drewburchfield/project-bootstrap/quality-gate

Quality Gate

Structured review and remediation before merge. Review methodology is host-agnostic (Review Suite). Execution uses the current host’s adapter.

Deep docs (load on demand):

• references/review-suite/SPEC.md — tiers, pass artifacts, capability probe
• references/review-suite/passes/*.md — seven lenses
• references/review-suite/host-adapters/{claude-code,codex,grok,opencode,agy}.md
• references/review-suite/synthesis.md
• references/linear-integration.md, references/github-cli.md (issue/PR ops)

Modes

Prefer ship-loop when the user wants CI watch + greenlit merge ownership.

0. Capability probe (every run)

Probe before review. Never invent tools.

host:          claude-code | codex | grok | opencode | agy | other
git/gh:        available?
issue_tracker: from .claude/project-meta.json (linear|jira|github|standalone|none)
agent_suite:   host can run isolated multi-lens review (Claude: pr-review-toolkit; else subagents)
external_bots: from project-meta.review.externalBots or profile; default empty
clawpatch:     command -v clawpatch (Codex path)
require_suite: project-meta.review.requireAgentSuite (default false)

project-meta optional review block:

{
  "review": {
    "externalBots": ["devin", "coderabbit"],
    "botWaitSeconds": 60,
    "requireExternalBots": false,
    "requireAgentSuite": false
  }
}

Forbidden: unconditional sleep 60 (or any fixed wait) for Devin/CodeRabbit when external_bots is empty or bots are not configured.

If bots are configured: after push, wait up to botWaitSeconds (default 60), then fetch PR comments; if none arrive, proceed and note bots: timeout/none.

1. Prerequisites and issue detection

• Not on main/master for full PR modes (local may run on any branch with a diff).
• Read project-meta.json issueTracker when present; else standalone.
• Detect issue from branch / commits / tracker CLI when configured; --no-issue skips.
• Uncommitted work: do not silently mix into shipping; ask or commit if clearly part of the work.

Details: references/linear-integration.md, references/github-cli.md.

2. Review Suite (core)

Always load and follow references/review-suite/SPEC.md.

Host adapter selection

On Claude: prefer parallel toolkit agents mapped to suite pass ids (see claude-code adapter). If toolkit missing, fall back to T2 sequential using passes/*.md.

On non-Claude hosts: do not require /review-pr or pr-review-toolkit:* agent IDs. Run pass files sequentially (or host subagents) and emit artifacts.

Pass artifacts (required for Pass)

For each mandatory pass (code-correctness, silent-failures, tests, types if typed, comments, simplify, spec-conformance if a spec source exists), emit the artifact shape from SPEC (checked list, findings or none_found: true with non-empty checked).

Tier declaration (required every run)

Emit tier_declaration from SPEC: host, adapter, tier (T3|T2|T1|T0), execution_mode, passes_completed, passes_skipped, capabilities, gate_status (Pass|Blocked|Incomplete).

Synthesize with references/review-suite/synthesis.md.

Honesty

• Do not claim Review Suite / quality-gate ran unless T2/T3 with artifacts (or toolkit-mapped equivalent).
• Manual diff read alone = label as manual/T1, not suite Pass.
• If requireAgentSuite: true and suite cannot run → Blocked with config hint.

3. Workflow by mode

Local (--local)

1. Probe + collect local or unpushed diff.
2. Run Review Suite (adapter).
3. Remediate Critical/High (local commits only if user wants fixes).
4. Re-run suite on updated diff if needed.
5. Stop. No PR, no push, no merge.

Full PR (default / issue / --no-issue)

1. Preflight Review Suite on branch diff (before or as first pass).
2. Create or reuse PR (gh); see github-cli reference.
3. Run Review Suite on PR diff (adapter).
4. Remediate Critical/High; push intentional commits.
5. External bots only if configured (probe); then re-fetch comments and fold findings.
6. Re-run suite or targeted re-check on changed risk.
7. Verification: project-meta quality commands or inferred tests/lint/typecheck.
8. Present merge options; do not merge without user confirmation (unless ship-loop owns merge with greenlit).
9. On merge approve: squash default, update issue tracker if configured.

When ship-loop is driving: run this skill’s review + remediation steps only; hand merge/watch/deploy back to ship-loop.

4. Remediation

• Critical/High first; Medium deferred only with reason.
• Scoped fixes; clear commit messages.
• Draft-only for risky changes unless user approved auto-remediation.
• Deterministic formatter/linter fixes may apply when safe.

5. Output format

Quality gate
Mode: local | full | standalone
Host: <host> · Adapter: <adapter> · Tier: T2|T3|T1|T0
 
Capabilities
- agent_suite: yes|no
- external_bots: none | [devin,…] (waited|skipped|timeout)
- issue_tracker: linear|none|…
 
Gate: Pass | Blocked | Incomplete
 
Findings
- [High] title (path:line): evidence → fix
 
Passes
- code-correctness: N findings | none found (checked: …)
- …
 
Verification
- command: result
 
Next
- <one concrete step>

Include machine-readable tier_declaration + pass artifacts when writing to a file or when the host benefits from structure (YAML/JSON block is fine).

6. Hard rules

1. Follow Review Suite SPEC for Pass/Blocked/Incomplete.
2. No unconditional external-bot sleep.
3. No Claude-only tool IDs as required steps on non-Claude hosts.
4. No auto-merge without explicit user approval in this skill (ship-loop may greenlit separately).
5. No false “all agents ran” without actual suite execution.
6. Prefer ship-loop for babysit-until-shipped.

7. Drew / maximal pipeline (optional)

For full personal pipeline, set in project-meta:

{
  "review": {
    "externalBots": ["devin", "coderabbit"],
    "botWaitSeconds": 60,
    "requireAgentSuite": true
  }
}

Plugin defaults stay permissive so other harnesses complete T2 without bots.