parallel-review

Runtime required. This skill uses Canopy tree notation; canopy-runtime must be active.

Detect canopy-runtime — present if either:

• canopy-runtime/SKILL.md exists under .claude/skills/, .github/skills/, or .agents/skills/, OR
• a canopy-runtime marker block exists in CLAUDE.md or .github/copilot-instructions.md.

If neither is present — install canopy-runtime first (see the compatibility field for the source and install options), then re-invoke this skill.

Do not interpret the ## Tree without canopy-runtime active.

Parallel-review target: $ARGUMENTS

Tree

parallel-review
├── **EXPLORE_TARGET** << $ARGUMENTS >> context
├── SHOW_PLAN >> target | file_count | aspects
├── ASK << Proceed with parallel review? | Yes | No
├── IF << user chose No
│   └── END Cancelled.
├── Read `assets/constants/review-aspects.md` for per-aspect criteria
├── PARALLEL
│   ├── **REVIEW_ASPECT** << "security"    | context.files >> security_findings
│   ├── **REVIEW_ASPECT** << "performance" | context.files >> perf_findings
│   ├── **REVIEW_ASPECT** << "style"       | context.files >> style_findings
│   └── **REVIEW_ASPECT** << "correctness" | context.files >> correctness_findings
├── MERGE_ASPECT_FINDINGS << security_findings | perf_findings | style_findings | correctness_findings >> all_findings
├── REPORT_BY_SEVERITY << all_findings
└── VERIFY_EXPECTED << assets/verify/verify-expected.md

Rules

• Each bold-marked op call (**EXPLORE_TARGET**, **REVIEW_ASPECT**) dispatches as a subagent — its own context window, no access to parent state outside its << inputs
• PARALLEL children that are bold-marked op calls fan out as parallel subagents in a single agent turn
• Bindings (>>) in PARALLEL children are merged in declaration order regardless of completion order — MERGE_ASPECT_FINDINGS relies on stable names, not arrival timing
• A single subagent failure does not abort siblings (Promise.allSettled) — MERGE_ASPECT_FINDINGS handles missing/empty aspects gracefully
• Suggested fixes must be concrete: show the corrected code snippet, not just a description
• Only report what is observable in the provided files — do not speculate

Response: target | aspects_reviewed | critical | warnings | info | overall_verdict