| name | critique |
| description | The critique node of the canonical workflow — run two adversarial critics in parallel (implementer lens + user lens) against a freshly-planned spec, write their findings to .oh/tasks/<slug>/critique.md, and hand off to the approve gate. This is the EVIDENCE half of the critique→approve pair; it produces findings, it does not decide (the /approve gate decides). Runs on local artifacts only, before any GitHub-side state. TRIGGER when: a PRD/plan exists and needs adversarial review before the commitment gate; "critique this spec", "run the critics on <slug>", "review prd before we build". |
| argument-hint | <slug> [--weigh] |
Critique — two adversarial critics before commitment
The critique node of the canonical workflow (AGENTS.md § The Workflow), composed by /spec critique. It runs
the short adversarial feedback loop on a planned spec before anything is committed,
and writes the critique.md artifact the approve gate then decides on.
Core principle: surface risk while the spec is still the cheapest thing to revise.
Two critics with different framings review .oh/tasks/<slug>/prd.md in parallel —
symmetric critics waste context. Every finding is SEVERITY-tagged so the downstream
/approve gate can act on it mechanically. This node produces evidence; it does not
gate (that is /approve). Like the gate, it touches only local artifacts — no
GitHub-side state exists yet (critic-before-commitment). Formalizes /ship-spec Stage 3 as a
reusable node.
Inputs
| Arg | Meaning |
|---|
<slug> | The task slug — the critics read .oh/tasks/<slug>/prd.md; output is written to .oh/tasks/<slug>/critique.md. Required. |
If .oh/tasks/<slug>/prd.md is absent there is nothing to critique — print an error
pointing at /spec plan and emit no STATUS: token (honest exits: a missing
spec is a failure, not a clean critique).
Run the critics (parallel)
Launch two Agent tool calls in a single message (parallel execution) with
subagent_type: "critic". Both also receive the protected-paths cross-check: read
.claude/protected-paths.txt and flag any proposed deletion of an entry on that list
as SEVERITY: H unless the AC carries an explicit override note (the gate that would
have caught the v0.7 convergence regression, PR #212 / US-012).
Critic A — Implementer lens
You are an adversarial implementer reviewing a PRD before any code is written. Read
.oh/tasks/<slug>/prd.md. Read .claude/protected-paths.txt and treat its entries as
MUST-NOT-DELETE without an override note. Surface technical risks BEFORE
implementation. Focus on: (1) vague/unverifiable acceptance criteria; (2) missing
dependencies each story silently assumes; (3) pattern conflicts with this repo (read
AGENTS.md + the relevant .oh/skills/*/SKILL.md and sibling .oh/tasks/*/prd.json); (4) scope creep — "single
iteration" stories that are really 2+; (5) hidden destructive operations not
explicitly gated; (6) protected-path violations → SEVERITY: H + [PROTECTED-PATH].
Return:
CRITIC_A — IMPLEMENTER LENS
[SEVERITY: H/M/L] [STORY: US-NNN or *] [FINDING] | [EVIDENCE: file or AC text] | [RECOMMENDATION]
Critic B — User lens
You are an adversarial user reviewing a PRD before implementation. Read
.oh/tasks/<slug>/prd.md and .oh/context/USER.md (the single-developer / single-project
framing). Read .claude/protected-paths.txt and treat its entries as MUST-NOT-DELETE
without an override note. Surface scope and framing risks BEFORE the team commits.
Focus on: (1) scope ambiguity — what's missing from Non-Goals; (2) audience
misalignment vs. .oh/context/USER.md; (3) hidden expectations the PRD doesn't address;
(4) premature optimization — solving a problem the user doesn't have yet; (5) missing
rollback/escape hatch for destructive stories; (6) protected-path violations →
SEVERITY: H + [PROTECTED-PATH]. Return:
CRITIC_B — USER LENS
[SEVERITY: H/M/L] [STORY: US-NNN or *] [FINDING] | [EVIDENCE: file or PRD section] | [RECOMMENDATION]
Self-consistency weighting (optional, default-off)
Opt-in only. Default behavior is unchanged: without the flag this node runs exactly the
two single-pass critics above (K=1) and writes the same critique.md — zero added agent
calls, byte-for-byte the same cost model.
When invoked with --weigh (or CRITIQUE_WEIGH=1), each lens is sampled K times (K
default 1 = current behavior; opt-in K=2, hard cap 2) and the per-lens samples are routed
through /weigh (vote method) so a finding that fires in only 1 of K samples is
down-weighted relative to one that fires in all K, before the SEVERITY tally. This trades
up to 2× critic cost for a flakiness-resistant tally; it never raises a finding's severity,
only discounts low-consistency ones.
Invariants (the seam is schema-preserving):
- The
critique.md output schema below — ## Critic A, ## Critic B, ## Synthesis, the
SEVERITY tally, and the Recommendation — is identical in both modes; /approve and
/ship-spec Stage 4 parse it unchanged.
- A down-weighted finding is still listed (with a
consistency: n/K note), never silently
dropped; the SEVERITY auto-gate still fails closed on any unmitigated high.
K is capped at 2 — total critique cost never exceeds 2× the default 2-agent baseline.
Write .oh/tasks/<slug>/critique.md
Persist both critics' raw output plus a synthesis, in the exact shape /approve (and
/ship-spec Stage 4) parses:
# Critique — <slug>
Generated <date>; reviews `prd.md` post-/prd, pre-/ralph.
## Critic A — Implementer lens
<raw output>
## Critic B — User lens
<raw output>
## Synthesis
- **High-severity findings**: <count>
- **Medium-severity findings**: <count>
- **Recommendation**: PROCEED | REVISE-PRD | HALT
The Recommendation reflects the SEVERITY tally: any unmitigated SEVERITY: H (or
[PROTECTED-PATH]) → HALT; only M/L → PROCEED (note the acknowledged risks);
none → PROCEED. This node only records that judgment — the binding decision is the
/approve gate's.
What this skill does NOT do
- Decide. It produces
critique.md; the approve node reads it and emits the
APPROVED/DENIED verdict. Keep evidence and decision separate.
- Run symmetric critics. The two lenses MUST differ; identical framings add no
coverage.
- Touch GitHub-side state or the PRD. Output is
critique.md only. Revising
prd.md on a HALT is the plan node's job after /approve denies.
Memory Protocol
After a run, append to .oh/memory/<UTC-date>/log.md per .oh/skills/retro/references/memory-protocol.md:
## critique -- HH:MM UTC
- **Result**: OP
- **Spec**: <slug>
- **Findings**: H<n> / M<n> / L<n>
- **Recommendation**: PROCEED | REVISE-PRD | HALT
- **Observation**: <one sentence>
