49	+}
	50	+
	51	+// configFile is an existing file for Ask's check; its content is not read by
	52	+// the fake agent.
	53	+func configFile(t *testing.T) string {
	54	+ t.Helper()
	55	+ path := filepath.Join(t.TempDir(), "agent.bob.yaml")
	56	+ if err := os.WriteFile(path, []byte("model: m\n"), 0o644); err != nil {
	57	+ t.Fatal(err)
	58	+ }
	59	+ return path
	60	+}
	61	+
	62	+// recorder is a ui.Sink that remembers what it was shown and answers
	63	+// permission requests with `allow`.
	64	+type recorder struct {
	65	+ mu sync.Mutex
	66	+ text strings.Builder
	67	+ events []string
	68	+ allow bool
	69	+}
	70	+
	71	+func (r *recorder) log(s string) { r.mu.Lock(); r.events = append(r.events, s); r.mu.Unlock() }
	72	+func (r *recorder) Text(s string) { r.mu.Lock(); r.text.WriteString(s); r.mu.Unlock() }
	73	+func (r *recorder) Reasoning(s string) { r.log("thought:" + s) }
	74	+func (r *recorder) ToolStart(ev ui.ToolEvent) { r.log("start:" + ev.Title + ":" + ev.Tool) }
	75	+func (r *recorder) ToolRunning(string) { r.log("running") }
	76	+func (r *recorder) ToolEnd(_ string, res ui.ToolResult) {
	77	+ r.log(fmt.Sprintf("end:%s:%v", res.Output, res.Failed))
	78	+}
	79	+func (r *recorder) WorkDir() string { return "" }
	80	+func (r *recorder) Allow(_ context.Context, ev ui.ToolEvent) bool {
	81	+ r.log("allow?" + ev.Title)
	82	+ return r.allow
	83	+}
	84	+func (r *recorder) all() string {
	85	+ r.mu.Lock()
	86	+ defer r.mu.Unlock()
	87	+ return strings.Join(r.events, " \| ")
	88	+}
	89	+
	90	+// One ask: the prompt reaches the child, its text streams to the sink and is
	91	+// returned, its tool call is shown under bob's name with the permission asked
	92	+// of the asker, and its thoughts are shown as thoughts.
	93	+func TestAskStreamsTheAnswerAndTheToolCalls(t *testing.T) {
	94	+ p := fakePool(t, "tool")
	95	+ sink := &recorder{allow: true}
	96	+ answer, err := p.Ask(context.Background(), "bob", configFile(t), t.TempDir(), "what is 2+2?", sink)
	97	+ if err != nil {
	98	+ t.Fatal(err)
	99	+ }
	100	+ if answer != "You asked: what is 2+2? (tool allowed)" \|\| sink.text.String() != answer {
	101	+ t.Errorf("answer %q, streamed %q", answer, sink.text.String())
	102	+ }
	103	+ // The SDK delivers the child's requests (the permission) concurrently with
	104	+ // its notifications (the thought), so the thought's place may vary; the
	105	+ // tool call's own order may not: shown, asked, running, done.
	106	+ got := sink.all()
	107	+ if !strings.Contains(got, "thought:hmm") {
	108	+ t.Errorf("events %s: the thought is missing", got)
	109	+ }
	110	+ tool := strings.ReplaceAll(strings.ReplaceAll(got, "thought:hmm \| ", ""), " \| thought:hmm", "")
	111	+ if want := "start:bob · bash: ls:bob/bash \| allow?bob · bash: ls \| running \| end:a.txt:false"; tool != want {
	112	+ t.Errorf("tool events:\n got %s\nwant %s", tool, want)
	113	+ }
	114	+}
	115	+
	116	+// A refusal from the asker's user reaches the child as its reject option.
	117	+func TestAskForwardsARefusal(t *testing.T) {
	118	+ p := fakePool(t, "tool")
	119	+ answer, err := p.Ask(context.Background(), "bob", configFile(t), t.TempDir(), "q", &recorder{allow: false})
	120	+ if err != nil \|\| !strings.HasSuffix(answer, "(tool rejected)") {
	121	+ t.Errorf("answer %q, err %v, want the child to see the rejection", answer, err)
	122	+ }
	123	+}
	124	+
	125	+// The child is started once and reused, with a fresh session per ask.
	126	+func TestAskReusesTheChildWithAFreshSession(t *testing.T) {
	127	+ p := fakePool(t, "echo")
	128	+ cfg := configFile(t)
	129	+ first, err := p.Ask(context.Background(), "bob", cfg, t.TempDir(), "one", &recorder{})
	130	+ if err != nil {
	131	+ t.Fatal(err)
	132	+ }
	133	+ c := p.children["bob"]
	134	+ second, err := p.Ask(context.Background(), "bob", cfg, t.TempDir(), "two", &recorder{})
	135	+ if err != nil {
	136	+ t.Fatal(err)
	137	+ }
	138	+ if p.children["bob"] != c {
	139	+ t.Error("the child was started again")
	140	+ }
	141	+ if first == second \|\| !strings.Contains(second, "session 2") {
	142	+ t.Errorf("answers %q / %q, want two sessions", first, second)
	143	+ }
	144	+}
	145	+
	146	+// Cancelling the asker's context cancels the child's turn: Ask returns the
	147	+// context's error and what had streamed, and the child is still usable.
	148	+func TestAskIsCancellable(t *testing.T) {
	149	+ p := fakePool(t, "slow")
	150	+ ctx, cancel := context.WithCancel(context.Background())
	151	+ sink := &recorder{}
	152	+ go func() {
	153	+ for !strings.Contains(sink.textNow(), "working") {
	154	+ time.Sleep(10 * time.Millisecond)
	155	+ }
	156	+ cancel()
	157	+ }()
	158	+ answer, err := p.Ask(ctx, "bob", configFile(t), t.TempDir(), "q", sink)
	159	+ if err != context.Canceled \|\| answer != "working" {
	160	+ t.Errorf("answer %q, err %v, want the partial answer and context.Canceled", answer, err)
	161	+ }
	162	+ if !p.children["bob"].alive() {
	163	+ t.Error("the child died with the cancelled ask")
	164	+ }
	165	+}
	166	+
	167	+func (r *recorder) textNow() string { r.mu.Lock(); defer r.mu.Unlock(); return r.text.String() }
	168	+
	169	+// A child that fails to start reports its own last stderr line; a missing
	170	+// config file is reported before anything starts.
	171	+func TestAskReportsTheChildFailure(t *testing.T) {
	172	+ p := fakePool(t, "crash")
	173	+ _, err := p.Ask(context.Background(), "bob", configFile(t), t.TempDir(), "q", &recorder{})
	174	+ if err == nil \|\| !strings.Contains(err.Error(), "bob did not start") \|\| !strings.Contains(err.Error(), "[config error: bad.yaml: model must not be empty ]") {
	175	+ t.Errorf("err = %v", err)
	176	+ }
	177	+
	178	+ _, err = fakePool(t, "echo").Ask(context.Background(), "bob", filepath.Join(t.TempDir(), "nope.yaml"), t.TempDir(), "q", &recorder{})
	179	+ if err == nil \|\| !strings.Contains(err.Error(), "bob: config file") {
	180	+ t.Errorf("missing config: err = %v", err)
	181	+ }
	182	+}
	183	+
	184	+// An agent asked through MaxDepth levels cannot ask further.
	185	+func TestAskRespectsTheDepthLimit(t *testing.T) {
	186	+ t.Setenv(DepthEnv, fmt.Sprint(MaxDepth))
	187	+ if _, err := fakePool(t, "echo").Ask(context.Background(), "bob", configFile(t), t.TempDir(), "q", &recorder{}); err == nil \|\| !strings.Contains(err.Error(), "cannot ask further") {
	188	+ t.Errorf("err = %v", err)
	189	+ }
	190	+}
	191	+
	192	+// --- the fake agent ----------------------------------------------------------
	193	+
	194	+type fakeAgent struct {
	195	+ mode string
	196	+ conn *sdk.AgentSideConnection
	197	+ sessions int
	198	+}
	199	+
	200	+func (a *fakeAgent) send(ctx context.Context, sid sdk.SessionId, u sdk.SessionUpdate) {
	201	+ _ = a.conn.SessionUpdate(ctx, sdk.SessionNotification{SessionId: sid, Update: u})
	202	+}
	203	+
	204	+func (a *fakeAgent) Prompt(ctx context.Context, p sdk.PromptRequest) (sdk.PromptResponse, error) {
	205	+ question := ""
	206	+ if len(p.Prompt) > 0 && p.Prompt[0].Text != nil {
	207	+ question = p.Prompt[0].Text.Text
	208	+ }
	209	+ switch a.mode {
	210	+ case "echo":
	211	+ a.send(ctx, p.SessionId, sdk.UpdateAgentMessageText(fmt.Sprintf("session %d: %s", a.sessions, question)))
	212	+ case "slow":
	213	+ a.send(ctx, p.SessionId, sdk.UpdateAgentMessageText("working"))
	214	+ <-ctx.Done()
	215	+ return sdk.PromptResponse{StopReason: sdk.StopReasonCancelled}, nil
	216	+ case "tool":
	217	+ a.send(ctx, p.SessionId, sdk.UpdateAgentThoughtText("hmm"))
	218	+ a.send(ctx, p.SessionId, sdk.StartToolCall("call_1", "bash: ls", sdk.WithStartKind(sdk.ToolKindExecute), sdk.WithStartStatus(sdk.ToolCallStatusPending)))
	219	+ resp, err := a.conn.RequestPermission(ctx, sdk.RequestPermissionRequest{
	220	+ SessionId: p.SessionId,
	221	+ ToolCall: sdk.ToolCallUpdate{ToolCallId: "call_1", Title: sdk.Ptr("bash: ls")},
	222	+ Options: []sdk.PermissionOption{
	223	+ {Kind: sdk.PermissionOptionKindAllowOnce, Name: "Allow", OptionId: "allow"},
	224	+ {Kind: sdk.PermissionOptionKindRejectOnce, Name: "Reject", OptionId: "reject"},
	225	+ },
	226	+ })
	227	+ verdict := "rejected"
	228	+ if err == nil && resp.Outcome.Selected != nil && resp.Outcome.Selected.OptionId == "allow" {
	229	+ verdict = "allowed"
	230	+ a.send(ctx, p.SessionId, sdk.UpdateToolCall("call_1", sdk.WithUpdateStatus(sdk.ToolCallStatusInProgress)))
	231	+ a.send(ctx, p.SessionId, sdk.UpdateToolCall("call_1", sdk.WithUpdateStatus(sdk.ToolCallStatusCompleted), sdk.WithUpdateContent([]sdk.ToolCallContent{sdk.ToolContent(sdk.TextBlock("a.txt"))})))
	232	+ }
	233	+ a.send(ctx, p.SessionId, sdk.UpdateAgentMessageText("You asked: "+question+" (tool "+verdict+")"))
	234	+ }
	235	+ return sdk.PromptResponse{StopReason: sdk.StopReasonEndTurn}, nil
	236	+}
	237	+
	238	+func (a *fakeAgent) NewSession(context.Context, sdk.NewSessionRequest) (sdk.NewSessionResponse, error) {
	239	+ a.sessions++
	240	+ return sdk.NewSessionResponse{SessionId: sdk.SessionId(fmt.Sprintf("s%d", a.sessions))}, nil
	241	+}
	242	+
	243	+func (a *fakeAgent) Initialize(context.Context, sdk.InitializeRequest) (sdk.InitializeResponse, error) {
	244	+ return sdk.InitializeResponse{ProtocolVersion: sdk.ProtocolVersionNumber}, nil
	245	+}
	246	+
	247	+func (a *fakeAgent) Authenticate(context.Context, sdk.AuthenticateRequest) (sdk.AuthenticateResponse, error) {
	248	+ return sdk.AuthenticateResponse{}, nil
	249	+}
	250	+func (a *fakeAgent) Logout(context.Context, sdk.LogoutRequest) (sdk.LogoutResponse, error) {
	251	+ return sdk.LogoutResponse{}, nil
	252	+}
	253	+func (a *fakeAgent) Cancel(context.Context, sdk.CancelNotification) error { return nil }
	254	+func (a *fakeAgent) CloseSession(context.Context, sdk.CloseSessionRequest) (sdk.CloseSessionResponse, error) {
	255	+ return sdk.CloseSessionResponse{}, nil
	256	+}
	257	+func (a *fakeAgent) ListSessions(context.Context, sdk.ListSessionsRequest) (sdk.ListSessionsResponse, error) {
	258	+ return sdk.ListSessionsResponse{}, nil
	259	+}
	260	+func (a *fakeAgent) ResumeSession(context.Context, sdk.ResumeSessionRequest) (sdk.ResumeSessionResponse, error) {
	261	+ return sdk.ResumeSessionResponse{}, nil
	262	+}
	263	+func (a *fakeAgent) SetSessionConfigOption(context.Context, sdk.SetSessionConfigOptionRequest) (sdk.SetSessionConfigOptionResponse, error) {
	264	+ return sdk.SetSessionConfigOptionResponse{}, nil
	265	+}
	266	+func (a *fakeAgent) SetSessionMode(context.Context, sdk.SetSessionModeRequest) (sdk.SetSessionModeResponse, error) {
	267	+ return sdk.SetSessionModeResponse{}, nil
	268	+}

++package ask
++
++import (
++	"context"
++	"fmt"
++	"os"
++	"os/exec"
++	"path/filepath"
++	"strings"
++	"sync"
++	"testing"
++	"time"
++
++	"mm/internal/ui"
++
++	sdk "github.com/coder/acp-go-sdk"
++)
++
++// The child agent of these tests is THIS test binary, re-run with
++// ASK_FAKE_AGENT set: TestFakeAgent then serves ACP on stdin/stdout like
++// `mm -acp` would, with behaviour picked by the variable's value. No model,
++// no network, and the real client, pool, pipes and process handling.
++func TestFakeAgent(t *testing.T) {
++	mode := os.Getenv("ASK_FAKE_AGENT")
++	if mode == "" {
++		t.Skip("helper process only")
++	}
++	if mode == "crash" {
++		fmt.Fprintln(os.Stderr, "[config error: bad.yaml: model must not be empty ]")
++		os.Exit(1)
++	}
++	a := &fakeAgent{mode: mode}
++	a.conn = sdk.NewAgentSideConnection(a, os.Stdout, os.Stdin)
++	<-a.conn.Done()
++	os.Exit(0)
++}
++
++// fakePool is a pool whose children are the fake agent in the given mode.
++func fakePool(t *testing.T, mode string) *Pool {
++	t.Helper()
++	p := NewPool()
++	p.command = func(string) *exec.Cmd {
++		cmd := exec.Command(os.Args[0], "-test.run=^TestFakeAgent$")
++		cmd.Env = append(os.Environ(), "ASK_FAKE_AGENT="+mode)
++		return cmd
++	}
++	t.Cleanup(p.Close)
++	return p
++}
++
++// configFile is an existing file for Ask's check; its content is not read by
++// the fake agent.
++func configFile(t *testing.T) string {
++	t.Helper()
++	path := filepath.Join(t.TempDir(), "agent.bob.yaml")
++	if err := os.WriteFile(path, []byte("model: m\n"), 0o644); err != nil {
++		t.Fatal(err)
++	}
++	return path
++}
++
++// recorder is a ui.Sink that remembers what it was shown and answers
++// permission requests with `allow`.
++type recorder struct {
++	mu     sync.Mutex
++	text   strings.Builder
++	events []string
++	allow  bool
++}
++
++func (r *recorder) log(s string)              { r.mu.Lock(); r.events = append(r.events, s); r.mu.Unlock() }
++func (r *recorder) Text(s string)             { r.mu.Lock(); r.text.WriteString(s); r.mu.Unlock() }
++func (r *recorder) Reasoning(s string)        { r.log("thought:" + s) }
++func (r *recorder) ToolStart(ev ui.ToolEvent) { r.log("start:" + ev.Title + ":" + ev.Tool) }
++func (r *recorder) ToolRunning(string)        { r.log("running") }
++func (r *recorder) ToolEnd(_ string, res ui.ToolResult) {
++	r.log(fmt.Sprintf("end:%s:%v", res.Output, res.Failed))
++}
++func (r *recorder) WorkDir() string { return "" }
++func (r *recorder) Allow(_ context.Context, ev ui.ToolEvent) bool {
++	r.log("allow?" + ev.Title)
++	return r.allow
++}
++func (r *recorder) all() string {
++	r.mu.Lock()
++	defer r.mu.Unlock()
++	return strings.Join(r.events, " | ")
++}
++
++// One ask: the prompt reaches the child, its text streams to the sink and is
++// returned, its tool call is shown under bob's name with the permission asked
++// of the asker, and its thoughts are shown as thoughts.
++func TestAskStreamsTheAnswerAndTheToolCalls(t *testing.T) {
++	p := fakePool(t, "tool")
++	sink := &recorder{allow: true}
++	answer, err := p.Ask(context.Background(), "bob", configFile(t), t.TempDir(), "what is 2+2?", sink)
++	if err != nil {
++		t.Fatal(err)
++	}
++	if answer != "You asked: what is 2+2? (tool allowed)" || sink.text.String() != answer {
++		t.Errorf("answer %q, streamed %q", answer, sink.text.String())
++	}
++	// The SDK delivers the child's requests (the permission) concurrently with
++	// its notifications (the thought), so the thought's place may vary; the
++	// tool call's own order may not: shown, asked, running, done.
++	got := sink.all()
++	if !strings.Contains(got, "thought:hmm") {
++		t.Errorf("events %s: the thought is missing", got)
++	}
++	tool := strings.ReplaceAll(strings.ReplaceAll(got, "thought:hmm | ", ""), " | thought:hmm", "")
++	if want := "start:bob · bash: ls:bob/bash | allow?bob · bash: ls | running | end:a.txt:false"; tool != want {
++		t.Errorf("tool events:\n got %s\nwant %s", tool, want)
++	}
++}
++
++// A refusal from the asker's user reaches the child as its reject option.
++func TestAskForwardsARefusal(t *testing.T) {
++	p := fakePool(t, "tool")
++	answer, err := p.Ask(context.Background(), "bob", configFile(t), t.TempDir(), "q", &recorder{allow: false})
++	if err != nil || !strings.HasSuffix(answer, "(tool rejected)") {
++		t.Errorf("answer %q, err %v, want the child to see the rejection", answer, err)
++	}
++}
++
++// The child is started once and reused, with a fresh session per ask.
++func TestAskReusesTheChildWithAFreshSession(t *testing.T) {
++	p := fakePool(t, "echo")
++	cfg := configFile(t)
++	first, err := p.Ask(context.Background(), "bob", cfg, t.TempDir(), "one", &recorder{})
++	if err != nil {
++		t.Fatal(err)
++	}
++	c := p.children["bob"]
++	second, err := p.Ask(context.Background(), "bob", cfg, t.TempDir(), "two", &recorder{})
++	if err != nil {
++		t.Fatal(err)
++	}
++	if p.children["bob"] != c {
++		t.Error("the child was started again")
++	}
++	if first == second || !strings.Contains(second, "session 2") {
++		t.Errorf("answers %q / %q, want two sessions", first, second)
++	}
++}
++
++// Cancelling the asker's context cancels the child's turn: Ask returns the
++// context's error and what had streamed, and the child is still usable.
++func TestAskIsCancellable(t *testing.T) {
++	p := fakePool(t, "slow")
++	ctx, cancel := context.WithCancel(context.Background())
++	sink := &recorder{}
++	go func() {
++		for !strings.Contains(sink.textNow(), "working") {
++			time.Sleep(10 * time.Millisecond)
++		}
++		cancel()
++	}()
++	answer, err := p.Ask(ctx, "bob", configFile(t), t.TempDir(), "q", sink)
++	if err != context.Canceled || answer != "working" {
++		t.Errorf("answer %q, err %v, want the partial answer and context.Canceled", answer, err)
++	}
++	if !p.children["bob"].alive() {
++		t.Error("the child died with the cancelled ask")
++	}
++}
++
++func (r *recorder) textNow() string { r.mu.Lock(); defer r.mu.Unlock(); return r.text.String() }
++
++// A child that fails to start reports its own last stderr line; a missing
++// config file is reported before anything starts.
++func TestAskReportsTheChildFailure(t *testing.T) {
++	p := fakePool(t, "crash")
++	_, err := p.Ask(context.Background(), "bob", configFile(t), t.TempDir(), "q", &recorder{})
++	if err == nil || !strings.Contains(err.Error(), "bob did not start") || !strings.Contains(err.Error(), "[config error: bad.yaml: model must not be empty ]") {
++		t.Errorf("err = %v", err)
++	}
++
++	_, err = fakePool(t, "echo").Ask(context.Background(), "bob", filepath.Join(t.TempDir(), "nope.yaml"), t.TempDir(), "q", &recorder{})
++	if err == nil || !strings.Contains(err.Error(), "bob: config file") {
++		t.Errorf("missing config: err = %v", err)
++	}
++}
++
++// An agent asked through MaxDepth levels cannot ask further.
++func TestAskRespectsTheDepthLimit(t *testing.T) {
++	t.Setenv(DepthEnv, fmt.Sprint(MaxDepth))
++	if _, err := fakePool(t, "echo").Ask(context.Background(), "bob", configFile(t), t.TempDir(), "q", &recorder{}); err == nil || !strings.Contains(err.Error(), "cannot ask further") {
++		t.Errorf("err = %v", err)
++	}
++}
++
++// --- the fake agent ----------------------------------------------------------
++
++type fakeAgent struct {
++	mode     string
++	conn     *sdk.AgentSideConnection
++	sessions int
++}
++
++func (a *fakeAgent) send(ctx context.Context, sid sdk.SessionId, u sdk.SessionUpdate) {
++	_ = a.conn.SessionUpdate(ctx, sdk.SessionNotification{SessionId: sid, Update: u})
++}
++
++func (a *fakeAgent) Prompt(ctx context.Context, p sdk.PromptRequest) (sdk.PromptResponse, error) {
++	question := ""
++	if len(p.Prompt) > 0 && p.Prompt[0].Text != nil {
++		question = p.Prompt[0].Text.Text
++	}
++	switch a.mode {
++	case "echo":
++		a.send(ctx, p.SessionId, sdk.UpdateAgentMessageText(fmt.Sprintf("session %d: %s", a.sessions, question)))
++	case "slow":
++		a.send(ctx, p.SessionId, sdk.UpdateAgentMessageText("working"))
++		<-ctx.Done()
++		return sdk.PromptResponse{StopReason: sdk.StopReasonCancelled}, nil
++	case "tool":
++		a.send(ctx, p.SessionId, sdk.UpdateAgentThoughtText("hmm"))
++		a.send(ctx, p.SessionId, sdk.StartToolCall("call_1", "bash: ls", sdk.WithStartKind(sdk.ToolKindExecute), sdk.WithStartStatus(sdk.ToolCallStatusPending)))
++		resp, err := a.conn.RequestPermission(ctx, sdk.RequestPermissionRequest{
++			SessionId: p.SessionId,
++			ToolCall:  sdk.ToolCallUpdate{ToolCallId: "call_1", Title: sdk.Ptr("bash: ls")},
++			Options: []sdk.PermissionOption{
++				{Kind: sdk.PermissionOptionKindAllowOnce, Name: "Allow", OptionId: "allow"},
++				{Kind: sdk.PermissionOptionKindRejectOnce, Name: "Reject", OptionId: "reject"},
++			},
++		})
++		verdict := "rejected"
++		if err == nil && resp.Outcome.Selected != nil && resp.Outcome.Selected.OptionId == "allow" {
++			verdict = "allowed"
++			a.send(ctx, p.SessionId, sdk.UpdateToolCall("call_1", sdk.WithUpdateStatus(sdk.ToolCallStatusInProgress)))
++			a.send(ctx, p.SessionId, sdk.UpdateToolCall("call_1", sdk.WithUpdateStatus(sdk.ToolCallStatusCompleted), sdk.WithUpdateContent([]sdk.ToolCallContent{sdk.ToolContent(sdk.TextBlock("a.txt"))})))
++		}
++		a.send(ctx, p.SessionId, sdk.UpdateAgentMessageText("You asked: "+question+" (tool "+verdict+")"))
++	}
++	return sdk.PromptResponse{StopReason: sdk.StopReasonEndTurn}, nil
++}
++
++func (a *fakeAgent) NewSession(context.Context, sdk.NewSessionRequest) (sdk.NewSessionResponse, error) {
++	a.sessions++
++	return sdk.NewSessionResponse{SessionId: sdk.SessionId(fmt.Sprintf("s%d", a.sessions))}, nil
++}
++
++func (a *fakeAgent) Initialize(context.Context, sdk.InitializeRequest) (sdk.InitializeResponse, error) {
++	return sdk.InitializeResponse{ProtocolVersion: sdk.ProtocolVersionNumber}, nil
++}
++
++func (a *fakeAgent) Authenticate(context.Context, sdk.AuthenticateRequest) (sdk.AuthenticateResponse, error) {
++	return sdk.AuthenticateResponse{}, nil
++}
++func (a *fakeAgent) Logout(context.Context, sdk.LogoutRequest) (sdk.LogoutResponse, error) {
++	return sdk.LogoutResponse{}, nil
++}
++func (a *fakeAgent) Cancel(context.Context, sdk.CancelNotification) error { return nil }
++func (a *fakeAgent) CloseSession(context.Context, sdk.CloseSessionRequest) (sdk.CloseSessionResponse, error) {
++	return sdk.CloseSessionResponse{}, nil
++}
++func (a *fakeAgent) ListSessions(context.Context, sdk.ListSessionsRequest) (sdk.ListSessionsResponse, error) {
++	return sdk.ListSessionsResponse{}, nil
++}
++func (a *fakeAgent) ResumeSession(context.Context, sdk.ResumeSessionRequest) (sdk.ResumeSessionResponse, error) {
++	return sdk.ResumeSessionResponse{}, nil
++}
++func (a *fakeAgent) SetSessionConfigOption(context.Context, sdk.SetSessionConfigOptionRequest) (sdk.SetSessionConfigOptionResponse, error) {
++	return sdk.SetSessionConfigOptionResponse{}, nil
++}
++func (a *fakeAgent) SetSessionMode(context.Context, sdk.SetSessionModeRequest) (sdk.SetSessionModeResponse, error) {
++	return sdk.SetSessionModeResponse{}, nil
++}

added internal/ask/run.go +79 -0 Viewed

new file mode 100644
@@ -0,0 +1,79 @@
	1	+package ask
	2	+
	3	+import (
	4	+ "context"
	5	+ "errors"
	6	+ "fmt"
	7	+
	8	+ "mm/internal/command"
	9	+ "mm/internal/config"
	10	+ "mm/internal/ui"
	11	+
	12	+ "github.com/firebase/genkit/go/ai"
	13	+)
	14	+
	15	+// Runner is what a front end needs to run /ask: Pool in production, a fake
	16	+// in the front ends' tests.
	17	+type Runner interface {
	18	+ Run(ctx context.Context, req command.AskRequest, history []ai.Message, cwd string, sink ui.Sink) ([]ai.Message, error)
	19	+}
	20	+
	21	+var _ Runner = (*Pool)(nil)
	22	+
	23	+// Run is what the three front ends call for one /ask: it builds what the
	24	+// other agent receives (the question, and the conversation for
	25	+// /ask-with-context), asks it, and returns the exchange to append to the
	26	+// history — or an error, and then nothing to append: a failed or cancelled
	27	+// ask leaves the history as it was.
	28	+//
	29	+// The answer itself has already been shown, as it streamed, through sink.
	30	+//
	31	+// req, ok, err := command.ParseAsk(input)
	32	+// if ok && err == nil {
	33	+// exchange, err := pool.Run(ctx, req, messages, cwd, sink)
	34	+// if err == nil {
	35	+// messages = append(messages, exchange...)
	36	+// }
	37	+// }
	38	+func (p Pool) Run(ctx context.Context, req command.AskRequest, history []ai.Message, cwd string, sink ui.Sink) ([]*ai.Message, error) {
	39	+ path, ok := config.Cfg.MMConfigs[req.Agent]
	40	+ if !ok {
	41	+ return nil, fmt.Errorf("unknown agent %q", req.Agent)
	42	+ }
	43	+ transcript := ""
	44	+ if req.WithContext {
	45	+ transcript = Transcript(history, config.Cfg.MaxOutput)
	46	+ }
	47	+ answer, err := p.Ask(ctx, req.Agent, path, cwd, Prompt(req.Prompt, transcript), sink)
	48	+ if err != nil {
	49	+ return nil, err
	50	+ }
	51	+ if answer == "" {
	52	+ return nil, errors.New(req.Agent + " gave no answer")
	53	+ }
	54	+ return Exchange(req.Agent, Describe(path), req.WithContext, req.Prompt, answer, config.Cfg.MaxOutput), nil
	55	+}
	56	+
	57	+// Label is the one line a front end shows before the answer streams: who is
	58	+// asked, on what model, and how.
	59	+//
	60	+// ask.Label(req) // "🗣 asking bob (ollama/qwen3.5:2b), with the conversation…"
	61	+func Label(req command.AskRequest) string {
	62	+ line := fmt.Sprintf("🗣 asking %s (%s)", req.Agent, Describe(config.Cfg.MMConfigs[req.Agent]))
	63	+ if req.WithContext {
	64	+ line += ", with the conversation"
	65	+ }
	66	+ return line + "…"
	67	+}
	68	+
	69	+// Missing lists the configured agents whose file does not exist, for the
	70	+// start-up warnings: the agent starts anyway, and /ask-<name> says it again.
	71	+func Missing() []string {
	72	+ var warnings []string
	73	+ for _, name := range config.AgentNames() {
	74	+ if path := config.Cfg.MMConfigs[name]; !exists(path) {
	75	+ warnings = append(warnings, fmt.Sprintf("mmConfigs: %s → %s not found — /ask-%s will fail", name, path, name))
	76	+ }
	77	+ }
	78	+ return warnings
	79	+}

++package ask
++
++import (
++	"context"
++	"errors"
++	"fmt"
++
++	"mm/internal/command"
++	"mm/internal/config"
++	"mm/internal/ui"
++
++	"github.com/firebase/genkit/go/ai"
++)
++
++// Runner is what a front end needs to run /ask: Pool in production, a fake
++// in the front ends' tests.
++type Runner interface {
++	Run(ctx context.Context, req command.AskRequest, history []*ai.Message, cwd string, sink ui.Sink) ([]*ai.Message, error)
++}
++
++var _ Runner = (*Pool)(nil)
++
++// Run is what the three front ends call for one /ask: it builds what the
++// other agent receives (the question, and the conversation for
++// /ask-with-context), asks it, and returns the exchange to append to the
++// history — or an error, and then nothing to append: a failed or cancelled
++// ask leaves the history as it was.
++//
++// The answer itself has already been shown, as it streamed, through sink.
++//
++//	req, ok, err := command.ParseAsk(input)
++//	if ok && err == nil {
++//		exchange, err := pool.Run(ctx, req, messages, cwd, sink)
++//		if err == nil {
++//			messages = append(messages, exchange...)
++//		}
++//	}
++func (p *Pool) Run(ctx context.Context, req command.AskRequest, history []*ai.Message, cwd string, sink ui.Sink) ([]*ai.Message, error) {
++	path, ok := config.Cfg.MMConfigs[req.Agent]
++	if !ok {
++		return nil, fmt.Errorf("unknown agent %q", req.Agent)
++	}
++	transcript := ""
++	if req.WithContext {
++		transcript = Transcript(history, config.Cfg.MaxOutput)
++	}
++	answer, err := p.Ask(ctx, req.Agent, path, cwd, Prompt(req.Prompt, transcript), sink)
++	if err != nil {
++		return nil, err
++	}
++	if answer == "" {
++		return nil, errors.New(req.Agent + " gave no answer")
++	}
++	return Exchange(req.Agent, Describe(path), req.WithContext, req.Prompt, answer, config.Cfg.MaxOutput), nil
++}
++
++// Label is the one line a front end shows before the answer streams: who is
++// asked, on what model, and how.
++//
++//	ask.Label(req) // "🗣 asking bob (ollama/qwen3.5:2b), with the conversation…"
++func Label(req command.AskRequest) string {
++	line := fmt.Sprintf("🗣 asking %s (%s)", req.Agent, Describe(config.Cfg.MMConfigs[req.Agent]))
++	if req.WithContext {
++		line += ", with the conversation"
++	}
++	return line + "…"
++}
++
++// Missing lists the configured agents whose file does not exist, for the
++// start-up warnings: the agent starts anyway, and /ask-<name> says it again.
++func Missing() []string {
++	var warnings []string
++	for _, name := range config.AgentNames() {
++		if path := config.Cfg.MMConfigs[name]; !exists(path) {
++			warnings = append(warnings, fmt.Sprintf("mmConfigs: %s → %s not found — /ask-%s will fail", name, path, name))
++		}
++	}
++	return warnings
++}

added internal/ask/run_test.go +83 -0 Viewed

new file mode 100644
@@ -0,0 +1,83 @@
	1	+package ask
	2	+
	3	+import (
	4	+ "context"
	5	+ "os"
	6	+ "path/filepath"
	7	+ "strings"
	8	+ "testing"
	9	+
	10	+ "mm/internal/command"
	11	+ "mm/internal/config"
	12	+
	13	+ "github.com/firebase/genkit/go/ai"
	14	+)
	15	+
	16	+// withAgent configures one agent, bob, whose file exists.
	17	+func withAgent(t *testing.T) {
	18	+ t.Helper()
	19	+ saved := config.Cfg
	20	+ t.Cleanup(func() { config.Cfg = saved })
	21	+ path := filepath.Join(t.TempDir(), "agent.bob.yaml")
	22	+ if err := os.WriteFile(path, []byte("provider: ollama\nmodel: qwen3.5:2b\n"), 0o644); err != nil {
	23	+ t.Fatal(err)
	24	+ }
	25	+ config.Cfg.MMConfigs = map[string]string{"bob": path}
	26	+}
	27	+
	28	+// A plain ask sends the question alone and returns the marked pair.
	29	+func TestRunReturnsTheExchange(t *testing.T) {
	30	+ withAgent(t)
	31	+ history := []*ai.Message{ai.NewSystemTextMessage("s"), ai.NewUserTextMessage("PREVIOUS-QUESTION")}
	32	+ got, err := fakePool(t, "echo").Run(context.Background(), command.AskRequest{Agent: "bob", Prompt: "hi"}, history, t.TempDir(), &recorder{})
	33	+ if err != nil {
	34	+ t.Fatal(err)
	35	+ }
	36	+ if len(got) != 2 \|\| got[0].Text() != "[asked bob (ollama/qwen3.5:2b): hi]" \|\| got[1].Text() != "[bob's answer]\nsession 1: hi" {
	37	+ t.Errorf("exchange = %q / %q", got[0].Text(), got[1].Text())
	38	+ }
	39	+}
	40	+
	41	+// /ask-with-context sends the transcript before the question.
	42	+func TestRunWithContextSendsTheConversation(t *testing.T) {
	43	+ withAgent(t)
	44	+ history := []*ai.Message{ai.NewSystemTextMessage("s"), ai.NewUserTextMessage("PREVIOUS-QUESTION")}
	45	+ got, err := fakePool(t, "echo").Run(context.Background(), command.AskRequest{Agent: "bob", WithContext: true, Prompt: "hi"}, history, t.TempDir(), &recorder{})
	46	+ if err != nil {
	47	+ t.Fatal(err)
	48	+ }
	49	+ answer := got[1].Text()
	50	+ if !strings.Contains(answer, "User: PREVIOUS-QUESTION") \|\| !strings.Contains(answer, "[Question]\nhi") {
	51	+ t.Errorf("the child received:\n%s", answer)
	52	+ }
	53	+ if !strings.Contains(got[0].Text(), "with this conversation as context") {
	54	+ t.Errorf("question marker = %q", got[0].Text())
	55	+ }
	56	+}
	57	+
	58	+// A failed ask leaves nothing to append; an unknown agent is refused before
	59	+// anything starts; a missing file is listed for the start-up warnings.
	60	+func TestRunFailures(t *testing.T) {
	61	+ withAgent(t)
	62	+ if got, err := fakePool(t, "crash").Run(context.Background(), command.AskRequest{Agent: "bob", Prompt: "hi"}, nil, t.TempDir(), &recorder{}); err == nil \|\| got != nil {
	63	+ t.Errorf("crash: %v, %v — want an error and no exchange", got, err)
	64	+ }
	65	+ if _, err := fakePool(t, "echo").Run(context.Background(), command.AskRequest{Agent: "spock", Prompt: "hi"}, nil, t.TempDir(), &recorder{}); err == nil {
	66	+ t.Error("unknown agent accepted")
	67	+ }
	68	+ if len(Missing()) != 0 {
	69	+ t.Errorf("Missing() = %v with an existing file", Missing())
	70	+ }
	71	+ config.Cfg.MMConfigs["riker"] = filepath.Join(t.TempDir(), "nope.yaml")
	72	+ if m := Missing(); len(m) != 1 \|\| !strings.Contains(m[0], "riker") {
	73	+ t.Errorf("Missing() = %v", m)
	74	+ }
	75	+}
	76	+
	77	+// The label names the agent, its model, and the context when sent.
	78	+func TestLabel(t *testing.T) {
	79	+ withAgent(t)
	80	+ if got := Label(command.AskRequest{Agent: "bob", WithContext: true}); got != "🗣 asking bob (ollama/qwen3.5:2b), with the conversation…" {
	81	+ t.Errorf("Label = %q", got)
	82	+ }
	83	+}

++package ask
++
++import (
++	"context"
++	"os"
++	"path/filepath"
++	"strings"
++	"testing"
++
++	"mm/internal/command"
++	"mm/internal/config"
++
++	"github.com/firebase/genkit/go/ai"
++)
++
++// withAgent configures one agent, bob, whose file exists.
++func withAgent(t *testing.T) {
++	t.Helper()
++	saved := config.Cfg
++	t.Cleanup(func() { config.Cfg = saved })
++	path := filepath.Join(t.TempDir(), "agent.bob.yaml")
++	if err := os.WriteFile(path, []byte("provider: ollama\nmodel: qwen3.5:2b\n"), 0o644); err != nil {
++		t.Fatal(err)
++	}
++	config.Cfg.MMConfigs = map[string]string{"bob": path}
++}
++
++// A plain ask sends the question alone and returns the marked pair.
++func TestRunReturnsTheExchange(t *testing.T) {
++	withAgent(t)
++	history := []*ai.Message{ai.NewSystemTextMessage("s"), ai.NewUserTextMessage("PREVIOUS-QUESTION")}
++	got, err := fakePool(t, "echo").Run(context.Background(), command.AskRequest{Agent: "bob", Prompt: "hi"}, history, t.TempDir(), &recorder{})
++	if err != nil {
++		t.Fatal(err)
++	}
++	if len(got) != 2 || got[0].Text() != "[asked bob (ollama/qwen3.5:2b): hi]" || got[1].Text() != "[bob's answer]\nsession 1: hi" {
++		t.Errorf("exchange = %q / %q", got[0].Text(), got[1].Text())
++	}
++}
++
++// /ask-with-context sends the transcript before the question.
++func TestRunWithContextSendsTheConversation(t *testing.T) {
++	withAgent(t)
++	history := []*ai.Message{ai.NewSystemTextMessage("s"), ai.NewUserTextMessage("PREVIOUS-QUESTION")}
++	got, err := fakePool(t, "echo").Run(context.Background(), command.AskRequest{Agent: "bob", WithContext: true, Prompt: "hi"}, history, t.TempDir(), &recorder{})
++	if err != nil {
++		t.Fatal(err)
++	}
++	answer := got[1].Text()
++	if !strings.Contains(answer, "User: PREVIOUS-QUESTION") || !strings.Contains(answer, "[Question]\nhi") {
++		t.Errorf("the child received:\n%s", answer)
++	}
++	if !strings.Contains(got[0].Text(), "with this conversation as context") {
++		t.Errorf("question marker = %q", got[0].Text())
++	}
++}
++
++// A failed ask leaves nothing to append; an unknown agent is refused before
++// anything starts; a missing file is listed for the start-up warnings.
++func TestRunFailures(t *testing.T) {
++	withAgent(t)
++	if got, err := fakePool(t, "crash").Run(context.Background(), command.AskRequest{Agent: "bob", Prompt: "hi"}, nil, t.TempDir(), &recorder{}); err == nil || got != nil {
++		t.Errorf("crash: %v, %v — want an error and no exchange", got, err)
++	}
++	if _, err := fakePool(t, "echo").Run(context.Background(), command.AskRequest{Agent: "spock", Prompt: "hi"}, nil, t.TempDir(), &recorder{}); err == nil {
++		t.Error("unknown agent accepted")
++	}
++	if len(Missing()) != 0 {
++		t.Errorf("Missing() = %v with an existing file", Missing())
++	}
++	config.Cfg.MMConfigs["riker"] = filepath.Join(t.TempDir(), "nope.yaml")
++	if m := Missing(); len(m) != 1 || !strings.Contains(m[0], "riker") {
++		t.Errorf("Missing() = %v", m)
++	}
++}
++
++// The label names the agent, its model, and the context when sent.
++func TestLabel(t *testing.T) {
++	withAgent(t)
++	if got := Label(command.AskRequest{Agent: "bob", WithContext: true}); got != "🗣 asking bob (ollama/qwen3.5:2b), with the conversation…" {
++		t.Errorf("Label = %q", got)
++	}
++}

added internal/command/ask.go +106 -0 Viewed

new file mode 100644
@@ -0,0 +1,106 @@
	1	+package command
	2	+
	3	+import (
	4	+ "errors"
	5	+ "fmt"
	6	+ "slices"
	7	+ "strings"
	8	+)
	9	+
	10	+// The /ask family: ask another mm agent (config `mmConfigs`), with or without
	11	+// the current conversation as context. Generated by SetAgents, because the
	12	+// per-agent commands are named after the configured agents:
	13	+//
	14	+// /ask <name> <prompt> (terminal and window)
	15	+// /ask-with-context <name> <prompt> (terminal and window)
	16	+// /ask-<name> <prompt> (every front end, the editor included)
	17	+// /ask-with-context-<name> <prompt> (every front end)
	18	+//
	19	+// The editor gets the per-agent forms only: its `/` menu completes a command
	20	+// name, never an argument, so "/ask-bob" is the one way to list the agents
	21	+// there.
	22	+const (
	23	+ Ask = "ask"
	24	+ AskWithContext = "ask-with-context"
	25	+)
	26	+
	27	+// askCommands is the generated part of the table; empty when no agent is
	28	+// configured, so /help and the menus show no /ask at all.
	29	+var askCommands []Command
	30	+
	31	+// SetAgents generates the /ask family for the given agent names (sorted by
	32	+// the caller — config.AgentNames). Called once by main after the config is
	33	+// loaded; nil or empty removes the family.
	34	+//
	35	+// command.SetAgents([]string{"bob", "riker"})
	36	+// command.For(command.Editor) // … ask-bob, ask-with-context-bob, ask-riker, …
	37	+func SetAgents(names []string) {
	38	+ askCommands = nil
	39	+ if len(names) == 0 {
	40	+ return
	41	+ }
	42	+ list := strings.Join(names, "\|")
	43	+ askCommands = append(askCommands,
	44	+ Command{Name: Ask, Description: "Ask another mm agent (" + list + "); its answer stays in the history", Fronts: Terminal \| Window, TakesArg: true, Hint: "<agent> <prompt>"},
	45	+ Command{Name: AskWithContext, Description: "Same, sending the current conversation along", Fronts: Terminal \| Window, TakesArg: true, Hint: "<agent> <prompt>"},
	46	+ )
	47	+ for _, name := range names {
	48	+ askCommands = append(askCommands,
	49	+ Command{Name: Ask + "-" + name, Description: "Ask " + name + " (another mm agent); its answer stays in the history", Fronts: Terminal \| Editor \| Window, TakesArg: true, Hint: "<prompt>"},
	50	+ Command{Name: AskWithContext + "-" + name, Description: "Ask " + name + ", sending the current conversation along", Fronts: Terminal \| Editor \| Window, TakesArg: true, Hint: "<prompt>"},
	51	+ )
	52	+ }
	53	+}
	54	+
	55	+// AskRequest is one /ask, parsed.
	56	+type AskRequest struct {
	57	+ Agent string // a configured agent name
	58	+ WithContext bool // send the current conversation along
	59	+ Prompt string // what to ask, trimmed, never empty
	60	+}
	61	+
	62	+// ParseAsk recognises the /ask family. ok says the input IS an /ask command
	63	+// (whatever its validity); err then says what is wrong with it — an unknown
	64	+// agent, a missing prompt — in words the front end shows as is.
	65	+//
	66	+// req, ok, err := command.ParseAsk("/ask-bob what is 2+2?")
	67	+// // req == AskRequest{Agent: "bob", Prompt: "what is 2+2?"}, ok, err == nil
	68	+func ParseAsk(input string) (req AskRequest, ok bool, err error) {
	69	+ cmd, arg, found := Parse(input)
	70	+ if !found {
	71	+ return AskRequest{}, false, nil
	72	+ }
	73	+ switch {
	74	+ case cmd.Name == Ask \|\| cmd.Name == AskWithContext:
	75	+ req.WithContext = cmd.Name == AskWithContext
	76	+ req.Agent, req.Prompt, _ = strings.Cut(arg, " ")
	77	+ if req.Agent == "" {
	78	+ return req, true, fmt.Errorf("usage: /%s <agent> <prompt> — agents: %s", cmd.Name, strings.Join(agentNames(), ", "))
	79	+ }
	80	+ if !slices.Contains(agentNames(), req.Agent) {
	81	+ return req, true, fmt.Errorf("unknown agent %q — agents: %s", req.Agent, strings.Join(agentNames(), ", "))
	82	+ }
	83	+ case strings.HasPrefix(cmd.Name, AskWithContext+"-"):
	84	+ req.WithContext, req.Agent, req.Prompt = true, strings.TrimPrefix(cmd.Name, AskWithContext+"-"), arg
	85	+ case strings.HasPrefix(cmd.Name, Ask+"-"):
	86	+ req.Agent, req.Prompt = strings.TrimPrefix(cmd.Name, Ask+"-"), arg
	87	+ default:
	88	+ return AskRequest{}, false, nil
	89	+ }
	90	+ req.Prompt = strings.TrimSpace(req.Prompt)
	91	+ if req.Prompt == "" {
	92	+ return req, true, errors.New("nothing to ask: type the question after the command")
	93	+ }
	94	+ return req, true, nil
	95	+}
	96	+
	97	+// agentNames lists the agents SetAgents was given, in its order.
	98	+func agentNames() []string {
	99	+ var names []string
	100	+ for _, c := range askCommands {
	101	+ if name, found := strings.CutPrefix(c.Name, Ask+"-"); found && !strings.HasPrefix(c.Name, AskWithContext) {
	102	+ names = append(names, name)
	103	+ }
	104	+ }
	105	+ return names
	106	+}

++package command
++
++import (
++	"errors"
++	"fmt"
++	"slices"
++	"strings"
++)
++
++// The /ask family: ask another mm agent (config `mmConfigs`), with or without
++// the current conversation as context. Generated by SetAgents, because the
++// per-agent commands are named after the configured agents:
++//
++//	/ask <name> <prompt>                (terminal and window)
++//	/ask-with-context <name> <prompt>   (terminal and window)
++//	/ask-<name> <prompt>                (every front end, the editor included)
++//	/ask-with-context-<name> <prompt>   (every front end)
++//
++// The editor gets the per-agent forms only: its `/` menu completes a command
++// name, never an argument, so "/ask-bob" is the one way to list the agents
++// there.
++const (
++	Ask            = "ask"
++	AskWithContext = "ask-with-context"
++)
++
++// askCommands is the generated part of the table; empty when no agent is
++// configured, so /help and the menus show no /ask at all.
++var askCommands []Command
++
++// SetAgents generates the /ask family for the given agent names (sorted by
++// the caller — config.AgentNames). Called once by main after the config is
++// loaded; nil or empty removes the family.
++//
++//	command.SetAgents([]string{"bob", "riker"})
++//	command.For(command.Editor) // … ask-bob, ask-with-context-bob, ask-riker, …
++func SetAgents(names []string) {
++	askCommands = nil
++	if len(names) == 0 {
++		return
++	}
++	list := strings.Join(names, "|")
++	askCommands = append(askCommands,
++		Command{Name: Ask, Description: "Ask another mm agent (" + list + "); its answer stays in the history", Fronts: Terminal | Window, TakesArg: true, Hint: "<agent> <prompt>"},
++		Command{Name: AskWithContext, Description: "Same, sending the current conversation along", Fronts: Terminal | Window, TakesArg: true, Hint: "<agent> <prompt>"},
++	)
++	for _, name := range names {
++		askCommands = append(askCommands,
++			Command{Name: Ask + "-" + name, Description: "Ask " + name + " (another mm agent); its answer stays in the history", Fronts: Terminal | Editor | Window, TakesArg: true, Hint: "<prompt>"},
++			Command{Name: AskWithContext + "-" + name, Description: "Ask " + name + ", sending the current conversation along", Fronts: Terminal | Editor | Window, TakesArg: true, Hint: "<prompt>"},
++		)
++	}
++}
++
++// AskRequest is one /ask, parsed.
++type AskRequest struct {
++	Agent       string // a configured agent name
++	WithContext bool   // send the current conversation along
++	Prompt      string // what to ask, trimmed, never empty
++}
++
++// ParseAsk recognises the /ask family. ok says the input IS an /ask command
++// (whatever its validity); err then says what is wrong with it — an unknown
++// agent, a missing prompt — in words the front end shows as is.
++//
++//	req, ok, err := command.ParseAsk("/ask-bob what is 2+2?")
++//	// req == AskRequest{Agent: "bob", Prompt: "what is 2+2?"}, ok, err == nil
++func ParseAsk(input string) (req AskRequest, ok bool, err error) {
++	cmd, arg, found := Parse(input)
++	if !found {
++		return AskRequest{}, false, nil
++	}
++	switch {
++	case cmd.Name == Ask || cmd.Name == AskWithContext:
++		req.WithContext = cmd.Name == AskWithContext
++		req.Agent, req.Prompt, _ = strings.Cut(arg, " ")
++		if req.Agent == "" {
++			return req, true, fmt.Errorf("usage: /%s <agent> <prompt> — agents: %s", cmd.Name, strings.Join(agentNames(), ", "))
++		}
++		if !slices.Contains(agentNames(), req.Agent) {
++			return req, true, fmt.Errorf("unknown agent %q — agents: %s", req.Agent, strings.Join(agentNames(), ", "))
++		}
++	case strings.HasPrefix(cmd.Name, AskWithContext+"-"):
++		req.WithContext, req.Agent, req.Prompt = true, strings.TrimPrefix(cmd.Name, AskWithContext+"-"), arg
++	case strings.HasPrefix(cmd.Name, Ask+"-"):
++		req.Agent, req.Prompt = strings.TrimPrefix(cmd.Name, Ask+"-"), arg
++	default:
++		return AskRequest{}, false, nil
++	}
++	req.Prompt = strings.TrimSpace(req.Prompt)
++	if req.Prompt == "" {
++		return req, true, errors.New("nothing to ask: type the question after the command")
++	}
++	return req, true, nil
++}
++
++// agentNames lists the agents SetAgents was given, in its order.
++func agentNames() []string {
++	var names []string
++	for _, c := range askCommands {
++		if name, found := strings.CutPrefix(c.Name, Ask+"-"); found && !strings.HasPrefix(c.Name, AskWithContext) {
++			names = append(names, name)
++		}
++	}
++	return names
++}

added internal/command/ask_test.go +99 -0 Viewed

new file mode 100644
@@ -0,0 +1,99 @@
	1	+package command
	2	+
	3	+import (
	4	+ "strings"
	5	+ "testing"
	6	+)
	7	+
	8	+// withAgents installs an /ask family for one test and removes it afterwards:
	9	+// the generated commands are package state, shared by every test.
	10	+func withAgents(t *testing.T, names ...string) {
	11	+ t.Helper()
	12	+ SetAgents(names)
	13	+ t.Cleanup(func() { SetAgents(nil) })
	14	+}
	15	+
	16	+// No agent configured: no /ask anywhere, so /help and the menus stay as they
	17	+// were and "/ask bob hi" is a question for the model.
	18	+func TestNoAgentsNoAskCommands(t *testing.T) {
	19	+ SetAgents(nil)
	20	+ for _, c := range All() {
	21	+ if strings.HasPrefix(c.Name, Ask) {
	22	+ t.Errorf("unexpected %s with no agents", c.Slash())
	23	+ }
	24	+ }
	25	+ if _, ok, _ := ParseAsk("/ask bob hi"); ok {
	26	+ t.Error("ParseAsk recognised /ask with no agents configured")
	27	+ }
	28	+}
	29	+
	30	+// The editor gets the per-agent forms only (its menu completes names, not
	31	+// arguments); the terminal and the window get the generic forms too.
	32	+func TestAskCommandsPerFront(t *testing.T) {
	33	+ withAgents(t, "bob", "riker")
	34	+ names := func(f Front) []string {
	35	+ var out []string
	36	+ for _, c := range For(f) {
	37	+ if strings.HasPrefix(c.Name, Ask) {
	38	+ out = append(out, c.Name)
	39	+ }
	40	+ }
	41	+ return out
	42	+ }
	43	+ perAgent := "ask-bob ask-with-context-bob ask-riker ask-with-context-riker"
	44	+ if got := strings.Join(names(Editor), " "); got != perAgent {
	45	+ t.Errorf("Editor = %q, want %q", got, perAgent)
	46	+ }
	47	+ want := "ask ask-with-context " + perAgent
	48	+ for _, f := range []Front{Terminal, Window} {
	49	+ if got := strings.Join(names(f), " "); got != want {
	50	+ t.Errorf("front %d = %q, want %q", f, got, want)
	51	+ }
	52	+ }
	53	+ for _, c := range For(Editor) {
	54	+ if strings.HasPrefix(c.Name, Ask) && (c.Hint != "<prompt>" \|\| !c.TakesArg) {
	55	+ t.Errorf("%s: Hint %q TakesArg %v, want <prompt> and true", c.Slash(), c.Hint, c.TakesArg)
	56	+ }
	57	+ }
	58	+ if !strings.Contains(Usage(Terminal), "/ask-bob <prompt>") \|\| !strings.Contains(Usage(Terminal), "/ask <agent> <prompt>") {
	59	+ t.Errorf("Usage does not show the hints:\n%s", Usage(Terminal))
	60	+ }
	61	+}
	62	+
	63	+// Every accepted form, and every error the front ends show as is.
	64	+func TestParseAsk(t *testing.T) {
	65	+ withAgents(t, "bob", "riker")
	66	+ for _, tc := range []struct {
	67	+ in string
	68	+ ok bool
	69	+ want AskRequest
	70	+ wantErr string
	71	+ }{
	72	+ {"/ask-bob what is 2+2?", true, AskRequest{Agent: "bob", Prompt: "what is 2+2?"}, ""},
	73	+ {" /ask-with-context-riker review it ", true, AskRequest{Agent: "riker", WithContext: true, Prompt: "review it"}, ""},
	74	+ {"/ask bob what is 2+2?", true, AskRequest{Agent: "bob", Prompt: "what is 2+2?"}, ""},
	75	+ {"/ask-with-context riker review it", true, AskRequest{Agent: "riker", WithContext: true, Prompt: "review it"}, ""},
	76	+ {"/ask-bob", true, AskRequest{Agent: "bob"}, "nothing to ask"},
	77	+ {"/ask bob", true, AskRequest{Agent: "bob"}, "nothing to ask"},
	78	+ {"/ask", true, AskRequest{}, "usage: /ask <agent> <prompt> — agents: bob, riker"},
	79	+ {"/ask spock hi", true, AskRequest{Agent: "spock", Prompt: "hi"}, `unknown agent "spock" — agents: bob, riker`},
	80	+ {"/ask-spock hi", false, AskRequest{}, ""},
	81	+ {"/compact", false, AskRequest{}, ""},
	82	+ {"ask bob hi", false, AskRequest{}, ""},
	83	+ } {
	84	+ got, ok, err := ParseAsk(tc.in)
	85	+ if ok != tc.ok {
	86	+ t.Errorf("%q: ok = %v, want %v", tc.in, ok, tc.ok)
	87	+ continue
	88	+ }
	89	+ if tc.wantErr == "" && err != nil {
	90	+ t.Errorf("%q: unexpected error %v", tc.in, err)
	91	+ }
	92	+ if tc.wantErr != "" && (err == nil \|\| !strings.Contains(err.Error(), tc.wantErr)) {
	93	+ t.Errorf("%q: error %v, want one containing %q", tc.in, err, tc.wantErr)
	94	+ }
	95	+ if ok && got != tc.want {
	96	+ t.Errorf("%q: %+v, want %+v", tc.in, got, tc.want)
	97	+ }
	98	+ }
	99	+}

++package command
++
++import (
++	"strings"
++	"testing"
++)
++
++// withAgents installs an /ask family for one test and removes it afterwards:
++// the generated commands are package state, shared by every test.
++func withAgents(t *testing.T, names ...string) {
++	t.Helper()
++	SetAgents(names)
++	t.Cleanup(func() { SetAgents(nil) })
++}
++
++// No agent configured: no /ask anywhere, so /help and the menus stay as they
++// were and "/ask bob hi" is a question for the model.
++func TestNoAgentsNoAskCommands(t *testing.T) {
++	SetAgents(nil)
++	for _, c := range All() {
++		if strings.HasPrefix(c.Name, Ask) {
++			t.Errorf("unexpected %s with no agents", c.Slash())
++		}
++	}
++	if _, ok, _ := ParseAsk("/ask bob hi"); ok {
++		t.Error("ParseAsk recognised /ask with no agents configured")
++	}
++}
++
++// The editor gets the per-agent forms only (its menu completes names, not
++// arguments); the terminal and the window get the generic forms too.
++func TestAskCommandsPerFront(t *testing.T) {
++	withAgents(t, "bob", "riker")
++	names := func(f Front) []string {
++		var out []string
++		for _, c := range For(f) {
++			if strings.HasPrefix(c.Name, Ask) {
++				out = append(out, c.Name)
++			}
++		}
++		return out
++	}
++	perAgent := "ask-bob ask-with-context-bob ask-riker ask-with-context-riker"
++	if got := strings.Join(names(Editor), " "); got != perAgent {
++		t.Errorf("Editor = %q, want %q", got, perAgent)
++	}
++	want := "ask ask-with-context " + perAgent
++	for _, f := range []Front{Terminal, Window} {
++		if got := strings.Join(names(f), " "); got != want {
++			t.Errorf("front %d = %q, want %q", f, got, want)
++		}
++	}
++	for _, c := range For(Editor) {
++		if strings.HasPrefix(c.Name, Ask) && (c.Hint != "<prompt>" || !c.TakesArg) {
++			t.Errorf("%s: Hint %q TakesArg %v, want <prompt> and true", c.Slash(), c.Hint, c.TakesArg)
++		}
++	}
++	if !strings.Contains(Usage(Terminal), "/ask-bob <prompt>") || !strings.Contains(Usage(Terminal), "/ask <agent> <prompt>") {
++		t.Errorf("Usage does not show the hints:\n%s", Usage(Terminal))
++	}
++}
++
++// Every accepted form, and every error the front ends show as is.
++func TestParseAsk(t *testing.T) {
++	withAgents(t, "bob", "riker")
++	for _, tc := range []struct {
++		in      string
++		ok      bool
++		want    AskRequest
++		wantErr string
++	}{
++		{"/ask-bob what is 2+2?", true, AskRequest{Agent: "bob", Prompt: "what is 2+2?"}, ""},
++		{"  /ask-with-context-riker  review it  ", true, AskRequest{Agent: "riker", WithContext: true, Prompt: "review it"}, ""},
++		{"/ask bob what is 2+2?", true, AskRequest{Agent: "bob", Prompt: "what is 2+2?"}, ""},
++		{"/ask-with-context riker review it", true, AskRequest{Agent: "riker", WithContext: true, Prompt: "review it"}, ""},
++		{"/ask-bob", true, AskRequest{Agent: "bob"}, "nothing to ask"},
++		{"/ask bob", true, AskRequest{Agent: "bob"}, "nothing to ask"},
++		{"/ask", true, AskRequest{}, "usage: /ask <agent> <prompt> — agents: bob, riker"},
++		{"/ask spock hi", true, AskRequest{Agent: "spock", Prompt: "hi"}, `unknown agent "spock" — agents: bob, riker`},
++		{"/ask-spock hi", false, AskRequest{}, ""},
++		{"/compact", false, AskRequest{}, ""},
++		{"ask bob hi", false, AskRequest{}, ""},
++	} {
++		got, ok, err := ParseAsk(tc.in)
++		if ok != tc.ok {
++			t.Errorf("%q: ok = %v, want %v", tc.in, ok, tc.ok)
++			continue
++		}
++		if tc.wantErr == "" && err != nil {
++			t.Errorf("%q: unexpected error %v", tc.in, err)
++		}
++		if tc.wantErr != "" && (err == nil || !strings.Contains(err.Error(), tc.wantErr)) {
++			t.Errorf("%q: error %v, want one containing %q", tc.in, err, tc.wantErr)
++		}
++		if ok && got != tc.want {
++			t.Errorf("%q: %+v, want %+v", tc.in, got, tc.want)
++		}
++	}
++}

modified internal/command/command.go +11 -4 Viewed

@@ -51,6 +51,9 @@ type Command struct {
51	51	// TakesArg says "/name something" is still this command ("/resume <id>");
52	52	// without it, "/compact now" is a question for the model, as it always was.
53	53	TakesArg bool
	54	+ // Hint names the argument where a usage line or an editor shows it
	55	+ // ("<prompt>"); "" = the generic "[arg]".
	56	+ Hint string
54	57	}
55	58
56	59	// Slash is the command as typed.
@@ -71,13 +74,14 @@ var table = []Command{
71	74	{Name: Copy, Description: "Copy the last answer to the clipboard (/copy 2: the one before; /copy all: the whole conversation)", Fronts: Terminal \| Window, TakesArg: true},
72	75	}
73	76
74		-// All returns the whole table, in display order.
75		-func All() []Command { return append([]Command(nil), table...) }
	77	+// All returns the whole table, in display order: the fixed commands, then
	78	+// the /ask family generated from the configured agents (ask.go).
	79	+func All() []Command { return append(append([]Command(nil), table...), askCommands...) }
76	80
77	81	// For returns the commands front f offers, in display order.
78	82	func For(f Front) []Command {
79	83	var out []Command
80		- for _, c := range table {
	84	+ for _, c := range All() {
81	85	if c.Offered(f) {
82	86	out = append(out, c)
83	87	}
@@ -101,7 +105,7 @@ func Parse(input string) (cmd Command, arg string, ok bool) {
101	105	return Command{}, "", false
102	106	}
103	107	name, rest, _ := strings.Cut(input[1:], " ")
104		- for _, c := range table {
	108	+ for _, c := range All() {
105	109	if c.Name != name {
106	110	continue
107	111	}
@@ -142,6 +146,9 @@ func Usage(f Front) string {
142	146
143	147	// usage is how a command is typed, with its argument when it takes one.
144	148	func usage(c Command) string {
	149	+ if c.Hint != "" {
	150	+ return c.Slash() + " " + c.Hint
	151	+ }
145	152	if c.TakesArg {
146	153	return c.Slash() + " [arg]"
147	154	}

  	// TakesArg says "/name something" is still this command ("/resume <id>");
  	// without it, "/compact now" is a question for the model, as it always was.
  	TakesArg bool
++	// Hint names the argument where a usage line or an editor shows it
++	// ("<prompt>"); "" = the generic "[arg]".
++	Hint string
  }
  // Slash is the command as typed.
  	{Name: Copy, Description: "Copy the last answer to the clipboard (/copy 2: the one before; /copy all: the whole conversation)", Fronts: Terminal | Window, TakesArg: true},
  }
--// All returns the whole table, in display order.
++// All returns the whole table, in display order: the fixed commands, then
--func All() []Command { return append([]Command(nil), table...) }
++// the /ask family generated from the configured agents (ask.go).
++func All() []Command { return append(append([]Command(nil), table...), askCommands...) }
  // For returns the commands front f offers, in display order.
  func For(f Front) []Command {
  	var out []Command
--	for _, c := range table {
++	for _, c := range All() {
  		if c.Offered(f) {
  			out = append(out, c)
  		}
  		return Command{}, "", false
  	}
  	name, rest, _ := strings.Cut(input[1:], " ")
--	for _, c := range table {
++	for _, c := range All() {
  		if c.Name != name {
  			continue
  		}
  // usage is how a command is typed, with its argument when it takes one.
  func usage(c Command) string {
++	if c.Hint != "" {
++		return c.Slash() + " " + c.Hint
++	}
  	if c.TakesArg {
  		return c.Slash() + " [arg]"
  	}

added internal/config/agents.go +68 -0 Viewed

new file mode 100644
@@ -0,0 +1,68 @@
	1	+package config
	2	+
	3	+import (
	4	+ "fmt"
	5	+ "path/filepath"
	6	+ "regexp"
	7	+ "sort"
	8	+ "strings"
	9	+)
	10	+
	11	+// SessionsEnv names the variable that turns session persistence off for one
	12	+// run ("off"). The /ask machinery sets it on the agents it starts.
	13	+const SessionsEnv = "AGENT_SESSIONS"
	14	+
	15	+// agentName is the format of an mmConfigs name: it becomes the tail of a
	16	+// command name (/ask-bob), so it must be typable, unambiguous after a dash,
	17	+// and safe in an editor's `/` menu.
	18	+var agentName = regexp.MustCompile(`^[a-z0-9][a-z0-9-]*$`)
	19	+
	20	+// validateAgents checks the mmConfigs names and that each has a path. The
	21	+// files themselves are NOT required to exist: a missing one is a start-up
	22	+// warning (main), and /ask-<name> then says so — one absent sibling config
	23	+// must not stop the agent from starting.
	24	+func (c Config) validateAgents() error {
	25	+ for _, name := range AgentNames() {
	26	+ if !agentName.MatchString(name) {
	27	+ return fmt.Errorf("mmConfigs: %q is not a valid name (lowercase letters, digits and '-', starting with a letter or a digit)", name)
	28	+ }
	29	+ // "/ask-with-context-x" must mean one thing: the context form for x.
	30	+ if strings.HasPrefix(name, "with-context") {
	31	+ return fmt.Errorf("mmConfigs: %q is not a valid name (it would read as /ask-with-context-…)", name)
	32	+ }
	33	+ if c.MMConfigs[name] == "" {
	34	+ return fmt.Errorf("mmConfigs.%s: the path must not be empty", name)
	35	+ }
	36	+ }
	37	+ return nil
	38	+}
	39	+
	40	+// resolveAgentPaths makes every relative mmConfigs path absolute against the
	41	+// directory of the config file that declared it — the one anchor that does
	42	+// not depend on who started mm (a terminal, an editor, another mm).
	43	+func resolveAgentPaths(configPath string) {
	44	+ dir := filepath.Dir(configPath)
	45	+ for name, p := range Cfg.MMConfigs {
	46	+ if p == "" \|\| filepath.IsAbs(p) {
	47	+ continue
	48	+ }
	49	+ if abs, err := filepath.Abs(filepath.Join(dir, p)); err == nil {
	50	+ Cfg.MMConfigs[name] = abs
	51	+ }
	52	+ }
	53	+}
	54	+
	55	+// AgentNames returns the mmConfigs names, sorted: the order of /help, of the
	56	+// editor's menu and of the window's picker.
	57	+//
	58	+// for _, name := range config.AgentNames() {
	59	+// fmt.Println("/ask-" + name, config.Cfg.MMConfigs[name])
	60	+// }
	61	+func AgentNames() []string {
	62	+ names := make([]string, 0, len(Cfg.MMConfigs))
	63	+ for name := range Cfg.MMConfigs {
	64	+ names = append(names, name)
	65	+ }
	66	+ sort.Strings(names)
	67	+ return names
	68	+}

++package config
++
++import (
++	"fmt"
++	"path/filepath"
++	"regexp"
++	"sort"
++	"strings"
++)
++
++// SessionsEnv names the variable that turns session persistence off for one
++// run ("off"). The /ask machinery sets it on the agents it starts.
++const SessionsEnv = "AGENT_SESSIONS"
++
++// agentName is the format of an mmConfigs name: it becomes the tail of a
++// command name (/ask-bob), so it must be typable, unambiguous after a dash,
++// and safe in an editor's `/` menu.
++var agentName = regexp.MustCompile(`^[a-z0-9][a-z0-9-]*$`)
++
++// validateAgents checks the mmConfigs names and that each has a path. The
++// files themselves are NOT required to exist: a missing one is a start-up
++// warning (main), and /ask-<name> then says so — one absent sibling config
++// must not stop the agent from starting.
++func (c Config) validateAgents() error {
++	for _, name := range AgentNames() {
++		if !agentName.MatchString(name) {
++			return fmt.Errorf("mmConfigs: %q is not a valid name (lowercase letters, digits and '-', starting with a letter or a digit)", name)
++		}
++		// "/ask-with-context-x" must mean one thing: the context form for x.
++		if strings.HasPrefix(name, "with-context") {
++			return fmt.Errorf("mmConfigs: %q is not a valid name (it would read as /ask-with-context-…)", name)
++		}
++		if c.MMConfigs[name] == "" {
++			return fmt.Errorf("mmConfigs.%s: the path must not be empty", name)
++		}
++	}
++	return nil
++}
++
++// resolveAgentPaths makes every relative mmConfigs path absolute against the
++// directory of the config file that declared it — the one anchor that does
++// not depend on who started mm (a terminal, an editor, another mm).
++func resolveAgentPaths(configPath string) {
++	dir := filepath.Dir(configPath)
++	for name, p := range Cfg.MMConfigs {
++		if p == "" || filepath.IsAbs(p) {
++			continue
++		}
++		if abs, err := filepath.Abs(filepath.Join(dir, p)); err == nil {
++			Cfg.MMConfigs[name] = abs
++		}
++	}
++}
++
++// AgentNames returns the mmConfigs names, sorted: the order of /help, of the
++// editor's menu and of the window's picker.
++//
++//	for _, name := range config.AgentNames() {
++//		fmt.Println("/ask-" + name, config.Cfg.MMConfigs[name])
++//	}
++func AgentNames() []string {
++	names := make([]string, 0, len(Cfg.MMConfigs))
++	for name := range Cfg.MMConfigs {
++		names = append(names, name)
++	}
++	sort.Strings(names)
++	return names
++}

added internal/config/agents_test.go +96 -0 Viewed

new file mode 100644
@@ -0,0 +1,96 @@
	1	+package config
	2	+
	3	+import (
	4	+ "os"
	5	+ "path/filepath"
	6	+ "reflect"
	7	+ "strings"
	8	+ "testing"
	9	+)
	10	+
	11	+// mmConfigs is read, its relative paths are anchored on the config file's
	12	+// directory (not the current one), absolute paths are kept, and the names
	13	+// come back sorted.
	14	+func TestMMConfigsAreResolvedAgainstTheConfigFile(t *testing.T) {
	15	+ defaults := Cfg
	16	+ t.Cleanup(func() { Cfg = defaults })
	17	+
	18	+ dir := t.TempDir()
	19	+ path := filepath.Join(dir, "agent.yaml")
	20	+ yaml := "mmConfigs:\n riker: other/agent.riker.yaml\n bob: /abs/agent.bob.yaml\n"
	21	+ if err := os.WriteFile(path, []byte(yaml), 0o644); err != nil {
	22	+ t.Fatal(err)
	23	+ }
	24	+ if _, err := Load(path); err != nil {
	25	+ t.Fatal(err)
	26	+ }
	27	+ if got, want := Cfg.MMConfigs["riker"], filepath.Join(dir, "other", "agent.riker.yaml"); got != want {
	28	+ t.Errorf("riker = %q, want %q", got, want)
	29	+ }
	30	+ if got := Cfg.MMConfigs["bob"]; got != "/abs/agent.bob.yaml" {
	31	+ t.Errorf("bob = %q, want the absolute path unchanged", got)
	32	+ }
	33	+ if got := AgentNames(); !reflect.DeepEqual(got, []string{"bob", "riker"}) {
	34	+ t.Errorf("AgentNames = %v, want [bob riker]", got)
	35	+ }
	36	+}
	37	+
	38	+// A name that could not be a command name, or an empty path, is a config
	39	+// error naming the culprit; no mmConfigs at all is fine.
	40	+func TestMMConfigsValidation(t *testing.T) {
	41	+ defaults := Cfg
	42	+ t.Cleanup(func() { Cfg = defaults })
	43	+ dir := t.TempDir()
	44	+
	45	+ for _, tc := range []struct{ yaml, want string }{
	46	+ {"mmConfigs:\n Bob: a.yaml\n", `"Bob" is not a valid name`},
	47	+ {"mmConfigs:\n ask bob: a.yaml\n", `"ask bob" is not a valid name`},
	48	+ {"mmConfigs:\n -bob: a.yaml\n", `"-bob" is not a valid name`},
	49	+ {"mmConfigs:\n bob: \"\"\n", "mmConfigs.bob: the path must not be empty"},
	50	+ {"mmConfigs:\n with-context-bob: a.yaml\n", "it would read as /ask-with-context"},
	51	+ } {
	52	+ Cfg = defaults
	53	+ Cfg.MMConfigs = nil
	54	+ path := filepath.Join(dir, "bad.yaml")
	55	+ if err := os.WriteFile(path, []byte(tc.yaml), 0o644); err != nil {
	56	+ t.Fatal(err)
	57	+ }
	58	+ if _, err := Load(path); err == nil \|\| !strings.Contains(err.Error(), tc.want) {
	59	+ t.Errorf("%q: Load = %v, want an error containing %q", tc.yaml, err, tc.want)
	60	+ }
	61	+ }
	62	+
	63	+ Cfg = defaults
	64	+ Cfg.MMConfigs = nil
	65	+ if len(AgentNames()) != 0 {
	66	+ t.Errorf("no mmConfigs: AgentNames = %v, want none", AgentNames())
	67	+ }
	68	+}
	69	+
	70	+// AGENT_SESSIONS=off turns persistence off whatever the file says — what an
	71	+// agent started by /ask runs with; any other value leaves the file alone.
	72	+func TestSessionsEnvTurnsPersistenceOff(t *testing.T) {
	73	+ defaults := Cfg
	74	+ t.Cleanup(func() { Cfg = defaults })
	75	+ path := filepath.Join(t.TempDir(), "agent.yaml")
	76	+ if err := os.WriteFile(path, []byte("sessions:\n enabled: true\n"), 0o644); err != nil {
	77	+ t.Fatal(err)
	78	+ }
	79	+
	80	+ t.Setenv(SessionsEnv, "off")
	81	+ if _, err := Load(path); err != nil {
	82	+ t.Fatal(err)
	83	+ }
	84	+ if Cfg.Sessions.Enabled {
	85	+ t.Error("AGENT_SESSIONS=off: sessions still enabled")
	86	+ }
	87	+
	88	+ Cfg = defaults
	89	+ t.Setenv(SessionsEnv, "")
	90	+ if _, err := Load(path); err != nil {
	91	+ t.Fatal(err)
	92	+ }
	93	+ if !Cfg.Sessions.Enabled {
	94	+ t.Error("AGENT_SESSIONS unset: the file's sessions.enabled was ignored")
	95	+ }
	96	+}

++package config
++
++import (
++	"os"
++	"path/filepath"
++	"reflect"
++	"strings"
++	"testing"
++)
++
++// mmConfigs is read, its relative paths are anchored on the config file's
++// directory (not the current one), absolute paths are kept, and the names
++// come back sorted.
++func TestMMConfigsAreResolvedAgainstTheConfigFile(t *testing.T) {
++	defaults := Cfg
++	t.Cleanup(func() { Cfg = defaults })
++
++	dir := t.TempDir()
++	path := filepath.Join(dir, "agent.yaml")
++	yaml := "mmConfigs:\n  riker: other/agent.riker.yaml\n  bob: /abs/agent.bob.yaml\n"
++	if err := os.WriteFile(path, []byte(yaml), 0o644); err != nil {
++		t.Fatal(err)
++	}
++	if _, err := Load(path); err != nil {
++		t.Fatal(err)
++	}
++	if got, want := Cfg.MMConfigs["riker"], filepath.Join(dir, "other", "agent.riker.yaml"); got != want {
++		t.Errorf("riker = %q, want %q", got, want)
++	}
++	if got := Cfg.MMConfigs["bob"]; got != "/abs/agent.bob.yaml" {
++		t.Errorf("bob = %q, want the absolute path unchanged", got)
++	}
++	if got := AgentNames(); !reflect.DeepEqual(got, []string{"bob", "riker"}) {
++		t.Errorf("AgentNames = %v, want [bob riker]", got)
++	}
++}
++
++// A name that could not be a command name, or an empty path, is a config
++// error naming the culprit; no mmConfigs at all is fine.
++func TestMMConfigsValidation(t *testing.T) {
++	defaults := Cfg
++	t.Cleanup(func() { Cfg = defaults })
++	dir := t.TempDir()
++
++	for _, tc := range []struct{ yaml, want string }{
++		{"mmConfigs:\n  Bob: a.yaml\n", `"Bob" is not a valid name`},
++		{"mmConfigs:\n  ask bob: a.yaml\n", `"ask bob" is not a valid name`},
++		{"mmConfigs:\n  -bob: a.yaml\n", `"-bob" is not a valid name`},
++		{"mmConfigs:\n  bob: \"\"\n", "mmConfigs.bob: the path must not be empty"},
++		{"mmConfigs:\n  with-context-bob: a.yaml\n", "it would read as /ask-with-context"},
++	} {
++		Cfg = defaults
++		Cfg.MMConfigs = nil
++		path := filepath.Join(dir, "bad.yaml")
++		if err := os.WriteFile(path, []byte(tc.yaml), 0o644); err != nil {
++			t.Fatal(err)
++		}
++		if _, err := Load(path); err == nil || !strings.Contains(err.Error(), tc.want) {
++			t.Errorf("%q: Load = %v, want an error containing %q", tc.yaml, err, tc.want)
++		}
++	}
++
++	Cfg = defaults
++	Cfg.MMConfigs = nil
++	if len(AgentNames()) != 0 {
++		t.Errorf("no mmConfigs: AgentNames = %v, want none", AgentNames())
++	}
++}
++
++// AGENT_SESSIONS=off turns persistence off whatever the file says — what an
++// agent started by /ask runs with; any other value leaves the file alone.
++func TestSessionsEnvTurnsPersistenceOff(t *testing.T) {
++	defaults := Cfg
++	t.Cleanup(func() { Cfg = defaults })
++	path := filepath.Join(t.TempDir(), "agent.yaml")
++	if err := os.WriteFile(path, []byte("sessions:\n  enabled: true\n"), 0o644); err != nil {
++		t.Fatal(err)
++	}
++
++	t.Setenv(SessionsEnv, "off")
++	if _, err := Load(path); err != nil {
++		t.Fatal(err)
++	}
++	if Cfg.Sessions.Enabled {
++		t.Error("AGENT_SESSIONS=off: sessions still enabled")
++	}
++
++	Cfg = defaults
++	t.Setenv(SessionsEnv, "")
++	if _, err := Load(path); err != nil {
++		t.Fatal(err)
++	}
++	if !Cfg.Sessions.Enabled {
++		t.Error("AGENT_SESSIONS unset: the file's sessions.enabled was ignored")
++	}
++}

modified internal/config/config.go +16 -0 Viewed

@@ -157,6 +157,12 @@ type Config struct {
157	157	// the terminal always has. Only the window mode reads it today; the
158	158	// editor has its own dialog and the line mode has no gatekeeper.
159	159	Permissions string `yaml:"permissions"`
	160	+
	161	+ // MMConfigs: the other mm agents this one can ask (`/ask-<name>`), name
	162	+ // -> path of that agent's YAML file. A relative path is relative to THIS
	163	+ // file's directory, and Load makes it absolute (agents.go). Each name
	164	+ // becomes a command name, hence its format.
	165	+ MMConfigs map[string]string `yaml:"mmConfigs"`
160	166	}
161	167
162	168	// The two values of Permissions.
@@ -516,6 +522,7 @@ func Load(cliPath string) (string, error) {
516	522	return "", fmt.Errorf("%s: %w", path, err)
517	523	}
518	524	applyEnv()
	525	+ resolveAgentPaths(path)
519	526	if err := Cfg.validate(); err != nil {
520	527	return "", fmt.Errorf("%s: %w", path, err)
521	528	}
@@ -533,6 +540,12 @@ func applyEnv() {
533	540	if v := os.Getenv("AGENT_MODEL"); v != "" {
534	541	Cfg.Model = v
535	542	}
	543	+ // AGENT_SESSIONS=off: an agent started by another one's /ask keeps no
	544	+ // session files — it runs in the asker's project, and its sessions would
	545	+ // show up in the asker's /sessions and /resume.
	546	+ if os.Getenv(SessionsEnv) == "off" {
	547	+ Cfg.Sessions.Enabled = false
	548	+ }
536	549	}
537	550
538	551	// validate checks the language servers: every one named and runnable, every
@@ -619,5 +632,8 @@ func (c Config) validate() error {
619	632	if err := c.MCP.Validate(); err != nil {
620	633	return err
621	634	}
	635	+ if err := c.validateAgents(); err != nil {
	636	+ return err
	637	+ }
622	638	return c.LSP.validate()
623	639	}

  	// the terminal always has. Only the window mode reads it today; the
  	// editor has its own dialog and the line mode has no gatekeeper.
  	Permissions string `yaml:"permissions"`
++
++	// MMConfigs: the other mm agents this one can ask (`/ask-<name>`), name
++	// -> path of that agent's YAML file. A relative path is relative to THIS
++	// file's directory, and Load makes it absolute (agents.go). Each name
++	// becomes a command name, hence its format.
++	MMConfigs map[string]string `yaml:"mmConfigs"`
  }
  // The two values of Permissions.
  		return "", fmt.Errorf("%s: %w", path, err)
  	}
  	applyEnv()
++	resolveAgentPaths(path)
  	if err := Cfg.validate(); err != nil {
  		return "", fmt.Errorf("%s: %w", path, err)
  	}
  	if v := os.Getenv("AGENT_MODEL"); v != "" {
  		Cfg.Model = v
  	}
++	// AGENT_SESSIONS=off: an agent started by another one's /ask keeps no
++	// session files — it runs in the asker's project, and its sessions would
++	// show up in the asker's /sessions and /resume.
++	if os.Getenv(SessionsEnv) == "off" {
++		Cfg.Sessions.Enabled = false
++	}
  }
  // validate checks the language servers: every one named and runnable, every
  	if err := c.MCP.Validate(); err != nil {
  		return err
  	}
++	if err := c.validateAgents(); err != nil {
++		return err
++	}
  	return c.LSP.validate()
  }

modified internal/store/store.go +2 -1 Viewed

@@ -64,7 +64,7 @@ type Session struct {
64	64	type Failure struct {
65	65	At time.Time `json:"at"`
66	66	Turn int `json:"turn"` // user questions in the history when it happened
67		- Where string `json:"where"` // FailGeneration, FailCompaction or FailResume
	67	+ Where string `json:"where"` // FailGeneration, FailCompaction, FailResume or FailAsk
68	68	Message string `json:"message"`
69	69	Detail string `json:"detail,omitempty"`
70	70	Question string `json:"question,omitempty"`
@@ -78,6 +78,7 @@ const (
78	78	FailGeneration = "generation" // the model's answer to a question (engine.Generate)
79	79	FailCompaction = "compaction" // the summary request of a compaction (automatic or /compact)
80	80	FailResume = "resume" // a /resume that could not load its session
	81	+ FailAsk = "ask" // an /ask to another agent (internal/ask)
81	82	)
82	83
83	84	// maxErrors bounds the trail: a session that fails on every turn must not

  type Failure struct {
  	At       time.Time `json:"at"`
  	Turn     int       `json:"turn"`  // user questions in the history when it happened
--	Where    string    `json:"where"` // FailGeneration, FailCompaction or FailResume
++	Where    string    `json:"where"` // FailGeneration, FailCompaction, FailResume or FailAsk
  	Message  string    `json:"message"`
  	Detail   string    `json:"detail,omitempty"`
  	Question string    `json:"question,omitempty"`
  	FailGeneration = "generation" // the model's answer to a question (engine.Generate)
  	FailCompaction = "compaction" // the summary request of a compaction (automatic or /compact)
  	FailResume     = "resume"     // a /resume that could not load its session
++	FailAsk        = "ask"        // an /ask to another agent (internal/ask)
  )
  // maxErrors bounds the trail: a session that fails on every turn must not

added internal/tui/ask.go +51 -0 Viewed

new file mode 100644
@@ -0,0 +1,51 @@
	1	+// The window's side of /ask: another mm agent answers in the conversation,
	2	+// its tool calls shown as tool calls, and its answer joins the history for
	3	+// the next question. Run like a turn — on a goroutine, settled by the loop
	4	+// goroutine — so Esc and Ctrl-C stop it the same way.
	5	+
	6	+package tui
	7	+
	8	+import (
	9	+ "context"
	10	+ "fmt"
	11	+
	12	+ "mm/internal/ask"
	13	+ "mm/internal/command"
	14	+ "mm/internal/store"
	15	+
	16	+ "github.com/firebase/genkit/go/ai"
	17	+)
	18	+
	19	+// startAsk shows the command as typed and asks req.Agent on a goroutine.
	20	+func (r *runner) startAsk(p Prompt, req command.AskRequest) {
	21	+ r.t.AddUser(p.Display())
	22	+ r.t.Notice(ask.Label(req))
	23	+ ctx, cancel := context.WithCancel(context.Background())
	24	+ r.cancel = cancel
	25	+ r.cancelled.Store(false)
	26	+ r.busy.Store("asking " + req.Agent)
	27	+ r.streamed.Store(false)
	28	+ msgs := append([]*ai.Message(nil), r.messages...)
	29	+ go func() {
	30	+ exchange, err := r.opts.Asker.Run(ctx, req, msgs, r.opts.Cwd, &sink{r: r})
	31	+ r.results <- outcome{kind: "ask", messages: append(msgs, exchange...), err: err, cancelled: ctx.Err() != nil, line: req.Agent}
	32	+ r.wake()
	33	+ }()
	34	+}
	35	+
	36	+// applyAsk folds a finished /ask into the session: the exchange is kept only
	37	+// when the other agent answered.
	38	+func (r *runner) applyAsk(res outcome) {
	39	+ r.t.Close()
	40	+ switch {
	41	+ case res.err != nil && res.cancelled:
	42	+ r.t.Notice("🛑 Ask aborted — nothing kept.")
	43	+ case res.err != nil:
	44	+ r.t.Error("[error: " + res.err.Error() + "]")
	45	+ r.record(store.NewFailure(r.messages, store.FailAsk, res.err.Error(), res.err))
	46	+ default:
	47	+ r.messages = res.messages
	48	+ r.save()
	49	+ r.t.Notice(fmt.Sprintf("🗣 %s's answer is in the history for your next question", res.line))
	50	+ }
	51	+}

++// The window's side of /ask: another mm agent answers in the conversation,
++// its tool calls shown as tool calls, and its answer joins the history for
++// the next question. Run like a turn — on a goroutine, settled by the loop
++// goroutine — so Esc and Ctrl-C stop it the same way.
++
++package tui
++
++import (
++	"context"
++	"fmt"
++
++	"mm/internal/ask"
++	"mm/internal/command"
++	"mm/internal/store"
++
++	"github.com/firebase/genkit/go/ai"
++)
++
++// startAsk shows the command as typed and asks req.Agent on a goroutine.
++func (r *runner) startAsk(p Prompt, req command.AskRequest) {
++	r.t.AddUser(p.Display())
++	r.t.Notice(ask.Label(req))
++	ctx, cancel := context.WithCancel(context.Background())
++	r.cancel = cancel
++	r.cancelled.Store(false)
++	r.busy.Store("asking " + req.Agent)
++	r.streamed.Store(false)
++	msgs := append([]*ai.Message(nil), r.messages...)
++	go func() {
++		exchange, err := r.opts.Asker.Run(ctx, req, msgs, r.opts.Cwd, &sink{r: r})
++		r.results <- outcome{kind: "ask", messages: append(msgs, exchange...), err: err, cancelled: ctx.Err() != nil, line: req.Agent}
++		r.wake()
++	}()
++}
++
++// applyAsk folds a finished /ask into the session: the exchange is kept only
++// when the other agent answered.
++func (r *runner) applyAsk(res outcome) {
++	r.t.Close()
++	switch {
++	case res.err != nil && res.cancelled:
++		r.t.Notice("🛑 Ask aborted — nothing kept.")
++	case res.err != nil:
++		r.t.Error("[error: " + res.err.Error() + "]")
++		r.record(store.NewFailure(r.messages, store.FailAsk, res.err.Error(), res.err))
++	default:
++		r.messages = res.messages
++		r.save()
++		r.t.Notice(fmt.Sprintf("🗣 %s's answer is in the history for your next question", res.line))
++	}
++}

added internal/tui/ask_test.go +95 -0 Viewed

new file mode 100644
@@ -0,0 +1,95 @@
	1	+package tui
	2	+
	3	+import (
	4	+ "context"
	5	+ "errors"
	6	+ "strings"
	7	+ "testing"
	8	+
	9	+ "mm/internal/command"
	10	+ "mm/internal/config"
	11	+ "mm/internal/ui"
	12	+
	13	+ "github.com/firebase/genkit/go/ai"
	14	+)
	15	+
	16	+// fakeAsker plays the other agent: it shows one tool call and its answer
	17	+// through the sink, or fails, or waits to be cancelled.
	18	+type fakeAsker struct {
	19	+ answer string
	20	+ fail error
	21	+ block bool
	22	+}
	23	+
	24	+func (a fakeAsker) Run(ctx context.Context, req command.AskRequest, _ []ai.Message, _ string, sink ui.Sink) ([]*ai.Message, error) {
	25	+ sink.ToolStart(ui.ToolEvent{ID: "ask-1", Tool: req.Agent + "/bash", Title: req.Agent + " · bash: ls"})
	26	+ sink.ToolEnd("ask-1", ui.ToolResult{Output: "a.txt"})
	27	+ if a.block {
	28	+ <-ctx.Done()
	29	+ return nil, ctx.Err()
	30	+ }
	31	+ if a.fail != nil {
	32	+ return nil, a.fail
	33	+ }
	34	+ sink.Text(a.answer)
	35	+ return []*ai.Message{ai.NewUserTextMessage("[asked " + req.Agent + ": " + req.Prompt + "]"), ai.NewModelTextMessage(a.answer)}, nil
	36	+}
	37	+
	38	+// askRunner is a test runner with one configured agent, bob.
	39	+func askRunner(t testing.T, a fakeAsker) (runner, Transcript) {
	40	+ t.Helper()
	41	+ r, tr := newTestRunner(t, &fakeModel{})
	42	+ config.Cfg.MMConfigs = map[string]string{"bob": "/nowhere/agent.bob.yaml"}
	43	+ command.SetAgents([]string{"bob"})
	44	+ t.Cleanup(func() { command.SetAgents(nil) })
	45	+ r.opts.Asker = a
	46	+ return r, tr
	47	+}
	48	+
	49	+// /ask-bob runs as a job ("asking bob"), shows bob's tool call and answer,
	50	+// and keeps the exchange in the history.
	51	+func TestWindowAskKeepsTheExchange(t *testing.T) {
	52	+ r, tr := askRunner(t, &fakeAsker{answer: "four"})
	53	+ before := len(r.messages)
	54	+ r.submit(promptOf("/ask-bob what is 2+2?"))
	55	+ if r.busyLabel() != "asking bob" {
	56	+ t.Fatalf("busy %q, want asking bob", r.busyLabel())
	57	+ }
	58	+ waitIdle(t, r)
	59	+ if len(r.messages) != before+2 \|\| r.messages[len(r.messages)-1].Text() != "four" {
	60	+ t.Errorf("history: %d messages, last %q", len(r.messages), r.messages[len(r.messages)-1].Text())
	61	+ }
	62	+ var kinds []Kind
	63	+ for _, e := range tr.Entries() {
	64	+ kinds = append(kinds, e.Kind)
	65	+ }
	66	+ if !strings.Contains(notices(tr), "🗣 asking bob") \|\| !strings.Contains(notices(tr), "bob's answer is in the history") \|\| tr.Tools() != 1 {
	67	+ t.Errorf("transcript kinds %v, notices:\n%s", kinds, notices(tr))
	68	+ }
	69	+}
	70	+
	71	+// Esc during an ask stops it and keeps nothing; a failure is an error line
	72	+// and keeps nothing; a bad command is refused without starting anything.
	73	+func TestWindowAskAbortFailureAndMisuse(t *testing.T) {
	74	+ r, tr := askRunner(t, &fakeAsker{block: true})
	75	+ before := len(r.messages)
	76	+ r.submit(promptOf("/ask-bob hi"))
	77	+ r.abort()
	78	+ waitIdle(t, r)
	79	+ if len(r.messages) != before \|\| !strings.Contains(notices(tr), "Ask aborted — nothing kept") {
	80	+ t.Errorf("abort: %d messages, notices:\n%s", len(r.messages), notices(tr))
	81	+ }
	82	+
	83	+ r, tr = askRunner(t, &fakeAsker{fail: errors.New("bob did not start: boom")})
	84	+ r.submit(promptOf("/ask bob hi"))
	85	+ waitIdle(t, r)
	86	+ if len(r.messages) != before \|\| !strings.Contains(notices(tr), "[error: bob did not start: boom]") {
	87	+ t.Errorf("failure: notices:\n%s", notices(tr))
	88	+ }
	89	+
	90	+ r, tr = askRunner(t, &fakeAsker{})
	91	+ r.submit(promptOf("/ask spock hi"))
	92	+ if r.isBusy() \|\| !strings.Contains(notices(tr), `unknown agent "spock"`) {
	93	+ t.Errorf("misuse: busy %v, notices:\n%s", r.isBusy(), notices(tr))
	94	+ }
	95	+}

++package tui
++
++import (
++	"context"
++	"errors"
++	"strings"
++	"testing"
++
++	"mm/internal/command"
++	"mm/internal/config"
++	"mm/internal/ui"
++
++	"github.com/firebase/genkit/go/ai"
++)
++
++// fakeAsker plays the other agent: it shows one tool call and its answer
++// through the sink, or fails, or waits to be cancelled.
++type fakeAsker struct {
++	answer string
++	fail   error
++	block  bool
++}
++
++func (a *fakeAsker) Run(ctx context.Context, req command.AskRequest, _ []*ai.Message, _ string, sink ui.Sink) ([]*ai.Message, error) {
++	sink.ToolStart(ui.ToolEvent{ID: "ask-1", Tool: req.Agent + "/bash", Title: req.Agent + " · bash: ls"})
++	sink.ToolEnd("ask-1", ui.ToolResult{Output: "a.txt"})
++	if a.block {
++		<-ctx.Done()
++		return nil, ctx.Err()
++	}
++	if a.fail != nil {
++		return nil, a.fail
++	}
++	sink.Text(a.answer)
++	return []*ai.Message{ai.NewUserTextMessage("[asked " + req.Agent + ": " + req.Prompt + "]"), ai.NewModelTextMessage(a.answer)}, nil
++}
++
++// askRunner is a test runner with one configured agent, bob.
++func askRunner(t *testing.T, a *fakeAsker) (*runner, *Transcript) {
++	t.Helper()
++	r, tr := newTestRunner(t, &fakeModel{})
++	config.Cfg.MMConfigs = map[string]string{"bob": "/nowhere/agent.bob.yaml"}
++	command.SetAgents([]string{"bob"})
++	t.Cleanup(func() { command.SetAgents(nil) })
++	r.opts.Asker = a
++	return r, tr
++}
++
++// /ask-bob runs as a job ("asking bob"), shows bob's tool call and answer,
++// and keeps the exchange in the history.
++func TestWindowAskKeepsTheExchange(t *testing.T) {
++	r, tr := askRunner(t, &fakeAsker{answer: "four"})
++	before := len(r.messages)
++	r.submit(promptOf("/ask-bob what is 2+2?"))
++	if r.busyLabel() != "asking bob" {
++		t.Fatalf("busy %q, want asking bob", r.busyLabel())
++	}
++	waitIdle(t, r)
++	if len(r.messages) != before+2 || r.messages[len(r.messages)-1].Text() != "four" {
++		t.Errorf("history: %d messages, last %q", len(r.messages), r.messages[len(r.messages)-1].Text())
++	}
++	var kinds []Kind
++	for _, e := range tr.Entries() {
++		kinds = append(kinds, e.Kind)
++	}
++	if !strings.Contains(notices(tr), "🗣 asking bob") || !strings.Contains(notices(tr), "bob's answer is in the history") || tr.Tools() != 1 {
++		t.Errorf("transcript kinds %v, notices:\n%s", kinds, notices(tr))
++	}
++}
++
++// Esc during an ask stops it and keeps nothing; a failure is an error line
++// and keeps nothing; a bad command is refused without starting anything.
++func TestWindowAskAbortFailureAndMisuse(t *testing.T) {
++	r, tr := askRunner(t, &fakeAsker{block: true})
++	before := len(r.messages)
++	r.submit(promptOf("/ask-bob hi"))
++	r.abort()
++	waitIdle(t, r)
++	if len(r.messages) != before || !strings.Contains(notices(tr), "Ask aborted — nothing kept") {
++		t.Errorf("abort: %d messages, notices:\n%s", len(r.messages), notices(tr))
++	}
++
++	r, tr = askRunner(t, &fakeAsker{fail: errors.New("bob did not start: boom")})
++	r.submit(promptOf("/ask bob hi"))
++	waitIdle(t, r)
++	if len(r.messages) != before || !strings.Contains(notices(tr), "[error: bob did not start: boom]") {
++		t.Errorf("failure: notices:\n%s", notices(tr))
++	}
++
++	r, tr = askRunner(t, &fakeAsker{})
++	r.submit(promptOf("/ask spock hi"))
++	if r.isBusy() || !strings.Contains(notices(tr), `unknown agent "spock"`) {
++		t.Errorf("misuse: busy %v, notices:\n%s", r.isBusy(), notices(tr))
++	}
++}

modified internal/tui/commands.go +8 -0 Viewed

@@ -68,6 +68,14 @@ func (r *runner) abort() {
68	68	// dispatch runs a slash command, or starts a turn. A command is read from
69	69	// the typed text: a paste token is never part of a command name.
70	70	func (r *runner) dispatch(p Prompt) {
	71	+ if req, isAsk, err := command.ParseAsk(p.Typed); isAsk {
	72	+ if err != nil {
	73	+ r.t.Error("⚠️ " + err.Error())
	74	+ return
	75	+ }
	76	+ r.startAsk(p, req)
	77	+ return
	78	+ }
71	79	cmd, arg, ok := command.Parse(p.Typed)
72	80	if !ok {
73	81	r.startTurn(p)

  // dispatch runs a slash command, or starts a turn. A command is read from
  // the typed text: a paste token is never part of a command name.
  func (r *runner) dispatch(p Prompt) {
++	if req, isAsk, err := command.ParseAsk(p.Typed); isAsk {
++		if err != nil {
++			r.t.Error("⚠️ " + err.Error())
++			return
++		}
++		r.startAsk(p, req)
++		return
++	}
  	cmd, arg, ok := command.Parse(p.Typed)
  	if !ok {
  		r.startTurn(p)

modified internal/tui/turn.go +12 -3 Viewed

@@ -7,6 +7,7 @@ import (
7	7	"sync"
8	8	"sync/atomic"
9	9
	10	+ "mm/internal/ask"
10	11	"mm/internal/compact"
11	12	"mm/internal/config"
12	13	"mm/internal/detector"
@@ -55,6 +56,9 @@ type Options struct {
55	56	// Banner is the start-up text main printed (or would have); shown as
56	57	// the first notice, since stdout is the window now.
57	58	Banner string
	59	+ // Asker runs the /ask commands (another mm agent); nil when none is
	60	+ // configured — command.ParseAsk then recognises none.
	61	+ Asker ask.Runner
58	62	}
59	63
60	64	// Copier is what /copy writes to: clipboard.Copier, or a fake in tests.
@@ -65,7 +69,7 @@ type Copier interface {
65	69	// outcome is what a background job — a turn or a compaction — reports back
66	70	// to the loop goroutine.
67	71	type outcome struct {
68		- kind string // "turn" or "compact"
	72	+ kind string // "turn", "compact" or "ask"
69	73	messages []*ai.Message
70	74	before int
71	75	resp *ai.ModelResponse
@@ -95,7 +99,7 @@ type runner struct {
95	99	clip Copier // set by newApp; nil means /copy reports that it cannot
96	100
97	101	messages []*ai.Message
98		- busy atomic.Value // string: "", "thinking", "compressing"
	102	+ busy atomic.Value // string: "", "thinking", "compressing", "asking <agent>"
99	103	quit atomic.Bool
100	104	cancel context.CancelFunc
101	105	// cancelled says the job in progress has already been told to stop. A
@@ -129,7 +133,8 @@ func newRunner(opts Options, t Transcript, wake func()) runner {
129	133	return r
130	134	}
131	135
132		-// busyLabel is what the rule shows: "", "thinking" or "compressing".
	136	+// busyLabel is what the rule shows: "", "thinking", "compressing" or
	137	+// "asking <agent>".
133	138	func (r *runner) busyLabel() string { return r.busy.Load().(string) }
134	139
135	140	// isBusy reports whether a job is running.
@@ -255,6 +260,10 @@ func (r *runner) settle() {
255	260
256	261	// apply folds one outcome into the session and the transcript.
257	262	func (r *runner) apply(res outcome) {
	263	+ if res.kind == "ask" {
	264	+ r.applyAsk(res)
	265	+ return
	266	+ }
258	267	if res.kind == "compact" {
259	268	r.messages = res.messages
260	269	switch {

  	"sync"
  	"sync/atomic"
++	"mm/internal/ask"
  	"mm/internal/compact"
  	"mm/internal/config"
  	"mm/internal/detector"
  	// Banner is the start-up text main printed (or would have); shown as
  	// the first notice, since stdout is the window now.
  	Banner string
++	// Asker runs the /ask commands (another mm agent); nil when none is
++	// configured — command.ParseAsk then recognises none.
++	Asker ask.Runner
  }
  // Copier is what /copy writes to: clipboard.Copier, or a fake in tests.
  // outcome is what a background job — a turn or a compaction — reports back
  // to the loop goroutine.
  type outcome struct {
--	kind     string // "turn" or "compact"
++	kind     string // "turn", "compact" or "ask"
  	messages []*ai.Message
  	before   int
  	resp     *ai.ModelResponse
  	clip Copier // set by newApp; nil means /copy reports that it cannot
  	messages []*ai.Message
--	busy     atomic.Value // string: "", "thinking", "compressing"
++	busy     atomic.Value // string: "", "thinking", "compressing", "asking <agent>"
  	quit     atomic.Bool
  	cancel   context.CancelFunc
  	// cancelled says the job in progress has already been told to stop. A
  	return r
  }
--// busyLabel is what the rule shows: "", "thinking" or "compressing".
++// busyLabel is what the rule shows: "", "thinking", "compressing" or
++// "asking <agent>".
  func (r *runner) busyLabel() string { return r.busy.Load().(string) }
  // isBusy reports whether a job is running.
  // apply folds one outcome into the session and the transcript.
  func (r *runner) apply(res outcome) {
++	if res.kind == "ask" {
++		r.applyAsk(res)
++		return
++	}
  	if res.kind == "compact" {
  		r.messages = res.messages
  		switch {

modified internal/tui/view.go +1 -1 Viewed

@@ -78,7 +78,7 @@ type View struct {
78	78	header string // the header line
79	79	status string // the counters on the right of the status bar
80	80	version string // shown before them while the bar is wide enough
81		- busy string // "", "thinking", "compressing"
	81	+ busy string // "", "thinking", "compressing", "asking <agent>"
82	82	ask *permission
83	83
84	84	// input is what has been typed but not sent, as lines. A long paste is

  	header  string // the header line
  	status  string // the counters on the right of the status bar
  	version string // shown before them while the bar is wide enough
--	busy    string // "", "thinking", "compressing"
++	busy    string // "", "thinking", "compressing", "asking <agent>"
  	ask     *permission
  	// input is what has been typed but not sent, as lines. A long paste is

modified main.go +15 -5 Viewed

@@ -38,6 +38,7 @@ import (
38	38
39	39	"mm/internal/acp"
40	40	"mm/internal/agent"
	41	+ "mm/internal/ask"
41	42	"mm/internal/command"
42	43	"mm/internal/config"
43	44	"mm/internal/detector"
@@ -162,6 +163,13 @@ func main() {
162	163	// is made absolute here, so the banner can say exactly where it looked.
163	164	config.Cfg.SkillsDir = resolveSkillsDir(config.Cfg.SkillsDir, path, *acpMode)
164	165
	166	+ // The other agents this one can ask (`mmConfigs`): their /ask-<name>
	167	+ // commands join the shared table before any front end reads it, and their
	168	+ // processes start on the first ask and stop when this one exits.
	169	+ command.SetAgents(config.AgentNames())
	170	+ asker := ask.NewPool()
	171	+ defer asker.Close()
	172	+
165	173	window, banner := windowMode((tuiMode \|\| config.Cfg.TUI) && !acpMode)
166	174
167	175	e, err := engine.New(ctx)
@@ -181,6 +189,7 @@ func main() {
181	189	set := assembleTools(e, loopDetector)
182	190	defer set.close()
183	191	agentTools := set.tools
	192	+ set.warnings = append(set.warnings, ask.Missing()...)
184	193
185	194	if path == "" {
186	195	path = "built-in defaults"
@@ -198,19 +207,19 @@ func main() {
198	207	// Three front ends, one wiring: everything above — config, engine, tools,
199	208	// detector — is strictly identical whichever loop runs below.
200	209	if *acpMode {
201		- acp.Run(ctx, e, config.Cfg.System, agentTools)
	210	+ acp.Run(ctx, e, config.Cfg.System, agentTools, asker)
202	211	return
203	212	}
204	213	if window {
205		- runWindow(ctx, e, agentTools, loopDetector, strings.TrimSpace(banner.String()))
	214	+ runWindow(ctx, e, agentTools, loopDetector, strings.TrimSpace(banner.String()), asker)
206	215	return
207	216	}
208		- agent.Run(ctx, e, config.Cfg.System, agentTools, loopDetector)
	217	+ agent.Run(ctx, e, config.Cfg.System, agentTools, loopDetector, asker)
209	218	}
210	219
211	220	// runWindow starts the single-window front end, falling back to the line
212	221	// mode — banner first — if the terminal cannot be opened after all.
213		-func runWindow(ctx context.Context, e engine.Engine, agentTools []ai.ToolRef, d detector.LoopDetector, banner string) {
	222	+func runWindow(ctx context.Context, e engine.Engine, agentTools []ai.ToolRef, d detector.LoopDetector, banner string, asker ask.Runner) {
214	223	cwd, _ := os.Getwd()
215	224	err := tui.Run(ctx, tui.Options{
216	225	Model: e,
@@ -226,6 +235,7 @@ func runWindow(ctx context.Context, e *engine.Engine, agentTools []ai.ToolRef, d
226	235	Skills: skills.Declared(),
227	236	Version: version.Short(),
228	237	Banner: banner,
	238	+ Asker: asker,
229	239	})
230	240	if err == nil {
231	241	return
@@ -235,7 +245,7 @@ func runWindow(ctx context.Context, e *engine.Engine, agentTools []ai.ToolRef, d
235	245	say("[tui: %v — running the line mode]", err)
236	246	}
237	247	say("%s", banner)
238		- agent.Run(ctx, e, config.Cfg.System, agentTools, d)
	248	+ agent.Run(ctx, e, config.Cfg.System, agentTools, d, asker)
239	249	}
240	250
241	251	// toolSet is what assembleTools hands main: the tools, their names for the

  	"mm/internal/acp"
  	"mm/internal/agent"
++	"mm/internal/ask"
  	"mm/internal/command"
  	"mm/internal/config"
  	"mm/internal/detector"
  	// is made absolute here, so the banner can say exactly where it looked.
  	config.Cfg.SkillsDir = resolveSkillsDir(config.Cfg.SkillsDir, path, *acpMode)
++	// The other agents this one can ask (`mmConfigs`): their /ask-<name>
++	// commands join the shared table before any front end reads it, and their
++	// processes start on the first ask and stop when this one exits.
++	command.SetAgents(config.AgentNames())
++	asker := ask.NewPool()
++	defer asker.Close()
++
  	window, banner := windowMode((*tuiMode || config.Cfg.TUI) && !*acpMode)
  	e, err := engine.New(ctx)
  	set := assembleTools(e, loopDetector)
  	defer set.close()
  	agentTools := set.tools
++	set.warnings = append(set.warnings, ask.Missing()...)
  	if path == "" {
  		path = "built-in defaults"
  	// Three front ends, one wiring: everything above — config, engine, tools,
  	// detector — is strictly identical whichever loop runs below.
  	if *acpMode {
--		acp.Run(ctx, e, config.Cfg.System, agentTools)
++		acp.Run(ctx, e, config.Cfg.System, agentTools, asker)
  		return
  	}
  	if window {
--		runWindow(ctx, e, agentTools, loopDetector, strings.TrimSpace(banner.String()))
++		runWindow(ctx, e, agentTools, loopDetector, strings.TrimSpace(banner.String()), asker)
  		return
  	}
--	agent.Run(ctx, e, config.Cfg.System, agentTools, loopDetector)
++	agent.Run(ctx, e, config.Cfg.System, agentTools, loopDetector, asker)
  }
  // runWindow starts the single-window front end, falling back to the line
  // mode — banner first — if the terminal cannot be opened after all.
--func runWindow(ctx context.Context, e *engine.Engine, agentTools []ai.ToolRef, d *detector.LoopDetector, banner string) {
++func runWindow(ctx context.Context, e *engine.Engine, agentTools []ai.ToolRef, d *detector.LoopDetector, banner string, asker ask.Runner) {
  	cwd, _ := os.Getwd()
  	err := tui.Run(ctx, tui.Options{
  		Model:     e,
  		Skills:    skills.Declared(),
  		Version:   version.Short(),
  		Banner:    banner,
++		Asker:     asker,
  	})
  	if err == nil {
  		return
  		say("[tui: %v — running the line mode]", err)
  	}
  	say("%s", banner)
--	agent.Run(ctx, e, config.Cfg.System, agentTools, d)
++	agent.Run(ctx, e, config.Cfg.System, agentTools, d, asker)
  }
  // toolSet is what assembleTools hands main: the tools, their names for the