improve(skills/orchestrator): Nemotron COMPLEX review cycle (hour 13 UTC)

[raelli]

Automated self-improvement pass on src/skills/orchestrator/index.js. Reviewed by Nemotron COMPLEX — tests pass.

Change

Coerce a non-string query value to a string via safeStringify before pushing it onto the LLM messages array. The OpenAI/LiteLLM API requires content to be a string; passing a raw object or number would silently produce a malformed request. The default query = "" already makes this a no-op for the common case; the coercion only fires when a caller supplies a non-string value.

Why only this change

The Nemotron draft also suggested:

• Catching callLLM errors in the orchestrator — rejected because dispatchSSE already centralises error handling and emits a spec-compliant { error: { code, message } } frame on throw; catching internally would swallow the real message and emit a fake-success frame instead.
• Catching skill.handle errors — same reasoning as above.
• Swapping Object.hasOwn for hasOwnProperty.call — Node 18+ is already required (husky 9 constraint), so this is churn with no correctness gain.
• Handling a pre-parsed-object rawArgs — the OpenAI tool-call contract always delivers a JSON string here; adding untested handling for a hypothetical violates the "no new features" constraint.

The review verdict was null (router timed out at 90 s), so only the unambiguous, well-scoped suggestion was applied.

750/750 tests pass.

---

Generated by Claude Code