raytsystem Run Review

Inputs and outputs

• Accept one bounded target, exact file/source references, rubric, and stop condition.
• Return PASS or sorted Critical/High/Medium findings with file:line, impact, evidence, and minimal fix.

Write scope

• Keep the reviewer read-only.
• Never edit files, acquire writer leases, promote, create Git refs, or perform external actions.
• Let the main agent resolve contradictions and own all writes.

Preflight

1. Run uv run raytsystem agent preflight --skill raytsystem-run-review --write --json locally.
2. Run agent subagent-check before sending any excerpt to a reviewer surface.
3. Supply only the minimal non-sensitive target; do not leak the intended answer or suspected fix.

Workflow

1. Inspect target contracts, implementation, tests, and declared gate evidence.
2. Reproduce suspected failures read-only when safe.
3. Rank findings by concrete impact; omit style-only commentary unless requested.
4. Return summaries with source references, not raw logs.

Validation

• Verify source-of-truth boundaries, generation binding, idempotency, recovery, skipped gates, and docs/code agreement.
• Require evidence for every finding and distinguish untested risk from confirmed failure.
• Exercise evals m3-review-golden and m3-review-adversarial.

Recovery

• If quota/tool access ends, return reviewed scope, unresolved files/questions, and exact next read-only check.
• Do not call partial review success.

Stop and approval conditions

• Stop before writes, private hosted transfer, secrets, promotion, external mutation, or scope expansion.
• Report unavailable when an independent surface cannot safely receive inputs; let the main agent continue sequentially.