| name | review-and-fix |
| description | Run adversarial swarm review then auto-fix all findings in parallel. Use when the user says "review and fix", "swarm fix", or wants review and fix in one shot. |
| allowed-tools | Task, TaskOutput, Bash, Read, Write, Edit, Glob, Grep, TeamCreate, TeamDelete, SendMessage, TaskCreate, TaskUpdate, TaskList, TaskGet, Agent, Skill |
| skills | ["dancing-bear-rules"] |
Review and Fix
Delegates to workflows/code/review-and-fix.yaml. Phase 1 runs the adversarial swarm review
(diff fetch, concern sweep, per-file validation, consolidation, fact-check, human gate,
inline comment posting). Phase 2 triages findings by fix type, fans out code, test, and
qlty fixes in parallel, verifies everything is green, resolves posted threads, and
presents a summary for final approval.
When to Use
- User says "review and fix", "swarm fix", or "review then auto-fix"
- One-shot review + remediation before merge
- Existing review findings need automated fixing (set
skip_review: "true")
Derive Params from Context
GITHUB_TOKEN= gh pr view --json number -q .number
If the command returns a number, use it. If the branch has no open PR, ask the user.
Invocation
IMPORTANT: Use the /workflow skill — do NOT call ./bin/workflow run --execute directly. It only writes dispatch files and exits (status=pending). The /workflow skill is what actually spawns agents, waits for results, and handles human gates.
Skill(skill="workflow", args="--workflow workflows/code/review-and-fix.yaml --params pr_number=314 test_cmd='python3 -m unittest tests/grafana/ -f -q' source_root=grafana/ test_path=tests/grafana/")
Skip Phase 1 if review already ran and findings are in /tmp/review-and-fix-{pr_number}/outputs/:
Skill(skill="workflow", args="--workflow workflows/code/review-and-fix.yaml --params pr_number=314 skip_review=true")
Params
| Param | Default | Description |
|---|
pr_number | "" | PR number (required; auto-detected from branch if blank) |
test_cmd | "" | e.g. python3 -m unittest tests/grafana/ -f -q |
source_root | "" | e.g. grafana/ |
test_path | "" | e.g. tests/grafana/ |
min_coverage | "80" | Coverage threshold (percent, no % sign) |
auth_domains | "github,qlty" | Auth domains to pre-check |
skip_review | "false" | Set "true" to jump straight to Phase 2 |
Workflow Stages
Phase 1 — Review
- init — create workspace directories (
outputs/, outputs/diffs/, outputs/findings/, validation/)
- fetch-pr-context — PR metadata, per-file diffs, commit history, PR description
- concern-sweep — load domain guides, filter to concerns triggered by this diff
- enumerate-targets — cross-reference concerns with per-file diffs; produce manifest
- validate-concerns (parallel fan-out) — one agent per manifest entry; write per-finding JSON
- cross-unit-check (parallel with validate-concerns) — cross-file consistency check
- consolidate — merge, de-duplicate, and sort findings; write
consolidated.json + summary
- fact-check-findings — verify line numbers, evidence quotes, and severity against guide files
- human-gate — present findings; user approves, edits, or discards
- post-comments — post approved findings as inline GitHub PR review comments
Phase 2 — Fix
- triage-findings — group findings into code/test/docs/qlty buckets; fetch SQ issues; write
fix-manifest.json
- fix-code (worker_queue, parallel) — fix logic bugs and style issues in source files; run ruff after each
- fix-tests (worker_queue, parallel with fix-code) — add missing tests, strengthen weak assertions; run python3 -m unittest per fix
- fix-qlty (parallel with fix-code and fix-tests) — fix new-code SQ issues; verify gate after
- verify-fixes — run ruff, test suite, and SQ gate; confirm all green
- resolve-threads — resolve GitHub threads for findings that were fixed
- human-gate-fixes — present fix summary; user approves for merge or requests another round
