| name | csw-review |
| description | Use to run a rigorous multi-lane review before declaring work done — for any broad, risky, shared, security-sensitive, or release-facing change. Uses native scheduling for independent review lanes and gates on ALL of them passing with a single binding verdict. |
csw-review — all-or-nothing review gate
Run independent review lanes in parallel and accept the work only if EVERY lane
passes. One failing lane = overall REJECTION.
Lanes (independent; parallel when useful)
Use native task subagents to run these as independent workers, routing each to the
verifier agent with a lane-specific brief. Use /fleet for user-visible parallel
execution and /tasks to inspect or cancel work. Enforce non-mutation with host
deny/available-tool policy before launch; a review prompt is not a security boundary.
- F1 Compliance — every planned task done? every acceptance criterion met?
plan/output format honored?
- F2 Code quality — diagnostics clean; idioms match the repo; no dead code; no
avoidable risk; no AI-slop.
- F3 Real manual QA — every QA scenario actually executed through its channel;
evidence artifacts present; cleanup receipts recorded.
- F4 Scope fidelity — nothing shipped beyond Must-have; nothing Must-NOT
introduced.
Each lane's worker prompt must be self-contained: include the goal, the diff (or
how to obtain it), the success criteria, and the evidence locations.
Distrust and re-verify
Worker verdicts are claims. Before accepting an APPROVE, spot-check it yourself:
re-read the diff, re-run the cited tests, open the cited evidence artifacts.
Replay command-backed criteria with csw-runtime verify --id <C0NN> -- <argv...>;
bind file-backed real-surface evidence with csw-runtime artifact --id <C0NN> --path <workspace-file> --summary "<observed outcome>". Free-text evidence cannot
pass a criterion.
All-or-nothing gate
Combine the lanes:
- If any lane returns REJECTION (or hedges — "looks good but…", "approve if…"),
the overall verdict is REJECTION. List every blocking finding and loop:
fix → re-run the affected lane(s) → re-gate.
- Only when all four lanes return clean APPROVE is the verdict
UNCONDITIONAL APPROVAL.
Binding outcome
Record the result. If a goal runtime is active, completion still requires
csw-runtime complete to pass (criteria + machine receipts + zero blockers) — this review
is the human-facing quality gate layered on top of the oracle, not a replacement
for it. Unresolved review findings should be filed as blockers
(csw-runtime blocker add) so they block completion until addressed.
