/code-review-runner

Deterministic code review with LLM-powered findings. Two-tier architecture:

• T0 (deterministic): best-practices-* validators, ruff lint, compile() check, file limits
• T1 (LLM): codex/scillm review with structured findings prompt

Each finding is scored by severity. Suggested fixes remain advisory because the runner does not apply them; current-tree compilation or DoD results cannot validate a proposed fix. Critical and major findings fail the review until they are reconciled.

SciLLM requests use the active proxy key from the environment or running proxy container and always send X-Caller-Skill: code-review-runner. A 401 from a stale environment key triggers one retry with the running proxy container key. The Codex backend uses the current one-shot gpt-5.5 route and omits temperature and max_tokens as required by the SciLLM paved path. An HTTP failure, empty assistant response, or malformed findings payload makes the review status error and the CLI exits nonzero; zero provider output must never become PASS.