Add explicit qwen_chat thinking_mode wire policy (server_default/enabled/disabled)

[Yifan Yang (Yif-Yang)]

Resolves the tension between #28 and #90.

The problem

chat_template_kwargs is a vLLM/SGLang extension, not part of the OpenAI chat-completions API. OpenAI, Azure, and strict OpenAI-compatible gateways reject the unknown top-level body field with HTTP 400, and non-Qwen vLLM models served with it can emit <think> output with no <answer> tag — the acc=0.000 reported in #28.

c31c50b fixed that by only emitting the field when thinking was enabled. That closed #28, but it also means enable_thinking=false omits the field rather than sending false, so the client has no supported way to explicitly disable thinking — the request in #90 (and #109).

Both issues are legitimate, and neither is satisfiable by a boolean: the protocol has three states, and which one is correct depends on the serving stack, not on a user preference.

The change

A three-state, role-scoped thinking_mode:

mode	on the wire
server_default (default)	chat_template_kwargs not sent — the server's chat template decides
enabled	chat_template_kwargs: {"enable_thinking": true}
disabled	chat_template_kwargs: {"enable_thinking": false}

• use local llm train ,acc is 0 #28 stays fixed: server_default sends exactly the bytes every existing deployment sends today.
• Clarification request: Qwen enable_thinking setting used for reported ALFWorld results #90 is satisfied: disabled sends the explicit false.
• Zero config regressions: the legacy enable_thinking boolean keeps its historical wire meaning (true → send true, false → omit). No existing YAML, env var, or programmatic call changes behavior.
• Setting both keys to conflicting values raises ValueError rather than silently picking a winner, and unknown tokens raise too — a typo must never silently flip a reproducibility control.
• None still means "do not modify" in configure_qwen_chat, so a missing YAML key cannot clear env-derived configuration.

Because server_default delegates a result-affecting choice to the server's chat template, the backend warns once per role when a request goes out under it, and the resolved per-role mode is recorded in the run's config.json under resolved_qwen_thinking_modes — so a run's thinking policy is always recoverable from its artifacts.

Rejected alternatives

• Default to disabled — would re-break use local llm train ,acc is 0 #28 for strict gateways with no config change.
• Sniff base_url for localhost — deployment topology is not a capability signal; self-hosted vLLM often sits behind a public domain.
• Match on model name — capability belongs to the serving stack and template, not the model ID; aliases are common and hosted Qwen gateways may still reject vLLM extensions.
• Catch the 400 and retry without the field — silently abandons the reproducibility guarantee the user explicitly asked for.

Surface

model.qwen_chat_thinking_mode plus optimizer_/target_ variants, wired through YAML, --cfg-options, scripts/train.py CLI, eval_only.py, and QWEN_CHAT_THINKING_MODE env (role-scoped variants take precedence). Docs updated in docs/reference/config.md, docs/guide/configuration.md, and .env.example — including the contradiction between "local vLLM endpoint" and "OpenAI-compatible" wording, which is precisely why the wire policy cannot be inferred and must be explicit.

Tests

20 new tests covering all three payload states, the one-time warning, role/shared env precedence, legacy-key mapping, conflict detection, and invalid-token rejection. Full suite: 1134 passed, 10 skipped.

Supersedes #109 — credit to dylan (@LESdylan) for identifying the gap; the landed fix is the three-state enum rather than an unconditional emit, so #28 does not regress.