| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Cross-cutting rule covering subagent hygiene:
- Decision table: parallel-safe (read-only investigation, independent
test failures, multi-domain research) vs sequential-with-review
(plan execution, coupled edits) vs never-parallel (concurrent writes)
vs don't-subagent-at-all (target known, work fits in ~10 tool calls).
- Prompt contract: scope / pasted context / explicit 'do NOT' list /
required output format. Missing any field yields shallow work.
- Context-pollution rule: subagents absorb noise the main thread
shouldn't carry. When one fails, dispatch a fix-agent with the
failure report — do not retry in the orchestrator.
- Review-gate cadence: after each sequential task, or every ~3 in
parallel batches.
- Anti-patterns including parallel implementation on overlapping
files, broad 'fix all the tests' prompts, timeout-tuning to mask
flakes, and letting the agent decide scope.
Added one-line cross-references to subagents.md from debug/,
review-code/, and finish-branch/ SKILL.md — the skills that most
rely on delegation.
Clean-room synthesis from NeoLabHQ/context-engineering-kit's
subagent-driven-development pattern (MIT).
|
|
|
improvements
Renamed review-pr → review-code (the skill accepts PR, SHA range,
current branch, staged changes — "pr" was understating scope).
Rewrote SKILL.md with YAML frontmatter (previously header-style) and
merged useful patterns from two sources:
From obra/superpowers skills/requesting-code-review:
- Intent-vs-delivery grading (given plan/ADR/ticket)
- Mandatory Strengths section (three minimum)
- Per-issue Critical/Important/Minor severity (per-criterion
PASS/WARN/FAIL retained; complementary axes)
- Required verdict + 1-2 sentence reasoning
- Multi-input support (PR / SHA range / current branch / --staged)
- Sub-agent dispatch recommendation for heavy reviews
- Concrete filled-in example output
From the claude-plugins-official code-review plugin:
- Phase 0 eligibility gate (skip closed/draft/auto/trivial/already-reviewed)
- CLAUDE.md traversal + adherence criterion (reads root + per-directory
CLAUDE.md files; audits the diff against stated rules)
- Multi-perspective Phase 2: five passes (CLAUDE.md adherence, shallow
bug scan, git history context, prior PR comments, in-scope code
comments). For large reviews, dispatch as parallel sub-agents.
- Confidence filter (High/Medium/Low; drop Low before reporting)
- False-positive categories explicitly enumerated (pre-existing issues
on unmodified lines, lint/typecheck issues CI handles,
senior-wouldn't-call-out nitpicks, silenced issues with valid reason,
intentional scope changes, unmodified-line issues, framework-behavior
tests)
- Trust-CI discipline (don't run builds yourself)
Substance from the original review-pr kept verbatim:
- DeepSat-specific criteria (security, TDD evidence, conventions,
no-AI-attribution, API contracts, architecture layering, root-cause
discipline)
Size: 60 lines → 347 lines. Growth is structural (added phases, added
example, added perspectives, added filters) not verbose — each section
earns its lines.
NOT adopted from the plugin:
- GitHub comment output format (plugin posts PR comments; review-code
outputs a markdown report the user can paste if they want)
- "Generated with Claude Code" footer (violates no-AI-attribution rule)
- Specific 0/25/50/75/100 confidence scale (Critical/Important/Minor
covers the same signal with less ceremony)
Makefile SKILLS updated: review-pr → review-code. Old
~/.claude/skills/review-pr symlink removed; make install creates the
new one at ~/.claude/skills/review-code.
|
