bots-garden/mini-me · 931b4ab

added .tickets/issues/0016-check-and-review-the-compress-mechanism.yaml +13 -0 Viewed

new file mode 100644
@@ -0,0 +1,13 @@
	1	+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
	2	+id: 16
	3	+title: Check and review the compress mechanism
	4	+state: open
	5	+author:
	6	+ name: k33g
	7	+ email: ph.charriere@gmail.com
	8	+createdAt: 2026-09-17T06:17:29.481Z
	9	+updatedAt: 2026-09-17T06:17:29.481Z
	10	+labels: []
	11	+body: ""
	12	+tasks: []
	13	+comments: []

++# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
++id: 16
++title: Check and review the compress mechanism
++state: open
++author:
++  name: k33g
++  email: ph.charriere@gmail.com
++createdAt: 2026-09-17T06:17:29.481Z
++updatedAt: 2026-09-17T06:17:29.481Z
++labels: []
++body: ""
++tasks: []
++comments: []

modified README.md +92 -1 Viewed

@@ -1,4 +1,95 @@
1	1	# Mini Me
2	2
3		-## ACP with Mini Me and Zed
	3	+> No, no. We don't gnaw on our kitty.
4	4
	5	+Mini Me (`mm`) is a minimalist coding agent for local LLMs, written in Go, that runs in your terminal or inside an editor such as Zed.
	6	+
	7	+## What is mini-me?
	8	+
	9	+Mini Me started as the preparation of a Tools-in-Action talk for Devoxx Belgium 2026, where I explain how an agent loop works with a single tool: `bash`. The model asks for a command, the agent runs it, the output goes back to the model, and so on until the model answers.
	10	+
	11	+From there it grew into the coding agent I wanted for local models (Docker Model Runner, llama.cpp): small enough to be read in an afternoon, and careful about what a small model does wrong:
	12	+
	13	+- hallucinations — the agent shows the real output of every command before the model comments on it, and the prompt forbids stating anything a command did not return;
	14	+- silent failures — a failing command, a refused edit or a rejected permission goes back to the model as text, so it reads why and tries something else;
	15	+- endless loops — the same action with the same result, repeated, is detected and the model is told to change approach; a watchdog catches a connection that went silent.
	16	+
	17	+What makes Mini Me a bit special is that it was designed from the start to be used from an editor through the [Agent Client Protocol (ACP)](https://agentclientprotocol.com): the same loop runs in a terminal REPL, in a full-screen terminal window, and inside Zed.
	18	+
	19	+## Install
	20	+
	21	+You need Go 1.25+ and a model server that can emit tool calls: Docker Model Runner (default) or Llamacpp.
	22	+
	23	+```bash
	24	+git clone https://git.rickub.com/bots-garden/mini-me.git
	25	+cd mini-me
	26	+./install.sh # go build, then copies mm to /usr/local/bin
	27	+mm -h
	28	+```
	29	+> `install.sh` works only on macOS and Linux
	30	+
	31	+Or download a binary from the [releases page](https://rickub.com/bots-garden/mini-me/releases) and put it on your `PATH`.
	32	+
	33	+Then pick a configuration file. Two are shipped: `agent.dmr.yaml` for Docker Model Runner and `agent.llamacpp.yaml` for llama.cpp. Copy one to `agent.yaml` in your project and adjust the `model:` and `system:` keys to taste.
	34	+
	35	+## Use
	36	+
	37	+The base command is `mm`. Run it from the directory of the project you want to work on: that is where commands run and where files are read and written.
	38	+
	39	+### 1. Terminal, line by line
	40	+
	41	+```bash
	42	+mm # reads ./agent.yaml
	43	+mm agent.dmr.yaml # or name the file
	44	+```
	45	+
	46	+Type a question at the `>` prompt. Each command the model runs is shown with its output, and every turn ends with a recap (`⚙ 2 command(s)`). Useful commands: `/help`, `/new`, `/compact`, `/quit`.
	47	+
	48	+### 2. Terminal, one window
	49	+
	50	+```bash
	51	+mm -tui
	52	+```
	53	+
	54	+The conversation above, a box to type in below. Enter sends, Ctrl-J adds a line, Esc cancels a turn, F1 lists the keys. With `permissions: ask` in the YAML, the agent asks before running a command or writing a file.
	55	+
	56	+### 3. Inside an editor (ACP)
	57	+
	58	+```bash
	59	+mm -acp
	60	+```
	61	+
	62	+Speaks the Agent Client Protocol on stdin/stdout. You never run this one by hand: the editor launches it. See the next section.
	63	+
	64	+## Use with the Zed editor
	65	+
	66	+1. Install `mm` on your `PATH` (see above).
	67	+2. In Zed, open `settings.json` and register the agent:
	68	+
	69	+ ```json
	70	+ {
	71	+ "agent_servers": {
	72	+ "mini-me": {
	73	+ "type": "custom",
	74	+ "command": "mm",
	75	+ "args": ["-acp"],
	76	+ "env": { "AGENT_CONFIG": "/absolute/path/to/agent.yaml" }
	77	+ }
	78	+ }
	79	+ }
	80	+ ```
	81	+
	82	+ `AGENT_CONFIG` must be an absolute path: Zed starts `mm` in the project directory, and a relative path would not be found. `"type": "custom"` is required.
	83	+
	84	+3. Open the Agent panel, pick mini-me, and talk to it. Every `bash`, `read_file`, `write_file` and `edit_file` call shows up as a tool call, and Zed asks you to allow it before it runs. File changes appear as real diffs.
	85	+4. Handy things: Escape cancels a turn; `/new` and `/compact` are in the slash-command menu; `@file` attaches a file for the model to read; with `sessions.enabled: true` in the YAML, Reload Agent brings the whole conversation back.
	86	+5. When something looks wrong, open `dev: open acp logs`: the banner, warnings and the agent's `[acp] …` trail go there.
	87	+
	88	+### Other ACP editors
	89	+
	90	+Anything that speaks ACP can host `mm -acp` the same way: JetBrains IDEs (IntelliJ IDEA, PyCharm, WebStorm…), Neovim (CodeCompanion, agentic.nvim, avante.nvim), Emacs (agent-shell), Visual Studio Code through third-party ACP extensions, Sublime Text and Qt Creator through their ACP plugins. The registration details differ; the agent side does not.
	91	+
	92	+## Learn more
	93	+
	94	+- [Documentation](docs/README.md) — tutorials, how-to guides, reference and explanations, in English and French.
	95	+- The `skills/` directory: Markdown procedures the agent can load with its `read_skill` tool.

  # Mini Me
--## ACP with Mini Me and Zed
++> No, no. We don't gnaw on our kitty.
++**Mini Me** (`mm`) is a minimalist coding agent for **local LLMs**, written in Go, that runs in your terminal or inside an editor such as **Zed**.
++
++## What is mini-me?
++
++Mini Me started as the preparation of a *Tools-in-Action* talk for **Devoxx Belgium 2026**, where I explain how an **agent loop** works with a single tool: `bash`. The model asks for a command, the agent runs it, the output goes back to the model, and so on until the model answers.
++
++From there it grew into the coding agent I wanted for local models (Docker Model Runner, llama.cpp): small enough to be read in an afternoon, and careful about what a small model does wrong:
++
++- **hallucinations** — the agent shows the real output of every command before the model comments on it, and the prompt forbids stating anything a command did not return;
++- **silent failures** — a failing command, a refused edit or a rejected permission goes back to the model as text, so it reads why and tries something else;
++- **endless loops** — the same action with the same result, repeated, is detected and the model is told to change approach; a watchdog catches a connection that went silent.
++
++What makes Mini Me a bit special is that it was designed from the start to be used **from an editor through the [Agent Client Protocol (ACP)](https://agentclientprotocol.com)**: the same loop runs in a terminal REPL, in a full-screen terminal window, and inside Zed.
++
++## Install
++
++You need Go 1.25+ and a model server that can emit tool calls: Docker Model Runner (default) or Llamacpp.
++
++```bash
++git clone https://git.rickub.com/bots-garden/mini-me.git
++cd mini-me
++./install.sh          # go build, then copies mm to /usr/local/bin
++mm -h
++```
++> `install.sh` works only on macOS and Linux
++
++Or download a binary from the [releases page](https://rickub.com/bots-garden/mini-me/releases) and put it on your `PATH`.
++
++Then pick a configuration file. Two are shipped: `agent.dmr.yaml` for Docker Model Runner and `agent.llamacpp.yaml` for llama.cpp. Copy one to `agent.yaml` in your project and adjust the `model:` and `system:` keys to taste.
++
++## Use
++
++The base command is `mm`. Run it **from the directory of the project you want to work on**: that is where commands run and where files are read and written.
++
++### 1. Terminal, line by line
++
++```bash
++mm                    # reads ./agent.yaml
++mm agent.dmr.yaml     # or name the file
++```
++
++Type a question at the `>` prompt. Each command the model runs is shown with its output, and every turn ends with a recap (`⚙ 2 command(s)`). Useful commands: `/help`, `/new`, `/compact`, `/quit`.
++
++### 2. Terminal, one window
++
++```bash
++mm -tui
++```
++
++The conversation above, a box to type in below. **Enter** sends, **Ctrl-J** adds a line, **Esc** cancels a turn, **F1** lists the keys. With `permissions: ask` in the YAML, the agent asks before running a command or writing a file.
++
++### 3. Inside an editor (ACP)
++
++```bash
++mm -acp
++```
++
++Speaks the Agent Client Protocol on stdin/stdout. You never run this one by hand: the editor launches it. See the next section.
++
++## Use with the Zed editor
++
++1. Install `mm` on your `PATH` (see above).
++2. In Zed, open `settings.json` and register the agent:
++
++   ```json
++   {
++     "agent_servers": {
++       "mini-me": {
++         "type": "custom",
++         "command": "mm",
++         "args": ["-acp"],
++         "env": { "AGENT_CONFIG": "/absolute/path/to/agent.yaml" }
++       }
++     }
++   }
++   ```
++
++   `AGENT_CONFIG` must be an **absolute** path: Zed starts `mm` in the project directory, and a relative path would not be found. `"type": "custom"` is required.
++
++3. Open the Agent panel, pick **mini-me**, and talk to it. Every `bash`, `read_file`, `write_file` and `edit_file` call shows up as a tool call, and Zed asks you to allow it before it runs. File changes appear as real diffs.
++4. Handy things: **Escape** cancels a turn; `/new` and `/compact` are in the slash-command menu; `@file` attaches a file for the model to read; with `sessions.enabled: true` in the YAML, **Reload Agent** brings the whole conversation back.
++5. When something looks wrong, open `dev: open acp logs`: the banner, warnings and the agent's `[acp] …` trail go there.
++
++### Other ACP editors
++
++Anything that speaks ACP can host `mm -acp` the same way: **JetBrains IDEs** (IntelliJ IDEA, PyCharm, WebStorm…), **Neovim** (CodeCompanion, agentic.nvim, avante.nvim), **Emacs** (agent-shell), **Visual Studio Code** through third-party ACP extensions, Sublime Text and Qt Creator through their ACP plugins. The registration details differ; the agent side does not.
++
++## Learn more
++
++- [Documentation](docs/README.md) — tutorials, how-to guides, reference and explanations, in English and French.
++- The `skills/` directory: Markdown procedures the agent can load with its `read_skill` tool.

modified agent.dmr.yaml +30 -101 Viewed

@@ -1,85 +1,44 @@
1	1	# Agent settings.
2	2	#
3		-# Every key is optional: what you leave out keeps its built-in default
4		-# (see internal/config/config.go). Nothing here requires recompiling —
5		-# `go run .` picks the file up at startup.
6		-#
7		-# Use another file with: AGENT_CONFIG=./fast.yaml go run .
8		-
9		-# Which server is behind the URL. Left out, it means `dmr`: every agent.yaml
10		-# written before this key existed keeps working unchanged. `llamacpp` speaks
11		-# the same OpenAI-compatible protocol, but with its own port, its own way of
12		-# refusing tool calls (no --jinja) and its own words in error messages — the
13		-# provider is what turns those into one useful line. See agent.llamacpp.yaml.
14		-# AGENT_PROVIDER overrides this key for one run.
15	3	provider: dmr
16	4
17		-# The chat model, in the provider's own naming (DMR: `ai/…`, `hf.co/…`). It
18		-# MUST be able to emit tool_calls, otherwise the agent never runs a single
19		-# command. AGENT_MODEL overrides it.
20	5	model: huggingface.co/jetbrains/mellum2-12b-a2.5b-instruct-gguf-q4_k_m:Q4_K_M
	6	+baseUrl: http://localhost:12434/engines/v1
	7	+fallback: http://host.docker.internal:12434/engines/v1
	8	+
21	9
22		-# Endpoint. `baseUrl` is tried first; if it does not answer, `fallback` is used
23		-# — handy from a container or a sandbox, where the host is reachable as
24		-# host.docker.internal. Both are LEFT OUT here on purpose: each provider has its
25		-# defaults, and an explicit value is always respected — so with DMR's URL
26		-# written in this file, `AGENT_PROVIDER=llamacpp ./bob agent.yaml` still talked
27		-# to port 12434. Leaving them out keeps the file provider-neutral. These are the
28		-# values the `dmr` provider fills in:
29		-# baseUrl: http://localhost:12434/engines/v1
30		-# fallback: http://host.docker.internal:12434/engines/v1
31		-# AGENT_BASE_URL — or, for dmr, the older DMR_BASE_URL — overrides both without
32		-# probing. `fallback: ""` (an explicit empty string) disables the fallback.
33		-
34		-# The NAME of the environment variable holding the API key — never the key
35		-# itself, so this file can be committed and shown on screen. DMR ignores the
36		-# key, so nothing is needed here; llama-server checks one only when started
37		-# with --api-key (default variable: LLAMA_API_KEY).
38		-# apiKeyEnv: LLAMA_API_KEY
39		-
40		-# How many tokens the server actually serves. The model does not know this:
41		-# it is what the operator started the server with (`llama-server -c 32768`,
42		-# `docker model configure`). Shown at start-up next to its origin, because a
43		-# number without an origin never gets corrected. 0 = ask the server
44		-# (llama-server tells its n_ctx on /props), else "unknown".
	10	+# How many tokens the server actually serves.
	11	+# 0 = ask the server
45	12	contextWindow: 0
46	13
47		-# Max number of characters a tool returns to the model. Context safeguard:
48		-# beyond that, the output is truncated (beginning + end kept).
	14	+# Max number of characters a tool returns to the model.
	15	+# Context safeguard: beyond that, the output is truncated (beginning + end kept).
49	16	maxOutput: 16000
50	17
51		-# Max number of model ↔ tools round trips for a single question. Raise it for
52		-# tasks that need long chains of commands.
	18	+# Max number of model <-> tools round trips for a single question.
	19	+# Raise it for tasks that need long chains of commands.
53	20	maxTurns: 40
54	21
55		-# The built-in file tools: read_file, write_file, edit_file. This is the switch
56		-# this part exists for. `true`: the model edits files through tools it can SEE
57		-# in its tool list. `false`: the agent is part 09 again — bash and read_skill —
58		-# and edits files through the `edit` CLI if it is on the PATH. Same binary,
59		-# same prompts, two set-ups: measured side by side, that is the comparison
60		-# (in part 07, `read_skill` as a tool was loaded 3/3 times where a catalogue
61		-# in the prompt plus `cat` managed 1/11). Paths are relative to the current
62		-# directory and are not confined to it.
	22	+# The `bash` tool.
	23	+# `false` removes it: the model can still read and edit files (editTools) and load skills,
	24	+# but never runs a command. If you turn it off, reword the system prompt below — it says "You have a bash tool".
	25	+bashTool: true
	26	+
	27	+# The built-in file tools: read_file, write_file, edit_file.
63	28	editTools: true
64	29
65		-# Where the `read_skill` tool looks for its markdown procedures. The path is
66		-# relative to the CURRENT DIRECTORY, so run the agent from the directory that
67		-# holds skills/ — not from src/. When the directory has no *.md file, the tool
68		-# is not declared at all.
	30	+# Where the `read_skill` tool looks for its markdown procedures.
	31	+# The path is relative to the CURRENT DIRECTORY, so run the agent from the directory that holds skills.
69	32	skillsDir: skills
70	33
71		-# Persistent sessions, per project. When enabled, every conversation is saved
72		-# as one JSON file under `dir` (relative to the directory the agent is started
73		-# from; the editor's project in ACP mode), rewritten after each turn. In the
74		-# terminal, `/sessions` lists them and `/resume` continues the last one; in an
75		-# editor, "Reload Agent" (ACP session/load) replays one. `/new` always starts
76		-# a fresh session and file. Add `.mm/` to the project's .gitignore.
	34	+# Persistent sessions, per project.
	35	+# When enabled, every conversation is saved as one JSON file under `dir`
77	36	sessions:
78	37	enabled: false
79	38	dir: .mm/sessions
80	39
81		-# How many lines of a command's output are echoed to the terminal. "Show me
82		-# that file" means show it: running `cat` is not displaying it. 0 disables it.
	40	+# How many lines of a command's output are echoed to the terminal.
	41	+# 0 disables it.
83	42	previewLines: 20
84	43
85	44	# The system prompt: what the agent is, and what it is allowed to do.
@@ -155,64 +114,34 @@ system: \|
155	114	# we want precise and reproducible answers, not creativity.
156	115	sampling:
157	116	temperature: 0.0
158		- # Une seule commande par tour. Sans ça, le modèle peut en demander
159		- # plusieurs d'un coup : Genkit les exécute EN PARALLÈLE, dans le même
160		- # répertoire, et on ne sait plus quelle sortie appartient à quelle
161		- # commande. Pour le détecteur de boucle, c'est pire : deux appels du même
162		- # tour s'enregistrent dans un ordre indéterminé, et « la même action deux
163		- # fois de suite » perd son sens. Mettre `true` pour retrouver le
164		- # comportement par défaut de l'API.
165	117	parallel_tool_calls: false
166	118	top_p: 0.9
167	119	max_tokens: 4096
168	120
169	121	watchdogTimeout: 30s
170	122
171		-# Context compression, carried over from part 08 (its CONTEXT_WINDOW.md has the
172		-# reasoning). OFF by default: with `enabled: false` the agent behaves exactly
173		-# as before. Why it exists: nothing in the agent ever shortens the history —
174		-# measured with the fake engine of part 03, the messages sent to the model grew
175		-# 2 → 5 → 7 → 9 over four requests, and every `bash` output (up to maxOutput
176		-# characters) and every skill read stays until /quit. On a local model the
177		-# window is fixed at load time and the prefill runs on the laptop, so a long
178		-# session first shows up as a false "[watchdog: connection hang]", then as an
179		-# error from the server.
180		-#
181		-# The window it measures against is the top-level `contextWindow` above — ONE
182		-# key for the banner and for the trigger, so the two cannot disagree. When
183		-# that key is 0, the value the provider's probe learned from the server is
184		-# used (llama-server tells its n_ctx on /props; DMR tells nothing). When both
185		-# are unknown, only `maxMessages` can trigger, and the agent says so at
186		-# start-up.
	123	+# Context compression
187	124	context:
188	125	enabled: false
189	126
190		- # Compress when the history reaches this share of the window. 75 leaves a
191		- # quarter for the next question, the outputs of its turns and the answer —
192		- # a single `bash` output can be maxOutput characters, about 4-5k tokens.
	127	+ # Compress when the history reaches this share of the window.
193	128	threshold: 75
194	129
195		- # Fallback on the message count, for when the window is unknown or the
196		- # estimate is off. One command costs 2 messages (call + response), so 80 is
197		- # roughly 30 commands of history.
	130	+ # Fallback on the message count, for when the window is unknown.
	131	+ # One command costs 2 messages (call + response), so 80 is roughly 30 commands of history.
198	132	maxMessages: 80
199	133
200		- # Question turns kept raw at the end. The recent turns are where the model
201		- # works; summarising them makes it re-run what it just did. Three covers
202		- # "do X", "now fix it", "and test it" — the usual shape of a demo.
	134	+ # How many recent questions (with their commands and answers) are kept as-is instead of being summarised.
	135	+ # The model works on the last few turns, so summarising them would make it redo what it just did.
203	136	keepLastTurns: 3
204	137
205		- # max_tokens of the summary request. Seven sections of one line per item
206		- # fit in far less; the cap stops a runaway model from filling the window it
207		- # was asked to empty.
	138	+ # max_tokens of the summary request.
208	139	summaryMaxTokens: 1200
209	140
210		- # Replaces the built-in summary prompt (08-context-compression/CONTEXT_WINDOW.md
211		- # § 4.1) when set. Empty = built-in.
	141	+ # Replaces the built-in summary prompt when set.
	142	+ # Empty = built-in.
212	143	prompt: ""
213	144
214		- # One line after each compression: what was replaced, what was kept,
215		- # tokens before → after. Worth showing on a screen, noise in a log.
216	145	showStats: true
217	146
218	147	displayCommands: true

  # Agent settings.
  #
--# Every key is optional: what you leave out keeps its built-in default
--# (see internal/config/config.go). Nothing here requires recompiling —
--# `go run .` picks the file up at startup.
--#
--# Use another file with:  AGENT_CONFIG=./fast.yaml go run .
--
--# Which server is behind the URL. Left out, it means `dmr`: every agent.yaml
--# written before this key existed keeps working unchanged. `llamacpp` speaks
--# the same OpenAI-compatible protocol, but with its own port, its own way of
--# refusing tool calls (no --jinja) and its own words in error messages — the
--# provider is what turns those into one useful line. See agent.llamacpp.yaml.
--# AGENT_PROVIDER overrides this key for one run.
  provider: dmr
--# The chat model, in the provider's own naming (DMR: `ai/…`, `hf.co/…`). It
--# MUST be able to emit tool_calls, otherwise the agent never runs a single
--# command. AGENT_MODEL overrides it.
  model: huggingface.co/jetbrains/mellum2-12b-a2.5b-instruct-gguf-q4_k_m:Q4_K_M
++baseUrl: http://localhost:12434/engines/v1
++fallback: http://host.docker.internal:12434/engines/v1
++
--# Endpoint. `baseUrl` is tried first; if it does not answer, `fallback` is used
++# How many tokens the server actually serves.
--# — handy from a container or a sandbox, where the host is reachable as
++# 0 = ask the server
--# host.docker.internal. Both are LEFT OUT here on purpose: each provider has its
--# defaults, and an explicit value is always respected — so with DMR's URL
--# written in this file, `AGENT_PROVIDER=llamacpp ./bob agent.yaml` still talked
--# to port 12434. Leaving them out keeps the file provider-neutral. These are the
--# values the `dmr` provider fills in:
--#   baseUrl: http://localhost:12434/engines/v1
--#   fallback: http://host.docker.internal:12434/engines/v1
--# AGENT_BASE_URL — or, for dmr, the older DMR_BASE_URL — overrides both without
--# probing. `fallback: ""` (an explicit empty string) disables the fallback.
--
--# The NAME of the environment variable holding the API key — never the key
--# itself, so this file can be committed and shown on screen. DMR ignores the
--# key, so nothing is needed here; llama-server checks one only when started
--# with --api-key (default variable: LLAMA_API_KEY).
--# apiKeyEnv: LLAMA_API_KEY
--
--# How many tokens the server actually serves. The model does not know this:
--# it is what the operator started the server with (`llama-server -c 32768`,
--# `docker model configure`). Shown at start-up next to its origin, because a
--# number without an origin never gets corrected. 0 = ask the server
--# (llama-server tells its n_ctx on /props), else "unknown".
  contextWindow: 0
--# Max number of characters a tool returns to the model. Context safeguard:
++# Max number of characters a tool returns to the model.
--# beyond that, the output is truncated (beginning + end kept).
++# Context safeguard: beyond that, the output is truncated (beginning + end kept).
  maxOutput: 16000
--# Max number of model ↔ tools round trips for a single question. Raise it for
++# Max number of model <-> tools round trips for a single question.
--# tasks that need long chains of commands.
++# Raise it for tasks that need long chains of commands.
  maxTurns: 40
--# The built-in file tools: read_file, write_file, edit_file. This is the switch
++# The `bash` tool.
--# this part exists for. `true`: the model edits files through tools it can SEE
++# `false` removes it: the model can still read and edit files (editTools) and load skills,
--# in its tool list. `false`: the agent is part 09 again — bash and read_skill —
++# but never runs a command. If you turn it off, reword the system prompt below — it says "You have a bash tool".
--# and edits files through the `edit` CLI if it is on the PATH. Same binary,
++bashTool: true
--# same prompts, two set-ups: measured side by side, that is the comparison
++
--# (in part 07, `read_skill` as a tool was loaded 3/3 times where a catalogue
++# The built-in file tools: read_file, write_file, edit_file.
--# in the prompt plus `cat` managed 1/11). Paths are relative to the current
--# directory and are not confined to it.
  editTools: true
--# Where the `read_skill` tool looks for its markdown procedures. The path is
++# Where the `read_skill` tool looks for its markdown procedures.
--# relative to the CURRENT DIRECTORY, so run the agent from the directory that
++# The path is relative to the CURRENT DIRECTORY, so run the agent from the directory that holds skills.
--# holds skills/ — not from src/. When the directory has no *.md file, the tool
--# is not declared at all.
  skillsDir: skills
--# Persistent sessions, per project. When enabled, every conversation is saved
++# Persistent sessions, per project.
--# as one JSON file under `dir` (relative to the directory the agent is started
++# When enabled, every conversation is saved as one JSON file under `dir`
--# from; the editor's project in ACP mode), rewritten after each turn. In the
--# terminal, `/sessions` lists them and `/resume` continues the last one; in an
--# editor, "Reload Agent" (ACP session/load) replays one. `/new` always starts
--# a fresh session and file. Add `.mm/` to the project's .gitignore.
  sessions:
    enabled: false
    dir: .mm/sessions
--# How many lines of a command's output are echoed to the terminal. "Show me
++# How many lines of a command's output are echoed to the terminal.
--# that file" means show it: running `cat` is not displaying it. 0 disables it.
++# 0 disables it.
  previewLines: 20
  # The system prompt: what the agent is, and what it is allowed to do.
  # we want precise and reproducible answers, not creativity.
  sampling:
    temperature: 0.0
--  # Une seule commande par tour. Sans ça, le modèle peut en demander
--  # plusieurs d'un coup : Genkit les exécute EN PARALLÈLE, dans le même
--  # répertoire, et on ne sait plus quelle sortie appartient à quelle
--  # commande. Pour le détecteur de boucle, c'est pire : deux appels du même
--  # tour s'enregistrent dans un ordre indéterminé, et « la même action deux
--  # fois de suite » perd son sens. Mettre `true` pour retrouver le
--  # comportement par défaut de l'API.
    parallel_tool_calls: false
    top_p: 0.9
    max_tokens: 4096
  watchdogTimeout: 30s
--# Context compression, carried over from part 08 (its CONTEXT_WINDOW.md has the
++# Context compression
--# reasoning). OFF by default: with `enabled: false` the agent behaves exactly
--# as before. Why it exists: nothing in the agent ever shortens the history —
--# measured with the fake engine of part 03, the messages sent to the model grew
--# 2 → 5 → 7 → 9 over four requests, and every `bash` output (up to maxOutput
--# characters) and every skill read stays until /quit. On a local model the
--# window is fixed at load time and the prefill runs on the laptop, so a long
--# session first shows up as a false "[watchdog: connection hang]", then as an
--# error from the server.
--#
--# The window it measures against is the top-level `contextWindow` above — ONE
--# key for the banner and for the trigger, so the two cannot disagree. When
--# that key is 0, the value the provider's probe learned from the server is
--# used (llama-server tells its n_ctx on /props; DMR tells nothing). When both
--# are unknown, only `maxMessages` can trigger, and the agent says so at
--# start-up.
  context:
    enabled: false
--  # Compress when the history reaches this share of the window. 75 leaves a
++  # Compress when the history reaches this share of the window.
--  # quarter for the next question, the outputs of its turns and the answer —
--  # a single `bash` output can be maxOutput characters, about 4-5k tokens.
    threshold: 75
--  # Fallback on the message count, for when the window is unknown or the
++  # Fallback on the message count, for when the window is unknown.
--  # estimate is off. One command costs 2 messages (call + response), so 80 is
++  # One command costs 2 messages (call + response), so 80 is roughly 30 commands of history.
--  # roughly 30 commands of history.
    maxMessages: 80
--  # Question turns kept raw at the end. The recent turns are where the model
++  # How many recent questions (with their commands and answers) are kept as-is instead of being summarised.
--  # works; summarising them makes it re-run what it just did. Three covers
++  # The model works on the last few turns, so summarising them would make it redo what it just did.
--  # "do X", "now fix it", "and test it" — the usual shape of a demo.
    keepLastTurns: 3
--  # max_tokens of the summary request. Seven sections of one line per item
++  # max_tokens of the summary request.
--  # fit in far less; the cap stops a runaway model from filling the window it
--  # was asked to empty.
    summaryMaxTokens: 1200
--  # Replaces the built-in summary prompt (08-context-compression/CONTEXT_WINDOW.md
++  # Replaces the built-in summary prompt when set.
--  # § 4.1) when set. Empty = built-in.
++  # Empty = built-in.
    prompt: ""
--  # One line after each compression: what was replaced, what was kept,
--  # tokens before → after. Worth showing on a screen, noise in a log.
    showStats: true
  displayCommands: true

modified agent.llamacpp.yaml +44 -73 Viewed

@@ -1,91 +1,59 @@
1		-# Same agent, served by llama.cpp's `llama-server` instead of Docker Model Runner.
	1	+# Agent settings — same agent, served by llama.cpp's `llama-server` instead of Docker Model Runner.
2	2	#
3		-# llama-server -hf poolside/Laguna-XS-2.1-GGUF:Q4_K_M \
4		-# -a poolside/Laguna-XS-2.1-GGUF:Q4_K_M --jinja -c 32768 --port 8080
5		-# ./bob agent.llamacpp.yaml # from demo/, so that skills/ is found
	3	+# llama serve -hf JetBrains/Mellum2-12B-A2.5B-Instruct-GGUF-Q4_K_M:Q4_K_M
6	4	#
7		-# `-hf <user>/<repo>:<quant>` downloads the GGUF from Hugging Face (the quant tag
8		-# is optional and defaults to Q4_K_M; here the repo has exactly one such file,
9		-# Laguna-XS-2.1-Q4_K_M.gguf, 20.3 GB). `-a` gives the served model the SAME name
10		-# as the `model:` key below: llama-server routes requests on the `model` field,
11		-# and without an alias it exposes the file's name on /v1/models, not the repo's.
12		-# `--jinja` is not optional: without it llama-server refuses the `tools`
13		-# parameter, and this agent is nothing but tool calls. The agent recognises that
14		-# refusal and says so in one line instead of printing the server's stack.
15		-# `-c` is the context the server SERVES (the model accepts up to 262,144); it is
16		-# what shows up under `ctx:` at start-up, read from GET /props.
17		-# Laguna XS 2.1 is a 33B MoE with 3B active parameters, built for coding; its
18		-# card says the chat template does tool calling under --jinja, and that it
19		-# needs a recent llama.cpp (the card points at PR #25165 — check it is merged
20		-# in your build, or build that branch).
21		-
22	5	provider: llamacpp
23	6
24		-# Must match what GET /v1/models returns — the `-a` alias above. llama-server
25		-# routes on this field; a name it does not serve is an error, not a fallback.
26		-#model: poolside/Laguna-XS-2.1-GGUF:Q4_K_M
27		-#model: unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_M
28	7	model: jetbrains/mellum2-12b-a2.5b-instruct-gguf-q4_k_m:Q4_K_M
29	8
30	9	# llama-server's default is 127.0.0.1:8080; the OpenAI routes sit under /v1.
31		-# Leave it out to get exactly this value.
32	10	baseUrl: http://127.0.0.1:8080/v1
33	11
34		-# No fallback here: llama-server is wherever you started it.
	12	+# No fallback: llama-server is wherever you started it.
35	13	fallback: ""
36	14
37	15	# Only when llama-server was started with --api-key; unset otherwise.
38	16	# apiKeyEnv: LLAMA_API_KEY
39	17
40		-# The built-in file tools: read_file, write_file, edit_file. This is the switch
41		-# this part exists for. `true`: the model edits files through tools it can SEE
42		-# in its tool list. `false`: the agent is part 09 again — bash and read_skill —
43		-# and edits files through the `edit` CLI if it is on the PATH. Same binary,
44		-# same prompts, two set-ups: measured side by side, that is the comparison
45		-# (in part 07, `read_skill` as a tool was loaded 3/3 times where a catalogue
46		-# in the prompt plus `cat` managed 1/11). Paths are relative to the current
47		-# directory and are not confined to it.
	18	+# The `bash` tool.
	19	+# `false` removes it: the model can still read and edit files (editTools) and load skills,
	20	+# but never runs a command. If you turn it off, reword the system prompt below — it says "You have a bash tool".
	21	+bashTool: true
	22	+
	23	+# The built-in file tools: read_file, write_file, edit_file.
48	24	editTools: true
49	25
50		-# 0 = read the served size from /props. Set it when the server hides /props
51		-# (a reverse proxy) or when you know better.
	26	+# How many tokens the server actually serves.
	27	+# 0 = ask the server (llama-server tells it on /props)
52	28	contextWindow: 0
53	29
	30	+# Max number of characters a tool returns to the model.
	31	+# Context safeguard: beyond that, the output is truncated (beginning + end kept).
54	32	maxOutput: 16000
	33	+
	34	+# Max number of model <-> tools round trips for a single question.
	35	+# Raise it for tasks that need long chains of commands.
55	36	maxTurns: 40
	37	+
	38	+# Where the `read_skill` tool looks for its markdown procedures.
	39	+# The path is relative to the CURRENT DIRECTORY, so run the agent from the directory that holds skills.
56	40	skillsDir: skills
57	41
58		-# Persistent sessions, per project — OFF by default.
59		-#
60		-# How to turn them on:
61		-#
62		-# 1. Set `enabled: true` below (keep `dir` unless you want another place).
63		-# 2. Restart the agent from the project directory:
64		-# mm agent.llamacpp.yaml
65		-# or, from Zed, point AGENT_CONFIG at this file (absolute path) and use
66		-# "Reload Agent" / open a new thread so the new process reads the key.
67		-# 3. Check the banner: it now shows a `sessions:` line with the directory
68		-# and the number of earlier sessions found there.
69		-# 4. Add `.mm/` to the project's .gitignore (already done in this repository).
70		-#
71		-# What it does once on: every conversation is saved as one JSON file
72		-# `<dir>/<id>.json`, rewritten after each turn, `/compact` and `/new`. `dir`
73		-# is relative to the directory the agent is started from — the editor's
74		-# project in ACP mode — so each project keeps its own sessions; an absolute
75		-# path works too. A start always opens a NEW session: in the terminal,
76		-# `/sessions` lists the stored ones and `/resume` continues the latest
77		-# (`/resume <id>` a given one); in an editor, "Reload Agent" (ACP
78		-# session/load) replays the whole thread, tool calls included, without
79		-# running anything again. `/new` always starts a fresh session and file.
80		-# Write failures are logged, never fatal; with `enabled: false` nothing is
81		-# written and `/sessions` says so.
	42	+# Persistent sessions, per project.
	43	+# When enabled, every conversation is saved as one JSON file under `dir`.
	44	+# `/sessions` lists them, `/resume` continues one; in an editor, "Reload Agent" replays the thread.
82	45	sessions:
83	46	enabled: true
84	47	dir: .mm/sessions
85	48
	49	+# How many lines of a command's output are echoed to the terminal.
	50	+# 0 disables it.
86	51	previewLines: 20
	52	+
87	53	displayCommands: true
88	54
	55	+# The system prompt: what the agent is, and what it is allowed to do.
	56	+# This is the knob to play with — it is the shortest path to changing behaviour.
89	57	system: \|
90	58	Your name is Bob.
91	59	You are a coding agent working in a terminal.
@@ -153,32 +121,35 @@ system: \|
153	121	rule you cannot guess. Stop every job you started before you finish, and say
154	122	which ones you left running.
155	123
156		-# Same sampling as the DMR file. llama-server honours `parallel_tool_calls`
157		-# (off by default on its side too) and `max_tokens`.
	124	+# Generation settings (OpenAI API keys). Low temperature for a coding agent:
	125	+# we want precise and reproducible answers, not creativity.
158	126	sampling:
159	127	temperature: 0.0
160	128	parallel_tool_calls: false
161	129	top_p: 0.9
162	130	max_tokens: 4096
163	131
164		-# llama-server processes the whole prompt before the first token; on a laptop a
165		-# 32k context can take a while. Same watchdog as DMR, raise it if it fires.
	132	+# llama-server processes the whole prompt before the first token;
	133	+# on a laptop a 32k context can take a while. Raise it if it fires.
166	134	watchdogTimeout: 30s
167	135
168		-# Context compression (from part 08). ON here, because this is the one set-up
169		-# where the agent knows the window without being told: `contextWindow: 0`
170		-# above means "read n_ctx from /props", and that number is what `threshold`
171		-# applies to — so the banner's `ctx: 32768 (/props)` and the trigger agree by
172		-# construction. With a 32k window and a 33B MoE, the whole history is
173		-# re-read at every turn; compressing at 75 % keeps the prefill — and the
174		-# watchdog — inside the 30 s above. Set `enabled: false` to get the exact
175		-# part-07 behaviour back; `/compact` still works.
	136	+# Context compression.
	137	+# ON here: the agent knows the window (read from /props), so `threshold` has a number to apply to.
176	138	context:
177	139	enabled: true
	140	+
	141	+ # Compress when the history reaches this share of the window.
178	142	threshold: 75
179		- # Kept as a net for a reverse proxy that hides /props: then the window is
180		- # unknown and only this can trigger.
	143	+
	144	+ # Fallback on the message count, for when the window is unknown.
	145	+ # One command costs 2 messages (call + response), so 80 is roughly 30 commands of history.
181	146	maxMessages: 80
	147	+
	148	+ # How many recent questions (with their commands and answers) are kept as-is instead of being summarised.
	149	+ # The model works on the last few turns, so summarising them would make it redo what it just did.
182	150	keepLastTurns: 3
	151	+
	152	+ # max_tokens of the summary request.
183	153	summaryMaxTokens: 1200
	154	+
184	155	showStats: true

--# Same agent, served by llama.cpp's `llama-server` instead of Docker Model Runner.
++# Agent settings — same agent, served by llama.cpp's `llama-server` instead of Docker Model Runner.
  #
--#   llama-server -hf poolside/Laguna-XS-2.1-GGUF:Q4_K_M \
++#   llama serve -hf JetBrains/Mellum2-12B-A2.5B-Instruct-GGUF-Q4_K_M:Q4_K_M
--#                -a poolside/Laguna-XS-2.1-GGUF:Q4_K_M --jinja -c 32768 --port 8080
--#   ./bob agent.llamacpp.yaml          # from demo/, so that skills/ is found
  #
--# `-hf <user>/<repo>:<quant>` downloads the GGUF from Hugging Face (the quant tag
--# is optional and defaults to Q4_K_M; here the repo has exactly one such file,
--# Laguna-XS-2.1-Q4_K_M.gguf, 20.3 GB). `-a` gives the served model the SAME name
--# as the `model:` key below: llama-server routes requests on the `model` field,
--# and without an alias it exposes the file's name on /v1/models, not the repo's.
--# `--jinja` is not optional: without it llama-server refuses the `tools`
--# parameter, and this agent is nothing but tool calls. The agent recognises that
--# refusal and says so in one line instead of printing the server's stack.
--# `-c` is the context the server SERVES (the model accepts up to 262,144); it is
--# what shows up under `ctx:` at start-up, read from GET /props.
--# Laguna XS 2.1 is a 33B MoE with 3B active parameters, built for coding; its
--# card says the chat template does tool calling under --jinja, and that it
--# needs a recent llama.cpp (the card points at PR #25165 — check it is merged
--# in your build, or build that branch).
--
  provider: llamacpp
--# Must match what GET /v1/models returns — the `-a` alias above. llama-server
--# routes on this field; a name it does not serve is an error, not a fallback.
--#model: poolside/Laguna-XS-2.1-GGUF:Q4_K_M
--#model: unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_M
  model: jetbrains/mellum2-12b-a2.5b-instruct-gguf-q4_k_m:Q4_K_M
  # llama-server's default is 127.0.0.1:8080; the OpenAI routes sit under /v1.
--# Leave it out to get exactly this value.
  baseUrl: http://127.0.0.1:8080/v1
--# No fallback here: llama-server is wherever you started it.
++# No fallback: llama-server is wherever you started it.
  fallback: ""
  # Only when llama-server was started with --api-key; unset otherwise.
  # apiKeyEnv: LLAMA_API_KEY
--# The built-in file tools: read_file, write_file, edit_file. This is the switch
++# The `bash` tool.
--# this part exists for. `true`: the model edits files through tools it can SEE
++# `false` removes it: the model can still read and edit files (editTools) and load skills,
--# in its tool list. `false`: the agent is part 09 again — bash and read_skill —
++# but never runs a command. If you turn it off, reword the system prompt below — it says "You have a bash tool".
--# and edits files through the `edit` CLI if it is on the PATH. Same binary,
++bashTool: true
--# same prompts, two set-ups: measured side by side, that is the comparison
++
--# (in part 07, `read_skill` as a tool was loaded 3/3 times where a catalogue
++# The built-in file tools: read_file, write_file, edit_file.
--# in the prompt plus `cat` managed 1/11). Paths are relative to the current
--# directory and are not confined to it.
  editTools: true
--# 0 = read the served size from /props. Set it when the server hides /props
++# How many tokens the server actually serves.
--# (a reverse proxy) or when you know better.
++# 0 = ask the server (llama-server tells it on /props)
  contextWindow: 0
++# Max number of characters a tool returns to the model.
++# Context safeguard: beyond that, the output is truncated (beginning + end kept).
  maxOutput: 16000
++
++# Max number of model <-> tools round trips for a single question.
++# Raise it for tasks that need long chains of commands.
  maxTurns: 40
++
++# Where the `read_skill` tool looks for its markdown procedures.
++# The path is relative to the CURRENT DIRECTORY, so run the agent from the directory that holds skills.
  skillsDir: skills
--# Persistent sessions, per project — OFF by default.
++# Persistent sessions, per project.
--#
++# When enabled, every conversation is saved as one JSON file under `dir`.
--# How to turn them on:
++# `/sessions` lists them, `/resume` continues one; in an editor, "Reload Agent" replays the thread.
--#
--#   1. Set `enabled: true` below (keep `dir` unless you want another place).
--#   2. Restart the agent from the project directory:
--#        mm agent.llamacpp.yaml
--#      or, from Zed, point AGENT_CONFIG at this file (absolute path) and use
--#      "Reload Agent" / open a new thread so the new process reads the key.
--#   3. Check the banner: it now shows a `sessions:` line with the directory
--#      and the number of earlier sessions found there.
--#   4. Add `.mm/` to the project's .gitignore (already done in this repository).
--#
--# What it does once on: every conversation is saved as one JSON file
--# `<dir>/<id>.json`, rewritten after each turn, `/compact` and `/new`. `dir`
--# is relative to the directory the agent is started from — the editor's
--# project in ACP mode — so each project keeps its own sessions; an absolute
--# path works too. A start always opens a NEW session: in the terminal,
--# `/sessions` lists the stored ones and `/resume` continues the latest
--# (`/resume <id>` a given one); in an editor, "Reload Agent" (ACP
--# session/load) replays the whole thread, tool calls included, without
--# running anything again. `/new` always starts a fresh session and file.
--# Write failures are logged, never fatal; with `enabled: false` nothing is
--# written and `/sessions` says so.
  sessions:
    enabled: true
    dir: .mm/sessions
++# How many lines of a command's output are echoed to the terminal.
++# 0 disables it.
  previewLines: 20
++
  displayCommands: true
++# The system prompt: what the agent is, and what it is allowed to do.
++# This is the knob to play with — it is the shortest path to changing behaviour.
  system: |
    Your name is Bob.
    You are a coding agent working in a terminal.
    rule you cannot guess. Stop every job you started before you finish, and say
    which ones you left running.
--# Same sampling as the DMR file. llama-server honours `parallel_tool_calls`
++# Generation settings (OpenAI API keys). Low temperature for a coding agent:
--# (off by default on its side too) and `max_tokens`.
++# we want precise and reproducible answers, not creativity.
  sampling:
    temperature: 0.0
    parallel_tool_calls: false
    top_p: 0.9
    max_tokens: 4096
--# llama-server processes the whole prompt before the first token; on a laptop a
++# llama-server processes the whole prompt before the first token;
--# 32k context can take a while. Same watchdog as DMR, raise it if it fires.
++# on a laptop a 32k context can take a while. Raise it if it fires.
  watchdogTimeout: 30s
--# Context compression (from part 08). ON here, because this is the one set-up
++# Context compression.
--# where the agent knows the window without being told: `contextWindow: 0`
++# ON here: the agent knows the window (read from /props), so `threshold` has a number to apply to.
--# above means "read n_ctx from /props", and that number is what `threshold`
--# applies to — so the banner's `ctx: 32768 (/props)` and the trigger agree by
--# construction. With a 32k window and a 33B MoE, the whole history is
--# re-read at every turn; compressing at 75 % keeps the prefill — and the
--# watchdog — inside the 30 s above. Set `enabled: false` to get the exact
--# part-07 behaviour back; `/compact` still works.
  context:
    enabled: true
++
++  # Compress when the history reaches this share of the window.
    threshold: 75
--  # Kept as a net for a reverse proxy that hides /props: then the window is
++
--  # unknown and only this can trigger.
++  # Fallback on the message count, for when the window is unknown.
++  # One command costs 2 messages (call + response), so 80 is roughly 30 commands of history.
    maxMessages: 80
++
++  # How many recent questions (with their commands and answers) are kept as-is instead of being summarised.
++  # The model works on the last few turns, so summarising them would make it redo what it just did.
    keepLastTurns: 3
++
++  # max_tokens of the summary request.
    summaryMaxTokens: 1200
++
    showStats: true

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

@@ -43,7 +43,7 @@ Printed on stdout in terminal mode, on stderr in ACP mode. Dimmed when stdout is
43	43
44	44	\| Line \| Content \|
45	45	\|------\|---------\|
46		-\| `[warning: …]` \| Zero or more probe warnings: server unreachable, model not in the server's list, no chat template, compression that can never trigger. \|
	46	+\| `[warning: …]` \| Zero or more probe warnings: server unreachable, model not in the server's list, no chat template, compression that can never trigger, no tool declared at all. \|
47	47	\| `config: … \\| provider: … \\| model: … \\| ctx: … \\| tools: … \\| skills: N` \| Configuration path, provider name, model identifier, context window with its origin (`config`, `/props`) or `unknown`, declared tool names, number of skills found. \|
48	48	\| `Agent ready. Commands: …` \| Terminal mode only. \|
49	49	\| `Shortcut: Ctrl+C …` \| Terminal mode only. \|

  | Line | Content |
  |------|---------|
--| `[warning: …]` | Zero or more probe warnings: server unreachable, model not in the server's list, no chat template, compression that can never trigger. |
++| `[warning: …]` | Zero or more probe warnings: server unreachable, model not in the server's list, no chat template, compression that can never trigger, no tool declared at all. |
  | `config: … \| provider: … \| model: … \| ctx: … \| tools: … \| skills: N` | Configuration path, provider name, model identifier, context window with its origin (`config`, `/props`) or `unknown`, declared tool names, number of skills found. |
  | `Agent ready. Commands: …` | Terminal mode only. |
  | `Shortcut: Ctrl+C …` | Terminal mode only. |

modified docs/en/reference/configuration.md +1 -0 Viewed

@@ -18,6 +18,7 @@
18	18	\| `displayCommands` \| bool \| `false` \| Print the numbered list of commands and file operations after each answer. \|
19	19	\| `system` \| string \| built-in prompt \| The system prompt, first message of every history. \|
20	20	\| `skillsDir` \| string \| `skills` \| Directory holding the skills, as `<name>.md` files or `<name>/SKILL.md` directories. A relative path is resolved from the directory `mm` is started in (terminal mode, or ACP mode without a config file), or next to the config file (ACP mode with one); the banner warns with the absolute path when nothing is found there. \|
	21	+\| `bashTool` \| bool \| `true` \| Declare `bash`. `false` removes it: the model can still read and edit files (`editTools`) and load skills, but never runs a command. Reword a system prompt that still announces a bash tool. \|
21	22	\| `editTools` \| bool \| `true` \| Declare `read_file`, `write_file`, `edit_file`. `false` leaves `bash` and `read_skill` only. \|
22	23	\| `sampling` \| map \| `temperature: 0.0`, `top_p: 0.9`, `max_tokens: 4096` \| Generation settings, OpenAI API key names, passed to the server. The shipped `agent.yaml` adds `parallel_tool_calls: false`. \|
23	24	\| `watchdogTimeout` \| duration \| `20s` \| Silence tolerated between two streamed tokens before the connection is declared hung. Doubled for summary requests. \|

  | `displayCommands` | bool | `false` | Print the numbered list of commands and file operations after each answer. |
  | `system` | string | built-in prompt | The system prompt, first message of every history. |
  | `skillsDir` | string | `skills` | Directory holding the skills, as `<name>.md` files or `<name>/SKILL.md` directories. A relative path is resolved from the directory `mm` is started in (terminal mode, or ACP mode without a config file), or next to the config file (ACP mode with one); the banner warns with the absolute path when nothing is found there. |
++| `bashTool` | bool | `true` | Declare `bash`. `false` removes it: the model can still read and edit files (`editTools`) and load skills, but never runs a command. Reword a system prompt that still announces a bash tool. |
  | `editTools` | bool | `true` | Declare `read_file`, `write_file`, `edit_file`. `false` leaves `bash` and `read_skill` only. |
  | `sampling` | map | `temperature: 0.0`, `top_p: 0.9`, `max_tokens: 4096` | Generation settings, OpenAI API key names, passed to the server. The shipped `agent.yaml` adds `parallel_tool_calls: false`. |
  | `watchdogTimeout` | duration | `20s` | Silence tolerated between two streamed tokens before the connection is declared hung. Doubled for summary requests. |

modified docs/en/reference/tools.md +1 -1 Viewed

@@ -4,7 +4,7 @@
4	4
5	5	## `bash`
6	6
7		-Always declared.
	7	+Declared when `bashTool: true` (the default).
8	8
9	9	\| Field \| Type \| Description \|
10	10	\|-------\|------\|-------------\|

  ## `bash`
--Always declared.
++Declared when `bashTool: true` (the default).
  | Field | Type | Description |
  |-------|------|-------------|

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

@@ -43,7 +43,7 @@ Affichée sur stdout en mode terminal, sur stderr en mode ACP. Grisée quand std
43	43
44	44	\| Ligne \| Contenu \|
45	45	\|-------\|---------\|
46		-\| `[warning: …]` \| Zéro ou plusieurs avertissements de la sonde : serveur injoignable, modèle absent de la liste du serveur, pas de template de chat, compression qui ne peut jamais se déclencher. \|
	46	+\| `[warning: …]` \| Zéro ou plusieurs avertissements de la sonde : serveur injoignable, modèle absent de la liste du serveur, pas de template de chat, compression qui ne peut jamais se déclencher, aucun outil déclaré. \|
47	47	\| `config: … \\| provider: … \\| model: … \\| ctx: … \\| tools: … \\| skills: N` \| Chemin de la configuration, nom du fournisseur, identifiant du modèle, fenêtre de contexte avec son origine (`config`, `/props`) ou `unknown`, noms des outils déclarés, nombre de skills trouvés. \|
48	48	\| `Agent ready. Commands: …` \| Mode terminal seulement. \|
49	49	\| `Shortcut: Ctrl+C …` \| Mode terminal seulement. \|

  | Ligne | Contenu |
  |-------|---------|
--| `[warning: …]` | Zéro ou plusieurs avertissements de la sonde : serveur injoignable, modèle absent de la liste du serveur, pas de template de chat, compression qui ne peut jamais se déclencher. |
++| `[warning: …]` | Zéro ou plusieurs avertissements de la sonde : serveur injoignable, modèle absent de la liste du serveur, pas de template de chat, compression qui ne peut jamais se déclencher, aucun outil déclaré. |
  | `config: … \| provider: … \| model: … \| ctx: … \| tools: … \| skills: N` | Chemin de la configuration, nom du fournisseur, identifiant du modèle, fenêtre de contexte avec son origine (`config`, `/props`) ou `unknown`, noms des outils déclarés, nombre de skills trouvés. |
  | `Agent ready. Commands: …` | Mode terminal seulement. |
  | `Shortcut: Ctrl+C …` | Mode terminal seulement. |

modified docs/fr/reference/configuration.md +1 -0 Viewed

@@ -18,6 +18,7 @@
18	18	\| `displayCommands` \| booléen \| `false` \| Affiche la liste numérotée des commandes et opérations sur fichiers après chaque réponse. \|
19	19	\| `system` \| chaîne \| prompt intégré \| Le prompt système, premier message de chaque historique. \|
20	20	\| `skillsDir` \| chaîne \| `skills` \| Répertoire des skills, sous forme de fichiers `<nom>.md` ou de répertoires `<nom>/SKILL.md`. Un chemin relatif est résolu depuis le répertoire de lancement de `mm` (mode terminal, ou mode ACP sans fichier de configuration), ou à côté du fichier de configuration (mode ACP avec fichier) ; la bannière avertit avec le chemin absolu quand rien n'y est trouvé. \|
	21	+\| `bashTool` \| booléen \| `true` \| Déclare `bash`. `false` le retire : le modèle peut encore lire et modifier des fichiers (`editTools`) et charger des skills, mais n'exécute jamais de commande. Reformulez un prompt système qui annonce encore un outil bash. \|
21	22	\| `editTools` \| booléen \| `true` \| Déclare `read_file`, `write_file`, `edit_file`. `false` ne laisse que `bash` et `read_skill`. \|
22	23	\| `sampling` \| dictionnaire \| `temperature: 0.0`, `top_p: 0.9`, `max_tokens: 4096` \| Paramètres de génération, noms de l'API OpenAI, transmis au serveur. L'`agent.yaml` livré ajoute `parallel_tool_calls: false`. \|
23	24	\| `watchdogTimeout` \| durée \| `20s` \| Silence toléré entre deux tokens diffusés avant de déclarer la connexion bloquée. Doublé pour les requêtes de résumé. \|

  | `displayCommands` | booléen | `false` | Affiche la liste numérotée des commandes et opérations sur fichiers après chaque réponse. |
  | `system` | chaîne | prompt intégré | Le prompt système, premier message de chaque historique. |
  | `skillsDir` | chaîne | `skills` | Répertoire des skills, sous forme de fichiers `<nom>.md` ou de répertoires `<nom>/SKILL.md`. Un chemin relatif est résolu depuis le répertoire de lancement de `mm` (mode terminal, ou mode ACP sans fichier de configuration), ou à côté du fichier de configuration (mode ACP avec fichier) ; la bannière avertit avec le chemin absolu quand rien n'y est trouvé. |
++| `bashTool` | booléen | `true` | Déclare `bash`. `false` le retire : le modèle peut encore lire et modifier des fichiers (`editTools`) et charger des skills, mais n'exécute jamais de commande. Reformulez un prompt système qui annonce encore un outil bash. |
  | `editTools` | booléen | `true` | Déclare `read_file`, `write_file`, `edit_file`. `false` ne laisse que `bash` et `read_skill`. |
  | `sampling` | dictionnaire | `temperature: 0.0`, `top_p: 0.9`, `max_tokens: 4096` | Paramètres de génération, noms de l'API OpenAI, transmis au serveur. L'`agent.yaml` livré ajoute `parallel_tool_calls: false`. |
  | `watchdogTimeout` | durée | `20s` | Silence toléré entre deux tokens diffusés avant de déclarer la connexion bloquée. Doublé pour les requêtes de résumé. |

modified docs/fr/reference/tools.md +1 -1 Viewed

@@ -4,7 +4,7 @@
4	4
5	5	## `bash`
6	6
7		-Toujours déclaré.
	7	+Déclaré quand `bashTool: true` (la valeur par défaut).
8	8
9	9	\| Champ \| Type \| Description \|
10	10	\|-------\|------\|-------------\|

  ## `bash`
--Toujours déclaré.
++Déclaré quand `bashTool: true` (la valeur par défaut).
  | Champ | Type | Description |
  |-------|------|-------------|

modified internal/config/config.go +9 -0 Viewed

@@ -73,6 +73,14 @@ type Config struct {
73	73	// tool. When it holds no *.md file, the tool is not declared at all.
74	74	SkillsDir string `yaml:"skillsDir"`
75	75
	76	+ // BashTool: declare the `bash` tool. On by default — it is the one tool
	77	+ // the talk is built on. Off, the model can only read and edit files (with
	78	+ // EditTools) and load skills: an agent that explains and patches code but
	79	+ // never runs a command, for a machine where a shell in the model's hands is
	80	+ // not wanted. The system prompt is the operator's: one that still says
	81	+ // "you have a bash tool" should be reworded with it.
	82	+ BashTool bool `yaml:"bashTool"`
	83	+
76	84	// EditTools: declare the built-in file tools (read_file, write_file,
77	85	// edit_file). Off, the agent has bash and read_skill only, exactly like
78	86	// part 09 — and edits files through the `edit` CLI if it is on the PATH.
@@ -184,6 +192,7 @@ var Cfg = Config{
184	192	// needs. `displayCommands: true` in the YAML turns it on.
185	193	DisplayCommands: false,
186	194	SkillsDir: "skills",
	195	+ BashTool: true,
187	196	EditTools: true,
188	197	System: `You are a coding agent working in a terminal.
189	198	You have a "bash" tool to run shell commands.

  	// tool. When it holds no *.md file, the tool is not declared at all.
  	SkillsDir string `yaml:"skillsDir"`
++	// BashTool: declare the `bash` tool. On by default — it is the one tool
++	// the talk is built on. Off, the model can only read and edit files (with
++	// EditTools) and load skills: an agent that explains and patches code but
++	// never runs a command, for a machine where a shell in the model's hands is
++	// not wanted. The system prompt is the operator's: one that still says
++	// "you have a bash tool" should be reworded with it.
++	BashTool bool `yaml:"bashTool"`
++
  	// EditTools: declare the built-in file tools (read_file, write_file,
  	// edit_file). Off, the agent has bash and read_skill only, exactly like
  	// part 09 — and edits files through the `edit` CLI if it is on the PATH.
  	// needs. `displayCommands: true` in the YAML turns it on.
  	DisplayCommands: false,
  	SkillsDir:       "skills",
++	BashTool:        true,
  	EditTools:       true,
  	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 +36 -0 Viewed

@@ -72,3 +72,39 @@ func TestTUIAndPermissionsConfig(t *testing.T) {
72	72	t.Errorf("Load(bad.yaml) = %v, want a permissions validation error", err)
73	73	}
74	74	}
	75	+
	76	+// bash is declared by default, so every YAML written before `bashTool`
	77	+// existed keeps its tool; `bashTool: false` is read; a YAML that says nothing
	78	+// about it leaves the default alone even when it sets other keys.
	79	+func TestBashToolConfig(t *testing.T) {
	80	+ defaults := Cfg
	81	+ t.Cleanup(func() { Cfg = defaults })
	82	+
	83	+ if !Cfg.BashTool {
	84	+ t.Fatal("defaults: bashTool off, want on")
	85	+ }
	86	+
	87	+ dir := t.TempDir()
	88	+ off := filepath.Join(dir, "off.yaml")
	89	+ if err := os.WriteFile(off, []byte("bashTool: false\n"), 0o644); err != nil {
	90	+ t.Fatal(err)
	91	+ }
	92	+ if _, err := Load(off); err != nil {
	93	+ t.Fatal(err)
	94	+ }
	95	+ if Cfg.BashTool {
	96	+ t.Error("after off.yaml: bashTool still on")
	97	+ }
	98	+
	99	+ Cfg = defaults
	100	+ silent := filepath.Join(dir, "silent.yaml")
	101	+ if err := os.WriteFile(silent, []byte("maxTurns: 3\n"), 0o644); err != nil {
	102	+ t.Fatal(err)
	103	+ }
	104	+ if _, err := Load(silent); err != nil {
	105	+ t.Fatal(err)
	106	+ }
	107	+ if !Cfg.BashTool {
	108	+ t.Error("after silent.yaml: bashTool off, want the default kept")
	109	+ }
	110	+}

  		t.Errorf("Load(bad.yaml) = %v, want a permissions validation error", err)
  	}
  }
++
++// bash is declared by default, so every YAML written before `bashTool`
++// existed keeps its tool; `bashTool: false` is read; a YAML that says nothing
++// about it leaves the default alone even when it sets other keys.
++func TestBashToolConfig(t *testing.T) {
++	defaults := Cfg
++	t.Cleanup(func() { Cfg = defaults })
++
++	if !Cfg.BashTool {
++		t.Fatal("defaults: bashTool off, want on")
++	}
++
++	dir := t.TempDir()
++	off := filepath.Join(dir, "off.yaml")
++	if err := os.WriteFile(off, []byte("bashTool: false\n"), 0o644); err != nil {
++		t.Fatal(err)
++	}
++	if _, err := Load(off); err != nil {
++		t.Fatal(err)
++	}
++	if Cfg.BashTool {
++		t.Error("after off.yaml: bashTool still on")
++	}
++
++	Cfg = defaults
++	silent := filepath.Join(dir, "silent.yaml")
++	if err := os.WriteFile(silent, []byte("maxTurns: 3\n"), 0o644); err != nil {
++		t.Fatal(err)
++	}
++	if _, err := Load(silent); err != nil {
++		t.Fatal(err)
++	}
++	if !Cfg.BashTool {
++		t.Error("after silent.yaml: bashTool off, want the default kept")
++	}
++}

modified main.go +20 -5 Viewed

@@ -158,16 +158,21 @@ func runWindow(ctx context.Context, e *engine.Engine, agentTools []ai.ToolRef, d
158	158	agent.Run(ctx, e, config.Cfg.System, agentTools, d)
159	159	}
160	160
161		-// assembleTools declares the tools every front end gets: `bash` always;
162		-// `read_skill` only when the skills directory holds something to read (its
	161	+// assembleTools declares the tools every front end gets: `bash` unless
	162	+// `bashTool: false` turned it off; `read_skill` only when the skills
	163	+// directory holds something to read (its
163	164	// description carries the catalogue, so declaring it empty would advertise a
164	165	// tool that can do nothing); the file tools behind `editTools` — the same
165	166	// binary must run as part 09 did (bash + the `edit` CLI) and as this part
166	167	// (built-in tools), so the two can be compared on the same prompts. It
167	168	// returns the tools, their names for the banner, and the skill count.
168	169	func assembleTools(e engine.Engine, d detector.LoopDetector) ([]ai.ToolRef, []string, int) {
169		- agentTools := []ai.ToolRef{tools.Bash(e.G, d)}
170		- toolNames := []string{"bash"}
	170	+ var agentTools []ai.ToolRef
	171	+ var toolNames []string
	172	+ if config.Cfg.BashTool {
	173	+ agentTools = append(agentTools, tools.Bash(e.G, d))
	174	+ toolNames = append(toolNames, "bash")
	175	+ }
171	176	skillCount := len(skills.List(config.Cfg.SkillsDir))
172	177	if skillTool := tools.ReadSkill(e.G, config.Cfg.SkillsDir, d); skillTool != nil {
173	178	agentTools = append(agentTools, skillTool)
@@ -211,11 +216,21 @@ func printBanner(ctx context.Context, e *engine.Engine, path string, toolNames [
211	216	info.Warnings = append(info.Warnings,
212	217	fmt.Sprintf("no skills found in %s (expected <name>.md or <name>/SKILL.md there)", config.Cfg.SkillsDir))
213	218	}
	219	+ // No tool at all is a legal set-up — a chat with the model — but rarely the
	220	+ // intended one; say so where the "tools:" column would otherwise be empty.
	221	+ if len(toolNames) == 0 {
	222	+ info.Warnings = append(info.Warnings,
	223	+ "no tool is declared (bashTool and editTools are both off, and there are no skills): the model can only answer, not act")
	224	+ }
214	225	for _, w := range info.Warnings {
215	226	say("%s[warning: %s]%s", dim, w, off)
216	227	}
	228	+ toolCol := strings.Join(toolNames, ", ")
	229	+ if toolCol == "" {
	230	+ toolCol = "none"
	231	+ }
217	232	say("%sconfig: %s \| provider: %s \| model: %s \| ctx: %s \| tools: %s \| skills: %d%s",
218		- dim, path, e.Backend.Provider, config.Cfg.Model, ctxCol, strings.Join(toolNames, ", "), skillCount, off)
	233	+ dim, path, e.Backend.Provider, config.Cfg.Model, ctxCol, toolCol, skillCount, off)
219	234	}
220	235
221	236	// windowMode decides whether the window front end runs, and where the banner

  	agent.Run(ctx, e, config.Cfg.System, agentTools, d)
  }
--// assembleTools declares the tools every front end gets: `bash` always;
++// assembleTools declares the tools every front end gets: `bash` unless
--// `read_skill` only when the skills directory holds something to read (its
++// `bashTool: false` turned it off; `read_skill` only when the skills
++// directory holds something to read (its
  // description carries the catalogue, so declaring it empty would advertise a
  // tool that can do nothing); the file tools behind `editTools` — the same
  // binary must run as part 09 did (bash + the `edit` CLI) and as this part
  // (built-in tools), so the two can be compared on the same prompts. It
  // returns the tools, their names for the banner, and the skill count.
  func assembleTools(e *engine.Engine, d *detector.LoopDetector) ([]ai.ToolRef, []string, int) {
--	agentTools := []ai.ToolRef{tools.Bash(e.G, d)}
++	var agentTools []ai.ToolRef
--	toolNames := []string{"bash"}
++	var toolNames []string
++	if config.Cfg.BashTool {
++		agentTools = append(agentTools, tools.Bash(e.G, d))
++		toolNames = append(toolNames, "bash")
++	}
  	skillCount := len(skills.List(config.Cfg.SkillsDir))
  	if skillTool := tools.ReadSkill(e.G, config.Cfg.SkillsDir, d); skillTool != nil {
  		agentTools = append(agentTools, skillTool)
  		info.Warnings = append(info.Warnings,
  			fmt.Sprintf("no skills found in %s (expected <name>.md or <name>/SKILL.md there)", config.Cfg.SkillsDir))
  	}
++	// No tool at all is a legal set-up — a chat with the model — but rarely the
++	// intended one; say so where the "tools:" column would otherwise be empty.
++	if len(toolNames) == 0 {
++		info.Warnings = append(info.Warnings,
++			"no tool is declared (bashTool and editTools are both off, and there are no skills): the model can only answer, not act")
++	}
  	for _, w := range info.Warnings {
  		say("%s[warning: %s]%s", dim, w, off)
  	}
++	toolCol := strings.Join(toolNames, ", ")
++	if toolCol == "" {
++		toolCol = "none"
++	}
  	say("%sconfig: %s | provider: %s | model: %s | ctx: %s | tools: %s | skills: %d%s",
--		dim, path, e.Backend.Provider, config.Cfg.Model, ctxCol, strings.Join(toolNames, ", "), skillCount, off)
++		dim, path, e.Backend.Provider, config.Cfg.Model, ctxCol, toolCol, skillCount, off)
  }
  // windowMode decides whether the window front end runs, and where the banner

📦 Mini Me