bots-garden/mini-me · #49

modified .tickets/issues/0165-how-to-unload-models-when-using-decision-models.yaml +2 -2 Viewed

@@ -6,10 +6,10 @@ author:
6	6	name: k33g
7	7	email: ph.charriere@gmail.com
8	8	createdAt: 2026-10-26T05:00:19.639Z
9		-updatedAt: 2026-10-26T14:41:47.750Z
	9	+updatedAt: 2026-10-26T19:50:00.000Z
10	10	labels:
11	11	- decision models
12		-milestone: v1.0.3
	12	+milestone: v1.0.4
13	13	epic: Decision models
14	14	body: ""
15	15	tasks: []

    name: k33g
    email: ph.charriere@gmail.com
  createdAt: 2026-10-26T05:00:19.639Z
--updatedAt: 2026-10-26T14:41:47.750Z
++updatedAt: 2026-10-26T19:50:00.000Z
  labels:
    - decision models
--milestone: v1.0.3
++milestone: v1.0.4
  epic: Decision models
  body: ""
  tasks: []

modified .tickets/issues/0166-think-on-how-to-simplify-usage-of-decision-model.yaml +2 -2 Viewed

@@ -6,10 +6,10 @@ author:
6	6	name: k33g
7	7	email: ph.charriere@gmail.com
8	8	createdAt: 2026-10-26T05:01:03.034Z
9		-updatedAt: 2026-10-26T14:42:01.348Z
	9	+updatedAt: 2026-10-26T19:50:00.000Z
10	10	labels:
11	11	- decision models
12		-milestone: v1.0.3
	12	+milestone: v1.0.4
13	13	epic: Decision models
14	14	body: ""
15	15	tasks: []

    name: k33g
    email: ph.charriere@gmail.com
  createdAt: 2026-10-26T05:01:03.034Z
--updatedAt: 2026-10-26T14:42:01.348Z
++updatedAt: 2026-10-26T19:50:00.000Z
  labels:
    - decision models
--milestone: v1.0.3
++milestone: v1.0.4
  epic: Decision models
  body: ""
  tasks: []

modified .tickets/issues/0176-openai-sampling-settings-bug.yaml +2 -2 Viewed

@@ -6,11 +6,11 @@ author:
6	6	name: k33g
7	7	email: ph.charriere@gmail.com
8	8	createdAt: 2026-10-26T13:50:22.583Z
9		-updatedAt: 2026-10-26T14:40:39.691Z
	9	+updatedAt: 2026-10-26T19:45:00.000Z
10	10	labels:
11	11	- kind::bug
12	12	- To be tested
13		-milestone: v1.0.2
	13	+milestone: v1.0.3
14	14	body: \|
15	15	I generated the `agent.yaml` file with this command: `mm init`
16	16	the result is:

    name: k33g
    email: ph.charriere@gmail.com
  createdAt: 2026-10-26T13:50:22.583Z
--updatedAt: 2026-10-26T14:40:39.691Z
++updatedAt: 2026-10-26T19:45:00.000Z
  labels:
    - kind::bug
    - To be tested
--milestone: v1.0.2
++milestone: v1.0.3
  body: |
    I generated the `agent.yaml` file with this command: `mm init`
    the result is:

added .tickets/issues/0177-mm-init-openai-compatible-services.yaml +103 -0 Viewed

new file mode 100644
@@ -0,0 +1,103 @@
	1	+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
	2	+id: 177
	3	+title: "mm init: OpenAI-compatible services — ask the endpoint and the API key's variable, list or type the model"
	4	+state: closed
	5	+author:
	6	+ name: k33g
	7	+ email: ph.charriere@gmail.com
	8	+createdAt: 2026-10-26T17:00:00.000Z
	9	+updatedAt: 2026-10-26T19:45:00.000Z
	10	+labels:
	11	+ - kind::feature
	12	+milestone: v1.0.3
	13	+related:
	14	+ - 176
	15	+ - 121
	16	+body: \|
	17	+ ## Need (2026-10-26)
	18	+ In `mm init`, `provider: openai` is the way to write an agent for ANY OpenAI-compatible service — OpenAI, Cerebras, Hugging Face, Groq… — not only api.openai.com. For that provider, `mm init` should:
	19	+ - ask the endpoint (default `https://api.openai.com/v1`) and write it as `baseUrl`;
	20	+ - ask the name of the environment variable that holds the API key (default `OPENAI_API_KEY`; e.g. `CEREBRAS_API_KEY`, `HF_TOKEN`) and write it as `apiKeyEnv`;
	21	+ - if that variable is set, list the models of that endpoint with the key; otherwise (or when the list fails) let the user type the model ID, saying why there is no list.
	22	+
	23	+ ## Today (checked in internal/agentinit on 2026-10-26)
	24	+ - The endpoint can only be changed with `--base-url`; no question.
	25	+ - The key is always read from `OPENAI_API_KEY` (`facts.go`, `Models`); `apiKeyEnv` is never asked nor written.
	26	+ - The model list is fetched with a 5-second timeout, the one used to probe local servers.
	27	+
	28	+ ## Reported by a user
	29	+ - "The first time I choose OpenAI in mm init, no model is offered. The second time, I get the whole list." Likely cause (not reproduced: no network to OpenAI in the sandbox): the 5 s timeout on a first, cold connection to api.openai.com (DNS + TLS); and the window does not show why the list is empty — only the line mode prints `ModelsError()`.
	30	+ - `gpt-4o-mini` works; `gpt-6-luna` fails: `400 … Function tools with reasoning_effort are not supported for gpt-6-luna in /v1/chat/completions. To use function tools, use /v1/responses or set reasoning_effort to 'none'.` mm speaks `/v1/chat/completions` only and always sends tools, so the fix is configuration: `sampling: reasoning_effort: none` (sent as written since ticket 176). Not verified (no key).
	31	+
	32	+ ## Plan
	33	+ 1. Two questions, for `provider: openai` only, between the server and the model: the endpoint, then the key's variable name (`--base-url` and a new `--api-key-env` answer them). A new endpoint or variable forgets the model.
	34	+ 2. The model list uses the key read from that variable; no key → no request, the model question becomes a text question that says the variable is not set; a failed list → the same, with the error in the question (window and line mode alike).
	35	+ 3. A longer timeout for the remote list (the `openai` provider) than for the local probes.
	36	+ 4. The written file: `baseUrl` (unless it is the default), `apiKeyEnv` (unless `OPENAI_API_KEY`), and a commented `# reasoning_effort: none` line in `sampling` with why.
	37	+ 5. Tests against a fake server (key sent, no key → text question, the error in the prompt, the file's keys); docs EN + FR (`mm init` how-to and CLI reference).
	38	+tasks:
	39	+ - id: 1
	40	+ title: "Endpoint and key-variable questions for provider openai (+ --api-key-env)"
	41	+ state: closed
	42	+ priority: none
	43	+ author:
	44	+ name: k33g
	45	+ email: ph.charriere@gmail.com
	46	+ createdAt: 2026-10-26T17:00:00.000Z
	47	+ - id: 2
	48	+ title: "Model list with that key; no key or a failed list: type the model ID, the reason shown"
	49	+ state: closed
	50	+ priority: none
	51	+ author:
	52	+ name: k33g
	53	+ email: ph.charriere@gmail.com
	54	+ createdAt: 2026-10-26T17:00:00.000Z
	55	+ - id: 3
	56	+ title: "Longer timeout for the remote model list"
	57	+ state: closed
	58	+ priority: none
	59	+ author:
	60	+ name: k33g
	61	+ email: ph.charriere@gmail.com
	62	+ createdAt: 2026-10-26T17:00:00.000Z
	63	+ - id: 4
	64	+ title: "Written file: baseUrl, apiKeyEnv, a commented reasoning_effort: none"
	65	+ state: closed
	66	+ priority: none
	67	+ author:
	68	+ name: k33g
	69	+ email: ph.charriere@gmail.com
	70	+ createdAt: 2026-10-26T17:00:00.000Z
	71	+ - id: 5
	72	+ title: "Tests and docs EN + FR"
	73	+ state: closed
	74	+ priority: none
	75	+ author:
	76	+ name: k33g
	77	+ email: ph.charriere@gmail.com
	78	+ createdAt: 2026-10-26T17:00:00.000Z
	79	+comments:
	80	+ - id: 1
	81	+ author:
	82	+ name: k33g
	83	+ email: ph.charriere@gmail.com
	84	+ createdAt: 2026-10-26T18:00:00.000Z
	85	+ body: \|
	86	+ Done (tasks 1–5).
	87	+
	88	+ - Questions (`internal/agentinit/flow.go`): for `provider: openai` only, `endpoint` (Text, default `https://api.openai.com/v1`, must start with http(s)://, trailing / removed) and `keyenv` (Text, default `OPENAI_API_KEY`, a shell variable name), between the server and the model. Flags: `--base-url` (existing) and the new `--api-key-env`. A new endpoint or variable forgets the model; another server forgets both (unless given by a flag). The OpenAI API's own address and `OPENAI_API_KEY` are stored as "default" (no `baseUrl` / `apiKeyEnv` written).
	89	+ - Model list (`facts.go`): `ListModels(ctx, provider, base, key)`; for openai the key is read from the named variable, and the service is not called at all when it is empty. Timeout 20 s for openai, 5 s for the local probes (was 5 s for all).
	90	+ - Why there is no list is now in the model question itself (`Model ID? (CEREBRAS_API_KEY is not set: no list — type it)` or `Model ID? (no list from <base>: <error>)`), so the window shows it too; the line mode's separate ℹ️ line and the now unused `Flow.ModelsError` were removed.
	91	+ - Written file (`render.go`): `baseUrl` and `apiKeyEnv` when not the defaults, the key comment names the variable, and a commented `# reasoning_effort: none` with its reason at the end of `sampling`.
	92	+ - Tests: `internal/agentinit/openai_flow_test.go`, 7 tests against a fake service (questions only for openai; list with the named key; no key → text question, service not called; 401 → reason in the question; answers checked; what a change forgets; the file's keys). `go test ./...`, `go vet ./...`, gofmt clean.
	93	+ - End to end with the binary and a local fake service: line mode without the key → "FAKE_SVC_KEY is not set: no list — type it", file written and loaded; `--yes` with the key → the service's models listed, `baseUrl`, `apiKeyEnv` and the hint in the file.
	94	+ - Docs EN + FR: reference/cli (`--base-url`, `--api-key-env`, `--model` rows), how-to/create-an-agent (variant "An OpenAI-compatible service"), how-to/run-with-an-openai-compatible-server (the mm init line, the reasoning_effort variant).
	95	+
	96	+ NOT verified: against the real OpenAI, Cerebras or Hugging Face APIs (no keys, no network in the sandbox); that 20 s fixes the reported "no model on the first run" (cause not reproduced); `reasoning_effort: none` with gpt-6-luna (reported by a user, not tried); the window driven live (its tests pass).
	97	+ - id: 2
	98	+ author:
	99	+ name: k33g
	100	+ email: ph.charriere@gmail.com
	101	+ createdAt: 2026-10-26T19:00:00.000Z
	102	+ body: \|
	103	+ Tested by the user with OpenAI and Mistral AI: works. Follow-up fixed: the window's text answers (endpoint, key variable, file) now have a caret — ←→, Home/End (Ctrl-A/Ctrl-E), Backspace before / Delete after the caret, typing inserts at it (`field` in internal/tui/initwin.go, 2 tests; docs EN + FR). Closed.

++# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
++id: 177
++title: "mm init: OpenAI-compatible services — ask the endpoint and the API key's variable, list or type the model"
++state: closed
++author:
++  name: k33g
++  email: ph.charriere@gmail.com
++createdAt: 2026-10-26T17:00:00.000Z
++updatedAt: 2026-10-26T19:45:00.000Z
++labels:
++  - kind::feature
++milestone: v1.0.3
++related:
++  - 176
++  - 121
++body: |
++  ## Need (2026-10-26)
++  In `mm init`, `provider: openai` is the way to write an agent for ANY OpenAI-compatible service — OpenAI, Cerebras, Hugging Face, Groq… — not only api.openai.com. For that provider, `mm init` should:
++  - ask the **endpoint** (default `https://api.openai.com/v1`) and write it as `baseUrl`;
++  - ask the **name of the environment variable** that holds the API key (default `OPENAI_API_KEY`; e.g. `CEREBRAS_API_KEY`, `HF_TOKEN`) and write it as `apiKeyEnv`;
++  - if that variable is set, **list the models** of that endpoint with the key; otherwise (or when the list fails) **let the user type the model ID**, saying why there is no list.
++
++  ## Today (checked in internal/agentinit on 2026-10-26)
++  - The endpoint can only be changed with `--base-url`; no question.
++  - The key is always read from `OPENAI_API_KEY` (`facts.go`, `Models`); `apiKeyEnv` is never asked nor written.
++  - The model list is fetched with a **5-second timeout**, the one used to probe local servers.
++
++  ## Reported by a user
++  - "The first time I choose OpenAI in mm init, no model is offered. The second time, I get the whole list." Likely cause (not reproduced: no network to OpenAI in the sandbox): the 5 s timeout on a first, cold connection to api.openai.com (DNS + TLS); and the **window does not show why** the list is empty — only the line mode prints `ModelsError()`.
++  - `gpt-4o-mini` works; `gpt-6-luna` fails: `400 … Function tools with reasoning_effort are not supported for gpt-6-luna in /v1/chat/completions. To use function tools, use /v1/responses or set reasoning_effort to 'none'.` mm speaks `/v1/chat/completions` only and always sends tools, so the fix is configuration: `sampling: reasoning_effort: none` (sent as written since ticket 176). Not verified (no key).
++
++  ## Plan
++  1. Two questions, for `provider: openai` only, between the server and the model: the endpoint, then the key's variable name (`--base-url` and a new `--api-key-env` answer them). A new endpoint or variable forgets the model.
++  2. The model list uses the key read from that variable; no key → no request, the model question becomes a text question that says the variable is not set; a failed list → the same, with the error in the question (window and line mode alike).
++  3. A longer timeout for the remote list (the `openai` provider) than for the local probes.
++  4. The written file: `baseUrl` (unless it is the default), `apiKeyEnv` (unless `OPENAI_API_KEY`), and a commented `# reasoning_effort: none` line in `sampling` with why.
++  5. Tests against a fake server (key sent, no key → text question, the error in the prompt, the file's keys); docs EN + FR (`mm init` how-to and CLI reference).
++tasks:
++  - id: 1
++    title: "Endpoint and key-variable questions for provider openai (+ --api-key-env)"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-26T17:00:00.000Z
++  - id: 2
++    title: "Model list with that key; no key or a failed list: type the model ID, the reason shown"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-26T17:00:00.000Z
++  - id: 3
++    title: "Longer timeout for the remote model list"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-26T17:00:00.000Z
++  - id: 4
++    title: "Written file: baseUrl, apiKeyEnv, a commented reasoning_effort: none"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-26T17:00:00.000Z
++  - id: 5
++    title: "Tests and docs EN + FR"
++    state: closed
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-26T17:00:00.000Z
++comments:
++  - id: 1
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-26T18:00:00.000Z
++    body: |
++      Done (tasks 1–5).
++
++      - **Questions** (`internal/agentinit/flow.go`): for `provider: openai` only, `endpoint` (Text, default `https://api.openai.com/v1`, must start with http(s)://, trailing / removed) and `keyenv` (Text, default `OPENAI_API_KEY`, a shell variable name), between the server and the model. Flags: `--base-url` (existing) and the new `--api-key-env`. A new endpoint or variable forgets the model; another server forgets both (unless given by a flag). The OpenAI API's own address and `OPENAI_API_KEY` are stored as "default" (no `baseUrl` / `apiKeyEnv` written).
++      - **Model list** (`facts.go`): `ListModels(ctx, provider, base, key)`; for openai the key is read from the named variable, and the service is not called at all when it is empty. Timeout 20 s for openai, 5 s for the local probes (was 5 s for all).
++      - **Why there is no list** is now in the model question itself (`Model ID? (CEREBRAS_API_KEY is not set: no list — type it)` or `Model ID? (no list from <base>: <error>)`), so the window shows it too; the line mode's separate ℹ️ line and the now unused `Flow.ModelsError` were removed.
++      - **Written file** (`render.go`): `baseUrl` and `apiKeyEnv` when not the defaults, the key comment names the variable, and a commented `# reasoning_effort: none` with its reason at the end of `sampling`.
++      - **Tests**: `internal/agentinit/openai_flow_test.go`, 7 tests against a fake service (questions only for openai; list with the named key; no key → text question, service not called; 401 → reason in the question; answers checked; what a change forgets; the file's keys). `go test ./...`, `go vet ./...`, gofmt clean.
++      - **End to end** with the binary and a local fake service: line mode without the key → "FAKE_SVC_KEY is not set: no list — type it", file written and loaded; `--yes` with the key → the service's models listed, `baseUrl`, `apiKeyEnv` and the hint in the file.
++      - **Docs** EN + FR: reference/cli (`--base-url`, `--api-key-env`, `--model` rows), how-to/create-an-agent (variant "An OpenAI-compatible service"), how-to/run-with-an-openai-compatible-server (the mm init line, the reasoning_effort variant).
++
++      NOT verified: against the real OpenAI, Cerebras or Hugging Face APIs (no keys, no network in the sandbox); that 20 s fixes the reported "no model on the first run" (cause not reproduced); `reasoning_effort: none` with gpt-6-luna (reported by a user, not tried); the window driven live (its tests pass).
++  - id: 2
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-26T19:00:00.000Z
++    body: |
++      Tested by the user with OpenAI and Mistral AI: works. Follow-up fixed: the window's text answers (endpoint, key variable, file) now have a caret — ←→, Home/End (Ctrl-A/Ctrl-E), Backspace before / Delete after the caret, typing inserts at it (`field` in internal/tui/initwin.go, 2 tests; docs EN + FR). Closed.

added .tickets/issues/0178-support-openai-responses-api.yaml +102 -0 Viewed

new file mode 100644
@@ -0,0 +1,102 @@
	1	+# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
	2	+id: 178
	3	+title: "Support OpenAI's Responses API, to keep reasoning and tools together (gpt-6-luna)"
	4	+state: open
	5	+author:
	6	+ name: k33g
	7	+ email: ph.charriere@gmail.com
	8	+createdAt: 2026-10-26T19:30:00.000Z
	9	+updatedAt: 2026-10-26T19:30:00.000Z
	10	+labels:
	11	+ - kind::feature
	12	+epic: Providers (LLM Engine)
	13	+related:
	14	+ - 176
	15	+ - 177
	16	+body: \|
	17	+ ## The problem (reported 2026-10-26)
	18	+ With `provider: openai`, `gpt-4o-mini` works but `gpt-6-luna` answers:
	19	+
	20	+ ```
	21	+ 400 Bad Request — "Function tools with reasoning_effort are not supported for gpt-6-luna in
	22	+ /v1/chat/completions. To use function tools, use /v1/responses or set reasoning_effort to 'none'."
	23	+ ```
	24	+
	25	+ - OpenAI has two APIs: Chat Completions (`/v1/chat/completions`, the one every "OpenAI-compatible" server speaks) and Responses (`/v1/responses`, the newer one OpenAI recommends for its reasoning models).
	26	+ - mm speaks only Chat Completions, through Genkit's `compat_oai` plugin.
	27	+ - An mm agent always sends tools (bash, the file tools…). For this model, OpenAI refuses tools + active reasoning on Chat Completions; the model's default reasoning applies when the file sets none.
	28	+
	29	+ ## What already works (configuration only)
	30	+ `sampling: reasoning_effort: none` — sent as written since ticket 176, written commented by `mm init` since ticket 177. The model then works with tools but does not reason. Not tried with gpt-6-luna (no key in the sandbox).
	31	+
	32	+ ## The goal
	33	+ Keep the reasoning AND the tools on such models: let mm speak the Responses API.
	34	+
	35	+ ## What exists (checked 2026-10-26)
	36	+ - Genkit Go v1.13.1, `plugins/compat_oai`: Chat Completions only (no Responses code; one comment links the Responses docs).
	37	+ - `github.com/openai/openai-go` v1.8.2, already in `go.mod`, has a `responses` package (`client.Responses.New` / `NewStreaming`).
	38	+ - So: a Genkit model of our own (`genkit.DefineModel` with mm's generate function) that converts Genkit's request and response to and from the Responses API with openai-go — the same seam `engine.Provider` already gives (one implementation per wire protocol).
	39	+
	40	+ ## What it must keep working (mm relies on all of it)
	41	+ - Streaming text chunks (watchdog, line mode, window, ACP) — Responses streams typed events (`response.output_text.delta`, `response.function_call_arguments.delta`, `response.completed`…), not Chat Completions chunks.
	42	+ - Tool calls and their results across the turns (`function_call` / `function_call_output` items instead of `tool_calls` / `role: tool` messages); the full conversation returned to mm (`engine.Generate`'s WrapModel capture).
	43	+ - The reasoning between turns: with `store: false`, the reasoning items must be sent back (`include: ["reasoning.encrypted_content"]`) or the model loses its thread after each tool call; or `previous_response_id` with `store: true` (the conversation then lives on OpenAI's side — to decide).
	44	+ - Reasoning shown as thoughts (`showThinking`), usage counts (`usage_update`, context %), the struggle signals, `structuredOutput` (`text.format: json_schema`), the context compression's summary request, `max_output_tokens` (Responses' name for the limit).
	45	+
	46	+ ## Options to decide
	47	+ - A — a new provider `openai-responses` (explicit, nothing changes for the existing `openai` files).
	48	+ - B — a key on the openai provider, e.g. `api: responses` (one provider, two wire protocols).
	49	+ - C — automatic: Responses for OpenAI's reasoning models only — rejected for now: a list of model names would age fast, and other OpenAI-compatible services do not all serve `/v1/responses` (which do — vLLM, Ollama, llama-server… — is NOT verified).
	50	+
	51	+ ## Not done here
	52	+ Nothing implemented. A real check needs an OpenAI key and a reasoning model.
	53	+tasks:
	54	+ - id: 1
	55	+ title: "Decide: a new provider (A) or an api key on openai (B); store false + encrypted reasoning, or previous_response_id"
	56	+ state: open
	57	+ priority: none
	58	+ author:
	59	+ name: k33g
	60	+ email: ph.charriere@gmail.com
	61	+ createdAt: 2026-10-26T19:30:00.000Z
	62	+ - id: 2
	63	+ title: "Check which OpenAI-compatible servers serve /v1/responses (vLLM, Ollama, llama-server…)"
	64	+ state: open
	65	+ priority: none
	66	+ author:
	67	+ name: k33g
	68	+ email: ph.charriere@gmail.com
	69	+ createdAt: 2026-10-26T19:30:00.000Z
	70	+ - id: 3
	71	+ title: "A Genkit model on openai-go's responses package: request/response conversion, streaming events, tool calls"
	72	+ state: open
	73	+ priority: none
	74	+ author:
	75	+ name: k33g
	76	+ email: ph.charriere@gmail.com
	77	+ createdAt: 2026-10-26T19:30:00.000Z
	78	+ - id: 4
	79	+ title: "Reasoning items carried between turns; thoughts, usage, structuredOutput, compaction, max_output_tokens"
	80	+ state: open
	81	+ priority: none
	82	+ author:
	83	+ name: k33g
	84	+ email: ph.charriere@gmail.com
	85	+ createdAt: 2026-10-26T19:30:00.000Z
	86	+ - id: 5
	87	+ title: "Tests against a fake /v1/responses server; a live check with a reasoning model"
	88	+ state: open
	89	+ priority: none
	90	+ author:
	91	+ name: k33g
	92	+ email: ph.charriere@gmail.com
	93	+ createdAt: 2026-10-26T19:30:00.000Z
	94	+ - id: 6
	95	+ title: "mm init and docs EN + FR"
	96	+ state: open
	97	+ priority: none
	98	+ author:
	99	+ name: k33g
	100	+ email: ph.charriere@gmail.com
	101	+ createdAt: 2026-10-26T19:30:00.000Z
	102	+comments: []

++# Managed by IssueSpec. Hand edits are welcome; keep the schema valid.
++id: 178
++title: "Support OpenAI's Responses API, to keep reasoning and tools together (gpt-6-luna)"
++state: open
++author:
++  name: k33g
++  email: ph.charriere@gmail.com
++createdAt: 2026-10-26T19:30:00.000Z
++updatedAt: 2026-10-26T19:30:00.000Z
++labels:
++  - kind::feature
++epic: Providers (LLM Engine)
++related:
++  - 176
++  - 177
++body: |
++  ## The problem (reported 2026-10-26)
++  With `provider: openai`, `gpt-4o-mini` works but `gpt-6-luna` answers:
++
++  ```
++  400 Bad Request — "Function tools with reasoning_effort are not supported for gpt-6-luna in
++  /v1/chat/completions. To use function tools, use /v1/responses or set reasoning_effort to 'none'."
++  ```
++
++  - OpenAI has two APIs: **Chat Completions** (`/v1/chat/completions`, the one every "OpenAI-compatible" server speaks) and **Responses** (`/v1/responses`, the newer one OpenAI recommends for its reasoning models).
++  - mm speaks **only Chat Completions**, through Genkit's `compat_oai` plugin.
++  - An mm agent **always sends tools** (bash, the file tools…). For this model, OpenAI refuses tools + active reasoning on Chat Completions; the model's default reasoning applies when the file sets none.
++
++  ## What already works (configuration only)
++  `sampling: reasoning_effort: none` — sent as written since ticket 176, written commented by `mm init` since ticket 177. The model then works with tools **but does not reason**. Not tried with gpt-6-luna (no key in the sandbox).
++
++  ## The goal
++  Keep the reasoning AND the tools on such models: let mm speak the **Responses API**.
++
++  ## What exists (checked 2026-10-26)
++  - Genkit Go v1.13.1, `plugins/compat_oai`: Chat Completions only (no Responses code; one comment links the Responses docs).
++  - `github.com/openai/openai-go` v1.8.2, already in `go.mod`, has a `responses` package (`client.Responses.New` / `NewStreaming`).
++  - So: a Genkit model of our own (`genkit.DefineModel` with mm's generate function) that converts Genkit's request and response to and from the Responses API with openai-go — the same seam `engine.Provider` already gives (one implementation per wire protocol).
++
++  ## What it must keep working (mm relies on all of it)
++  - Streaming text chunks (watchdog, line mode, window, ACP) — Responses streams typed events (`response.output_text.delta`, `response.function_call_arguments.delta`, `response.completed`…), not Chat Completions chunks.
++  - Tool calls and their results across the turns (`function_call` / `function_call_output` items instead of `tool_calls` / `role: tool` messages); the full conversation returned to mm (`engine.Generate`'s WrapModel capture).
++  - **The reasoning between turns**: with `store: false`, the reasoning items must be sent back (`include: ["reasoning.encrypted_content"]`) or the model loses its thread after each tool call; or `previous_response_id` with `store: true` (the conversation then lives on OpenAI's side — to decide).
++  - Reasoning shown as thoughts (`showThinking`), usage counts (`usage_update`, context %), the struggle signals, `structuredOutput` (`text.format: json_schema`), the context compression's summary request, `max_output_tokens` (Responses' name for the limit).
++
++  ## Options to decide
++  - **A — a new provider** `openai-responses` (explicit, nothing changes for the existing `openai` files).
++  - **B — a key on the openai provider**, e.g. `api: responses` (one provider, two wire protocols).
++  - **C — automatic**: Responses for OpenAI's reasoning models only — rejected for now: a list of model names would age fast, and other OpenAI-compatible services do not all serve `/v1/responses` (which do — vLLM, Ollama, llama-server… — is NOT verified).
++
++  ## Not done here
++  Nothing implemented. A real check needs an OpenAI key and a reasoning model.
++tasks:
++  - id: 1
++    title: "Decide: a new provider (A) or an api key on openai (B); store false + encrypted reasoning, or previous_response_id"
++    state: open
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-26T19:30:00.000Z
++  - id: 2
++    title: "Check which OpenAI-compatible servers serve /v1/responses (vLLM, Ollama, llama-server…)"
++    state: open
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-26T19:30:00.000Z
++  - id: 3
++    title: "A Genkit model on openai-go's responses package: request/response conversion, streaming events, tool calls"
++    state: open
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-26T19:30:00.000Z
++  - id: 4
++    title: "Reasoning items carried between turns; thoughts, usage, structuredOutput, compaction, max_output_tokens"
++    state: open
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-26T19:30:00.000Z
++  - id: 5
++    title: "Tests against a fake /v1/responses server; a live check with a reasoning model"
++    state: open
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-26T19:30:00.000Z
++  - id: 6
++    title: "mm init and docs EN + FR"
++    state: open
++    priority: none
++    author:
++      name: k33g
++      email: ph.charriere@gmail.com
++    createdAt: 2026-10-26T19:30:00.000Z
++comments: []

modified .tickets/milestones.yaml +3 -0 Viewed

@@ -47,3 +47,6 @@ milestones:
47	47	- title: v1.0.3
48	48	state: open
49	49	dueOn: 2026-10-31
	50	+ - title: v1.0.4
	51	+ state: open
	52	+ dueOn: 2026-10-31

    - title: v1.0.3
      state: open
      dueOn: 2026-10-31
++  - title: v1.0.4
++    state: open
++    dueOn: 2026-10-31

added demos/open-ai-compliant/README.md +19 -0 Viewed

new file mode 100644
@@ -0,0 +1,19 @@
	1	+# OpenAI compliant provider
	2	+
	3	+## OpenAI
	4	+
	5	+```bash
	6	+export OPENAI_API_KEY=sk-proj.....GO-kA
	7	+mm init
	8	+# name of the API key: OPENAI_API_KEY
	9	+# default endpoint: https://api.openai.com/v1
	10	+```
	11	+
	12	+## MistralAI
	13	+
	14	+```bash
	15	+export MISTRAL_API_KEY=mstrl_Ks......_3RFL8e
	16	+mm init
	17	+# name of the API key: MISTRAL_API_KEY
	18	+# default endpoint: https://api.mistral.ai/v1
	19	+```
		\ No newline at end of file

++# OpenAI compliant provider
++
++## OpenAI
++
++```bash
++export OPENAI_API_KEY=sk-proj.....GO-kA
++mm init
++# name of the API key: OPENAI_API_KEY
++# default endpoint: https://api.openai.com/v1
++```
++
++## MistralAI
++
++```bash
++export MISTRAL_API_KEY=mstrl_Ks......_3RFL8e
++mm init
++# name of the API key: MISTRAL_API_KEY
++# default endpoint: https://api.mistral.ai/v1
++```
 \ No newline at end of file

added demos/open-ai-compliant/agent.mistral.yaml +99 -0 Viewed

new file mode 100644
@@ -0,0 +1,99 @@
	1	+# developer — a coding agent: reads and changes the project, runs commands.
	2	+#
	3	+# Written by `mm init`. Every value says why it has that value; change
	4	+# them as you watch the agent work (guides/agent-settings-for-small-models.md).
	5	+#
	6	+# model family: other
	7	+# usage: developer
	8	+# machine: small — a laptop, models up to about 8B
	9	+# model facts: none — openai said nothing about this model beyond its name
	10	+#
	11	+# Where the numbers come from (the tag before their comment):
	12	+# [guide] the settings guide's row for this machine
	13	+# [rule] a rule of thumb, not specific to this model
	14	+#
	15	+# mm -config <this file> or mm -tui <this file>
	16	+
	17	+provider: openai
	18	+# The model. Chosen for it (family other): sampling (temperature, top_p,
	19	+# max_tokens). Change the model and redo them together — running mm init
	20	+# again is the quick way.
	21	+model: "ministral-8b-2512"
	22	+baseUrl: "https://api.mistral.ai/v1"
	23	+# The API key is read from MISTRAL_API_KEY: export it — it never goes in this file.
	24	+apiKeyEnv: MISTRAL_API_KEY
	25	+
	26	+# 0 = unknown: the generic openai provider cannot ask the server for its
	27	+# window, so compression counts messages (context.maxMessages) instead — set
	28	+# the window your server serves
	29	+contextWindow: 0
	30	+# your answer: every command and file change waits for your answer
	31	+yolo: false
	32	+# usage developer
	33	+bashTool: true
	34	+editTools: true
	35	+planTool: true
	36	+askUserTool: true
	37	+# no skills: add a skills/ folder and set skillsDir to use some (see
	38	+# docs/en/how-to/add-a-skill.md)
	39	+skillsDir: no-skills
	40	+# usage developer
	41	+maxTurns: 40
	42	+# [guide] settings guide, row "Small model (≤ 8B), short tasks"
	43	+maxOutput: 8000
	44	+firstTokenTimeout: 3m
	45	+# [guide] aligned on bashTimeout, as the settings guide says when bash is on
	46	+# (section 3)
	47	+watchdogTimeout: 30s
	48	+# [guide] settings guide, row "Small model (≤ 8B), short tasks"
	49	+bashTimeout: 30s
	50	+
	51	+sampling:
	52	+ # [rule] precise temperature for a developer agent without reasoning (family
	53	+ # other)
	54	+ temperature: 0.2
	55	+ # [rule] family other: skills/create-agent rule when nothing is published:
	56	+ # 0.7 / 0.9, 0.2 for a code agent without reasoning
	57	+ top_p: 0.9
	58	+ # [rule] room for an answer or a whole file in one tool call
	59	+ max_tokens: 4096
	60	+ # one tool call per reply: small models run a check before the change it
	61	+ # checks has landed otherwise
	62	+ parallel_tool_calls: false
	63	+ # Uncomment if the API answers "Function tools with reasoning_effort are not
	64	+ # supported … set reasoning_effort to 'none'" (seen with gpt-6-luna): mm uses
	65	+ # /v1/chat/completions and always sends tools. It turns the reasoning off.
	66	+ # reasoning_effort: none
	67	+
	68	+context:
	69	+ # compress the history before the window is full (settings guide, every row)
	70	+ enabled: true
	71	+ # mm's default: 75 % of the window the server serves, a quarter left for the
	72	+ # next question and its tool outputs
	73	+ threshold: 75
	74	+
	75	+agentsMd:
	76	+ # An AGENTS.md in the project directory is added to the system prompt.
	77	+ enabled: true
	78	+
	79	+acp:
	80	+ name: developer
	81	+ title: "Developer (developer)"
	82	+
	83	+system: \|
	84	+ You are a coding agent working in a terminal, in the user's project.
	85	+ You have a bash tool, file tools (read_file, write_file, edit_file) and a
	86	+ plan tool.
	87	+ Read before you change: look at the files and run the commands that tell
	88	+ you how the project works. Make the smallest change that does the job,
	89	+ then run the tests or the build to check it.
	90	+ Never state the output of a command you did not run in this answer.
	91	+ When a task has several steps, make a plan and update it as you go.
	92	+
	93	+# The family of this model is not in mm's table: the sampling is a generic
	94	+# starting point — check the model card.
	95	+# The server said nothing about this model beyond its name: the settings
	96	+# come from the family table.
	97	+
	98	+# Next: skills (docs/en/how-to/add-a-skill.md), MCP servers (docs/en/tutorials/mcp-getting-started.md),
	99	+# hooks (docs/en/how-to/guard-the-agent-with-hooks.md), sessions (docs/en/how-to/keep-sessions.md).

++# developer — a coding agent: reads and changes the project, runs commands.
++#
++# Written by `mm init`. Every value says why it has that value; change
++# them as you watch the agent work (guides/agent-settings-for-small-models.md).
++#
++#   model family: other
++#   usage:        developer
++#   machine:      small — a laptop, models up to about 8B
++#   model facts:  none — openai said nothing about this model beyond its name
++#
++# Where the numbers come from (the tag before their comment):
++#   [guide]    the settings guide's row for this machine
++#   [rule]     a rule of thumb, not specific to this model
++#
++#   mm -config <this file>          or   mm -tui <this file>
++
++provider: openai
++# The model. Chosen for it (family other): sampling (temperature, top_p,
++# max_tokens). Change the model and redo them together — running mm init
++# again is the quick way.
++model: "ministral-8b-2512"
++baseUrl: "https://api.mistral.ai/v1"
++# The API key is read from MISTRAL_API_KEY: export it — it never goes in this file.
++apiKeyEnv: MISTRAL_API_KEY
++
++# 0 = unknown: the generic openai provider cannot ask the server for its
++# window, so compression counts messages (context.maxMessages) instead — set
++# the window your server serves
++contextWindow: 0
++# your answer: every command and file change waits for your answer
++yolo: false
++# usage developer
++bashTool: true
++editTools: true
++planTool: true
++askUserTool: true
++# no skills: add a skills/ folder and set skillsDir to use some (see
++# docs/en/how-to/add-a-skill.md)
++skillsDir: no-skills
++# usage developer
++maxTurns: 40
++# [guide] settings guide, row "Small model (≤ 8B), short tasks"
++maxOutput: 8000
++firstTokenTimeout: 3m
++# [guide] aligned on bashTimeout, as the settings guide says when bash is on
++# (section 3)
++watchdogTimeout: 30s
++# [guide] settings guide, row "Small model (≤ 8B), short tasks"
++bashTimeout: 30s
++
++sampling:
++  # [rule] precise temperature for a developer agent without reasoning (family
++  # other)
++  temperature: 0.2
++  # [rule] family other: skills/create-agent rule when nothing is published:
++  # 0.7 / 0.9, 0.2 for a code agent without reasoning
++  top_p: 0.9
++  # [rule] room for an answer or a whole file in one tool call
++  max_tokens: 4096
++  # one tool call per reply: small models run a check before the change it
++  # checks has landed otherwise
++  parallel_tool_calls: false
++  # Uncomment if the API answers "Function tools with reasoning_effort are not
++  # supported … set reasoning_effort to 'none'" (seen with gpt-6-luna): mm uses
++  # /v1/chat/completions and always sends tools. It turns the reasoning off.
++  # reasoning_effort: none
++
++context:
++  # compress the history before the window is full (settings guide, every row)
++  enabled: true
++  # mm's default: 75 % of the window the server serves, a quarter left for the
++  # next question and its tool outputs
++  threshold: 75
++
++agentsMd:
++  # An AGENTS.md in the project directory is added to the system prompt.
++  enabled: true
++
++acp:
++  name: developer
++  title: "Developer (developer)"
++
++system: |
++  You are a coding agent working in a terminal, in the user's project.
++  You have a bash tool, file tools (read_file, write_file, edit_file) and a
++  plan tool.
++  Read before you change: look at the files and run the commands that tell
++  you how the project works. Make the smallest change that does the job,
++  then run the tests or the build to check it.
++  Never state the output of a command you did not run in this answer.
++  When a task has several steps, make a plan and update it as you go.
++
++# The family of this model is not in mm's table: the sampling is a generic
++# starting point — check the model card.
++# The server said nothing about this model beyond its name: the settings
++# come from the family table.
++
++# Next: skills (docs/en/how-to/add-a-skill.md), MCP servers (docs/en/tutorials/mcp-getting-started.md),
++# hooks (docs/en/how-to/guard-the-agent-with-hooks.md), sessions (docs/en/how-to/keep-sessions.md).

renamed demos/open-ai-compliant/agent.yaml +8 -4 Viewed

similarity index 90%
rename from demos/open-ai/agent.yaml
rename to demos/open-ai-compliant/agent.yaml
@@ -19,13 +19,13 @@ provider: openai
19	19	# The model. Chosen for it (family openai): sampling (temperature,
20	20	# max_completion_tokens). Change the model and redo them together — running
21	21	# mm init again is the quick way.
22		-model: "gpt-5.4"
	22	+model: "gpt-4.1"
23	23	# No baseUrl: openai's defaults apply. For the record:
24	24	# baseUrl: "https://api.openai.com/v1"
25	25	# (openai has no fallback address)
26		-# The key, when the server needs one, is read from OPENAI_API_KEY (or the
27		-# variable apiKeyEnv names): export it — it never goes in this file.
28		-# The default is the OpenAI API: set baseUrl for any other OpenAI-compatible server.
	26	+# The API key is read from OPENAI_API_KEY (or the variable apiKeyEnv names):
	27	+# export it — it never goes in this file.
	28	+# The default is the OpenAI API: set baseUrl for any other OpenAI-compatible service.
29	29
30	30	# 0 = unknown: the generic openai provider cannot ask the server for its
31	31	# window, so compression counts messages (context.maxMessages) instead — set
@@ -62,6 +62,10 @@ sampling:
62	62	# one tool call per reply: small models run a check before the change it
63	63	# checks has landed otherwise
64	64	parallel_tool_calls: false
	65	+ # Uncomment if the API answers "Function tools with reasoning_effort are not
	66	+ # supported … set reasoning_effort to 'none'" (seen with gpt-6-luna): mm uses
	67	+ # /v1/chat/completions and always sends tools. It turns the reasoning off.
	68	+ # reasoning_effort: none
65	69
66	70	context:
67	71	# compress the history before the window is full (settings guide, every row)

  # The model. Chosen for it (family openai): sampling (temperature,
  # max_completion_tokens). Change the model and redo them together — running
  # mm init again is the quick way.
--model: "gpt-5.4"
++model: "gpt-4.1"
  # No baseUrl: openai's defaults apply. For the record:
  # baseUrl: "https://api.openai.com/v1"
  # (openai has no fallback address)
--# The key, when the server needs one, is read from OPENAI_API_KEY (or the
++# The API key is read from OPENAI_API_KEY (or the variable apiKeyEnv names):
--# variable apiKeyEnv names): export it — it never goes in this file.
++# export it — it never goes in this file.
--# The default is the OpenAI API: set baseUrl for any other OpenAI-compatible server.
++# The default is the OpenAI API: set baseUrl for any other OpenAI-compatible service.
  # 0 = unknown: the generic openai provider cannot ask the server for its
  # window, so compression counts messages (context.maxMessages) instead — set
    # one tool call per reply: small models run a check before the change it
    # checks has landed otherwise
    parallel_tool_calls: false
++  # Uncomment if the API answers "Function tools with reasoning_effort are not
++  # supported … set reasoning_effort to 'none'" (seen with gpt-6-luna): mm uses
++  # /v1/chat/completions and always sends tools. It turns the reasoning off.
++  # reasoning_effort: none
  context:
    # compress the history before the window is full (settings guide, every row)

modified docs/en/how-to/create-an-agent.md +2 -1 Viewed

@@ -51,7 +51,7 @@ This guide shows how to write a new agent.yaml for a model and a job, with mm's
51	51	↑↓ choose · Enter next · Esc back · PgUp/PgDn preview · Ctrl-C quit
52	52	```
53	53
54		- `↑` `↓` choose, `Enter` answers and goes to the next question, `Esc` goes back one question, `PgUp` `PgDn` scroll the preview, `Ctrl-C` quits without writing. A text question (the name, the file) is typed; `Ctrl-U` clears it. On a screen narrower than 100 columns, `Tab` shows the preview in place of the question. When every question is answered, `Enter` writes the file.
	54	+ `↑` `↓` choose, `Enter` answers and goes to the next question, `Esc` goes back one question, `PgUp` `PgDn` scroll the preview, `Ctrl-C` quits without writing. A text question (the name, the file; the endpoint and the key's variable for `openai`) is typed; `←` `→` move the cursor in it, `Home` `End` go to its ends, `Delete` removes the character after the cursor, `Ctrl-U` clears it. On a screen narrower than 100 columns, `Tab` shows the preview in place of the question. When every question is answered, `Enter` writes the file.
55	55
56	56	The questions: the model server, the model (from the server's list), its family (detected from the name), what the agent is for (developer, reviewer, translator, writer, assistant, researcher, docs, json, coordinator), which machine runs the model (small, medium, large), whether to ask before each command or file change, the agent's name and the file.
57	57
@@ -79,6 +79,7 @@ This guide shows how to write a new agent.yaml for a model and a job, with mm's
79	79	```
80	80
81	81	- See the file without writing it. `-o -` prints it on stdout.
	82	+- An OpenAI-compatible service (OpenAI, Cerebras, Hugging Face…). Choose `openai`: two more questions follow, the endpoint (`https://api.openai.com/v1` by default) and the name of the variable that holds the API key (`OPENAI_API_KEY` by default, `CEREBRAS_API_KEY`, `HF_TOKEN`…). Export the key before running `mm init`: with it, the model question lists the service's models; without it, you type the model's ID (the question says the variable is not set). Flags: `--provider openai --base-url https://api.cerebras.ai/v1 --api-key-env CEREBRAS_API_KEY`. The file gets `baseUrl` and `apiKeyEnv`, and a commented `reasoning_effort: none` in `sampling`, to uncomment if the API answers "Function tools with reasoning_effort are not supported".
82	83	- A server elsewhere. `--base-url http://192.168.1.20:17434/v1`; it is written into the file as `baseUrl`. Without it, the file has no `baseUrl` and records the provider's default address and fallback as comments, ready to uncomment.
83	84	- A server that does not answer. mm says so and asks for the model's name: the settings then come from the family table only.
84	85	- An existing file. mm never overwrites it, and says so before asking anything: choose another name with `-o`.

      ↑↓ choose · Enter next · Esc back · PgUp/PgDn preview · Ctrl-C quit
     ```
--   `↑` `↓` choose, `Enter` answers and goes to the next question, `Esc` goes back one question, `PgUp` `PgDn` scroll the preview, `Ctrl-C` quits without writing. A text question (the name, the file) is typed; `Ctrl-U` clears it. On a screen narrower than 100 columns, `Tab` shows the preview in place of the question. When every question is answered, `Enter` writes the file.
++   `↑` `↓` choose, `Enter` answers and goes to the next question, `Esc` goes back one question, `PgUp` `PgDn` scroll the preview, `Ctrl-C` quits without writing. A text question (the name, the file; the endpoint and the key's variable for `openai`) is typed; `←` `→` move the cursor in it, `Home` `End` go to its ends, `Delete` removes the character after the cursor, `Ctrl-U` clears it. On a screen narrower than 100 columns, `Tab` shows the preview in place of the question. When every question is answered, `Enter` writes the file.
     The questions: the model server, the model (from the server's list), its family (detected from the name), what the agent is for (developer, reviewer, translator, writer, assistant, researcher, docs, json, coordinator), which machine runs the model (small, medium, large), whether to ask before each command or file change, the agent's name and the file.
    ```
  - **See the file without writing it.** `-o -` prints it on stdout.
++- **An OpenAI-compatible service (OpenAI, Cerebras, Hugging Face…).** Choose `openai`: two more questions follow, the endpoint (`https://api.openai.com/v1` by default) and the name of the variable that holds the API key (`OPENAI_API_KEY` by default, `CEREBRAS_API_KEY`, `HF_TOKEN`…). Export the key before running `mm init`: with it, the model question lists the service's models; without it, you type the model's ID (the question says the variable is not set). Flags: `--provider openai --base-url https://api.cerebras.ai/v1 --api-key-env CEREBRAS_API_KEY`. The file gets `baseUrl` and `apiKeyEnv`, and a commented `reasoning_effort: none` in `sampling`, to uncomment if the API answers "Function tools with reasoning_effort are not supported".
  - **A server elsewhere.** `--base-url http://192.168.1.20:17434/v1`; it is written into the file as `baseUrl`. Without it, the file has no `baseUrl` and records the provider's default address and fallback as comments, ready to uncomment.
  - **A server that does not answer.** mm says so and asks for the model's name: the settings then come from the family table only.
  - **An existing file.** mm never overwrites it, and says so before asking anything: choose another name with `-o`.

modified docs/en/how-to/run-with-an-openai-compatible-server.md +2 -1 Viewed

@@ -29,11 +29,12 @@ This guide shows how to run mini-me on a model server that has no dedicated prov
29	29	🤖 The command printed: `hello-from-openai-provider`
30	30	```
31	31
32		-`mm init --provider openai --base-url <address>` writes such a file for you (see [Create an agent with `mm init`](create-an-agent.md)).
	32	+`mm init` writes such a file for you: choose `openai`, then give the endpoint and the name of the variable that holds the key (or `--provider openai --base-url <address> --api-key-env <VARIABLE>`) (see [Create an agent with `mm init`](create-an-agent.md)).
33	33
34	34	## Variants
35	35
36	36	- The OpenAI API. No `baseUrl`, `OPENAI_API_KEY` exported, a model such as `gpt-5-mini`. Its reasoning models accept only the default temperature and take `max_completion_tokens` instead of `max_tokens`: write `temperature: 1.0` and `max_completion_tokens` in `sampling`. With `provider: openai` mm sends that block as written, without its own defaults (`top_p`, `max_tokens`), so remove a key the API still refuses and nothing else gets in the way. `mm init` writes such a block. (The request's content was checked against a local server; no call to the OpenAI API was made in this repository.)
	37	+- "Function tools with reasoning_effort are not supported". Some OpenAI models (seen: `gpt-6-luna`) refuse tools on `/v1/chat/completions` while they reason; mm uses that route and always sends tools. Add `reasoning_effort: none` to `sampling` (it turns the reasoning off). Not verified in this repository: reported by a user.
37	38	- The key is refused. The start-up warning says `openai: <baseUrl> refused the key (HTTP 401) — export OPENAI_API_KEY`.
38	39	- Nothing answers. The warning says `openai: nothing answers at <baseUrl> — check baseUrl: …`: the address must be the OpenAI-compatible one, usually ending in `/v1`.
39	40	- A server mm has a provider for. Prefer it (`dmr`, `llamacpp`, `llmman`, `ollama`): those know how the server reads a long prompt, turns reasoning off for a summary, unloads a model, and tell the window size. The generic provider does none of that.

     🤖 The command printed: `hello-from-openai-provider`
     ```
--`mm init --provider openai --base-url <address>` writes such a file for you (see [Create an agent with `mm init`](create-an-agent.md)).
++`mm init` writes such a file for you: choose `openai`, then give the endpoint and the name of the variable that holds the key (or `--provider openai --base-url <address> --api-key-env <VARIABLE>`) (see [Create an agent with `mm init`](create-an-agent.md)).
  ## Variants
  - **The OpenAI API.** No `baseUrl`, `OPENAI_API_KEY` exported, a model such as `gpt-5-mini`. Its reasoning models accept only the default temperature and take `max_completion_tokens` instead of `max_tokens`: write `temperature: 1.0` and `max_completion_tokens` in `sampling`. With `provider: openai` mm sends that block as written, without its own defaults (`top_p`, `max_tokens`), so remove a key the API still refuses and nothing else gets in the way. `mm init` writes such a block. (The request's content was checked against a local server; no call to the OpenAI API was made in this repository.)
++- **"Function tools with reasoning_effort are not supported".** Some OpenAI models (seen: `gpt-6-luna`) refuse tools on `/v1/chat/completions` while they reason; mm uses that route and always sends tools. Add `reasoning_effort: none` to `sampling` (it turns the reasoning off). Not verified in this repository: reported by a user.
  - **The key is refused.** The start-up warning says `openai: <baseUrl> refused the key (HTTP 401) — export OPENAI_API_KEY`.
  - **Nothing answers.** The warning says `openai: nothing answers at <baseUrl> — check baseUrl: …`: the address must be the OpenAI-compatible one, usually ending in `/v1`.
  - **A server mm has a provider for.** Prefer it (`dmr`, `llamacpp`, `llmman`, `ollama`): those know how the server reads a long prompt, turns reasoning off for a summary, unloads a model, and tell the window size. The generic provider does none of that.

modified docs/en/reference/cli.md +4 -3 Viewed

@@ -192,8 +192,9 @@ A turn cancelled by the user reports nothing. `demos/struggle-signals/` triggers
192	192	\| Flag \| Question when absent \| Default \|
193	193	\|------\|----------------------\|---------\|
194	194	\| `--provider` \| Model server: `llmman`, `ollama`, `dmr`, `llamacpp`, `openai` \| the first local one whose server answers, else `llmman` (`openai` is never probed: its default address is the OpenAI API) \|
195		-\| `--base-url` \| — (not asked) \| the provider's address, or its fallback when it does not answer \|
196		-\| `--model` \| Model: a list of the server's models (`/api/tags` for llmman and Ollama, `/models` for the others), or a name to type when the server does not answer \| the first of the list \|
	195	+\| `--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 \|
	196	+\| `--api-key-env` \| `openai` only: Environment variable that holds the API key \| `OPENAI_API_KEY`. Written as `apiKeyEnv` unless it is the default \|
	197	+\| `--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 \|
197	198	\| `--family` \| Model family \| detected from the model's name, then from the server's family name \|
198	199	\| `--usage` \| What the agent is for: `developer`, `reviewer`, `translator`, `writer`, `assistant`, `researcher`, `docs`, `json`, `coordinator` \| `developer` \|
199	200	\| `--machine` \| Which machine runs the model: `small`, `medium`, `large` \| `small` \|
@@ -206,7 +207,7 @@ A turn cancelled by the user reports nothing. `demos/struggle-signals/` triggers
206	207
207	208	\| Aspect \| Behaviour \|
208	209	\|--------\|-----------\|
209		-\| Window \| When stdin and stdout are terminals (and neither `--yes` nor `--line`): a full-screen window — steps (✓ answered, ▸ current, · to come, each answer beside its step), the question with the highlighted choice's description (models are listed by their short name, as in the window's header — `Qwen3.5-4B:Q4_K_M` — and the highlighted model's full identifier is shown under the list), and the file as it would be written with that choice (preview beside the question from 100 columns, in its place with `Tab` below). Keys: `↑` `↓` `Home` `End` choose, `Enter` answers, `y` / `n` answer the yes-or-no question, `Esc` takes the previous answer back (answers given by flags are skipped), `PgUp` `PgDn` scroll the preview, `Ctrl-U` clears a text answer, `Ctrl-C` quits without writing (`nothing written`, exit 1). A refused answer shows on the bottom rule. Once every question is answered, `Enter` writes the file. \|
	210	+\| Window \| When stdin and stdout are terminals (and neither `--yes` nor `--line`): a full-screen window — steps (✓ answered, ▸ current, · to come, each answer beside its step), the question with the highlighted choice's description (models are listed by their short name, as in the window's header — `Qwen3.5-4B:Q4_K_M` — and the highlighted model's full identifier is shown under the list), and the file as it would be written with that choice (preview beside the question from 100 columns, in its place with `Tab` below). Keys: `↑` `↓` `Home` `End` choose, `Enter` answers, `y` / `n` answer the yes-or-no question, `Esc` takes the previous answer back (answers given by flags are skipped), `PgUp` `PgDn` scroll the preview, `←` `→` move the cursor in a text answer (`Home` `End` or `Ctrl-A` `Ctrl-E` to its ends; typing inserts at the cursor, `Backspace` deletes before it, `Delete` after it), `Ctrl-U` clears it, `Ctrl-C` quits without writing (`nothing written`, exit 1). A refused answer shows on the bottom rule. Once every question is answered, `Enter` writes the file. \|
210	211	\| Line mode \| Otherwise: one question per screen line, numbered choices; a number, an option's text or another text answers; an empty line or `/skip` takes the default. Usages and machines are listed with a one-line description first. \|
211	212	\| Model facts \| llmman and Ollama: `POST /api/show` — the native window, the GGUF's sampling (`general.sampling.*`, Ollama's Modelfile parameters), the chat template's reasoning switch (`enable_thinking`, `reasoning_effort`) and its default. llama-server: `GET /props` — `n_ctx` and the chat template. Docker Model Runner and `openai`: none — the family table decides; `openai` lists its models with `OPENAI_API_KEY`. \|
212	213	\| Where each value comes from \| The server's facts, then the family table, then the usage, then the machine row; the table is `internal/agentinit/presets.yaml`, each entry with its source and, per sampling block, its `origin` (`measured`, `card` or `rule`). \|

  | Flag | Question when absent | Default |
  |------|----------------------|---------|
  | `--provider` | Model server: `llmman`, `ollama`, `dmr`, `llamacpp`, `openai` | the first local one whose server answers, else `llmman` (`openai` is never probed: its default address is the OpenAI API) |
--| `--base-url` | — (not asked) | the provider's address, or its fallback when it does not answer |
++| `--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 |
--| `--model` | Model: a list of the server's models (`/api/tags` for llmman and Ollama, `/models` for the others), or a name to type when the server does not answer | the first of the list |
++| `--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 |
  | `--usage` | What the agent is for: `developer`, `reviewer`, `translator`, `writer`, `assistant`, `researcher`, `docs`, `json`, `coordinator` | `developer` |
  | `--machine` | Which machine runs the model: `small`, `medium`, `large` | `small` |
  | Aspect | Behaviour |
  |--------|-----------|
--| Window | When stdin and stdout are terminals (and neither `--yes` nor `--line`): a full-screen window — steps (✓ answered, ▸ current, · to come, each answer beside its step), the question with the highlighted choice's description (models are listed by their short name, as in the window's header — `Qwen3.5-4B:Q4_K_M` — and the highlighted model's full identifier is shown under the list), and the file as it would be written with that choice (preview beside the question from 100 columns, in its place with `Tab` below). Keys: `↑` `↓` `Home` `End` choose, `Enter` answers, `y` / `n` answer the yes-or-no question, `Esc` takes the previous answer back (answers given by flags are skipped), `PgUp` `PgDn` scroll the preview, `Ctrl-U` clears a text answer, `Ctrl-C` quits without writing (`nothing written`, exit 1). A refused answer shows on the bottom rule. Once every question is answered, `Enter` writes the file. |
++| Window | When stdin and stdout are terminals (and neither `--yes` nor `--line`): a full-screen window — steps (✓ answered, ▸ current, · to come, each answer beside its step), the question with the highlighted choice's description (models are listed by their short name, as in the window's header — `Qwen3.5-4B:Q4_K_M` — and the highlighted model's full identifier is shown under the list), and the file as it would be written with that choice (preview beside the question from 100 columns, in its place with `Tab` below). Keys: `↑` `↓` `Home` `End` choose, `Enter` answers, `y` / `n` answer the yes-or-no question, `Esc` takes the previous answer back (answers given by flags are skipped), `PgUp` `PgDn` scroll the preview, `←` `→` move the cursor in a text answer (`Home` `End` or `Ctrl-A` `Ctrl-E` to its ends; typing inserts at the cursor, `Backspace` deletes before it, `Delete` after it), `Ctrl-U` clears it, `Ctrl-C` quits without writing (`nothing written`, exit 1). A refused answer shows on the bottom rule. Once every question is answered, `Enter` writes the file. |
  | Line mode | Otherwise: one question per screen line, numbered choices; a number, an option's text or another text answers; an empty line or `/skip` takes the default. Usages and machines are listed with a one-line description first. |
  | Model facts | llmman and Ollama: `POST /api/show` — the native window, the GGUF's sampling (`general.sampling.*`, Ollama's Modelfile parameters), the chat template's reasoning switch (`enable_thinking`, `reasoning_effort`) and its default. llama-server: `GET /props` — `n_ctx` and the chat template. Docker Model Runner and `openai`: none — the family table decides; `openai` lists its models with `OPENAI_API_KEY`. |
  | Where each value comes from | The server's facts, then the family table, then the usage, then the machine row; the table is `internal/agentinit/presets.yaml`, each entry with its source and, per sampling block, its `origin` (`measured`, `card` or `rule`). |

modified docs/fr/how-to/create-an-agent.md +2 -1 Viewed

@@ -51,7 +51,7 @@ Ce guide montre comment écrire un nouvel agent.yaml pour un modèle et une tâc
51	51	↑↓ choose · Enter next · Esc back · PgUp/PgDn preview · Ctrl-C quit
52	52	```
53	53
54		- `↑` `↓` choisissent, `Entrée` répond et passe à la question suivante, `Échap` revient d'une question en arrière, `PgUp` `PgDn` font défiler l'aperçu, `Ctrl-C` quitte sans rien écrire. Une question de texte (le nom, le fichier) se tape ; `Ctrl-U` l'efface. Sur un écran de moins de 100 colonnes, `Tab` affiche l'aperçu à la place de la question. Quand toutes les questions ont une réponse, `Entrée` écrit le fichier.
	54	+ `↑` `↓` choisissent, `Entrée` répond et passe à la question suivante, `Échap` revient d'une question en arrière, `PgUp` `PgDn` font défiler l'aperçu, `Ctrl-C` quitte sans rien écrire. Une question de texte (le nom, le fichier ; l'endpoint et la variable de la clé pour `openai`) se tape ; `←` `→` déplacent le curseur, `Début` `Fin` vont à ses bouts, `Suppr` efface le caractère après le curseur, `Ctrl-U` efface tout. Sur un écran de moins de 100 colonnes, `Tab` affiche l'aperçu à la place de la question. Quand toutes les questions ont une réponse, `Entrée` écrit le fichier.
55	55
56	56	Les questions : le serveur de modèles, le modèle (dans la liste du serveur), sa famille (détectée d'après le nom), à quoi sert l'agent (developer, reviewer, translator, writer, assistant, researcher, docs, json, coordinator), quelle machine fait tourner le modèle (small, medium, large), s'il faut demander avant chaque commande ou modification de fichier, le nom de l'agent et le fichier.
57	57
@@ -79,6 +79,7 @@ Ce guide montre comment écrire un nouvel agent.yaml pour un modèle et une tâc
79	79	```
80	80
81	81	- Voir le fichier sans l'écrire. `-o -` l'affiche sur la sortie standard.
	82	+- Un service compatible OpenAI (OpenAI, Cerebras, Hugging Face…). Choisissez `openai` : deux questions de plus suivent, l'endpoint (`https://api.openai.com/v1` par défaut) et le nom de la variable qui contient la clé d'API (`OPENAI_API_KEY` par défaut, `CEREBRAS_API_KEY`, `HF_TOKEN`…). Exportez la clé avant de lancer `mm init` : avec elle, la question du modèle liste les modèles du service ; sans elle, vous tapez l'ID du modèle (la question dit que la variable n'est pas renseignée). Options : `--provider openai --base-url https://api.cerebras.ai/v1 --api-key-env CEREBRAS_API_KEY`. Le fichier reçoit `baseUrl` et `apiKeyEnv`, et un `reasoning_effort: none` en commentaire dans `sampling`, à décommenter si l'API répond « Function tools with reasoning_effort are not supported ».
82	83	- Un serveur ailleurs. `--base-url http://192.168.1.20:17434/v1` ; il est écrit dans le fichier comme `baseUrl`. Sans lui, le fichier n'a pas de `baseUrl` et garde en commentaire l'adresse par défaut du fournisseur et son fallback, prêts à être décommentés.
83	84	- Un serveur qui ne répond pas. mm le dit et demande le nom du modèle : les réglages viennent alors de la seule table des familles.
84	85	- Un fichier existant. mm ne l'écrase jamais, et le dit avant de poser la moindre question : choisissez un autre nom avec `-o`.

      ↑↓ choose · Enter next · Esc back · PgUp/PgDn preview · Ctrl-C quit
     ```
--   `↑` `↓` choisissent, `Entrée` répond et passe à la question suivante, `Échap` revient d'une question en arrière, `PgUp` `PgDn` font défiler l'aperçu, `Ctrl-C` quitte sans rien écrire. Une question de texte (le nom, le fichier) se tape ; `Ctrl-U` l'efface. Sur un écran de moins de 100 colonnes, `Tab` affiche l'aperçu à la place de la question. Quand toutes les questions ont une réponse, `Entrée` écrit le fichier.
++   `↑` `↓` choisissent, `Entrée` répond et passe à la question suivante, `Échap` revient d'une question en arrière, `PgUp` `PgDn` font défiler l'aperçu, `Ctrl-C` quitte sans rien écrire. Une question de texte (le nom, le fichier ; l'endpoint et la variable de la clé pour `openai`) se tape ; `←` `→` déplacent le curseur, `Début` `Fin` vont à ses bouts, `Suppr` efface le caractère après le curseur, `Ctrl-U` efface tout. Sur un écran de moins de 100 colonnes, `Tab` affiche l'aperçu à la place de la question. Quand toutes les questions ont une réponse, `Entrée` écrit le fichier.
     Les questions : le serveur de modèles, le modèle (dans la liste du serveur), sa famille (détectée d'après le nom), à quoi sert l'agent (developer, reviewer, translator, writer, assistant, researcher, docs, json, coordinator), quelle machine fait tourner le modèle (small, medium, large), s'il faut demander avant chaque commande ou modification de fichier, le nom de l'agent et le fichier.
    ```
  - **Voir le fichier sans l'écrire.** `-o -` l'affiche sur la sortie standard.
++- **Un service compatible OpenAI (OpenAI, Cerebras, Hugging Face…).** Choisissez `openai` : deux questions de plus suivent, l'endpoint (`https://api.openai.com/v1` par défaut) et le nom de la variable qui contient la clé d'API (`OPENAI_API_KEY` par défaut, `CEREBRAS_API_KEY`, `HF_TOKEN`…). Exportez la clé avant de lancer `mm init` : avec elle, la question du modèle liste les modèles du service ; sans elle, vous tapez l'ID du modèle (la question dit que la variable n'est pas renseignée). Options : `--provider openai --base-url https://api.cerebras.ai/v1 --api-key-env CEREBRAS_API_KEY`. Le fichier reçoit `baseUrl` et `apiKeyEnv`, et un `reasoning_effort: none` en commentaire dans `sampling`, à décommenter si l'API répond « Function tools with reasoning_effort are not supported ».
  - **Un serveur ailleurs.** `--base-url http://192.168.1.20:17434/v1` ; il est écrit dans le fichier comme `baseUrl`. Sans lui, le fichier n'a pas de `baseUrl` et garde en commentaire l'adresse par défaut du fournisseur et son fallback, prêts à être décommentés.
  - **Un serveur qui ne répond pas.** mm le dit et demande le nom du modèle : les réglages viennent alors de la seule table des familles.
  - **Un fichier existant.** mm ne l'écrase jamais, et le dit avant de poser la moindre question : choisissez un autre nom avec `-o`.

modified docs/fr/how-to/run-with-an-openai-compatible-server.md +2 -1 Viewed

@@ -29,11 +29,12 @@ Ce guide montre comment faire tourner mini-me sur un serveur de modèles qui n'a
29	29	🤖 The command printed: `hello-from-openai-provider`
30	30	```
31	31
32		-`mm init --provider openai --base-url <adresse>` écrit ce fichier pour vous (voir [Créer un agent avec `mm init`](create-an-agent.md)).
	32	+`mm init` écrit ce fichier pour vous : choisissez `openai`, puis donnez l'endpoint et le nom de la variable qui contient la clé (ou `--provider openai --base-url <adresse> --api-key-env <VARIABLE>`) (voir [Créer un agent avec `mm init`](create-an-agent.md)).
33	33
34	34	## Variantes
35	35
36	36	- L'API d'OpenAI. Pas de `baseUrl`, `OPENAI_API_KEY` exportée, un modèle comme `gpt-5-mini`. Ses modèles à raisonnement n'acceptent que la température par défaut et prennent `max_completion_tokens` au lieu de `max_tokens` : écrivez `temperature: 1.0` et `max_completion_tokens` dans `sampling`. Avec `provider: openai`, mm envoie ce bloc tel quel, sans ses propres valeurs par défaut (`top_p`, `max_tokens`) : retirez seulement une clé que l'API refuserait encore. `mm init` écrit un tel bloc. (Le contenu de la requête a été vérifié contre un serveur local ; aucun appel à l'API OpenAI n'a été fait dans ce dépôt.)
	37	+- « Function tools with reasoning_effort are not supported ». Certains modèles OpenAI (vu : `gpt-6-luna`) refusent les outils sur `/v1/chat/completions` pendant qu'ils raisonnent ; mm utilise cette route et envoie toujours des outils. Ajoutez `reasoning_effort: none` à `sampling` (cela coupe le raisonnement). Non vérifié dans ce dépôt : signalé par un utilisateur.
37	38	- La clé est refusée. L'avertissement de démarrage indique `openai: <baseUrl> refused the key (HTTP 401) — export OPENAI_API_KEY`.
38	39	- Rien ne répond. L'avertissement indique `openai: nothing answers at <baseUrl> — check baseUrl: …` : l'adresse doit être celle compatible OpenAI, qui finit généralement par `/v1`.
39	40	- Un serveur pour lequel mm a un fournisseur. Préférez-le (`dmr`, `llamacpp`, `llmman`, `ollama`) : ceux-là savent comment le serveur lit un long prompt, coupent le raisonnement pour un résumé, déchargent un modèle et donnent la taille de la fenêtre. Le fournisseur générique ne fait rien de tout cela.

     🤖 The command printed: `hello-from-openai-provider`
     ```
--`mm init --provider openai --base-url <adresse>` écrit ce fichier pour vous (voir [Créer un agent avec `mm init`](create-an-agent.md)).
++`mm init` écrit ce fichier pour vous : choisissez `openai`, puis donnez l'endpoint et le nom de la variable qui contient la clé (ou `--provider openai --base-url <adresse> --api-key-env <VARIABLE>`) (voir [Créer un agent avec `mm init`](create-an-agent.md)).
  ## Variantes
  - **L'API d'OpenAI.** Pas de `baseUrl`, `OPENAI_API_KEY` exportée, un modèle comme `gpt-5-mini`. Ses modèles à raisonnement n'acceptent que la température par défaut et prennent `max_completion_tokens` au lieu de `max_tokens` : écrivez `temperature: 1.0` et `max_completion_tokens` dans `sampling`. Avec `provider: openai`, mm envoie ce bloc tel quel, sans ses propres valeurs par défaut (`top_p`, `max_tokens`) : retirez seulement une clé que l'API refuserait encore. `mm init` écrit un tel bloc. (Le contenu de la requête a été vérifié contre un serveur local ; aucun appel à l'API OpenAI n'a été fait dans ce dépôt.)
++- **« Function tools with reasoning_effort are not supported ».** Certains modèles OpenAI (vu : `gpt-6-luna`) refusent les outils sur `/v1/chat/completions` pendant qu'ils raisonnent ; mm utilise cette route et envoie toujours des outils. Ajoutez `reasoning_effort: none` à `sampling` (cela coupe le raisonnement). Non vérifié dans ce dépôt : signalé par un utilisateur.
  - **La clé est refusée.** L'avertissement de démarrage indique `openai: <baseUrl> refused the key (HTTP 401) — export OPENAI_API_KEY`.
  - **Rien ne répond.** L'avertissement indique `openai: nothing answers at <baseUrl> — check baseUrl: …` : l'adresse doit être celle compatible OpenAI, qui finit généralement par `/v1`.
  - **Un serveur pour lequel mm a un fournisseur.** Préférez-le (`dmr`, `llamacpp`, `llmman`, `ollama`) : ceux-là savent comment le serveur lit un long prompt, coupent le raisonnement pour un résumé, déchargent un modèle et donnent la taille de la fenêtre. Le fournisseur générique ne fait rien de tout cela.

modified docs/fr/reference/cli.md +4 -3 Viewed

@@ -192,8 +192,9 @@ Un tour annulé par l'utilisateur ne signale rien. `demos/struggle-signals/` dé
192	192	\| Option \| Question quand elle est absente \| Défaut \|
193	193	\|--------\|---------------------------------\|--------\|
194	194	\| `--provider` \| Serveur de modèles : `llmman`, `ollama`, `dmr`, `llamacpp`, `openai` \| le premier serveur local qui répond, sinon `llmman` (`openai` n'est jamais sondé : son adresse par défaut est l'API d'OpenAI) \|
195		-\| `--base-url` \| — (pas posée) \| l'adresse du fournisseur, ou son adresse de secours quand elle ne répond pas \|
196		-\| `--model` \| Modèle : la liste des modèles du serveur (`/api/tags` pour llmman et Ollama, `/models` pour les autres), ou un nom à taper quand le serveur ne répond pas \| le premier de la liste \|
	195	+\| `--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 \|
	196	+\| `--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 \|
	197	+\| `--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 \|
197	198	\| `--family` \| Famille du modèle \| détectée d'après le nom du modèle, puis d'après le nom de famille donné par le serveur \|
198	199	\| `--usage` \| À quoi sert l'agent : `developer`, `reviewer`, `translator`, `writer`, `assistant`, `researcher`, `docs`, `json`, `coordinator` \| `developer` \|
199	200	\| `--machine` \| Quelle machine fait tourner le modèle : `small`, `medium`, `large` \| `small` \|
@@ -206,7 +207,7 @@ Un tour annulé par l'utilisateur ne signale rien. `demos/struggle-signals/` dé
206	207
207	208	\| Aspect \| Comportement \|
208	209	\|--------\|--------------\|
209		-\| Fenêtre \| Quand l'entrée et la sortie standard sont des terminaux (et ni `--yes` ni `--line`) : une fenêtre plein écran — les étapes (✓ répondue, ▸ en cours, · à venir, chaque réponse à côté de son étape), la question avec la description du choix surligné (les modèles sont listés par leur nom court, comme dans l'en-tête de la fenêtre — `Qwen3.5-4B:Q4_K_M` — et l'identifiant complet du modèle surligné s'affiche sous la liste), et le fichier tel qu'il serait écrit avec ce choix (aperçu à côté de la question à partir de 100 colonnes, à sa place avec `Tab` en dessous). Touches : `↑` `↓` `Début` `Fin` choisissent, `Entrée` répond, `y` / `n` répondent à la question oui ou non, `Échap` reprend la réponse précédente (les réponses données par option sont sautées), `PgUp` `PgDn` font défiler l'aperçu, `Ctrl-U` efface une réponse de texte, `Ctrl-C` quitte sans rien écrire (`nothing written`, code 1). Une réponse refusée s'affiche sur le trait du bas. Quand toutes les questions ont une réponse, `Entrée` écrit le fichier. \|
	210	+\| Fenêtre \| Quand l'entrée et la sortie standard sont des terminaux (et ni `--yes` ni `--line`) : une fenêtre plein écran — les étapes (✓ répondue, ▸ en cours, · à venir, chaque réponse à côté de son étape), la question avec la description du choix surligné (les modèles sont listés par leur nom court, comme dans l'en-tête de la fenêtre — `Qwen3.5-4B:Q4_K_M` — et l'identifiant complet du modèle surligné s'affiche sous la liste), et le fichier tel qu'il serait écrit avec ce choix (aperçu à côté de la question à partir de 100 colonnes, à sa place avec `Tab` en dessous). Touches : `↑` `↓` `Début` `Fin` choisissent, `Entrée` répond, `y` / `n` répondent à la question oui ou non, `Échap` reprend la réponse précédente (les réponses données par option sont sautées), `PgUp` `PgDn` font défiler l'aperçu, `←` `→` déplacent le curseur dans une réponse de texte (`Début` `Fin` ou `Ctrl-A` `Ctrl-E` pour aller à ses bouts ; la saisie s'insère au curseur, `Retour arrière` efface avant lui, `Suppr` après), `Ctrl-U` efface une réponse de texte, `Ctrl-C` quitte sans rien écrire (`nothing written`, code 1). Une réponse refusée s'affiche sur le trait du bas. Quand toutes les questions ont une réponse, `Entrée` écrit le fichier. \|
210	211	\| Mode ligne \| Sinon : une question par ligne, choix numérotés ; un numéro, le texte d'une option ou un autre texte y répond ; une ligne vide ou `/skip` prend le défaut. Les usages et les machines sont d'abord listés avec une description d'une ligne. \|
211	212	\| Faits sur le modèle \| llmman et Ollama : `POST /api/show` — la fenêtre native, l'échantillonnage du GGUF (`general.sampling.*`, les paramètres du Modelfile d'Ollama), l'interrupteur de raisonnement du chat template (`enable_thinking`, `reasoning_effort`) et son défaut. llama-server : `GET /props` — `n_ctx` et le chat template. Docker Model Runner et `openai` : rien — la table des familles décide ; `openai` liste ses modèles avec `OPENAI_API_KEY`. \|
212	213	\| Origine de chaque valeur \| Les faits du serveur, puis la table des familles, puis l'usage, puis la ligne de la machine ; la table est `internal/agentinit/presets.yaml`, chaque entrée avec sa source et, par bloc d'échantillonnage, son `origin` (`measured`, `card` ou `rule`). \|

  | Option | Question quand elle est absente | Défaut |
  |--------|---------------------------------|--------|
  | `--provider` | Serveur de modèles : `llmman`, `ollama`, `dmr`, `llamacpp`, `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` | — (pas posée) | l'adresse du fournisseur, ou son adresse de secours quand elle ne répond pas |
++| `--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 |
--| `--model` | Modèle : la liste des modèles du serveur (`/api/tags` pour llmman et Ollama, `/models` pour les autres), ou un nom à taper quand le serveur ne répond pas | le premier de la liste |
++| `--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 |
  | `--usage` | À quoi sert l'agent : `developer`, `reviewer`, `translator`, `writer`, `assistant`, `researcher`, `docs`, `json`, `coordinator` | `developer` |
  | `--machine` | Quelle machine fait tourner le modèle : `small`, `medium`, `large` | `small` |
  | Aspect | Comportement |
  |--------|--------------|
--| Fenêtre | Quand l'entrée et la sortie standard sont des terminaux (et ni `--yes` ni `--line`) : une fenêtre plein écran — les étapes (✓ répondue, ▸ en cours, · à venir, chaque réponse à côté de son étape), la question avec la description du choix surligné (les modèles sont listés par leur nom court, comme dans l'en-tête de la fenêtre — `Qwen3.5-4B:Q4_K_M` — et l'identifiant complet du modèle surligné s'affiche sous la liste), et le fichier tel qu'il serait écrit avec ce choix (aperçu à côté de la question à partir de 100 colonnes, à sa place avec `Tab` en dessous). Touches : `↑` `↓` `Début` `Fin` choisissent, `Entrée` répond, `y` / `n` répondent à la question oui ou non, `Échap` reprend la réponse précédente (les réponses données par option sont sautées), `PgUp` `PgDn` font défiler l'aperçu, `Ctrl-U` efface une réponse de texte, `Ctrl-C` quitte sans rien écrire (`nothing written`, code 1). Une réponse refusée s'affiche sur le trait du bas. Quand toutes les questions ont une réponse, `Entrée` écrit le fichier. |
++| Fenêtre | Quand l'entrée et la sortie standard sont des terminaux (et ni `--yes` ni `--line`) : une fenêtre plein écran — les étapes (✓ répondue, ▸ en cours, · à venir, chaque réponse à côté de son étape), la question avec la description du choix surligné (les modèles sont listés par leur nom court, comme dans l'en-tête de la fenêtre — `Qwen3.5-4B:Q4_K_M` — et l'identifiant complet du modèle surligné s'affiche sous la liste), et le fichier tel qu'il serait écrit avec ce choix (aperçu à côté de la question à partir de 100 colonnes, à sa place avec `Tab` en dessous). Touches : `↑` `↓` `Début` `Fin` choisissent, `Entrée` répond, `y` / `n` répondent à la question oui ou non, `Échap` reprend la réponse précédente (les réponses données par option sont sautées), `PgUp` `PgDn` font défiler l'aperçu, `←` `→` déplacent le curseur dans une réponse de texte (`Début` `Fin` ou `Ctrl-A` `Ctrl-E` pour aller à ses bouts ; la saisie s'insère au curseur, `Retour arrière` efface avant lui, `Suppr` après), `Ctrl-U` efface une réponse de texte, `Ctrl-C` quitte sans rien écrire (`nothing written`, code 1). Une réponse refusée s'affiche sur le trait du bas. Quand toutes les questions ont une réponse, `Entrée` écrit le fichier. |
  | Mode ligne | Sinon : une question par ligne, choix numérotés ; un numéro, le texte d'une option ou un autre texte y répond ; une ligne vide ou `/skip` prend le défaut. Les usages et les machines sont d'abord listés avec une description d'une ligne. |
  | Faits sur le modèle | llmman et Ollama : `POST /api/show` — la fenêtre native, l'échantillonnage du GGUF (`general.sampling.*`, les paramètres du Modelfile d'Ollama), l'interrupteur de raisonnement du chat template (`enable_thinking`, `reasoning_effort`) et son défaut. llama-server : `GET /props` — `n_ctx` et le chat template. Docker Model Runner et `openai` : rien — la table des familles décide ; `openai` liste ses modèles avec `OPENAI_API_KEY`. |
  | Origine de chaque valeur | Les faits du serveur, puis la table des familles, puis l'usage, puis la ligne de la machine ; la table est `internal/agentinit/presets.yaml`, chaque entrée avec sa source et, par bloc d'échantillonnage, son `origin` (`measured`, `card` ou `rule`). |

modified guides/raccourcis-clavier-tui.md +2 -1 Viewed

@@ -109,7 +109,8 @@ Quand l'agent vous pose une question (son outil `ask_user`), un petit formulaire
109	109	\| `Échap` \| Revient d'une question en arrière, en reprenant sa réponse. \|
110	110	\| `PgUp` `PgDn` \| Font défiler l'aperçu. \|
111	111	\| `Tab` \| Sur un écran de moins de 100 colonnes : affiche l'aperçu à la place de la question, et inversement. \|
112		-\| `Ctrl-U` \| Efface une réponse de texte (le nom, le fichier). \|
	112	+\| `←` `→` \| Dans une réponse de texte (l'endpoint, la variable de la clé, le nom, le fichier) : déplacent le curseur ; la saisie s'insère là. `Début` `Fin` (ou `Ctrl-A` `Ctrl-E`) vont au début et à la fin ; `Retour arrière` efface avant le curseur, `Suppr` après. \|
	113	+\| `Ctrl-U` \| Efface une réponse de texte. \|
113	114	\| `Ctrl-C` \| Quitte sans rien écrire. \|
114	115
115	116	## Bon à savoir

  | `Échap` | Revient d'une question en arrière, en reprenant sa réponse. |
  | `PgUp` `PgDn` | Font défiler l'aperçu. |
  | `Tab` | Sur un écran de moins de 100 colonnes : affiche l'aperçu à la place de la question, et inversement. |
--| `Ctrl-U` | Efface une réponse de texte (le nom, le fichier). |
++| `←` `→` | Dans une réponse de texte (l'endpoint, la variable de la clé, le nom, le fichier) : déplacent le curseur ; la saisie s'insère là. `Début` `Fin` (ou `Ctrl-A` `Ctrl-E`) vont au début et à la fin ; `Retour arrière` efface avant le curseur, `Suppr` après. |
++| `Ctrl-U` | Efface une réponse de texte. |
  | `Ctrl-C` | Quitte sans rien écrire. |
  ## Bon à savoir

modified guides/tui-keyboard-shortcuts.md +2 -1 Viewed

@@ -109,7 +109,8 @@ When the agent asks you something (its `ask_user` tool), a small form replaces t
109	109	\| `Esc` \| Go back one question, taking its answer back. \|
110	110	\| `PgUp` `PgDn` \| Scroll the preview. \|
111	111	\| `Tab` \| On a screen narrower than 100 columns: show the preview in place of the question, and back. \|
112		-\| `Ctrl-U` \| Clear a text answer (the name, the file). \|
	112	+\| `←` `→` \| In a text answer (the endpoint, the key's variable, the name, the file): move the cursor; typing inserts there. `Home` `End` (or `Ctrl-A` `Ctrl-E`) go to the start and the end; `Backspace` deletes before the cursor, `Delete` after it. \|
	113	+\| `Ctrl-U` \| Clear a text answer. \|
113	114	\| `Ctrl-C` \| Quit without writing anything. \|
114	115
115	116	## Good to know

  | `Esc` | Go back one question, taking its answer back. |
  | `PgUp` `PgDn` | Scroll the preview. |
  | `Tab` | On a screen narrower than 100 columns: show the preview in place of the question, and back. |
--| `Ctrl-U` | Clear a text answer (the name, the file). |
++| `←` `→` | In a text answer (the endpoint, the key's variable, the name, the file): move the cursor; typing inserts there. `Home` `End` (or `Ctrl-A` `Ctrl-E`) go to the start and the end; `Backspace` deletes before the cursor, `Delete` after it. |
++| `Ctrl-U` | Clear a text answer. |
  | `Ctrl-C` | Quit without writing anything. |
  ## Good to know

modified internal/agentinit/decide.go +10 -7 Viewed

@@ -7,14 +7,17 @@ import (
7	7
8	8	// Choices are the wizard's answers.
9	9	type Choices struct {
10		- Provider string // dmr, llmman, ollama, llamacpp
	10	+ Provider string // llmman, ollama, dmr, llamacpp, openai
11	11	BaseURL string // "" = the provider's default
12		- Model string
13		- Usage string
14		- Machine string
15		- Family string // "" = detected from the model's name
16		- Name string // acp.name
17		- Yolo *bool // nil = the usage's default
	12	+ // APIKeyEnv names the variable holding the API key of an
	13	+ // OpenAI-compatible service (provider openai); "" = OPENAI_API_KEY.
	14	+ APIKeyEnv string
	15	+ Model string
	16	+ Usage string
	17	+ Machine string
	18	+ Family string // "" = detected from the model's name
	19	+ Name string // acp.name
	20	+ Yolo *bool // nil = the usage's default
18	21	}
19	22
20	23	// Setting is one value of the generated file with the reason it has that

  // Choices are the wizard's answers.
  type Choices struct {
--	Provider string // dmr, llmman, ollama, llamacpp
++	Provider string // llmman, ollama, dmr, llamacpp, openai
  	BaseURL  string // "" = the provider's default
--	Model    string
++	// APIKeyEnv names the variable holding the API key of an
--	Usage    string
++	// OpenAI-compatible service (provider openai); "" = OPENAI_API_KEY.
--	Machine  string
++	APIKeyEnv string
--	Family   string // "" = detected from the model's name
++	Model     string
--	Name     string // acp.name
++	Usage     string
--	Yolo     *bool  // nil = the usage's default
++	Machine   string
++	Family    string // "" = detected from the model's name
++	Name      string // acp.name
++	Yolo      *bool  // nil = the usage's default
  }
  // Setting is one value of the generated file with the reason it has that

modified internal/agentinit/facts.go +30 -4 Viewed

@@ -89,12 +89,38 @@ func call(ctx context.Context, method, url string, body any, timeout time.Durati
89	89	}
90	90
91	91	// Models lists the models a server serves: llmman's and Ollama's /api/tags,
92		-// the OpenAI /models of the others. Sorted.
	92	+// the OpenAI /models of the others. Sorted. For `openai` the key is read
	93	+// from OPENAI_API_KEY; ListModels takes it from elsewhere.
93	94	//
94	95	// names, err := agentinit.Models(ctx, "llmman", "http://127.0.0.1:17434/v1")
95	96	func Models(ctx context.Context, provider, base string) ([]string, error) {
	97	+ key := ""
96	98	if provider == "openai" {
97		- ctx = withKey(ctx, os.Getenv("OPENAI_API_KEY"))
	99	+ key = os.Getenv("OPENAI_API_KEY")
	100	+ }
	101	+ return ListModels(ctx, provider, base, key)
	102	+}
	103	+
	104	+// Timeouts of a model list: a local server answers at once or is not
	105	+// there — the first question probes four of them, so it must stay short; a
	106	+// remote service (provider openai) may take seconds on a first, cold
	107	+// connection (DNS, TLS). Ticket 177: with 5 s, the OpenAI list came back
	108	+// empty on a first `mm init` and full on the second.
	109	+const (
	110	+ localListTimeout = 5 * time.Second
	111	+ remoteListTimeout = 20 * time.Second
	112	+)
	113	+
	114	+// ListModels is Models with the API key given (sent as a Bearer token when
	115	+// not empty) — the key of an OpenAI-compatible service, read from the
	116	+// variable the user named.
	117	+//
	118	+// names, err := agentinit.ListModels(ctx, "openai", "https://api.cerebras.ai/v1", os.Getenv("CEREBRAS_API_KEY"))
	119	+func ListModels(ctx context.Context, provider, base, key string) ([]string, error) {
	120	+ ctx = withKey(ctx, key)
	121	+ timeout := localListTimeout
	122	+ if provider == "openai" {
	123	+ timeout = remoteListTimeout
98	124	}
99	125	var names []string
100	126	switch provider {
@@ -104,7 +130,7 @@ func Models(ctx context.Context, provider, base string) ([]string, error) {
104	130	Name string `json:"name"`
105	131	} `json:"models"`
106	132	}
107		- if err := call(ctx, http.MethodGet, root(base)+"/api/tags", nil, 5*time.Second, &tags); err != nil {
	133	+ if err := call(ctx, http.MethodGet, root(base)+"/api/tags", nil, timeout, &tags); err != nil {
108	134	return nil, err
109	135	}
110	136	for _, m := range tags.Models {
@@ -116,7 +142,7 @@ func Models(ctx context.Context, provider, base string) ([]string, error) {
116	142	ID string `json:"id"`
117	143	} `json:"data"`
118	144	}
119		- if err := call(ctx, http.MethodGet, strings.TrimRight(base, "/")+"/models", nil, 5*time.Second, &list); err != nil {
	145	+ if err := call(ctx, http.MethodGet, strings.TrimRight(base, "/")+"/models", nil, timeout, &list); err != nil {
120	146	return nil, err
121	147	}
122	148	for _, m := range list.Data {

  }
  // Models lists the models a server serves: llmman's and Ollama's /api/tags,
--// the OpenAI /models of the others. Sorted.
++// the OpenAI /models of the others. Sorted. For `openai` the key is read
++// from OPENAI_API_KEY; ListModels takes it from elsewhere.
  //
  //	names, err := agentinit.Models(ctx, "llmman", "http://127.0.0.1:17434/v1")
  func Models(ctx context.Context, provider, base string) ([]string, error) {
++	key := ""
  	if provider == "openai" {
--		ctx = withKey(ctx, os.Getenv("OPENAI_API_KEY"))
++		key = os.Getenv("OPENAI_API_KEY")
++	}
++	return ListModels(ctx, provider, base, key)
++}
++
++// Timeouts of a model list: a local server answers at once or is not
++// there — the first question probes four of them, so it must stay short; a
++// remote service (provider openai) may take seconds on a first, cold
++// connection (DNS, TLS). Ticket 177: with 5 s, the OpenAI list came back
++// empty on a first `mm init` and full on the second.
++const (
++	localListTimeout  = 5 * time.Second
++	remoteListTimeout = 20 * time.Second
++)
++
++// ListModels is Models with the API key given (sent as a Bearer token when
++// not empty) — the key of an OpenAI-compatible service, read from the
++// variable the user named.
++//
++//	names, err := agentinit.ListModels(ctx, "openai", "https://api.cerebras.ai/v1", os.Getenv("CEREBRAS_API_KEY"))
++func ListModels(ctx context.Context, provider, base, key string) ([]string, error) {
++	ctx = withKey(ctx, key)
++	timeout := localListTimeout
++	if provider == "openai" {
++		timeout = remoteListTimeout
  	}
  	var names []string
  	switch provider {
  				Name string `json:"name"`
  			} `json:"models"`
  		}
--		if err := call(ctx, http.MethodGet, root(base)+"/api/tags", nil, 5*time.Second, &tags); err != nil {
++		if err := call(ctx, http.MethodGet, root(base)+"/api/tags", nil, timeout, &tags); err != nil {
  			return nil, err
  		}
  		for _, m := range tags.Models {
  				ID string `json:"id"`
  			} `json:"data"`
  		}
--		if err := call(ctx, http.MethodGet, strings.TrimRight(base, "/")+"/models", nil, 5*time.Second, &list); err != nil {
++		if err := call(ctx, http.MethodGet, strings.TrimRight(base, "/")+"/models", nil, timeout, &list); err != nil {
  			return nil, err
  		}
  		for _, m := range list.Data {

modified internal/agentinit/flow.go +114 -15 Viewed