| name | review-work |
| description | Post-implementation review orchestrator. Launches parallel review sub-agents covering goal/constraint verification, code quality, security, hands-on QA, and context mining. ALL must pass for review to pass. MUST USE after completing significant implementation work. |
review-work
Five-agent parallel review orchestrator. Triggers: "review work", "review my work", "review changes", "QA my work", "verify implementation", "check my work", "validate changes", "post-implementation review".
When to use
After ANY significant implementation work — feature, bug fix, refactor — before declaring it done.
Workflow
Fire these five agents in parallel via task (independent, all read-only):
task(agent: "oracle", tasks: [
{ id: "rv-goal", description: "Goal & constraint verification", assignment: "<assignment 1 below>" },
{ id: "rv-quality", description: "Code quality review", assignment: "<assignment 2 below>" },
{ id: "rv-security", description: "Security review", assignment: "<assignment 3 below>" }
])
task(agent: "task", tasks: [
{ id: "rv-qa", description: "Hands-on QA execution", assignment: "<assignment 4 below>" },
{ id: "rv-context", description: "Context mining (git/PRs/issues)", assignment: "<assignment 5 below>" }
])
Assignments (templates)
1. Goal & constraint verification (oracle)
Map the user's original request and explicit constraints to actual code changes. For each sub-requirement:
- [ACHIEVED / MISSED / PARTIAL] Requirement description
- Evidence: specific code reference or gap
For each explicit constraint (e.g. "no new dependencies", "preserve API"):
- [PASS / FAIL] Constraint
- Evidence
2. Code quality review (oracle)
Review the diff for:
- Naming clarity
- Function size & cohesion
- Error handling completeness
- Edge cases (null, empty, large, concurrent)
- Test coverage of new behavior
- Adherence to existing codebase patterns
- AI slop (see
ai-slop-remover skill for patterns)
Flag findings as critical / high / medium / low.
3. Security review (oracle)
Review the diff for:
- Input validation at trust boundaries
- Authn / authz changes
- Secret handling
- Injection (SQL, command, prompt, path)
- SSRF / open-redirect / CSRF
- Information disclosure in errors / logs
- Dependency security (new deps only)
Flag findings as
critical / high / medium / low.
4. Hands-on QA execution (task)
Run the actual feature end-to-end. For CLI: invoke it. For API: hit the endpoints. For UI: drive the browser. Show concrete output. Record:
- Happy path: works / doesn't
- Error path: works / doesn't
- Edge cases tested
- Diagnostics / build / tests: all green
5. Context mining (task)
Look at recent project context that might inform the review:
- Recent git commits — what conventions were established?
- Open / recently-closed issues — known gotchas in this area?
- Open PRs — any conflicting work in flight?
- Documentation / READMEs — anything the change should update?
Pass criteria
ALL FIVE must pass:
- Goal: every requirement ACHIEVED, every constraint PASS
- Quality: no
critical or high findings
- Security: no
critical or high findings
- QA: happy path + error path + all relevant edge cases verified
- Context: change is consistent with recent project trajectory
If ANY fails → fix → re-run THE SAME review session for the failing agent, not a fresh one.
iter-1 stub. Iter-2 will expand each assignment template with concrete checklists.