Local reviewer sends reasoning_effort to every ollama model, 400-ing on models without thinking support

[atomantic]

Problem

runToolFreeLocalCompletion in server/services/codeReview.js attaches reasoning_effort to the OpenAI-compatible request whenever normalizeReviewerEffort(effort, backend) returns a level:

const resolvedEffort = normalizeReviewerEffort(effort, backend) || null
...
...(resolvedEffort ? { reasoning_effort: resolvedEffort } : {}),

That gate is keyed on the backend, not on the model. Ollama translates reasoning_effort into its thinking parameter and hard-rejects it for a model that does not advertise thinking support:

ollama API error 400: {"error":{"message":"\"<model>\" does not support thinking",
"type":"invalid_request_error","param":null,"code":null}}

So a claim/PR run pinned to ollama[<non-thinking model>]~effort=low can never get a verdict: run-local-code-review.mjs exits 1 with ok:false, and a REQUIRED reviewer configured that way makes every run review-blocked, leaving PRs open and unmergeable. Reproduced deterministically (two consecutive attempts) on a real claim run against a gemma-family coder GGUF; PR #6049 is the PR it blocked.

The doc comment on runLocalCodeReview already states the intended behavior — "Omitted from the body entirely when unset or not a level this backend accepts — a non-reasoning [model] …" — so this is the implementation not matching its own documented contract, and the contract itself should say model, not backend.

Note the failure is easy to miss in the claim-comment path: runLocalClaimCommentReview returns early on an empty comment list before it ever reaches the API, so the tool-free comment gate passes with the same model/effort that then 400s on the code review.

Work

• Resolve thinking support per model before attaching reasoning_effort. PortOS already knows this: server/services/localLlm.js maps the ollama capability thinking -> reasoning (see the capability map around line 580), so the model list carries the answer.
• When the resolved model does not support it, drop reasoning_effort from the body and proceed with the review rather than failing — a non-reasoning model still returns usable findings.
• Fail closed only when the capability genuinely cannot be resolved AND the request still errors; keep the existing ok:false shape for that case so review-blocked still works.
• Correct the runLocalCodeReview doc comment to say the omission is decided per model.

Acceptance criteria

• runLocalCodeReview({ backend: 'ollama', model: <non-thinking model>, effort: 'low', diff }) returns ok:true with findings instead of a 400.
• A test pins the body shape both ways: reasoning_effort present for a thinking-capable model, absent for one without the capability.
• A model whose capabilities cannot be fetched does not regress the current behavior for thinking-capable models.

Files

• server/services/codeReview.js (runToolFreeLocalCompletion, runLocalCodeReview, normalizeReviewerEffort)
• server/services/localLlm.js (capability map, already present)
• server/services/codeReview.test.js