code-review-gate

Every code change outside a pipeline must clear this gate before it counts as done. Gates must be green before the Reviewer runs, and a CRITICAL security finding always blocks.

Contract

• Input: a working directory with changed files.
• Output: a gate report (PASS/FAIL per gate), then a Reviewer score (1–10) with structured findings, then a verdict line.
• Tool boundary: this skill reports and routes; it never auto-fixes failing gates or findings. The Reviewer sees only the changed files, never the full codebase.
• Done when: Phase 3 prints a PASS verdict line, or a BLOCK with the specific finding to fix.

Steps

1. Phase 1 — Quality Gates. Invoke the /quality-gate skill. If any gate fails, show the failing gate and its error output, stop, surface the fix, and wait for a re-run. Phase 2 begins after every gate is green.
2. Phase 2 — Reviewer. Collect the changed files with the commands in references/diff-verdict-and-rules.md, then dispatch the Reviewer subagent — coding-pipeline:reviewer (or reviewer in a flat ~/.claude/agents install) — handing it the full content of each changed file, the one-line gate summary, and the original request verbatim (the Reviewer needs it to judge change discipline). The dispatched subagent already carries its Reviewer persona; do not read the agent file into the main context. It covers the Security Deep-Dive checklist, language-specific checks, the test-falsifiability check, and the change-discipline check (CD1–CD7) defined in that reference.
3. Phase 3 — Verdict. Apply the verdict table in references/diff-verdict-and-rules.md: score ≥ 8.0 with no CRITICAL passes; any CRITICAL blocks; below 8.0 surfaces findings for a decision. After any fix, restart from Phase 1.

References

• references/diff-verdict-and-rules.md — diff collection commands, Reviewer payload, falsifiability check, verdict table, and anti-patterns.
• coding-pipeline/references/change-discipline.md — CD1–CD7 rows, severities, and worked ❌/✅ examples (single source of truth).