Csw Review

csw-review — all-or-nothing review gate

Use this skill before declaring risky or broad work complete. Review lanes may run independently through native scheduling, but the conductor owns the final verdict: worker conclusions are claims until their diffs, commands, and artifacts are re-inspected.

Read the focused references:

• Compliance lanes — goal, plan, code quality, manual QA, compatibility, and operational checks.
• Evidence review — receipt freshness, artifact identity, command provenance, real-surface interpretation, and cleanup.
• Security and scope — authority, trust, data, prompt injection, worktree isolation, and must-NOT boundaries.
• Verdict template — finding format, severity, re-review loop, and UNCONDITIONAL APPROVAL gate.
• Draft-plan and completed-work review — mode gates, atomic task checks, plan verdicts, and completed-delivery review.
• Review worker packets — self-contained read-only lane packets, native /tasks interpretation, and conductor synthesis.
• Manual QA and context mining — diff reconstruction plus CLI, HTTP, web, TUI, package, and artifact scenarios.

Review preconditions

Do not begin a completion review without:

1. user-visible goal and current approved scope;
2. success criteria and their revisions;
3. complete diff including non-ignored untracked paths;
4. test, build, package, or install commands from owned sources;
5. real manual QA artifact index;
6. known limitations, blockers, and cleanup receipt;
7. plan or decision record for non-trivial work.

If a required input is absent, return REJECTION with the missing review input.

Worker isolation and scheduling

Independent lanes may use native task workers. Use /fleet only when the user wants visible parallelism and /tasks for inspection or cancellation.

Review workers are read-only. Enforce that through the host deny/available-tool policy; “do not edit” prose is not a security boundary. Every worker packet is self-contained and names goal, scope, diff, criteria, evidence, lane questions, and required verdict.

Do not give a reviewer authority to fix findings in place. Review and remediation are separate phases.

Lane F1: compliance and plan fidelity

Check:

• every approved must-have maps to implementation and evidence;
• every must-NOT remains absent;
• ordered tasks and acceptance scenarios were completed;
• criteria match the current goal rather than an earlier revision;
• required docs, migration, compatibility, and cleanup work is present;
• no task was marked complete solely from prose or a worker claim.

Missing required work is rejection even when existing tests pass.

Lane F2: code quality and maintainability

Inspect the actual diff and surrounding code:

• behavior matches local architecture and conventions;
• tests fail for the intended pre-change reason and prove useful contracts;
• no dead branches, broad catches, duplicate logic, needless wrappers, or generated churn entered the change;
• types, error handling, resource ownership, and boundaries are explicit;
• comments explain non-obvious intent rather than narrating syntax;
• diagnostics, lint, format, build, focused tests, and full relevant tests pass;
• public changes have compatibility coverage.

Re-run material commands yourself. A pasted output excerpt is not enough.

Lane F3: real manual QA

For each success criterion, inspect the actual user/operator channel:

• CLI output and exit status;
• HTTP request, status, headers, body, and side effects;
• browser interaction, viewport/state captures, keyboard, and accessibility;
• terminal dimensions, resize, scrolling, color-off, and CJK where relevant;
• installed package or plugin discovery from the packaged candidate;
• generated artifact identity and content.

Confirm fixture, build identity, scenario steps, expected result, evidence path, and cleanup. A unit suite cannot substitute for channel evidence.

Lane F4: evidence integrity

Audit every machine receipt:

• command proof uses the exact runtime invocation injected at session start with verify --id <C0NN> -- <approved argv...>;

• real-surface files use that invocation with artifact --id <C0NN> --path <file> --summary "<result>";

• correct criterion and current revision;

• successful exit or current artifact identity;

• workspace freshness when available;

• ignored inputs bound separately;

• command argv sourced from approved repository or plan surfaces;

• no raw secret or source body persisted;

• artifact is nonempty, current, and actually supports the summarized claim;

• no stale evidence after later edits.

Receipts do not authenticate against a malicious same-user editor. Review the real workspace, command source, and artifact rather than treating state JSON as a trust oracle.

Lane F5: security and authority

Review in proportion to risk:

• user authority for writes, external messages, deployment, publication, or data changes;
• validation and handling of untrusted input;
• secret exposure in config, logs, artifacts, commands, and test fixtures;
• permission, path, symlink, injection, process, and network boundaries;
• trusted-command provenance for runtime verification;
• bounded timeout, cancellation, teardown, and partial-failure behavior;
• dependency or generated artifact changes;
• host-enforced worker restrictions and isolated worktrees.

Security uncertainty on a material boundary is rejection, not an optional note.

Lane F6: scope fidelity and integration

Compare request, approved plan, and final diff:

• no opportunistic feature, refactor, format sweep, dependency change, or public contract beyond must-have;
• no required behavior quietly deferred;
• unrelated user changes remain preserved;
• nested repositories and generated outputs are handled deliberately;
• cross-task integrations and imports are complete;
• docs and release surfaces describe current behavior without overclaiming.

An individually reasonable change can still fail because it was not authorized.

Evidence replay order

Review in dependency order:

1. current workspace and diff;
2. criterion revision and scope;
3. command provenance;
4. focused checks;
5. full relevant regression;
6. package/install/release surfaces;
7. real manual QA artifacts;
8. cleanup and process absence;
9. completion-oracle state.

If any source changes during remediation, invalidate and replay affected later steps.

Finding format

Every blocking finding includes:

ID and severity:
Lane:
Criterion or scope item:
Path/evidence:
Observed:
Expected:
User or system impact:
Required correction:
Required re-verification:

Use BLOCKER, MAJOR, MINOR, or NOTE. A minor can still block when the approved completion bar explicitly requires it.

Overall verdict

• Any lane with a blocker, unresolved material gap, missing proof, hedge, or conditional approval makes the overall verdict REJECTION.
• APPROVE IF, looks good but, and “follow up later” are rejection.
• Only all lanes clean yields UNCONDITIONAL APPROVAL.

Record unresolved findings as runtime blockers when a goal is active. Free-text evidence cannot pass a criterion.

Remediation loop

For rejection:

1. preserve all findings;
2. route fixes through the approved plan and csw-work;
3. rerun focused tests and affected real surfaces;
4. refresh receipts and cleanup;
5. re-run every affected lane, plus integration lanes when shared surfaces changed;
6. synthesize a new complete verdict.

Reviewers do not lower severity merely because remediation is inconvenient.

Binding outcome

UNCONDITIONAL APPROVAL is the human-facing quality gate, not a replacement for the goal runtime. Final completion still requires every current criterion to have valid machine evidence, zero open blockers, and the injected runtime invocation's complete to succeed.

The review report names reviewed commit/worktree state, lanes, findings, commands, artifact index, cleanup, verdict, and the exact next action.