| name | orc-review-verify |
| description | Review/verify worker for ORC Phases 5–6 (Opus 5 medium). Two modes: a REVIEW pass — examine the changed files against the code pattern + constraints, create/update tests, and classify EVERY finding on the P0–P3 severity ladder (P0/P1 gate ship, P2/P3 advisory); a VERIFY pass — run the build + full test suite and check each intent-spec definition-of-done criterion PLUS the pattern's enforceable validation-gate lines, reporting pass/fail; or an opt-in SECURITY pass — sweep the changed files against a supplied OWASP/STRIDE checklist. Returns findings[] with severity, a tests tally, and a verify result — it reports; the orchestrator owns the auto-fix-once loop. Dispatched in-pipeline after execution; distinct from the standalone /orc-verify (which checks only git-modified changes with no orchestrator). ALWAYS a spawned subagent — not for direct user invocation.
|
orc-review-verify
One entry point: spawned via the Task tool with subagent.md framing + the
slice, which points to core.md. core.md is the AUTHORITATIVE spec (full input
slice, both procedures, return contract); the summary below orients — on any
conflict, core.md wins.
What the worker does (summary)
- phase=review: examine
changed_files[] against code_pattern + constraints[];
create/update tests for the changed surface; independently re-check each
invariants[] rule AND each enforceable validation_gate[] line against the
diff (a violation is P0), plus any fe_rules[] pack rules on FE diffs
(file:line findings, P1–P3 by impact, never auto-P0); classify EVERY
finding on the ladder — P0 (failing tests/build, unmet acceptance
criteria, runtime errors, invariant violations — caller auto-fixes, no ask),
P1 (correctness/security risk, constraint violations — caller asks the
user before fixing), P2 (maintainability — advisory fix-batch offer),
P3 (cosmetic — advisory, counted). Never fix P2/P3; P0/P1 fixes are the
caller's call. Evidence-or-advisory: every P0–P2 finding carries
file:line + the offending line(s) quoted VERBATIM; an unanchored finding is
AUTO-P3 (never gates, never triggers a fix) — the caller spot-checks quotes
before acting on P0/P1.
- phase=verify (two halves, v0.33.0): FIRST the TDD gate — run the
plan's TDD suite (
tdd_suite[] in the slice) plus the build + full test
suite; green is the definition-of-done for non-exempt requirements, and each
red TDD test is a P0 tied to its requirement (the caller owns the
tdd_loop_max repair loop). THEN the adversarial review — attack the
green implementation: edge cases the spec missed, error paths, contract
violations, race/ordering, workflow breaks (dead wiring, broken commands) —
findings on the same P0–P3 ladder. Also fold the enforceable
validation_gate[] lines into the criteria set, then check EACH criterion
individually (pass/fail, nothing invented; an unmet gate line = unmet
criterion = P0), each with per-criterion evidence (the test/output line or
file:line that proves it); report failures precisely. You don't fix — the
caller owns the auto-fix-once loop.
- phase=security (opt-in): sweep only
changed_files[] against the supplied
security_checklist[] (wrap Semgrep if installed, never install it); findings
on the same ladder (exploitable-in-diff = P0, hardening gap = P1,
defense-in-depth = P2/P3). Report-only, no result verdict.
Return shape (summary — full contract in core.md)
{ phase, findings[]{severity: P0|P1|P2|P3, location "file:line", quote (verbatim, required P0–P2; unanchored ⇒ AUTO-P3), description, criterion}, criteria[] (verify only){criterion, result, evidence}, tests{added,updated,passing}, result (verify only): passed|failed, actual_model, actual_effort, failure_reason }
Validation checkpoint before returning: every finding carries a severity
from the P0–P3 ladder; every P0–P2 finding carries file:line + a VERBATIM
quote (else downgrade it to P3 yourself);
verify sets result + per-criterion evidence; actual_model is quoted
VERBATIM (never inferred). A
malformed return is treated as failure by the caller.
Fixed models (from ../../references/effort-and-mode.md): review on the OpenSpec/self
path = Opus 5 medium; review on the Superpowers path is delegated to the
Superpowers review skill instead (Sonnet 4.6 medium). Verify = Opus 5 medium,
always this subskill.
