review-crucible

Review Crucible

Review Crucible is a source-neutral review workflow for Claude Code, Codex, and Amp. Its behavior is defined by the canonical references in references/. Load only the references needed for the task.

Operating Contract

• Review changed behavior, invariants, and integration contracts.
• Apply local project instructions before generic guidance.
• Treat false positives as review debt; validate before reporting.
• Separate correctness/safety review from simplification.
• Prefer review quality over latency and cost. Follow references/model-policy.md so review passes use the strongest verified host model and the highest supported effort.
• Use current documentation when a finding depends on external library, framework, SDK, API, CLI, build-tool, or cloud-service behavior.
• Preserve author intent unless it conflicts with security, data integrity, or explicit project rules.

Required References

• references/review-orchestration.md: scope discovery, pass selection, targeted activation, host routing, and validation ownership.
• references/reviewer-roles.md: canonical behavior for each reviewer role.
• references/review-rubric.md: severity, confidence, false-positive filters, validation gate, and output format.
• references/agent-roster.md: role names and host check/agent mapping.
• references/model-policy.md: strongest-model and highest-effort defaults.
• references/compatibility.md: Claude Code, Codex, and Amp limits.
• references/github-commenting.md: inline PR comment discipline.
• references/rust-best-practices-integration.md: Rust handoff rules.

Workflow

1. Establish scope: PR, branch diff, commit range, named files, or recent git changes.
2. If PR metadata exists, run the preflight gate from reviewer-roles.md.
3. Read applicable AGENTS.md, CLAUDE.md, contribution, testing, release, and language/framework instruction files.
4. Summarize intended behavior and classify changed files by risk area.
5. Select passes using review-orchestration.md.
6. Run focused reviewers in parallel when the active host supports and permits subagents; otherwise run the same passes sequentially.
7. Validate candidates through finding-validator or parent/orchestrator logic when the host cannot run a validator subagent.
8. Report only through review-rubric.md.
9. Post GitHub comments only when requested, following github-commenting.md.
10. Run simplification only after correctness/safety review, following reviewer-roles.md#code-simplifier.

Host Notes

• Claude Code: plugin commands and agents/*.md can launch focused reviewers. The generated agents preload this skill so references are available.
• Codex: install this skill and .codex/agents/*.toml custom agents for reviewer fan-out. Codex only spawns subagents when explicitly asked.
• Amp: install .agents/checks/*.md and the Amp AGENTS.md snippet. Amp checks run as separate review subagents, but the parent thread owns final deduplication, validation, fixes, and comment posting.

If a host cannot access a reference at runtime, say which reference was unavailable and continue only within the documented host limit.