| name | validate |
| description | Produce PASS/WARN/FAIL verdicts for artifacts, plans, code, PRs, or gates — including quick readiness/sanity checks before commit (absorbs vibe) and completion audits. Triggers: "validate an artifact", "PASS/WARN/FAIL verdict", "readiness / completion audit". |
$validate — Canonical Validator Skill
Loop position: move 6 (prove acceptance) of the operating loop — the driving adapter for the validate_acceptance port: every Given/When/Then must map to a passing test before a bead closes.
Role: validator. Input = artifact (plan, spec, code, PR, fitness gate). Output = verdict.v1 (PASS / WARN / FAIL with rationale + findings).
Status (2026-05-08): introduced ADDITIVE in Phase 1 (m6v5.D.1 / soc-78s2v). Existing validators (councilvibe/pre-mortem/red-team/review/eval-outcomes plus retired pr-validate and validation lanes) stayed until Phase 2 shim conversion (m6v5.D.2). Fix-C smoke (soc-wb2aa) gates Phase 2.
$validate is a driving adapter for the validate_acceptance port in the
Intent-to-Loop Hexagon.
When the artifact contains a hexagon: block, preserve the bounded context,
context packet, guard adapters, and done state in the verdict.
When the artifact claims DONE/closed/green, apply the
Completion-Claim Kernel
before returning PASS.
A verdict is re-plan evidence, not just a retry trigger. Under $rpi, a
FAIL/WARN (and its findings) surfaces UP to the orchestrator's
Agile Re-Plan Loop:
the remaining waves may be refactored, inserted, dropped, or reordered in
response — not only the failed objective re-cranked. Under --auto that pivot is
autonomous. Looping a failed objective forever without asking whether the plan
should change is the waterfall anti-pattern.
Modes (≤8 per Fix-F mode-flag budget)
| Mode | Purpose | Replaces (post-Phase 2) |
|---|
| (default) | 2-judge multi-judge consensus on any artifact | $council default |
--quick | Inline single-agent structured review | $council --quick |
--deep | 4-judge thorough review | $council --deep |
--mixed | Cross-vendor (Claude + Codex), N×2 judges | $council --mixed |
--debate | Adversarial 2-round refinement | $council --debate, $red-team |
--mode=post-impl | Code-readiness pipeline (complexity → bug-hunt → council) | vibe |
--mode=pre-impl [--target=X] | Plan/spec validation; target ∈ {scenario,fitness,ratchet,scope,skill,health} | $pre-mortem, $eval-outcomes, $goals measure, $flywheel, $scope, $skill-auditor, ao doctor |
--mode=pr | PR-shape verdict (diff review + acceptance check) | $review |
Mode-budget assertion: 8 modes. Adding a 9th requires demoting an existing one OR refusing the addition (per Fix-F § continuous CI gate).
Folded skills (cp-ki8): validation + pr-validate retired into these modes
The retired validation and pr-validate lanes were the Phase-1 placeholders for --mode=post-impl
and --mode=pr; both are now retired (cp-ki8) and their load-bearing contract folded
here so no capability is lost:
--mode=post-impl (was the validation lane) — full close-out + no-self-grading invariant.
Beyond the inline complexity → bug-hunt → council pipeline, this mode owns the
validate_acceptance port: every Given/When/Then from the intent issue must map to a
passing test (criterion→test roll-up; activity logs do not close beads), and the
acceptance verdict must be produced by a blind, context-isolated sub-agent judge that
did not author the code (author ≠ validator — ag-9jle.5 / ag-lmdx.4). Refuse to
certify acceptance when judge_id == author_id; the only escape is an inline-fallback
self-grade that is stamped as waived, not independently validated. Apply the
Completion-Claim Kernel
before accepting any DONE/closed/green claim. For epic-scope close-out this mode may
delegate to vibe, $post-mortem, and $forge rather than inlining them.--mode=pr (was the pr-validate lane) — submission-readiness checks. In addition to the
diff/acceptance verdict, run, in order: (1) upstream alignment FIRST (BLOCKING —
git rev-list --count HEAD..origin/main; fail if many commits behind or merge would
conflict), (2) CONTRIBUTING.md compliance (BLOCKING), (3) isolation — single commit
type + thematic files + atomic scope, (4) scope-creep containment, (5) quality gate
(tests/lint, non-blocking). On FAIL, emit remediation steps (split-by-type cherry-pick,
rebase-on-upstream) so the verdict is actionable.
Folded triggers (ag-s43tg wave 1): vibe + bead-completion-audit route here
vibe → --mode=post-impl. Use when doing a quick readiness or sanity check
that code is ready to commit or ship, short of a full review — the post-impl
pipeline (complexity → bug-hunt → council) is the vibe check.bead-completion-audit → --mode=post-impl close-out. Use when
auditing closed beads for real shipped evidence, acceptance proof, and truthful
closeout — the Completion-Claim Kernel and the no-self-grading invariant above
own this audit.
Quick Start
$validate path/to/plan.md
$validate --quick path/to/plan.md
$validate --deep path/to/spec.md
$validate --mode=pre-impl path/to/plan.md
$validate --mode=post-impl recent
$validate --mode=pr 123
$validate --mode=pre-impl --target=fitness
Default uses runtime-native subagent spawning. Falls back to --quick (inline) when no multi-agent capability detected.
Execution
Step 1: Resolve mode + target
Parse --mode and --target. Default mode is multi-judge. Validate combinations:
| Mode | Allowed --target |
|---|
| default, --quick, --deep, --mixed, --debate | n/a |
| --mode=post-impl | n/a (pipeline scope is recent code changes) |
| --mode=pre-impl | scenario, fitness, ratchet, scope, skill, health (default: pre-mortem on plan) |
| --mode=pr | n/a (PR ID/path is positional) |
Reject invalid combinations (e.g., --mode=pr --target=fitness).
Step 2: Load artifact + context
ARTIFACT="${1:-recent}"
For --mode=pre-impl, also load:
.agents/planning-rules/*.md (compiled planning rules).agents/findings/registry.jsonl (active findings).agents/pre-mortem-checks/*.md (compiled prevention)
For --mode=post-impl, run pre-checks:
- complexity audit (radon for python, gocyclo for go)
- bug-hunt sweep (skill-body convention; no
$review skill needed)
For --mode=pr, fetch the PR diff (gh pr diff <id> or path).
Step 3: Determine spawn backend
spawn_agent available → Codex sub-agentTeamCreate available → Claude native teamtask (read-only skill tool, OpenCode) → opencode subagent- None → fall back to
--quick (inline single-agent)
Log selected backend in the verdict frontmatter.
Step 4: Run judges
| Mode | Judges | Perspectives |
|---|
| default | 2 | independent (no labeled perspectives) |
| --deep | 4 | missing-requirements, feasibility, scope, spec-completeness |
| --mixed | 2N (default N=3) | same N perspectives across Claude + Codex |
| --debate | 2+ rounds | adversarial; 2 rounds with critique-rebuttal |
| --quick | 0 (inline self) | structured review |
| --mode=post-impl | 2 + pipeline | complexity → bug-hunt → 2-judge council |
| --mode=pre-impl | 2-4 | per target preset |
| --mode=pr | 2 | diff-review + acceptance-check |
Each judge gets:
- artifact path
- relevant context (planning rules, findings)
- council FAIL pattern check prompt (top 8)
- temporal interrogation prompt (--deep + --target=plan)
Step 5: Mandatory checks (auto-trigger)
For --mode=pre-impl --target=plan:
- temporal interrogation (auto for plans with 5+ files or 3+ deps)
- error & rescue map
- council FAIL pattern check (top 8)
- test pyramid coverage check
- input validation check (enum-like fields)
For --mode=post-impl:
- L0/L1/L2 coverage check on changed files
For --mode=pre-impl --target=fitness:
- read GOALS.md
- evaluate each gate against current state
- report PASS/WARN/FAIL per gate + aggregate
Step 6: Consolidate to verdict
Each judge returns a per-judge result. Consolidate:
- PASS only if all judges PASS (or majority for --deep)
- WARN if any judge raises a warning the others don't dispute
- FAIL if any judge raises a blocker the others don't override
Step 7: Write verdict
Output path: .agents/council/YYYY-MM-DD-validate-<topic-slug>.md
---
id: validate-YYYY-MM-DD-<slug>
type: verdict
date: YYYY-MM-DD
mode: <mode>
target: <target or n/a>
artifact: <path>
backend: <codex-subagents | claude-teams | opencode | inline>
---
# Validate Verdict — <topic>
## Council Verdict: PASS / WARN / FAIL
| Failure mode | Risk | Severity | Addressed? |
|---|---|---|---|
| ... | ... | ... | ... |
## Pseudocode Fixes (when WARN/FAIL)
(copy-pastable into affected issues per pre-mortem 4.6 contract)
## FAIL Pattern Check
(top 8 patterns — status per pattern)
## Verdict
PASS — proceed
WARN — review concerns, accept risk, or apply fixes
FAIL — block; revise artifact and rerun
The exact heading ## Council Verdict: PASS / WARN / FAIL is mandatory — downstream validators and ledger readers parse it with anchored regex.
Step 8: Persist findings (when applicable)
For --mode=pre-impl reusable findings: append to .agents/findings/registry.jsonl (atomic temp+rename).
Step 9: Report
- Verdict (PASS/WARN/FAIL).
- Key concerns (when not PASS).
- Output path.
- Recommended next action.
--target taxonomy (pre-impl)
--target | What gets graded | Replaces |
|---|
| (default) | Plan/spec for an upcoming $implement | $pre-mortem |
| scenario | Holdout scenario gate | $eval-outcomes |
| fitness | GOALS.md fitness gates | $goals measure, ao goals measure |
| ratchet | Brownian Ratchet checkpoint | $flywheel, ao ratchet status |
| scope | Frozen-dirs declaration | $scope |
| skill | SKILL.md hygiene + audit | $skill-auditor, $heal-skill (audit half) |
| health | Repo health probe | ao doctor |
Each target has its own inline check rubric until Phase 2 extraction.
Validation discipline (2026-06-09, cards 6–10, cp-hhd7)
Verdict form — the gate parses these lines anchored
VERDICT: PASS
(blank line)
COMMANDS RUN:
<actual commands + verbatim output snippets>
REASONS:
- bullet citing a COMMANDS RUN line
A verdict with no COMMANDS RUN: section is unverified — reject it and
dispatch a fresh validator. A verdict whose COMMANDS RUN: lists only commands
the author ran (not the judge) is a counterfeit judge — treat as FAIL and
re-route to a genuinely independent validator. No ## headings or parentheticals
on VERDICT: or COMMANDS RUN: lines; the gate parses them anchored.
Cross-family floor (POLICY → gate icb6 enforces; this skill supports)
For assurance closes (the control-plane verdict-gate, cp-icb6), the floor is
≥2 verdicts from ≥2 distinct model families, author family excluded, fail-closed.
This skill supports that policy via --mixed mode and the verdict form above;
the policy itself lives in the gate, not here. A same-model council is valid for
non-assurance decisions (design brainstorms, quick checks) — do not refuse those.
Tier mapping:
- STRICT (irreversible, security, production close): Codex + strong-Gemini (A1); or Fable + Codex (A2).
- ROUTINE (everyday close): Codex + non-author-Claude (A3); or Fable + non-author (A4).
The A7 ruling (2026-06-09, memory validation-family-policy-risk-tiered):
Gemini is currently benched for STRICT validation — use Codex + Fable for A1/A2
tiers. Gemini may return for STRICT when Bo graduates it from the bench. Do not
present Gemini paths as live for tier A1/A2 until then.
Judge empirically on a differentiating fixture (card 9, cp-8720)
When two implementations of the same intent exist, do NOT award based on
authorship or surface aesthetics. Run both on a differentiating fixture (an
input that exposes their behavioral difference), record the outputs verbatim, and
graft the loser's unique assets onto the winner. "My worker wrote it" is not
evidence.
Dispatch record first (card 3, cp-hhtu)
Before dispatching a validator, register intent on the bead graph (update status,
assign actor). Two parallel validators on the same bead produce a dedup incident,
not a cross-family quorum. Check for an existing actor before spawning.
Judges re-measure; they do not read (card 8)
A judge re-runs the cited commands on the actual artifacts. It does not read the
author's evidence file and agree. Attest judge_source: <model> inside COMMANDS RUN
so the gate can confirm the judge identity. A judge that ran nothing is a reader, not
a verifier — discard its verdict.
Evidence figures are measured, never inferred (the cp-801l lesson)
A worker's evidence file may only contain numbers and outputs that were captured —
pasted verbatim from a command's output — never reconstructed from memory. The
canonical failure: "36 checks — 35 pass" stated with confidence was inference; the
measured reality was 36 run / 34 pass / 1 fail / 1 skip, on a different commit.
- Validators: treat any uncited figure (a count, a pass-rate, a timing, a commit
hash) as unverifiable → FAIL until a log is produced or the claim is corrected.
- Corrections happen by APPENDED erratum — a dated erratum block crediting the
source measurement — never by silently editing the original figure. A silently
edited evidence file is indistinguishable from a fabricated one.
Constraints (one-role-per-skill)
- One role: validator. Output is always a verdict. Never mutates code (delegates to
$implement for fixes).
- No new modes without dropping/merging an existing one (Fix-F mode-budget cap = 8).
- Verdict heading is regex-anchored — do not alter the
## Council Verdict: ... text format.
See Also
skills/rpi/SKILL.md — orchestrator that fires $validate --mode=pre-impl after $planskills/curate/SKILL.md — miner role (paired canonical skill)schemas/verdict.v1.schema.json — output contractpre-land-refuters — the mutate-shared-trunk pawl: a self-administered DONE is a claim, so an unbiased fresh-context refuter (model-agnostic by default; multi-model ≥2 families opt-in for the highest-irreversibility doors) attacks it at the push boundary regardless of complexity (complexity scales the panel's depth, never exempts the gate); its CONFIRMED verdict is enforced executably by scripts/reconcile-pr.sh
Reference Documents
- references/validate.feature — Executable spec: verdict.v1 PASS/WARN/FAIL for any artifact, --mode selects shape, 8-mode budget (soc-qk4b)
Reference library (incl. rescued vibe references, ag-s43tg)
