bots-garden/mini-me · f81c3b6

modified .tickets/issues/0210-acp-create-mode-mm-acp-create-mm-init-from-an-acp.yaml +2 -2 Viewed

@@ -175,6 +175,8 @@ tasks:
175	175	name: k33g
176	176	email: ph.charriere@gmail.com
177	177	createdAt: 2026-11-14T05:41:20.000Z
	178	+ updatedAt: 2026-11-14T05:56:48.000Z
	179	+ closedAt: 2026-11-14T05:56:48.000Z
178	180	comments:
179	181	- id: 1
180	182	author:
@@ -194,8 +196,6 @@ comments:
194	196	client by hand), the acceptance "tried in two real clients",
195	197	permission buttons for choices in conversation mode, and session
196	198	config options. qlty not run (not installed in this sandbox).
197		- updatedAt: 2026-11-14T05:56:48.000Z
198		- closedAt: 2026-11-14T05:56:48.000Z
199	199	- id: 2
200	200	author:
201	201	name: k33g

        name: k33g
        email: ph.charriere@gmail.com
      createdAt: 2026-11-14T05:41:20.000Z
++    updatedAt: 2026-11-14T05:56:48.000Z
++    closedAt: 2026-11-14T05:56:48.000Z
  comments:
    - id: 1
      author:
        client by hand), the acceptance "tried in two real clients",
        permission buttons for choices in conversation mode, and session
        config options. qlty not run (not installed in this sandbox).
--    updatedAt: 2026-11-14T05:56:48.000Z
--    closedAt: 2026-11-14T05:56:48.000Z
    - id: 2
      author:
        name: k33g

modified .tickets/issues/0211-check-if-llamacpp-1-0-0-is-working.yaml +2 -2 Viewed

@@ -6,9 +6,9 @@ author:
6	6	name: k33g
7	7	email: ph.charriere@gmail.com
8	8	createdAt: 2026-11-14T05:56:02.153Z
9		-updatedAt: 2026-11-14T06:35:10.000Z
	9	+updatedAt: 2026-11-14T18:45:26.000Z
10	10	labels: []
11		-milestone: v1.1.0
	11	+milestone: v2.0.0
12	12	epic: Providers (LLM Engine)
13	13	body: ""
14	14	tasks: []

    name: k33g
    email: ph.charriere@gmail.com
  createdAt: 2026-11-14T05:56:02.153Z
--updatedAt: 2026-11-14T06:35:10.000Z
++updatedAt: 2026-11-14T18:45:26.000Z
  labels: []
--milestone: v1.1.0
++milestone: v2.0.0
  epic: Providers (LLM Engine)
  body: ""
  tasks: []

modified .tickets/issues/0212-mm-dockerinit-know-the-mm-folder-project-layout.yaml +2 -2 Viewed

@@ -137,6 +137,8 @@ tasks:
137	137	name: k33g
138	138	email: ph.charriere@gmail.com
139	139	createdAt: 2026-11-14T06:29:41.000Z
	140	+ updatedAt: 2026-11-14T06:37:57.000Z
	141	+ closedAt: 2026-11-14T06:37:57.000Z
140	142	comments:
141	143	- id: 1
142	144	author:
@@ -159,8 +161,6 @@ comments:
159	161	Also asked in the same session: `mm sbxinit` now prints
160	162	`Run the sandbox <name>: sbx env run --force` (and the line in the
161	163	sbxenv.yaml comments).
162		- updatedAt: 2026-11-14T06:37:57.000Z
163		- closedAt: 2026-11-14T06:37:57.000Z
164	164	- id: 2
165	165	author:
166	166	name: k33g

        name: k33g
        email: ph.charriere@gmail.com
      createdAt: 2026-11-14T06:29:41.000Z
++    updatedAt: 2026-11-14T06:37:57.000Z
++    closedAt: 2026-11-14T06:37:57.000Z
  comments:
    - id: 1
      author:
        Also asked in the same session: `mm sbxinit` now prints
        `Run the sandbox <name>: sbx env run --force` (and the line in the
        sbxenv.yaml comments).
--    updatedAt: 2026-11-14T06:37:57.000Z
--    closedAt: 2026-11-14T06:37:57.000Z
    - id: 2
      author:
        name: k33g

modified .tickets/issues/0214-mm-acp-create-start-ollama-models-are-listed-fro.yaml +63 -7 Viewed

@@ -1,12 +1,12 @@
1	1	# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2	2	id: 214
3	3	title: "mm acp-create (/start): Ollama models are listed from the container address instead of the local one"
4		-state: open
	4	+state: closed
5	5	author:
6	6	name: k33g
7	7	email: ph.charriere@gmail.com
8	8	createdAt: 2026-11-14T07:40:50.000Z
9		-updatedAt: 2026-11-14T07:47:59.000Z
	9	+updatedAt: 2026-11-14T18:14:17.000Z
10	10	labels:
11	11	- bug
12	12	- kind::bug
@@ -76,44 +76,54 @@ body: \|
76	76	tasks:
77	77	- id: 1
78	78	title: Reproduce - mm acp-create from Zed and mm init in a terminal, Ollama on the host, note the address used
79		- state: open
	79	+ state: closed
80	80	priority: high
81	81	author:
82	82	name: k33g
83	83	email: ph.charriere@gmail.com
84	84	createdAt: 2026-11-14T07:40:50.000Z
	85	+ updatedAt: 2026-11-14T18:14:17.000Z
	86	+ closedAt: 2026-11-14T18:14:17.000Z
85	87	- id: 2
86	88	title: Find why the local address is not taken (probe timeout, localhost vs 127.0.0.1 / IPv6, environment of the editor's process)
87		- state: open
	89	+ state: closed
88	90	priority: high
89	91	author:
90	92	name: k33g
91	93	email: ph.charriere@gmail.com
92	94	createdAt: 2026-11-14T07:40:50.000Z
	95	+ updatedAt: 2026-11-14T18:14:17.000Z
	96	+ closedAt: 2026-11-14T18:14:17.000Z
93	97	- id: 3
94	98	title: Fix - the wizard lists the models from the local address when mm is not in a container
95		- state: open
	99	+ state: closed
96	100	priority: high
97	101	author:
98	102	name: k33g
99	103	email: ph.charriere@gmail.com
100	104	createdAt: 2026-11-14T07:40:50.000Z
	105	+ updatedAt: 2026-11-14T17:54:42.000Z
	106	+ closedAt: 2026-11-14T17:54:42.000Z
101	107	- id: 4
102	108	title: Show the address the models were listed from in the model question (both modes)
103		- state: open
	109	+ state: closed
104	110	priority: medium
105	111	author:
106	112	name: k33g
107	113	email: ph.charriere@gmail.com
108	114	createdAt: 2026-11-14T07:40:50.000Z
	115	+ updatedAt: 2026-11-14T17:54:42.000Z
	116	+ closedAt: 2026-11-14T17:54:42.000Z
109	117	- id: 5
110	118	title: Tests - a local server answering slowly or only on 127.0.0.1 keeps the local address
111		- state: open
	119	+ state: closed
112	120	priority: medium
113	121	author:
114	122	name: k33g
115	123	email: ph.charriere@gmail.com
116	124	createdAt: 2026-11-14T07:40:50.000Z
	125	+ updatedAt: 2026-11-14T17:54:42.000Z
	126	+ closedAt: 2026-11-14T17:54:42.000Z
117	127	comments:
118	128	- id: 1
119	129	author:
@@ -128,3 +138,49 @@ comments:
128	138	it: look at what is specific to Ollama — its `/v1/models` answering
129	139	later than 2 s or not 200, Ollama listening on IPv4 only while DMR
130	140	answers on both, or `OLLAMA_HOST` on the user's machine.
	141	+ - id: 2
	142	+ author:
	143	+ name: k33g
	144	+ email: ph.charriere@gmail.com
	145	+ createdAt: 2026-11-14T17:54:42.000Z
	146	+ body: \|
	147	+ Fixed on `fix/wizard-local-address` (uncommitted), without waiting for
	148	+ the exact cause on the user's machine: the fix covers hypotheses 1
	149	+ (slow probe), 2 (IPv4 only) and 4 (Docker Desktop resolving
	150	+ host.docker.internal on the host, which hid the failed probe).
	151	+
	152	+ - The wizard (mm init and mm acp-create) no longer takes the engine's
	153	+ resolution (2 s probe of `/models`, then a silent switch to the
	154	+ container fallback). `resolveBase` → `AGENT_BASE_URL` (or the
	155	+ legacy variable) when set; else the first local address whose model
	156	+ list answers (the provider's own list endpoint, `/api/tags` for
	157	+ Ollama, 5 s): the default, then the same with `127.0.0.1` when it
	158	+ names `localhost`; the fallback only when none answers; a provider
	159	+ without a fallback is not probed. `engine.EnvBaseURL` added.
	160	+ - The model question names the address: `Model? (listed from …)`.
	161	+ - The written file is unchanged (no `baseUrl` for the default).
	162	+ - Runtime (`mm acp` with the written file) still uses the engine's
	163	+ resolution — not touched here.
	164	+
	165	+ Tests: `TestChooseBasePrefersALocalAddress`, `TestLocalAddresses`,
	166	+ `TestResolveBaseKeepsASlowLocalServer` (a server whose `/v1/models`
	167	+ takes 3 s: the engine's probe takes the fallback, the wizard keeps the
	168	+ local address), `TestResolveBaseHonoursAgentBaseURL`.
	169	+
	170	+ End to end, real binary, `mm acp-create --provider ollama`, a fake
	171	+ Ollama on 127.0.0.1:11434 whose `/v1/models` takes 3 s: before →
	172	+ `Model ID? (no list from http://host.docker.internal:11434/v1 …)`;
	173	+ after → `Model? (listed from http://localhost:11434/v1)`.
	174	+
	175	+ Not done: tasks 1 and 2 (reproduce on the user's Mac with the real
	176	+ Ollama, and name the exact cause) — the curl checks given in the
	177	+ conversation would tell.
	178	+ - id: 3
	179	+ author:
	180	+ name: k33g
	181	+ email: ph.charriere@gmail.com
	182	+ createdAt: 2026-11-14T18:14:17.000Z
	183	+ body: \|
	184	+ Tested by the user with the real Ollama: it works (local models
	185	+ listed). Closed. The exact cause on the user's machine was not named
	186	+ (tasks 1–2 closed with the user's test as the reproduction).

  # Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
  id: 214
  title: "mm acp-create (/start): Ollama models are listed from the container address instead of the local one"
--state: open
++state: closed
  author:
    name: k33g
    email: ph.charriere@gmail.com
  createdAt: 2026-11-14T07:40:50.000Z
--updatedAt: 2026-11-14T07:47:59.000Z
++updatedAt: 2026-11-14T18:14:17.000Z
  labels:
    - bug
    - kind::bug
  tasks:
    - id: 1
      title: Reproduce - mm acp-create from Zed and mm init in a terminal, Ollama on the host, note the address used
--    state: open
++    state: closed
      priority: high
      author:
        name: k33g
        email: ph.charriere@gmail.com
      createdAt: 2026-11-14T07:40:50.000Z
++    updatedAt: 2026-11-14T18:14:17.000Z
++    closedAt: 2026-11-14T18:14:17.000Z
    - id: 2
      title: Find why the local address is not taken (probe timeout, localhost vs 127.0.0.1 / IPv6, environment of the editor's process)
--    state: open
++    state: closed
      priority: high
      author:
        name: k33g
        email: ph.charriere@gmail.com
      createdAt: 2026-11-14T07:40:50.000Z
++    updatedAt: 2026-11-14T18:14:17.000Z
++    closedAt: 2026-11-14T18:14:17.000Z
    - id: 3
      title: Fix - the wizard lists the models from the local address when mm is not in a container
--    state: open
++    state: closed
      priority: high
      author:
        name: k33g
        email: ph.charriere@gmail.com
      createdAt: 2026-11-14T07:40:50.000Z
++    updatedAt: 2026-11-14T17:54:42.000Z
++    closedAt: 2026-11-14T17:54:42.000Z
    - id: 4
      title: Show the address the models were listed from in the model question (both modes)
--    state: open
++    state: closed
      priority: medium
      author:
        name: k33g
        email: ph.charriere@gmail.com
      createdAt: 2026-11-14T07:40:50.000Z
++    updatedAt: 2026-11-14T17:54:42.000Z
++    closedAt: 2026-11-14T17:54:42.000Z
    - id: 5
      title: Tests - a local server answering slowly or only on 127.0.0.1 keeps the local address
--    state: open
++    state: closed
      priority: medium
      author:
        name: k33g
        email: ph.charriere@gmail.com
      createdAt: 2026-11-14T07:40:50.000Z
++    updatedAt: 2026-11-14T17:54:42.000Z
++    closedAt: 2026-11-14T17:54:42.000Z
  comments:
    - id: 1
      author:
        it: look at what is specific to Ollama — its `/v1/models` answering
        later than 2 s or not 200, Ollama listening on IPv4 only while DMR
        answers on both, or `OLLAMA_HOST` on the user's machine.
++  - id: 2
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T17:54:42.000Z
++    body: |
++      Fixed on `fix/wizard-local-address` (uncommitted), without waiting for
++      the exact cause on the user's machine: the fix covers hypotheses 1
++      (slow probe), 2 (IPv4 only) and 4 (Docker Desktop resolving
++      host.docker.internal on the host, which hid the failed probe).
++
++      - The wizard (mm init and mm acp-create) no longer takes the engine's
++        resolution (2 s probe of `/models`, then a silent switch to the
++        container fallback). `resolveBase` → `AGENT_BASE_URL` (or the
++        legacy variable) when set; else the first local address whose model
++        list answers (the provider's own list endpoint, `/api/tags` for
++        Ollama, 5 s): the default, then the same with `127.0.0.1` when it
++        names `localhost`; the fallback only when none answers; a provider
++        without a fallback is not probed. `engine.EnvBaseURL` added.
++      - The model question names the address: `Model? (listed from …)`.
++      - The written file is unchanged (no `baseUrl` for the default).
++      - Runtime (`mm acp` with the written file) still uses the engine's
++        resolution — not touched here.
++
++      Tests: `TestChooseBasePrefersALocalAddress`, `TestLocalAddresses`,
++      `TestResolveBaseKeepsASlowLocalServer` (a server whose `/v1/models`
++      takes 3 s: the engine's probe takes the fallback, the wizard keeps the
++      local address), `TestResolveBaseHonoursAgentBaseURL`.
++
++      End to end, real binary, `mm acp-create --provider ollama`, a fake
++      Ollama on 127.0.0.1:11434 whose `/v1/models` takes 3 s: before →
++      `Model ID? (no list from http://host.docker.internal:11434/v1 …)`;
++      after → `Model? (listed from http://localhost:11434/v1)`.
++
++      Not done: tasks 1 and 2 (reproduce on the user's Mac with the real
++      Ollama, and name the exact cause) — the curl checks given in the
++      conversation would tell.
++  - id: 3
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T18:14:17.000Z
++    body: |
++      Tested by the user with the real Ollama: it works (local models
++      listed). Closed. The exact cause on the user's machine was not named
++      (tasks 1–2 closed with the user's test as the reproduction).

modified .tickets/issues/0215-mm-acp-create-start-test-every-provider-address.yaml +34 -5 Viewed

@@ -1,12 +1,12 @@
1	1	# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2	2	id: 215
3	3	title: "mm acp-create (/start): test every provider - address used and model list"
4		-state: open
	4	+state: closed
5	5	author:
6	6	name: k33g
7	7	email: ph.charriere@gmail.com
8	8	createdAt: 2026-11-14T07:41:16.000Z
9		-updatedAt: 2026-11-14T08:07:53.000Z
	9	+updatedAt: 2026-11-14T18:45:13.000Z
10	10	labels:
11	11	- kind::chore
12	12	- To be tested
@@ -79,12 +79,14 @@ tasks:
79	79	closedAt: 2026-11-14T07:47:59.000Z
80	80	- id: 3
81	81	title: llamacpp - llama-server on 127.0.0.1:8080
82		- state: open
	82	+ state: closed
83	83	priority: medium
84	84	author:
85	85	name: k33g
86	86	email: ph.charriere@gmail.com
87	87	createdAt: 2026-11-14T07:41:16.000Z
	88	+ updatedAt: 2026-11-14T18:24:44.000Z
	89	+ closedAt: 2026-11-14T18:24:44.000Z
88	90	- id: 4
89	91	title: kronk
90	92	state: closed
@@ -107,20 +109,24 @@ tasks:
107	109	closedAt: 2026-11-14T08:07:53.000Z
108	110	- id: 6
109	111	title: Same checks with mm in a container or a Docker Sandbox (the fallback must be taken there)
110		- state: open
	112	+ state: closed
111	113	priority: medium
112	114	author:
113	115	name: k33g
114	116	email: ph.charriere@gmail.com
115	117	createdAt: 2026-11-14T07:41:16.000Z
	118	+ updatedAt: 2026-11-14T18:45:13.000Z
	119	+ closedAt: 2026-11-14T18:45:13.000Z
116	120	- id: 7
117	121	title: Write the results in a comment (version, client, address used, list ok or not); a ticket per bug found
118		- state: open
	122	+ state: closed
119	123	priority: low
120	124	author:
121	125	name: k33g
122	126	email: ph.charriere@gmail.com
123	127	createdAt: 2026-11-14T07:41:16.000Z
	128	+ updatedAt: 2026-11-14T18:45:13.000Z
	129	+ closedAt: 2026-11-14T18:45:13.000Z
124	130	comments:
125	131	- id: 1
126	132	author:
@@ -148,3 +154,26 @@ comments:
148	154	body: \|
149	155	User (2026-11-14): openai works too. Still to try: llamacpp, and
150	156	inside a container / sandbox.
	157	+ - id: 4
	158	+ author:
	159	+ name: k33g
	160	+ email: ph.charriere@gmail.com
	161	+ createdAt: 2026-11-14T18:24:44.000Z
	162	+ body: \|
	163	+ User (2026-11-14): llamacpp works too. Every provider now checked
	164	+ on the host: dmr, llmman, kronk, openai, llamacpp, and ollama after
	165	+ the fix of #214. Still to try: inside a container / Docker Sandbox
	166	+ (task 6).
	167	+ - id: 5
	168	+ author:
	169	+ name: k33g
	170	+ email: ph.charriere@gmail.com
	171	+ createdAt: 2026-11-14T18:45:13.000Z
	172	+ body: \|
	173	+ Closed at the user's request. Summary: on the host, every provider
	174	+ lists its models in `/start` (dmr, llmman, kronk, openai, llamacpp;
	175	+ ollama after #214). In a container, the fallback works for ollama and
	176	+ llmman; llamacpp had none — added by #217 (on a branch, not in a
	177	+ release yet: a container built by `mm dockerinit` runs the latest
	178	+ release, v1.0.9, which does not have it). Docker Sandbox not tried.
	179	+

  # Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
  id: 215
  title: "mm acp-create (/start): test every provider - address used and model list"
--state: open
++state: closed
  author:
    name: k33g
    email: ph.charriere@gmail.com
  createdAt: 2026-11-14T07:41:16.000Z
--updatedAt: 2026-11-14T08:07:53.000Z
++updatedAt: 2026-11-14T18:45:13.000Z
  labels:
    - kind::chore
    - To be tested
      closedAt: 2026-11-14T07:47:59.000Z
    - id: 3
      title: llamacpp - llama-server on 127.0.0.1:8080
--    state: open
++    state: closed
      priority: medium
      author:
        name: k33g
        email: ph.charriere@gmail.com
      createdAt: 2026-11-14T07:41:16.000Z
++    updatedAt: 2026-11-14T18:24:44.000Z
++    closedAt: 2026-11-14T18:24:44.000Z
    - id: 4
      title: kronk
      state: closed
      closedAt: 2026-11-14T08:07:53.000Z
    - id: 6
      title: Same checks with mm in a container or a Docker Sandbox (the fallback must be taken there)
--    state: open
++    state: closed
      priority: medium
      author:
        name: k33g
        email: ph.charriere@gmail.com
      createdAt: 2026-11-14T07:41:16.000Z
++    updatedAt: 2026-11-14T18:45:13.000Z
++    closedAt: 2026-11-14T18:45:13.000Z
    - id: 7
      title: Write the results in a comment (version, client, address used, list ok or not); a ticket per bug found
--    state: open
++    state: closed
      priority: low
      author:
        name: k33g
        email: ph.charriere@gmail.com
      createdAt: 2026-11-14T07:41:16.000Z
++    updatedAt: 2026-11-14T18:45:13.000Z
++    closedAt: 2026-11-14T18:45:13.000Z
  comments:
    - id: 1
      author:
      body: |
        User (2026-11-14): **openai works** too. Still to try: llamacpp, and
        inside a container / sandbox.
++  - id: 4
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T18:24:44.000Z
++    body: |
++      User (2026-11-14): **llamacpp works** too. Every provider now checked
++      on the host: dmr, llmman, kronk, openai, llamacpp, and ollama after
++      the fix of #214. Still to try: inside a container / Docker Sandbox
++      (task 6).
++  - id: 5
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T18:45:13.000Z
++    body: |
++      Closed at the user's request. Summary: on the host, every provider
++      lists its models in `/start` (dmr, llmman, kronk, openai, llamacpp;
++      ollama after #214). In a container, the fallback works for ollama and
++      llmman; llamacpp had none — added by #217 (on a branch, not in a
++      release yet: a container built by `mm dockerinit` runs the latest
++      release, v1.0.9, which does not have it). Docker Sandbox not tried.
++

modified .tickets/issues/0216-mm-dockerinit-i-have-no-name-in-the-container-th.yaml +11 -4 Viewed

@@ -1,12 +1,12 @@
1	1	# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
2	2	id: 216
3	3	title: "mm dockerinit: \"I have no name!\" in the container - the mm account has uid 1000, compose runs as the user's uid"
4		-state: open
	4	+state: closed
5	5	author:
6	6	name: k33g
7	7	email: ph.charriere@gmail.com
8	8	createdAt: 2026-11-14T17:31:27.000Z
9		-updatedAt: 2026-11-14T17:35:14.000Z
	9	+updatedAt: 2026-11-14T17:50:32.000Z
10	10	labels:
11	11	- bug
12	12	- kind::bug
@@ -143,6 +143,8 @@ tasks:
143	143	name: k33g
144	144	email: ph.charriere@gmail.com
145	145	createdAt: 2026-11-14T17:31:27.000Z
	146	+ updatedAt: 2026-11-14T17:35:14.000Z
	147	+ closedAt: 2026-11-14T17:35:14.000Z
146	148	comments:
147	149	- id: 1
148	150	author:
@@ -170,5 +172,10 @@ comments:
170	172	identity. Unrelated to the fix — the docs say to set `user.name` /
171	173	`user.email` or the `GIT_*` variables. Not tried: `mm tui` in the
172	174	container (no terminal here), ssh, Docker Desktop on macOS.
173		- updatedAt: 2026-11-14T17:35:14.000Z
174		- closedAt: 2026-11-14T17:35:14.000Z
	175	+ - id: 2
	176	+ author:
	177	+ name: k33g
	178	+ email: ph.charriere@gmail.com
	179	+ createdAt: 2026-11-14T17:50:32.000Z
	180	+ body: \|
	181	+ Closed at the user's request.

  # Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
  id: 216
  title: "mm dockerinit: \"I have no name!\" in the container - the mm account has uid 1000, compose runs as the user's uid"
--state: open
++state: closed
  author:
    name: k33g
    email: ph.charriere@gmail.com
  createdAt: 2026-11-14T17:31:27.000Z
--updatedAt: 2026-11-14T17:35:14.000Z
++updatedAt: 2026-11-14T17:50:32.000Z
  labels:
    - bug
    - kind::bug
        name: k33g
        email: ph.charriere@gmail.com
      createdAt: 2026-11-14T17:31:27.000Z
++    updatedAt: 2026-11-14T17:35:14.000Z
++    closedAt: 2026-11-14T17:35:14.000Z
  comments:
    - id: 1
      author:
        identity. Unrelated to the fix — the docs say to set `user.name` /
        `user.email` or the `GIT_*` variables. Not tried: `mm tui` in the
        container (no terminal here), ssh, Docker Desktop on macOS.
--    updatedAt: 2026-11-14T17:35:14.000Z
++  - id: 2
--    closedAt: 2026-11-14T17:35:14.000Z
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T17:50:32.000Z
++    body: |
++      Closed at the user's request.

added .tickets/issues/0217-llamacpp-fall-back-on-host-docker-internal-8080.yaml +154 -0 Viewed

new file mode 100644
@@ -0,0 +1,154 @@
	1	+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
	2	+id: 217
	3	+title: "llamacpp: fall back on host.docker.internal:8080 like the other local providers"
	4	+state: closed
	5	+author:
	6	+ name: k33g
	7	+ email: ph.charriere@gmail.com
	8	+createdAt: 2026-11-14T18:36:24.000Z
	9	+updatedAt: 2026-11-14T18:45:13.000Z
	10	+labels:
	11	+ - enhancement
	12	+ - kind::feature
	13	+milestone: v1.1.0
	14	+epic: Providers (LLM Engine)
	15	+related:
	16	+ - 215
	17	+body: \|
	18	+ ## Seen (user, 2026-11-14, while testing #215)
	19	+
	20	+ With mm in a container, the fallback address works for ollama and
	21	+ llmman, not for llamacpp.
	22	+
	23	+ ## Why (code read)
	24	+
	25	+ 1. llamacpp declares no fallback ([internal/engine/openai_compat.go](internal/engine/openai_compat.go)):
	26	+ dmr, ollama, llmman and kronk fall back on
	27	+ `http://host.docker.internal:<port>/...` when their default does not
	28	+ answer; llamacpp only has `http://127.0.0.1:8080/v1`. In a container,
	29	+ that is the container itself, and nothing else is tried. The container
	30	+ guide says so ("set `baseUrl: http://host.docker.internal:8080/v1`").
	31	+ The shipped `agent.llamacpp.yaml` even opts out on purpose
	32	+ (`fallback: ""`, "llama-server is wherever you started it").
	33	+ 2. llama-server listens on `127.0.0.1` by default. On Linux and in a
	34	+ Docker Sandbox, `host.docker.internal` is the host's gateway, not its
	35	+ loopback: start it with `--host 0.0.0.0` (as Kronk needs
	36	+ `--api-host 0.0.0.0:11435`). On Docker Desktop (Mac) the loopback is
	37	+ reached anyway (Ollama on 127.0.0.1 works there) — not checked for
	38	+ llama-server.
	39	+ 3. **Docker Sandbox: port 8080 is not in the mini-me kit's network
	40	+ policy** (12434, 11434, 17434 only): `sbx policy allow network
	41	+ host.docker.internal:8080`, or a change in the kit (other repository,
	42	+ mini-me-kit) — out of scope here.
	43	+
	44	+ ## Change (agreed with the user)
	45	+
	46	+ - llamacpp's fallback: `http://host.docker.internal:8080/v1`, the same
	47	+ rule as the others (taken when the default does not answer `/models`
	48	+ within 2 s; `fallback: ""` in a file still turns it off).
	49	+ - Its start hint mentions `--host 0.0.0.0` for mm in a container or a
	50	+ sandbox.
	51	+ - `agent.llamacpp.yaml` drops `fallback: ""`, so that a copy of it gets
	52	+ the fallback. Re-litigates that file's explicit choice — the user asked
	53	+ for llamacpp to behave like the others. `config.samples/` and
	54	+ `demos/struggle-signals/` (pinned to a fake server) are left alone.
	55	+ - `mm init` writes the fallback in the file's comment, as for the others
	56	+ (ticket 121); the wizard's address choice (#214) now also tries it.
	57	+
	58	+ ## Acceptance criteria
	59	+
	60	+ - mm in a container, no `baseUrl`, llama-server on the host started with
	61	+ `--host 0.0.0.0` → mm reaches it through `host.docker.internal:8080`.
	62	+ - On the host, nothing changes when llama-server answers on 127.0.0.1.
	63	+ - Tests and docs EN+FR.
	64	+tasks:
	65	+ - id: 1
	66	+ title: Engine - llamacpp falls back on http://host.docker.internal:8080/v1, like the other local providers
	67	+ state: closed
	68	+ priority: high
	69	+ author:
	70	+ name: k33g
	71	+ email: ph.charriere@gmail.com
	72	+ createdAt: 2026-11-14T18:36:24.000Z
	73	+ updatedAt: 2026-11-14T18:39:34.000Z
	74	+ closedAt: 2026-11-14T18:39:34.000Z
	75	+ - id: 2
	76	+ title: startHint - say --host 0.0.0.0 for mm in a container or a sandbox
	77	+ state: closed
	78	+ priority: medium
	79	+ author:
	80	+ name: k33g
	81	+ email: ph.charriere@gmail.com
	82	+ createdAt: 2026-11-14T18:36:24.000Z
	83	+ updatedAt: 2026-11-14T18:39:34.000Z
	84	+ closedAt: 2026-11-14T18:39:34.000Z
	85	+ - id: 3
	86	+ title: agent.llamacpp.yaml - drop fallback empty (it opted out on purpose)
	87	+ state: closed
	88	+ priority: medium
	89	+ author:
	90	+ name: k33g
	91	+ email: ph.charriere@gmail.com
	92	+ createdAt: 2026-11-14T18:36:24.000Z
	93	+ updatedAt: 2026-11-14T18:39:34.000Z
	94	+ closedAt: 2026-11-14T18:39:34.000Z
	95	+ - id: 4
	96	+ title: Tests - provider resolution, mm init's comment, the wizard's address choice
	97	+ state: closed
	98	+ priority: medium
	99	+ author:
	100	+ name: k33g
	101	+ email: ph.charriere@gmail.com
	102	+ createdAt: 2026-11-14T18:36:24.000Z
	103	+ updatedAt: 2026-11-14T18:39:34.000Z
	104	+ closedAt: 2026-11-14T18:39:34.000Z
	105	+ - id: 5
	106	+ title: Docs EN+FR - configuration (providers table), run-in-a-container, run-with-llama-cpp, sandbox (port 8080 policy)
	107	+ state: closed
	108	+ priority: medium
	109	+ author:
	110	+ name: k33g
	111	+ email: ph.charriere@gmail.com
	112	+ createdAt: 2026-11-14T18:36:24.000Z
	113	+ updatedAt: 2026-11-14T18:39:34.000Z
	114	+ closedAt: 2026-11-14T18:39:34.000Z
	115	+comments:
	116	+ - id: 1
	117	+ author:
	118	+ name: k33g
	119	+ email: ph.charriere@gmail.com
	120	+ createdAt: 2026-11-14T18:39:34.000Z
	121	+ body: \|
	122	+ Implemented on `feature/llamacpp-fallback` (uncommitted).
	123	+
	124	+ - Engine: llamacpp's fallback `http://host.docker.internal:8080/v1`;
	125	+ start hint `… --port 8080 (add --host 0.0.0.0 when mm runs in a
	126	+ container or a sandbox)`.
	127	+ - `agent.llamacpp.yaml`: `fallback: ""` removed, the comment says what
	128	+ the fallback is and how to turn it off. `config.samples/` and
	129	+ `demos/struggle-signals/` keep their `fallback: ""`.
	130	+ - Tests updated where they asserted "no fallback"
	131	+ (`TestResolveURLPrecedence`: falls back when baseUrl is down, keeps
	132	+ it when up; `TestConfigCompat` pins `fallback: ""`;
	133	+ `TestDefaultURLs`; `TestRenderRecordsTheDefaultURLs`: mm init now
	134	+ writes the fallback comment for llamacpp). Whole suite green.
	135	+ - Docs EN+FR: configuration (providers table, agent.llamacpp.yaml),
	136	+ mcp-configuration (EN), cli (warning example), run-in-a-container,
	137	+ run-with-llama-cpp (container / sandbox variant), sandbox guide.
	138	+
	139	+ Side effect, same as dmr/ollama/llmman already: on the host with
	140	+ llama-server down, the start-up warning names the fallback address
	141	+ (`nothing answers at http://host.docker.internal:8080/v1`).
	142	+
	143	+ Not tried with a real llama-server in a container or a sandbox.
	144	+ - id: 2
	145	+ author:
	146	+ name: k33g
	147	+ email: ph.charriere@gmail.com
	148	+ createdAt: 2026-11-14T18:45:13.000Z
	149	+ body: \|
	150	+ Closed at the user's request. mm init writes the fallback comment for
	151	+ llamacpp (checked by the user). In a container, checked only with
	152	+ v1.0.9 (the released binary the image downloads), which does not have
	153	+ the fallback yet: the fix reaches containers with the next release.
	154	+

++# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
++id: 217
++title: "llamacpp: fall back on host.docker.internal:8080 like the other local providers"
++state: closed
++author:
++  name: k33g
++  email: ph.charriere@gmail.com
++createdAt: 2026-11-14T18:36:24.000Z
++updatedAt: 2026-11-14T18:45:13.000Z
++labels:
++  - enhancement
++  - kind::feature
++milestone: v1.1.0
++epic: Providers (LLM Engine)
++related:
++  - 215
++body: |
++  ## Seen (user, 2026-11-14, while testing #215)
++
++  With mm in a container, the fallback address works for **ollama** and
++  **llmman**, not for **llamacpp**.
++
++  ## Why (code read)
++
++  1. **llamacpp declares no fallback** ([internal/engine/openai_compat.go](internal/engine/openai_compat.go)):
++     dmr, ollama, llmman and kronk fall back on
++     `http://host.docker.internal:<port>/...` when their default does not
++     answer; llamacpp only has `http://127.0.0.1:8080/v1`. In a container,
++     that is the container itself, and nothing else is tried. The container
++     guide says so ("set `baseUrl: http://host.docker.internal:8080/v1`").
++     The shipped `agent.llamacpp.yaml` even opts out on purpose
++     (`fallback: ""`, "llama-server is wherever you started it").
++  2. **llama-server listens on `127.0.0.1` by default.** On Linux and in a
++     Docker Sandbox, `host.docker.internal` is the host's gateway, not its
++     loopback: start it with `--host 0.0.0.0` (as Kronk needs
++     `--api-host 0.0.0.0:11435`). On Docker Desktop (Mac) the loopback is
++     reached anyway (Ollama on 127.0.0.1 works there) — not checked for
++     llama-server.
++  3. **Docker Sandbox: port 8080 is not in the mini-me kit's network
++     policy** (12434, 11434, 17434 only): `sbx policy allow network
++     host.docker.internal:8080`, or a change in the kit (other repository,
++     mini-me-kit) — out of scope here.
++
++  ## Change (agreed with the user)
++
++  - llamacpp's fallback: `http://host.docker.internal:8080/v1`, the same
++    rule as the others (taken when the default does not answer `/models`
++    within 2 s; `fallback: ""` in a file still turns it off).
++  - Its start hint mentions `--host 0.0.0.0` for mm in a container or a
++    sandbox.
++  - `agent.llamacpp.yaml` drops `fallback: ""`, so that a copy of it gets
++    the fallback. Re-litigates that file's explicit choice — the user asked
++    for llamacpp to behave like the others. `config.samples/` and
++    `demos/struggle-signals/` (pinned to a fake server) are left alone.
++  - `mm init` writes the fallback in the file's comment, as for the others
++    (ticket 121); the wizard's address choice (#214) now also tries it.
++
++  ## Acceptance criteria
++
++  - mm in a container, no `baseUrl`, llama-server on the host started with
++    `--host 0.0.0.0` → mm reaches it through `host.docker.internal:8080`.
++  - On the host, nothing changes when llama-server answers on 127.0.0.1.
++  - Tests and docs EN+FR.
++tasks:
++  - id: 1
++    title: Engine - llamacpp falls back on http://host.docker.internal:8080/v1, like the other local providers
++    state: closed
++    priority: high
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T18:36:24.000Z
++    updatedAt: 2026-11-14T18:39:34.000Z
++    closedAt: 2026-11-14T18:39:34.000Z
++  - id: 2
++    title: startHint - say --host 0.0.0.0 for mm in a container or a sandbox
++    state: closed
++    priority: medium
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T18:36:24.000Z
++    updatedAt: 2026-11-14T18:39:34.000Z
++    closedAt: 2026-11-14T18:39:34.000Z
++  - id: 3
++    title: agent.llamacpp.yaml - drop fallback empty (it opted out on purpose)
++    state: closed
++    priority: medium
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T18:36:24.000Z
++    updatedAt: 2026-11-14T18:39:34.000Z
++    closedAt: 2026-11-14T18:39:34.000Z
++  - id: 4
++    title: Tests - provider resolution, mm init's comment, the wizard's address choice
++    state: closed
++    priority: medium
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T18:36:24.000Z
++    updatedAt: 2026-11-14T18:39:34.000Z
++    closedAt: 2026-11-14T18:39:34.000Z
++  - id: 5
++    title: Docs EN+FR - configuration (providers table), run-in-a-container, run-with-llama-cpp, sandbox (port 8080 policy)
++    state: closed
++    priority: medium
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T18:36:24.000Z
++    updatedAt: 2026-11-14T18:39:34.000Z
++    closedAt: 2026-11-14T18:39:34.000Z
++comments:
++  - id: 1
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T18:39:34.000Z
++    body: |
++      Implemented on `feature/llamacpp-fallback` (uncommitted).
++
++      - Engine: llamacpp's fallback `http://host.docker.internal:8080/v1`;
++        start hint `… --port 8080 (add --host 0.0.0.0 when mm runs in a
++        container or a sandbox)`.
++      - `agent.llamacpp.yaml`: `fallback: ""` removed, the comment says what
++        the fallback is and how to turn it off. `config.samples/` and
++        `demos/struggle-signals/` keep their `fallback: ""`.
++      - Tests updated where they asserted "no fallback"
++        (`TestResolveURLPrecedence`: falls back when baseUrl is down, keeps
++        it when up; `TestConfigCompat` pins `fallback: ""`;
++        `TestDefaultURLs`; `TestRenderRecordsTheDefaultURLs`: mm init now
++        writes the fallback comment for llamacpp). Whole suite green.
++      - Docs EN+FR: configuration (providers table, agent.llamacpp.yaml),
++        mcp-configuration (EN), cli (warning example), run-in-a-container,
++        run-with-llama-cpp (container / sandbox variant), sandbox guide.
++
++      Side effect, same as dmr/ollama/llmman already: on the host with
++      llama-server down, the start-up warning names the fallback address
++      (`nothing answers at http://host.docker.internal:8080/v1`).
++
++      Not tried with a real llama-server in a container or a sandbox.
++  - id: 2
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T18:45:13.000Z
++    body: |
++      Closed at the user's request. mm init writes the fallback comment for
++      llamacpp (checked by the user). In a container, checked only with
++      v1.0.9 (the released binary the image downloads), which does not have
++      the fallback yet: the fix reaches containers with the next release.
++

added .tickets/issues/0218-llamacpp-llama-server-s-default-port-will-change.yaml +114 -0 Viewed

new file mode 100644
@@ -0,0 +1,114 @@
	1	+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
	2	+id: 218
	3	+title: "llamacpp: llama-server's default port will change from 8080 to 9931"
	4	+state: open
	5	+author:
	6	+ name: k33g
	7	+ email: ph.charriere@gmail.com
	8	+createdAt: 2026-11-14T18:37:23.000Z
	9	+updatedAt: 2026-11-14T18:37:23.000Z
	10	+labels:
	11	+ - kind::chore
	12	+milestone: v2.0.0
	13	+epic: Providers (LLM Engine)
	14	+related:
	15	+ - 217
	16	+body: \|
	17	+ ## Notice (seen by the user, 2026-11-14)
	18	+
	19	+ llama-server prints at start-up:
	20	+
	21	+ ```text
	22	+ llama_server: NOTICE: server default port will be changed to :9931 in a future release
	23	+ ```
	24	+
	25	+ The llama.cpp version that printed it, and the release that will make
	26	+ the change, are not recorded yet.
	27	+
	28	+ ## What depends on 8080 in mini-me (read 2026-11-14)
	29	+
	30	+ - The engine ([internal/engine/openai_compat.go](internal/engine/openai_compat.go)):
	31	+ llamacpp's default address `http://127.0.0.1:8080/v1`, its fallback
	32	+ `http://host.docker.internal:8080/v1` (ticket #217) and its start hint
	33	+ (`llama-server -m <model.gguf> --jinja --port 8080`).
	34	+ - Shipped files: `agent.llamacpp.yaml` (`baseUrl:
	35	+ http://127.0.0.1:8080/v1`), `config.samples/agent.llamacpp.*.yaml`,
	36	+ and the demos run on llama-server (e.g. `demos/qwen-style/`,
	37	+ `demos/decision-skills/`) — to check one by one: most `8080` in
	38	+ `demos/` are web apps the agents build, unrelated.
	39	+ - Docs EN+FR: `how-to/run-with-llama-cpp.md`,
	40	+ `reference/configuration.md` (providers table, `agent.llamacpp.yaml`),
	41	+ `reference/cli.md` (the warning example), `how-to/run-in-a-container.md`,
	42	+ `how-to/run-in-a-docker-sandbox.md` (`sbx policy allow network
	43	+ host.docker.internal:8080`), `how-to/verify-tool-order.md`.
	44	+
	45	+ ## Who is affected
	46	+
	47	+ Only someone who starts llama-server without `--port`: the docs and
	48	+ the start hint always pass `--port 8080`, and a file with an explicit
	49	+ `baseUrl` keeps working. With the new default, an agent with no
	50	+ `baseUrl` would find nothing at 8080.
	51	+
	52	+ ## To decide
	53	+
	54	+ - Follow the new default (`9931`) once the release is out — and say so
	55	+ in the release notes, since files without `baseUrl` change server.
	56	+ - Keep `8080` and keep passing `--port 8080` everywhere.
	57	+ - Or, during a transition, try both on the host: the wizard already
	58	+ tries several local addresses in order (`localAddresses`, #214); the
	59	+ engine's own resolution tries one default and one fallback.
	60	+
	61	+ ## Done when
	62	+
	63	+ The release is identified, the choice is made and written here, and the
	64	+ engine, the shipped files and the docs agree on it.
	65	+tasks:
	66	+ - id: 1
	67	+ title: Find the llama.cpp release that changes the default port, and whether it is already out
	68	+ state: open
	69	+ priority: high
	70	+ author:
	71	+ name: k33g
	72	+ email: ph.charriere@gmail.com
	73	+ createdAt: 2026-11-14T18:37:23.000Z
	74	+ - id: 2
	75	+ title: Decide - follow the new default (9931), keep 8080, or try both during a transition
	76	+ state: open
	77	+ priority: high
	78	+ author:
	79	+ name: k33g
	80	+ email: ph.charriere@gmail.com
	81	+ createdAt: 2026-11-14T18:37:23.000Z
	82	+ - id: 3
	83	+ title: Engine - default address, fallback and start hint of llamacpp
	84	+ state: open
	85	+ priority: medium
	86	+ author:
	87	+ name: k33g
	88	+ email: ph.charriere@gmail.com
	89	+ createdAt: 2026-11-14T18:37:23.000Z
	90	+ - id: 4
	91	+ title: Shipped files - agent.llamacpp.yaml, config.samples/agent.llamacpp.*, demos using llamacpp
	92	+ state: open
	93	+ priority: medium
	94	+ author:
	95	+ name: k33g
	96	+ email: ph.charriere@gmail.com
	97	+ createdAt: 2026-11-14T18:37:23.000Z
	98	+ - id: 5
	99	+ title: Docs EN+FR - run-with-llama-cpp, configuration (providers table), cli (warning example), container and sandbox guides
	100	+ state: open
	101	+ priority: medium
	102	+ author:
	103	+ name: k33g
	104	+ email: ph.charriere@gmail.com
	105	+ createdAt: 2026-11-14T18:37:23.000Z
	106	+ - id: 6
	107	+ title: Docker Sandbox - the port in the mini-me kit's network policy, if 8080 is ever added there (mini-me-kit repository)
	108	+ state: open
	109	+ priority: low
	110	+ author:
	111	+ name: k33g
	112	+ email: ph.charriere@gmail.com
	113	+ createdAt: 2026-11-14T18:37:23.000Z
	114	+comments: []

++# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
++id: 218
++title: "llamacpp: llama-server's default port will change from 8080 to 9931"
++state: open
++author:
++  name: k33g
++  email: ph.charriere@gmail.com
++createdAt: 2026-11-14T18:37:23.000Z
++updatedAt: 2026-11-14T18:37:23.000Z
++labels:
++  - kind::chore
++milestone: v2.0.0
++epic: Providers (LLM Engine)
++related:
++  - 217
++body: |
++  ## Notice (seen by the user, 2026-11-14)
++
++  llama-server prints at start-up:
++
++  ```text
++  llama_server: NOTICE: server default port will be changed to :9931 in a future release
++  ```
++
++  The llama.cpp version that printed it, and the release that will make
++  the change, are not recorded yet.
++
++  ## What depends on 8080 in mini-me (read 2026-11-14)
++
++  - **The engine** ([internal/engine/openai_compat.go](internal/engine/openai_compat.go)):
++    llamacpp's default address `http://127.0.0.1:8080/v1`, its fallback
++    `http://host.docker.internal:8080/v1` (ticket #217) and its start hint
++    (`llama-server -m <model.gguf> --jinja --port 8080`).
++  - **Shipped files**: `agent.llamacpp.yaml` (`baseUrl:
++    http://127.0.0.1:8080/v1`), `config.samples/agent.llamacpp.*.yaml`,
++    and the demos run on llama-server (e.g. `demos/qwen-style/`,
++    `demos/decision-skills/`) — to check one by one: most `8080` in
++    `demos/` are web apps the agents build, unrelated.
++  - **Docs EN+FR**: `how-to/run-with-llama-cpp.md`,
++    `reference/configuration.md` (providers table, `agent.llamacpp.yaml`),
++    `reference/cli.md` (the warning example), `how-to/run-in-a-container.md`,
++    `how-to/run-in-a-docker-sandbox.md` (`sbx policy allow network
++    host.docker.internal:8080`), `how-to/verify-tool-order.md`.
++
++  ## Who is affected
++
++  Only someone who starts llama-server **without** `--port`: the docs and
++  the start hint always pass `--port 8080`, and a file with an explicit
++  `baseUrl` keeps working. With the new default, an agent with no
++  `baseUrl` would find nothing at 8080.
++
++  ## To decide
++
++  - Follow the new default (`9931`) once the release is out — and say so
++    in the release notes, since files without `baseUrl` change server.
++  - Keep `8080` and keep passing `--port 8080` everywhere.
++  - Or, during a transition, try both on the host: the wizard already
++    tries several local addresses in order (`localAddresses`, #214); the
++    engine's own resolution tries one default and one fallback.
++
++  ## Done when
++
++  The release is identified, the choice is made and written here, and the
++  engine, the shipped files and the docs agree on it.
++tasks:
++  - id: 1
++    title: Find the llama.cpp release that changes the default port, and whether it is already out
++    state: open
++    priority: high
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T18:37:23.000Z
++  - id: 2
++    title: Decide - follow the new default (9931), keep 8080, or try both during a transition
++    state: open
++    priority: high
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T18:37:23.000Z
++  - id: 3
++    title: Engine - default address, fallback and start hint of llamacpp
++    state: open
++    priority: medium
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T18:37:23.000Z
++  - id: 4
++    title: Shipped files - agent.llamacpp.yaml, config.samples/agent.llamacpp.*, demos using llamacpp
++    state: open
++    priority: medium
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T18:37:23.000Z
++  - id: 5
++    title: Docs EN+FR - run-with-llama-cpp, configuration (providers table), cli (warning example), container and sandbox guides
++    state: open
++    priority: medium
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T18:37:23.000Z
++  - id: 6
++    title: Docker Sandbox - the port in the mini-me kit's network policy, if 8080 is ever added there (mini-me-kit repository)
++    state: open
++    priority: low
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-11-14T18:37:23.000Z
++comments: []

modified agent.llamacpp.yaml +4 -2 Viewed

@@ -9,8 +9,10 @@ model: jetbrains/mellum2-12b-a2.5b-instruct-gguf-q4_k_m:Q4_K_M
9	9	# llama-server's default is 127.0.0.1:8080; the OpenAI routes sit under /v1.
10	10	baseUrl: http://127.0.0.1:8080/v1
11	11
12		-# No fallback: llama-server is wherever you started it.
13		-fallback: ""
	12	+# No fallback line: when baseUrl does not answer, mm tries llama-server on the
	13	+# host, http://host.docker.internal:8080/v1 — what mm in a container or a
	14	+# sandbox needs (start llama-server with --host 0.0.0.0 there). fallback: ""
	15	+# turns that off.
14	16
15	17	# Only when llama-server was started with --api-key; unset otherwise.
16	18	# apiKeyEnv: LLAMA_API_KEY

  # llama-server's default is 127.0.0.1:8080; the OpenAI routes sit under /v1.
  baseUrl: http://127.0.0.1:8080/v1
--# No fallback: llama-server is wherever you started it.
++# No fallback line: when baseUrl does not answer, mm tries llama-server on the
--fallback: ""
++# host, http://host.docker.internal:8080/v1 — what mm in a container or a
++# sandbox needs (start llama-server with --host 0.0.0.0 there). fallback: ""
++# turns that off.
  # Only when llama-server was started with --api-key; unset otherwise.
  # apiKeyEnv: LLAMA_API_KEY

modified docs/en/how-to/run-in-a-container.md +2 -2 Viewed

@@ -68,9 +68,9 @@ The project is mounted at the same path as on your machine, so the paths the edi
68	68
69	69	## Reach the model
70	70
71		-From the container, `localhost` is the container. `dmr`, `llmman` and `ollama` fall back on `host.docker.internal` by themselves when their default address does not answer, as long as the agent file sets no `baseUrl`. The compose file makes that name known on Linux too.
	71	+From the container, `localhost` is the container. `dmr`, `llmman`, `ollama`, `kronk` and `llamacpp` fall back on `host.docker.internal` by themselves when their default address does not answer, as long as the agent file sets no `baseUrl`. The compose file makes that name known on Linux too.
72	72
73		-- `llamacpp` has no such fallback: set `baseUrl: http://host.docker.internal:8080/v1`.
	73	+- `llamacpp`: on Linux, start llama-server with `--host 0.0.0.0`, as it listens on the loopback only by default (see the last point). Files written before v1.1.0 may say `fallback: ""`: remove that line.
74	74	- An agent file with a `baseUrl` on `localhost` or `127.0.0.1`: replace the host with `host.docker.internal`, or add a `fallback`.
75	75	- A server on Linux that only listens on `127.0.0.1`: make it listen on an address the container can reach. `host.docker.internal` is then the host's gateway, not its loopback.
76	76

  ## Reach the model
--From the container, `localhost` is the container. `dmr`, `llmman` and `ollama` fall back on `host.docker.internal` by themselves when their default address does not answer, as long as the agent file sets no `baseUrl`. The compose file makes that name known on Linux too.
++From the container, `localhost` is the container. `dmr`, `llmman`, `ollama`, `kronk` and `llamacpp` fall back on `host.docker.internal` by themselves when their default address does not answer, as long as the agent file sets no `baseUrl`. The compose file makes that name known on Linux too.
--- **`llamacpp`** has no such fallback: set `baseUrl: http://host.docker.internal:8080/v1`.
++- **`llamacpp`**: on Linux, start llama-server with `--host 0.0.0.0`, as it listens on the loopback only by default (see the last point). Files written before v1.1.0 may say `fallback: ""`: remove that line.
  - **An agent file with a `baseUrl` on `localhost` or `127.0.0.1`:** replace the host with `host.docker.internal`, or add a `fallback`.
  - **A server on Linux that only listens on `127.0.0.1`:** make it listen on an address the container can reach. `host.docker.internal` is then the host's gateway, not its loopback.

modified docs/en/how-to/run-in-a-docker-sandbox.md +1 -1 Viewed

@@ -203,7 +203,7 @@ Inside a sandbox `localhost` is the sandbox itself. Your model server keeps runn
203	203
204	204	- The `mini-me` kit's network policy allows `host.docker.internal:12434` (Docker Model Runner), `:11434` (Ollama) and `:17434` (llmman). `agent.dmr.yaml` already falls back to `http://host.docker.internal:12434/engines/v1`; the configurations shipped in `demos/use-with-sbx/` use llmman (`fallback: http://host.docker.internal:17434/v1`).
205	205	- Kronk (`:11435`) is not in that list: allow `localhost:11435` in the policy, and start Kronk on the host with `--api-host 0.0.0.0:11435` — it listens on the loopback only otherwise. See [How to run mini-me on Kronk](run-with-kronk.md).
206		-- For another port — llama-server's `8080`, for instance — add it to the kit's `network-policy` capability, or allow it for one sandbox:
	206	+- For another port — llama-server's `8080`, for instance, which mm tries on its own (`http://host.docker.internal:8080/v1`) when started on the host with `--host 0.0.0.0` — add it to the kit's `network-policy` capability, or allow it for one sandbox:
207	207
208	208	```bash
209	209	sbx policy allow network host.docker.internal:8080

  - The `mini-me` kit's network policy allows **`host.docker.internal:12434`** (Docker Model Runner), **`:11434`** (Ollama) and **`:17434`** (llmman). `agent.dmr.yaml` already falls back to `http://host.docker.internal:12434/engines/v1`; the configurations shipped in `demos/use-with-sbx/` use llmman (`fallback: http://host.docker.internal:17434/v1`).
  - Kronk (`:11435`) is not in that list: allow `localhost:11435` in the policy, and start Kronk on the host with `--api-host 0.0.0.0:11435` — it listens on the loopback only otherwise. See [How to run mini-me on Kronk](run-with-kronk.md).
--- For another port — llama-server's `8080`, for instance — add it to the kit's `network-policy` capability, or allow it for one sandbox:
++- For another port — llama-server's `8080`, for instance, which mm tries on its own (`http://host.docker.internal:8080/v1`) when started on the host with `--host 0.0.0.0` — add it to the kit's `network-policy` capability, or allow it for one sandbox:
    ```bash
    sbx policy allow network host.docker.internal:8080

modified docs/en/how-to/run-with-llama-cpp.md +1 -0 Viewed

@@ -23,6 +23,7 @@ This guide shows how to point `mm` at a `llama-server` you start yourself. It as
23	23	## Variants
24	24
25	25	- Different port or host. Set `baseUrl` in the YAML to `http://<host>:<port>/v1`, or override it for one run with `AGENT_BASE_URL=http://<host>:<port>/v1 ./mm agent.llamacpp.yaml`. An environment override is taken as-is, without the fallback probe.
	26	+- From a container or a Docker Sandbox. When `baseUrl` does not answer, mm tries `http://host.docker.internal:8080/v1`, llama-server on the host. Start it there with `--host 0.0.0.0`: by default it listens on the loopback only, which a container on Linux or a sandbox cannot reach. In a sandbox, also allow the port: `sbx policy allow network host.docker.internal:8080`. `fallback: ""` turns the fallback off. See [run mini-me in a container](run-in-a-container.md) and [run mini-me in a Docker Sandbox](run-in-a-docker-sandbox.md).
26	27	- `autoPull: true`. llama-server cannot download a model over its API: it serves the one it was started with. `-hf <repo>:<quant>` (step 1) is what downloads it. With `autoPull` on and a `model:` the server does not list, mm says so at start-up and goes on. See [`autoPull`](../reference/configuration.md).
27	28	- Server started with `--api-key`. Export the key in `LLAMA_API_KEY` (the provider's default variable) or name another variable under `apiKeyEnv`. The key itself never goes in the YAML.
28	29	- Keep `agent.yaml` and switch provider for one run. `AGENT_PROVIDER=llamacpp ./mm` works as long as `agent.yaml` leaves `baseUrl` empty; an explicit DMR URL in the file would still be honoured and point at port 12434.

  ## Variants
  - **Different port or host.** Set `baseUrl` in the YAML to `http://<host>:<port>/v1`, or override it for one run with `AGENT_BASE_URL=http://<host>:<port>/v1 ./mm agent.llamacpp.yaml`. An environment override is taken as-is, without the fallback probe.
++- **From a container or a Docker Sandbox.** When `baseUrl` does not answer, mm tries `http://host.docker.internal:8080/v1`, llama-server on the host. Start it there with `--host 0.0.0.0`: by default it listens on the loopback only, which a container on Linux or a sandbox cannot reach. In a sandbox, also allow the port: `sbx policy allow network host.docker.internal:8080`. `fallback: ""` turns the fallback off. See [run mini-me in a container](run-in-a-container.md) and [run mini-me in a Docker Sandbox](run-in-a-docker-sandbox.md).
  - **`autoPull: true`.** llama-server cannot download a model over its API: it serves the one it was started with. `-hf <repo>:<quant>` (step 1) is what downloads it. With `autoPull` on and a `model:` the server does not list, mm says so at start-up and goes on. See [`autoPull`](../reference/configuration.md).
  - **Server started with `--api-key`.** Export the key in `LLAMA_API_KEY` (the provider's default variable) or name another variable under `apiKeyEnv`. The key itself never goes in the YAML.
  - **Keep `agent.yaml` and switch provider for one run.** `AGENT_PROVIDER=llamacpp ./mm` works as long as `agent.yaml` leaves `baseUrl` empty; an explicit DMR URL in the file would still be honoured and point at port 12434.

modified docs/en/reference/cli.md +2 -2 Viewed

@@ -104,7 +104,7 @@ Printed on stdout in terminal mode, on stderr in ACP mode. Dimmed when stdout is
104	104	### Example
105	105
106	106	```
107		-[warning: llamacpp: nothing answers at http://127.0.0.1:8080/v1 — start it: llama-server -m <model.gguf> --jinja --port 8080]
	107	+[warning: llamacpp: nothing answers at http://host.docker.internal:8080/v1 — start it: llama-server -m <model.gguf> --jinja --port 8080 (add --host 0.0.0.0 when mm runs in a container or a sandbox)]
108	108	mm v0.1.7 \| config: agent.llamacpp.yaml \| provider: llamacpp \| model: jetbrains/mellum2-12b-a2.5b-instruct-gguf-q4_k_m:Q4_K_M \| ctx: unknown \| tools: bash, read_file, write_file, edit_file \| skills: 0 \| yolo: on
109	109	```
110	110
@@ -240,7 +240,7 @@ An agent file outside the workspace only gives a warning: see [when the agent fi
240	240	\|------\|----------------------\|---------\|
241	241	\| `--project yes\\|no` \| Keep everything in a `.mm` folder? — asked first. `yes`: the file is `.mm/agent.yaml` (the folder is created), its paths are written from the project root (`skillsDir: .mm/skills`), and `mm` / `mm tui` started there find it with no argument \| `no` \|
242	242	\| `--provider` \| Model server: `llmman`, `ollama`, `dmr`, `llamacpp`, `kronk`, `openai` \| the first local one whose server answers, else `llmman` (`openai` is never probed: its default address is the OpenAI API) \|
243		-\| `--base-url` \| `openai` only: Endpoint — the service's OpenAI-compatible address (OpenAI, Cerebras, Hugging Face…); for the other providers, not asked \| `https://api.openai.com/v1` for `openai`; otherwise the provider's address, or its fallback when it does not answer. Written as `baseUrl` unless it is the default \|
	243	+\| `--base-url` \| `openai` only: Endpoint — the service's OpenAI-compatible address (OpenAI, Cerebras, Hugging Face…); for the other providers, not asked \| `https://api.openai.com/v1` for `openai`; otherwise `AGENT_BASE_URL` when set, else the first of the provider's local addresses whose model list answers — its default, then the same with `127.0.0.1` when the default names `localhost` — and its container fallback (`host.docker.internal`) only when none does. The model question names the address its list comes from (`Model? (listed from …)`). Written as `baseUrl` unless it is the default \|
244	244	\| `--api-key-env` \| `openai` only: Environment variable that holds the API key \| `OPENAI_API_KEY`. Written as `apiKeyEnv` unless it is the default \|
245	245	\| `--model` \| Model: a list of the server's models (`/api/tags` for llmman and Ollama, `/models` for the others; 5 s for a local server, 20 s for `openai`), or a name to type when there is no list — the question says why (`<VARIABLE> is not set: no list`, or the server's error). For `openai` the list is asked with the key read from the variable named above, and not asked at all when that variable is empty \| the first of the list \|
246	246	\| `--family` \| Model family \| detected from the model's name, then from the server's family name \|

  ### Example
  ```
--[warning: llamacpp: nothing answers at http://127.0.0.1:8080/v1 — start it: llama-server -m <model.gguf> --jinja --port 8080]
++[warning: llamacpp: nothing answers at http://host.docker.internal:8080/v1 — start it: llama-server -m <model.gguf> --jinja --port 8080 (add --host 0.0.0.0 when mm runs in a container or a sandbox)]
  mm v0.1.7 | config: agent.llamacpp.yaml | provider: llamacpp | model: jetbrains/mellum2-12b-a2.5b-instruct-gguf-q4_k_m:Q4_K_M | ctx: unknown | tools: bash, read_file, write_file, edit_file | skills: 0 | yolo: on
  ```
  |------|----------------------|---------|
  | `--project yes\|no` | Keep everything in a `.mm` folder? — asked first. `yes`: the file is `.mm/agent.yaml` (the folder is created), its paths are written from the project root (`skillsDir: .mm/skills`), and `mm` / `mm tui` started there find it with no argument | `no` |
  | `--provider` | Model server: `llmman`, `ollama`, `dmr`, `llamacpp`, `kronk`, `openai` | the first local one whose server answers, else `llmman` (`openai` is never probed: its default address is the OpenAI API) |
--| `--base-url` | `openai` only: Endpoint — the service's OpenAI-compatible address (OpenAI, Cerebras, Hugging Face…); for the other providers, not asked | `https://api.openai.com/v1` for `openai`; otherwise the provider's address, or its fallback when it does not answer. Written as `baseUrl` unless it is the default |
++| `--base-url` | `openai` only: Endpoint — the service's OpenAI-compatible address (OpenAI, Cerebras, Hugging Face…); for the other providers, not asked | `https://api.openai.com/v1` for `openai`; otherwise `AGENT_BASE_URL` when set, else the first of the provider's local addresses whose model list answers — its default, then the same with `127.0.0.1` when the default names `localhost` — and its container fallback (`host.docker.internal`) only when none does. The model question names the address its list comes from (`Model? (listed from …)`). Written as `baseUrl` unless it is the default |
  | `--api-key-env` | `openai` only: Environment variable that holds the API key | `OPENAI_API_KEY`. Written as `apiKeyEnv` unless it is the default |
  | `--model` | Model: a list of the server's models (`/api/tags` for llmman and Ollama, `/models` for the others; 5 s for a local server, 20 s for `openai`), or a name to type when there is no list — the question says why (`<VARIABLE> is not set: no list`, or the server's error). For `openai` the list is asked with the key read from the variable named above, and not asked at all when that variable is empty | the first of the list |
  | `--family` | Model family | detected from the model's name, then from the server's family name |

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

@@ -197,7 +197,7 @@ tools:
197	197	\| Provider \| `baseUrl` \| `fallback` \| `apiKeyEnv` \| Key required \|
198	198	\|----------\|-----------\|------------\|-------------\|--------------\|
199	199	\| `dmr` \| `http://localhost:12434/engines/v1` \| `http://host.docker.internal:12434/engines/v1` \| none \| no \|
200		-\| `llamacpp` \| `http://127.0.0.1:8080/v1` \| none \| `LLAMA_API_KEY` \| no \|
	200	+\| `llamacpp` \| `http://127.0.0.1:8080/v1` \| `http://host.docker.internal:8080/v1` \| `LLAMA_API_KEY` \| no \|
201	201	\| `llmman` \| `http://127.0.0.1:17434/v1` \| `http://host.docker.internal:17434/v1` \| `LLMMAN_API_KEY` \| no \|
202	202	\| `ollama` \| `http://localhost:11434/v1` \| `http://host.docker.internal:11434/v1` \| `OLLAMA_API_KEY` \| no \|
203	203	\| `kronk` \| `http://127.0.0.1:11435/v1` \| `http://host.docker.internal:11435/v1` \| `KRONK_TOKEN` \| no — only Kronk's `authenticated` and `full-protected` modes check it \|
@@ -477,7 +477,7 @@ A start-up warning, not an error, is printed when `context.enabled` is `true`, t
477	477	\| File \| Purpose \|
478	478	\|------\|---------\|
479	479	\| `agent.yaml` \| Docker Model Runner set-up: `provider: dmr`, `maxTurns: 40`, `displayCommands: true`, compression off, a system prompt named "Bob" that describes the skills and file-editing rules. \|
480		-\| `agent.llamacpp.yaml` \| llama.cpp set-up: `provider: llamacpp`, `baseUrl: http://127.0.0.1:8080/v1`, `fallback: ""`, compression on. \|
	480	+\| `agent.llamacpp.yaml` \| llama.cpp set-up: `provider: llamacpp`, `baseUrl: http://127.0.0.1:8080/v1`, compression on. \|
481	481
482	482	### Example
483	483

  | Provider | `baseUrl` | `fallback` | `apiKeyEnv` | Key required |
  |----------|-----------|------------|-------------|--------------|
  | `dmr` | `http://localhost:12434/engines/v1` | `http://host.docker.internal:12434/engines/v1` | none | no |
--| `llamacpp` | `http://127.0.0.1:8080/v1` | none | `LLAMA_API_KEY` | no |
++| `llamacpp` | `http://127.0.0.1:8080/v1` | `http://host.docker.internal:8080/v1` | `LLAMA_API_KEY` | no |
  | `llmman` | `http://127.0.0.1:17434/v1` | `http://host.docker.internal:17434/v1` | `LLMMAN_API_KEY` | no |
  | `ollama` | `http://localhost:11434/v1` | `http://host.docker.internal:11434/v1` | `OLLAMA_API_KEY` | no |
  | `kronk` | `http://127.0.0.1:11435/v1` | `http://host.docker.internal:11435/v1` | `KRONK_TOKEN` | no — only Kronk's `authenticated` and `full-protected` modes check it |
  | File | Purpose |
  |------|---------|
  | `agent.yaml` | Docker Model Runner set-up: `provider: dmr`, `maxTurns: 40`, `displayCommands: true`, compression off, a system prompt named "Bob" that describes the skills and file-editing rules. |
--| `agent.llamacpp.yaml` | llama.cpp set-up: `provider: llamacpp`, `baseUrl: http://127.0.0.1:8080/v1`, `fallback: ""`, compression on. |
++| `agent.llamacpp.yaml` | llama.cpp set-up: `provider: llamacpp`, `baseUrl: http://127.0.0.1:8080/v1`, compression on. |
  ### Example

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

@@ -78,7 +78,7 @@ The shipped `agent.lsp.yaml` is `agent.dmr.yaml` with this block on (gopls activ
78	78	\| Provider \| `baseUrl` \| `fallback` \| `apiKeyEnv` \| Key required \|
79	79	\|----------\|-----------\|------------\|-------------\|--------------\|
80	80	\| `dmr` \| `http://localhost:12434/engines/v1` \| `http://host.docker.internal:12434/engines/v1` \| none \| no \|
81		-\| `llamacpp` \| `http://127.0.0.1:8080/v1` \| none \| `LLAMA_API_KEY` \| no \|
	81	+\| `llamacpp` \| `http://127.0.0.1:8080/v1` \| `http://host.docker.internal:8080/v1` \| `LLAMA_API_KEY` \| no \|
82	82	\| `llmman` \| `http://127.0.0.1:17434/v1` \| `http://host.docker.internal:17434/v1` \| `LLMMAN_API_KEY` \| no \|
83	83	\| `ollama` \| `http://localhost:11434/v1` \| `http://host.docker.internal:11434/v1` \| `OLLAMA_API_KEY` \| no \|
84	84
@@ -114,7 +114,7 @@ A start-up warning, not an error, is printed when `context.enabled` is `true`, t
114	114	\| File \| Purpose \|
115	115	\|------\|---------\|
116	116	\| `agent.yaml` \| Docker Model Runner set-up: `provider: dmr`, `maxTurns: 40`, `displayCommands: true`, compression off, a system prompt named "Bob" that describes the skills and file-editing rules. \|
117		-\| `agent.llamacpp.yaml` \| llama.cpp set-up: `provider: llamacpp`, `baseUrl: http://127.0.0.1:8080/v1`, `fallback: ""`, compression on. \|
	117	+\| `agent.llamacpp.yaml` \| llama.cpp set-up: `provider: llamacpp`, `baseUrl: http://127.0.0.1:8080/v1`, compression on. \|
118	118
119	119	### Example
120	120

  | Provider | `baseUrl` | `fallback` | `apiKeyEnv` | Key required |
  |----------|-----------|------------|-------------|--------------|
  | `dmr` | `http://localhost:12434/engines/v1` | `http://host.docker.internal:12434/engines/v1` | none | no |
--| `llamacpp` | `http://127.0.0.1:8080/v1` | none | `LLAMA_API_KEY` | no |
++| `llamacpp` | `http://127.0.0.1:8080/v1` | `http://host.docker.internal:8080/v1` | `LLAMA_API_KEY` | no |
  | `llmman` | `http://127.0.0.1:17434/v1` | `http://host.docker.internal:17434/v1` | `LLMMAN_API_KEY` | no |
  | `ollama` | `http://localhost:11434/v1` | `http://host.docker.internal:11434/v1` | `OLLAMA_API_KEY` | no |
  | File | Purpose |
  |------|---------|
  | `agent.yaml` | Docker Model Runner set-up: `provider: dmr`, `maxTurns: 40`, `displayCommands: true`, compression off, a system prompt named "Bob" that describes the skills and file-editing rules. |
--| `agent.llamacpp.yaml` | llama.cpp set-up: `provider: llamacpp`, `baseUrl: http://127.0.0.1:8080/v1`, `fallback: ""`, compression on. |
++| `agent.llamacpp.yaml` | llama.cpp set-up: `provider: llamacpp`, `baseUrl: http://127.0.0.1:8080/v1`, compression on. |
  ### Example

modified docs/fr/how-to/run-in-a-container.md +2 -2 Viewed

@@ -68,9 +68,9 @@ Le projet est monté au même chemin que sur votre machine : les chemins envoyé
68	68
69	69	## Joindre le modèle
70	70
71		-Depuis le conteneur, `localhost` désigne le conteneur. `dmr`, `llmman` et `ollama` se rabattent d'eux-mêmes sur `host.docker.internal` quand leur adresse par défaut ne répond pas, tant que le fichier d'agent ne définit pas de `baseUrl`. Le fichier compose fait aussi connaître ce nom sous Linux.
	71	+Depuis le conteneur, `localhost` désigne le conteneur. `dmr`, `llmman`, `ollama`, `kronk` et `llamacpp` se rabattent d'eux-mêmes sur `host.docker.internal` quand leur adresse par défaut ne répond pas, tant que le fichier d'agent ne définit pas de `baseUrl`. Le fichier compose fait aussi connaître ce nom sous Linux.
72	72
73		-- `llamacpp` n'a pas ce repli : définissez `baseUrl: http://host.docker.internal:8080/v1`.
	73	+- `llamacpp` : sous Linux, lancez llama-server avec `--host 0.0.0.0`, car par défaut il n'écoute que sur la boucle locale (voir le dernier point). Les fichiers écrits avant la v1.1.0 peuvent contenir `fallback: ""` : retirez cette ligne.
74	74	- Un fichier d'agent avec un `baseUrl` sur `localhost` ou `127.0.0.1` : remplacez l'hôte par `host.docker.internal`, ou ajoutez un `fallback`.
75	75	- Un serveur sous Linux qui n'écoute que sur `127.0.0.1` : faites-le écouter sur une adresse joignable par le conteneur. `host.docker.internal` y désigne la passerelle de l'hôte, pas sa boucle locale.
76	76

  ## Joindre le modèle
--Depuis le conteneur, `localhost` désigne le conteneur. `dmr`, `llmman` et `ollama` se rabattent d'eux-mêmes sur `host.docker.internal` quand leur adresse par défaut ne répond pas, tant que le fichier d'agent ne définit pas de `baseUrl`. Le fichier compose fait aussi connaître ce nom sous Linux.
++Depuis le conteneur, `localhost` désigne le conteneur. `dmr`, `llmman`, `ollama`, `kronk` et `llamacpp` se rabattent d'eux-mêmes sur `host.docker.internal` quand leur adresse par défaut ne répond pas, tant que le fichier d'agent ne définit pas de `baseUrl`. Le fichier compose fait aussi connaître ce nom sous Linux.
--- **`llamacpp`** n'a pas ce repli : définissez `baseUrl: http://host.docker.internal:8080/v1`.
++- **`llamacpp`** : sous Linux, lancez llama-server avec `--host 0.0.0.0`, car par défaut il n'écoute que sur la boucle locale (voir le dernier point). Les fichiers écrits avant la v1.1.0 peuvent contenir `fallback: ""` : retirez cette ligne.
  - **Un fichier d'agent avec un `baseUrl` sur `localhost` ou `127.0.0.1` :** remplacez l'hôte par `host.docker.internal`, ou ajoutez un `fallback`.
  - **Un serveur sous Linux qui n'écoute que sur `127.0.0.1` :** faites-le écouter sur une adresse joignable par le conteneur. `host.docker.internal` y désigne la passerelle de l'hôte, pas sa boucle locale.

modified docs/fr/how-to/run-in-a-docker-sandbox.md +1 -1 Viewed

@@ -203,7 +203,7 @@ Dans un sandbox, `localhost` est le sandbox lui-même. Votre serveur de modèles
203	203
204	204	- La politique réseau du kit `mini-me` autorise `host.docker.internal:12434` (Docker Model Runner), `:11434` (Ollama) et `:17434` (llmman). `agent.dmr.yaml` se replie déjà sur `http://host.docker.internal:12434/engines/v1` ; les configurations livrées dans `demos/use-with-sbx/` utilisent llmman (`fallback: http://host.docker.internal:17434/v1`).
205	205	- Kronk (`:11435`) n'est pas dans cette liste : autorisez `localhost:11435` dans la politique, et lancez Kronk sur l'hôte avec `--api-host 0.0.0.0:11435` — sinon il n'écoute que sur la boucle locale. Voir [Faire tourner mini-me sur Kronk](run-with-kronk.md).
206		-- Pour un autre port — le `8080` de llama-server, par exemple — ajoutez-le à la capacité `network-policy` du kit, ou autorisez-le pour un seul sandbox :
	206	+- Pour un autre port — le `8080` de llama-server, par exemple, que mm essaie de lui-même (`http://host.docker.internal:8080/v1`) quand il est lancé sur l'hôte avec `--host 0.0.0.0` — ajoutez-le à la capacité `network-policy` du kit, ou autorisez-le pour un seul sandbox :
207	207
208	208	```bash
209	209	sbx policy allow network host.docker.internal:8080

  - La politique réseau du kit `mini-me` autorise **`host.docker.internal:12434`** (Docker Model Runner), **`:11434`** (Ollama) et **`:17434`** (llmman). `agent.dmr.yaml` se replie déjà sur `http://host.docker.internal:12434/engines/v1` ; les configurations livrées dans `demos/use-with-sbx/` utilisent llmman (`fallback: http://host.docker.internal:17434/v1`).
  - Kronk (`:11435`) n'est pas dans cette liste : autorisez `localhost:11435` dans la politique, et lancez Kronk sur l'hôte avec `--api-host 0.0.0.0:11435` — sinon il n'écoute que sur la boucle locale. Voir [Faire tourner mini-me sur Kronk](run-with-kronk.md).
--- Pour un autre port — le `8080` de llama-server, par exemple — ajoutez-le à la capacité `network-policy` du kit, ou autorisez-le pour un seul sandbox :
++- Pour un autre port — le `8080` de llama-server, par exemple, que mm essaie de lui-même (`http://host.docker.internal:8080/v1`) quand il est lancé sur l'hôte avec `--host 0.0.0.0` — ajoutez-le à la capacité `network-policy` du kit, ou autorisez-le pour un seul sandbox :
    ```bash
    sbx policy allow network host.docker.internal:8080

modified docs/fr/how-to/run-with-llama-cpp.md +1 -0 Viewed

@@ -23,6 +23,7 @@ Ce guide montre comment pointer `mm` vers un `llama-server` que vous démarrez v
23	23	## Variantes
24	24
25	25	- Autre port ou autre hôte. Mettez `baseUrl` à `http://<hôte>:<port>/v1` dans le YAML, ou surchargez pour une seule exécution avec `AGENT_BASE_URL=http://<hôte>:<port>/v1 ./mm agent.llamacpp.yaml`. Une surcharge par l'environnement est prise telle quelle, sans sonde de repli.
	26	+- Depuis un conteneur ou un Docker Sandbox. Quand `baseUrl` ne répond pas, mm essaie `http://host.docker.internal:8080/v1`, llama-server sur l'hôte. Lancez-le avec `--host 0.0.0.0` : par défaut il n'écoute que sur la boucle locale, qu'un conteneur sous Linux ou un sandbox ne peuvent pas joindre. Dans un sandbox, autorisez aussi le port : `sbx policy allow network host.docker.internal:8080`. `fallback: ""` coupe le repli. Voir [faire tourner mini-me dans un conteneur](run-in-a-container.md) et [faire tourner mini-me dans un Docker Sandbox](run-in-a-docker-sandbox.md).
26	27	- `autoPull: true`. llama-server ne sait pas télécharger un modèle par son API : il sert celui avec lequel il a été lancé. C'est `-hf <dépôt>:<quant>` (étape 1) qui le télécharge. Avec `autoPull` actif et un `model:` que le serveur ne liste pas, mm le dit au démarrage et continue. Voir [`autoPull`](../reference/configuration.md).
27	28	- Serveur démarré avec `--api-key`. Exportez la clé dans `LLAMA_API_KEY` (la variable par défaut du fournisseur) ou nommez une autre variable dans `apiKeyEnv`. La clé elle-même ne va jamais dans le YAML.
28	29	- Garder `agent.yaml` et changer de fournisseur pour une exécution. `AGENT_PROVIDER=llamacpp ./mm` fonctionne tant qu'`agent.yaml` laisse `baseUrl` vide ; une URL DMR explicite dans le fichier serait respectée et pointerait toujours sur le port 12434.

  ## Variantes
  - **Autre port ou autre hôte.** Mettez `baseUrl` à `http://<hôte>:<port>/v1` dans le YAML, ou surchargez pour une seule exécution avec `AGENT_BASE_URL=http://<hôte>:<port>/v1 ./mm agent.llamacpp.yaml`. Une surcharge par l'environnement est prise telle quelle, sans sonde de repli.
++- **Depuis un conteneur ou un Docker Sandbox.** Quand `baseUrl` ne répond pas, mm essaie `http://host.docker.internal:8080/v1`, llama-server sur l'hôte. Lancez-le avec `--host 0.0.0.0` : par défaut il n'écoute que sur la boucle locale, qu'un conteneur sous Linux ou un sandbox ne peuvent pas joindre. Dans un sandbox, autorisez aussi le port : `sbx policy allow network host.docker.internal:8080`. `fallback: ""` coupe le repli. Voir [faire tourner mini-me dans un conteneur](run-in-a-container.md) et [faire tourner mini-me dans un Docker Sandbox](run-in-a-docker-sandbox.md).
  - **`autoPull: true`.** llama-server ne sait pas télécharger un modèle par son API : il sert celui avec lequel il a été lancé. C'est `-hf <dépôt>:<quant>` (étape 1) qui le télécharge. Avec `autoPull` actif et un `model:` que le serveur ne liste pas, mm le dit au démarrage et continue. Voir [`autoPull`](../reference/configuration.md).
  - **Serveur démarré avec `--api-key`.** Exportez la clé dans `LLAMA_API_KEY` (la variable par défaut du fournisseur) ou nommez une autre variable dans `apiKeyEnv`. La clé elle-même ne va jamais dans le YAML.
  - **Garder `agent.yaml` et changer de fournisseur pour une exécution.** `AGENT_PROVIDER=llamacpp ./mm` fonctionne tant qu'`agent.yaml` laisse `baseUrl` vide ; une URL DMR explicite dans le fichier serait respectée et pointerait toujours sur le port 12434.

modified docs/fr/reference/cli.md +2 -2 Viewed

@@ -104,7 +104,7 @@ Affichée sur stdout en mode terminal, sur stderr en mode ACP. Grisée quand std
104	104	### Exemple
105	105
106	106	```
107		-[warning: llamacpp: nothing answers at http://127.0.0.1:8080/v1 — start it: llama-server -m <model.gguf> --jinja --port 8080]
	107	+[warning: llamacpp: nothing answers at http://host.docker.internal:8080/v1 — start it: llama-server -m <model.gguf> --jinja --port 8080 (add --host 0.0.0.0 when mm runs in a container or a sandbox)]
108	108	mm v0.1.7 \| config: agent.llamacpp.yaml \| provider: llamacpp \| model: jetbrains/mellum2-12b-a2.5b-instruct-gguf-q4_k_m:Q4_K_M \| ctx: unknown \| tools: bash, read_file, write_file, edit_file \| skills: 0 \| yolo: on
109	109	```
110	110
@@ -240,7 +240,7 @@ Un fichier d'agent hors du workspace ne donne qu'un avertissement : voir [quand
240	240	\|--------\|---------------------------------\|--------\|
241	241	\| `--project yes\\|no` \| Tout ranger dans un dossier `.mm` ? — posée en premier. `yes` : le fichier est `.mm/agent.yaml` (le dossier est créé), ses chemins sont écrits depuis la racine du projet (`skillsDir: .mm/skills`), et `mm` / `mm tui` lancés là le trouvent sans argument \| `no` \|
242	242	\| `--provider` \| Serveur de modèles : `llmman`, `ollama`, `dmr`, `llamacpp`, `kronk`, `openai` \| le premier serveur local qui répond, sinon `llmman` (`openai` n'est jamais sondé : son adresse par défaut est l'API d'OpenAI) \|
243		-\| `--base-url` \| `openai` seulement : Endpoint — l'adresse compatible OpenAI du service (OpenAI, Cerebras, Hugging Face…) ; pour les autres fournisseurs, pas posée \| `https://api.openai.com/v1` pour `openai` ; sinon l'adresse du fournisseur, ou son adresse de secours quand elle ne répond pas. Écrite comme `baseUrl` sauf si c'est la valeur par défaut \|
	243	+\| `--base-url` \| `openai` seulement : Endpoint — l'adresse compatible OpenAI du service (OpenAI, Cerebras, Hugging Face…) ; pour les autres fournisseurs, pas posée \| `https://api.openai.com/v1` pour `openai` ; sinon `AGENT_BASE_URL` s'il est défini, sinon la première des adresses locales du fournisseur dont la liste de modèles répond — son adresse par défaut, puis la même avec `127.0.0.1` quand elle nomme `localhost` — et son adresse de secours pour conteneur (`host.docker.internal`) seulement quand aucune ne répond. La question du modèle nomme l'adresse d'où vient sa liste (`Model? (listed from …)`). Écrite comme `baseUrl` sauf si c'est la valeur par défaut \|
244	244	\| `--api-key-env` \| `openai` seulement : variable d'environnement qui contient la clé d'API \| `OPENAI_API_KEY`. Écrite comme `apiKeyEnv` sauf si c'est la valeur par défaut \|
245	245	\| `--model` \| Modèle : la liste des modèles du serveur (`/api/tags` pour llmman et Ollama, `/models` pour les autres ; 5 s pour un serveur local, 20 s pour `openai`), ou un nom à taper quand il n'y a pas de liste — la question dit pourquoi (`<VARIABLE> is not set: no list`, ou l'erreur du serveur). Pour `openai`, la liste est demandée avec la clé lue dans la variable nommée plus haut, et pas demandée du tout quand cette variable est vide \| le premier de la liste \|
246	246	\| `--family` \| Famille du modèle \| détectée d'après le nom du modèle, puis d'après le nom de famille donné par le serveur \|

  ### Exemple
  ```
--[warning: llamacpp: nothing answers at http://127.0.0.1:8080/v1 — start it: llama-server -m <model.gguf> --jinja --port 8080]
++[warning: llamacpp: nothing answers at http://host.docker.internal:8080/v1 — start it: llama-server -m <model.gguf> --jinja --port 8080 (add --host 0.0.0.0 when mm runs in a container or a sandbox)]
  mm v0.1.7 | config: agent.llamacpp.yaml | provider: llamacpp | model: jetbrains/mellum2-12b-a2.5b-instruct-gguf-q4_k_m:Q4_K_M | ctx: unknown | tools: bash, read_file, write_file, edit_file | skills: 0 | yolo: on
  ```
  |--------|---------------------------------|--------|
  | `--project yes\|no` | Tout ranger dans un dossier `.mm` ? — posée en premier. `yes` : le fichier est `.mm/agent.yaml` (le dossier est créé), ses chemins sont écrits depuis la racine du projet (`skillsDir: .mm/skills`), et `mm` / `mm tui` lancés là le trouvent sans argument | `no` |
  | `--provider` | Serveur de modèles : `llmman`, `ollama`, `dmr`, `llamacpp`, `kronk`, `openai` | le premier serveur local qui répond, sinon `llmman` (`openai` n'est jamais sondé : son adresse par défaut est l'API d'OpenAI) |
--| `--base-url` | `openai` seulement : Endpoint — l'adresse compatible OpenAI du service (OpenAI, Cerebras, Hugging Face…) ; pour les autres fournisseurs, pas posée | `https://api.openai.com/v1` pour `openai` ; sinon l'adresse du fournisseur, ou son adresse de secours quand elle ne répond pas. Écrite comme `baseUrl` sauf si c'est la valeur par défaut |
++| `--base-url` | `openai` seulement : Endpoint — l'adresse compatible OpenAI du service (OpenAI, Cerebras, Hugging Face…) ; pour les autres fournisseurs, pas posée | `https://api.openai.com/v1` pour `openai` ; sinon `AGENT_BASE_URL` s'il est défini, sinon la première des adresses locales du fournisseur dont la liste de modèles répond — son adresse par défaut, puis la même avec `127.0.0.1` quand elle nomme `localhost` — et son adresse de secours pour conteneur (`host.docker.internal`) seulement quand aucune ne répond. La question du modèle nomme l'adresse d'où vient sa liste (`Model? (listed from …)`). Écrite comme `baseUrl` sauf si c'est la valeur par défaut |
  | `--api-key-env` | `openai` seulement : variable d'environnement qui contient la clé d'API | `OPENAI_API_KEY`. Écrite comme `apiKeyEnv` sauf si c'est la valeur par défaut |
  | `--model` | Modèle : la liste des modèles du serveur (`/api/tags` pour llmman et Ollama, `/models` pour les autres ; 5 s pour un serveur local, 20 s pour `openai`), ou un nom à taper quand il n'y a pas de liste — la question dit pourquoi (`<VARIABLE> is not set: no list`, ou l'erreur du serveur). Pour `openai`, la liste est demandée avec la clé lue dans la variable nommée plus haut, et pas demandée du tout quand cette variable est vide | le premier de la liste |
  | `--family` | Famille du modèle | détectée d'après le nom du modèle, puis d'après le nom de famille donné par le serveur |

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

@@ -197,7 +197,7 @@ tools:
197	197	\| Fournisseur \| `baseUrl` \| `fallback` \| `apiKeyEnv` \| Clé requise \|
198	198	\|-------------\|-----------\|------------\|-------------\|-------------\|
199	199	\| `dmr` \| `http://localhost:12434/engines/v1` \| `http://host.docker.internal:12434/engines/v1` \| aucune \| non \|
200		-\| `llamacpp` \| `http://127.0.0.1:8080/v1` \| aucun \| `LLAMA_API_KEY` \| non \|
	200	+\| `llamacpp` \| `http://127.0.0.1:8080/v1` \| `http://host.docker.internal:8080/v1` \| `LLAMA_API_KEY` \| non \|
201	201	\| `llmman` \| `http://127.0.0.1:17434/v1` \| `http://host.docker.internal:17434/v1` \| `LLMMAN_API_KEY` \| non \|
202	202	\| `ollama` \| `http://localhost:11434/v1` \| `http://host.docker.internal:11434/v1` \| `OLLAMA_API_KEY` \| non \|
203	203	\| `kronk` \| `http://127.0.0.1:11435/v1` \| `http://host.docker.internal:11435/v1` \| `KRONK_TOKEN` \| non — seuls les modes `authenticated` et `full-protected` de Kronk la vérifient \|
@@ -477,7 +477,7 @@ Un avertissement au démarrage, et non une erreur, est affiché quand `context.e
477	477	\| Fichier \| Rôle \|
478	478	\|---------\|------\|
479	479	\| `agent.yaml` \| Configuration Docker Model Runner : `provider: dmr`, `maxTurns: 40`, `displayCommands: true`, compression désactivée, un prompt système nommé « Bob » qui décrit les skills et les règles d'édition de fichiers. \|
480		-\| `agent.llamacpp.yaml` \| Configuration llama.cpp : `provider: llamacpp`, `baseUrl: http://127.0.0.1:8080/v1`, `fallback: ""`, compression activée. \|
	480	+\| `agent.llamacpp.yaml` \| Configuration llama.cpp : `provider: llamacpp`, `baseUrl: http://127.0.0.1:8080/v1`, compression activée. \|
481	481
482	482	### Exemple
483	483

  | Fournisseur | `baseUrl` | `fallback` | `apiKeyEnv` | Clé requise |
  |-------------|-----------|------------|-------------|-------------|
  | `dmr` | `http://localhost:12434/engines/v1` | `http://host.docker.internal:12434/engines/v1` | aucune | non |
--| `llamacpp` | `http://127.0.0.1:8080/v1` | aucun | `LLAMA_API_KEY` | non |
++| `llamacpp` | `http://127.0.0.1:8080/v1` | `http://host.docker.internal:8080/v1` | `LLAMA_API_KEY` | non |
  | `llmman` | `http://127.0.0.1:17434/v1` | `http://host.docker.internal:17434/v1` | `LLMMAN_API_KEY` | non |
  | `ollama` | `http://localhost:11434/v1` | `http://host.docker.internal:11434/v1` | `OLLAMA_API_KEY` | non |
  | `kronk` | `http://127.0.0.1:11435/v1` | `http://host.docker.internal:11435/v1` | `KRONK_TOKEN` | non — seuls les modes `authenticated` et `full-protected` de Kronk la vérifient |
  | Fichier | Rôle |
  |---------|------|
  | `agent.yaml` | Configuration Docker Model Runner : `provider: dmr`, `maxTurns: 40`, `displayCommands: true`, compression désactivée, un prompt système nommé « Bob » qui décrit les skills et les règles d'édition de fichiers. |
--| `agent.llamacpp.yaml` | Configuration llama.cpp : `provider: llamacpp`, `baseUrl: http://127.0.0.1:8080/v1`, `fallback: ""`, compression activée. |
++| `agent.llamacpp.yaml` | Configuration llama.cpp : `provider: llamacpp`, `baseUrl: http://127.0.0.1:8080/v1`, compression activée. |
  ### Exemple

modified internal/agentinit/decide_test.go +1 -1 Viewed

@@ -170,7 +170,7 @@ func TestRenderRecordsTheDefaultURLs(t *testing.T) {
170	170	t.Errorf("llmman's default URLs missing:\n%s", llmman)
171	171	}
172	172	llamacpp := Render(Decide(p, Choices{Provider: "llamacpp", Model: "qwen3", Usage: "assistant", Machine: "small"}, Facts{}))
173		- if !strings.Contains(llamacpp, "# baseUrl: \"http://127.0.0.1:8080/v1\"\n# (llamacpp has no fallback address)\n") {
	173	+ if !strings.Contains(llamacpp, "# baseUrl: \"http://127.0.0.1:8080/v1\"\n# fallback: \"http://host.docker.internal:8080/v1\" # tried when baseUrl does not answer\n") {
174	174	t.Errorf("llamacpp's default URL missing:\n%s", llamacpp)
175	175	}
176	176	own := Render(Decide(p, Choices{Provider: "llmman", BaseURL: "http://box:17434/v1", Model: "qwen3", Usage: "assistant", Machine: "small"}, Facts{}))

  		t.Errorf("llmman's default URLs missing:\n%s", llmman)
  	}
  	llamacpp := Render(Decide(p, Choices{Provider: "llamacpp", Model: "qwen3", Usage: "assistant", Machine: "small"}, Facts{}))
--	if !strings.Contains(llamacpp, "# baseUrl: \"http://127.0.0.1:8080/v1\"\n# (llamacpp has no fallback address)\n") {
++	if !strings.Contains(llamacpp, "# baseUrl: \"http://127.0.0.1:8080/v1\"\n# fallback: \"http://host.docker.internal:8080/v1\"   # tried when baseUrl does not answer\n") {
  		t.Errorf("llamacpp's default URL missing:\n%s", llamacpp)
  	}
  	own := Render(Decide(p, Choices{Provider: "llmman", BaseURL: "http://box:17434/v1", Model: "qwen3", Usage: "assistant", Machine: "small"}, Facts{}))

modified internal/agentinit/flow.go +5 -2 Viewed

@@ -222,6 +222,9 @@ func (f *Flow) Step(key string) Step {
222	222	st.Kind, st.Prompt = elicit.Text, f.typeModelPrompt()
223	223	break
224	224	}
	225	+ // Where the list comes from (ticket 214): a wrong address — the
	226	+ // container one on the host — shows at once.
	227	+ st.Prompt = fmt.Sprintf("Model? (listed from %s)", f.Base())
225	228	for _, m := range models {
226	229	st.Options = append(st.Options, Option{Value: m, Label: m, Detail: modelDetail(f.c.Provider, m)})
227	230	}
@@ -578,7 +581,7 @@ func (f *Flow) Result() (Choices, Facts, string) {
578	581	// Base is the resolved server address ("" before a provider is chosen).
579	582	func (f *Flow) Base() string {
580	583	if f.base == "" && f.c.Provider != "" {
581		- f.base = resolveBase(f.c)
	584	+ f.base = resolveBase(f.ctx, f.c)
582	585	}
583	586	return f.base
584	587	}
@@ -628,7 +631,7 @@ func (f *Flow) answers(name string) bool {
628	631	if f.reachable == nil {
629	632	f.reachable = map[string]bool{}
630	633	for _, p := range local {
631		- _, err := Models(f.ctx, p, resolveBase(Choices{Provider: p}))
	634	+ _, err := Models(f.ctx, p, resolveBase(f.ctx, Choices{Provider: p}))
632	635	f.reachable[p] = err == nil
633	636	}
634	637	}

  			st.Kind, st.Prompt = elicit.Text, f.typeModelPrompt()
  			break
  		}
++		// Where the list comes from (ticket 214): a wrong address — the
++		// container one on the host — shows at once.
++		st.Prompt = fmt.Sprintf("Model? (listed from %s)", f.Base())
  		for _, m := range models {
  			st.Options = append(st.Options, Option{Value: m, Label: m, Detail: modelDetail(f.c.Provider, m)})
  		}
  // Base is the resolved server address ("" before a provider is chosen).
  func (f *Flow) Base() string {
  	if f.base == "" && f.c.Provider != "" {
--		f.base = resolveBase(f.c)
++		f.base = resolveBase(f.ctx, f.c)
  	}
  	return f.base
  }
  	if f.reachable == nil {
  		f.reachable = map[string]bool{}
  		for _, p := range local {
--			_, err := Models(f.ctx, p, resolveBase(Choices{Provider: p}))
++			_, err := Models(f.ctx, p, resolveBase(f.ctx, Choices{Provider: p}))
  			f.reachable[p] = err == nil
  		}
  	}

modified internal/agentinit/wizard.go +61 -3 Viewed

@@ -367,12 +367,70 @@ func (w *lineWizard) ask(st Step) string {
367	367	}
368	368	}
369	369
370		-// resolveBase is the server's address: the flag's, else the provider's
371		-// default (or its fallback, when the default does not answer).
372		-func resolveBase(c Choices) string {
	370	+// resolveBase is the server's address the wizard lists models from: the
	371	+// flag's, else the one the environment imposes (AGENT_BASE_URL…), else the
	372	+// first that answers of the provider's local addresses, then its fallback
	373	+// (chooseBase).
	374	+//
	375	+// Ticket 214: it used to take the engine's own resolution, whose probe gives
	376	+// the default address 2 s on /models and switches to the container fallback
	377	+// (host.docker.internal) without a word. mm on a Mac with Docker Desktop —
	378	+// where that name resolves on the host too — then listed Ollama's models
	379	+// from the container address. The written file is not concerned: it carries
	380	+// no baseUrl for the default address.
	381	+func resolveBase(ctx context.Context, c Choices) string {
373	382	if c.BaseURL != "" {
374	383	return c.BaseURL
375	384	}
	385	+ if env := engine.EnvBaseURL(c.Provider); env != "" {
	386	+ return env
	387	+ }
	388	+ if base, fallback, ok := engine.DefaultURLs(c.Provider); ok {
	389	+ return chooseBase(base, fallback, func(url string) bool {
	390	+ _, err := Models(ctx, c.Provider, url)
	391	+ return err == nil
	392	+ })
	393	+ }
	394	+ return engineBase(c)
	395	+}
	396	+
	397	+// chooseBase picks the address to list models from: the first that lists
	398	+// them among the local ones (localAddresses), else the fallback when it
	399	+// does, else the default — whose error the model question then shows. A
	400	+// provider without a fallback (llamacpp, openai) is not probed at all: there
	401	+// is nothing to choose, and openai's default is a remote service nobody
	402	+// asked mm to call yet.
	403	+func chooseBase(base, fallback string, lists func(string) bool) string {
	404	+ if fallback == "" {
	405	+ return base
	406	+ }
	407	+ for _, url := range localAddresses(base) {
	408	+ if lists(url) {
	409	+ return url
	410	+ }
	411	+ }
	412	+ if lists(fallback) {
	413	+ return fallback
	414	+ }
	415	+ return base
	416	+}
	417	+
	418	+// localAddresses is the default address, then — when it names localhost —
	419	+// the same with 127.0.0.1: a server listening on IPv4 only does not answer
	420	+// a localhost that resolves to ::1 first.
	421	+//
	422	+// localAddresses("http://localhost:11434/v1") // [http://localhost:11434/v1 http://127.0.0.1:11434/v1]
	423	+func localAddresses(base string) []string {
	424	+ out := []string{base}
	425	+ if strings.Contains(base, "://localhost:") \|\| strings.HasSuffix(base, "://localhost") {
	426	+ out = append(out, strings.Replace(base, "://localhost", "://127.0.0.1", 1))
	427	+ }
	428	+ return out
	429	+}
	430	+
	431	+// engineBase is the engine's own resolution, for a provider the wizard has
	432	+// no addresses for.
	433	+func engineBase(c Choices) string {
376	434	p, err := engine.Lookup(c.Provider)
377	435	if err != nil {
378	436	return ""

  	}
  }
--// resolveBase is the server's address: the flag's, else the provider's
++// resolveBase is the server's address the wizard lists models from: the
--// default (or its fallback, when the default does not answer).
++// flag's, else the one the environment imposes (AGENT_BASE_URL…), else the
--func resolveBase(c Choices) string {
++// first that answers of the provider's local addresses, then its fallback
++// (chooseBase).
++//
++// Ticket 214: it used to take the engine's own resolution, whose probe gives
++// the default address 2 s on /models and switches to the container fallback
++// (host.docker.internal) without a word. mm on a Mac with Docker Desktop —
++// where that name resolves on the host too — then listed Ollama's models
++// from the container address. The written file is not concerned: it carries
++// no baseUrl for the default address.
++func resolveBase(ctx context.Context, c Choices) string {
  	if c.BaseURL != "" {
  		return c.BaseURL
  	}
++	if env := engine.EnvBaseURL(c.Provider); env != "" {
++		return env
++	}
++	if base, fallback, ok := engine.DefaultURLs(c.Provider); ok {
++		return chooseBase(base, fallback, func(url string) bool {
++			_, err := Models(ctx, c.Provider, url)
++			return err == nil
++		})
++	}
++	return engineBase(c)
++}
++
++// chooseBase picks the address to list models from: the first that lists
++// them among the local ones (localAddresses), else the fallback when it
++// does, else the default — whose error the model question then shows. A
++// provider without a fallback (llamacpp, openai) is not probed at all: there
++// is nothing to choose, and openai's default is a remote service nobody
++// asked mm to call yet.
++func chooseBase(base, fallback string, lists func(string) bool) string {
++	if fallback == "" {
++		return base
++	}
++	for _, url := range localAddresses(base) {
++		if lists(url) {
++			return url
++		}
++	}
++	if lists(fallback) {
++		return fallback
++	}
++	return base
++}
++
++// localAddresses is the default address, then — when it names localhost —
++// the same with 127.0.0.1: a server listening on IPv4 only does not answer
++// a localhost that resolves to ::1 first.
++//
++//	localAddresses("http://localhost:11434/v1") // [http://localhost:11434/v1 http://127.0.0.1:11434/v1]
++func localAddresses(base string) []string {
++	out := []string{base}
++	if strings.Contains(base, "://localhost:") || strings.HasSuffix(base, "://localhost") {
++		out = append(out, strings.Replace(base, "://localhost", "://127.0.0.1", 1))
++	}
++	return out
++}
++
++// engineBase is the engine's own resolution, for a provider the wizard has
++// no addresses for.
++func engineBase(c Choices) string {
  	p, err := engine.Lookup(c.Provider)
  	if err != nil {
  		return ""

modified internal/agentinit/wizard_test.go +102 -0 Viewed

@@ -3,13 +3,18 @@ package agentinit
3	3	import (
4	4	"bytes"
5	5	"context"
	6	+ "encoding/json"
	7	+ "net/http"
	8	+ "net/http/httptest"
6	9	"os"
7	10	"path/filepath"
8	11	"strings"
9	12	"testing"
	13	+ "time"
10	14
11	15	"mm/internal/config"
12	16	"mm/internal/elicit"
	17	+ "mm/internal/engine"
13	18	)
14	19
15	20	func llmmanStub(t *testing.T) string {
@@ -247,3 +252,100 @@ func TestMainToStdoutAndBadAnswers(t *testing.T) {
247	252	t.Errorf("bad usage: exit %d:\n%s", code, stdout.String())
248	253	}
249	254	}
	255	+
	256	+// Ticket 214: the wizard lists the models from a local address whenever one
	257	+// answers — the default, then the same with 127.0.0.1 — and takes the
	258	+// container fallback only when none does. A provider without a fallback is
	259	+// not probed.
	260	+func TestChooseBasePrefersALocalAddress(t *testing.T) {
	261	+ const base, ipv4, fallback = "http://localhost:11434/v1", "http://127.0.0.1:11434/v1", "http://host.docker.internal:11434/v1"
	262	+ answering := func(urls ...string) (func(string) bool, *[]string) {
	263	+ var asked []string
	264	+ return func(url string) bool {
	265	+ asked = append(asked, url)
	266	+ for _, u := range urls {
	267	+ if u == url {
	268	+ return true
	269	+ }
	270	+ }
	271	+ return false
	272	+ }, &asked
	273	+ }
	274	+ cases := []struct {
	275	+ name string
	276	+ answer []string
	277	+ want string
	278	+ wantAsked int
	279	+ }{
	280	+ {"the default answers, the fallback too (Docker Desktop on a Mac)", []string{base, fallback}, base, 1},
	281	+ {"only 127.0.0.1 answers (a server on IPv4 only)", []string{ipv4, fallback}, ipv4, 2},
	282	+ {"only the fallback answers (mm in a container)", []string{fallback}, fallback, 3},
	283	+ {"nothing answers: the default, whose error the question shows", nil, base, 3},
	284	+ }
	285	+ for _, c := range cases {
	286	+ lists, asked := answering(c.answer...)
	287	+ if got := chooseBase(base, fallback, lists); got != c.want \|\| len(*asked) != c.wantAsked {
	288	+ t.Errorf("%s: got %s after asking %v, want %s after %d", c.name, got, *asked, c.want, c.wantAsked)
	289	+ }
	290	+ }
	291	+ lists, asked := answering()
	292	+ if got := chooseBase("https://api.openai.com/v1", "", lists); got != "https://api.openai.com/v1" \|\| len(*asked) != 0 {
	293	+ t.Errorf("no fallback: got %s, asked %v — want the default, nothing asked", got, *asked)
	294	+ }
	295	+}
	296	+
	297	+func TestLocalAddresses(t *testing.T) {
	298	+ for in, want := range map[string]string{
	299	+ "http://localhost:11434/v1": "http://localhost:11434/v1 http://127.0.0.1:11434/v1",
	300	+ "http://localhost:12434/engines/v1": "http://localhost:12434/engines/v1 http://127.0.0.1:12434/engines/v1",
	301	+ "http://127.0.0.1:17434/v1": "http://127.0.0.1:17434/v1",
	302	+ } {
	303	+ if got := strings.Join(localAddresses(in), " "); got != want {
	304	+ t.Errorf("localAddresses(%s) = %s, want %s", in, got, want)
	305	+ }
	306	+ }
	307	+}
	308	+
	309	+// The address comes from the model list, not the engine's 2 s probe of
	310	+// /models: a server whose /v1/models is slow but whose list answers is kept
	311	+// — before ticket 214, the probe gave up and the fallback was taken.
	312	+func TestResolveBaseKeepsASlowLocalServer(t *testing.T) {
	313	+ slow := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
	314	+ switch r.URL.Path {
	315	+ case "/api/tags":
	316	+ _ = json.NewEncoder(w).Encode(map[string]any{"models": []any{map[string]any{"name": "qwen3.5:4b"}}})
	317	+ case "/v1/models":
	318	+ time.Sleep(3 * time.Second) // longer than the engine's probe
	319	+ _ = json.NewEncoder(w).Encode(map[string]any{"data": []any{}})
	320	+ default:
	321	+ http.NotFound(w, r)
	322	+ }
	323	+ }))
	324	+ t.Cleanup(slow.Close)
	325	+ base := slow.URL + "/v1"
	326	+ // The engine's own resolution — what the wizard used before — gives up
	327	+ // on this server and takes the fallback.
	328	+ fallback := "http://host.docker.internal:1/v1"
	329	+ p, err := engine.Lookup("ollama")
	330	+ if err != nil {
	331	+ t.Fatal(err)
	332	+ }
	333	+ if b, err := p.Resolve(config.Config{Provider: "ollama", Model: "probe", BaseURL: base, Fallback: &fallback}); err != nil \|\| b.BaseURL != fallback {
	334	+ t.Fatalf("the engine's probe kept %s (%v): the test server is not slow enough to show ticket 214", b.BaseURL, err)
	335	+ }
	336	+ got := chooseBase(base, fallback, func(url string) bool {
	337	+ _, err := Models(context.Background(), "ollama", url)
	338	+ return err == nil
	339	+ })
	340	+ if got != base {
	341	+ t.Errorf("got %s, want the local server %s", got, base)
	342	+ }
	343	+}
	344	+
	345	+// An address imposed by the environment is taken as is, like the engine does.
	346	+func TestResolveBaseHonoursAgentBaseURL(t *testing.T) {
	347	+ t.Setenv("AGENT_BASE_URL", "http://example.invalid:9/v1")
	348	+ if got := resolveBase(context.Background(), Choices{Provider: "ollama"}); got != "http://example.invalid:9/v1" {
	349	+ t.Errorf("got %s", got)
	350	+ }
	351	+}

  import (
  	"bytes"
  	"context"
++	"encoding/json"
++	"net/http"
++	"net/http/httptest"
  	"os"
  	"path/filepath"
  	"strings"
  	"testing"
++	"time"
  	"mm/internal/config"
  	"mm/internal/elicit"
++	"mm/internal/engine"
  )
  func llmmanStub(t *testing.T) string {
  		t.Errorf("bad usage: exit %d:\n%s", code, stdout.String())
  	}
  }
++
++// Ticket 214: the wizard lists the models from a local address whenever one
++// answers — the default, then the same with 127.0.0.1 — and takes the
++// container fallback only when none does. A provider without a fallback is
++// not probed.
++func TestChooseBasePrefersALocalAddress(t *testing.T) {
++	const base, ipv4, fallback = "http://localhost:11434/v1", "http://127.0.0.1:11434/v1", "http://host.docker.internal:11434/v1"
++	answering := func(urls ...string) (func(string) bool, *[]string) {
++		var asked []string
++		return func(url string) bool {
++			asked = append(asked, url)
++			for _, u := range urls {
++				if u == url {
++					return true
++				}
++			}
++			return false
++		}, &asked
++	}
++	cases := []struct {
++		name      string
++		answer    []string
++		want      string
++		wantAsked int
++	}{
++		{"the default answers, the fallback too (Docker Desktop on a Mac)", []string{base, fallback}, base, 1},
++		{"only 127.0.0.1 answers (a server on IPv4 only)", []string{ipv4, fallback}, ipv4, 2},
++		{"only the fallback answers (mm in a container)", []string{fallback}, fallback, 3},
++		{"nothing answers: the default, whose error the question shows", nil, base, 3},
++	}
++	for _, c := range cases {
++		lists, asked := answering(c.answer...)
++		if got := chooseBase(base, fallback, lists); got != c.want || len(*asked) != c.wantAsked {
++			t.Errorf("%s: got %s after asking %v, want %s after %d", c.name, got, *asked, c.want, c.wantAsked)
++		}
++	}
++	lists, asked := answering()
++	if got := chooseBase("https://api.openai.com/v1", "", lists); got != "https://api.openai.com/v1" || len(*asked) != 0 {
++		t.Errorf("no fallback: got %s, asked %v — want the default, nothing asked", got, *asked)
++	}
++}
++
++func TestLocalAddresses(t *testing.T) {
++	for in, want := range map[string]string{
++		"http://localhost:11434/v1":         "http://localhost:11434/v1 http://127.0.0.1:11434/v1",
++		"http://localhost:12434/engines/v1": "http://localhost:12434/engines/v1 http://127.0.0.1:12434/engines/v1",
++		"http://127.0.0.1:17434/v1":         "http://127.0.0.1:17434/v1",
++	} {
++		if got := strings.Join(localAddresses(in), " "); got != want {
++			t.Errorf("localAddresses(%s) = %s, want %s", in, got, want)
++		}
++	}
++}
++
++// The address comes from the model list, not the engine's 2 s probe of
++// /models: a server whose /v1/models is slow but whose list answers is kept
++// — before ticket 214, the probe gave up and the fallback was taken.
++func TestResolveBaseKeepsASlowLocalServer(t *testing.T) {
++	slow := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
++		switch r.URL.Path {
++		case "/api/tags":
++			_ = json.NewEncoder(w).Encode(map[string]any{"models": []any{map[string]any{"name": "qwen3.5:4b"}}})
++		case "/v1/models":
++			time.Sleep(3 * time.Second) // longer than the engine's probe
++			_ = json.NewEncoder(w).Encode(map[string]any{"data": []any{}})
++		default:
++			http.NotFound(w, r)
++		}
++	}))
++	t.Cleanup(slow.Close)
++	base := slow.URL + "/v1"
++	// The engine's own resolution — what the wizard used before — gives up
++	// on this server and takes the fallback.
++	fallback := "http://host.docker.internal:1/v1"
++	p, err := engine.Lookup("ollama")
++	if err != nil {
++		t.Fatal(err)
++	}
++	if b, err := p.Resolve(config.Config{Provider: "ollama", Model: "probe", BaseURL: base, Fallback: &fallback}); err != nil || b.BaseURL != fallback {
++		t.Fatalf("the engine's probe kept %s (%v): the test server is not slow enough to show ticket 214", b.BaseURL, err)
++	}
++	got := chooseBase(base, fallback, func(url string) bool {
++		_, err := Models(context.Background(), "ollama", url)
++		return err == nil
++	})
++	if got != base {
++		t.Errorf("got %s, want the local server %s", got, base)
++	}
++}
++
++// An address imposed by the environment is taken as is, like the engine does.
++func TestResolveBaseHonoursAgentBaseURL(t *testing.T) {
++	t.Setenv("AGENT_BASE_URL", "http://example.invalid:9/v1")
++	if got := resolveBase(context.Background(), Choices{Provider: "ollama"}); got != "http://example.invalid:9/v1" {
++		t.Errorf("got %s", got)
++	}
++}

modified internal/engine/openai_compat.go +20 -1 Viewed

@@ -87,6 +87,20 @@ func DefaultURLs(provider string) (baseURL, fallback string, ok bool) {
87	87	return p.baseURL, p.fallback, true
88	88	}
89	89
	90	+// EnvBaseURL is the address the environment imposes on a provider —
	91	+// AGENT_BASE_URL, else the provider's legacy variable (DMR_BASE_URL) — or
	92	+// "" when none is set. Resolve takes it as is, without probing; so does
	93	+// `mm init` (ticket 214).
	94	+func EnvBaseURL(provider string) string {
	95	+ if env := os.Getenv("AGENT_BASE_URL"); env != "" {
	96	+ return env
	97	+ }
	98	+ if p, found := registry[provider].(*openaiCompat); found && p.legacyBaseURLEnv != "" {
	99	+ return os.Getenv(p.legacyBaseURLEnv)
	100	+ }
	101	+ return ""
	102	+}
	103	+
90	104	func init() {
91	105	register(&openaiCompat{
92	106	name: "dmr",
@@ -104,11 +118,16 @@ func init() {
104	118	register(&openaiCompat{
105	119	name: "llamacpp",
106	120	baseURL: "http://127.0.0.1:8080/v1", // llama-server's default host:port, plus the OpenAI prefix
	121	+ // From a container or a sandbox, like the other local servers
	122	+ // (ticket 217). llama-server listens on the loopback only unless
	123	+ // started with --host 0.0.0.0, which Linux and a Docker Sandbox need
	124	+ // for this name to reach it (Docker Desktop reaches the loopback).
	125	+ fallback: "http://host.docker.internal:8080/v1",
107	126	// llama-server only checks a key when started with --api-key; most
108	127	// local set-ups are not, so an unset LLAMA_API_KEY is normal.
109	128	apiKeyEnv: "LLAMA_API_KEY",
110	129	dummyKey: "not-needed",
111		- startHint: "start it: llama-server -m <model.gguf> --jinja --port 8080",
	130	+ startHint: "start it: llama-server -m <model.gguf> --jinja --port 8080 (add --host 0.0.0.0 when mm runs in a container or a sandbox)",
112	131	notFoundHint: "llama-server serves one model; check baseUrl ends with /v1 (model: %s)",
113	132	probe: probeLlamaCpp,
114	133	progress: true,

  	return p.baseURL, p.fallback, true
  }
++// EnvBaseURL is the address the environment imposes on a provider —
++// AGENT_BASE_URL, else the provider's legacy variable (DMR_BASE_URL) — or
++// "" when none is set. Resolve takes it as is, without probing; so does
++// `mm init` (ticket 214).
++func EnvBaseURL(provider string) string {
++	if env := os.Getenv("AGENT_BASE_URL"); env != "" {
++		return env
++	}
++	if p, found := registry[provider].(*openaiCompat); found && p.legacyBaseURLEnv != "" {
++		return os.Getenv(p.legacyBaseURLEnv)
++	}
++	return ""
++}
++
  func init() {
  	register(&openaiCompat{
  		name:             "dmr",
  	register(&openaiCompat{
  		name:    "llamacpp",
  		baseURL: "http://127.0.0.1:8080/v1", // llama-server's default host:port, plus the OpenAI prefix
++		// From a container or a sandbox, like the other local servers
++		// (ticket 217). llama-server listens on the loopback only unless
++		// started with --host 0.0.0.0, which Linux and a Docker Sandbox need
++		// for this name to reach it (Docker Desktop reaches the loopback).
++		fallback: "http://host.docker.internal:8080/v1",
  		// llama-server only checks a key when started with --api-key; most
  		// local set-ups are not, so an unset LLAMA_API_KEY is normal.
  		apiKeyEnv:    "LLAMA_API_KEY",
  		dummyKey:     "not-needed",
--		startHint:    "start it: llama-server -m <model.gguf> --jinja --port 8080",
++		startHint:    "start it: llama-server -m <model.gguf> --jinja --port 8080 (add --host 0.0.0.0 when mm runs in a container or a sandbox)",
  		notFoundHint: "llama-server serves one model; check baseUrl ends with /v1 (model: %s)",
  		probe:        probeLlamaCpp,
  		progress:     true,

modified internal/engine/provider_test.go +11 -6 Viewed

@@ -108,10 +108,13 @@ func TestResolveURLPrecedence(t *testing.T) {
108	108	{"an explicit empty fallback disables it", "dmr", config.Config{BaseURL: down, Fallback: ptr("")}, nil, down, "not-needed"},
109	109	{"AGENT_BASE_URL wins without probing", "dmr", config.Config{BaseURL: up.URL, Fallback: ptr("")}, map[string]string{"AGENT_BASE_URL": down}, down, "not-needed"},
110	110	{"DMR_BASE_URL still works for dmr", "dmr", config.Config{BaseURL: up.URL, Fallback: ptr("")}, map[string]string{"DMR_BASE_URL": down}, down, "not-needed"},
111		- {"llamacpp default is llama-server's port", "llamacpp", config.Config{}, nil, "http://127.0.0.1:8080/v1", "not-needed"},
112		- {"llamacpp ignores DMR_BASE_URL", "llamacpp", config.Config{}, map[string]string{"DMR_BASE_URL": down}, "http://127.0.0.1:8080/v1", "not-needed"},
113		- {"llamacpp reads LLAMA_API_KEY when set", "llamacpp", config.Config{}, map[string]string{"LLAMA_API_KEY": "secret"}, "http://127.0.0.1:8080/v1", "secret"},
114		- {"apiKeyEnv names another variable", "llamacpp", config.Config{APIKeyEnv: "MY_KEY"}, map[string]string{"MY_KEY": "k2"}, "http://127.0.0.1:8080/v1", "k2"},
	111	+ // llamacpp falls back on the host like the other local servers
	112	+ // since ticket 217 (its default address: TestDefaultURLs).
	113	+ {"llamacpp falls back to the host's llama-server", "llamacpp", config.Config{BaseURL: down}, nil, "http://host.docker.internal:8080/v1", "not-needed"},
	114	+ {"llamacpp keeps a reachable baseUrl", "llamacpp", config.Config{BaseURL: up.URL}, nil, up.URL, "not-needed"},
	115	+ {"llamacpp ignores DMR_BASE_URL", "llamacpp", config.Config{BaseURL: up.URL}, map[string]string{"DMR_BASE_URL": down}, up.URL, "not-needed"},
	116	+ {"llamacpp reads LLAMA_API_KEY when set", "llamacpp", config.Config{BaseURL: up.URL}, map[string]string{"LLAMA_API_KEY": "secret"}, up.URL, "secret"},
	117	+ {"apiKeyEnv names another variable", "llamacpp", config.Config{BaseURL: up.URL, APIKeyEnv: "MY_KEY"}, map[string]string{"MY_KEY": "k2"}, up.URL, "k2"},
115	118	{"llmman falls back to the host's daemon", "llmman", config.Config{BaseURL: down}, nil, "http://host.docker.internal:17434/v1", "not-needed"},
116	119	{"llmman keeps a reachable baseUrl", "llmman", config.Config{BaseURL: up.URL}, nil, up.URL, "not-needed"},
117	120	{"llmman reads LLMMAN_API_KEY when set", "llmman", config.Config{BaseURL: up.URL}, map[string]string{"LLMMAN_API_KEY": "k3", "LLAMA_API_KEY": "secret"}, up.URL, "k3"},
@@ -445,7 +448,9 @@ func TestConfigCompat(t *testing.T) {
445	448	}
446	449
447	450	config.Cfg = saved
448		- llama := write("llama.yaml", "provider: llamacpp\nmodel: qwen2.5-coder\ncontextWindow: 16384\n")
	451	+ // fallback: "" keeps the default address whether llama-server runs or
	452	+ // not (it has a fallback since ticket 217).
	453	+ llama := write("llama.yaml", "provider: llamacpp\nmodel: qwen2.5-coder\ncontextWindow: 16384\nfallback: \"\"\n")
449	454	if _, err := config.Load(llama); err != nil {
450	455	t.Fatalf("Load(llamacpp config): %v", err)
451	456	}
@@ -559,7 +564,7 @@ func TestDefaultURLs(t *testing.T) {
559	564	{"ollama", "http://localhost:11434/v1", "http://host.docker.internal:11434/v1"},
560	565	{"kronk", "http://127.0.0.1:11435/v1", "http://host.docker.internal:11435/v1"},
561	566	{"dmr", "http://localhost:12434/engines/v1", "http://host.docker.internal:12434/engines/v1"},
562		- {"llamacpp", "http://127.0.0.1:8080/v1", ""},
	567	+ {"llamacpp", "http://127.0.0.1:8080/v1", "http://host.docker.internal:8080/v1"},
563	568	{"openai", "https://api.openai.com/v1", ""},
564	569	}
565	570	for _, c := range cases {

  		{"an explicit empty fallback disables it", "dmr", config.Config{BaseURL: down, Fallback: ptr("")}, nil, down, "not-needed"},
  		{"AGENT_BASE_URL wins without probing", "dmr", config.Config{BaseURL: up.URL, Fallback: ptr("")}, map[string]string{"AGENT_BASE_URL": down}, down, "not-needed"},
  		{"DMR_BASE_URL still works for dmr", "dmr", config.Config{BaseURL: up.URL, Fallback: ptr("")}, map[string]string{"DMR_BASE_URL": down}, down, "not-needed"},
--		{"llamacpp default is llama-server's port", "llamacpp", config.Config{}, nil, "http://127.0.0.1:8080/v1", "not-needed"},
++		// llamacpp falls back on the host like the other local servers
--		{"llamacpp ignores DMR_BASE_URL", "llamacpp", config.Config{}, map[string]string{"DMR_BASE_URL": down}, "http://127.0.0.1:8080/v1", "not-needed"},
++		// since ticket 217 (its default address: TestDefaultURLs).
--		{"llamacpp reads LLAMA_API_KEY when set", "llamacpp", config.Config{}, map[string]string{"LLAMA_API_KEY": "secret"}, "http://127.0.0.1:8080/v1", "secret"},
++		{"llamacpp falls back to the host's llama-server", "llamacpp", config.Config{BaseURL: down}, nil, "http://host.docker.internal:8080/v1", "not-needed"},
--		{"apiKeyEnv names another variable", "llamacpp", config.Config{APIKeyEnv: "MY_KEY"}, map[string]string{"MY_KEY": "k2"}, "http://127.0.0.1:8080/v1", "k2"},
++		{"llamacpp keeps a reachable baseUrl", "llamacpp", config.Config{BaseURL: up.URL}, nil, up.URL, "not-needed"},
++		{"llamacpp ignores DMR_BASE_URL", "llamacpp", config.Config{BaseURL: up.URL}, map[string]string{"DMR_BASE_URL": down}, up.URL, "not-needed"},
++		{"llamacpp reads LLAMA_API_KEY when set", "llamacpp", config.Config{BaseURL: up.URL}, map[string]string{"LLAMA_API_KEY": "secret"}, up.URL, "secret"},
++		{"apiKeyEnv names another variable", "llamacpp", config.Config{BaseURL: up.URL, APIKeyEnv: "MY_KEY"}, map[string]string{"MY_KEY": "k2"}, up.URL, "k2"},
  		{"llmman falls back to the host's daemon", "llmman", config.Config{BaseURL: down}, nil, "http://host.docker.internal:17434/v1", "not-needed"},
  		{"llmman keeps a reachable baseUrl", "llmman", config.Config{BaseURL: up.URL}, nil, up.URL, "not-needed"},
  		{"llmman reads LLMMAN_API_KEY when set", "llmman", config.Config{BaseURL: up.URL}, map[string]string{"LLMMAN_API_KEY": "k3", "LLAMA_API_KEY": "secret"}, up.URL, "k3"},
  	}
  	config.Cfg = saved
--	llama := write("llama.yaml", "provider: llamacpp\nmodel: qwen2.5-coder\ncontextWindow: 16384\n")
++	// fallback: "" keeps the default address whether llama-server runs or
++	// not (it has a fallback since ticket 217).
++	llama := write("llama.yaml", "provider: llamacpp\nmodel: qwen2.5-coder\ncontextWindow: 16384\nfallback: \"\"\n")
  	if _, err := config.Load(llama); err != nil {
  		t.Fatalf("Load(llamacpp config): %v", err)
  	}
  		{"ollama", "http://localhost:11434/v1", "http://host.docker.internal:11434/v1"},
  		{"kronk", "http://127.0.0.1:11435/v1", "http://host.docker.internal:11435/v1"},
  		{"dmr", "http://localhost:12434/engines/v1", "http://host.docker.internal:12434/engines/v1"},
--		{"llamacpp", "http://127.0.0.1:8080/v1", ""},
++		{"llamacpp", "http://127.0.0.1:8080/v1", "http://host.docker.internal:8080/v1"},
  		{"openai", "https://api.openai.com/v1", ""},
  	}
  	for _, c := range cases {

Merge feature/llamacpp-fallback into main