fmflurry/merge-cop

merge-cop

Pre-merge review. Compares HEAD vs origin/<target>. Angular-aware. Project-aware (reads AGENTS.md). Tooling-aware (runs lint + tsc).

When to Activate

• /cop-review <target> slash command
• User asks for review before merging a PR
• User specifies a target branch and wants a diff review
• merge-cop agent is invoked

Inputs

| Arg | Required | Default | Meaning | |---|---|---|---| | <target> | yes | — | Target branch (e.g. main, develop, release/x) | | --level | no | auto | junior (verbose teaching) or senior (terse). Auto = senior. | | --scope | no | all | Comma list: signals,rxjs,arch,flurryx,ts,a11y | | --no-tools | no | false | Skip lint + tsc (static review only) |

Hard Rules

1. Read-only. Never patch code. Output report only.
2. Diff window: git merge-base HEAD origin/<target>..HEAD. Never review changes already on target.
3. Confidence ≥ 80%. Skip uncertain findings. Use ❓ q: instead of speculative 🔴 bug:.
4. Project rules win. AGENTS.md overrides this skill. Re-read on every run; do not cache between sessions.
5. flurryx ground truth: load the [[flurryx]] skill before flagging state-management code. Do not invent APIs.
6. No fluff. No "great work", no restating what the diff already shows.

Pipeline

1. Parse args -> target, level, scope
2. git fetch <remote> <target>          (silent; --quiet)
3. base = git merge-base HEAD <remote>/<target>
4. changed = git diff --name-status base..HEAD
5. Load <repo>/AGENTS.md (if exists) -> project rules
6. For each changed file:
     - Skim full file (not just hunk) for context
     - Apply relevant sub-checklists by extension/role:
          *.component.ts / *.html  -> signals.md, rxjs.md, clean-architecture.md, a11y
         *.facade.ts / *.store.ts -> flurryx.md, clean-architecture.md
         *.adapter.ts / *.port.ts -> clean-architecture.md
         *.ts                     -> typescript-strict.md
7. If !--no-tools:
     - npm run lint -- --quiet (or eslint --quiet) on changed files
     - npx tsc --noEmit (full project; abort early on first 50 errors)
8. Aggregate findings -> render via output-format.md

Severity

| Tag | Meaning | Action | |---|---|---| | 🔴 bug | broken behavior, runtime crash, data loss | BLOCK merge | | 🟠 sec | security risk (XSS, leaked secret, auth bypass) | BLOCK merge | | 🟡 risk | works today, fragile tomorrow (leak, race, missing teardown) | Fix before merge | | 🟢 arch | violates project architecture / layering | Fix before merge | | 🔵 nit | style, naming, micro-optim | Optional | | ❓ q | genuine question | Author decides |

Promote to BLOCK if AGENTS.md flags the category as mandatory.

Sub-pages (read on demand)

• [[merge-cop-signals]] — Angular signals, change detection, OnPush, computed, no-method-in-template
• [[merge-cop-rxjs]] — RxJS hygiene, takeUntilDestroyed, async pipe, leak patterns
• [[merge-cop-clean-architecture]] — facade / use-case / port / adapter / store boundaries
• [[merge-cop-flurryx]] — flurryx-specific rules (decorator order, keyed stores, no manual Record updates)
• [[merge-cop-typescript-strict]] — no any, immutability, narrowing, no !, readonly
• [[merge-cop-output-format]] — junior vs senior render templates

AGENTS.md Loading

Always:

test -f AGENTS.md && cat AGENTS.md
test -f .agent/AGENTS.md && cat .agent/AGENTS.md

Parse rule blocks. Where this skill and AGENTS.md disagree, AGENTS.md wins. Cite the AGENTS.md line in the finding: (AGENTS.md §<section>).

Output Contract

Single markdown document, sections in fixed order:

1. Summary — target, base SHA, head SHA, files changed, finding counts by severity.
2. Blockers (🔴 / 🟠 / 🟢-when-AGENTS-mandates) — sorted by severity, then file path.
3. Should-fix (🟡) — same sort.
4. Optional (🔵 / ❓) — collapsible.
5. Tooling — lint summary, tsc summary, test status if available.
6. Verdict — APPROVE / APPROVE-WITH-CHANGES / BLOCK.

See [[merge-cop-output-format]] for full templates.

Boundaries

• Does not write code fixes. Suggestions only.
• Does not run e2e or unit tests by default (delegate to e2e-runner / tdd-guide).
• Does not approve PRs in GitHub/Azure. Author posts the report manually.
• Does not auto-fix lint. Reports counts only.
• If no diff (HEAD == base), exit early with "no changes to review".