Pattern Review

Skill name: pattern-review

Run the pattern-review gate for agentic development work. Use before human review or at PR boundaries to review a repo branch diff against active .pattern-review rules, write a persisted pattern-review.md report, resolve findings, waivers, and manual review items, and rerun until the gate passes or is explicitly waived.

Use this as a gate, not a checklist. Review the current Git repository's branch diff against active pattern rules.

Deterministic helper:

python3 /Users/trent.brown/agentic-development-workflow/scripts/pattern_tool.py review-inputs --cwd <repo>

Workflow:

• Run the helper to collect repo context, base/head, changed files, rule sources, duplicate-ID overrides, and deterministic trigger results.
• If no applicable .pattern-review directory is found after the normal scope walk, fail and tell the caller to run /pattern-init <intended-scope>.
• Display the .pattern-review directories found in the rule stack before evaluating rules so the caller can spot an unexpected scope.
• If no active rules are found in an existing scope, write PASS_WITH_WARNINGS and state that the gate is not meaningfully configured; recommend pattern-extract.
• Evaluate rules whose trigger result is true.
• For agentic triggers, decide whether the rule triggers and record a brief rationale whether it does or does not.
• For triggered agentic reviews, inspect the changed files and return one of pass, fail, warning, not_applicable, or needs_manual_review. Every non-pass result needs file/line evidence or a note explaining why evidence is not file-specific.
• For triggered manual reviews, block until a human disposition is recorded for the current merge-base/head context.
• Treat blocker failures and unresolved manual items as FAIL. Treat warnings as PASS_WITH_WARNINGS. Explicit persisted waivers produce PASS_WITH_WAIVERS, which is an acceptable passing verdict.
• Write the final human-readable report before declaring the gate passed. For normal workflow branches, use docs/issues/<branch>/pattern-review.md. Include review context, rule stack, triggered rules, findings, waivers, manual review dispositions, and final verdict. Reference rules by ID, title, source path, and hash; do not dump full rule text by default.
• If the gate fails, fix the implementation or record a manual review/waiver, then rerun. Do not emit learning events for ordinary findings; emit them only when the gate exposes a rule-system weakness.