bots-garden/mini-me · 07f685d

modified .tickets/issues/0092-how-to-unload-a-model.yaml +2 -2 Viewed

@@ -1,12 +1,12 @@
1	1	# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2	2	id: 92
3	3	title: How to unload a model
4		-state: open
	4	+state: closed
5	5	author:
6	6	name: k33g
7	7	email: ph.charriere@gmail.com
8	8	createdAt: 2026-10-17T14:24:18.694Z
9		-updatedAt: 2026-10-18T16:00:00.000Z
	9	+updatedAt: 2026-10-18T19:30:22.673Z
10	10	labels:
11	11	- priority::urgent
12	12	milestone: v0.7.0

  # Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
  id: 92
  title: How to unload a model
--state: open
++state: closed
  author:
    name: k33g
    email: ph.charriere@gmail.com
  createdAt: 2026-10-17T14:24:18.694Z
--updatedAt: 2026-10-18T16:00:00.000Z
++updatedAt: 2026-10-18T19:30:22.673Z
  labels:
    - priority::urgent
  milestone: v0.7.0

deleted .tickets/issues/0101-unload-a-model-command-and-tool.yaml +0 -116 Viewed

deleted file mode 100644
@@ -1,116 +0,0 @@
1		-# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2		-id: 101
3		-title: Unload a model from memory — /unload command and unload_model tool
4		-state: open
5		-author:
6		- name: k33g
7		- email: ph.charriere@gmail.com
8		-createdAt: 2026-10-18T16:30:00.000Z
9		-updatedAt: 2026-10-18T16:30:00.000Z
10		-labels:
11		- - kind::feature
12		-milestone: v0.7.0
13		-body: \|
14		- Follows ticket 92 (research: how each server unloads a model). Give mm the
15		- capacity to unload a model from the server's memory, for the llmman, ollama
16		- and llamacpp providers, in three forms:
17		-
18		- 1. a slash command `/unload [model]`, in the line mode, the window (-tui)
19		- AND over ACP (an editor's slash-command menu);
20		- 2. a built-in tool `unload_model`, like `ask_agent`, that the model calls
21		- itself when its instructions say so — e.g. "unload your model before you
22		- hand the task to another agent", or before handing control back;
23		- 3. a demo on llmman in `demos/unload-models/`.
24		-
25		- The capacity can be turned off in the configuration file.
26		-
27		- Proposal (to confirm before coding):
28		- - Key `unloadModels` (bool, default true). `false`: no `/unload` command
29		- (not listed, not advertised over ACP) and no `unload_model` tool.
30		- - `/unload` alone unloads THIS agent's model; `/unload <model>` another one
31		- served by the same server. The next question reloads the model, as usual.
32		- - Tool `unload_model {model?: string}`: empty = this agent's model. The
33		- output says what happened, as text (unloaded / was not loaded / not
34		- supported by this server, and why), never a Go error.
35		- - Per provider (ticket 92 comment 1):
36		- - llmman, ollama: POST <root>/api/generate {"model": m, "keep_alive": 0}
37		- (root = baseUrl without /v1); success = done_reason "unload".
38		- - llamacpp: POST <root>/models/unload {"model": m} — router mode only; a
39		- single-model llama-server has no such route: say so, and point to
40		- --sleep-idle-seconds.
41		- - dmr: out of scope for now — "not supported by this provider".
42		- - Permission: `unload_model` changes nothing on disk; it is gated like
43		- `ask_agent` (asked when `yolo: false`)? To decide.
44		- - Hooks: the tool goes through preToolCall/postToolCall like every tool.
45		-
46		- Open questions:
47		- - Default on or off? (proposal: on, since unloading is harmless — the next
48		- request reloads.)
49		- - Should the tool be allowed to unload a model other than its own, or only
50		- its own (safer for a small model that might unload the wrong one)?
51		- - Permission with `yolo: false`: ask, or never ask?
52		-tasks:
53		- - id: 1
54		- title: "Agree on the key name and default, the tool's scope (own model only or any), and its permission"
55		- state: open
56		- priority: none
57		- author:
58		- name: k33g
59		- email: ph.charriere@gmail.com
60		- createdAt: 2026-10-18T16:30:00.000Z
61		- updatedAt: 2026-10-18T16:30:00.000Z
62		- - id: 2
63		- title: "engine: Unload(ctx, model) per provider — llmman/ollama /api/generate keep_alive 0, llamacpp /models/unload, others not supported — with tests on a fake server"
64		- state: open
65		- priority: none
66		- author:
67		- name: k33g
68		- email: ph.charriere@gmail.com
69		- createdAt: 2026-10-18T16:30:00.000Z
70		- updatedAt: 2026-10-18T16:30:00.000Z
71		- - id: 3
72		- title: "config key (unloadModels), validation and banner"
73		- state: open
74		- priority: none
75		- author:
76		- name: k33g
77		- email: ph.charriere@gmail.com
78		- createdAt: 2026-10-18T16:30:00.000Z
79		- updatedAt: 2026-10-18T16:30:00.000Z
80		- - id: 4
81		- title: "/unload [model] in the command table for Terminal \| Window \| Editor, handled by the three front ends"
82		- state: open
83		- priority: none
84		- author:
85		- name: k33g
86		- email: ph.charriere@gmail.com
87		- createdAt: 2026-10-18T16:30:00.000Z
88		- updatedAt: 2026-10-18T16:30:00.000Z
89		- - id: 5
90		- title: "built-in tool unload_model (reserved name, declared unless unloadModels is false)"
91		- state: open
92		- priority: none
93		- author:
94		- name: k33g
95		- email: ph.charriere@gmail.com
96		- createdAt: 2026-10-18T16:30:00.000Z
97		- updatedAt: 2026-10-18T16:30:00.000Z
98		- - id: 6
99		- title: "demos/unload-models/ on llmman: a coordinator that unloads its model before handing a task to another agent (different model), READMEs EN+FR, measured"
100		- state: open
101		- priority: none
102		- author:
103		- name: k33g
104		- email: ph.charriere@gmail.com
105		- createdAt: 2026-10-18T16:30:00.000Z
106		- updatedAt: 2026-10-18T16:30:00.000Z
107		- - id: 7
108		- title: "Docs EN+FR (reference: configuration, tools, slash-commands, acp; how-to), quality gate, .memory"
109		- state: open
110		- priority: none
111		- author:
112		- name: k33g
113		- email: ph.charriere@gmail.com
114		- createdAt: 2026-10-18T16:30:00.000Z
115		- updatedAt: 2026-10-18T16:30:00.000Z
116		-comments: []

--# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
--id: 101
--title: Unload a model from memory — /unload command and unload_model tool
--state: open
--author:
--  name: k33g
--  email: ph.charriere@gmail.com
--createdAt: 2026-10-18T16:30:00.000Z
--updatedAt: 2026-10-18T16:30:00.000Z
--labels:
--  - kind::feature
--milestone: v0.7.0
--body: |
--  Follows ticket 92 (research: how each server unloads a model). Give mm the
--  capacity to unload a model from the server's memory, for the llmman, ollama
--  and llamacpp providers, in three forms:
--
--  1. a slash command `/unload [model]`, in the line mode, the window (-tui)
--     AND over ACP (an editor's slash-command menu);
--  2. a built-in tool `unload_model`, like `ask_agent`, that the model calls
--     itself when its instructions say so — e.g. "unload your model before you
--     hand the task to another agent", or before handing control back;
--  3. a demo on llmman in `demos/unload-models/`.
--
--  The capacity can be turned off in the configuration file.
--
--  Proposal (to confirm before coding):
--  - Key `unloadModels` (bool, default true). `false`: no `/unload` command
--    (not listed, not advertised over ACP) and no `unload_model` tool.
--  - `/unload` alone unloads THIS agent's model; `/unload <model>` another one
--    served by the same server. The next question reloads the model, as usual.
--  - Tool `unload_model {model?: string}`: empty = this agent's model. The
--    output says what happened, as text (unloaded / was not loaded / not
--    supported by this server, and why), never a Go error.
--  - Per provider (ticket 92 comment 1):
--    - llmman, ollama: POST <root>/api/generate {"model": m, "keep_alive": 0}
--      (root = baseUrl without /v1); success = done_reason "unload".
--    - llamacpp: POST <root>/models/unload {"model": m} — router mode only; a
--      single-model llama-server has no such route: say so, and point to
--      --sleep-idle-seconds.
--    - dmr: out of scope for now — "not supported by this provider".
--  - Permission: `unload_model` changes nothing on disk; it is gated like
--    `ask_agent` (asked when `yolo: false`)? To decide.
--  - Hooks: the tool goes through preToolCall/postToolCall like every tool.
--
--  Open questions:
--  - Default on or off? (proposal: on, since unloading is harmless — the next
--    request reloads.)
--  - Should the tool be allowed to unload a model other than its own, or only
--    its own (safer for a small model that might unload the wrong one)?
--  - Permission with `yolo: false`: ask, or never ask?
--tasks:
--  - id: 1
--    title: "Agree on the key name and default, the tool's scope (own model only or any), and its permission"
--    state: open
--    priority: none
--    author:
--      name: k33g
--      email: ph.charriere@gmail.com
--    createdAt: 2026-10-18T16:30:00.000Z
--    updatedAt: 2026-10-18T16:30:00.000Z
--  - id: 2
--    title: "engine: Unload(ctx, model) per provider — llmman/ollama /api/generate keep_alive 0, llamacpp /models/unload, others not supported — with tests on a fake server"
--    state: open
--    priority: none
--    author:
--      name: k33g
--      email: ph.charriere@gmail.com
--    createdAt: 2026-10-18T16:30:00.000Z
--    updatedAt: 2026-10-18T16:30:00.000Z
--  - id: 3
--    title: "config key (unloadModels), validation and banner"
--    state: open
--    priority: none
--    author:
--      name: k33g
--      email: ph.charriere@gmail.com
--    createdAt: 2026-10-18T16:30:00.000Z
--    updatedAt: 2026-10-18T16:30:00.000Z
--  - id: 4
--    title: "/unload [model] in the command table for Terminal | Window | Editor, handled by the three front ends"
--    state: open
--    priority: none
--    author:
--      name: k33g
--      email: ph.charriere@gmail.com
--    createdAt: 2026-10-18T16:30:00.000Z
--    updatedAt: 2026-10-18T16:30:00.000Z
--  - id: 5
--    title: "built-in tool unload_model (reserved name, declared unless unloadModels is false)"
--    state: open
--    priority: none
--    author:
--      name: k33g
--      email: ph.charriere@gmail.com
--    createdAt: 2026-10-18T16:30:00.000Z
--    updatedAt: 2026-10-18T16:30:00.000Z
--  - id: 6
--    title: "demos/unload-models/ on llmman: a coordinator that unloads its model before handing a task to another agent (different model), READMEs EN+FR, measured"
--    state: open
--    priority: none
--    author:
--      name: k33g
--      email: ph.charriere@gmail.com
--    createdAt: 2026-10-18T16:30:00.000Z
--    updatedAt: 2026-10-18T16:30:00.000Z
--  - id: 7
--    title: "Docs EN+FR (reference: configuration, tools, slash-commands, acp; how-to), quality gate, .memory"
--    state: open
--    priority: none
--    author:
--      name: k33g
--      email: ph.charriere@gmail.com
--    createdAt: 2026-10-18T16:30:00.000Z
--    updatedAt: 2026-10-18T16:30:00.000Z
--comments: []

added .tickets/issues/0101-unload-a-model-from-memory-unload-command-and-un.yaml +197 -0 Viewed

new file mode 100644
@@ -0,0 +1,197 @@
	1	+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
	2	+id: 101
	3	+title: Unload a model from memory — /unload command and unload_model tool
	4	+state: closed
	5	+author:
	6	+ name: k33g
	7	+ email: ph.charriere@gmail.com
	8	+createdAt: 2026-10-18T16:30:00.000Z
	9	+updatedAt: 2026-10-18T19:30:11.700Z
	10	+labels:
	11	+ - kind::feature
	12	+milestone: v0.7.0
	13	+related:
	14	+ - 92
	15	+body: \|
	16	+ Follows ticket 92 (research: how each server unloads a model). Give mm the
	17	+ capacity to unload a model from the server's memory, for the llmman, ollama
	18	+ and llamacpp providers, in three forms:
	19	+
	20	+ 1. a slash command `/unload [model]`, in the line mode, the window (-tui)
	21	+ AND over ACP (an editor's slash-command menu);
	22	+ 2. a built-in tool `unload_model`, like `ask_agent`, that the model calls
	23	+ itself when its instructions say so — e.g. "unload your model before you
	24	+ hand the task to another agent", or before handing control back;
	25	+ 3. a demo on llmman in `demos/unload-models/`.
	26	+
	27	+ The capacity can be turned off in the configuration file.
	28	+
	29	+ Proposal (to confirm before coding):
	30	+ - Key `unloadModels` (bool, default true). `false`: no `/unload` command
	31	+ (not listed, not advertised over ACP) and no `unload_model` tool.
	32	+ - `/unload` alone unloads THIS agent's model; `/unload <model>` another one
	33	+ served by the same server. The next question reloads the model, as usual.
	34	+ - Tool `unload_model {model?: string}`: empty = this agent's model. The
	35	+ output says what happened, as text (unloaded / was not loaded / not
	36	+ supported by this server, and why), never a Go error.
	37	+ - Per provider (ticket 92 comment 1):
	38	+ - llmman, ollama: POST <root>/api/generate {"model": m, "keep_alive": 0}
	39	+ (root = baseUrl without /v1); success = done_reason "unload".
	40	+ - llamacpp: POST <root>/models/unload {"model": m} — router mode only; a
	41	+ single-model llama-server has no such route: say so, and point to
	42	+ --sleep-idle-seconds.
	43	+ - dmr: out of scope for now — "not supported by this provider".
	44	+ - Permission: `unload_model` changes nothing on disk; it is gated like
	45	+ `ask_agent` (asked when `yolo: false`)? To decide.
	46	+ - Hooks: the tool goes through preToolCall/postToolCall like every tool.
	47	+
	48	+ Open questions:
	49	+ - Default on or off? (proposal: on, since unloading is harmless — the next
	50	+ request reloads.)
	51	+ - Should the tool be allowed to unload a model other than its own, or only
	52	+ its own (safer for a small model that might unload the wrong one)?
	53	+ - Permission with `yolo: false`: ask, or never ask?
	54	+tasks:
	55	+ - id: 1
	56	+ title: Agree on the key name and default, the tool's scope (own model only or any), and its permission
	57	+ state: closed
	58	+ priority: none
	59	+ author:
	60	+ name: k33g
	61	+ email: ph.charriere@gmail.com
	62	+ createdAt: 2026-10-18T16:30:00.000Z
	63	+ updatedAt: 2026-10-18T16:30:00.000Z
	64	+ - id: 2
	65	+ title: "engine: Unload(ctx, model) per provider — llmman/ollama /api/generate keep_alive 0, llamacpp /models/unload, others not supported — with tests on a fake server"
	66	+ state: closed
	67	+ priority: none
	68	+ author:
	69	+ name: k33g
	70	+ email: ph.charriere@gmail.com
	71	+ createdAt: 2026-10-18T16:30:00.000Z
	72	+ updatedAt: 2026-10-18T16:30:00.000Z
	73	+ - id: 3
	74	+ title: config key (unloadModels), validation and banner
	75	+ state: closed
	76	+ priority: none
	77	+ author:
	78	+ name: k33g
	79	+ email: ph.charriere@gmail.com
	80	+ createdAt: 2026-10-18T16:30:00.000Z
	81	+ updatedAt: 2026-10-18T16:30:00.000Z
	82	+ - id: 4
	83	+ title: /unload [model] in the command table for Terminal \| Window \| Editor, handled by the three front ends
	84	+ state: closed
	85	+ priority: none
	86	+ author:
	87	+ name: k33g
	88	+ email: ph.charriere@gmail.com
	89	+ createdAt: 2026-10-18T16:30:00.000Z
	90	+ updatedAt: 2026-10-18T16:30:00.000Z
	91	+ - id: 5
	92	+ title: built-in tool unload_model (reserved name, declared unless unloadModels is false)
	93	+ state: closed
	94	+ priority: none
	95	+ author:
	96	+ name: k33g
	97	+ email: ph.charriere@gmail.com
	98	+ createdAt: 2026-10-18T16:30:00.000Z
	99	+ updatedAt: 2026-10-18T16:30:00.000Z
	100	+ - id: 6
	101	+ title: "demos/unload-models/ on llmman: a coordinator that unloads its model before handing a task to another agent (different model), READMEs EN+FR, measured"
	102	+ state: closed
	103	+ priority: none
	104	+ author:
	105	+ name: k33g
	106	+ email: ph.charriere@gmail.com
	107	+ createdAt: 2026-10-18T16:30:00.000Z
	108	+ updatedAt: 2026-10-18T16:30:00.000Z
	109	+ - id: 7
	110	+ title: "Docs EN+FR (reference: configuration, tools, slash-commands, acp; how-to), quality gate, .memory"
	111	+ state: open
	112	+ priority: none
	113	+ author:
	114	+ name: k33g
	115	+ email: ph.charriere@gmail.com
	116	+ createdAt: 2026-10-18T16:30:00.000Z
	117	+ updatedAt: 2026-10-18T16:30:00.000Z
	118	+comments:
	119	+ - id: 1
	120	+ author:
	121	+ name: k33g
	122	+ email: ph.charriere@gmail.com
	123	+ createdAt: 2026-10-18T16:45:00.000Z
	124	+ body: \|
	125	+ Decided by the user: `unloadModels` ON by default; the tool may unload ANY
	126	+ model (`model` empty = this agent's); never asks for permission, even with
	127	+ `yolo: false` (like read_file: nothing changes on disk, the next request
	128	+ reloads). Branch created by the user.
	129	+ - id: 2
	130	+ author:
	131	+ name: k33g
	132	+ email: ph.charriere@gmail.com
	133	+ createdAt: 2026-10-18T18:00:00.000Z
	134	+ body: \|
	135	+ Done on feature/unload-model (tasks 2-6; task 7 open: quality gate not run,
	136	+ qlty absent from this sandbox — gofmt/go vet clean, go test ./... green).
	137	+ - engine/unload.go: Engine.Unload(ctx, model) → one line; per provider:
	138	+ llmman/ollama = /api/ps lookup first (a keep_alive 0 generate on a model
	139	+ that is NOT loaded would load it), then POST /api/generate keep_alive 0,
	140	+ done_reason "unload"; llamacpp = POST /models/unload (router mode; a 404
	141	+ explains single-model mode and --sleep-idle-seconds); dmr = not supported.
	142	+ - config unloadModels (default true, user's choice, confirmed after being
	143	+ told every existing agent gets one more tool); reserved name unload_model.
	144	+ - /unload [model]: command.SetUnload, Terminal\|Window\|Editor; REPL, window
	145	+ (off the event loop) and ACP handlers. Tool unload_model (any model, empty
	146	+ = own), never gated, not beside a structuredOutput.
	147	+ - Finding: unloading the agent's OWN model through the tool lasts one step
	148	+ (it needs its model to read the result) — said in the tool description;
	149	+ the demo unloads the asked agent's model instead.
	150	+ - Live: /unload over ACP on llmman (unloaded → "nothing to unload" →
	151	+ unknown model); demos/guard, demos/guards and structured-output-pizza
	152	+ still work with the default. demos/unload-models: MiniCPM5 2B coordinator
	153	+ + Qwen3.5 9B expert, unload_model called 4/5 (after a prompt fix; first
	154	+ prompt skipped it), expert gone from /api/ps each time it was called.
	155	+ - Docs EN+FR: reference configuration / slash-commands / tools, how-to
	156	+ unload-a-model, indexes, README, demo READMEs.
	157	+ - id: 3
	158	+ author:
	159	+ name: k33g
	160	+ email: ph.charriere@gmail.com
	161	+ createdAt: 2026-10-18T19:30:00.000Z
	162	+ body: \|
	163	+ Follow-ups (user's requests):
	164	+ - "unload the main model before the other agent starts": the model cannot
	165	+ (it is needed until the ask_agent call and again to read the answer).
	166	+ User chose a preAsk hook in the demo over an mmConfigs option:
	167	+ demos/unload-models/hooks/unload-self.js (keep_alive 0 on the asking
	168	+ agent's model, read from the hook input "model"). Measured with /api/ps
	169	+ every 0.5 s on 3 questions: coordinator alone → nothing → expert ALONE →
	170	+ both (coordinator reloads to read) → coordinator; hook 3/3, tool 3/3.
	171	+ - "/unload not visible in -tui": the "/" picker drew only the first 8
	172	+ choices (pickerRows) and never scrolled; /unload is the 10th command.
	173	+ Pre-existing bug (it also hid /ask-<name> commands). Fixed: the window
	174	+ follows the selection (tui/draw.go pickerFirst, + test).
	175	+ - "unload_model calls not visible over ACP": over the real protocol the
	176	+ editor receives the tool_call ("unload_model: <model>", pending →
	177	+ in_progress → completed) — checked with a scripted client on the demo.
	178	+ Most likely the editor runs an older installed mm; not reproduced here.
	179	+ - id: 4
	180	+ author:
	181	+ name: k33g
	182	+ email: ph.charriere@gmail.com
	183	+ createdAt: 2026-10-18T21:30:00.000Z
	184	+ body: \|
	185	+ Demo follow-ups (user's requests):
	186	+ - Models: coordinator Qwen3.5 9B, expert Qwen3.5 4B (was Ornith 1.5 9B /
	187	+ Qwen3.5 9B). With preAsk + unload_model: 7/7 clean, but both models
	188	+ loaded together ~18 s per question (the coordinator reloads to read the
	189	+ answer before it can call unload_model).
	190	+ - "never two models loaded at once": hooks/unload-self.js became
	191	+ hooks/unload.js self\|asked; a postAsk hook (unload.js asked, model read
	192	+ from ask.agentModel) unloads the expert before the coordinator reloads.
	193	+ unload_model step removed from the prompt (kept for "free the expert's
	194	+ memory" requests; expert model name in the prompt). Measured 9/9:
	195	+ 9B → nothing → 4B → nothing → 9B; "unload the expert" 4/4. Prompt
	196	+ needed "ask_agent is ALWAYS your first action" (short prompt let the
	197	+ coordinator answer yield itself, 1/7).

++# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
++id: 101
++title: Unload a model from memory — /unload command and unload_model tool
++state: closed
++author:
++  name: k33g
++  email: ph.charriere@gmail.com
++createdAt: 2026-10-18T16:30:00.000Z
++updatedAt: 2026-10-18T19:30:11.700Z
++labels:
++  - kind::feature
++milestone: v0.7.0
++related:
++  - 92
++body: |
++  Follows ticket 92 (research: how each server unloads a model). Give mm the
++  capacity to unload a model from the server's memory, for the llmman, ollama
++  and llamacpp providers, in three forms:
++
++  1. a slash command `/unload [model]`, in the line mode, the window (-tui)
++     AND over ACP (an editor's slash-command menu);
++  2. a built-in tool `unload_model`, like `ask_agent`, that the model calls
++     itself when its instructions say so — e.g. "unload your model before you
++     hand the task to another agent", or before handing control back;
++  3. a demo on llmman in `demos/unload-models/`.
++
++  The capacity can be turned off in the configuration file.
++
++  Proposal (to confirm before coding):
++  - Key `unloadModels` (bool, default true). `false`: no `/unload` command
++    (not listed, not advertised over ACP) and no `unload_model` tool.
++  - `/unload` alone unloads THIS agent's model; `/unload <model>` another one
++    served by the same server. The next question reloads the model, as usual.
++  - Tool `unload_model {model?: string}`: empty = this agent's model. The
++    output says what happened, as text (unloaded / was not loaded / not
++    supported by this server, and why), never a Go error.
++  - Per provider (ticket 92 comment 1):
++    - llmman, ollama: POST <root>/api/generate {"model": m, "keep_alive": 0}
++      (root = baseUrl without /v1); success = done_reason "unload".
++    - llamacpp: POST <root>/models/unload {"model": m} — router mode only; a
++      single-model llama-server has no such route: say so, and point to
++      --sleep-idle-seconds.
++    - dmr: out of scope for now — "not supported by this provider".
++  - Permission: `unload_model` changes nothing on disk; it is gated like
++    `ask_agent` (asked when `yolo: false`)? To decide.
++  - Hooks: the tool goes through preToolCall/postToolCall like every tool.
++
++  Open questions:
++  - Default on or off? (proposal: on, since unloading is harmless — the next
++    request reloads.)
++  - Should the tool be allowed to unload a model other than its own, or only
++    its own (safer for a small model that might unload the wrong one)?
++  - Permission with `yolo: false`: ask, or never ask?
++tasks:
++  - id: 1
++    title: Agree on the key name and default, the tool's scope (own model only or any), and its permission
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-18T16:30:00.000Z
++    updatedAt: 2026-10-18T16:30:00.000Z
++  - id: 2
++    title: "engine: Unload(ctx, model) per provider — llmman/ollama /api/generate keep_alive 0, llamacpp /models/unload, others not supported — with tests on a fake server"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-18T16:30:00.000Z
++    updatedAt: 2026-10-18T16:30:00.000Z
++  - id: 3
++    title: config key (unloadModels), validation and banner
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-18T16:30:00.000Z
++    updatedAt: 2026-10-18T16:30:00.000Z
++  - id: 4
++    title: /unload [model] in the command table for Terminal | Window | Editor, handled by the three front ends
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-18T16:30:00.000Z
++    updatedAt: 2026-10-18T16:30:00.000Z
++  - id: 5
++    title: built-in tool unload_model (reserved name, declared unless unloadModels is false)
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-18T16:30:00.000Z
++    updatedAt: 2026-10-18T16:30:00.000Z
++  - id: 6
++    title: "demos/unload-models/ on llmman: a coordinator that unloads its model before handing a task to another agent (different model), READMEs EN+FR, measured"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-18T16:30:00.000Z
++    updatedAt: 2026-10-18T16:30:00.000Z
++  - id: 7
++    title: "Docs EN+FR (reference: configuration, tools, slash-commands, acp; how-to), quality gate, .memory"
++    state: open
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-18T16:30:00.000Z
++    updatedAt: 2026-10-18T16:30:00.000Z
++comments:
++  - id: 1
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-18T16:45:00.000Z
++    body: |
++      Decided by the user: `unloadModels` ON by default; the tool may unload ANY
++      model (`model` empty = this agent's); never asks for permission, even with
++      `yolo: false` (like read_file: nothing changes on disk, the next request
++      reloads). Branch created by the user.
++  - id: 2
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-18T18:00:00.000Z
++    body: |
++      Done on feature/unload-model (tasks 2-6; task 7 open: quality gate not run,
++      qlty absent from this sandbox — gofmt/go vet clean, go test ./... green).
++      - engine/unload.go: Engine.Unload(ctx, model) → one line; per provider:
++        llmman/ollama = /api/ps lookup first (a keep_alive 0 generate on a model
++        that is NOT loaded would load it), then POST /api/generate keep_alive 0,
++        done_reason "unload"; llamacpp = POST /models/unload (router mode; a 404
++        explains single-model mode and --sleep-idle-seconds); dmr = not supported.
++      - config unloadModels (default true, user's choice, confirmed after being
++        told every existing agent gets one more tool); reserved name unload_model.
++      - /unload [model]: command.SetUnload, Terminal|Window|Editor; REPL, window
++        (off the event loop) and ACP handlers. Tool unload_model (any model, empty
++        = own), never gated, not beside a structuredOutput.
++      - Finding: unloading the agent's OWN model through the tool lasts one step
++        (it needs its model to read the result) — said in the tool description;
++        the demo unloads the asked agent's model instead.
++      - Live: /unload over ACP on llmman (unloaded → "nothing to unload" →
++        unknown model); demos/guard, demos/guards and structured-output-pizza
++        still work with the default. demos/unload-models: MiniCPM5 2B coordinator
++        + Qwen3.5 9B expert, unload_model called 4/5 (after a prompt fix; first
++        prompt skipped it), expert gone from /api/ps each time it was called.
++      - Docs EN+FR: reference configuration / slash-commands / tools, how-to
++        unload-a-model, indexes, README, demo READMEs.
++  - id: 3
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-18T19:30:00.000Z
++    body: |
++      Follow-ups (user's requests):
++      - "unload the main model before the other agent starts": the model cannot
++        (it is needed until the ask_agent call and again to read the answer).
++        User chose a preAsk hook in the demo over an mmConfigs option:
++        demos/unload-models/hooks/unload-self.js (keep_alive 0 on the asking
++        agent's model, read from the hook input "model"). Measured with /api/ps
++        every 0.5 s on 3 questions: coordinator alone → nothing → expert ALONE →
++        both (coordinator reloads to read) → coordinator; hook 3/3, tool 3/3.
++      - "/unload not visible in -tui": the "/" picker drew only the first 8
++        choices (pickerRows) and never scrolled; /unload is the 10th command.
++        Pre-existing bug (it also hid /ask-<name> commands). Fixed: the window
++        follows the selection (tui/draw.go pickerFirst, + test).
++      - "unload_model calls not visible over ACP": over the real protocol the
++        editor receives the tool_call ("unload_model: <model>", pending →
++        in_progress → completed) — checked with a scripted client on the demo.
++        Most likely the editor runs an older installed mm; not reproduced here.
++  - id: 4
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-18T21:30:00.000Z
++    body: |
++      Demo follow-ups (user's requests):
++      - Models: coordinator Qwen3.5 9B, expert Qwen3.5 4B (was Ornith 1.5 9B /
++        Qwen3.5 9B). With preAsk + unload_model: 7/7 clean, but both models
++        loaded together ~18 s per question (the coordinator reloads to read the
++        answer before it can call unload_model).
++      - "never two models loaded at once": hooks/unload-self.js became
++        hooks/unload.js self|asked; a postAsk hook (unload.js asked, model read
++        from ask.agentModel) unloads the expert before the coordinator reloads.
++        unload_model step removed from the prompt (kept for "free the expert's
++        memory" requests; expert model name in the prompt). Measured 9/9:
++        9B → nothing → 4B → nothing → 9B; "unload the expert" 4/4. Prompt
++        needed "ask_agent is ALWAYS your first action" (short prompt let the
++        coordinator answer yield itself, 1/7).

modified README.md +2 -0 Viewed

@@ -120,6 +120,7 @@ Anything that speaks ACP can host `mm -acp` the same way: JetBrains IDEs (In
120	120	- Guard models: have every question checked by guard models (IBM's Granite Guardian), one agent per risk, before your assistant answers — a `preGeneration` hook asks them without the conversation history, and the assistant refuses or answers according to their verdicts. See [Screen questions with guard models](docs/en/how-to/screen-questions-with-guard-models.md), `demos/guard/` and `demos/guards/`.
121	121	- Plans: the built-in `plan` tool lets the agent turn a list of tasks into a tracked plan and show its progress after each answer. See [Manage plans and tasks](docs/en/how-to/manage-plans-and-tasks.md).
122	122	- Save an answer: `/save` writes the last answer to a Markdown file; with `saveAnswerTool: true`, the model does it itself when you ask in your own words (`save_answer`). Never overwrites a file. See [Save an answer to a file](docs/en/how-to/save-an-answer.md).
	123	+- Unload a model: `/unload [model]` frees a model from the server's memory — in the terminal, the window and an editor —, and the `unload_model` tool lets the agent do it itself, e.g. on another agent's model once it has answered; llmman, Ollama and llama-server in router mode. `unloadModels: false` turns both off. See [Free a model's memory](docs/en/how-to/unload-a-model.md) and `demos/unload-models/`.
123	124	- Docker Sandboxes: the sandbox kit lives in its own repository, [mini-me-kit](https://rickub.com/bots-garden/mini-me-kit); `sbx/` explains how to use it (`sbx run …`), with two mixins (`go-toolchain`, `cli-tools`) and ready-made configurations — see `sbx/README.md` and [Run mini-me in a Docker Sandbox](docs/en/how-to/run-in-a-docker-sandbox.md).
124	125
125	126	## Demos
@@ -152,6 +153,7 @@ The `demos/` directory holds ready-made set-ups (most with a README explaining h
152	153	\| `demos/programming-answers-2/` \| Tuto: an agent that answers with mini tutorials in the conversation and saves one to Markdown only when asked (`fetch_content` + `save_answer`), Mellum2.1 12B Thinking on llmman, measured results \|
153	154	\| `demos/guard/` \| one guard model (Granite Guardian 3.2 3B-A800M on llmman) answering `Yes`/`No` to "is this question harmful?"; why its `system:` line is a risk name, not a prompt; custom risks \|
154	155	\| `demos/guards/` \| an assistant (Gemma 4 E4B) protected by seven guard agents, one per risk: a Node.js `preGeneration` hook asks them in parallel without the history, the assistant refuses or answers according to their verdicts; `displayHooks: true`; measured results \|
	156	+\| `demos/unload-models/` \| a coordinator (Qwen3.5 9B) hands programming questions to an expert (Qwen3.5 4B) with `ask_agent`, a `preAsk` hook unloads the coordinator's own model before the expert starts and a `postAsk` hook unloads the expert's once it has answered — the two models are never loaded together; `/unload` and `unload_model` on request; llmman, measured results \|
155	157	\| `demos/declared-tools/` \| an ops desk with no `bash` and no skills, acting only through four tools declared under `tools:` — release, version bump, service logs, slow requests — covering every parameter type (string, enum, boolean, integer, number), a safe default (`dryRun: true`) and an optional argument; Mellum2 12B on DMR, measured results \|
156	158
157	159	## Learn more

  - **Guard models**: have every question checked by guard models (IBM's Granite Guardian), one agent per risk, before your assistant answers — a `preGeneration` hook asks them without the conversation history, and the assistant refuses or answers according to their verdicts. See [Screen questions with guard models](docs/en/how-to/screen-questions-with-guard-models.md), `demos/guard/` and `demos/guards/`.
  - **Plans**: the built-in `plan` tool lets the agent turn a list of tasks into a tracked plan and show its progress after each answer. See [Manage plans and tasks](docs/en/how-to/manage-plans-and-tasks.md).
  - **Save an answer**: `/save` writes the last answer to a Markdown file; with `saveAnswerTool: true`, the model does it itself when you ask in your own words (`save_answer`). Never overwrites a file. See [Save an answer to a file](docs/en/how-to/save-an-answer.md).
++- **Unload a model**: `/unload [model]` frees a model from the server's memory — in the terminal, the window and an editor —, and the `unload_model` tool lets the agent do it itself, e.g. on another agent's model once it has answered; llmman, Ollama and llama-server in router mode. `unloadModels: false` turns both off. See [Free a model's memory](docs/en/how-to/unload-a-model.md) and `demos/unload-models/`.
  - **Docker Sandboxes**: the sandbox kit lives in its own repository, [mini-me-kit](https://rickub.com/bots-garden/mini-me-kit); `sbx/` explains how to use it (`sbx run …`), with two mixins (`go-toolchain`, `cli-tools`) and ready-made configurations — see `sbx/README.md` and [Run mini-me in a Docker Sandbox](docs/en/how-to/run-in-a-docker-sandbox.md).
  ## Demos
  | `demos/programming-answers-2/` | **Tuto**: an agent that answers with mini tutorials in the conversation and saves one to Markdown only when asked (`fetch_content` + `save_answer`), Mellum2.1 12B Thinking on llmman, measured results |
  | `demos/guard/` | one guard model (Granite Guardian 3.2 3B-A800M on llmman) answering `Yes`/`No` to "is this question harmful?"; why its `system:` line is a risk name, not a prompt; custom risks |
  | `demos/guards/` | an assistant (Gemma 4 E4B) protected by seven guard agents, one per risk: a Node.js `preGeneration` hook asks them in parallel without the history, the assistant refuses or answers according to their verdicts; `displayHooks: true`; measured results |
++| `demos/unload-models/` | a coordinator (Qwen3.5 9B) hands programming questions to an expert (Qwen3.5 4B) with `ask_agent`, a `preAsk` hook unloads the coordinator's own model before the expert starts and a `postAsk` hook unloads the expert's once it has answered — the two models are never loaded together; `/unload` and `unload_model` on request; llmman, measured results |
  | `demos/declared-tools/` | an ops desk with no `bash` and no skills, acting only through four tools declared under `tools:` — release, version bump, service logs, slow requests — covering every parameter type (string, enum, boolean, integer, number), a safe default (`dryRun: true`) and an optional argument; Mellum2 12B on DMR, measured results |
  ## Learn more

modified demos/guards/README.en.md +4 -0 Viewed

@@ -30,6 +30,10 @@ The hook runs the guards with the same `mm` as the assistant: mini-me exports it
30	30
31	31	## How it works
32	32
	33	+![The demos/guards flow: the question goes through a preGeneration hook that asks seven guard agents, then to the assistant](guards-flow.svg)
	34	+
	35	+Editable source: `guards-flow.drawio` (draw.io); `guards-flow.svg` is the same picture.
	36	+
33	37	```text
34	38	user's question
35	39	│

  ## How it works
++![The demos/guards flow: the question goes through a preGeneration hook that asks seven guard agents, then to the assistant](guards-flow.svg)
++
++Editable source: `guards-flow.drawio` (draw.io); `guards-flow.svg` is the same picture.
++
  ```text
  user's question
     │

modified demos/guards/README.fr.md +4 -0 Viewed

@@ -30,6 +30,10 @@ Le hook lance les gardes avec le même `mm` que l'assistant : mini-me exporte so
30	30
31	31	## Comment ça marche
32	32
	33	+![Le flux de demos/guards : la question passe par un hook preGeneration qui interroge sept agents de garde, puis va à l'assistant](guards-flow.svg)
	34	+
	35	+Source modifiable : `guards-flow.drawio` (draw.io) ; `guards-flow.svg` est la même image.
	36	+
33	37	```text
34	38	question de l'utilisateur
35	39	│

  ## Comment ça marche
++![Le flux de demos/guards : la question passe par un hook preGeneration qui interroge sept agents de garde, puis va à l'assistant](guards-flow.svg)
++
++Source modifiable : `guards-flow.drawio` (draw.io) ; `guards-flow.svg` est la même image.
++
  ```text
  question de l'utilisateur
     │

added demos/guards/guards-flow.drawio +30 -0 Viewed

new file mode 100644
@@ -0,0 +1,30 @@
	1	+<mxfile host="app.diagrams.net" agent="drawio-fancy-diagram-skill" version="24.7.17" type="device">
	2	+ <diagram name="demos/guards" id="guards-flow">
	3	+ <mxGraphModel dx="1400" dy="800" grid="0" gridSize="10" guides="1" tooltips="1" connect="1" arrows="1" fold="1" page="1" pageScale="1" pageWidth="1200" pageHeight="680" math="0" shadow="0">
	4	+ <root>
	5	+ <mxCell id="0"/>
	6	+ <mxCell id="1" parent="0"/>
	7	+ <mxCell id="bg" value="" style="rounded=0;whiteSpace=wrap;html=1;fillColor=#FFFFFF;gradientColor=#EEF2FA;strokeColor=none;" vertex="1" parent="1"><mxGeometry x="0" y="0" width="1200" height="680" as="geometry"/></mxCell>
	8	+ <mxCell id="title" value="demos/guards — every question screened first" style="text;html=1;align=center;verticalAlign=middle;fontSize=24;fontStyle=1;fontColor=#1A237E;" vertex="1" parent="1"><mxGeometry x="200" y="18" width="800" height="30" as="geometry"/></mxCell>
	9	+ <mxCell id="subtitle" value="a preGeneration hook asks seven guard agents, the question alone, before the assistant answers" style="text;html=1;align=center;verticalAlign=middle;fontSize=13;fontStyle=2;fontColor=#546E7A;" vertex="1" parent="1"><mxGeometry x="150" y="48" width="900" height="20" as="geometry"/></mxCell>
	10	+ <mxCell id="user" value="🧑<br><b>User</b><br><font style="font-size:11px">asks a question<br>terminal · -tui · editor</font>" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#E3F2FD;gradientColor=#BBDEFB;strokeColor=#1565C0;strokeWidth=2;shadow=1;fontSize=15;fontColor=#0D47A1;" vertex="1" parent="1"><mxGeometry x="40" y="140" width="190" height="120" as="geometry"/></mxCell>
	11	+ <mxCell id="hook" value="🔍<br><b>preGeneration hook</b><br><font style="font-size:11px">hooks/guards.js<br>no verdict → stop (fail closed)</font>" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#E1F5FE;gradientColor=#B3E5FC;strokeColor=#0277BD;strokeWidth=2;shadow=1;fontSize=15;fontColor=#01579B;" vertex="1" parent="1"><mxGeometry x="330" y="140" width="250" height="120" as="geometry"/></mxCell>
	12	+ <mxCell id="assistant" value="🤖<br><b>Assistant</b><br><font style="font-size:11px">Gemma 4 E4B · no tools<br>all No → answers · a Yes → refuses</font>" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#F3E5F5;gradientColor=#E1BEE7;strokeColor=#6A1B9A;strokeWidth=3;shadow=1;fontSize=15;fontColor=#4A148C;" vertex="1" parent="1"><mxGeometry x="900" y="140" width="260" height="120" as="geometry"/></mxCell>
	13	+ <mxCell id="badge" value="★ decides" style="rounded=1;arcSize=50;whiteSpace=wrap;html=1;fillColor=#6A1B9A;strokeColor=none;fontSize=11;fontStyle=1;fontColor=#FFFFFF;" vertex="1" parent="1"><mxGeometry x="1060" y="128" width="88" height="24" as="geometry"/></mxCell>
	14	+ <mxCell id="guards" value="7 guard agents · Granite Guardian 3.2 3B · in parallel, 1.2 s in total" style="rounded=1;arcSize=4;whiteSpace=wrap;html=1;fillColor=#FAFAFA;strokeColor=#90A4AE;strokeWidth=2;dashed=1;verticalAlign=top;align=center;fontSize=13;fontStyle=3;fontColor=#607D8B;spacingTop=8;" vertex="1" parent="1"><mxGeometry x="200" y="330" width="800" height="300" as="geometry"/></mxCell>
	15	+ <mxCell id="g0" value="<b>harm</b>" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#ECEFF1;gradientColor=#CFD8DC;strokeColor=#455A64;strokeWidth=2;shadow=1;fontSize=14;fontColor=#263238;" vertex="1" parent="1"><mxGeometry x="236" y="380" width="170" height="64" as="geometry"/></mxCell>
	16	+ <mxCell id="g1" value="<b>jailbreak</b>" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#ECEFF1;gradientColor=#CFD8DC;strokeColor=#455A64;strokeWidth=2;shadow=1;fontSize=14;fontColor=#263238;" vertex="1" parent="1"><mxGeometry x="422" y="380" width="170" height="64" as="geometry"/></mxCell>
	17	+ <mxCell id="g2" value="<b>violence</b>" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#ECEFF1;gradientColor=#CFD8DC;strokeColor=#455A64;strokeWidth=2;shadow=1;fontSize=14;fontColor=#263238;" vertex="1" parent="1"><mxGeometry x="608" y="380" width="170" height="64" as="geometry"/></mxCell>
	18	+ <mxCell id="g3" value="<b>unethical_behavior</b>" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#ECEFF1;gradientColor=#CFD8DC;strokeColor=#455A64;strokeWidth=2;shadow=1;fontSize=14;fontColor=#263238;" vertex="1" parent="1"><mxGeometry x="794" y="380" width="170" height="64" as="geometry"/></mxCell>
	19	+ <mxCell id="g4" value="<b>social_bias</b>" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#ECEFF1;gradientColor=#CFD8DC;strokeColor=#455A64;strokeWidth=2;shadow=1;fontSize=14;fontColor=#263238;" vertex="1" parent="1"><mxGeometry x="329" y="460" width="170" height="64" as="geometry"/></mxCell>
	20	+ <mxCell id="g5" value="<b>profanity</b>" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#ECEFF1;gradientColor=#CFD8DC;strokeColor=#455A64;strokeWidth=2;shadow=1;fontSize=14;fontColor=#263238;" vertex="1" parent="1"><mxGeometry x="515" y="460" width="170" height="64" as="geometry"/></mxCell>
	21	+ <mxCell id="g6" value="<b>sexual_content</b>" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#ECEFF1;gradientColor=#CFD8DC;strokeColor=#455A64;strokeWidth=2;shadow=1;fontSize=14;fontColor=#263238;" vertex="1" parent="1"><mxGeometry x="701" y="460" width="170" height="64" as="geometry"/></mxCell>
	22	+ <mxCell id="notes" value="each guard: mm -config guards/&lt;risk&gt;.yaml -p="&lt;question&gt;"<br>system: &lt;risk name&gt; · a fresh conversation, never the history<br><i><font color="#78909C" style="font-size:11px">answers "Yes" or "No" + &lt;confidence&gt;</font></i>" style="text;html=1;align=center;verticalAlign=middle;fontSize=12;fontColor=#455A64;" vertex="1" parent="1"><mxGeometry x="250" y="550" width="700" height="70" as="geometry"/></mxCell>
	23	+ <mxCell id="e1" value="question" style="edgeStyle=orthogonalEdgeStyle;rounded=1;html=1;strokeColor=#37474F;strokeWidth=2;endArrow=block;endFill=1;fontSize=12;fontStyle=1;fontColor=#263238;labelBackgroundColor=#FFFFFF;" edge="1" parent="1" source="user" target="hook"><mxGeometry relative="1" as="geometry"></mxGeometry></mxCell>
	24	+ <mxCell id="e2" value="question + safety report<br><font style="font-size:10px;font-weight:normal" color="#607D8B">[context added by a hook]</font>" style="edgeStyle=orthogonalEdgeStyle;rounded=1;html=1;strokeColor=#37474F;strokeWidth=2;endArrow=block;endFill=1;fontSize=12;fontStyle=1;fontColor=#263238;labelBackgroundColor=#FFFFFF;" edge="1" parent="1" source="hook" target="assistant"><mxGeometry relative="1" as="geometry"></mxGeometry></mxCell>
	25	+ <mxCell id="e3" value="question alone<br><font style="font-size:10px;font-weight:normal" color="#607D8B">← Yes / No per risk</font>" style="edgeStyle=orthogonalEdgeStyle;rounded=1;html=1;strokeColor=#37474F;strokeWidth=2;endArrow=block;endFill=1;fontSize=12;fontStyle=1;fontColor=#263238;labelBackgroundColor=#FFFFFF;startArrow=block;startFill=1;align=left;spacingLeft=14;" edge="1" parent="1" source="hook" target="guards"><mxGeometry relative="1" as="geometry"></mxGeometry></mxCell>
	26	+ <mxCell id="e4" value="answer · or refusal" style="edgeStyle=orthogonalEdgeStyle;rounded=1;html=1;strokeColor=#37474F;strokeWidth=2;endArrow=block;endFill=1;fontSize=12;fontStyle=1;fontColor=#263238;labelBackgroundColor=#FFFFFF;exitX=0.5;exitY=0;entryX=0.5;entryY=0;" edge="1" parent="1" source="assistant" target="user"><mxGeometry relative="1" as="geometry"><Array as="points"><mxPoint x="1030" y="100"/><mxPoint x="135" y="100"/></Array></mxGeometry></mxCell>
	27	+ </root>
	28	+ </mxGraphModel>
	29	+ </diagram>
	30	+</mxfile>

++<mxfile host="app.diagrams.net" agent="drawio-fancy-diagram-skill" version="24.7.17" type="device">
++  <diagram name="demos/guards" id="guards-flow">
++    <mxGraphModel dx="1400" dy="800" grid="0" gridSize="10" guides="1" tooltips="1" connect="1" arrows="1" fold="1" page="1" pageScale="1" pageWidth="1200" pageHeight="680" math="0" shadow="0">
++      <root>
++        <mxCell id="0"/>
++        <mxCell id="1" parent="0"/>
++        <mxCell id="bg" value="" style="rounded=0;whiteSpace=wrap;html=1;fillColor=#FFFFFF;gradientColor=#EEF2FA;strokeColor=none;" vertex="1" parent="1"><mxGeometry x="0" y="0" width="1200" height="680" as="geometry"/></mxCell>
++        <mxCell id="title" value="demos/guards — every question screened first" style="text;html=1;align=center;verticalAlign=middle;fontSize=24;fontStyle=1;fontColor=#1A237E;" vertex="1" parent="1"><mxGeometry x="200" y="18" width="800" height="30" as="geometry"/></mxCell>
++        <mxCell id="subtitle" value="a preGeneration hook asks seven guard agents, the question alone, before the assistant answers" style="text;html=1;align=center;verticalAlign=middle;fontSize=13;fontStyle=2;fontColor=#546E7A;" vertex="1" parent="1"><mxGeometry x="150" y="48" width="900" height="20" as="geometry"/></mxCell>
++        <mxCell id="user" value="🧑&lt;br&gt;&lt;b&gt;User&lt;/b&gt;&lt;br&gt;&lt;font style=&quot;font-size:11px&quot;&gt;asks a question&lt;br&gt;terminal · -tui · editor&lt;/font&gt;" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#E3F2FD;gradientColor=#BBDEFB;strokeColor=#1565C0;strokeWidth=2;shadow=1;fontSize=15;fontColor=#0D47A1;" vertex="1" parent="1"><mxGeometry x="40" y="140" width="190" height="120" as="geometry"/></mxCell>
++        <mxCell id="hook" value="🔍&lt;br&gt;&lt;b&gt;preGeneration hook&lt;/b&gt;&lt;br&gt;&lt;font style=&quot;font-size:11px&quot;&gt;hooks/guards.js&lt;br&gt;no verdict → stop (fail closed)&lt;/font&gt;" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#E1F5FE;gradientColor=#B3E5FC;strokeColor=#0277BD;strokeWidth=2;shadow=1;fontSize=15;fontColor=#01579B;" vertex="1" parent="1"><mxGeometry x="330" y="140" width="250" height="120" as="geometry"/></mxCell>
++        <mxCell id="assistant" value="🤖&lt;br&gt;&lt;b&gt;Assistant&lt;/b&gt;&lt;br&gt;&lt;font style=&quot;font-size:11px&quot;&gt;Gemma 4 E4B · no tools&lt;br&gt;all No → answers · a Yes → refuses&lt;/font&gt;" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#F3E5F5;gradientColor=#E1BEE7;strokeColor=#6A1B9A;strokeWidth=3;shadow=1;fontSize=15;fontColor=#4A148C;" vertex="1" parent="1"><mxGeometry x="900" y="140" width="260" height="120" as="geometry"/></mxCell>
++        <mxCell id="badge" value="★ decides" style="rounded=1;arcSize=50;whiteSpace=wrap;html=1;fillColor=#6A1B9A;strokeColor=none;fontSize=11;fontStyle=1;fontColor=#FFFFFF;" vertex="1" parent="1"><mxGeometry x="1060" y="128" width="88" height="24" as="geometry"/></mxCell>
++        <mxCell id="guards" value="7 guard agents · Granite Guardian 3.2 3B · in parallel, 1.2 s in total" style="rounded=1;arcSize=4;whiteSpace=wrap;html=1;fillColor=#FAFAFA;strokeColor=#90A4AE;strokeWidth=2;dashed=1;verticalAlign=top;align=center;fontSize=13;fontStyle=3;fontColor=#607D8B;spacingTop=8;" vertex="1" parent="1"><mxGeometry x="200" y="330" width="800" height="300" as="geometry"/></mxCell>
++        <mxCell id="g0" value="&lt;b&gt;harm&lt;/b&gt;" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#ECEFF1;gradientColor=#CFD8DC;strokeColor=#455A64;strokeWidth=2;shadow=1;fontSize=14;fontColor=#263238;" vertex="1" parent="1"><mxGeometry x="236" y="380" width="170" height="64" as="geometry"/></mxCell>
++        <mxCell id="g1" value="&lt;b&gt;jailbreak&lt;/b&gt;" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#ECEFF1;gradientColor=#CFD8DC;strokeColor=#455A64;strokeWidth=2;shadow=1;fontSize=14;fontColor=#263238;" vertex="1" parent="1"><mxGeometry x="422" y="380" width="170" height="64" as="geometry"/></mxCell>
++        <mxCell id="g2" value="&lt;b&gt;violence&lt;/b&gt;" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#ECEFF1;gradientColor=#CFD8DC;strokeColor=#455A64;strokeWidth=2;shadow=1;fontSize=14;fontColor=#263238;" vertex="1" parent="1"><mxGeometry x="608" y="380" width="170" height="64" as="geometry"/></mxCell>
++        <mxCell id="g3" value="&lt;b&gt;unethical_behavior&lt;/b&gt;" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#ECEFF1;gradientColor=#CFD8DC;strokeColor=#455A64;strokeWidth=2;shadow=1;fontSize=14;fontColor=#263238;" vertex="1" parent="1"><mxGeometry x="794" y="380" width="170" height="64" as="geometry"/></mxCell>
++        <mxCell id="g4" value="&lt;b&gt;social_bias&lt;/b&gt;" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#ECEFF1;gradientColor=#CFD8DC;strokeColor=#455A64;strokeWidth=2;shadow=1;fontSize=14;fontColor=#263238;" vertex="1" parent="1"><mxGeometry x="329" y="460" width="170" height="64" as="geometry"/></mxCell>
++        <mxCell id="g5" value="&lt;b&gt;profanity&lt;/b&gt;" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#ECEFF1;gradientColor=#CFD8DC;strokeColor=#455A64;strokeWidth=2;shadow=1;fontSize=14;fontColor=#263238;" vertex="1" parent="1"><mxGeometry x="515" y="460" width="170" height="64" as="geometry"/></mxCell>
++        <mxCell id="g6" value="&lt;b&gt;sexual_content&lt;/b&gt;" style="rounded=1;arcSize=14;whiteSpace=wrap;html=1;fillColor=#ECEFF1;gradientColor=#CFD8DC;strokeColor=#455A64;strokeWidth=2;shadow=1;fontSize=14;fontColor=#263238;" vertex="1" parent="1"><mxGeometry x="701" y="460" width="170" height="64" as="geometry"/></mxCell>
++        <mxCell id="notes" value="each guard: mm -config guards/&amp;lt;risk&amp;gt;.yaml -p=&quot;&amp;lt;question&amp;gt;&quot;&lt;br&gt;system: &amp;lt;risk name&amp;gt; · a fresh conversation, never the history&lt;br&gt;&lt;i&gt;&lt;font color=&quot;#78909C&quot; style=&quot;font-size:11px&quot;&gt;answers &quot;Yes&quot; or &quot;No&quot; + &amp;lt;confidence&amp;gt;&lt;/font&gt;&lt;/i&gt;" style="text;html=1;align=center;verticalAlign=middle;fontSize=12;fontColor=#455A64;" vertex="1" parent="1"><mxGeometry x="250" y="550" width="700" height="70" as="geometry"/></mxCell>
++        <mxCell id="e1" value="question" style="edgeStyle=orthogonalEdgeStyle;rounded=1;html=1;strokeColor=#37474F;strokeWidth=2;endArrow=block;endFill=1;fontSize=12;fontStyle=1;fontColor=#263238;labelBackgroundColor=#FFFFFF;" edge="1" parent="1" source="user" target="hook"><mxGeometry relative="1" as="geometry"></mxGeometry></mxCell>
++        <mxCell id="e2" value="question + safety report&lt;br&gt;&lt;font style=&quot;font-size:10px;font-weight:normal&quot; color=&quot;#607D8B&quot;&gt;[context added by a hook]&lt;/font&gt;" style="edgeStyle=orthogonalEdgeStyle;rounded=1;html=1;strokeColor=#37474F;strokeWidth=2;endArrow=block;endFill=1;fontSize=12;fontStyle=1;fontColor=#263238;labelBackgroundColor=#FFFFFF;" edge="1" parent="1" source="hook" target="assistant"><mxGeometry relative="1" as="geometry"></mxGeometry></mxCell>
++        <mxCell id="e3" value="question alone&lt;br&gt;&lt;font style=&quot;font-size:10px;font-weight:normal&quot; color=&quot;#607D8B&quot;&gt;← Yes / No per risk&lt;/font&gt;" style="edgeStyle=orthogonalEdgeStyle;rounded=1;html=1;strokeColor=#37474F;strokeWidth=2;endArrow=block;endFill=1;fontSize=12;fontStyle=1;fontColor=#263238;labelBackgroundColor=#FFFFFF;startArrow=block;startFill=1;align=left;spacingLeft=14;" edge="1" parent="1" source="hook" target="guards"><mxGeometry relative="1" as="geometry"></mxGeometry></mxCell>
++        <mxCell id="e4" value="answer · or refusal" style="edgeStyle=orthogonalEdgeStyle;rounded=1;html=1;strokeColor=#37474F;strokeWidth=2;endArrow=block;endFill=1;fontSize=12;fontStyle=1;fontColor=#263238;labelBackgroundColor=#FFFFFF;exitX=0.5;exitY=0;entryX=0.5;entryY=0;" edge="1" parent="1" source="assistant" target="user"><mxGeometry relative="1" as="geometry"><Array as="points"><mxPoint x="1030" y="100"/><mxPoint x="135" y="100"/></Array></mxGeometry></mxCell>
++      </root>
++    </mxGraphModel>
++  </diagram>
++</mxfile>

added demos/guards/guards-flow.png +0 -0 Viewed

new file mode 100644
		Binary files /dev/null and b/demos/guards/guards-flow.png differ

new file mode 100644
	Binary files /dev/null and b/demos/guards/guards-flow.png differ		Binary files /dev/null and b/demos/guards/guards-flow.png differ

added demos/guards/guards-flow.svg +117 -0 Viewed

new file mode 100644
@@ -0,0 +1,117 @@
	1	+<svg xmlns="http://www.w3.org/2000/svg" width="1200" height="680" viewBox="0 0 1200 680" font-family="DejaVu Sans, Verdana, sans-serif">
	2	+ <defs>
	3	+ <linearGradient id="gBg" x1="0" y1="0" x2="0" y2="1">
	4	+ <stop offset="0%" stop-color="#FFFFFF"/>
	5	+ <stop offset="100%" stop-color="#EEF2FA"/>
	6	+ </linearGradient>
	7	+ <linearGradient id="gBlue" x1="0" y1="0" x2="0" y2="1">
	8	+ <stop offset="0%" stop-color="#E3F2FD"/>
	9	+ <stop offset="100%" stop-color="#BBDEFB"/>
	10	+ </linearGradient>
	11	+ <linearGradient id="gBlue2" x1="0" y1="0" x2="0" y2="1">
	12	+ <stop offset="0%" stop-color="#E1F5FE"/>
	13	+ <stop offset="100%" stop-color="#B3E5FC"/>
	14	+ </linearGradient>
	15	+ <linearGradient id="gPurple" x1="0" y1="0" x2="0" y2="1">
	16	+ <stop offset="0%" stop-color="#F3E5F5"/>
	17	+ <stop offset="100%" stop-color="#E1BEE7"/>
	18	+ </linearGradient>
	19	+ <linearGradient id="gGrey" x1="0" y1="0" x2="0" y2="1">
	20	+ <stop offset="0%" stop-color="#ECEFF1"/>
	21	+ <stop offset="100%" stop-color="#CFD8DC"/>
	22	+ </linearGradient>
	23	+ <filter id="shadow" x="-30%" y="-30%" width="160%" height="160%">
	24	+ <feDropShadow dx="2" dy="3" stdDeviation="4" flood-color="#000000" flood-opacity="0.25"/>
	25	+ </filter>
	26	+ <marker id="arrow" viewBox="0 0 10 10" refX="9" refY="5" markerWidth="8" markerHeight="8" orient="auto-start-reverse">
	27	+ <path d="M0,0 L10,5 L0,10 z" fill="#37474F"/>
	28	+ </marker>
	29	+ </defs>
	30	+
	31	+ <rect x="0" y="0" width="1200" height="680" fill="url(#gBg)"/>
	32	+
	33	+ <text x="600" y="38" text-anchor="middle" font-size="25" font-weight="700" fill="#1A237E">demos/guards — every question screened first</text>
	34	+ <text x="600" y="61" text-anchor="middle" font-size="14" font-style="italic" fill="#546E7A">a preGeneration hook asks seven guard agents, the question alone, before the assistant answers</text>
	35	+
	36	+ <!-- Answer: Assistant -> User (above the row) -->
	37	+ <polyline points="1030,140 1030,100 135,100 135,138" fill="none" stroke="#37474F" stroke-width="2" marker-end="url(#arrow)"/>
	38	+ <rect x="505" y="88" width="190" height="24" rx="4" fill="#FFFFFF"/>
	39	+ <text x="600" y="105" text-anchor="middle" font-size="12.5" font-weight="700" fill="#263238">answer · or refusal</text>
	40	+
	41	+ <!-- User -->
	42	+ <g filter="url(#shadow)">
	43	+ <rect x="40" y="140" width="190" height="120" rx="16" fill="url(#gBlue)" stroke="#1565C0" stroke-width="2"/>
	44	+ </g>
	45	+ <text x="135" y="182" text-anchor="middle" font-size="24">🧑</text>
	46	+ <text x="135" y="210" text-anchor="middle" font-size="16" font-weight="700" fill="#0D47A1">User</text>
	47	+ <text x="135" y="230" text-anchor="middle" font-size="11.5" fill="#0D47A1">asks a question</text>
	48	+ <text x="135" y="246" text-anchor="middle" font-size="11.5" fill="#0D47A1">terminal · -tui · editor</text>
	49	+
	50	+ <!-- User -> Hook -->
	51	+ <line x1="230" y1="200" x2="328" y2="200" stroke="#37474F" stroke-width="2" marker-end="url(#arrow)"/>
	52	+ <rect x="243" y="190" width="72" height="20" rx="4" fill="#FFFFFF"/>
	53	+ <text x="279" y="204" text-anchor="middle" font-size="12" font-weight="700" fill="#263238">question</text>
	54	+
	55	+ <!-- Hook -->
	56	+ <g filter="url(#shadow)">
	57	+ <rect x="330" y="140" width="250" height="120" rx="16" fill="url(#gBlue2)" stroke="#0277BD" stroke-width="2"/>
	58	+ </g>
	59	+ <text x="455" y="178" text-anchor="middle" font-size="22">🔍</text>
	60	+ <text x="455" y="204" text-anchor="middle" font-size="16" font-weight="700" fill="#01579B">preGeneration hook</text>
	61	+ <text x="455" y="223" text-anchor="middle" font-size="11.5" fill="#01579B">hooks/guards.js</text>
	62	+ <text x="455" y="240" text-anchor="middle" font-size="11.5" fill="#01579B">no verdict → stop (fail closed)</text>
	63	+
	64	+ <!-- Hook -> Assistant -->
	65	+ <line x1="580" y1="200" x2="898" y2="200" stroke="#37474F" stroke-width="2" marker-end="url(#arrow)"/>
	66	+ <rect x="632" y="180" width="216" height="40" rx="4" fill="#FFFFFF"/>
	67	+ <text x="740" y="196" text-anchor="middle" font-size="12.5" font-weight="700" fill="#263238">question + safety report</text>
	68	+ <text x="740" y="212" text-anchor="middle" font-size="10.5" fill="#607D8B">[context added by a hook]</text>
	69	+
	70	+ <!-- Assistant (highlighted) -->
	71	+ <g filter="url(#shadow)">
	72	+ <rect x="900" y="140" width="260" height="120" rx="16" fill="url(#gPurple)" stroke="#6A1B9A" stroke-width="3"/>
	73	+ </g>
	74	+ <rect x="1060" y="128" width="88" height="24" rx="12" fill="#6A1B9A"/>
	75	+ <text x="1104" y="144.5" text-anchor="middle" font-size="11" font-weight="700" fill="#FFFFFF">★ decides</text>
	76	+ <text x="1030" y="180" text-anchor="middle" font-size="22">🤖</text>
	77	+ <text x="1030" y="206" text-anchor="middle" font-size="16" font-weight="700" fill="#4A148C">Assistant</text>
	78	+ <text x="1030" y="225" text-anchor="middle" font-size="11.5" fill="#4A148C">Gemma 4 E4B · no tools</text>
	79	+ <text x="1030" y="242" text-anchor="middle" font-size="11.5" fill="#4A148C">all No → answers · a Yes → refuses</text>
	80	+
	81	+ <!-- Hook <-> guards -->
	82	+ <line x1="455" y1="262" x2="455" y2="328" stroke="#37474F" stroke-width="2" marker-start="url(#arrow)" marker-end="url(#arrow)"/>
	83	+ <rect x="468" y="274" width="170" height="40" rx="4" fill="#FFFFFF"/>
	84	+ <text x="474" y="290" font-size="12.5" font-weight="700" fill="#263238">question alone</text>
	85	+ <text x="474" y="306" font-size="10.5" fill="#607D8B">← Yes / No per risk</text>
	86	+
	87	+ <!-- Guards container -->
	88	+ <rect x="200" y="330" width="800" height="300" rx="10" fill="#FAFAFA" stroke="#90A4AE" stroke-width="2" stroke-dasharray="7,5"/>
	89	+ <text x="600" y="355" text-anchor="middle" font-size="13" font-style="italic" font-weight="700" fill="#607D8B">7 guard agents · Granite Guardian 3.2 3B · in parallel, 1.2 s in total</text>
	90	+
	91	+ <!-- Row 1 -->
	92	+ <g filter="url(#shadow)">
	93	+ <rect x="236" y="380" width="170" height="64" rx="14" fill="url(#gGrey)" stroke="#455A64" stroke-width="2"/>
	94	+ <rect x="422" y="380" width="170" height="64" rx="14" fill="url(#gGrey)" stroke="#455A64" stroke-width="2"/>
	95	+ <rect x="608" y="380" width="170" height="64" rx="14" fill="url(#gGrey)" stroke="#455A64" stroke-width="2"/>
	96	+ <rect x="794" y="380" width="170" height="64" rx="14" fill="url(#gGrey)" stroke="#455A64" stroke-width="2"/>
	97	+ </g>
	98	+ <text x="321" y="417" text-anchor="middle" font-size="14" font-weight="700" fill="#263238">harm</text>
	99	+ <text x="507" y="417" text-anchor="middle" font-size="14" font-weight="700" fill="#263238">jailbreak</text>
	100	+ <text x="693" y="417" text-anchor="middle" font-size="14" font-weight="700" fill="#263238">violence</text>
	101	+ <text x="879" y="417" text-anchor="middle" font-size="14" font-weight="700" fill="#263238">unethical_behavior</text>
	102	+
	103	+ <!-- Row 2 -->
	104	+ <g filter="url(#shadow)">
	105	+ <rect x="329" y="460" width="170" height="64" rx="14" fill="url(#gGrey)" stroke="#455A64" stroke-width="2"/>
	106	+ <rect x="515" y="460" width="170" height="64" rx="14" fill="url(#gGrey)" stroke="#455A64" stroke-width="2"/>
	107	+ <rect x="701" y="460" width="170" height="64" rx="14" fill="url(#gGrey)" stroke="#455A64" stroke-width="2"/>
	108	+ </g>
	109	+ <text x="414" y="497" text-anchor="middle" font-size="14" font-weight="700" fill="#263238">social_bias</text>
	110	+ <text x="600" y="497" text-anchor="middle" font-size="14" font-weight="700" fill="#263238">profanity</text>
	111	+ <text x="786" y="497" text-anchor="middle" font-size="14" font-weight="700" fill="#263238">sexual_content</text>
	112	+
	113	+ <!-- Container notes -->
	114	+ <text x="600" y="566" text-anchor="middle" font-size="12" fill="#455A64">each guard: mm -config guards/<risk>.yaml -p="<question>"</text>
	115	+ <text x="600" y="586" text-anchor="middle" font-size="12" fill="#455A64">system: <risk name> · a fresh conversation, never the history</text>
	116	+ <text x="600" y="610" text-anchor="middle" font-size="11" font-style="italic" fill="#78909C">answers "Yes" or "No" + <confidence></text>
	117	+</svg>

++<svg xmlns="http://www.w3.org/2000/svg" width="1200" height="680" viewBox="0 0 1200 680" font-family="DejaVu Sans, Verdana, sans-serif">
++  <defs>
++    <linearGradient id="gBg" x1="0" y1="0" x2="0" y2="1">
++      <stop offset="0%" stop-color="#FFFFFF"/>
++      <stop offset="100%" stop-color="#EEF2FA"/>
++    </linearGradient>
++    <linearGradient id="gBlue" x1="0" y1="0" x2="0" y2="1">
++      <stop offset="0%" stop-color="#E3F2FD"/>
++      <stop offset="100%" stop-color="#BBDEFB"/>
++    </linearGradient>
++    <linearGradient id="gBlue2" x1="0" y1="0" x2="0" y2="1">
++      <stop offset="0%" stop-color="#E1F5FE"/>
++      <stop offset="100%" stop-color="#B3E5FC"/>
++    </linearGradient>
++    <linearGradient id="gPurple" x1="0" y1="0" x2="0" y2="1">
++      <stop offset="0%" stop-color="#F3E5F5"/>
++      <stop offset="100%" stop-color="#E1BEE7"/>
++    </linearGradient>
++    <linearGradient id="gGrey" x1="0" y1="0" x2="0" y2="1">
++      <stop offset="0%" stop-color="#ECEFF1"/>
++      <stop offset="100%" stop-color="#CFD8DC"/>
++    </linearGradient>
++    <filter id="shadow" x="-30%" y="-30%" width="160%" height="160%">
++      <feDropShadow dx="2" dy="3" stdDeviation="4" flood-color="#000000" flood-opacity="0.25"/>
++    </filter>
++    <marker id="arrow" viewBox="0 0 10 10" refX="9" refY="5" markerWidth="8" markerHeight="8" orient="auto-start-reverse">
++      <path d="M0,0 L10,5 L0,10 z" fill="#37474F"/>
++    </marker>
++  </defs>
++
++  <rect x="0" y="0" width="1200" height="680" fill="url(#gBg)"/>
++
++  <text x="600" y="38" text-anchor="middle" font-size="25" font-weight="700" fill="#1A237E">demos/guards — every question screened first</text>
++  <text x="600" y="61" text-anchor="middle" font-size="14" font-style="italic" fill="#546E7A">a preGeneration hook asks seven guard agents, the question alone, before the assistant answers</text>
++
++  <!-- Answer: Assistant -> User (above the row) -->
++  <polyline points="1030,140 1030,100 135,100 135,138" fill="none" stroke="#37474F" stroke-width="2" marker-end="url(#arrow)"/>
++  <rect x="505" y="88" width="190" height="24" rx="4" fill="#FFFFFF"/>
++  <text x="600" y="105" text-anchor="middle" font-size="12.5" font-weight="700" fill="#263238">answer · or refusal</text>
++
++  <!-- User -->
++  <g filter="url(#shadow)">
++    <rect x="40" y="140" width="190" height="120" rx="16" fill="url(#gBlue)" stroke="#1565C0" stroke-width="2"/>
++  </g>
++  <text x="135" y="182" text-anchor="middle" font-size="24">🧑</text>
++  <text x="135" y="210" text-anchor="middle" font-size="16" font-weight="700" fill="#0D47A1">User</text>
++  <text x="135" y="230" text-anchor="middle" font-size="11.5" fill="#0D47A1">asks a question</text>
++  <text x="135" y="246" text-anchor="middle" font-size="11.5" fill="#0D47A1">terminal · -tui · editor</text>
++
++  <!-- User -> Hook -->
++  <line x1="230" y1="200" x2="328" y2="200" stroke="#37474F" stroke-width="2" marker-end="url(#arrow)"/>
++  <rect x="243" y="190" width="72" height="20" rx="4" fill="#FFFFFF"/>
++  <text x="279" y="204" text-anchor="middle" font-size="12" font-weight="700" fill="#263238">question</text>
++
++  <!-- Hook -->
++  <g filter="url(#shadow)">
++    <rect x="330" y="140" width="250" height="120" rx="16" fill="url(#gBlue2)" stroke="#0277BD" stroke-width="2"/>
++  </g>
++  <text x="455" y="178" text-anchor="middle" font-size="22">🔍</text>
++  <text x="455" y="204" text-anchor="middle" font-size="16" font-weight="700" fill="#01579B">preGeneration hook</text>
++  <text x="455" y="223" text-anchor="middle" font-size="11.5" fill="#01579B">hooks/guards.js</text>
++  <text x="455" y="240" text-anchor="middle" font-size="11.5" fill="#01579B">no verdict → stop (fail closed)</text>
++
++  <!-- Hook -> Assistant -->
++  <line x1="580" y1="200" x2="898" y2="200" stroke="#37474F" stroke-width="2" marker-end="url(#arrow)"/>
++  <rect x="632" y="180" width="216" height="40" rx="4" fill="#FFFFFF"/>
++  <text x="740" y="196" text-anchor="middle" font-size="12.5" font-weight="700" fill="#263238">question + safety report</text>
++  <text x="740" y="212" text-anchor="middle" font-size="10.5" fill="#607D8B">[context added by a hook]</text>
++
++  <!-- Assistant (highlighted) -->
++  <g filter="url(#shadow)">
++    <rect x="900" y="140" width="260" height="120" rx="16" fill="url(#gPurple)" stroke="#6A1B9A" stroke-width="3"/>
++  </g>
++  <rect x="1060" y="128" width="88" height="24" rx="12" fill="#6A1B9A"/>
++  <text x="1104" y="144.5" text-anchor="middle" font-size="11" font-weight="700" fill="#FFFFFF">★ decides</text>
++  <text x="1030" y="180" text-anchor="middle" font-size="22">🤖</text>
++  <text x="1030" y="206" text-anchor="middle" font-size="16" font-weight="700" fill="#4A148C">Assistant</text>
++  <text x="1030" y="225" text-anchor="middle" font-size="11.5" fill="#4A148C">Gemma 4 E4B · no tools</text>
++  <text x="1030" y="242" text-anchor="middle" font-size="11.5" fill="#4A148C">all No → answers · a Yes → refuses</text>
++
++  <!-- Hook <-> guards -->
++  <line x1="455" y1="262" x2="455" y2="328" stroke="#37474F" stroke-width="2" marker-start="url(#arrow)" marker-end="url(#arrow)"/>
++  <rect x="468" y="274" width="170" height="40" rx="4" fill="#FFFFFF"/>
++  <text x="474" y="290" font-size="12.5" font-weight="700" fill="#263238">question alone</text>
++  <text x="474" y="306" font-size="10.5" fill="#607D8B">← Yes / No per risk</text>
++
++  <!-- Guards container -->
++  <rect x="200" y="330" width="800" height="300" rx="10" fill="#FAFAFA" stroke="#90A4AE" stroke-width="2" stroke-dasharray="7,5"/>
++  <text x="600" y="355" text-anchor="middle" font-size="13" font-style="italic" font-weight="700" fill="#607D8B">7 guard agents · Granite Guardian 3.2 3B · in parallel, 1.2 s in total</text>
++
++  <!-- Row 1 -->
++  <g filter="url(#shadow)">
++    <rect x="236" y="380" width="170" height="64" rx="14" fill="url(#gGrey)" stroke="#455A64" stroke-width="2"/>
++    <rect x="422" y="380" width="170" height="64" rx="14" fill="url(#gGrey)" stroke="#455A64" stroke-width="2"/>
++    <rect x="608" y="380" width="170" height="64" rx="14" fill="url(#gGrey)" stroke="#455A64" stroke-width="2"/>
++    <rect x="794" y="380" width="170" height="64" rx="14" fill="url(#gGrey)" stroke="#455A64" stroke-width="2"/>
++  </g>
++  <text x="321" y="417" text-anchor="middle" font-size="14" font-weight="700" fill="#263238">harm</text>
++  <text x="507" y="417" text-anchor="middle" font-size="14" font-weight="700" fill="#263238">jailbreak</text>
++  <text x="693" y="417" text-anchor="middle" font-size="14" font-weight="700" fill="#263238">violence</text>
++  <text x="879" y="417" text-anchor="middle" font-size="14" font-weight="700" fill="#263238">unethical_behavior</text>
++
++  <!-- Row 2 -->
++  <g filter="url(#shadow)">
++    <rect x="329" y="460" width="170" height="64" rx="14" fill="url(#gGrey)" stroke="#455A64" stroke-width="2"/>
++    <rect x="515" y="460" width="170" height="64" rx="14" fill="url(#gGrey)" stroke="#455A64" stroke-width="2"/>
++    <rect x="701" y="460" width="170" height="64" rx="14" fill="url(#gGrey)" stroke="#455A64" stroke-width="2"/>
++  </g>
++  <text x="414" y="497" text-anchor="middle" font-size="14" font-weight="700" fill="#263238">social_bias</text>
++  <text x="600" y="497" text-anchor="middle" font-size="14" font-weight="700" fill="#263238">profanity</text>
++  <text x="786" y="497" text-anchor="middle" font-size="14" font-weight="700" fill="#263238">sexual_content</text>
++
++  <!-- Container notes -->
++  <text x="600" y="566" text-anchor="middle" font-size="12" fill="#455A64">each guard: mm -config guards/&lt;risk&gt;.yaml -p="&lt;question&gt;"</text>
++  <text x="600" y="586" text-anchor="middle" font-size="12" fill="#455A64">system: &lt;risk name&gt; · a fresh conversation, never the history</text>
++  <text x="600" y="610" text-anchor="middle" font-size="11" font-style="italic" fill="#78909C">answers "Yes" or "No" + &lt;confidence&gt;</text>
++</svg>

added demos/unload-models/README.en.md +86 -0 Viewed

new file mode 100644
@@ -0,0 +1,86 @@
	1	+# mini-me demo: unloading models
	2	+
	3	+[Version française](README.fr.md)
	4	+
	5	+A coordinator hands each programming question to an expert, and the two models take turns in the server's memory without ever being loaded together: a `preAsk` hook unloads the coordinator's model right before the expert starts, and a `postAsk` hook unloads the expert's model once it has answered, before the coordinator is reloaded to read the answer. The models are served by [llmman](https://github.com/llmmanorg/llmman).
	6	+
	7	+## What is here
	8	+
	9	+\| File \| Role \|
	10	+\|---\|---\|
	11	+\| `agent.yaml` \| The coordinator (Qwen3.5 9B): `ask_agent` to the expert, then the answer. The two hooks. `/unload` and the `unload_model` tool work here too. `displayHooks: true` shows the hook runs. \|
	12	+\| `hooks/unload.js` \| The hook (Node.js). `unload.js self` (`preAsk`): unloads the coordinator's own model right before mm starts the expert. `unload.js asked` (`postAsk`): unloads the expert's model once it has answered. \|
	13	+\| `agent.expert.yaml` \| The expert (Qwen3.5 4B): answers programming questions, no tools at all (`unloadModels: false`). \|
	14	+
	15	+## Run the demo
	16	+
	17	+```bash
	18	+llmman pull hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M
	19	+llmman pull hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
	20	+cd demos/unload-models
	21	+mm -tui agent.yaml
	22	+```
	23	+
	24	+Ask a programming question, then look at what the server holds:
	25	+
	26	+```bash
	27	+llmman ps # or: curl -s http://127.0.0.1:17434/api/ps
	28	+```
	29	+
	30	+After the answer, only the coordinator's Qwen3.5 9B is loaded. In the terminal the hand-off shows as:
	31	+
	32	+```text
	33	+🗣 ask_agent: expert (llmman/hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M) — Explain what a closure is in Ja…
	34	+ 🪝 1. preAsk expert · node "$MM_CONFIG_DIR/hooks/unload.js" self → continue
	35	+ 🪝 2. postAsk expert · node "$MM_CONFIG_DIR/hooks/unload.js" asked → continue
	36	+```
	37	+
	38	+By hand: `/unload` unloads the coordinator's own model, `/unload <model>` any other one of the same server — in the terminal, the window and an editor. Asking the coordinator to free the expert's memory ("Unload the expert's model.", "Libère la mémoire de l'expert.") makes it call `unload_model` on the expert's model (4/4 measured): its name is written in the prompt, as the model cannot guess how the server names a model.
	39	+
	40	+## How it works
	41	+
	42	+What the server holds during one question (measured with `/api/ps` twice a second):
	43	+
	44	+\| Moment \| Loaded \|
	45	+\|---\|---\|
	46	+\| the coordinator reads the question and calls `ask_agent` \| Qwen3.5 9B \|
	47	+\| the `preAsk` hook unloads the coordinator \| — \|
	48	+\| the expert works \| Qwen3.5 4B alone \|
	49	+\| the `postAsk` hook unloads the expert \| — \|
	50	+\| the coordinator reads the answer (the server reloads it) \| Qwen3.5 9B alone \|
	51	+
	52	+Why hooks, and not the model. Both unloads must happen at a moment when the coordinator's model is not running:
	53	+
	54	+- Before the expert: the coordinator's model has just decided to call `ask_agent`, and it is needed again only to read the answer. The model cannot unload itself — `unload_model` on its own model lasts until its next step.
	55	+- After the expert: if the coordinator called `unload_model` on the expert's model, it would already be running again — reloaded beside the 4B — when it does so. The two models would be loaded together for the whole step. (That is what the previous version of this demo did; see below.)
	56	+
	57	+mm runs the `preAsk` hooks right before it starts the asked agent and the `postAsk` hooks right after its answer, before the asking agent reads it. `hooks/unload.js` reads the model to unload from its input — `"model": "llmman/<model>"` for `self`, `"ask": {"agentModel": "llmman/<model>"}` for `asked` — and sends `keep_alive: 0` to the server. `matcher: ^expert$` keeps both hooks to this hand-off. `onError: warn`: if an unload fails, the hand-off goes on.
	58	+
	59	+The hooks are also certain: a model can skip an instruction, a hook always runs.
	60	+
	61	+## Measured results (2026-10-18, `mm -p`, llmman)
	62	+
	63	+Coordinator Qwen3.5 9B, expert Qwen3.5 4B, both hooks on. "Never together" means `/api/ps` (sampled twice a second) never showed both models at once during the question: 9B → nothing → 4B → nothing → 9B.
	64	+
	65	+\| Question \| `ask_agent` \| `preAsk` unloads the 9B \| `postAsk` unloads the 4B \| never together \| Time \|
	66	+\|---\|---\|---\|---\|---\|---\|
	67	+\| How do I reverse a string in Go? \| ✓ \| ✓ \| ✓ \| ✓ \| 62 s \|
	68	+\| What is a closure in JavaScript? \| ✓ \| ✓ \| ✓ \| ✓ \| 65 s \|
	69	+\| How do I read a file line by line in Python? \| ✓ \| ✓ \| ✓ \| ✓ \| 49 s \|
	70	+\| Explain Rust ownership briefly. \| ✓ \| ✓ \| ✓ \| ✓ \| 82 s \|
	71	+\| Comment trier un tableau en Java ? \| ✓ \| ✓ \| ✓ \| ✓ \| 119 s \|
	72	+\| How do I make an HTTP GET request in Go? \| ✓ \| ✓ \| ✓ \| ✓ \| 106 s \|
	73	+\| What does the yield keyword do in Python? (×3) \| ✓ ✓ ✓ \| ✓ ✓ ✓ \| ✓ ✓ ✓ \| ✓ ✓ ✓ \| 57 s, 60 s, 85 s \|
	74	+
	75	+9 out of 9. The French question is answered in French. A greeting ("Hello, who are you?") is answered in 2 s without any tool. The times include the loads: each question loads the 9B twice and the 4B once, and vary a lot from one run to the next.
	76	+
	77	+The prompt. Without the `unload_model` step, a short prompt ("1. Call ask_agent… 2. Then give the expert's answer") let the coordinator answer the `yield` question by itself, without the expert (1 time out of 7). Saying that `ask_agent` is ALWAYS the first action, "even for a question you know the answer to", gave the table above.
	78	+
	79	+Earlier versions. With the coordinator freeing the expert through `unload_model` (no `postAsk` hook), the 9B and the 4B were loaded together for ~18 s per question while the coordinator read the answer — 7/7 questions otherwise clean. Before that, with a MiniCPM5 2B coordinator and a Qwen3.5 9B expert, the coordinator skipped `unload_model` once in five.
	80	+
	81	+## Good to know
	82	+
	83	+- `postAsk` runs after every hand-off to the expert, answered, failed or cancelled: the expert's model is unloaded even when its turn failed.
	84	+- The next question reloads the 4B: a hand-off costs one load of each model, plus a reload of the coordinator. Keeping both loaded is faster, when the memory allows it.
	85	+- llmman also unloads by itself a model idle for five minutes (`keep_alive`); the hooks free the memory at once, which matters when the next model needs it right away.
	86	+- Turn off `/unload` and `unload_model` with `unloadModels: false`, as the expert does: an agent that only answers does not need one more tool. The hooks do not depend on that key.

++# mini-me demo: unloading models
++
++*[Version française](README.fr.md)*
++
++A coordinator hands each programming question to an expert, and the two models take turns in the server's memory without ever being loaded together: a `preAsk` hook unloads the coordinator's model right before the expert starts, and a `postAsk` hook unloads the expert's model once it has answered, before the coordinator is reloaded to read the answer. The models are served by [llmman](https://github.com/llmmanorg/llmman).
++
++## What is here
++
++| File | Role |
++|---|---|
++| `agent.yaml` | The **coordinator** (Qwen3.5 9B): `ask_agent` to the expert, then the answer. The two hooks. `/unload` and the `unload_model` tool work here too. `displayHooks: true` shows the hook runs. |
++| `hooks/unload.js` | The hook (Node.js). `unload.js self` (`preAsk`): unloads the coordinator's own model right before mm starts the expert. `unload.js asked` (`postAsk`): unloads the expert's model once it has answered. |
++| `agent.expert.yaml` | The **expert** (Qwen3.5 4B): answers programming questions, no tools at all (`unloadModels: false`). |
++
++## Run the demo
++
++```bash
++llmman pull hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M
++llmman pull hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
++cd demos/unload-models
++mm -tui agent.yaml
++```
++
++Ask a programming question, then look at what the server holds:
++
++```bash
++llmman ps        # or: curl -s http://127.0.0.1:17434/api/ps
++```
++
++After the answer, only the coordinator's Qwen3.5 9B is loaded. In the terminal the hand-off shows as:
++
++```text
++🗣 ask_agent: expert (llmman/hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M) — Explain what a closure is in Ja…
++  🪝 1. preAsk expert · node "$MM_CONFIG_DIR/hooks/unload.js" self → continue
++  🪝 2. postAsk expert · node "$MM_CONFIG_DIR/hooks/unload.js" asked → continue
++```
++
++By hand: `/unload` unloads the coordinator's own model, `/unload <model>` any other one of the same server — in the terminal, the window and an editor. Asking the coordinator to free the expert's memory ("Unload the expert's model.", "Libère la mémoire de l'expert.") makes it call `unload_model` on the expert's model (4/4 measured): its name is written in the prompt, as the model cannot guess how the server names a model.
++
++## How it works
++
++What the server holds during one question (measured with `/api/ps` twice a second):
++
++| Moment | Loaded |
++|---|---|
++| the coordinator reads the question and calls `ask_agent` | Qwen3.5 9B |
++| the `preAsk` hook unloads the coordinator | — |
++| the expert works | Qwen3.5 4B alone |
++| the `postAsk` hook unloads the expert | — |
++| the coordinator reads the answer (the server reloads it) | Qwen3.5 9B alone |
++
++**Why hooks, and not the model.** Both unloads must happen at a moment when the coordinator's model is not running:
++
++- *Before the expert*: the coordinator's model has just decided to call `ask_agent`, and it is needed again only to read the answer. The model cannot unload itself — `unload_model` on its own model lasts until its next step.
++- *After the expert*: if the coordinator called `unload_model` on the expert's model, it would already be running again — reloaded beside the 4B — when it does so. The two models would be loaded together for the whole step. (That is what the previous version of this demo did; see below.)
++
++mm runs the `preAsk` hooks right before it starts the asked agent and the `postAsk` hooks right after its answer, before the asking agent reads it. `hooks/unload.js` reads the model to unload from its input — `"model": "llmman/<model>"` for `self`, `"ask": {"agentModel": "llmman/<model>"}` for `asked` — and sends `keep_alive: 0` to the server. `matcher: ^expert$` keeps both hooks to this hand-off. `onError: warn`: if an unload fails, the hand-off goes on.
++
++The hooks are also certain: a model can skip an instruction, a hook always runs.
++
++## Measured results (2026-10-18, `mm -p`, llmman)
++
++Coordinator Qwen3.5 9B, expert Qwen3.5 4B, both hooks on. "Never together" means `/api/ps` (sampled twice a second) never showed both models at once during the question: 9B → nothing → 4B → nothing → 9B.
++
++| Question | `ask_agent` | `preAsk` unloads the 9B | `postAsk` unloads the 4B | never together | Time |
++|---|---|---|---|---|---|
++| How do I reverse a string in Go? | ✓ | ✓ | ✓ | ✓ | 62 s |
++| What is a closure in JavaScript? | ✓ | ✓ | ✓ | ✓ | 65 s |
++| How do I read a file line by line in Python? | ✓ | ✓ | ✓ | ✓ | 49 s |
++| Explain Rust ownership briefly. | ✓ | ✓ | ✓ | ✓ | 82 s |
++| Comment trier un tableau en Java ? | ✓ | ✓ | ✓ | ✓ | 119 s |
++| How do I make an HTTP GET request in Go? | ✓ | ✓ | ✓ | ✓ | 106 s |
++| What does the yield keyword do in Python? (×3) | ✓ ✓ ✓ | ✓ ✓ ✓ | ✓ ✓ ✓ | ✓ ✓ ✓ | 57 s, 60 s, 85 s |
++
++9 out of 9. The French question is answered in French. A greeting ("Hello, who are you?") is answered in 2 s without any tool. The times include the loads: each question loads the 9B twice and the 4B once, and vary a lot from one run to the next.
++
++**The prompt.** Without the `unload_model` step, a short prompt ("1. Call ask_agent… 2. Then give the expert's answer") let the coordinator answer the `yield` question by itself, without the expert (1 time out of 7). Saying that `ask_agent` is ALWAYS the first action, "even for a question you know the answer to", gave the table above.
++
++**Earlier versions.** With the coordinator freeing the expert through `unload_model` (no `postAsk` hook), the 9B and the 4B were loaded together for ~18 s per question while the coordinator read the answer — 7/7 questions otherwise clean. Before that, with a MiniCPM5 2B coordinator and a Qwen3.5 9B expert, the coordinator skipped `unload_model` once in five.
++
++## Good to know
++
++- **`postAsk` runs after every hand-off to the expert**, answered, failed or cancelled: the expert's model is unloaded even when its turn failed.
++- **The next question reloads the 4B**: a hand-off costs one load of each model, plus a reload of the coordinator. Keeping both loaded is faster, when the memory allows it.
++- **llmman also unloads by itself** a model idle for five minutes (`keep_alive`); the hooks free the memory at once, which matters when the next model needs it right away.
++- **Turn off `/unload` and `unload_model`** with `unloadModels: false`, as the expert does: an agent that only answers does not need one more tool. The hooks do not depend on that key.

added demos/unload-models/README.fr.md +86 -0 Viewed

new file mode 100644
@@ -0,0 +1,86 @@
	1	+# Démo mini-me : décharger des modèles
	2	+
	3	+[English version](README.en.md)
	4	+
	5	+Un coordinateur confie chaque question de programmation à un expert, et les deux modèles se relaient dans la mémoire du serveur sans jamais y être chargés ensemble : un hook `preAsk` décharge le modèle du coordinateur juste avant le lancement de l'expert, et un hook `postAsk` décharge le modèle de l'expert une fois sa réponse donnée, avant que le coordinateur soit rechargé pour la lire. Les modèles sont servis par [llmman](https://github.com/llmmanorg/llmman).
	6	+
	7	+## Ce qu'il y a ici
	8	+
	9	+\| Fichier \| Rôle \|
	10	+\|---\|---\|
	11	+\| `agent.yaml` \| Le coordinateur (Qwen3.5 9B) : `ask_agent` vers l'expert, puis la réponse. Les deux hooks. `/unload` et l'outil `unload_model` fonctionnent aussi ici. `displayHooks: true` affiche l'exécution des hooks. \|
	12	+\| `hooks/unload.js` \| Le hook (Node.js). `unload.js self` (`preAsk`) : décharge le modèle du coordinateur juste avant que mm lance l'expert. `unload.js asked` (`postAsk`) : décharge le modèle de l'expert une fois sa réponse donnée. \|
	13	+\| `agent.expert.yaml` \| L'expert (Qwen3.5 4B) : répond aux questions de programmation, sans aucun outil (`unloadModels: false`). \|
	14	+
	15	+## Lancer la démo
	16	+
	17	+```bash
	18	+llmman pull hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M
	19	+llmman pull hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
	20	+cd demos/unload-models
	21	+mm -tui agent.yaml
	22	+```
	23	+
	24	+Pose une question de programmation, puis regarde ce que le serveur garde en mémoire :
	25	+
	26	+```bash
	27	+llmman ps # ou : curl -s http://127.0.0.1:17434/api/ps
	28	+```
	29	+
	30	+Après la réponse, seul le Qwen3.5 9B du coordinateur est chargé. Dans le terminal, le relais s'affiche ainsi :
	31	+
	32	+```text
	33	+🗣 ask_agent: expert (llmman/hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M) — Explain what a closure is in Ja…
	34	+ 🪝 1. preAsk expert · node "$MM_CONFIG_DIR/hooks/unload.js" self → continue
	35	+ 🪝 2. postAsk expert · node "$MM_CONFIG_DIR/hooks/unload.js" asked → continue
	36	+```
	37	+
	38	+À la main : `/unload` décharge le modèle du coordinateur, `/unload <modèle>` n'importe quel autre modèle du même serveur — dans le terminal, la fenêtre et un éditeur. Demander au coordinateur de libérer la mémoire de l'expert (« Unload the expert's model. », « Libère la mémoire de l'expert. ») lui fait appeler `unload_model` sur le modèle de l'expert (4 sur 4 mesurés) : son nom est écrit dans le prompt, car le modèle ne peut pas deviner comment le serveur nomme un modèle.
	39	+
	40	+## Comment ça marche
	41	+
	42	+Ce que le serveur garde en mémoire pendant une question (mesuré avec `/api/ps` deux fois par seconde) :
	43	+
	44	+\| Moment \| Chargé \|
	45	+\|---\|---\|
	46	+\| le coordinateur lit la question et appelle `ask_agent` \| Qwen3.5 9B \|
	47	+\| le hook `preAsk` décharge le coordinateur \| — \|
	48	+\| l'expert travaille \| Qwen3.5 4B seul \|
	49	+\| le hook `postAsk` décharge l'expert \| — \|
	50	+\| le coordinateur lit la réponse (le serveur le recharge) \| Qwen3.5 9B seul \|
	51	+
	52	+Pourquoi des hooks, et pas le modèle. Les deux déchargements doivent avoir lieu à un moment où le modèle du coordinateur ne tourne pas :
	53	+
	54	+- Avant l'expert : le modèle du coordinateur vient de décider d'appeler `ask_agent`, et il n'en a de nouveau besoin que pour lire la réponse. Le modèle ne peut pas se décharger lui-même — `unload_model` sur son propre modèle ne dure que jusqu'à son étape suivante.
	55	+- Après l'expert : si le coordinateur appelait `unload_model` sur le modèle de l'expert, il tournerait déjà de nouveau — rechargé à côté du 4B — au moment de le faire. Les deux modèles seraient chargés ensemble pendant toute cette étape. (C'est ce que faisait la version précédente de cette démo ; voir plus bas.)
	56	+
	57	+mm exécute les hooks `preAsk` juste avant de lancer l'agent interrogé et les hooks `postAsk` juste après sa réponse, avant que l'agent qui l'a interrogé la lise. `hooks/unload.js` lit dans son entrée le modèle à décharger — `"model": "llmman/<modèle>"` pour `self`, `"ask": {"agentModel": "llmman/<modèle>"}` pour `asked` — et envoie `keep_alive: 0` au serveur. `matcher: ^expert$` limite les deux hooks à ce relais. `onError: warn` : si un déchargement échoue, le relais continue.
	58	+
	59	+Les hooks sont en plus certains : un modèle peut sauter une consigne, un hook s'exécute toujours.
	60	+
	61	+## Résultats mesurés (2026-10-18, `mm -p`, llmman)
	62	+
	63	+Coordinateur Qwen3.5 9B, expert Qwen3.5 4B, les deux hooks actifs. « Jamais ensemble » : `/api/ps` (interrogé deux fois par seconde) n'a jamais montré les deux modèles à la fois pendant la question : 9B → rien → 4B → rien → 9B.
	64	+
	65	+\| Question \| `ask_agent` \| `preAsk` décharge le 9B \| `postAsk` décharge le 4B \| jamais ensemble \| Durée \|
	66	+\|---\|---\|---\|---\|---\|---\|
	67	+\| How do I reverse a string in Go? \| ✓ \| ✓ \| ✓ \| ✓ \| 62 s \|
	68	+\| What is a closure in JavaScript? \| ✓ \| ✓ \| ✓ \| ✓ \| 65 s \|
	69	+\| How do I read a file line by line in Python? \| ✓ \| ✓ \| ✓ \| ✓ \| 49 s \|
	70	+\| Explain Rust ownership briefly. \| ✓ \| ✓ \| ✓ \| ✓ \| 82 s \|
	71	+\| Comment trier un tableau en Java ? \| ✓ \| ✓ \| ✓ \| ✓ \| 119 s \|
	72	+\| How do I make an HTTP GET request in Go? \| ✓ \| ✓ \| ✓ \| ✓ \| 106 s \|
	73	+\| What does the yield keyword do in Python? (×3) \| ✓ ✓ ✓ \| ✓ ✓ ✓ \| ✓ ✓ ✓ \| ✓ ✓ ✓ \| 57 s, 60 s, 85 s \|
	74	+
	75	+9 sur 9. La question en français reçoit une réponse en français. Une salutation (« Hello, who are you? ») reçoit une réponse en 2 s sans aucun outil. Les durées comprennent les chargements : chaque question charge le 9B deux fois et le 4B une fois, et elles varient beaucoup d'une exécution à l'autre.
	76	+
	77	+Le prompt. Sans l'étape `unload_model`, un prompt court (« 1. Call ask_agent… 2. Then give the expert's answer ») a laissé le coordinateur répondre seul à la question sur `yield`, sans l'expert (1 fois sur 7). Préciser que `ask_agent` est TOUJOURS la première action, « even for a question you know the answer to », a donné le tableau ci-dessus.
	78	+
	79	+Versions précédentes. Quand le coordinateur libérait l'expert avec `unload_model` (sans hook `postAsk`), le 9B et le 4B étaient chargés ensemble pendant ~18 s par question, le temps que le coordinateur lise la réponse — 7 questions sur 7 sans autre défaut. Avant cela, avec un coordinateur MiniCPM5 2B et un expert Qwen3.5 9B, le coordinateur a sauté `unload_model` une fois sur cinq.
	80	+
	81	+## Ce qu'il faut savoir
	82	+
	83	+- `postAsk` s'exécute après chaque relais vers l'expert, qu'il ait répondu, échoué ou été annulé : le modèle de l'expert est déchargé même quand son tour a échoué.
	84	+- La question suivante recharge le 4B : un relais coûte un chargement de chaque modèle, plus un rechargement du coordinateur. Garder les deux chargés est plus rapide, quand la mémoire le permet.
	85	+- llmman décharge aussi tout seul un modèle inactif depuis cinq minutes (`keep_alive`) ; les hooks libèrent la mémoire tout de suite, ce qui compte quand le modèle suivant en a besoin aussitôt.
	86	+- Désactive `/unload` et `unload_model` avec `unloadModels: false`, comme l'expert : un agent qui ne fait que répondre n'a pas besoin d'un outil de plus. Les hooks ne dépendent pas de cette clé.

++# Démo mini-me : décharger des modèles
++
++*[English version](README.en.md)*
++
++Un coordinateur confie chaque question de programmation à un expert, et les deux modèles se relaient dans la mémoire du serveur sans jamais y être chargés ensemble : un hook `preAsk` décharge le modèle du coordinateur juste avant le lancement de l'expert, et un hook `postAsk` décharge le modèle de l'expert une fois sa réponse donnée, avant que le coordinateur soit rechargé pour la lire. Les modèles sont servis par [llmman](https://github.com/llmmanorg/llmman).
++
++## Ce qu'il y a ici
++
++| Fichier | Rôle |
++|---|---|
++| `agent.yaml` | Le **coordinateur** (Qwen3.5 9B) : `ask_agent` vers l'expert, puis la réponse. Les deux hooks. `/unload` et l'outil `unload_model` fonctionnent aussi ici. `displayHooks: true` affiche l'exécution des hooks. |
++| `hooks/unload.js` | Le hook (Node.js). `unload.js self` (`preAsk`) : décharge le modèle du coordinateur juste avant que mm lance l'expert. `unload.js asked` (`postAsk`) : décharge le modèle de l'expert une fois sa réponse donnée. |
++| `agent.expert.yaml` | L'**expert** (Qwen3.5 4B) : répond aux questions de programmation, sans aucun outil (`unloadModels: false`). |
++
++## Lancer la démo
++
++```bash
++llmman pull hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M
++llmman pull hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
++cd demos/unload-models
++mm -tui agent.yaml
++```
++
++Pose une question de programmation, puis regarde ce que le serveur garde en mémoire :
++
++```bash
++llmman ps        # ou : curl -s http://127.0.0.1:17434/api/ps
++```
++
++Après la réponse, seul le Qwen3.5 9B du coordinateur est chargé. Dans le terminal, le relais s'affiche ainsi :
++
++```text
++🗣 ask_agent: expert (llmman/hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M) — Explain what a closure is in Ja…
++  🪝 1. preAsk expert · node "$MM_CONFIG_DIR/hooks/unload.js" self → continue
++  🪝 2. postAsk expert · node "$MM_CONFIG_DIR/hooks/unload.js" asked → continue
++```
++
++À la main : `/unload` décharge le modèle du coordinateur, `/unload <modèle>` n'importe quel autre modèle du même serveur — dans le terminal, la fenêtre et un éditeur. Demander au coordinateur de libérer la mémoire de l'expert (« Unload the expert's model. », « Libère la mémoire de l'expert. ») lui fait appeler `unload_model` sur le modèle de l'expert (4 sur 4 mesurés) : son nom est écrit dans le prompt, car le modèle ne peut pas deviner comment le serveur nomme un modèle.
++
++## Comment ça marche
++
++Ce que le serveur garde en mémoire pendant une question (mesuré avec `/api/ps` deux fois par seconde) :
++
++| Moment | Chargé |
++|---|---|
++| le coordinateur lit la question et appelle `ask_agent` | Qwen3.5 9B |
++| le hook `preAsk` décharge le coordinateur | — |
++| l'expert travaille | Qwen3.5 4B seul |
++| le hook `postAsk` décharge l'expert | — |
++| le coordinateur lit la réponse (le serveur le recharge) | Qwen3.5 9B seul |
++
++**Pourquoi des hooks, et pas le modèle.** Les deux déchargements doivent avoir lieu à un moment où le modèle du coordinateur ne tourne pas :
++
++- *Avant l'expert* : le modèle du coordinateur vient de décider d'appeler `ask_agent`, et il n'en a de nouveau besoin que pour lire la réponse. Le modèle ne peut pas se décharger lui-même — `unload_model` sur son propre modèle ne dure que jusqu'à son étape suivante.
++- *Après l'expert* : si le coordinateur appelait `unload_model` sur le modèle de l'expert, il tournerait déjà de nouveau — rechargé à côté du 4B — au moment de le faire. Les deux modèles seraient chargés ensemble pendant toute cette étape. (C'est ce que faisait la version précédente de cette démo ; voir plus bas.)
++
++mm exécute les hooks `preAsk` juste avant de lancer l'agent interrogé et les hooks `postAsk` juste après sa réponse, avant que l'agent qui l'a interrogé la lise. `hooks/unload.js` lit dans son entrée le modèle à décharger — `"model": "llmman/<modèle>"` pour `self`, `"ask": {"agentModel": "llmman/<modèle>"}` pour `asked` — et envoie `keep_alive: 0` au serveur. `matcher: ^expert$` limite les deux hooks à ce relais. `onError: warn` : si un déchargement échoue, le relais continue.
++
++Les hooks sont en plus certains : un modèle peut sauter une consigne, un hook s'exécute toujours.
++
++## Résultats mesurés (2026-10-18, `mm -p`, llmman)
++
++Coordinateur Qwen3.5 9B, expert Qwen3.5 4B, les deux hooks actifs. « Jamais ensemble » : `/api/ps` (interrogé deux fois par seconde) n'a jamais montré les deux modèles à la fois pendant la question : 9B → rien → 4B → rien → 9B.
++
++| Question | `ask_agent` | `preAsk` décharge le 9B | `postAsk` décharge le 4B | jamais ensemble | Durée |
++|---|---|---|---|---|---|
++| How do I reverse a string in Go? | ✓ | ✓ | ✓ | ✓ | 62 s |
++| What is a closure in JavaScript? | ✓ | ✓ | ✓ | ✓ | 65 s |
++| How do I read a file line by line in Python? | ✓ | ✓ | ✓ | ✓ | 49 s |
++| Explain Rust ownership briefly. | ✓ | ✓ | ✓ | ✓ | 82 s |
++| Comment trier un tableau en Java ? | ✓ | ✓ | ✓ | ✓ | 119 s |
++| How do I make an HTTP GET request in Go? | ✓ | ✓ | ✓ | ✓ | 106 s |
++| What does the yield keyword do in Python? (×3) | ✓ ✓ ✓ | ✓ ✓ ✓ | ✓ ✓ ✓ | ✓ ✓ ✓ | 57 s, 60 s, 85 s |
++
++9 sur 9. La question en français reçoit une réponse en français. Une salutation (« Hello, who are you? ») reçoit une réponse en 2 s sans aucun outil. Les durées comprennent les chargements : chaque question charge le 9B deux fois et le 4B une fois, et elles varient beaucoup d'une exécution à l'autre.
++
++**Le prompt.** Sans l'étape `unload_model`, un prompt court (« 1. Call ask_agent… 2. Then give the expert's answer ») a laissé le coordinateur répondre seul à la question sur `yield`, sans l'expert (1 fois sur 7). Préciser que `ask_agent` est TOUJOURS la première action, « even for a question you know the answer to », a donné le tableau ci-dessus.
++
++**Versions précédentes.** Quand le coordinateur libérait l'expert avec `unload_model` (sans hook `postAsk`), le 9B et le 4B étaient chargés ensemble pendant ~18 s par question, le temps que le coordinateur lise la réponse — 7 questions sur 7 sans autre défaut. Avant cela, avec un coordinateur MiniCPM5 2B et un expert Qwen3.5 9B, le coordinateur a sauté `unload_model` une fois sur cinq.
++
++## Ce qu'il faut savoir
++
++- **`postAsk` s'exécute après chaque relais vers l'expert**, qu'il ait répondu, échoué ou été annulé : le modèle de l'expert est déchargé même quand son tour a échoué.
++- **La question suivante recharge le 4B** : un relais coûte un chargement de chaque modèle, plus un rechargement du coordinateur. Garder les deux chargés est plus rapide, quand la mémoire le permet.
++- **llmman décharge aussi tout seul** un modèle inactif depuis cinq minutes (`keep_alive`) ; les hooks libèrent la mémoire tout de suite, ce qui compte quand le modèle suivant en a besoin aussitôt.
++- **Désactive `/unload` et `unload_model`** avec `unloadModels: false`, comme l'expert : un agent qui ne fait que répondre n'a pas besoin d'un outil de plus. Les hooks ne dépendent pas de cette clé.

added demos/unload-models/agent.expert.yaml +37 -0 Viewed

new file mode 100644
@@ -0,0 +1,37 @@
	1	+# The expert of demos/unload-models: answers programming questions. Asked by
	2	+# agent.yaml through ask_agent; its model is unloaded after each answer.
	3	+
	4	+provider: llmman
	5	+model: hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
	6	+baseUrl: http://127.0.0.1:17434/v1
	7	+fallback: http://host.docker.internal:17434/v1
	8	+
	9	+sampling:
	10	+ temperature: 0.2
	11	+ max_tokens: 2048
	12	+ chat_template_kwargs:
	13	+ enable_thinking: false
	14	+
	15	+watchdogTimeout: 5m
	16	+firstTokenTimeout: 3m
	17	+maxTurns: 2
	18	+
	19	+# The expert only answers: no tools at all, unload_model included.
	20	+unloadModels: false
	21	+bashTool: false
	22	+editTools: false
	23	+planTool: false
	24	+fetchContentTool: false
	25	+askUserTool: false
	26	+skillsDir: no-skills
	27	+
	28	+agentsMd:
	29	+ enabled: false
	30	+
	31	+acp:
	32	+ name: expert
	33	+ title: Programming expert
	34	+
	35	+system: \|
	36	+ You are a programming expert. Answer the question clearly and briefly:
	37	+ a short explanation, then one short code example when it helps.

++# The expert of demos/unload-models: answers programming questions. Asked by
++# agent.yaml through ask_agent; its model is unloaded after each answer.
++
++provider: llmman
++model: hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
++baseUrl: http://127.0.0.1:17434/v1
++fallback: http://host.docker.internal:17434/v1
++
++sampling:
++  temperature: 0.2
++  max_tokens: 2048
++  chat_template_kwargs:
++    enable_thinking: false
++
++watchdogTimeout: 5m
++firstTokenTimeout: 3m
++maxTurns: 2
++
++# The expert only answers: no tools at all, unload_model included.
++unloadModels: false
++bashTool: false
++editTools: false
++planTool: false
++fetchContentTool: false
++askUserTool: false
++skillsDir: no-skills
++
++agentsMd:
++  enabled: false
++
++acp:
++  name: expert
++  title: Programming expert
++
++system: |
++  You are a programming expert. Answer the question clearly and briefly:
++  a short explanation, then one short code example when it helps.

added demos/unload-models/agent.yaml +108 -0 Viewed

new file mode 100644
@@ -0,0 +1,108 @@
	1	+# Unload models — a coordinator hands each programming question to an expert,
	2	+# and the two models take turns in memory, never loaded together:
	3	+#
	4	+# hooks/unload.js self preAsk unloads the COORDINATOR's model right
	5	+# before the expert starts (only mm can do
	6	+# it at that moment)
	7	+# hooks/unload.js asked postAsk unloads the EXPERT's model once it has
	8	+# answered, before the coordinator is
	9	+# reloaded to read the answer
	10	+#
	11	+# coordinator Qwen3.5 9B hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M (this file)
	12	+# expert Qwen3.5 4B hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M (agent.expert.yaml)
	13	+#
	14	+# llmman pull hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M
	15	+# llmman pull hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
	16	+# llmman serve
	17	+#
	18	+# cd demos/unload-models
	19	+# mm -tui agent.yaml or mm -config agent.yaml -p "…"
	20	+#
	21	+# Watch the server between questions: llmman ps (or curl :17434/api/ps).
	22	+# `/unload` unloads the coordinator's own model by hand, `/unload <model>`
	23	+# any other one — in the terminal, the window and an editor.
	24	+
	25	+provider: llmman
	26	+model: hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M
	27	+baseUrl: http://127.0.0.1:17434/v1
	28	+fallback: http://host.docker.internal:17434/v1
	29	+
	30	+mmConfigs:
	31	+ expert:
	32	+ path: agent.expert.yaml
	33	+ description: >-
	34	+ Programming expert. Answers any question about code, languages,
	35	+ tools and how to do something in a program.
	36	+ yolo: true
	37	+
	38	+# /unload and unload_model are on by default (unloadModels: true); written
	39	+# here because this demo is about them. The hooks do the hand-off's unloads:
	40	+# a step of the model would come too late (the coordinator is already
	41	+# reloaded beside the expert) and a small model may skip it. unload_model
	42	+# stays for when the user asks to free a model.
	43	+unloadModels: true
	44	+
	45	+sampling:
	46	+ # Low temperature: routing wants a steady hand.
	47	+ temperature: 0.4
	48	+ top_p: 0.8
	49	+ top_k: 20
	50	+ max_tokens: 2048
	51	+ parallel_tool_calls: false
	52	+ chat_template_kwargs:
	53	+ enable_thinking: false
	54	+
	55	+# During ask_agent every event of the expert counts as a sign of life, but
	56	+# loading a 9B model is silent for a while.
	57	+watchdogTimeout: 5m
	58	+firstTokenTimeout: 3m
	59	+
	60	+# ask the expert, answer: two steps, one spare.
	61	+maxTurns: 3
	62	+maxOutput: 20000
	63	+displayCommands: true
	64	+displayHooks: true
	65	+
	66	+hooks:
	67	+ # onError: warn — an unload is a saving, not a check: the hand-off goes on.
	68	+ preAsk:
	69	+ - matcher: ^expert$
	70	+ command: node "$MM_CONFIG_DIR/hooks/unload.js" self
	71	+ timeout: 30s
	72	+ onError: warn
	73	+ postAsk:
	74	+ - matcher: ^expert$
	75	+ command: node "$MM_CONFIG_DIR/hooks/unload.js" asked
	76	+ timeout: 30s
	77	+ onError: warn
	78	+
	79	+yolo: true
	80	+bashTool: false
	81	+editTools: false
	82	+planTool: false
	83	+fetchContentTool: false
	84	+askUserTool: false
	85	+skillsDir: no-skills
	86	+
	87	+agentsMd:
	88	+ enabled: false
	89	+
	90	+acp:
	91	+ name: unloader
	92	+ title: 🧹 Unload models
	93	+
	94	+system: \|
	95	+ You are a coordinator. You do not answer programming questions yourself.
	96	+
	97	+ For every question about programming, even one you know the answer to,
	98	+ do these two steps, in order:
	99	+ 1. Call ask_agent with agent "expert" and the user's question. It is
	100	+ ALWAYS your first action: never answer a programming question before
	101	+ the expert has answered it.
	102	+ 2. Then give the expert's answer to the user, in the user's language.
	103	+
	104	+ If the user asks you to free or unload a model, call unload_model. The
	105	+ expert's model is "hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M".
	106	+
	107	+ For anything else (a greeting, a question about you), answer in one short
	108	+ sentence, without any tool.

++# Unload models — a coordinator hands each programming question to an expert,
++# and the two models take turns in memory, never loaded together:
++#
++#   hooks/unload.js self   preAsk   unloads the COORDINATOR's model right
++#                                   before the expert starts (only mm can do
++#                                   it at that moment)
++#   hooks/unload.js asked  postAsk  unloads the EXPERT's model once it has
++#                                   answered, before the coordinator is
++#                                   reloaded to read the answer
++#
++#   coordinator  Qwen3.5 9B hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M (this file)
++#   expert       Qwen3.5 4B hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M (agent.expert.yaml)
++#
++#   llmman pull hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M
++#   llmman pull hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
++#   llmman serve
++#
++#   cd demos/unload-models
++#   mm -tui agent.yaml           or   mm -config agent.yaml -p "…"
++#
++# Watch the server between questions: llmman ps (or curl :17434/api/ps).
++# `/unload` unloads the coordinator's own model by hand, `/unload <model>`
++# any other one — in the terminal, the window and an editor.
++
++provider: llmman
++model: hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M
++baseUrl: http://127.0.0.1:17434/v1
++fallback: http://host.docker.internal:17434/v1
++
++mmConfigs:
++  expert:
++    path: agent.expert.yaml
++    description: >-
++      Programming expert. Answers any question about code, languages,
++      tools and how to do something in a program.
++    yolo: true
++
++# /unload and unload_model are on by default (unloadModels: true); written
++# here because this demo is about them. The hooks do the hand-off's unloads:
++# a step of the model would come too late (the coordinator is already
++# reloaded beside the expert) and a small model may skip it. unload_model
++# stays for when the user asks to free a model.
++unloadModels: true
++
++sampling:
++  # Low temperature: routing wants a steady hand.
++  temperature: 0.4
++  top_p: 0.8
++  top_k: 20
++  max_tokens: 2048
++  parallel_tool_calls: false
++  chat_template_kwargs:
++    enable_thinking: false
++
++# During ask_agent every event of the expert counts as a sign of life, but
++# loading a 9B model is silent for a while.
++watchdogTimeout: 5m
++firstTokenTimeout: 3m
++
++# ask the expert, answer: two steps, one spare.
++maxTurns: 3
++maxOutput: 20000
++displayCommands: true
++displayHooks: true
++
++hooks:
++  # onError: warn — an unload is a saving, not a check: the hand-off goes on.
++  preAsk:
++    - matcher: ^expert$
++      command: node "$MM_CONFIG_DIR/hooks/unload.js" self
++      timeout: 30s
++      onError: warn
++  postAsk:
++    - matcher: ^expert$
++      command: node "$MM_CONFIG_DIR/hooks/unload.js" asked
++      timeout: 30s
++      onError: warn
++
++yolo: true
++bashTool: false
++editTools: false
++planTool: false
++fetchContentTool: false
++askUserTool: false
++skillsDir: no-skills
++
++agentsMd:
++  enabled: false
++
++acp:
++  name: unloader
++  title: 🧹 Unload models
++
++system: |
++  You are a coordinator. You do not answer programming questions yourself.
++
++  For every question about programming, even one you know the answer to,
++  do these two steps, in order:
++  1. Call ask_agent with agent "expert" and the user's question. It is
++     ALWAYS your first action: never answer a programming question before
++     the expert has answered it.
++  2. Then give the expert's answer to the user, in the user's language.
++
++  If the user asks you to free or unload a model, call unload_model. The
++  expert's model is "hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M".
++
++  For anything else (a greeting, a question about you), answer in one short
++  sentence, without any tool.

added demos/unload-models/hooks/unload.js +58 -0 Viewed

new file mode 100644
@@ -0,0 +1,58 @@
	1	+#!/usr/bin/env node
	2	+// preAsk / postAsk hook: unloads one model so that two models are never in
	3	+// memory together during a hand-off.
	4	+//
	5	+// node unload.js self preAsk: the ASKING agent's own model, right before
	6	+// mm starts the asked agent. Only mm can do it at
	7	+// that moment: the model has just decided to call
	8	+// ask_agent, and it is needed again only to read the
	9	+// answer — the server reloads it then by itself. The
	10	+// model cannot do it (unload_model on its own model
	11	+// lasts until its next step).
	12	+// node unload.js asked postAsk: the ASKED agent's model, once it has
	13	+// answered and before the asking agent reads the
	14	+// answer — so the asker's model reloads alone.
	15	+//
	16	+// Input: {"model": "llmman/<model>", "ask": {"agentModel": "llmman/<model>"},
	17	+// ...}. Server: $LLMMAN_URL, else llmman's default address, else the one a
	18	+// Docker sandbox reaches the host by. Prints nothing: the hand-off goes on
	19	+// whatever happens (onError: warn in agent.yaml), an unload is a saving, not
	20	+// a check.
	21	+
	22	+const fs = require("fs");
	23	+
	24	+const input = JSON.parse(fs.readFileSync(0, "utf8"));
	25	+const which = process.argv[2];
	26	+const ref = which === "asked" ? input.ask && input.ask.agentModel : input.model;
	27	+if (!ref) {
	28	+ console.error(`unload.js ${which}: no model in the input`);
	29	+ process.exit(1);
	30	+}
	31	+const model = ref.replace(/^[^/]+\//, ""); // drop the "llmman/" provider prefix
	32	+
	33	+const servers = process.env.LLMMAN_URL
	34	+ ? [process.env.LLMMAN_URL]
	35	+ : ["http://127.0.0.1:17434", "http://host.docker.internal:17434"];
	36	+
	37	+async function unload() {
	38	+ for (const server of servers) {
	39	+ try {
	40	+ const resp = await fetch(server + "/api/generate", {
	41	+ method: "POST",
	42	+ headers: { "Content-Type": "application/json" },
	43	+ body: JSON.stringify({ model, keep_alive: 0 }),
	44	+ signal: AbortSignal.timeout(20_000),
	45	+ });
	46	+ const reply = await resp.json();
	47	+ if (reply.done_reason === "unload") return;
	48	+ throw new Error(`${server}: not confirmed: ${JSON.stringify(reply)}`);
	49	+ } catch (err) {
	50	+ if (server === servers[servers.length - 1]) throw err;
	51	+ }
	52	+ }
	53	+}
	54	+
	55	+unload().catch((err) => {
	56	+ console.error(`could not unload ${model}: ${err.message}`);
	57	+ process.exit(1);
	58	+});

++#!/usr/bin/env node
++// preAsk / postAsk hook: unloads one model so that two models are never in
++// memory together during a hand-off.
++//
++//   node unload.js self    preAsk: the ASKING agent's own model, right before
++//                          mm starts the asked agent. Only mm can do it at
++//                          that moment: the model has just decided to call
++//                          ask_agent, and it is needed again only to read the
++//                          answer — the server reloads it then by itself. The
++//                          model cannot do it (unload_model on its own model
++//                          lasts until its next step).
++//   node unload.js asked   postAsk: the ASKED agent's model, once it has
++//                          answered and before the asking agent reads the
++//                          answer — so the asker's model reloads alone.
++//
++// Input: {"model": "llmman/<model>", "ask": {"agentModel": "llmman/<model>"},
++// ...}. Server: $LLMMAN_URL, else llmman's default address, else the one a
++// Docker sandbox reaches the host by. Prints nothing: the hand-off goes on
++// whatever happens (onError: warn in agent.yaml), an unload is a saving, not
++// a check.
++
++const fs = require("fs");
++
++const input = JSON.parse(fs.readFileSync(0, "utf8"));
++const which = process.argv[2];
++const ref = which === "asked" ? input.ask && input.ask.agentModel : input.model;
++if (!ref) {
++  console.error(`unload.js ${which}: no model in the input`);
++  process.exit(1);
++}
++const model = ref.replace(/^[^/]+\//, ""); // drop the "llmman/" provider prefix
++
++const servers = process.env.LLMMAN_URL
++  ? [process.env.LLMMAN_URL]
++  : ["http://127.0.0.1:17434", "http://host.docker.internal:17434"];
++
++async function unload() {
++  for (const server of servers) {
++    try {
++      const resp = await fetch(server + "/api/generate", {
++        method: "POST",
++        headers: { "Content-Type": "application/json" },
++        body: JSON.stringify({ model, keep_alive: 0 }),
++        signal: AbortSignal.timeout(20_000),
++      });
++      const reply = await resp.json();
++      if (reply.done_reason === "unload") return;
++      throw new Error(`${server}: not confirmed: ${JSON.stringify(reply)}`);
++    } catch (err) {
++      if (server === servers[servers.length - 1]) throw err;
++    }
++  }
++}
++
++unload().catch((err) => {
++  console.error(`could not unload ${model}: ${err.message}`);
++  process.exit(1);
++});

modified docs/en/README.md +2 -1 Viewed

@@ -27,6 +27,7 @@ This documentation is organised along the four [Diátaxis](https://diataxis.fr)
27	27	- [Add a skill](how-to/add-a-skill.md)
28	28	- [Declare a custom tool](how-to/declare-a-custom-tool.md)
29	29	- [Save an answer to a file](how-to/save-an-answer.md)
	30	+- [Free a model's memory](how-to/unload-a-model.md) — `/unload [model]` in every front end, the `unload_model` tool after another agent's answer, `unloadModels`
30	31	- [Measure whether the model picks the right skill](how-to/measure-skill-routing.md) — the routing eval: run it, read its table, use it as a before-and-after
31	32	- [Keep a long session inside the context window](how-to/manage-the-context-window.md)
32	33	- [Start a new session without restarting](how-to/start-a-new-session.md)
@@ -51,7 +52,7 @@ This documentation is organised along the four [Diátaxis](https://diataxis.fr)
51	52	- [Command line](reference/cli.md) — flags, environment variables, start-up banner
52	53	- [Slash commands](reference/slash-commands.md) — `/quit`, `/abort`, `/compact`, `/new`, `/sessions`, `/resume`, `/help`, keyboard shortcuts, window keys
53	54	- [Configuration file](reference/configuration.md) — every YAML key, its default and its validation rule
54		-- [Built-in tools](reference/tools.md) — `bash`, `read_skill`, `read_file`, `write_file`, `edit_file`, `lsp`, `fetch_content`, `ask_agent`, `ask_user`, `save_answer`, declared tools
	55	+- [Built-in tools](reference/tools.md) — `bash`, `read_skill`, `read_file`, `write_file`, `edit_file`, `lsp`, `fetch_content`, `ask_agent`, `ask_user`, `save_answer`, `unload_model`, declared tools
55	56	- [Skill file format](reference/skill-format.md)
56	57	- [Hooks](reference/hooks.md) — the six events, the JSON a hook reads and writes, `matcher`, `timeout`, `onError`, validation errors
57	58	- [Agent Client Protocol surface](reference/acp.md)

  - [Add a skill](how-to/add-a-skill.md)
  - [Declare a custom tool](how-to/declare-a-custom-tool.md)
  - [Save an answer to a file](how-to/save-an-answer.md)
++- [Free a model's memory](how-to/unload-a-model.md) — `/unload [model]` in every front end, the `unload_model` tool after another agent's answer, `unloadModels`
  - [Measure whether the model picks the right skill](how-to/measure-skill-routing.md) — the routing eval: run it, read its table, use it as a before-and-after
  - [Keep a long session inside the context window](how-to/manage-the-context-window.md)
  - [Start a new session without restarting](how-to/start-a-new-session.md)
  - [Command line](reference/cli.md) — flags, environment variables, start-up banner
  - [Slash commands](reference/slash-commands.md) — `/quit`, `/abort`, `/compact`, `/new`, `/sessions`, `/resume`, `/help`, keyboard shortcuts, window keys
  - [Configuration file](reference/configuration.md) — every YAML key, its default and its validation rule
--- [Built-in tools](reference/tools.md) — `bash`, `read_skill`, `read_file`, `write_file`, `edit_file`, `lsp`, `fetch_content`, `ask_agent`, `ask_user`, `save_answer`, declared tools
++- [Built-in tools](reference/tools.md) — `bash`, `read_skill`, `read_file`, `write_file`, `edit_file`, `lsp`, `fetch_content`, `ask_agent`, `ask_user`, `save_answer`, `unload_model`, declared tools
  - [Skill file format](reference/skill-format.md)
  - [Hooks](reference/hooks.md) — the six events, the JSON a hook reads and writes, `matcher`, `timeout`, `onError`, validation errors
  - [Agent Client Protocol surface](reference/acp.md)

added docs/en/how-to/unload-a-model.md +58 -0 Viewed

new file mode 100644
@@ -0,0 +1,58 @@
	1	+# How to free a model's memory
	2	+
	3	+This guide shows how to unload a model from the model server's memory without stopping the server: by hand with `/unload`, or by the agent itself with its `unload_model` tool — typically to free the memory of another agent's model once its answer is in. It assumes mm runs on llmman, Ollama, or llama-server in router mode. The model reloads by itself the next time something uses it.
	4	+
	5	+## Unload by hand
	6	+
	7	+Type, in the terminal, the window or an editor:
	8	+
	9	+```text
	10	+/unload
	11	+```
	12	+
	13	+The agent's own model is unloaded:
	14	+
	15	+```text
	16	+🧹 unloaded hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M
	17	+```
	18	+
	19	+Name another model served by the same server to unload that one instead:
	20	+
	21	+```text
	22	+/unload hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M
	23	+```
	24	+
	25	+A model that is not loaded gives `🧹 <model> was not loaded: nothing to unload`; nothing is sent to the server, so nothing is loaded either. Check what is loaded with `llmman ps`, `ollama ps`, or `curl <server>/api/ps`.
	26	+
	27	+## Let the agent unload a model
	28	+
	29	+The `unload_model` tool is declared by default. Tell the model when to use it, in its `system` prompt. The case it serves best is another agent's model, once that agent has answered:
	30	+
	31	+```yaml
	32	+mmConfigs:
	33	+ expert:
	34	+ path: agent.expert.yaml
	35	+ description: Programming expert.
	36	+
	37	+system: \|
	38	+ For every question about programming, do these three steps, in order:
	39	+ 1. Call ask_agent with agent "expert" and the user's question.
	40	+ 2. Call unload_model with model "hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M".
	41	+ It is ALWAYS your next action after ask_agent: never answer the user
	42	+ before you have called unload_model.
	43	+ 3. Then give the expert's answer to the user.
	44	+```
	45	+
	46	+Write the model's name exactly as the server lists it: the model cannot guess it. `demos/unload-models/` does the expert's unload with a `postAsk` hook instead (see the variants below); its coordinator still calls `unload_model` when the user asks it to free the expert's memory.
	47	+
	48	+## Variants
	49	+
	50	+- The agent's own model. `unload_model` with an empty `model` unloads it, but only until the agent's next step: the agent needs its model to read the tool's result, and the server reloads it. Use `/unload` between two questions instead — or, to free it while another agent works, a `preAsk` hook that unloads the model named in its input (`"model": "<provider>/<model>"`) right before the asked agent starts: `demos/unload-models/hooks/unload.js self`.
	51	+- A guarantee instead of an instruction. A small model may skip the step (measured: 4 times out of 5 with MiniCPM5 2B in the first version of `demos/unload-models`), and the agent's own model is already reloaded beside the other one when it calls the tool. For a certain unload, with the two models never loaded together, have a `postAsk` hook call the server itself — `curl -s <server>/api/generate -d '{"model":"…","keep_alive":0}'` — once the asked agent has answered and before the asking agent reads the answer: `demos/unload-models/hooks/unload.js asked` reads the asked agent's model from `ask.agentModel`. See [Control the hand-offs with hooks](control-agent-calls-with-hooks.md).
	52	+- llama-server. Only router mode (started without `-m`) can unload a model. A single-model server answers with the alternatives: `--sleep-idle-seconds N` frees the memory after N idle seconds, or stop the server.
	53	+- Turn it off. `unloadModels: false` removes both `/unload` and the tool — for an agent that must not have one more tool, such as a guard or a narrow specialist.
	54	+
	55	+## See also
	56	+
	57	+- [`unload_model` reference](../reference/tools.md#unload_model) and [`/unload`](../reference/slash-commands.md#commands)
	58	+- [`unloadModels`](../reference/configuration.md)

++# How to free a model's memory
++
++This guide shows how to unload a model from the model server's memory without stopping the server: by hand with `/unload`, or by the agent itself with its `unload_model` tool — typically to free the memory of another agent's model once its answer is in. It assumes mm runs on llmman, Ollama, or llama-server in router mode. The model reloads by itself the next time something uses it.
++
++## Unload by hand
++
++Type, in the terminal, the window or an editor:
++
++```text
++/unload
++```
++
++The agent's own model is unloaded:
++
++```text
++🧹 unloaded hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M
++```
++
++Name another model served by the same server to unload that one instead:
++
++```text
++/unload hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M
++```
++
++A model that is not loaded gives `🧹 <model> was not loaded: nothing to unload`; nothing is sent to the server, so nothing is loaded either. Check what is loaded with `llmman ps`, `ollama ps`, or `curl <server>/api/ps`.
++
++## Let the agent unload a model
++
++The `unload_model` tool is declared by default. Tell the model when to use it, in its `system` prompt. The case it serves best is another agent's model, once that agent has answered:
++
++```yaml
++mmConfigs:
++  expert:
++    path: agent.expert.yaml
++    description: Programming expert.
++
++system: |
++  For every question about programming, do these three steps, in order:
++  1. Call ask_agent with agent "expert" and the user's question.
++  2. Call unload_model with model "hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M".
++     It is ALWAYS your next action after ask_agent: never answer the user
++     before you have called unload_model.
++  3. Then give the expert's answer to the user.
++```
++
++Write the model's name exactly as the server lists it: the model cannot guess it. `demos/unload-models/` does the expert's unload with a `postAsk` hook instead (see the variants below); its coordinator still calls `unload_model` when the user asks it to free the expert's memory.
++
++## Variants
++
++- **The agent's own model.** `unload_model` with an empty `model` unloads it, but only until the agent's next step: the agent needs its model to read the tool's result, and the server reloads it. Use `/unload` between two questions instead — or, to free it while another agent works, a `preAsk` hook that unloads the model named in its input (`"model": "<provider>/<model>"`) right before the asked agent starts: `demos/unload-models/hooks/unload.js self`.
++- **A guarantee instead of an instruction.** A small model may skip the step (measured: 4 times out of 5 with MiniCPM5 2B in the first version of `demos/unload-models`), and the agent's own model is already reloaded beside the other one when it calls the tool. For a certain unload, with the two models never loaded together, have a `postAsk` hook call the server itself — `curl -s <server>/api/generate -d '{"model":"…","keep_alive":0}'` — once the asked agent has answered and before the asking agent reads the answer: `demos/unload-models/hooks/unload.js asked` reads the asked agent's model from `ask.agentModel`. See [Control the hand-offs with hooks](control-agent-calls-with-hooks.md).
++- **llama-server.** Only router mode (started without `-m`) can unload a model. A single-model server answers with the alternatives: `--sleep-idle-seconds N` frees the memory after N idle seconds, or stop the server.
++- **Turn it off.** `unloadModels: false` removes both `/unload` and the tool — for an agent that must not have one more tool, such as a guard or a narrow specialist.
++
++## See also
++
++- [`unload_model` reference](../reference/tools.md#unload_model) and [`/unload`](../reference/slash-commands.md#commands)
++- [`unloadModels`](../reference/configuration.md)

modified docs/en/reference/configuration.md +2 -1 Viewed

@@ -26,6 +26,7 @@
26	26	\| `fetchContentTool` \| bool \| `false` \| Declare `fetch_content`, which fetches web pages as Markdown. Off by default: it is the only built-in tool that reaches the network. \|
27	27	\| `askUserTool` \| bool \| `true` \| Declare `ask_user`, through which the model asks the user questions (pick one, pick several, type a text, yes or no) and waits for the answers. `false` removes it: the model cannot ask. Never declared beside a `structuredOutput`. Overridden by `AGENT_ASK_USER`; for an agent asked by another one, also by the asker's `mmConfigs.<name>.askUser`. See [tools](tools.md#ask_user). \|
28	28	\| `saveAnswerTool` \| bool \| `false` \| Declare `save_answer`, which writes an answer the model already gave to a new file: the model sends a file name, not the text again. Off by default. The `/save` command exists either way. See [tools](tools.md#save_answer). \|
	29	+\| `unloadModels` \| bool \| `true` \| The `/unload [model]` command (terminal, window, editor) and the `unload_model` tool, which free a model from the server's memory: llmman and Ollama, llama-server in router mode. `false` removes both. The tool is never declared beside a `structuredOutput`. See [tools](tools.md#unload_model) and [How to free a model's memory](../how-to/unload-a-model.md). \|
29	30	\| `sampling` \| map \| `temperature: 0.0`, `top_p: 0.9`, `max_tokens: 4096` \| Generation settings, OpenAI API key names, passed to the server. A key the OpenAI API does not declare is sent as it is, at the top level of the request — for example `chat_template_kwargs: {enable_thinking: false}`, which tells a llama.cpp server (`dmr`, `llamacpp`, `llmman`) not to reason; `reasoning_effort: none` does the same on Ollama. `messages`, `tools`, `tool_choice` (and the legacy `functions`, `function_call`) are ignored: Genkit builds them from the conversation. The shipped `agent.yaml` adds `parallel_tool_calls: false`. `max_tokens` also bounds one tool call: a `write_file` of a long document (or a thinking model's reasoning plus the call) can hit it, and the call is then cut off. mm hands it back to the model twice, asking for smaller pieces, then stops with `the model hit max_tokens (N) while writing a tool call`; raise the value if that happens. With streaming, mm also asks the server for token counts (`stream_options.include_usage`). \|
30	31	\| `watchdogTimeout` \| duration \| `20s` \| Silence tolerated between two streamed tokens before the connection is declared hung. Doubled for summary requests. \|
31	32	\| `firstTokenTimeout` \| duration \| `3m` \| Silence tolerated before a model call's first token, while the server reads the prompt (every round trip of a turn has one). Never shorter than `watchdogTimeout`; must be >= 0. With a llama.cpp server (`dmr`, `llamacpp`, `llmman` — whose local GGUF models run in llama-server) mm asks for `return_progress`: each progress report restarts this clock, and the spinner shows `Reading the prompt 12.0k/25.3k tokens`. Ollama sends no progress: this one wait covers the whole prompt (and the model's loading). On a timeout the watchdog says `no first token from the server after …` (before the first token) or `connection hang detected after …` (after). \|
@@ -128,7 +129,7 @@ The shipped `agent.lsp.yaml` is `agent.dmr.yaml` with this block on (gopls activ
128	129	\| Key \| Type \| Default \| Description \|
129	130	\|-----\|------\|---------\|-------------\|
130	131	\| `tools` \| list \| `[]` \| One entry per declared tool. \|
131		-\| `tools[].name` \| string \| — \| What the model calls. Required. Starts with a letter; letters, digits, `_`, `-`; at most 64 characters. Unique. Not the name of a built-in tool (`ask_agent`, `ask_user`, `bash`, `edit_file`, `fetch_content`, `lsp`, `plan`, `read_file`, `read_skill`, `save_answer`, `write_file`), even one this agent turns off. \|
	132	+\| `tools[].name` \| string \| — \| What the model calls. Required. Starts with a letter; letters, digits, `_`, `-`; at most 64 characters. Unique. Not the name of a built-in tool (`ask_agent`, `ask_user`, `bash`, `edit_file`, `fetch_content`, `lsp`, `plan`, `read_file`, `read_skill`, `save_answer`, `unload_model`, `write_file`), even one this agent turns off. \|
132	133	\| `tools[].description` \| string \| — \| What the tool does and when to use it, shown to the model. Required. A line `Required parameters: …` is appended. \|
133	134	\| `tools[].parameters` \| map name → parameter \| `{}` \| The arguments, in the order written. Names follow the `name` rules. \|
134	135	\| `tools[].parameters.<p>.type` \| string \| `string` \| `string`, `number`, `integer` or `boolean`. \|

  | `fetchContentTool` | bool | `false` | Declare `fetch_content`, which fetches web pages as Markdown. Off by default: it is the only built-in tool that reaches the network. |
  | `askUserTool` | bool | `true` | Declare `ask_user`, through which the model asks the user questions (pick one, pick several, type a text, yes or no) and waits for the answers. `false` removes it: the model cannot ask. Never declared beside a `structuredOutput`. Overridden by `AGENT_ASK_USER`; for an agent asked by another one, also by the asker's `mmConfigs.<name>.askUser`. See [tools](tools.md#ask_user). |
  | `saveAnswerTool` | bool | `false` | Declare `save_answer`, which writes an answer the model already gave to a new file: the model sends a file name, not the text again. Off by default. The `/save` command exists either way. See [tools](tools.md#save_answer). |
++| `unloadModels` | bool | `true` | The `/unload [model]` command (terminal, window, editor) and the `unload_model` tool, which free a model from the server's memory: llmman and Ollama, llama-server in router mode. `false` removes both. The tool is never declared beside a `structuredOutput`. See [tools](tools.md#unload_model) and [How to free a model's memory](../how-to/unload-a-model.md). |
  | `sampling` | map | `temperature: 0.0`, `top_p: 0.9`, `max_tokens: 4096` | Generation settings, OpenAI API key names, passed to the server. A key the OpenAI API does not declare is sent as it is, at the top level of the request — for example `chat_template_kwargs: {enable_thinking: false}`, which tells a llama.cpp server (`dmr`, `llamacpp`, `llmman`) not to reason; `reasoning_effort: none` does the same on Ollama. `messages`, `tools`, `tool_choice` (and the legacy `functions`, `function_call`) are ignored: Genkit builds them from the conversation. The shipped `agent.yaml` adds `parallel_tool_calls: false`. `max_tokens` also bounds one tool call: a `write_file` of a long document (or a thinking model's reasoning plus the call) can hit it, and the call is then cut off. mm hands it back to the model twice, asking for smaller pieces, then stops with `the model hit max_tokens (N) while writing a tool call`; raise the value if that happens. With streaming, mm also asks the server for token counts (`stream_options.include_usage`). |
  | `watchdogTimeout` | duration | `20s` | Silence tolerated between two streamed tokens before the connection is declared hung. Doubled for summary requests. |
  | `firstTokenTimeout` | duration | `3m` | Silence tolerated before a model call's first token, while the server reads the prompt (every round trip of a turn has one). Never shorter than `watchdogTimeout`; must be >= 0. With a llama.cpp server (`dmr`, `llamacpp`, `llmman` — whose local GGUF models run in llama-server) mm asks for `return_progress`: each progress report restarts this clock, and the spinner shows `Reading the prompt 12.0k/25.3k tokens`. Ollama sends no progress: this one wait covers the whole prompt (and the model's loading). On a timeout the watchdog says `no first token from the server after …` (before the first token) or `connection hang detected after …` (after). |
  | Key | Type | Default | Description |
  |-----|------|---------|-------------|
  | `tools` | list | `[]` | One entry per declared tool. |
--| `tools[].name` | string | — | What the model calls. Required. Starts with a letter; letters, digits, `_`, `-`; at most 64 characters. Unique. Not the name of a built-in tool (`ask_agent`, `ask_user`, `bash`, `edit_file`, `fetch_content`, `lsp`, `plan`, `read_file`, `read_skill`, `save_answer`, `write_file`), even one this agent turns off. |
++| `tools[].name` | string | — | What the model calls. Required. Starts with a letter; letters, digits, `_`, `-`; at most 64 characters. Unique. Not the name of a built-in tool (`ask_agent`, `ask_user`, `bash`, `edit_file`, `fetch_content`, `lsp`, `plan`, `read_file`, `read_skill`, `save_answer`, `unload_model`, `write_file`), even one this agent turns off. |
  | `tools[].description` | string | — | What the tool does and when to use it, shown to the model. Required. A line `Required parameters: …` is appended. |
  | `tools[].parameters` | map name → parameter | `{}` | The arguments, in the order written. Names follow the `name` rules. |
  | `tools[].parameters.<p>.type` | string | `string` | `string`, `number`, `integer` or `boolean`. |

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

@@ -15,6 +15,7 @@
15	15	\| `/help` \| Prints the commands of this front end, one per line with their description, aligned. \| Sent to the model as a new message. \|
16	16	\| `/copy [n\\|all]` \| Copies the last answer (`n`: the n-th from the end) to the clipboard through `pbcopy`/`wl-copy`/`xclip`/`xsel` or, failing that, OSC 52 when stdout is a terminal, and prints `📋 copied N chars (<road>)` or `⚠️ …`. `/copy all` is window-mode only: here it says so. \| Sent to the model as a new message. \|
17	17	\| `/save <file> [n]` \| Writes the last answer (`n`: the n-th from the end) as it is to a new file, creating the missing directories, and prints `💾 saved the last answer to <file> (N bytes)`. Never overwrites: an existing file gives `⚠️ <file> already exists — nothing was written; choose another name`. A file name with spaces is not supported. The model is not called. \| Sent to the model as a new message. \|
	18	+\| `/unload [model]` \| With `unloadModels` (default on): unloads a model from the server's memory — this agent's own, or the one named — and prints one line: `🧹 unloaded <model>`, `🧹 <model> was not loaded: nothing to unload`, or `🧹 cannot unload <model>: <reason>` (a provider or a server that has no way to). The history is not touched; the next question reloads the agent's model. \| Sent to the model as a new message. \|
18	19
19	20	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.
20	21
@@ -169,12 +170,13 @@ The notation works in the line mode and over ACP too, typed by hand: a `$skill`
169	170
170	171	## ACP mode
171	172
172		-Two commands exist in ACP mode: `/new` and `/compact`. They are advertised to the editor through an `available_commands_update` notification and intercepted by `session/prompt` when the prompt text, trimmed, is exactly the command. Each answers with one `agent_message_chunk` carrying the line the terminal would print, then `stopReason: end_turn`; the model is not consulted.
	173	+Two commands exist in ACP mode, three with `unloadModels` (the default): `/new`, `/compact` and `/unload`. They are advertised to the editor through an `available_commands_update` notification and intercepted by `session/prompt` when the prompt text, trimmed, is exactly the command (`/unload` may be followed by a model name). Each answers with one `agent_message_chunk` carrying the line the terminal would print, then `stopReason: end_turn`; the model is not consulted.
173	174
174	175	\| Command \| Over ACP \|
175	176	\|---------\|----------\|
176	177	\| `/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. \|
177	178	\| `/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. \|
	179	+\| `/unload [model]` \| With `unloadModels` (default on), advertised with the input hint `[model]`: same as in the terminal, the `🧹` line as the agent's text; the history is not touched. \|
178	180
179	181	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).
180	182

  | `/help` | Prints the commands of this front end, one per line with their description, aligned. | Sent to the model as a new message. |
  | `/copy [n\|all]` | Copies the last answer (`n`: the n-th from the end) to the clipboard through `pbcopy`/`wl-copy`/`xclip`/`xsel` or, failing that, OSC 52 when stdout is a terminal, and prints `📋 copied N chars (<road>)` or `⚠️ …`. `/copy all` is window-mode only: here it says so. | Sent to the model as a new message. |
  | `/save <file> [n]` | Writes the last answer (`n`: the n-th from the end) as it is to a new file, creating the missing directories, and prints `💾 saved the last answer to <file> (N bytes)`. Never overwrites: an existing file gives `⚠️ <file> already exists — nothing was written; choose another name`. A file name with spaces is not supported. The model is not called. | Sent to the model as a new message. |
++| `/unload [model]` | With `unloadModels` (default on): unloads a model from the server's memory — this agent's own, or the one named — and prints one line: `🧹 unloaded <model>`, `🧹 <model> was not loaded: nothing to unload`, or `🧹 cannot unload <model>: <reason>` (a provider or a server that has no way to). The history is not touched; the next question reloads the agent's model. | Sent to the model as a new message. |
  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.
  ## ACP mode
--Two commands exist in ACP mode: `/new` and `/compact`. They are advertised to the editor through an `available_commands_update` notification and intercepted by `session/prompt` when the prompt text, trimmed, is exactly the command. Each answers with one `agent_message_chunk` carrying the line the terminal would print, then `stopReason: end_turn`; the model is not consulted.
++Two commands exist in ACP mode, three with `unloadModels` (the default): `/new`, `/compact` and `/unload`. They are advertised to the editor through an `available_commands_update` notification and intercepted by `session/prompt` when the prompt text, trimmed, is exactly the command (`/unload` may be followed by a model name). Each answers with one `agent_message_chunk` carrying the line the terminal would print, then `stopReason: end_turn`; the model is not consulted.
  | Command | Over ACP |
  |---------|----------|
  | `/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. |
++| `/unload [model]` | With `unloadModels` (default on), advertised with the input hint `[model]`: same as in the terminal, the `🧹` line as the agent's text; the history is not touched. |
  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).

modified docs/en/reference/tools.md +20 -0 Viewed

@@ -268,6 +268,26 @@ Declared when `saveAnswerTool: true` (default `false`).
268	268	\| Terminal display \| `💾 save_answer: <path> (this answer\\|the previous answer)`, then the result. \|
269	269	\| ACP and window \| Tool call of kind `edit`, with the file's absolute path. \|
270	270
	271	+## `unload_model`
	272	+
	273	+Declared when `unloadModels` is true (the default), except beside a `structuredOutput`.
	274	+
	275	+\| Field \| Type \| Description \|
	276	+\|-------\|------\|-------------\|
	277	+\| `model` \| string \| The model to unload, as the server names it. Empty: the agent's own model. \|
	278	+
	279	+\| Aspect \| Behaviour \|
	280	+\|--------\|-----------\|
	281	+\| llmman, `ollama` \| Looks the model up in `GET /api/ps` first; when it is loaded, `POST /api/generate {"model": …, "keep_alive": 0}` and expects `done_reason: "unload"`. A model that is not loaded is never asked for: the request would load it first. \|
	282	+\| `llamacpp` \| `POST /models/unload {"model": …}`, which exists in router mode only (llama-server started without a model). A single-model server's 404 gives the reason and the alternatives: router mode, `--sleep-idle-seconds N`, stopping the server. \|
	283	+\| `dmr` \| `cannot unload <model>: the dmr provider cannot unload a model`. \|
	284	+\| Result \| One line: `unloaded <model>`, `<model> was not loaded: nothing to unload`, or `cannot unload <model>: <reason>` (unreachable server, the server's own error, no confirmation). \|
	285	+\| Own model \| Unloading the agent's own model lasts until its next step: the model is needed to read the result, and the server reloads it. \|
	286	+\| Permission \| Never asked, even with `yolo: false`: nothing changes on disk, and the model reloads by itself. \|
	287	+\| Terminal display \| `🧹 unload_model: <model>` (`(own model)` when empty), then the result. \|
	288	+\| ACP and window \| Tool call of kind `other`; a `cannot unload` result is a failed call. \|
	289	+\| Timeout \| 30 s per request to the server. \|
	290	+
271	291	## Declared tools
272	292
273	293	Declared from the `tools:` block of the agent file, one tool per entry (see [configuration](configuration.md#tools-keys)). Each has the entry's `name`; its fields are the entry's `parameters`.

  | Terminal display | `💾 save_answer: <path> (this answer\|the previous answer)`, then the result. |
  | ACP and window | Tool call of kind `edit`, with the file's absolute path. |
++## `unload_model`
++
++Declared when `unloadModels` is true (the default), except beside a `structuredOutput`.
++
++| Field | Type | Description |
++|-------|------|-------------|
++| `model` | string | The model to unload, as the server names it. Empty: the agent's own model. |
++
++| Aspect | Behaviour |
++|--------|-----------|
++| llmman, `ollama` | Looks the model up in `GET /api/ps` first; when it is loaded, `POST /api/generate {"model": …, "keep_alive": 0}` and expects `done_reason: "unload"`. A model that is not loaded is never asked for: the request would load it first. |
++| `llamacpp` | `POST /models/unload {"model": …}`, which exists in router mode only (llama-server started without a model). A single-model server's 404 gives the reason and the alternatives: router mode, `--sleep-idle-seconds N`, stopping the server. |
++| `dmr` | `cannot unload <model>: the dmr provider cannot unload a model`. |
++| Result | One line: `unloaded <model>`, `<model> was not loaded: nothing to unload`, or `cannot unload <model>: <reason>` (unreachable server, the server's own error, no confirmation). |
++| Own model | Unloading the agent's own model lasts until its next step: the model is needed to read the result, and the server reloads it. |
++| Permission | Never asked, even with `yolo: false`: nothing changes on disk, and the model reloads by itself. |
++| Terminal display | `🧹 unload_model: <model>` (`(own model)` when empty), then the result. |
++| ACP and window | Tool call of kind `other`; a `cannot unload` result is a failed call. |
++| Timeout | 30 s per request to the server. |
++
  ## Declared tools
  Declared from the `tools:` block of the agent file, one tool per entry (see [configuration](configuration.md#tools-keys)). Each has the entry's `name`; its fields are the entry's `parameters`.

modified docs/fr/README.md +2 -1 Viewed

@@ -27,6 +27,7 @@ Cette documentation est organisée selon les quatre quadrants [Diátaxis](https:
27	27	- [Ajouter un skill](how-to/add-a-skill.md)
28	28	- [Déclarer un outil personnalisé](how-to/declare-a-custom-tool.md)
29	29	- [Sauvegarder une réponse dans un fichier](how-to/save-an-answer.md)
	30	+- [Libérer la mémoire d'un modèle](how-to/unload-a-model.md) — `/unload [modèle]` dans toutes les façades, l'outil `unload_model` après la réponse d'un autre agent, `unloadModels`
30	31	- [Mesurer si le modèle choisit le bon skill](how-to/measure-skill-routing.md) — l'eval de routage : le lancer, lire son tableau, s'en servir en avant/après
31	32	- [Garder une longue session dans la fenêtre de contexte](how-to/manage-the-context-window.md)
32	33	- [Démarrer une nouvelle session sans relancer](how-to/start-a-new-session.md)
@@ -51,7 +52,7 @@ Cette documentation est organisée selon les quatre quadrants [Diátaxis](https:
51	52	- [Ligne de commande](reference/cli.md) — options, variables d'environnement, bannière de démarrage
52	53	- [Commandes slash](reference/slash-commands.md) — `/quit`, `/abort`, `/compact`, `/new`, `/sessions`, `/resume`, `/help`, raccourcis clavier, touches de la fenêtre
53	54	- [Fichier de configuration](reference/configuration.md) — chaque clé YAML, sa valeur par défaut et sa règle de validation
54		-- [Outils intégrés](reference/tools.md) — `bash`, `read_skill`, `read_file`, `write_file`, `edit_file`, `lsp`, `fetch_content`, `ask_agent`, `ask_user`, `save_answer`, outils déclarés
	55	+- [Outils intégrés](reference/tools.md) — `bash`, `read_skill`, `read_file`, `write_file`, `edit_file`, `lsp`, `fetch_content`, `ask_agent`, `ask_user`, `save_answer`, `unload_model`, outils déclarés
55	56	- [Format d'un fichier de skill](reference/skill-format.md)
56	57	- [Les hooks](reference/hooks.md) — les six événements, le JSON qu'un hook lit et écrit, `matcher`, `timeout`, `onError`, erreurs de validation
57	58	- [Surface Agent Client Protocol](reference/acp.md)

  - [Ajouter un skill](how-to/add-a-skill.md)
  - [Déclarer un outil personnalisé](how-to/declare-a-custom-tool.md)
  - [Sauvegarder une réponse dans un fichier](how-to/save-an-answer.md)
++- [Libérer la mémoire d'un modèle](how-to/unload-a-model.md) — `/unload [modèle]` dans toutes les façades, l'outil `unload_model` après la réponse d'un autre agent, `unloadModels`
  - [Mesurer si le modèle choisit le bon skill](how-to/measure-skill-routing.md) — l'eval de routage : le lancer, lire son tableau, s'en servir en avant/après
  - [Garder une longue session dans la fenêtre de contexte](how-to/manage-the-context-window.md)
  - [Démarrer une nouvelle session sans relancer](how-to/start-a-new-session.md)
  - [Ligne de commande](reference/cli.md) — options, variables d'environnement, bannière de démarrage
  - [Commandes slash](reference/slash-commands.md) — `/quit`, `/abort`, `/compact`, `/new`, `/sessions`, `/resume`, `/help`, raccourcis clavier, touches de la fenêtre
  - [Fichier de configuration](reference/configuration.md) — chaque clé YAML, sa valeur par défaut et sa règle de validation
--- [Outils intégrés](reference/tools.md) — `bash`, `read_skill`, `read_file`, `write_file`, `edit_file`, `lsp`, `fetch_content`, `ask_agent`, `ask_user`, `save_answer`, outils déclarés
++- [Outils intégrés](reference/tools.md) — `bash`, `read_skill`, `read_file`, `write_file`, `edit_file`, `lsp`, `fetch_content`, `ask_agent`, `ask_user`, `save_answer`, `unload_model`, outils déclarés
  - [Format d'un fichier de skill](reference/skill-format.md)
  - [Les hooks](reference/hooks.md) — les six événements, le JSON qu'un hook lit et écrit, `matcher`, `timeout`, `onError`, erreurs de validation
  - [Surface Agent Client Protocol](reference/acp.md)

added docs/fr/how-to/unload-a-model.md +58 -0 Viewed

new file mode 100644
@@ -0,0 +1,58 @@
	1	+# Libérer la mémoire d'un modèle
	2	+
	3	+Ce guide montre comment décharger un modèle de la mémoire du serveur sans arrêter le serveur : à la main avec `/unload`, ou par l'agent lui-même avec son outil `unload_model` — typiquement pour libérer la mémoire du modèle d'un autre agent une fois sa réponse obtenue. Il suppose que mm tourne sur llmman, Ollama, ou llama-server en mode router. Le modèle se recharge tout seul la prochaine fois qu'il sert.
	4	+
	5	+## Décharger à la main
	6	+
	7	+Tapez, dans le terminal, la fenêtre ou un éditeur :
	8	+
	9	+```text
	10	+/unload
	11	+```
	12	+
	13	+Le modèle de l'agent est déchargé :
	14	+
	15	+```text
	16	+🧹 unloaded hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M
	17	+```
	18	+
	19	+Nommez un autre modèle servi par le même serveur pour décharger celui-là :
	20	+
	21	+```text
	22	+/unload hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M
	23	+```
	24	+
	25	+Un modèle qui n'est pas chargé donne `🧹 <modèle> was not loaded: nothing to unload` ; rien n'est envoyé au serveur, donc rien n'est chargé non plus. Vérifiez ce qui est chargé avec `llmman ps`, `ollama ps`, ou `curl <serveur>/api/ps`.
	26	+
	27	+## Laisser l'agent décharger un modèle
	28	+
	29	+L'outil `unload_model` est déclaré par défaut. Dites au modèle quand s'en servir, dans son prompt `system`. Le cas où il sert le mieux est le modèle d'un autre agent, une fois que cet agent a répondu :
	30	+
	31	+```yaml
	32	+mmConfigs:
	33	+ expert:
	34	+ path: agent.expert.yaml
	35	+ description: Programming expert.
	36	+
	37	+system: \|
	38	+ For every question about programming, do these three steps, in order:
	39	+ 1. Call ask_agent with agent "expert" and the user's question.
	40	+ 2. Call unload_model with model "hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M".
	41	+ It is ALWAYS your next action after ask_agent: never answer the user
	42	+ before you have called unload_model.
	43	+ 3. Then give the expert's answer to the user.
	44	+```
	45	+
	46	+Écrivez le nom du modèle exactement comme le serveur le liste : le modèle ne peut pas le deviner. `demos/unload-models/` décharge plutôt l'expert avec un hook `postAsk` (voir les variantes ci-dessous) ; son coordinateur appelle toujours `unload_model` quand l'utilisateur lui demande de libérer la mémoire de l'expert.
	47	+
	48	+## Variantes
	49	+
	50	+- Le modèle de l'agent lui-même. `unload_model` avec un `model` vide le décharge, mais seulement jusqu'à l'étape suivante de l'agent : il a besoin de son modèle pour lire le résultat de l'outil, et le serveur le recharge. Utilisez plutôt `/unload` entre deux questions — ou, pour le libérer pendant qu'un autre agent travaille, un hook `preAsk` qui décharge le modèle nommé dans son entrée (`"model": "<fournisseur>/<modèle>"`) juste avant le lancement de l'agent interrogé : `demos/unload-models/hooks/unload.js self`.
	51	+- Une garantie plutôt qu'une consigne. Un petit modèle peut sauter l'étape (mesuré : 4 fois sur 5 avec MiniCPM5 2B dans la première version de `demos/unload-models`), et le modèle de l'agent est déjà rechargé à côté de l'autre quand il appelle l'outil. Pour un déchargement certain, sans jamais les deux modèles chargés ensemble, faites appeler le serveur par un hook `postAsk` — `curl -s <serveur>/api/generate -d '{"model":"…","keep_alive":0}'` — une fois que l'agent interrogé a répondu et avant que l'agent qui l'a interrogé lise la réponse : `demos/unload-models/hooks/unload.js asked` lit le modèle de l'agent interrogé dans `ask.agentModel`. Voir [Contrôler les relais avec des hooks](control-agent-calls-with-hooks.md).
	52	+- llama-server. Seul le mode router (lancé sans `-m`) peut décharger un modèle. Un serveur à modèle unique répond avec les alternatives : `--sleep-idle-seconds N` libère la mémoire après N secondes d'inactivité, ou arrêtez le serveur.
	53	+- Le désactiver. `unloadModels: false` retire à la fois `/unload` et l'outil — pour un agent qui ne doit pas avoir un outil de plus, comme une garde ou un spécialiste étroit.
	54	+
	55	+## Voir aussi
	56	+
	57	+- [Référence de `unload_model`](../reference/tools.md#unload_model) et de [`/unload`](../reference/slash-commands.md#commandes)
	58	+- [`unloadModels`](../reference/configuration.md)

++# Libérer la mémoire d'un modèle
++
++Ce guide montre comment décharger un modèle de la mémoire du serveur sans arrêter le serveur : à la main avec `/unload`, ou par l'agent lui-même avec son outil `unload_model` — typiquement pour libérer la mémoire du modèle d'un autre agent une fois sa réponse obtenue. Il suppose que mm tourne sur llmman, Ollama, ou llama-server en mode router. Le modèle se recharge tout seul la prochaine fois qu'il sert.
++
++## Décharger à la main
++
++Tapez, dans le terminal, la fenêtre ou un éditeur :
++
++```text
++/unload
++```
++
++Le modèle de l'agent est déchargé :
++
++```text
++🧹 unloaded hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M
++```
++
++Nommez un autre modèle servi par le même serveur pour décharger celui-là :
++
++```text
++/unload hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M
++```
++
++Un modèle qui n'est pas chargé donne `🧹 <modèle> was not loaded: nothing to unload` ; rien n'est envoyé au serveur, donc rien n'est chargé non plus. Vérifiez ce qui est chargé avec `llmman ps`, `ollama ps`, ou `curl <serveur>/api/ps`.
++
++## Laisser l'agent décharger un modèle
++
++L'outil `unload_model` est déclaré par défaut. Dites au modèle quand s'en servir, dans son prompt `system`. Le cas où il sert le mieux est le modèle d'un autre agent, une fois que cet agent a répondu :
++
++```yaml
++mmConfigs:
++  expert:
++    path: agent.expert.yaml
++    description: Programming expert.
++
++system: |
++  For every question about programming, do these three steps, in order:
++  1. Call ask_agent with agent "expert" and the user's question.
++  2. Call unload_model with model "hf.co/unsloth/Qwen3.5-9B-GGUF:Q4_K_M".
++     It is ALWAYS your next action after ask_agent: never answer the user
++     before you have called unload_model.
++  3. Then give the expert's answer to the user.
++```
++
++Écrivez le nom du modèle exactement comme le serveur le liste : le modèle ne peut pas le deviner. `demos/unload-models/` décharge plutôt l'expert avec un hook `postAsk` (voir les variantes ci-dessous) ; son coordinateur appelle toujours `unload_model` quand l'utilisateur lui demande de libérer la mémoire de l'expert.
++
++## Variantes
++
++- **Le modèle de l'agent lui-même.** `unload_model` avec un `model` vide le décharge, mais seulement jusqu'à l'étape suivante de l'agent : il a besoin de son modèle pour lire le résultat de l'outil, et le serveur le recharge. Utilisez plutôt `/unload` entre deux questions — ou, pour le libérer pendant qu'un autre agent travaille, un hook `preAsk` qui décharge le modèle nommé dans son entrée (`"model": "<fournisseur>/<modèle>"`) juste avant le lancement de l'agent interrogé : `demos/unload-models/hooks/unload.js self`.
++- **Une garantie plutôt qu'une consigne.** Un petit modèle peut sauter l'étape (mesuré : 4 fois sur 5 avec MiniCPM5 2B dans la première version de `demos/unload-models`), et le modèle de l'agent est déjà rechargé à côté de l'autre quand il appelle l'outil. Pour un déchargement certain, sans jamais les deux modèles chargés ensemble, faites appeler le serveur par un hook `postAsk` — `curl -s <serveur>/api/generate -d '{"model":"…","keep_alive":0}'` — une fois que l'agent interrogé a répondu et avant que l'agent qui l'a interrogé lise la réponse : `demos/unload-models/hooks/unload.js asked` lit le modèle de l'agent interrogé dans `ask.agentModel`. Voir [Contrôler les relais avec des hooks](control-agent-calls-with-hooks.md).
++- **llama-server.** Seul le mode router (lancé sans `-m`) peut décharger un modèle. Un serveur à modèle unique répond avec les alternatives : `--sleep-idle-seconds N` libère la mémoire après N secondes d'inactivité, ou arrêtez le serveur.
++- **Le désactiver.** `unloadModels: false` retire à la fois `/unload` et l'outil — pour un agent qui ne doit pas avoir un outil de plus, comme une garde ou un spécialiste étroit.
++
++## Voir aussi
++
++- [Référence de `unload_model`](../reference/tools.md#unload_model) et de [`/unload`](../reference/slash-commands.md#commandes)
++- [`unloadModels`](../reference/configuration.md)

modified docs/fr/reference/configuration.md +2 -1 Viewed

@@ -26,6 +26,7 @@
26	26	\| `fetchContentTool` \| booléen \| `false` \| Déclare `fetch_content`, qui récupère des pages web en Markdown. Désactivé par défaut : c'est le seul outil intégré qui accède au réseau. \|
27	27	\| `askUserTool` \| booléen \| `true` \| Déclare `ask_user`, par lequel le modèle pose des questions à l'utilisateur (choisir une option, en choisir plusieurs, taper un texte, oui ou non) et attend les réponses. `false` le retire : le modèle ne peut pas demander. Jamais déclaré à côté d'un `structuredOutput`. Surchargé par `AGENT_ASK_USER` ; pour un agent interrogé par un autre, aussi par le `mmConfigs.<nom>.askUser` de celui qui l'interroge. Voir [outils](tools.md#ask_user). \|
28	28	\| `saveAnswerTool` \| bool \| `false` \| Déclare `save_answer`, qui écrit dans un nouveau fichier une réponse que le modèle a déjà donnée : le modèle envoie un nom de fichier, pas le texte une seconde fois. Désactivé par défaut. La commande `/save` existe dans tous les cas. Voir les [outils](tools.md#save_answer). \|
	29	+\| `unloadModels` \| bool \| `true` \| La commande `/unload [modèle]` (terminal, fenêtre, éditeur) et l'outil `unload_model`, qui libèrent un modèle de la mémoire du serveur : llmman et Ollama, llama-server en mode router. `false` retire les deux. L'outil n'est jamais déclaré à côté d'une `structuredOutput`. Voir les [outils](tools.md#unload_model) et [Libérer la mémoire d'un modèle](../how-to/unload-a-model.md). \|
29	30	\| `sampling` \| dictionnaire \| `temperature: 0.0`, `top_p: 0.9`, `max_tokens: 4096` \| Paramètres de génération, noms de l'API OpenAI, transmis au serveur. Une clé que l'API OpenAI ne déclare pas est envoyée telle quelle, au premier niveau de la requête — par exemple `chat_template_kwargs: {enable_thinking: false}`, qui demande à un serveur llama.cpp (`dmr`, `llamacpp`, `llmman`) de ne pas raisonner ; `reasoning_effort: none` fait de même avec Ollama. `messages`, `tools`, `tool_choice` (et les anciens `functions`, `function_call`) sont ignorées : Genkit les construit à partir de la conversation. L'`agent.yaml` livré ajoute `parallel_tool_calls: false`. `max_tokens` borne aussi un appel d'outil : un `write_file` de long document (ou le raisonnement d'un modèle « thinking » plus l'appel) peut l'atteindre, et l'appel est alors coupé. mm le rend deux fois au modèle en lui demandant de procéder par morceaux, puis s'arrête avec `the model hit max_tokens (N) while writing a tool call` ; augmentez la valeur si cela arrive. En streaming, mm demande aussi au serveur le compte des tokens (`stream_options.include_usage`). \|
30	31	\| `watchdogTimeout` \| durée \| `20s` \| Silence toléré entre deux tokens diffusés avant de déclarer la connexion bloquée. Doublé pour les requêtes de résumé. \|
31	32	\| `firstTokenTimeout` \| durée \| `3m` \| Silence toléré avant le premier token d'un appel au modèle, pendant que le serveur lit le prompt (chaque aller-retour d'un tour en a un). Jamais plus court que `watchdogTimeout` ; doit être >= 0. Avec un serveur llama.cpp (`dmr`, `llamacpp`, `llmman` — dont les modèles GGUF locaux tournent dans llama-server), mm demande `return_progress` : chaque rapport d'avancement relance ce délai, et le spinner affiche `Reading the prompt 12.0k/25.3k tokens`. Ollama n'envoie pas d'avancement : cette seule attente couvre tout le prompt (et le chargement du modèle). À l'expiration, le watchdog dit `no first token from the server after …` (avant le premier token) ou `connection hang detected after …` (après). \|
@@ -128,7 +129,7 @@ Le fichier livré `agent.lsp.yaml` est `agent.dmr.yaml` avec ce bloc actif (gopl
128	129	\| Clé \| Type \| Défaut \| Description \|
129	130	\|-----\|------\|--------\|-------------\|
130	131	\| `tools` \| liste \| `[]` \| Une entrée par outil déclaré. \|
131		-\| `tools[].name` \| chaîne \| — \| Le nom que le modèle appelle. Obligatoire. Commence par une lettre ; lettres, chiffres, `_`, `-` ; 64 caractères au plus. Unique. Pas le nom d'un outil intégré (`ask_agent`, `ask_user`, `bash`, `edit_file`, `fetch_content`, `lsp`, `plan`, `read_file`, `read_skill`, `save_answer`, `write_file`), même désactivé dans cet agent. \|
	132	+\| `tools[].name` \| chaîne \| — \| Le nom que le modèle appelle. Obligatoire. Commence par une lettre ; lettres, chiffres, `_`, `-` ; 64 caractères au plus. Unique. Pas le nom d'un outil intégré (`ask_agent`, `ask_user`, `bash`, `edit_file`, `fetch_content`, `lsp`, `plan`, `read_file`, `read_skill`, `save_answer`, `unload_model`, `write_file`), même désactivé dans cet agent. \|
132	133	\| `tools[].description` \| chaîne \| — \| Ce que fait l'outil et quand s'en servir, montré au modèle. Obligatoire. Une ligne `Required parameters: …` y est ajoutée. \|
133	134	\| `tools[].parameters` \| map nom → paramètre \| `{}` \| Les arguments, dans l'ordre écrit. Les noms suivent les règles de `name`. \|
134	135	\| `tools[].parameters.<p>.type` \| chaîne \| `string` \| `string`, `number`, `integer` ou `boolean`. \|

  | `fetchContentTool` | booléen | `false` | Déclare `fetch_content`, qui récupère des pages web en Markdown. Désactivé par défaut : c'est le seul outil intégré qui accède au réseau. |
  | `askUserTool` | booléen | `true` | Déclare `ask_user`, par lequel le modèle pose des questions à l'utilisateur (choisir une option, en choisir plusieurs, taper un texte, oui ou non) et attend les réponses. `false` le retire : le modèle ne peut pas demander. Jamais déclaré à côté d'un `structuredOutput`. Surchargé par `AGENT_ASK_USER` ; pour un agent interrogé par un autre, aussi par le `mmConfigs.<nom>.askUser` de celui qui l'interroge. Voir [outils](tools.md#ask_user). |
  | `saveAnswerTool` | bool | `false` | Déclare `save_answer`, qui écrit dans un nouveau fichier une réponse que le modèle a déjà donnée : le modèle envoie un nom de fichier, pas le texte une seconde fois. Désactivé par défaut. La commande `/save` existe dans tous les cas. Voir les [outils](tools.md#save_answer). |
++| `unloadModels` | bool | `true` | La commande `/unload [modèle]` (terminal, fenêtre, éditeur) et l'outil `unload_model`, qui libèrent un modèle de la mémoire du serveur : llmman et Ollama, llama-server en mode router. `false` retire les deux. L'outil n'est jamais déclaré à côté d'une `structuredOutput`. Voir les [outils](tools.md#unload_model) et [Libérer la mémoire d'un modèle](../how-to/unload-a-model.md). |
  | `sampling` | dictionnaire | `temperature: 0.0`, `top_p: 0.9`, `max_tokens: 4096` | Paramètres de génération, noms de l'API OpenAI, transmis au serveur. Une clé que l'API OpenAI ne déclare pas est envoyée telle quelle, au premier niveau de la requête — par exemple `chat_template_kwargs: {enable_thinking: false}`, qui demande à un serveur llama.cpp (`dmr`, `llamacpp`, `llmman`) de ne pas raisonner ; `reasoning_effort: none` fait de même avec Ollama. `messages`, `tools`, `tool_choice` (et les anciens `functions`, `function_call`) sont ignorées : Genkit les construit à partir de la conversation. L'`agent.yaml` livré ajoute `parallel_tool_calls: false`. `max_tokens` borne aussi un appel d'outil : un `write_file` de long document (ou le raisonnement d'un modèle « thinking » plus l'appel) peut l'atteindre, et l'appel est alors coupé. mm le rend deux fois au modèle en lui demandant de procéder par morceaux, puis s'arrête avec `the model hit max_tokens (N) while writing a tool call` ; augmentez la valeur si cela arrive. En streaming, mm demande aussi au serveur le compte des tokens (`stream_options.include_usage`). |
  | `watchdogTimeout` | durée | `20s` | Silence toléré entre deux tokens diffusés avant de déclarer la connexion bloquée. Doublé pour les requêtes de résumé. |
  | `firstTokenTimeout` | durée | `3m` | Silence toléré avant le premier token d'un appel au modèle, pendant que le serveur lit le prompt (chaque aller-retour d'un tour en a un). Jamais plus court que `watchdogTimeout` ; doit être >= 0. Avec un serveur llama.cpp (`dmr`, `llamacpp`, `llmman` — dont les modèles GGUF locaux tournent dans llama-server), mm demande `return_progress` : chaque rapport d'avancement relance ce délai, et le spinner affiche `Reading the prompt 12.0k/25.3k tokens`. Ollama n'envoie pas d'avancement : cette seule attente couvre tout le prompt (et le chargement du modèle). À l'expiration, le watchdog dit `no first token from the server after …` (avant le premier token) ou `connection hang detected after …` (après). |
  | Clé | Type | Défaut | Description |
  |-----|------|--------|-------------|
  | `tools` | liste | `[]` | Une entrée par outil déclaré. |
--| `tools[].name` | chaîne | — | Le nom que le modèle appelle. Obligatoire. Commence par une lettre ; lettres, chiffres, `_`, `-` ; 64 caractères au plus. Unique. Pas le nom d'un outil intégré (`ask_agent`, `ask_user`, `bash`, `edit_file`, `fetch_content`, `lsp`, `plan`, `read_file`, `read_skill`, `save_answer`, `write_file`), même désactivé dans cet agent. |
++| `tools[].name` | chaîne | — | Le nom que le modèle appelle. Obligatoire. Commence par une lettre ; lettres, chiffres, `_`, `-` ; 64 caractères au plus. Unique. Pas le nom d'un outil intégré (`ask_agent`, `ask_user`, `bash`, `edit_file`, `fetch_content`, `lsp`, `plan`, `read_file`, `read_skill`, `save_answer`, `unload_model`, `write_file`), même désactivé dans cet agent. |
  | `tools[].description` | chaîne | — | Ce que fait l'outil et quand s'en servir, montré au modèle. Obligatoire. Une ligne `Required parameters: …` y est ajoutée. |
  | `tools[].parameters` | map nom → paramètre | `{}` | Les arguments, dans l'ordre écrit. Les noms suivent les règles de `name`. |
  | `tools[].parameters.<p>.type` | chaîne | `string` | `string`, `number`, `integer` ou `boolean`. |

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

@@ -14,6 +14,7 @@
14	14	\| `/resume [id]` \| Reprend une session enregistrée : la plus récemment mise à jour autre que la courante, en ignorant les vides, ou l'identifiant donné. Affiche `↩️ resumed <id> — N message(s), started with "<première question>"` (suivi de ` · ⚠️ N error(s)` quand la session a enregistré des erreurs) ; réinitialise le détecteur de boucles et le compte de tokens du serveur ; les tours suivants sont enregistrés dans cette session. Refusée avec une raison `⚠️` quand les sessions sont désactivées, qu'il n'y a pas de session antérieure, que l'identifiant est inconnu ou désigne la session courante. \| Envoyée au modèle comme nouveau message. \|
15	15	\| `/help` \| Affiche les commandes de cette façade, une par ligne avec leur description, alignées. \| Envoyée au modèle comme nouveau message. \|
16	16	\| `/save <fichier> [n]` \| Écrit la dernière réponse (`n` : la n-ième depuis la fin) telle quelle dans un nouveau fichier, en créant les répertoires manquants, et affiche `💾 saved the last answer to <fichier> (N bytes)`. N'écrase jamais : un fichier existant donne `⚠️ <fichier> already exists — nothing was written; choose another name`. Un nom de fichier avec des espaces n'est pas pris en charge. Le modèle n'est pas appelé. \| Envoyée au modèle comme nouveau message. \|
	17	+\| `/unload [modèle]` \| Avec `unloadModels` (activé par défaut) : décharge un modèle de la mémoire du serveur — celui de l'agent, ou celui nommé — et affiche une ligne : `🧹 unloaded <modèle>`, `🧹 <modèle> was not loaded: nothing to unload`, ou `🧹 cannot unload <modèle>: <raison>` (un fournisseur ou un serveur qui n'a aucun moyen de le faire). L'historique n'est pas touché ; la question suivante recharge le modèle de l'agent. \| Envoyée au modèle comme nouveau message. \|
17	18	\| `/copy [n\\|all]` \| Copie la dernière réponse (`n` : la n-ième depuis la fin) dans le presse-papiers via `pbcopy`/`wl-copy`/`xclip`/`xsel` ou, à défaut, OSC 52 quand stdout est un terminal, et affiche `📋 copied N chars (<chemin>)` ou `⚠️ …`. `/copy all` est réservé au mode fenêtre : ici la commande le dit. \| Envoyée au modèle comme nouveau message. \|
18	19
19	20	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.
@@ -169,12 +170,13 @@ La notation marche aussi en mode ligne et en ACP, tapée à la main : un `$skill
169	170
170	171	## Mode ACP
171	172
172		-Deux commandes existent en mode ACP : `/new` et `/compact`. Elles sont annoncées à l'éditeur par une notification `available_commands_update` et interceptées par `session/prompt` quand le texte du prompt, débarrassé de ses espaces, est exactement la commande. Chacune répond par un `agent_message_chunk` portant la ligne que le terminal afficherait, puis `stopReason: end_turn` ; le modèle n'est pas consulté.
	173	+Deux commandes existent en mode ACP, trois avec `unloadModels` (le défaut) : `/new`, `/compact` et `/unload`. Elles sont annoncées à l'éditeur par une notification `available_commands_update` et interceptées par `session/prompt` quand le texte du prompt, débarrassé de ses espaces, est exactement la commande (`/unload` peut être suivie d'un nom de modèle). Chacune répond par un `agent_message_chunk` portant la ligne que le terminal afficherait, puis `stopReason: end_turn` ; le modèle n'est pas consulté.
173	174
174	175	\| Commande \| En ACP \|
175	176	\|----------\|--------\|
176	177	\| `/new` \| Comme dans le terminal : historique ramené au prompt système, compte de tokens oublié ; l'identifiant de session, le `cwd` et les autorisations « toujours autoriser » sont conservés. \|
177	178	\| `/compact` \| Comme dans le terminal, toujours forcée : `🗜️ compressed …`, `🗜️ nothing to compact: …` ou `[compact: failed, history kept: …]`. Attend d'abord la fin d'un tour en cours. \|
	179	+\| `/unload [modèle]` \| Avec `unloadModels` (activé par défaut), annoncée avec l'indication de saisie `[model]` : comme dans le terminal, la ligne `🧹` comme texte de l'agent ; l'historique n'est pas touché. \|
178	180
179	181	Chaque skill y est proposé comme une commande à part entière : ACP n'a pas de sélecteur pour les skills, donc `/greetings Bob` dans le menu de l'éditeur est l'équivalent de `$greetings Bob` dans la fenêtre, et il est réécrit en exactement ça — voir la [référence ACP](acp.md#les-skills-comme-commandes).
180	182

  | `/resume [id]` | Reprend une session enregistrée : la plus récemment mise à jour autre que la courante, en ignorant les vides, ou l'identifiant donné. Affiche `↩️ resumed <id> — N message(s), started with "<première question>"` (suivi de ` · ⚠️ N error(s)` quand la session a enregistré des erreurs) ; réinitialise le détecteur de boucles et le compte de tokens du serveur ; les tours suivants sont enregistrés dans cette session. Refusée avec une raison `⚠️` quand les sessions sont désactivées, qu'il n'y a pas de session antérieure, que l'identifiant est inconnu ou désigne la session courante. | Envoyée au modèle comme nouveau message. |
  | `/help` | Affiche les commandes de cette façade, une par ligne avec leur description, alignées. | Envoyée au modèle comme nouveau message. |
  | `/save <fichier> [n]` | Écrit la dernière réponse (`n` : la n-ième depuis la fin) telle quelle dans un nouveau fichier, en créant les répertoires manquants, et affiche `💾 saved the last answer to <fichier> (N bytes)`. N'écrase jamais : un fichier existant donne `⚠️ <fichier> already exists — nothing was written; choose another name`. Un nom de fichier avec des espaces n'est pas pris en charge. Le modèle n'est pas appelé. | Envoyée au modèle comme nouveau message. |
++| `/unload [modèle]` | Avec `unloadModels` (activé par défaut) : décharge un modèle de la mémoire du serveur — celui de l'agent, ou celui nommé — et affiche une ligne : `🧹 unloaded <modèle>`, `🧹 <modèle> was not loaded: nothing to unload`, ou `🧹 cannot unload <modèle>: <raison>` (un fournisseur ou un serveur qui n'a aucun moyen de le faire). L'historique n'est pas touché ; la question suivante recharge le modèle de l'agent. | Envoyée au modèle comme nouveau message. |
  | `/copy [n\|all]` | Copie la dernière réponse (`n` : la n-ième depuis la fin) dans le presse-papiers via `pbcopy`/`wl-copy`/`xclip`/`xsel` ou, à défaut, OSC 52 quand stdout est un terminal, et affiche `📋 copied N chars (<chemin>)` ou `⚠️ …`. `/copy all` est réservé au mode fenêtre : ici la commande le dit. | Envoyée au modèle comme nouveau message. |
  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.
  ## Mode ACP
--Deux commandes existent en mode ACP : `/new` et `/compact`. Elles sont annoncées à l'éditeur par une notification `available_commands_update` et interceptées par `session/prompt` quand le texte du prompt, débarrassé de ses espaces, est exactement la commande. Chacune répond par un `agent_message_chunk` portant la ligne que le terminal afficherait, puis `stopReason: end_turn` ; le modèle n'est pas consulté.
++Deux commandes existent en mode ACP, trois avec `unloadModels` (le défaut) : `/new`, `/compact` et `/unload`. Elles sont annoncées à l'éditeur par une notification `available_commands_update` et interceptées par `session/prompt` quand le texte du prompt, débarrassé de ses espaces, est exactement la commande (`/unload` peut être suivie d'un nom de modèle). Chacune répond par un `agent_message_chunk` portant la ligne que le terminal afficherait, puis `stopReason: end_turn` ; le modèle n'est pas consulté.
  | Commande | En ACP |
  |----------|--------|
  | `/new` | Comme dans le terminal : historique ramené au prompt système, compte de tokens oublié ; l'identifiant de session, le `cwd` et les autorisations « toujours autoriser » sont conservés. |
  | `/compact` | Comme dans le terminal, toujours forcée : `🗜️ compressed …`, `🗜️ nothing to compact: …` ou `[compact: failed, history kept: …]`. Attend d'abord la fin d'un tour en cours. |
++| `/unload [modèle]` | Avec `unloadModels` (activé par défaut), annoncée avec l'indication de saisie `[model]` : comme dans le terminal, la ligne `🧹` comme texte de l'agent ; l'historique n'est pas touché. |
  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).

modified docs/fr/reference/tools.md +20 -0 Viewed

@@ -268,6 +268,26 @@ Déclaré quand `saveAnswerTool: true` (défaut `false`).
268	268	\| Affichage terminal \| `💾 save_answer: <path> (this answer\\|the previous answer)`, puis le résultat. \|
269	269	\| ACP et fenêtre \| Appel d'outil de type `edit`, avec le chemin absolu du fichier. \|
270	270
	271	+## `unload_model`
	272	+
	273	+Déclaré quand `unloadModels` vaut true (le défaut), sauf à côté d'une `structuredOutput`.
	274	+
	275	+\| Champ \| Type \| Description \|
	276	+\|-------\|------\|-------------\|
	277	+\| `model` \| chaîne \| Le modèle à décharger, tel que le serveur le nomme. Vide : le modèle de l'agent. \|
	278	+
	279	+\| Aspect \| Comportement \|
	280	+\|--------\|--------------\|
	281	+\| llmman, `ollama` \| Cherche d'abord le modèle dans `GET /api/ps` ; quand il est chargé, `POST /api/generate {"model": …, "keep_alive": 0}` et attend `done_reason: "unload"`. Un modèle non chargé n'est jamais demandé : la requête le chargerait d'abord. \|
	282	+\| `llamacpp` \| `POST /models/unload {"model": …}`, qui n'existe qu'en mode router (llama-server lancé sans modèle). Le 404 d'un serveur à modèle unique donne la raison et les alternatives : mode router, `--sleep-idle-seconds N`, arrêt du serveur. \|
	283	+\| `dmr` \| `cannot unload <modèle>: the dmr provider cannot unload a model`. \|
	284	+\| Résultat \| Une ligne : `unloaded <modèle>`, `<modèle> was not loaded: nothing to unload`, ou `cannot unload <modèle>: <raison>` (serveur injoignable, erreur du serveur lui-même, absence de confirmation). \|
	285	+\| Son propre modèle \| Décharger le modèle de l'agent ne dure que jusqu'à son étape suivante : le modèle est nécessaire pour lire le résultat, et le serveur le recharge. \|
	286	+\| Permission \| Jamais demandée, même avec `yolo: false` : rien ne change sur disque, et le modèle se recharge tout seul. \|
	287	+\| Affichage terminal \| `🧹 unload_model: <modèle>` (`(own model)` quand il est vide), puis le résultat. \|
	288	+\| ACP et fenêtre \| Appel d'outil de type `other` ; un résultat `cannot unload` est un appel en échec. \|
	289	+\| Délai \| 30 s par requête au serveur. \|
	290	+
271	291	## Outils déclarés
272	292
273	293	Déclarés à partir du bloc `tools:` du fichier d'agent, un outil par entrée (voir la [configuration](configuration.md#clés-de-tools)). Chacun porte le `name` de l'entrée ; ses champs sont les `parameters` de l'entrée.

  | Affichage terminal | `💾 save_answer: <path> (this answer\|the previous answer)`, puis le résultat. |
  | ACP et fenêtre | Appel d'outil de type `edit`, avec le chemin absolu du fichier. |
++## `unload_model`
++
++Déclaré quand `unloadModels` vaut true (le défaut), sauf à côté d'une `structuredOutput`.
++
++| Champ | Type | Description |
++|-------|------|-------------|
++| `model` | chaîne | Le modèle à décharger, tel que le serveur le nomme. Vide : le modèle de l'agent. |
++
++| Aspect | Comportement |
++|--------|--------------|
++| llmman, `ollama` | Cherche d'abord le modèle dans `GET /api/ps` ; quand il est chargé, `POST /api/generate {"model": …, "keep_alive": 0}` et attend `done_reason: "unload"`. Un modèle non chargé n'est jamais demandé : la requête le chargerait d'abord. |
++| `llamacpp` | `POST /models/unload {"model": …}`, qui n'existe qu'en mode router (llama-server lancé sans modèle). Le 404 d'un serveur à modèle unique donne la raison et les alternatives : mode router, `--sleep-idle-seconds N`, arrêt du serveur. |
++| `dmr` | `cannot unload <modèle>: the dmr provider cannot unload a model`. |
++| Résultat | Une ligne : `unloaded <modèle>`, `<modèle> was not loaded: nothing to unload`, ou `cannot unload <modèle>: <raison>` (serveur injoignable, erreur du serveur lui-même, absence de confirmation). |
++| Son propre modèle | Décharger le modèle de l'agent ne dure que jusqu'à son étape suivante : le modèle est nécessaire pour lire le résultat, et le serveur le recharge. |
++| Permission | Jamais demandée, même avec `yolo: false` : rien ne change sur disque, et le modèle se recharge tout seul. |
++| Affichage terminal | `🧹 unload_model: <modèle>` (`(own model)` quand il est vide), puis le résultat. |
++| ACP et fenêtre | Appel d'outil de type `other` ; un résultat `cannot unload` est un appel en échec. |
++| Délai | 30 s par requête au serveur. |
++
  ## Outils déclarés
  Déclarés à partir du bloc `tools:` du fichier d'agent, un outil par entrée (voir la [configuration](configuration.md#clés-de-tools)). Chacun porte le `name` de l'entrée ; ses champs sont les `parameters` de l'entrée.

modified internal/acp/acp.go +12 -0 Viewed

@@ -492,9 +492,21 @@ func (f front) slashCommand(ctx context.Context, sid sdk.SessionId, s session,
492	492	case command.Is(input, command.Compact):
493	493	return f.compactSession(ctx, sid, s), true
494	494	}
	495	+ if cmd, arg, ok := command.Parse(input); ok && cmd.Name == command.Unload {
	496	+ return f.unloadModel(ctx, sid, s, arg), true
	497	+ }
495	498	return sdk.PromptResponse{}, false
496	499	}
497	500
	501	+// unloadModel is /unload [model] over ACP: the engine's one line, sent as the
	502	+// agent's answer; the history is not touched.
	503	+func (f front) unloadModel(ctx context.Context, sid sdk.SessionId, s session, model string) sdk.PromptResponse {
	504	+ line := "🧹 " + f.e.Unload(ctx, model)
	505	+ fmt.Fprintf(os.Stderr, "[acp] session %s: /unload — %s\n", sid, line)
	506	+ _ = f.conn.SessionUpdate(ctx, sdk.SessionNotification{SessionId: sid, Update: sdk.UpdateAgentMessageText(line)})
	507	+ return sdk.PromptResponse{StopReason: sdk.StopReasonEndTurn, Usage: s.usage.totals()}
	508	+}
	509	+
498	510	// skillCommand turns a slash command that names a skill into the `$skill`
499	511	// notation, so one code path — skills.ExpandRefs, shared by all three front
500	512	// ends — decides what a named skill means.

  	case command.Is(input, command.Compact):
  		return f.compactSession(ctx, sid, s), true
  	}
++	if cmd, arg, ok := command.Parse(input); ok && cmd.Name == command.Unload {
++		return f.unloadModel(ctx, sid, s, arg), true
++	}
  	return sdk.PromptResponse{}, false
  }
++// unloadModel is /unload [model] over ACP: the engine's one line, sent as the
++// agent's answer; the history is not touched.
++func (f *front) unloadModel(ctx context.Context, sid sdk.SessionId, s *session, model string) sdk.PromptResponse {
++	line := "🧹 " + f.e.Unload(ctx, model)
++	fmt.Fprintf(os.Stderr, "[acp] session %s: /unload — %s\n", sid, line)
++	_ = f.conn.SessionUpdate(ctx, sdk.SessionNotification{SessionId: sid, Update: sdk.UpdateAgentMessageText(line)})
++	return sdk.PromptResponse{StopReason: sdk.StopReasonEndTurn, Usage: s.usage.totals()}
++}
++
  // skillCommand turns a slash command that names a skill into the `$skill`
  // notation, so one code path — skills.ExpandRefs, shared by all three front
  // ends — decides what a named skill means.

added internal/acp/unload_test.go +55 -0 Viewed

new file mode 100644
@@ -0,0 +1,55 @@
	1	+package acp
	2	+
	3	+import (
	4	+ "strings"
	5	+ "testing"
	6	+
	7	+ "mm/internal/command"
	8	+ "mm/internal/engine"
	9	+
	10	+ sdk "github.com/coder/acp-go-sdk"
	11	+)
	12	+
	13	+// With unloadModels on, /unload is advertised to the editor with an input
	14	+// hint, and session/prompt handles it: the engine's line comes back as the
	15	+// agent's text, the turn ends, and the history is untouched.
	16	+func TestUnloadOverACP(t *testing.T) {
	17	+ command.SetUnload(true)
	18	+ t.Cleanup(func() { command.SetUnload(false) })
	19	+
	20	+ advertised := false
	21	+ for _, c := range availableCommands() {
	22	+ if c.Name == command.Unload {
	23	+ advertised = c.Input != nil
	24	+ }
	25	+ }
	26	+ if !advertised {
	27	+ t.Error("/unload is not advertised with an input hint")
	28	+ }
	29	+
	30	+ w := newWire(t, "sys")
	31	+ f := frontOf(t, w)
	32	+ dmr, _ := engine.Lookup("dmr")
	33	+ f.e = &engine.Engine{Provider: dmr, Backend: engine.Backend{Provider: "dmr", Model: "m"}}
	34	+ w.send(1, "session/new", `{"cwd":"`+t.TempDir()+`","mcpServers":[]}`)
	35	+ _, resp := w.until(1)
	36	+ sid := resp["result"].(map[string]any)["sessionId"].(string)
	37	+
	38	+ w.send(2, "session/prompt", `{"sessionId":"`+sid+`","prompt":[{"type":"text","text":"/unload"}]}`)
	39	+ notes, resp := w.until(2)
	40	+ if stop := resp["result"].(map[string]any)["stopReason"]; stop != "end_turn" {
	41	+ t.Errorf("stopReason = %v", stop)
	42	+ }
	43	+ var text strings.Builder
	44	+ for _, n := range notes {
	45	+ if n["sessionUpdate"] == "agent_message_chunk" {
	46	+ text.WriteString(n["content"].(map[string]any)["text"].(string))
	47	+ }
	48	+ }
	49	+ if got := text.String(); got != "🧹 cannot unload m: the dmr provider cannot unload a model" {
	50	+ t.Errorf("agent text = %q", got)
	51	+ }
	52	+ if msgs := f.sessions[sdk.SessionId(sid)].messages; len(msgs) != 1 {
	53	+ t.Errorf("history: %d messages, want the system prompt alone", len(msgs))
	54	+ }
	55	+}

++package acp
++
++import (
++	"strings"
++	"testing"
++
++	"mm/internal/command"
++	"mm/internal/engine"
++
++	sdk "github.com/coder/acp-go-sdk"
++)
++
++// With unloadModels on, /unload is advertised to the editor with an input
++// hint, and session/prompt handles it: the engine's line comes back as the
++// agent's text, the turn ends, and the history is untouched.
++func TestUnloadOverACP(t *testing.T) {
++	command.SetUnload(true)
++	t.Cleanup(func() { command.SetUnload(false) })
++
++	advertised := false
++	for _, c := range availableCommands() {
++		if c.Name == command.Unload {
++			advertised = c.Input != nil
++		}
++	}
++	if !advertised {
++		t.Error("/unload is not advertised with an input hint")
++	}
++
++	w := newWire(t, "sys")
++	f := frontOf(t, w)
++	dmr, _ := engine.Lookup("dmr")
++	f.e = &engine.Engine{Provider: dmr, Backend: engine.Backend{Provider: "dmr", Model: "m"}}
++	w.send(1, "session/new", `{"cwd":"`+t.TempDir()+`","mcpServers":[]}`)
++	_, resp := w.until(1)
++	sid := resp["result"].(map[string]any)["sessionId"].(string)
++
++	w.send(2, "session/prompt", `{"sessionId":"`+sid+`","prompt":[{"type":"text","text":"/unload"}]}`)
++	notes, resp := w.until(2)
++	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(); got != "🧹 cannot unload m: the dmr provider cannot unload a model" {
++		t.Errorf("agent text = %q", got)
++	}
++	if msgs := f.sessions[sdk.SessionId(sid)].messages; len(msgs) != 1 {
++		t.Errorf("history: %d messages, want the system prompt alone", len(msgs))
++	}
++}

modified internal/agent/agent.go +2 -0 Viewed

@@ -562,6 +562,8 @@ func (l loop) historyCommand(input string, messages []ai.Message) ([]*ai.Messa
562	562	fmt.Println(copyAnswer(messages, arg, clipboard.New(terminalOSC52())))
563	563	case command.Save:
564	564	fmt.Println(saved.Command(modelAnswers(messages), arg))
	565	+ case command.Unload:
	566	+ fmt.Println("🧹 " + l.e.Unload(l.ctx, arg))
565	567	case command.Resume:
566	568	resumed, line, err := l.keep.Resume(arg, agentsmd.System(l.system, ""))
567	569	if err != nil {

  		fmt.Println(copyAnswer(messages, arg, clipboard.New(terminalOSC52())))
  	case command.Save:
  		fmt.Println(saved.Command(modelAnswers(messages), arg))
++	case command.Unload:
++		fmt.Println("🧹 " + l.e.Unload(l.ctx, arg))
  	case command.Resume:
  		resumed, line, err := l.keep.Resume(arg, agentsmd.System(l.system, ""))
  		if err != nil {

modified internal/command/command.go +7 -3 Viewed

@@ -41,6 +41,7 @@ const (
41	41	Help = "help"
42	42	Copy = "copy"
43	43	Save = "save"
	44	+ Unload = "unload"
44	45	)
45	46
46	47	// Command is one slash command: what it is called, what it does, who offers
@@ -76,9 +77,12 @@ var table = []Command{
76	77	{Name: Save, Description: "Save the last answer to a file, never overwriting one (/save notes.md 2: the answer before)", Fronts: Terminal \| Window, TakesArg: true, Hint: "<file> [n]"},
77	78	}
78	79
79		-// All returns the whole table, in display order: the fixed commands, then
80		-// the /ask family generated from the configured agents (ask.go).
81		-func All() []Command { return append(append([]Command(nil), table...), askCommands...) }
	80	+// All returns the whole table, in display order: the fixed commands, /unload
	81	+// when the configuration allows it (unload.go), then the /ask family generated
	82	+// from the configured agents (ask.go).
	83	+func All() []Command {
	84	+ return append(append(append([]Command(nil), table...), unloadCommands...), askCommands...)
	85	+}
82	86
83	87	// For returns the commands front f offers, in display order.
84	88	func For(f Front) []Command {

  	Help     = "help"
  	Copy     = "copy"
  	Save     = "save"
++	Unload   = "unload"
  )
  // Command is one slash command: what it is called, what it does, who offers
  	{Name: Save, Description: "Save the last answer to a file, never overwriting one (/save notes.md 2: the answer before)", Fronts: Terminal | Window, TakesArg: true, Hint: "<file> [n]"},
  }
--// All returns the whole table, in display order: the fixed commands, then
++// All returns the whole table, in display order: the fixed commands, /unload
--// the /ask family generated from the configured agents (ask.go).
++// when the configuration allows it (unload.go), then the /ask family generated
--func All() []Command { return append(append([]Command(nil), table...), askCommands...) }
++// from the configured agents (ask.go).
++func All() []Command {
++	return append(append(append([]Command(nil), table...), unloadCommands...), askCommands...)
++}
  // For returns the commands front f offers, in display order.
  func For(f Front) []Command {

added internal/command/unload.go +23 -0 Viewed

new file mode 100644
@@ -0,0 +1,23 @@
	1	+package command
	2	+
	3	+// unloadCommands holds /unload when config `unloadModels` allows it; empty
	4	+// otherwise, so /help and the menus do not show a command that is off.
	5	+var unloadCommands []Command
	6	+
	7	+// SetUnload adds /unload [model] to every front end, the editor included, or
	8	+// removes it. Called once by main after the config is loaded.
	9	+//
	10	+// command.SetUnload(config.Cfg.UnloadModels)
	11	+// cmd, arg, _ := command.Parse("/unload qwen3.5:2b") // Unload, "qwen3.5:2b"
	12	+func SetUnload(enabled bool) {
	13	+ unloadCommands = nil
	14	+ if enabled {
	15	+ unloadCommands = []Command{{
	16	+ Name: Unload,
	17	+ Description: "Unload a model from the server's memory: this agent's, or the one named (the next question reloads it)",
	18	+ Fronts: Terminal \| Editor \| Window,
	19	+ TakesArg: true,
	20	+ Hint: "[model]",
	21	+ }}
	22	+ }
	23	+}

++package command
++
++// unloadCommands holds /unload when config `unloadModels` allows it; empty
++// otherwise, so /help and the menus do not show a command that is off.
++var unloadCommands []Command
++
++// SetUnload adds /unload [model] to every front end, the editor included, or
++// removes it. Called once by main after the config is loaded.
++//
++//	command.SetUnload(config.Cfg.UnloadModels)
++//	cmd, arg, _ := command.Parse("/unload qwen3.5:2b") // Unload, "qwen3.5:2b"
++func SetUnload(enabled bool) {
++	unloadCommands = nil
++	if enabled {
++		unloadCommands = []Command{{
++			Name:        Unload,
++			Description: "Unload a model from the server's memory: this agent's, or the one named (the next question reloads it)",
++			Fronts:      Terminal | Editor | Window,
++			TakesArg:    true,
++			Hint:        "[model]",
++		}}
++	}
++}

added internal/command/unload_test.go +30 -0 Viewed

new file mode 100644
@@ -0,0 +1,30 @@
	1	+package command
	2	+
	3	+import "testing"
	4	+
	5	+func TestSetUnloadAddsAndRemovesTheCommand(t *testing.T) {
	6	+ t.Cleanup(func() { SetUnload(false) })
	7	+
	8	+ SetUnload(true)
	9	+ cmd, arg, ok := Parse("/unload qwen3.5:2b")
	10	+ if !ok \|\| cmd.Name != Unload \|\| arg != "qwen3.5:2b" {
	11	+ t.Fatalf("Parse = %+v %q %v", cmd, arg, ok)
	12	+ }
	13	+ if !Is("/unload", Unload) {
	14	+ t.Error("/unload alone is not the command")
	15	+ }
	16	+ for _, f := range []Front{Terminal, Window, Editor} {
	17	+ found := false
	18	+ for _, c := range For(f) {
	19	+ found = found \|\| c.Name == Unload
	20	+ }
	21	+ if !found {
	22	+ t.Errorf("front %d does not offer /unload", f)
	23	+ }
	24	+ }
	25	+
	26	+ SetUnload(false)
	27	+ if _, _, ok := Parse("/unload"); ok {
	28	+ t.Error("/unload still parsed with unloadModels off")
	29	+ }
	30	+}

++package command
++
++import "testing"
++
++func TestSetUnloadAddsAndRemovesTheCommand(t *testing.T) {
++	t.Cleanup(func() { SetUnload(false) })
++
++	SetUnload(true)
++	cmd, arg, ok := Parse("/unload qwen3.5:2b")
++	if !ok || cmd.Name != Unload || arg != "qwen3.5:2b" {
++		t.Fatalf("Parse = %+v %q %v", cmd, arg, ok)
++	}
++	if !Is("/unload", Unload) {
++		t.Error("/unload alone is not the command")
++	}
++	for _, f := range []Front{Terminal, Window, Editor} {
++		found := false
++		for _, c := range For(f) {
++			found = found || c.Name == Unload
++		}
++		if !found {
++			t.Errorf("front %d does not offer /unload", f)
++		}
++	}
++
++	SetUnload(false)
++	if _, _, ok := Parse("/unload"); ok {
++		t.Error("/unload still parsed with unloadModels off")
++	}
++}

modified internal/config/config.go +8 -0 Viewed

@@ -139,6 +139,13 @@ type Config struct {
139	139	// way.
140	140	SaveAnswerTool bool `yaml:"saveAnswerTool"`
141	141
	142	+ // UnloadModels: the /unload command and the `unload_model` tool, which
	143	+ // free a model from the server's memory (ticket 101; llmman, ollama,
	144	+ // llamacpp in router mode). On by default: unloading is harmless, the
	145	+ // next request reloads. The tool is never declared beside a
	146	+ // structuredOutput, like ask_user.
	147	+ UnloadModels bool `yaml:"unloadModels"`
	148	+
142	149	// Sampling: generation settings (OpenAI API keys; the plugin converts them.)
143	150	Sampling map[string]any `yaml:"sampling"`
144	151
@@ -534,6 +541,7 @@ var Cfg = Config{
534	541	EditTools: true,
535	542	PlanTool: true,
536	543	AskUserTool: true,
	544	+ UnloadModels: true,
537	545	// FetchContentTool stays false (the zero value): see its field comment.
538	546	System: `You are a coding agent working in a terminal.
539	547	You have a "bash" tool to run shell commands.

  	// way.
  	SaveAnswerTool bool `yaml:"saveAnswerTool"`
++	// UnloadModels: the /unload command and the `unload_model` tool, which
++	// free a model from the server's memory (ticket 101; llmman, ollama,
++	// llamacpp in router mode). On by default: unloading is harmless, the
++	// next request reloads. The tool is never declared beside a
++	// structuredOutput, like ask_user.
++	UnloadModels bool `yaml:"unloadModels"`
++
  	// Sampling: generation settings (OpenAI API keys; the plugin converts them.)
  	Sampling map[string]any `yaml:"sampling"`
  	EditTools:       true,
  	PlanTool:        true,
  	AskUserTool:     true,
++	UnloadModels:    true,
  	// FetchContentTool stays false (the zero value): see its field comment.
  	System: `You are a coding agent working in a terminal.
  You have a "bash" tool to run shell commands.

modified internal/config/config_test.go +20 -0 Viewed

@@ -238,3 +238,23 @@ func TestDisplayHooksConfig(t *testing.T) {
238	238	t.Error("displayHooks: true was not read")
239	239	}
240	240	}
	241	+
	242	+// /unload and unload_model are on unless the YAML turns them off.
	243	+func TestUnloadModelsConfig(t *testing.T) {
	244	+ defaults := Cfg
	245	+ t.Cleanup(func() { Cfg = defaults })
	246	+
	247	+ if !Cfg.UnloadModels {
	248	+ t.Fatal("unloadModels is off by default, want on")
	249	+ }
	250	+ off := filepath.Join(t.TempDir(), "off.yaml")
	251	+ if err := os.WriteFile(off, []byte("unloadModels: false\n"), 0o644); err != nil {
	252	+ t.Fatal(err)
	253	+ }
	254	+ if _, err := Load(off); err != nil {
	255	+ t.Fatal(err)
	256	+ }
	257	+ if Cfg.UnloadModels {
	258	+ t.Error("unloadModels: false was not read")
	259	+ }
	260	+}

  		t.Error("displayHooks: true was not read")
  	}
  }
++
++// /unload and unload_model are on unless the YAML turns them off.
++func TestUnloadModelsConfig(t *testing.T) {
++	defaults := Cfg
++	t.Cleanup(func() { Cfg = defaults })
++
++	if !Cfg.UnloadModels {
++		t.Fatal("unloadModels is off by default, want on")
++	}
++	off := filepath.Join(t.TempDir(), "off.yaml")
++	if err := os.WriteFile(off, []byte("unloadModels: false\n"), 0o644); err != nil {
++		t.Fatal(err)
++	}
++	if _, err := Load(off); err != nil {
++		t.Fatal(err)
++	}
++	if Cfg.UnloadModels {
++		t.Error("unloadModels: false was not read")
++	}
++}

modified internal/config/tools.go +1 -0 Viewed

@@ -44,6 +44,7 @@ var BuiltinToolNames = []string{
44	44	"read_file",
45	45	"read_skill",
46	46	"save_answer",
	47	+ "unload_model",
47	48	"write_file",
48	49	}
49	50

  	"read_file",
  	"read_skill",
  	"save_answer",
++	"unload_model",
  	"write_file",
  }

modified internal/engine/openai_compat.go +6 -0 Viewed

@@ -59,6 +59,9 @@ type openaiCompat struct {
59	59	// thinkingOff: how this server is told not to reason, for the summary of
60	60	// a compaction (usage.go). Zero value = no known way.
61	61	thinkingOff thinkingOff
	62	+
	63	+ // unload: how this server unloads a model (unload.go); nil = it cannot.
	64	+ unload unloadFunc
62	65	}
63	66
64	67	func init() {
@@ -86,6 +89,7 @@ func init() {
86	89	probe: probeLlamaCpp,
87	90	progress: true,
88	91	thinkingOff: thinkingOffTemplate,
	92	+ unload: unloadRouter,
89	93	})
90	94	register(&openaiCompat{
91	95	name: "llmman",
@@ -104,6 +108,7 @@ func init() {
104	108	// enable_thinking:false answers without reasoning).
105	109	progress: true,
106	110	thinkingOff: thinkingOffTemplate,
	111	+ unload: unloadKeepAlive,
107	112	})
108	113	register(&openaiCompat{
109	114	name: "ollama",
@@ -120,6 +125,7 @@ func init() {
120	125	// it), chat_template_kwargs ignored but reasoning_effort "none" obeyed.
121	126	progress: false,
122	127	thinkingOff: thinkingOffEffort,
	128	+ unload: unloadKeepAlive,
123	129	})
124	130	}
125	131

  	// thinkingOff: how this server is told not to reason, for the summary of
  	// a compaction (usage.go). Zero value = no known way.
  	thinkingOff thinkingOff
++
++	// unload: how this server unloads a model (unload.go); nil = it cannot.
++	unload unloadFunc
  }
  func init() {
  		probe:        probeLlamaCpp,
  		progress:     true,
  		thinkingOff:  thinkingOffTemplate,
++		unload:       unloadRouter,
  	})
  	register(&openaiCompat{
  		name:     "llmman",
  		// enable_thinking:false answers without reasoning).
  		progress:    true,
  		thinkingOff: thinkingOffTemplate,
++		unload:      unloadKeepAlive,
  	})
  	register(&openaiCompat{
  		name:     "ollama",
  		// it), chat_template_kwargs ignored but reasoning_effort "none" obeyed.
  		progress:    false,
  		thinkingOff: thinkingOffEffort,
++		unload:      unloadKeepAlive,
  	})
  }

added internal/engine/unload.go +196 -0 Viewed

new file mode 100644
@@ -0,0 +1,196 @@
	1	+package engine
	2	+
	3	+import (
	4	+ "bytes"
	5	+ "context"
	6	+ "encoding/json"
	7	+ "errors"
	8	+ "fmt"
	9	+ "io"
	10	+ "net/http"
	11	+ "slices"
	12	+ "strings"
	13	+ "time"
	14	+)
	15	+
	16	+// Unloading a model from the server's memory (ticket 101): /unload and the
	17	+// unload_model tool. The model comes back on the next request that names it,
	18	+// so unloading is harmless — it frees the RAM a model held, e.g. before
	19	+// another agent's model loads. Each server does it its own way (ticket 92):
	20	+//
	21	+// - llmman, Ollama: POST /api/generate {"model": m, "keep_alive": 0};
	22	+// - llama-server in router mode: POST /models/unload {"model": m};
	23	+// - a single-model llama-server, Docker Model Runner: no route.
	24	+
	25	+// unloadFunc unloads model from the server behind b and returns one line
	26	+// saying what happened. An error is a failure the caller words itself.
	27	+type unloadFunc func(ctx context.Context, b Backend, model string) (string, error)
	28	+
	29	+// Unloader is the optional side of a Provider that can unload a model.
	30	+type Unloader interface {
	31	+ Unload(ctx context.Context, b Backend, model string) (string, error)
	32	+}
	33	+
	34	+// unloadClient bounds one unload: the server answers once the backend process
	35	+// is stopped, which takes a few seconds, never minutes.
	36	+var unloadClient = &http.Client{Timeout: 30 * time.Second}
	37	+
	38	+// Unload unloads model (this engine's own model when empty) and returns ONE
	39	+// line for the user or the model: what was unloaded, that there was nothing
	40	+// to unload, or why it could not be done. Never an error: the line is the
	41	+// answer, the same way a failed command is text for the model.
	42	+//
	43	+// fmt.Println(e.Unload(ctx, "")) // "unloaded hf.co/…:Q4_K_M"
	44	+// fmt.Println(e.Unload(ctx, "qwen3.5:2b")) // another model of the same server
	45	+func (e *Engine) Unload(ctx context.Context, model string) string {
	46	+ model = strings.TrimSpace(model)
	47	+ if model == "" {
	48	+ model = e.Backend.Model
	49	+ }
	50	+ u, ok := e.Provider.(Unloader)
	51	+ if !ok {
	52	+ return fmt.Sprintf("cannot unload %s: the %s provider cannot unload a model", model, e.Backend.Provider)
	53	+ }
	54	+ line, err := u.Unload(ctx, e.Backend, model)
	55	+ if err != nil {
	56	+ return fmt.Sprintf("cannot unload %s: %v", model, err)
	57	+ }
	58	+ return line
	59	+}
	60	+
	61	+// Unload implements Unloader with the server's own way, when it has one.
	62	+func (p *openaiCompat) Unload(ctx context.Context, b Backend, model string) (string, error) {
	63	+ if p.unload == nil {
	64	+ return "", fmt.Errorf("the %s provider cannot unload a model", p.name)
	65	+ }
	66	+ return p.unload(ctx, b, model)
	67	+}
	68	+
	69	+// rootURL is the server's address without the OpenAI prefix: where the
	70	+// Ollama routes (/api/…) and llama-server's router routes (/models/…) live.
	71	+func rootURL(baseURL string) string {
	72	+ return strings.TrimSuffix(strings.TrimSuffix(baseURL, "/"), "/v1")
	73	+}
	74	+
	75	+// unloadKeepAlive is the llmman and Ollama way: a generate request with
	76	+// keep_alive 0. The model is first looked up in /api/ps: a generate for a
	77	+// model that is NOT loaded would load it only to unload it — minutes for a
	78	+// large one, and memory taken from whatever is running.
	79	+func unloadKeepAlive(ctx context.Context, b Backend, model string) (string, error) {
	80	+ root := rootURL(b.BaseURL)
	81	+ loaded, err := loadedModels(ctx, b, root)
	82	+ if err != nil {
	83	+ return "", err
	84	+ }
	85	+ if !slices.Contains(loaded, model) {
	86	+ return fmt.Sprintf("%s was not loaded: nothing to unload", model), nil
	87	+ }
	88	+ var reply struct {
	89	+ DoneReason string `json:"done_reason"`
	90	+ Error string `json:"error"`
	91	+ }
	92	+ if err := postJSON(ctx, b, root+"/api/generate", map[string]any{"model": model, "keep_alive": 0}, &reply); err != nil {
	93	+ return "", err
	94	+ }
	95	+ if reply.Error != "" {
	96	+ return "", fmt.Errorf("the server says: %s", reply.Error)
	97	+ }
	98	+ if reply.DoneReason != "unload" {
	99	+ return "", fmt.Errorf("the server did not confirm (done_reason %q)", reply.DoneReason)
	100	+ }
	101	+ return "unloaded " + model, nil
	102	+}
	103	+
	104	+// loadedModels lists the names /api/ps reports as loaded.
	105	+func loadedModels(ctx context.Context, b Backend, root string) ([]string, error) {
	106	+ var ps struct {
	107	+ Models []struct {
	108	+ Name string `json:"name"`
	109	+ } `json:"models"`
	110	+ }
	111	+ if err := getWithKey(ctx, b, root+"/api/ps", &ps); err != nil {
	112	+ return nil, err
	113	+ }
	114	+ names := make([]string, 0, len(ps.Models))
	115	+ for _, m := range ps.Models {
	116	+ names = append(names, m.Name)
	117	+ }
	118	+ return names, nil
	119	+}
	120	+
	121	+// unloadRouter is the llama-server way, in router mode only. A single-model
	122	+// llama-server has no /models/unload: its 404 is turned into the two ways it
	123	+// does free memory.
	124	+func unloadRouter(ctx context.Context, b Backend, model string) (string, error) {
	125	+ var reply struct {
	126	+ Success bool `json:"success"`
	127	+ Error any `json:"error"`
	128	+ }
	129	+ err := postJSON(ctx, b, rootURL(b.BaseURL)+"/models/unload", map[string]any{"model": model}, &reply)
	130	+ var status *httpStatusError
	131	+ switch {
	132	+ case errors.As(err, &status) && status.code == http.StatusNotFound:
	133	+ return "", fmt.Errorf("this llama-server serves a single model and has no unload route — start it in router mode (no -m), or with --sleep-idle-seconds N to free memory when idle, or stop it")
	134	+ case err != nil:
	135	+ return "", err
	136	+ case !reply.Success:
	137	+ return "", fmt.Errorf("the server did not confirm: %v", reply.Error)
	138	+ }
	139	+ return "unloaded " + model, nil
	140	+}
	141	+
	142	+// httpStatusError is a non-2xx answer, with the start of its body.
	143	+type httpStatusError struct {
	144	+ code int
	145	+ body string
	146	+}
	147	+
	148	+func (e *httpStatusError) Error() string {
	149	+ if e.body == "" {
	150	+ return fmt.Sprintf("the server answered %d", e.code)
	151	+ }
	152	+ return fmt.Sprintf("the server answered %d: %s", e.code, e.body)
	153	+}
	154	+
	155	+// postJSON posts body as JSON and decodes a 2xx answer into out.
	156	+func postJSON(ctx context.Context, b Backend, url string, body, out any) error {
	157	+ data, err := json.Marshal(body)
	158	+ if err != nil {
	159	+ return err
	160	+ }
	161	+ req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(data))
	162	+ if err != nil {
	163	+ return err
	164	+ }
	165	+ req.Header.Set("Content-Type", "application/json")
	166	+ return send(req, b, out)
	167	+}
	168	+
	169	+// getWithKey is GET url, decoded into out, with the backend's key.
	170	+func getWithKey(ctx context.Context, b Backend, url string, out any) error {
	171	+ req, err := http.NewRequestWithContext(ctx, http.MethodGet, url, nil)
	172	+ if err != nil {
	173	+ return err
	174	+ }
	175	+ return send(req, b, out)
	176	+}
	177	+
	178	+// send runs req with the backend's key and decodes a 2xx JSON answer.
	179	+func send(req *http.Request, b Backend, out any) error {
	180	+ if b.APIKey != "" {
	181	+ req.Header.Set("Authorization", "Bearer "+b.APIKey)
	182	+ }
	183	+ resp, err := unloadClient.Do(req)
	184	+ if err != nil {
	185	+ return fmt.Errorf("the server cannot be reached: %w", err)
	186	+ }
	187	+ defer func() { _ = resp.Body.Close() }()
	188	+ if resp.StatusCode/100 != 2 {
	189	+ body, _ := io.ReadAll(io.LimitReader(resp.Body, 300))
	190	+ return &httpStatusError{code: resp.StatusCode, body: strings.TrimSpace(string(body))}
	191	+ }
	192	+ if err := json.NewDecoder(resp.Body).Decode(out); err != nil {
	193	+ return fmt.Errorf("unreadable answer: %w", err)
	194	+ }
	195	+ return nil
	196	+}

++package engine
++
++import (
++	"bytes"
++	"context"
++	"encoding/json"
++	"errors"
++	"fmt"
++	"io"
++	"net/http"
++	"slices"
++	"strings"
++	"time"
++)
++
++// Unloading a model from the server's memory (ticket 101): /unload and the
++// unload_model tool. The model comes back on the next request that names it,
++// so unloading is harmless — it frees the RAM a model held, e.g. before
++// another agent's model loads. Each server does it its own way (ticket 92):
++//
++//   - llmman, Ollama: POST /api/generate {"model": m, "keep_alive": 0};
++//   - llama-server in router mode: POST /models/unload {"model": m};
++//   - a single-model llama-server, Docker Model Runner: no route.
++
++// unloadFunc unloads model from the server behind b and returns one line
++// saying what happened. An error is a failure the caller words itself.
++type unloadFunc func(ctx context.Context, b Backend, model string) (string, error)
++
++// Unloader is the optional side of a Provider that can unload a model.
++type Unloader interface {
++	Unload(ctx context.Context, b Backend, model string) (string, error)
++}
++
++// unloadClient bounds one unload: the server answers once the backend process
++// is stopped, which takes a few seconds, never minutes.
++var unloadClient = &http.Client{Timeout: 30 * time.Second}
++
++// Unload unloads model (this engine's own model when empty) and returns ONE
++// line for the user or the model: what was unloaded, that there was nothing
++// to unload, or why it could not be done. Never an error: the line is the
++// answer, the same way a failed command is text for the model.
++//
++//	fmt.Println(e.Unload(ctx, ""))                 // "unloaded hf.co/…:Q4_K_M"
++//	fmt.Println(e.Unload(ctx, "qwen3.5:2b"))       // another model of the same server
++func (e *Engine) Unload(ctx context.Context, model string) string {
++	model = strings.TrimSpace(model)
++	if model == "" {
++		model = e.Backend.Model
++	}
++	u, ok := e.Provider.(Unloader)
++	if !ok {
++		return fmt.Sprintf("cannot unload %s: the %s provider cannot unload a model", model, e.Backend.Provider)
++	}
++	line, err := u.Unload(ctx, e.Backend, model)
++	if err != nil {
++		return fmt.Sprintf("cannot unload %s: %v", model, err)
++	}
++	return line
++}
++
++// Unload implements Unloader with the server's own way, when it has one.
++func (p *openaiCompat) Unload(ctx context.Context, b Backend, model string) (string, error) {
++	if p.unload == nil {
++		return "", fmt.Errorf("the %s provider cannot unload a model", p.name)
++	}
++	return p.unload(ctx, b, model)
++}
++
++// rootURL is the server's address without the OpenAI prefix: where the
++// Ollama routes (/api/…) and llama-server's router routes (/models/…) live.
++func rootURL(baseURL string) string {
++	return strings.TrimSuffix(strings.TrimSuffix(baseURL, "/"), "/v1")
++}
++
++// unloadKeepAlive is the llmman and Ollama way: a generate request with
++// keep_alive 0. The model is first looked up in /api/ps: a generate for a
++// model that is NOT loaded would load it only to unload it — minutes for a
++// large one, and memory taken from whatever is running.
++func unloadKeepAlive(ctx context.Context, b Backend, model string) (string, error) {
++	root := rootURL(b.BaseURL)
++	loaded, err := loadedModels(ctx, b, root)
++	if err != nil {
++		return "", err
++	}
++	if !slices.Contains(loaded, model) {
++		return fmt.Sprintf("%s was not loaded: nothing to unload", model), nil
++	}
++	var reply struct {
++		DoneReason string `json:"done_reason"`
++		Error      string `json:"error"`
++	}
++	if err := postJSON(ctx, b, root+"/api/generate", map[string]any{"model": model, "keep_alive": 0}, &reply); err != nil {
++		return "", err
++	}
++	if reply.Error != "" {
++		return "", fmt.Errorf("the server says: %s", reply.Error)
++	}
++	if reply.DoneReason != "unload" {
++		return "", fmt.Errorf("the server did not confirm (done_reason %q)", reply.DoneReason)
++	}
++	return "unloaded " + model, nil
++}
++
++// loadedModels lists the names /api/ps reports as loaded.
++func loadedModels(ctx context.Context, b Backend, root string) ([]string, error) {
++	var ps struct {
++		Models []struct {
++			Name string `json:"name"`
++		} `json:"models"`
++	}
++	if err := getWithKey(ctx, b, root+"/api/ps", &ps); err != nil {
++		return nil, err
++	}
++	names := make([]string, 0, len(ps.Models))
++	for _, m := range ps.Models {
++		names = append(names, m.Name)
++	}
++	return names, nil
++}
++
++// unloadRouter is the llama-server way, in router mode only. A single-model
++// llama-server has no /models/unload: its 404 is turned into the two ways it
++// does free memory.
++func unloadRouter(ctx context.Context, b Backend, model string) (string, error) {
++	var reply struct {
++		Success bool `json:"success"`
++		Error   any  `json:"error"`
++	}
++	err := postJSON(ctx, b, rootURL(b.BaseURL)+"/models/unload", map[string]any{"model": model}, &reply)
++	var status *httpStatusError
++	switch {
++	case errors.As(err, &status) && status.code == http.StatusNotFound:
++		return "", fmt.Errorf("this llama-server serves a single model and has no unload route — start it in router mode (no -m), or with --sleep-idle-seconds N to free memory when idle, or stop it")
++	case err != nil:
++		return "", err
++	case !reply.Success:
++		return "", fmt.Errorf("the server did not confirm: %v", reply.Error)
++	}
++	return "unloaded " + model, nil
++}
++
++// httpStatusError is a non-2xx answer, with the start of its body.
++type httpStatusError struct {
++	code int
++	body string
++}
++
++func (e *httpStatusError) Error() string {
++	if e.body == "" {
++		return fmt.Sprintf("the server answered %d", e.code)
++	}
++	return fmt.Sprintf("the server answered %d: %s", e.code, e.body)
++}
++
++// postJSON posts body as JSON and decodes a 2xx answer into out.
++func postJSON(ctx context.Context, b Backend, url string, body, out any) error {
++	data, err := json.Marshal(body)
++	if err != nil {
++		return err
++	}
++	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(data))
++	if err != nil {
++		return err
++	}
++	req.Header.Set("Content-Type", "application/json")
++	return send(req, b, out)
++}
++
++// getWithKey is GET url, decoded into out, with the backend's key.
++func getWithKey(ctx context.Context, b Backend, url string, out any) error {
++	req, err := http.NewRequestWithContext(ctx, http.MethodGet, url, nil)
++	if err != nil {
++		return err
++	}
++	return send(req, b, out)
++}
++
++// send runs req with the backend's key and decodes a 2xx JSON answer.
++func send(req *http.Request, b Backend, out any) error {
++	if b.APIKey != "" {
++		req.Header.Set("Authorization", "Bearer "+b.APIKey)
++	}
++	resp, err := unloadClient.Do(req)
++	if err != nil {
++		return fmt.Errorf("the server cannot be reached: %w", err)
++	}
++	defer func() { _ = resp.Body.Close() }()
++	if resp.StatusCode/100 != 2 {
++		body, _ := io.ReadAll(io.LimitReader(resp.Body, 300))
++		return &httpStatusError{code: resp.StatusCode, body: strings.TrimSpace(string(body))}
++	}
++	if err := json.NewDecoder(resp.Body).Decode(out); err != nil {
++		return fmt.Errorf("unreadable answer: %w", err)
++	}
++	return nil
++}

added internal/engine/unload_test.go +139 -0 Viewed

new file mode 100644
@@ -0,0 +1,139 @@
	1	+package engine
	2	+
	3	+import (
	4	+ "context"
	5	+ "encoding/json"
	6	+ "net/http"
	7	+ "net/http/httptest"
	8	+ "strings"
	9	+ "sync"
	10	+ "testing"
	11	+)
	12	+
	13	+// fakeOllama is an Ollama-style server: /api/ps lists loaded, and a
	14	+// keep_alive 0 generate unloads the model it names.
	15	+type fakeOllama struct {
	16	+ mu sync.Mutex
	17	+ loaded []string
	18	+ generate []map[string]any
	19	+ reply string // the generate answer; "" = the real one
	20	+}
	21	+
	22	+func (f fakeOllama) server(t testing.T) *httptest.Server {
	23	+ t.Helper()
	24	+ srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
	25	+ f.mu.Lock()
	26	+ defer f.mu.Unlock()
	27	+ switch r.URL.Path {
	28	+ case "/api/ps":
	29	+ models := []map[string]any{}
	30	+ for _, m := range f.loaded {
	31	+ models = append(models, map[string]any{"name": m})
	32	+ }
	33	+ _ = json.NewEncoder(w).Encode(map[string]any{"models": models})
	34	+ case "/api/generate":
	35	+ var body map[string]any
	36	+ _ = json.NewDecoder(r.Body).Decode(&body)
	37	+ f.generate = append(f.generate, body)
	38	+ if f.reply != "" {
	39	+ _, _ = w.Write([]byte(f.reply))
	40	+ return
	41	+ }
	42	+ f.loaded = nil
	43	+ _, _ = w.Write([]byte(`{"model":"` + body["model"].(string) + `","done":true,"done_reason":"unload"}`))
	44	+ default:
	45	+ http.NotFound(w, r)
	46	+ }
	47	+ }))
	48	+ t.Cleanup(srv.Close)
	49	+ return srv
	50	+}
	51	+
	52	+func TestUnloadKeepAliveUnloadsALoadedModel(t *testing.T) {
	53	+ f := &fakeOllama{loaded: []string{"qwen3.5:2b"}}
	54	+ srv := f.server(t)
	55	+ line, err := unloadKeepAlive(context.Background(), Backend{BaseURL: srv.URL + "/v1"}, "qwen3.5:2b")
	56	+ if err != nil \|\| line != "unloaded qwen3.5:2b" {
	57	+ t.Fatalf("unload = %q, %v", line, err)
	58	+ }
	59	+ if len(f.generate) != 1 \|\| f.generate[0]["keep_alive"] != float64(0) \|\| f.generate[0]["model"] != "qwen3.5:2b" {
	60	+ t.Errorf("generate requests = %v, want one keep_alive 0 for the model", f.generate)
	61	+ }
	62	+}
	63	+
	64	+// A model that is not loaded is not asked for: a generate would load it
	65	+// first, only to unload it.
	66	+func TestUnloadKeepAliveSkipsAModelThatIsNotLoaded(t *testing.T) {
	67	+ f := &fakeOllama{loaded: []string{"other"}}
	68	+ srv := f.server(t)
	69	+ line, err := unloadKeepAlive(context.Background(), Backend{BaseURL: srv.URL + "/v1"}, "qwen3.5:2b")
	70	+ if err != nil \|\| line != "qwen3.5:2b was not loaded: nothing to unload" {
	71	+ t.Fatalf("unload = %q, %v", line, err)
	72	+ }
	73	+ if len(f.generate) != 0 {
	74	+ t.Errorf("a generate was sent for a model that was not loaded: %v", f.generate)
	75	+ }
	76	+}
	77	+
	78	+func TestUnloadKeepAliveReportsAServerThatDoesNotConfirm(t *testing.T) {
	79	+ f := &fakeOllama{loaded: []string{"m"}, reply: `{"error":"model is busy"}`}
	80	+ srv := f.server(t)
	81	+ if _, err := unloadKeepAlive(context.Background(), Backend{BaseURL: srv.URL + "/v1"}, "m"); err == nil \|\| !strings.Contains(err.Error(), "model is busy") {
	82	+ t.Errorf("err = %v, want the server's words", err)
	83	+ }
	84	+ f.reply = `{"done":true,"done_reason":"stop"}`
	85	+ if _, err := unloadKeepAlive(context.Background(), Backend{BaseURL: srv.URL + "/v1"}, "m"); err == nil \|\| !strings.Contains(err.Error(), "did not confirm") {
	86	+ t.Errorf("err = %v, want an unconfirmed unload", err)
	87	+ }
	88	+}
	89	+
	90	+func TestUnloadRouter(t *testing.T) {
	91	+ var got map[string]any
	92	+ srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
	93	+ if r.URL.Path != "/models/unload" {
	94	+ http.NotFound(w, r)
	95	+ return
	96	+ }
	97	+ _ = json.NewDecoder(r.Body).Decode(&got)
	98	+ _, _ = w.Write([]byte(`{"success":true}`))
	99	+ }))
	100	+ t.Cleanup(srv.Close)
	101	+ line, err := unloadRouter(context.Background(), Backend{BaseURL: srv.URL + "/v1"}, "gemma")
	102	+ if err != nil \|\| line != "unloaded gemma" \|\| got["model"] != "gemma" {
	103	+ t.Fatalf("unload = %q, %v; request %v", line, err, got)
	104	+ }
	105	+}
	106	+
	107	+// A single-model llama-server answers 404: the error says what it can do
	108	+// instead.
	109	+func TestUnloadRouterSingleModelServer(t *testing.T) {
	110	+ srv := httptest.NewServer(http.NotFoundHandler())
	111	+ t.Cleanup(srv.Close)
	112	+ _, err := unloadRouter(context.Background(), Backend{BaseURL: srv.URL + "/v1"}, "gemma")
	113	+ if err == nil \|\| !strings.Contains(err.Error(), "--sleep-idle-seconds") {
	114	+ t.Errorf("err = %v, want the single-model explanation", err)
	115	+ }
	116	+}
	117	+
	118	+// Engine.Unload: the engine's own model when none is named, one line in
	119	+// every case, and a provider without a way says so.
	120	+func TestEngineUnloadLines(t *testing.T) {
	121	+ f := &fakeOllama{loaded: []string{"own"}}
	122	+ srv := f.server(t)
	123	+ llmman, _ := Lookup("llmman")
	124	+ e := &Engine{Provider: llmman, Backend: Backend{Provider: "llmman", BaseURL: srv.URL + "/v1", Model: "own"}}
	125	+ if got := e.Unload(context.Background(), " "); got != "unloaded own" {
	126	+ t.Errorf("Unload(\"\") = %q, want the engine's own model unloaded", got)
	127	+ }
	128	+
	129	+ dmr, _ := Lookup("dmr")
	130	+ e = &Engine{Provider: dmr, Backend: Backend{Provider: "dmr", Model: "m"}}
	131	+ if got := e.Unload(context.Background(), ""); !strings.HasPrefix(got, "cannot unload m:") {
	132	+ t.Errorf("dmr Unload = %q, want a refusal", got)
	133	+ }
	134	+
	135	+ e = &Engine{Provider: llmman, Backend: Backend{Provider: "llmman", BaseURL: "http://127.0.0.1:1/v1", Model: "m"}}
	136	+ if got := e.Unload(context.Background(), ""); !strings.Contains(got, "cannot be reached") {
	137	+ t.Errorf("unreachable Unload = %q", got)
	138	+ }
	139	+}

++package engine
++
++import (
++	"context"
++	"encoding/json"
++	"net/http"
++	"net/http/httptest"
++	"strings"
++	"sync"
++	"testing"
++)
++
++// fakeOllama is an Ollama-style server: /api/ps lists loaded, and a
++// keep_alive 0 generate unloads the model it names.
++type fakeOllama struct {
++	mu       sync.Mutex
++	loaded   []string
++	generate []map[string]any
++	reply    string // the generate answer; "" = the real one
++}
++
++func (f *fakeOllama) server(t *testing.T) *httptest.Server {
++	t.Helper()
++	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
++		f.mu.Lock()
++		defer f.mu.Unlock()
++		switch r.URL.Path {
++		case "/api/ps":
++			models := []map[string]any{}
++			for _, m := range f.loaded {
++				models = append(models, map[string]any{"name": m})
++			}
++			_ = json.NewEncoder(w).Encode(map[string]any{"models": models})
++		case "/api/generate":
++			var body map[string]any
++			_ = json.NewDecoder(r.Body).Decode(&body)
++			f.generate = append(f.generate, body)
++			if f.reply != "" {
++				_, _ = w.Write([]byte(f.reply))
++				return
++			}
++			f.loaded = nil
++			_, _ = w.Write([]byte(`{"model":"` + body["model"].(string) + `","done":true,"done_reason":"unload"}`))
++		default:
++			http.NotFound(w, r)
++		}
++	}))
++	t.Cleanup(srv.Close)
++	return srv
++}
++
++func TestUnloadKeepAliveUnloadsALoadedModel(t *testing.T) {
++	f := &fakeOllama{loaded: []string{"qwen3.5:2b"}}
++	srv := f.server(t)
++	line, err := unloadKeepAlive(context.Background(), Backend{BaseURL: srv.URL + "/v1"}, "qwen3.5:2b")
++	if err != nil || line != "unloaded qwen3.5:2b" {
++		t.Fatalf("unload = %q, %v", line, err)
++	}
++	if len(f.generate) != 1 || f.generate[0]["keep_alive"] != float64(0) || f.generate[0]["model"] != "qwen3.5:2b" {
++		t.Errorf("generate requests = %v, want one keep_alive 0 for the model", f.generate)
++	}
++}
++
++// A model that is not loaded is not asked for: a generate would load it
++// first, only to unload it.
++func TestUnloadKeepAliveSkipsAModelThatIsNotLoaded(t *testing.T) {
++	f := &fakeOllama{loaded: []string{"other"}}
++	srv := f.server(t)
++	line, err := unloadKeepAlive(context.Background(), Backend{BaseURL: srv.URL + "/v1"}, "qwen3.5:2b")
++	if err != nil || line != "qwen3.5:2b was not loaded: nothing to unload" {
++		t.Fatalf("unload = %q, %v", line, err)
++	}
++	if len(f.generate) != 0 {
++		t.Errorf("a generate was sent for a model that was not loaded: %v", f.generate)
++	}
++}
++
++func TestUnloadKeepAliveReportsAServerThatDoesNotConfirm(t *testing.T) {
++	f := &fakeOllama{loaded: []string{"m"}, reply: `{"error":"model is busy"}`}
++	srv := f.server(t)
++	if _, err := unloadKeepAlive(context.Background(), Backend{BaseURL: srv.URL + "/v1"}, "m"); err == nil || !strings.Contains(err.Error(), "model is busy") {
++		t.Errorf("err = %v, want the server's words", err)
++	}
++	f.reply = `{"done":true,"done_reason":"stop"}`
++	if _, err := unloadKeepAlive(context.Background(), Backend{BaseURL: srv.URL + "/v1"}, "m"); err == nil || !strings.Contains(err.Error(), "did not confirm") {
++		t.Errorf("err = %v, want an unconfirmed unload", err)
++	}
++}
++
++func TestUnloadRouter(t *testing.T) {
++	var got map[string]any
++	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
++		if r.URL.Path != "/models/unload" {
++			http.NotFound(w, r)
++			return
++		}
++		_ = json.NewDecoder(r.Body).Decode(&got)
++		_, _ = w.Write([]byte(`{"success":true}`))
++	}))
++	t.Cleanup(srv.Close)
++	line, err := unloadRouter(context.Background(), Backend{BaseURL: srv.URL + "/v1"}, "gemma")
++	if err != nil || line != "unloaded gemma" || got["model"] != "gemma" {
++		t.Fatalf("unload = %q, %v; request %v", line, err, got)
++	}
++}
++
++// A single-model llama-server answers 404: the error says what it can do
++// instead.
++func TestUnloadRouterSingleModelServer(t *testing.T) {
++	srv := httptest.NewServer(http.NotFoundHandler())
++	t.Cleanup(srv.Close)
++	_, err := unloadRouter(context.Background(), Backend{BaseURL: srv.URL + "/v1"}, "gemma")
++	if err == nil || !strings.Contains(err.Error(), "--sleep-idle-seconds") {
++		t.Errorf("err = %v, want the single-model explanation", err)
++	}
++}
++
++// Engine.Unload: the engine's own model when none is named, one line in
++// every case, and a provider without a way says so.
++func TestEngineUnloadLines(t *testing.T) {
++	f := &fakeOllama{loaded: []string{"own"}}
++	srv := f.server(t)
++	llmman, _ := Lookup("llmman")
++	e := &Engine{Provider: llmman, Backend: Backend{Provider: "llmman", BaseURL: srv.URL + "/v1", Model: "own"}}
++	if got := e.Unload(context.Background(), "  "); got != "unloaded own" {
++		t.Errorf("Unload(\"\") = %q, want the engine's own model unloaded", got)
++	}
++
++	dmr, _ := Lookup("dmr")
++	e = &Engine{Provider: dmr, Backend: Backend{Provider: "dmr", Model: "m"}}
++	if got := e.Unload(context.Background(), ""); !strings.HasPrefix(got, "cannot unload m:") {
++		t.Errorf("dmr Unload = %q, want a refusal", got)
++	}
++
++	e = &Engine{Provider: llmman, Backend: Backend{Provider: "llmman", BaseURL: "http://127.0.0.1:1/v1", Model: "m"}}
++	if got := e.Unload(context.Background(), ""); !strings.Contains(got, "cannot be reached") {
++		t.Errorf("unreachable Unload = %q", got)
++	}
++}

added internal/tools/unload.go +69 -0 Viewed

new file mode 100644
@@ -0,0 +1,69 @@
	1	+package tools
	2	+
	3	+import (
	4	+ "context"
	5	+ "strings"
	6	+
	7	+ "mm/internal/detector"
	8	+ "mm/internal/spinner"
	9	+ "mm/internal/ui"
	10	+
	11	+ "github.com/firebase/genkit/go/ai"
	12	+ "github.com/firebase/genkit/go/genkit"
	13	+)
	14	+
	15	+// --- unload_model (ticket 101) -----------------------------------------------
	16	+//
	17	+// Frees a model from the server's memory. The model comes back by itself on
	18	+// the next request that names it, so the tool is never gated, even with
	19	+// yolo off: nothing changes on disk or in the project.
	20	+//
	21	+// Unloading the model that RUNS this agent lasts one step: the agent needs it
	22	+// to read the tool's result. The description says so, and points at the case
	23	+// that holds: the model of another agent, once its answer is in.
	24	+
	25	+type unloadModelInput struct {
	26	+ Model string `json:"model,omitempty" jsonschema_description:"The model to unload, exactly as the server names it. Empty: your own model."`
	27	+}
	28	+
	29	+// UnloadFunc unloads a model and says what happened in one line — the
	30	+// engine's Unload.
	31	+type UnloadFunc func(ctx context.Context, model string) string
	32	+
	33	+// UnloadModel declares the `unload_model` tool (config unloadModels).
	34	+//
	35	+// set.tools = append(set.tools, tools.UnloadModel(e.G, d, e.Unload))
	36	+func UnloadModel(g genkit.Genkit, d detector.LoopDetector, unload UnloadFunc) ai.ToolRef {
	37	+ description := `Unload a model from the model server's memory, to free RAM for another one. The model reloads by itself the next time something uses it.
	38	+Use it when your instructions say so, typically on the model of another agent once you have its answer and no longer need it.
	39	+Unloading your own model (model left empty) only lasts until your next step: you need it to read this result.
	40	+Returns one line: unloaded, not loaded (nothing to do), or why it could not be done.`
	41	+
	42	+ return genkit.DefineTool(g, "unload_model", description,
	43	+ func(tctx *ai.ToolContext, in unloadModelInput) (string, error) {
	44	+ title := "unload_model: " + target(in.Model)
	45	+ if s := ui.ActiveSink(); s != nil {
	46	+ ev := ui.ToolEvent{ID: ui.NextCallID(), Tool: "unload_model", Title: title, Kind: "other", Input: map[string]any{"model": in.Model}}
	47	+ s.ToolStart(ev)
	48	+ s.ToolRunning(ev.ID)
	49	+ out := unload(tctx, in.Model)
	50	+ s.ToolEnd(ev.ID, ui.ToolResult{Output: out, Failed: unloadFailed(out)})
	51	+ return record(d, "unload_model", in.Model, out)
	52	+ }
	53	+ announce("🧹 " + title)
	54	+ out := unload(tctx, in.Model)
	55	+ spinner.Suspend(func() { echoOutput(out) })
	56	+ return record(d, "unload_model", in.Model, out)
	57	+ })
	58	+}
	59	+
	60	+// target names the model in a title: "(own model)" when none is given.
	61	+func target(model string) string {
	62	+ if strings.TrimSpace(model) == "" {
	63	+ return "(own model)"
	64	+ }
	65	+ return model
	66	+}
	67	+
	68	+// unloadFailed reports whether the engine's line is a refusal.
	69	+func unloadFailed(line string) bool { return strings.HasPrefix(line, "cannot unload") }

++package tools
++
++import (
++	"context"
++	"strings"
++
++	"mm/internal/detector"
++	"mm/internal/spinner"
++	"mm/internal/ui"
++
++	"github.com/firebase/genkit/go/ai"
++	"github.com/firebase/genkit/go/genkit"
++)
++
++// --- unload_model (ticket 101) -----------------------------------------------
++//
++// Frees a model from the server's memory. The model comes back by itself on
++// the next request that names it, so the tool is never gated, even with
++// yolo off: nothing changes on disk or in the project.
++//
++// Unloading the model that RUNS this agent lasts one step: the agent needs it
++// to read the tool's result. The description says so, and points at the case
++// that holds: the model of another agent, once its answer is in.
++
++type unloadModelInput struct {
++	Model string `json:"model,omitempty" jsonschema_description:"The model to unload, exactly as the server names it. Empty: your own model."`
++}
++
++// UnloadFunc unloads a model and says what happened in one line — the
++// engine's Unload.
++type UnloadFunc func(ctx context.Context, model string) string
++
++// UnloadModel declares the `unload_model` tool (config unloadModels).
++//
++//	set.tools = append(set.tools, tools.UnloadModel(e.G, d, e.Unload))
++func UnloadModel(g *genkit.Genkit, d *detector.LoopDetector, unload UnloadFunc) ai.ToolRef {
++	description := `Unload a model from the model server's memory, to free RAM for another one. The model reloads by itself the next time something uses it.
++Use it when your instructions say so, typically on the model of another agent once you have its answer and no longer need it.
++Unloading your own model (model left empty) only lasts until your next step: you need it to read this result.
++Returns one line: unloaded, not loaded (nothing to do), or why it could not be done.`
++
++	return genkit.DefineTool(g, "unload_model", description,
++		func(tctx *ai.ToolContext, in unloadModelInput) (string, error) {
++			title := "unload_model: " + target(in.Model)
++			if s := ui.ActiveSink(); s != nil {
++				ev := ui.ToolEvent{ID: ui.NextCallID(), Tool: "unload_model", Title: title, Kind: "other", Input: map[string]any{"model": in.Model}}
++				s.ToolStart(ev)
++				s.ToolRunning(ev.ID)
++				out := unload(tctx, in.Model)
++				s.ToolEnd(ev.ID, ui.ToolResult{Output: out, Failed: unloadFailed(out)})
++				return record(d, "unload_model", in.Model, out)
++			}
++			announce("🧹 " + title)
++			out := unload(tctx, in.Model)
++			spinner.Suspend(func() { echoOutput(out) })
++			return record(d, "unload_model", in.Model, out)
++		})
++}
++
++// target names the model in a title: "(own model)" when none is given.
++func target(model string) string {
++	if strings.TrimSpace(model) == "" {
++		return "(own model)"
++	}
++	return model
++}
++
++// unloadFailed reports whether the engine's line is a refusal.
++func unloadFailed(line string) bool { return strings.HasPrefix(line, "cannot unload") }

added internal/tools/unload_test.go +62 -0 Viewed

new file mode 100644
@@ -0,0 +1,62 @@
	1	+package tools
	2	+
	3	+import (
	4	+ "context"
	5	+ "testing"
	6	+
	7	+ "mm/internal/detector"
	8	+ "mm/internal/ui"
	9	+
	10	+ "github.com/firebase/genkit/go/genkit"
	11	+)
	12	+
	13	+// unload_model hands the model to the engine as given — empty for its own —,
	14	+// shows a finished call, and never asks, even with yolo off: nothing changes
	15	+// on disk, and the model reloads by itself.
	16	+func TestUnloadModelNeverAsksAndPassesTheModelOn(t *testing.T) {
	17	+ yoloOff(t)
	18	+ g := genkit.Init(context.Background())
	19	+ var asked []string
	20	+ UnloadModel(g, detector.NewLoopDetector(10, 3), func(_ context.Context, model string) string {
	21	+ asked = append(asked, model)
	22	+ return "unloaded " + model
	23	+ })
	24	+ tool := genkit.LookupTool(g, "unload_model")
	25	+ sink := &recordingSink{allow: false}
	26	+ ui.SetActive(sink)
	27	+ t.Cleanup(func() { ui.SetActive(nil) })
	28	+
	29	+ out, err := tool.RunRaw(context.Background(), map[string]any{"model": "qwen3.5:2b"})
	30	+ if err != nil \|\| out != "unloaded qwen3.5:2b" {
	31	+ t.Fatalf("out = %v, %v", out, err)
	32	+ }
	33	+ if sink.asked != 0 {
	34	+ t.Errorf("asked %d time(s) for permission, want never", sink.asked)
	35	+ }
	36	+ if sink.started.Title != "unload_model: qwen3.5:2b" \|\| sink.ended.Failed {
	37	+ t.Errorf("shown as %+v → %+v", sink.started, sink.ended)
	38	+ }
	39	+
	40	+ if _, err := tool.RunRaw(context.Background(), map[string]any{}); err != nil {
	41	+ t.Fatal(err)
	42	+ }
	43	+ if len(asked) != 2 \|\| asked[1] != "" \|\| sink.started.Title != "unload_model: (own model)" {
	44	+ t.Errorf("own model: engine asked %q, title %q", asked, sink.started.Title)
	45	+ }
	46	+}
	47	+
	48	+func TestUnloadModelRefusalIsAFailedCall(t *testing.T) {
	49	+ g := genkit.Init(context.Background())
	50	+ UnloadModel(g, detector.NewLoopDetector(10, 3), func(context.Context, string) string {
	51	+ return "cannot unload m: the dmr provider cannot unload a model"
	52	+ })
	53	+ sink := &recordingSink{}
	54	+ ui.SetActive(sink)
	55	+ t.Cleanup(func() { ui.SetActive(nil) })
	56	+ if _, err := genkit.LookupTool(g, "unload_model").RunRaw(context.Background(), map[string]any{"model": "m"}); err != nil {
	57	+ t.Fatal(err)
	58	+ }
	59	+ if !sink.ended.Failed {
	60	+ t.Error("a refusal is not shown as a failed call")
	61	+ }
	62	+}

++package tools
++
++import (
++	"context"
++	"testing"
++
++	"mm/internal/detector"
++	"mm/internal/ui"
++
++	"github.com/firebase/genkit/go/genkit"
++)
++
++// unload_model hands the model to the engine as given — empty for its own —,
++// shows a finished call, and never asks, even with yolo off: nothing changes
++// on disk, and the model reloads by itself.
++func TestUnloadModelNeverAsksAndPassesTheModelOn(t *testing.T) {
++	yoloOff(t)
++	g := genkit.Init(context.Background())
++	var asked []string
++	UnloadModel(g, detector.NewLoopDetector(10, 3), func(_ context.Context, model string) string {
++		asked = append(asked, model)
++		return "unloaded " + model
++	})
++	tool := genkit.LookupTool(g, "unload_model")
++	sink := &recordingSink{allow: false}
++	ui.SetActive(sink)
++	t.Cleanup(func() { ui.SetActive(nil) })
++
++	out, err := tool.RunRaw(context.Background(), map[string]any{"model": "qwen3.5:2b"})
++	if err != nil || out != "unloaded qwen3.5:2b" {
++		t.Fatalf("out = %v, %v", out, err)
++	}
++	if sink.asked != 0 {
++		t.Errorf("asked %d time(s) for permission, want never", sink.asked)
++	}
++	if sink.started.Title != "unload_model: qwen3.5:2b" || sink.ended.Failed {
++		t.Errorf("shown as %+v → %+v", sink.started, sink.ended)
++	}
++
++	if _, err := tool.RunRaw(context.Background(), map[string]any{}); err != nil {
++		t.Fatal(err)
++	}
++	if len(asked) != 2 || asked[1] != "" || sink.started.Title != "unload_model: (own model)" {
++		t.Errorf("own model: engine asked %q, title %q", asked, sink.started.Title)
++	}
++}
++
++func TestUnloadModelRefusalIsAFailedCall(t *testing.T) {
++	g := genkit.Init(context.Background())
++	UnloadModel(g, detector.NewLoopDetector(10, 3), func(context.Context, string) string {
++		return "cannot unload m: the dmr provider cannot unload a model"
++	})
++	sink := &recordingSink{}
++	ui.SetActive(sink)
++	t.Cleanup(func() { ui.SetActive(nil) })
++	if _, err := genkit.LookupTool(g, "unload_model").RunRaw(context.Background(), map[string]any{"model": "m"}); err != nil {
++		t.Fatal(err)
++	}
++	if !sink.ended.Failed {
++		t.Error("a refusal is not shown as a failed call")
++	}
++}

modified internal/tui/commands.go +12 -0 Viewed

@@ -5,6 +5,7 @@
5	5	package tui
6	6
7	7	import (
	8	+ "context"
8	9	"fmt"
9	10
10	11	"mm/internal/agentsmd"
@@ -101,9 +102,20 @@ func (r *runner) dispatch(p Prompt) {
101	102	r.copy(arg)
102	103	case command.Save:
103	104	r.saveAnswer(arg)
	105	+ case command.Unload:
	106	+ r.unload(arg)
104	107	}
105	108	}
106	109
	110	+// unload runs /unload off the event loop — the server takes a few seconds to
	111	+// stop a model — and shows its one line as a notice.
	112	+func (r *runner) unload(model string) {
	113	+ go func() {
	114	+ r.t.Notice("🧹 " + r.opts.Model.Unload(context.Background(), model))
	115	+ r.wake()
	116	+ }()
	117	+}
	118	+
107	119	// help lists the window's commands and keys.
108	120	func (r *runner) help() {
109	121	r.t.Notice(command.Usage(command.Window) + "\n\nEnter sends · Ctrl-J new line · Esc cancels the turn · Ctrl-C cancels then quits (twice to leave a turn that will not stop) · Ctrl-L clears · Ctrl-Y copies the last answer · F2 mouse on/off (off: select text with the mouse as usual) · PgUp/PgDn scroll · Tab moves the focus · @path attaches a file")

  package tui
  import (
++	"context"
  	"fmt"
  	"mm/internal/agentsmd"
  		r.copy(arg)
  	case command.Save:
  		r.saveAnswer(arg)
++	case command.Unload:
++		r.unload(arg)
  	}
  }
++// unload runs /unload off the event loop — the server takes a few seconds to
++// stop a model — and shows its one line as a notice.
++func (r *runner) unload(model string) {
++	go func() {
++		r.t.Notice("🧹 " + r.opts.Model.Unload(context.Background(), model))
++		r.wake()
++	}()
++}
++
  // help lists the window's commands and keys.
  func (r *runner) help() {
  	r.t.Notice(command.Usage(command.Window) + "\n\nEnter sends · Ctrl-J new line · Esc cancels the turn · Ctrl-C cancels then quits (twice to leave a turn that will not stop) · Ctrl-L clears · Ctrl-Y copies the last answer · F2 mouse on/off (off: select text with the mouse as usual) · PgUp/PgDn scroll · Tab moves the focus · @path attaches a file")

modified internal/tui/draw.go +13 -3 Viewed

@@ -202,7 +202,7 @@ func (v *View) rightBar(leftWidth, width int) string {
202	202	}
203	203
204	204	// drawPicker paints the popup over the bottom of the conversation: one row
205		-// per choice, the highlighted one in reverse.
	205	+// per choice, the highlighted one in reverse, scrolled so that it is shown.
206	206	func (v *View) drawPicker(p painter) {
207	207	cs := v.choices()
208	208	if len(cs) == 0 {
@@ -222,10 +222,11 @@ func (v *View) drawPicker(p painter) {
222	222	width = max(width, labels+2+len([]rune(c.detail)))
223	223	}
224	224	width = min(width+4, p.width())
	225	+ first := pickerFirst(len(cs), rows, v.pick.selected)
225	226	for i := 0; i < rows; i++ {
226		- c := cs[i]
	227	+ c := cs[first+i]
227	228	style := stBody
228		- if i == v.pick.selected {
	229	+ if first+i == v.pick.selected {
229	230	style = stPick
230	231	}
231	232	text := " " + c.label + strings.Repeat(" ", max(labels-len([]rune(c.label))+2, 1)) + c.detail
@@ -233,3 +234,12 @@ func (v *View) drawPicker(p painter) {
233	234	p.textLimited(0, top+i, width, text, style)
234	235	}
235	236	}
	237	+
	238	+// pickerFirst is the first choice shown when only rows of n fit: 0 until the
	239	+// selection goes past the last row, then the window follows it — without
	240	+// it, the arrows selected choices nobody could see (/unload, the tenth
	241	+// command, was never shown).
	242	+func pickerFirst(n, rows, selected int) int {
	243	+ selected = min(selected, n-1)
	244	+ return max(0, selected-rows+1)
	245	+}

  }
  // drawPicker paints the popup over the bottom of the conversation: one row
--// per choice, the highlighted one in reverse.
++// per choice, the highlighted one in reverse, scrolled so that it is shown.
  func (v *View) drawPicker(p painter) {
  	cs := v.choices()
  	if len(cs) == 0 {
  		width = max(width, labels+2+len([]rune(c.detail)))
  	}
  	width = min(width+4, p.width())
++	first := pickerFirst(len(cs), rows, v.pick.selected)
  	for i := 0; i < rows; i++ {
--		c := cs[i]
++		c := cs[first+i]
  		style := stBody
--		if i == v.pick.selected {
++		if first+i == v.pick.selected {
  			style = stPick
  		}
  		text := "  " + c.label + strings.Repeat(" ", max(labels-len([]rune(c.label))+2, 1)) + c.detail
  		p.textLimited(0, top+i, width, text, style)
  	}
  }
++
++// pickerFirst is the first choice shown when only rows of n fit: 0 until the
++// selection goes past the last row, then the window follows it — without
++// it, the arrows selected choices nobody could see (/unload, the tenth
++// command, was never shown).
++func pickerFirst(n, rows, selected int) int {
++	selected = min(selected, n-1)
++	return max(0, selected-rows+1)
++}

modified internal/tui/pickers_test.go +18 -0 Viewed

@@ -195,3 +195,21 @@ func TestSlashIsACommandOnlyAtTheStart(t *testing.T) {
195	195	t.Errorf("the command popup opened mid-sentence: %v", labels(hs.v.choices()))
196	196	}
197	197	}
	198	+
	199	+// The popup shows pickerRows choices; once the selection goes past the last
	200	+// one, the window follows it.
	201	+func TestPickerFirstFollowsTheSelection(t *testing.T) {
	202	+ cases := []struct{ n, rows, selected, want int }{
	203	+ {3, 8, 2, 0}, // everything fits
	204	+ {10, 8, 0, 0}, // top
	205	+ {10, 8, 7, 0}, // last visible row
	206	+ {10, 8, 8, 1}, // one past: scroll by one
	207	+ {10, 8, 9, 2}, // the last choice, at the bottom
	208	+ {10, 8, 42, 2}, // a stale selection is clamped
	209	+ }
	210	+ for _, c := range cases {
	211	+ if got := pickerFirst(c.n, c.rows, c.selected); got != c.want {
	212	+ t.Errorf("pickerFirst(%d, %d, %d) = %d, want %d", c.n, c.rows, c.selected, got, c.want)
	213	+ }
	214	+ }
	215	+}

  		t.Errorf("the command popup opened mid-sentence: %v", labels(hs.v.choices()))
  	}
  }
++
++// The popup shows pickerRows choices; once the selection goes past the last
++// one, the window follows it.
++func TestPickerFirstFollowsTheSelection(t *testing.T) {
++	cases := []struct{ n, rows, selected, want int }{
++		{3, 8, 2, 0},   // everything fits
++		{10, 8, 0, 0},  // top
++		{10, 8, 7, 0},  // last visible row
++		{10, 8, 8, 1},  // one past: scroll by one
++		{10, 8, 9, 2},  // the last choice, at the bottom
++		{10, 8, 42, 2}, // a stale selection is clamped
++	}
++	for _, c := range cases {
++		if got := pickerFirst(c.n, c.rows, c.selected); got != c.want {
++			t.Errorf("pickerFirst(%d, %d, %d) = %d, want %d", c.n, c.rows, c.selected, got, c.want)
++		}
++	}
++}

modified internal/tui/turn.go +3 -0 Viewed

@@ -32,6 +32,9 @@ type Model interface {
32	32	Explain(err error) string
33	33	LastInputTokens() int
34	34	ForgetInputTokens()
	35	+ // Unload frees a model from the server's memory and says what happened
	36	+ // in one line (engine.Engine.Unload).
	37	+ Unload(ctx context.Context, model string) string
35	38	}
36	39
37	40	// Options is what Run needs from main: the same wiring as the two other

  	Explain(err error) string
  	LastInputTokens() int
  	ForgetInputTokens()
++	// Unload frees a model from the server's memory and says what happened
++	// in one line (engine.Engine.Unload).
++	Unload(ctx context.Context, model string) string
  }
  // Options is what Run needs from main: the same wiring as the two other

modified internal/tui/turn_test.go +25 -0 Viewed

@@ -10,6 +10,7 @@ import (
10	10	"testing"
11	11	"time"
12	12
	13	+ "mm/internal/command"
13	14	"mm/internal/config"
14	15	"mm/internal/detector"
15	16	mmhooks "mm/internal/hooks"
@@ -70,6 +71,12 @@ func (m model) Summarize(context.Context, []*ai.Message, int) (string, error) {
70	71	func (m model) Explain(err error) string { return "explained: " + err.Error() }
71	72	func (m model) LastInputTokens() int { return m.tokens }
72	73	func (m model) ForgetInputTokens() { m.tokens = 0 }
	74	+func (m model) Unload(_ context.Context, name string) string {
	75	+ if name == "" {
	76	+ name = "own"
	77	+ }
	78	+ return "unloaded " + name
	79	+}
73	80
74	81	func newTestRunner(t testing.T, m fakeModel) (runner, Transcript) {
75	82	t.Helper()
@@ -585,3 +592,21 @@ func TestRunnerHidesHooksWhenOff(t *testing.T) {
585	592	t.Errorf("hooks shown with displayHooks off:\n%s", notices(tr))
586	593	}
587	594	}
	595	+
	596	+// /unload runs off the event loop and shows the engine's line as a notice.
	597	+func TestRunnerUnload(t *testing.T) {
	598	+ command.SetUnload(true)
	599	+ t.Cleanup(func() { command.SetUnload(false) })
	600	+ r, tr := newTestRunner(t, &fakeModel{})
	601	+ r.submit(promptOf("/unload qwen3.5:2b"))
	602	+ deadline := time.Now().Add(5 * time.Second)
	603	+ for !strings.Contains(notices(tr), "🧹 unloaded qwen3.5:2b") {
	604	+ if time.Now().After(deadline) {
	605	+ t.Fatalf("notices:\n%s", notices(tr))
	606	+ }
	607	+ time.Sleep(5 * time.Millisecond)
	608	+ }
	609	+ if len(r.messages) != 1 {
	610	+ t.Errorf("the history changed: %d messages", len(r.messages))
	611	+ }
	612	+}

  	"testing"
  	"time"
++	"mm/internal/command"
  	"mm/internal/config"
  	"mm/internal/detector"
  	mmhooks "mm/internal/hooks"
  func (m model) Explain(err error) string { return "explained: " + err.Error() }
  func (m model) LastInputTokens() int     { return m.tokens }
  func (m model) ForgetInputTokens()       { m.tokens = 0 }
++func (m model) Unload(_ context.Context, name string) string {
++	if name == "" {
++		name = "own"
++	}
++	return "unloaded " + name
++}
  func newTestRunner(t *testing.T, m *fakeModel) (*runner, *Transcript) {
  	t.Helper()
  		t.Errorf("hooks shown with displayHooks off:\n%s", notices(tr))
  	}
  }
++
++// /unload runs off the event loop and shows the engine's line as a notice.
++func TestRunnerUnload(t *testing.T) {
++	command.SetUnload(true)
++	t.Cleanup(func() { command.SetUnload(false) })
++	r, tr := newTestRunner(t, &fakeModel{})
++	r.submit(promptOf("/unload qwen3.5:2b"))
++	deadline := time.Now().Add(5 * time.Second)
++	for !strings.Contains(notices(tr), "🧹 unloaded qwen3.5:2b") {
++		if time.Now().After(deadline) {
++			t.Fatalf("notices:\n%s", notices(tr))
++		}
++		time.Sleep(5 * time.Millisecond)
++	}
++	if len(r.messages) != 1 {
++		t.Errorf("the history changed: %d messages", len(r.messages))
++	}
++}

modified main.go +5 -0 Viewed

@@ -211,6 +211,7 @@ func main() {
211	211	// commands join the shared table before any front end reads it, and their
212	212	// processes start on the first ask and stop when this one exits.
213	213	command.SetAgents(config.AgentNames())
	214	+ command.SetUnload(config.Cfg.UnloadModels)
214	215	asker := ask.NewPool()
215	216	defer asker.Close()
216	217
@@ -393,6 +394,10 @@ func assembleTools(e engine.Engine, d detector.LoopDetector, asker tools.Agent
393	394	set.tools = append(set.tools, tools.SaveAnswer(e.G, d))
394	395	set.names = append(set.names, "save_answer")
395	396	}
	397	+ if config.Cfg.UnloadModels && config.Cfg.OutputSchema == nil {
	398	+ set.tools = append(set.tools, tools.UnloadModel(e.G, d, e.Unload))
	399	+ set.names = append(set.names, "unload_model")
	400	+ }
396	401	if config.Cfg.FetchContentTool {
397	402	// The product token "mm" is what the sites' robots.txt are read for.
398	403	fetcher := webmd.NewFetcher("mm/" + version.Short())

  	// 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())
++	command.SetUnload(config.Cfg.UnloadModels)
  	asker := ask.NewPool()
  	defer asker.Close()
  		set.tools = append(set.tools, tools.SaveAnswer(e.G, d))
  		set.names = append(set.names, "save_answer")
  	}
++	if config.Cfg.UnloadModels && config.Cfg.OutputSchema == nil {
++		set.tools = append(set.tools, tools.UnloadModel(e.G, d, e.Unload))
++		set.names = append(set.names, "unload_model")
++	}
  	if config.Cfg.FetchContentTool {
  		// The product token "mm" is what the sites' robots.txt are read for.
  		fetcher := webmd.NewFetcher("mm/" + version.Short())

Merge feature/unload-model into main