Skip to content

v0.3.0 - review intelligence + per-project settings

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 16 Mar 17:06
v0.3.0
6a072d9

Added

  • Per-project settings file (.claude/pr-review.local.md): configurable output language, default review source, default repo path, review dimensions per mode, fix-forward exclusion patterns, and developer profile tracking toggle. CLI flags override settings; settings override defaults.
  • gh pr diff fallback: when the target repo is not cloned locally, the skill now uses gh pr diff <PR#> --repo <owner/repo> to fetch actual file-level changes via GitHub API instead of silently relying on provider summaries.
  • Severity count validation: the skill now cross-checks the severity summary line against individual finding detail labels before finalizing, preventing mismatches (e.g., claiming MEDIUM 3 when one is actually LOW).
  • Fix-forward exclusion patterns: never auto-modify migration files, Dockerfiles, CI workflows, lockfiles, or environment files. These get "Manual fix recommended" instead.
  • Fix-forward commit hygiene: reads git log --oneline -10 to match repo commit convention, uses specific-file git add only (never broad staging), and never includes Co-Authored-By or similar attribution trailers.
  • Developer profile dedup: checks for existing ticket+date entries before appending to prevent duplicates across multiple review iterations.
  • Output template reference file (references/output-template.md): extracted from SKILL.md for progressive disclosure (SKILL.md reduced from 700 to 572 lines).

Changed

  • Quick mode depth reduced: now reviews top 5 dimensions (Bugs, Architecture, React/TypeScript, Error Handling, Performance) instead of all 9, producing meaningfully shorter reviews. Quick mode saves to [TICKET]-review-quick.md to avoid overwriting full reviews.
  • Independent review execution order enforced: Step 4 (independent code review) must complete entirely before reading any AI provider comments in Step 5. This prevents confirmation bias from Copilot/CodeRabbit data contaminating "independent" findings.
  • Step 1 clarified: now notes which AI providers have data available without reading their actual comments. Inline summaries visible in the PR doc must not influence Step 4.
  • Dirty detection relaxed: only staged changes (git diff --cached) trigger the dirty-repo block. Unstaged modifications (e.g., local .gitignore tweaks) are no longer considered dirty.
  • Repo access resolution chain: --repo-path → settings default_repo_path → CWD → gh pr diff fallback → graceful degradation with explicit disclosure.

Removed

  • Difficulty scale: removed the 1-5 difficulty rating and associated criteria table. It was a vestige from the CodeRabbit-only v0.1.0 era with no clear criteria for multi-provider reviews.

Fixed

  • Fix-forward no longer auto-modifies immutable framework files (Supabase migrations, Prisma migrations, etc.).
  • Fix-forward no longer uses broad git add commands that could stage unrelated user changes.
  • Fix-forward no longer includes Co-Authored-By attribution in generated commits.
  • Fix-forward now matches the target repo's commit message convention.
  • Quick mode no longer overwrites full review files (uses -quick suffix).

Release Boundary

  • 0.3.0 completes the MVP improvement plan (Phases 0-2): baseline measurement, bug fixes, intelligence improvements, and per-project settings.
  • Dedicated agent, PreToolUse hook, ${CLAUDE_PLUGIN_ROOT} paths, and full plugin restructure are deferred to 0.4.0 (Post-MVP Phase 3).
Assets 2
Loading