[watchme35]

Description

Fixes #402.

The base AIProvider.convert_request in providers/lib_llm_ext.py built the request body from model, messages, max_tokens, tools and tool_choice only, silently dropping request.reasoning_mode. As a result, the OpenAIAPI provider (and any provider delegating to the base implementation) sent no reasoning hint to OpenAI-compatible backends such as llama.cpp server, vLLM, or Strata.

Thinking-capable models then ran at the backend's default thinking level. Thinking tokens share the output budget, so long thinking chains exhaust it: finish_reason=length with empty content, and the user sees the "reasoning exceeded the token limit" message - which advises lowering reasoningMode, a setting that had no effect on this provider.

The fix forwards the mode as extra_body.reasoning_effort, the parameter these servers accept ("none" disables thinking). Plain OpenAI chat-completions endpoints ignore unknown fields, so existing users are unaffected. The OpenAI, OpenRouter and ASIonE providers already map the mode themselves:

• OpenRouter merges the base extra_body with its own reasoning object (both keys are consistent);
• ASIonE replaces extra_body with its thinking budget exactly as before.

Three unit tests are added for the base provider: forwarding medium and none, and omitting extra_body entirely when reasoning_mode is unset.

Review follow-up (thanks @timur-ashkenov): with the shipped default reasoningMode: medium, the base provider sends reasoning_effort to every model, including non-thinking models that reject the parameter (non-reasoning OpenAI models answer 400: "reasoning.effort is not supported with this model"). Added a supportsReasoningEffort config flag (default true): set to false, the parameter is omitted entirely, so such deployments keep the pre-PR request shape. OpenRouter and ASIOne are unaffected - they send their own reasoning parameters as before. Two more unit tests cover the omitted and sent cases.

How Has This Been Tested?

• Full unit suite on main @ fccc74d with the patch applied: pytest Autotests/unit -> 147 passed (142 pre-existing + 5 new).
• The identical mechanism (extra_body={"reasoning_effort": ...}) has been running in a production Omega v0.1.20 deployment (Docker, OpenAIAPI provider against a local llama.cpp/Strata proxy) since 2026-10-14, where it resolved the empty-response failures described in [OMEGA-522] OpenAIAPI provider drops reasoningMode from the request body (base AIProvider.convert_request ignores request.reasoning_mode) #402.

Checklist

• PR contains autogenerated code
• Self-review completed
• Test scenarios above are passed with the version of the code from PR

[timur-ashkenov]

@watchme35 With the default reasoningMode: medium, this now sends reasoning_effort to every model using the base provider, including non-thinking models. Could we preserve compatibility by allowing this parameter to be omitted for unsupported models, and add coverage for that case?

[watchme35]

Good catch — thanks for the review. Added a supportsReasoningEffort config flag (default true): set to false, the base provider omits reasoning_effort entirely, so deployments pointing OpenAIAPI at models that reject the parameter (non-reasoning OpenAI models answer 400: reasoning.effort is not supported with this model) keep working exactly as before this PR. OpenRouter and ASIOne are unaffected — they send their own reasoning parameters as before. Added unit tests for both the omitted and sent cases; full unit suite passes (147).