Code Quality Gate

Overview

Wraps cqa-analyzer — an offline static analyzer where no source leaves the machine — and converts its findings into one decision: did this diff introduce new issues on the lines it touched? Built for the pre-PR moment and for CI, where you need a deterministic yes/no plus a short, actionable finding list, not a wall of JSON.

The analyzer deliberately never invokes git, so it cannot know what changed. This skill supplies the missing piece: scripts/changed_lines_manifest.py turns a git diff into the changed-lines manifest the analyzer accepts, and scripts/run-gate.sh drives the whole flow with the correct flags.

Usage

Use this skill when:

• Preparing a PR and you want a quality read on the diff before reviewers see it.
• Wiring a pre-commit hook or a CI step that must fail the build only on new issues.
• The analyzer emitted JSON/SARIF and you need it triaged into "must-fix vs. accept".
• The user mentions code quality, cyclomatic complexity, changed-line gating, baselines, or SARIF.

Do not use it for runtime profiling or security-vulnerability scanning — it is a static quality gate, not a SAST or APM tool.

Core Concepts

• Changed-line gate. Findings are intersected with the lines the diff touched. A pre-existing issue on an untouched line does not fail the gate; a new one does. This is what makes it safe to drop into CI on a legacy repository.
• Baseline + new-findings-only. On a legacy repo, record today's findings once (--write-baseline), then gate only findings absent from that baseline. The analyzer hashes fingerprints, so the baseline contains no source text.
• Offline and deterministic. Same input → same output, no network. The wrapper always passes --offline, which makes the analyzer deny socket operations while it runs. A green local run predicts a green CI run.
• Two machine formats. json for caches and dashboards, sarif (2.1.0) for code-host annotations on the changed lines. text for humans.
• Configuration is pinned outside the tree being gated. A pull request could otherwise edit .code-quality.toml to disable a rule or exclude a path and pass the gate. --config <file> (use a file the PR cannot touch), --no-project-config (defaults only), or --expect-config-fingerprint <sha256> (analyzer exits 6 before scanning if the effective config differs) close that hole. Read the fingerprint once from configuration_fingerprint in a trusted JSON run and commit it to CI config.
• Analysis health. Every run reports what it could and could not analyze (analysis_health in JSON). Zero findings with parse failures means the scan was blind, not clean — --strict turns that into a non-zero exit.

Quick Reference

pipx install cqa-analyzer            # or: pip install cqa-analyzer

# One-time on a legacy repo: record the accepted floor (no gate)
scripts/run-gate.sh --write-baseline .code-quality-baseline.json

# Gate a branch: only NEW warning+ findings on CHANGED lines can fail (exit 4)
scripts/run-gate.sh --base origin/main --baseline .code-quality-baseline.json --format sarif > results.sarif

# Pre-commit: gate the lines you are about to commit
scripts/run-gate.sh --staged --baseline .code-quality-baseline.json

# Clean repo (no legacy debt): gate every changed-line finding, be strict about scan health
scripts/run-gate.sh --base origin/main --strict

# CI-hardened: the PR cannot weaken the gate by editing the repo's own config
scripts/run-gate.sh --base origin/main --baseline .code-quality-baseline.json \
  --expect-config-fingerprint "$CQA_CONFIG_FINGERPRINT" --strict --format sarif > results.sarif

Exit codes: 0 clean · 4 gate failed · 6 config fingerprint mismatch (both the analyzer's own) · 2 usage error · 3 environment error (missing analyzer, missing baseline) from the wrapper. --report-only remaps only 4; a 6 still fails.

Pre-PR Gate Checklist

Gate Progress:
- [ ] cqa-analyzer installed (or CQA_CMD set to an absolute path)
- [ ] Baseline exists (legacy repo) — or repo is clean and no baseline is needed
- [ ] In CI: config pinned (--expect-config-fingerprint or --config outside the tree)
- [ ] Ran: scripts/run-gate.sh --base origin/main [--baseline ...]
- [ ] analysis_health shows 0 parse failures (else the result is not trustworthy)
- [ ] Reviewed must-fix findings on changed lines
- [ ] SARIF uploaded if annotating the PR

Integration Patterns

• Pre-commit: a .pre-commit-hooks.yaml entry calling scripts/run-gate.sh --staged --baseline .code-quality-baseline.json blocks the commit before it reaches a PR.
• GitHub Actions: run with --format sarif > results.sarif and upload with github/codeql-action/upload-sarif so findings annotate the changed lines. Fetch enough history for the merge-base (fetch-depth: 0) and pin the config with --expect-config-fingerprint from a repository variable, so a PR cannot turn the gate off by editing .code-quality.toml.
• Agent-driven PR prep: the agent runs the gate with --format json, reads the selected findings, and reports the must-fix list to the user before opening the PR.

See references/output-contracts.md for the JSON/SARIF field shapes and references/changed-lines-manifest.md for the manifest schema the generator emits.

Common Mistakes

• With --report-only you MUST lead with the warning the wrapper prints: REPORT ONLY: this run exits 0 and will not block the build even when findings exist. Do not bury it in exit-code details.
• You MUST NOT gate a legacy repo without a baseline, because the first CI run drowns in pre-existing findings and the team disables the gate. Write a baseline once, then gate with --baseline.
• You MUST check analysis_health before trusting an empty result; zero findings with N parse failures means the scan was blind, not clean.
• You MUST NOT feed JSON to a code host expecting inline annotations; hosts annotate from SARIF. JSON is for caches and dashboards.
• You MUST NOT run an unpinned gate in CI on a repository where pull requests can edit .code-quality.toml; the gate is then advisory, not a gate.
• You MUST NOT claim a scan ran if the analyzer was not found. The wrapper exits 3 with an install hint; relay that, never fabricate findings.