What

reasoning_effort was only mapped to the binary enable_thinking toggle and otherwise reached Go-side templates — it was never forwarded to the backend. So jinja-templated models whose chat template keys on reasoning_effort (gpt-oss / Harmony, LFM2.5) could not be driven by it. In particular LFM2.5 ignores enable_thinking and keeps emitting <think>, but honors reasoning_effort.

This PR forwards the effective reasoning_effort to the backend as a chat_template_kwarg (mirroring how enable_thinking is already forwarded), and adds a config-level default so it can be pinned per model / per realtime pipeline.

Changes

• C++ (grpc-server.cpp): forward metadata["reasoning_effort"] → body_json["chat_template_kwargs"]["reasoning_effort"] in both the streaming and non-streaming predict paths, next to enable_thinking.
• Go (options.go): put the effective reasoning_effort into PredictOptions.Metadata.
• Config: ModelConfig.reasoning_effort (model default) and Pipeline.reasoning_effort (realtime), resolved by ModelConfig.ApplyReasoningEffort — a per-request value overrides the config default; none→disable, level→enable, and an operator's reasoning.disable: true still wins. request.go now uses that shared helper (no behavior change for the existing per-request path).
• Realtime: apply Pipeline.ReasoningEffort to the pipeline's LLM (per-session copy) and surface it on the template input.
• Registry entries + docs.

# model default
name: lfm2.5
reasoning_effort: none

# realtime pipeline (overrides the LLM's own)
name: gpt-realtime
pipeline:
  llm: lfm2.5
  reasoning_effort: none

Test plan

• ApplyReasoningEffort unit tests (request>config precedence, none/level mapping, operator-disable wins).
• gRPCPredictOpts forwards reasoning_effort into metadata (and omits when empty).
• request.go refactor — middleware tests green.
• realtime applyPipelineReasoning test.
• gofmt / golangci-lint clean on changed packages.
• Live check against LFM2.5: reasoning_effort: none suppresses <think> (C++ path — verifies on CI build / live).

Assisted-by: Claude:claude-opus-4-8 go test, golangci-lint

🤖 Generated with Claude Code