validation（validation 搜索）

/validation — Full Validation Phase Orchestrator

YOU MUST EXECUTE THIS WORKFLOW. Do not just describe it.

Strict Delegation Contract (default)

Validation delegates to /vibe, /post-mortem, /retro, and /forge (plus lifecycle skills /test, /deps, /review, /perf) via Skill(skill="<name>",...) calls — separate tool invocations. Strict delegation is the default.

Anti-pattern to reject: spawning judges via Agent() in place of /vibe, inlining post-mortem analysis, skipping /forge. See ../shared/references/strict-delegation-contract.md for the full contract and supported compression escapes (--quick, --no-retro, --no-forge, --no-lifecycle, --no-behavioral, --allow-critical-deps).

See .agents/learnings/2026-04-19-orchestrator-compression-anti-pattern.md for the live compression signature.

DAG — Execute This Sequentially

mkdir -p .agents/rpi
detect complexity from execution-packet or --complexity flag (default: standard)
detect ao CLI availability

Step 0: Load Prior Validation Context

Before running the validation pipeline, pull relevant learnings from prior reviews:

if command -v ao &>/dev/null; then
    ao lookup --query "<epic or goal context> validation review patterns" --limit 5 2>/dev/null || true
fi

Apply retrieved knowledge (mandatory when results returned): for each returned item, check applicability; if applicable, include as a known_risk (pattern + does-code-exhibit-it check); cite by filename when it influences a finding; record via ao metrics cite "<path>" --type applied. Skip silently if ao unavailable or returns no results.

*(orchestrator-owned: this knowledge retrieval is intentionally inline CLI, not a Skill() delegation. Do NOT expand into a separate /research --validation-context call — subsequent steps delegate to vibe/post-mortem/retro/forge.)*

Run every step in order. Do not stop between steps.

Step ordering precedence (STEPS 1 → 1.5 → 1.6 → 1.7 → 1.8 → 2 → …): STEP 1 (/vibe) runs first and determines whether the pipeline continues. STEPS 1.5 (four-surface closure), 1.6 (test pyramid), 1.7 (lifecycle checks), and 1.8 (behavioral) are separate orchestrator steps that run after vibe, not inline inside vibe. /vibe owns code quality; the surface/test/lifecycle/behavioral gates are additional closure checks layered on top.

STEP 1  ──  Skill(skill="vibe", args="recent [--quick]")
              Use --quick for fast/standard. Full council for full.
              PASS/WARN? → continue
              FAIL?      → write summary, output <promise>FAIL</promise>, stop
                           (validation cannot fix code — caller decides retry)

STEP 1.5 ── Four-Surface Closure (mandatory)
              Read `skills/validation/references/four-surface-closure.md` for the mandatory four-surface closure check.
              Check all four surfaces: Code, Documentation, Examples, Proof.
              All 4 pass? → continue
              if --strict-surfaces:
                Any surface fails? → FAIL, write summary, output <promise>FAIL</promise>, stop
              else (default):
                Code passes, others fail? → WARN, continue
                Code fails? → BLOCK, write summary, output <promise>FAIL</promise>, stop

STEP 1.6 ── Test pyramid coverage audit (advisory, append to summary)
              Check L0-L3 + BF1/BF4 per modified file. WARN only, not FAIL.

STEP 1.7 ── Lifecycle Checks (advisory except critical dependency findings)
              Skip entire step if: --no-lifecycle flag.
              Each sub-step uses --quick mode to limit context consumption.
              On budget expiry: skip remaining sub-steps, write [TIME-BOXED].

              a) if lifecycle tier >= minimal AND test_framework_detected:
                   Skill(skill="test", args="coverage --quick")
                   Append coverage delta to phase summary.

              b) if lifecycle tier >= standard AND dependency_manifest_exists:
                   Skill(skill="deps", args="vuln --quick")
                   CRITICAL vulns (CVSS >= 9.0): **FAIL** (block shipping). Opt-out: `--allow-critical-deps` for acknowledged risk acceptance.
                   Non-critical: advisory note only.

              c) if lifecycle tier >= standard:
                   Skill(skill="review", args="--diff --quick")
                   Append review findings to summary as advisory.

              d) if lifecycle tier == full AND modified_files_touch_hot_path:
                   Skill(skill="perf", args="profile --quick")
                   Append perf findings to summary as advisory.
                   Hot path detection: modified files match benchmark files
                   or patterns (handler, middleware, router, parser, engine,
                   worker, pool, codec).

STEP 1.8 ── Stage 4: Behavioral Validation (holdout scenarios + agent-built specs)
            Skip if: no .agents/holdout/ AND no .agents/specs/, or --no-behavioral
            Read `references/step-1.8-behavioral-validation.md` for full sub-steps.
            Loads holdout scenarios + agent specs → evaluator council → satisfaction gate.
            Evaluates each scenario and aggregates results into `satisfaction_score`
            (verdict schema field, `skills/council/schemas/verdict.json`: number 0.0-1.0,
            "Probabilistic satisfaction score (0.0 = unsatisfied, 1.0 = fully satisfied)").
            Per-dimension scores populate `satisfaction_breakdown`. The aggregated
            `satisfaction_score` seeds downstream gates and the phase summary.
            PASS/WARN? → continue | FAIL? → <promise>FAIL</promise>, stop

STEP 2  ──  if epic_id:
              Skill(skill="post-mortem", args="<epic-id> [--quick]")
            else:
              Skill(skill="post-mortem", args="recent [--quick]")
              Use --quick for fast/standard. Full council for full.
              PASS/WARN? → continue
              FAIL?      → write summary, output <promise>FAIL</promise>, stop

STEP 3  ──  if not --no-retro:
              Skill(skill="retro")

STEP 4  ──  if not --no-forge AND ao available:
              if [ -n "${CODEX_THREAD_ID:-}" ] || [ "${CODEX_INTERNAL_ORIGINATOR_OVERRIDE:-}" = "Codex Desktop" ]; then
                ao codex stop --auto-extract 2>/dev/null || true
              else
                ao forge transcript --last-session --queue --quiet 2>/dev/null || true
              fi

STEP 5  ──  write phase summary to .agents/rpi/phase-3-summary-YYYY-MM-DD-<slug>.md
              ao ratchet record vibe 2>/dev/null || true
              output <promise>DONE</promise>

That's it. Steps 1→2→3→4→5. No stopping between steps.

Setup Detail

Track state inline: epic_id, complexity, no_retro, no_forge, strict_surfaces, vibe_verdict, post_mortem_verdict. Load execution packet (if available): read complexity, contract_surfaces, and done_criteria from .agents/rpi/execution-packet.json. When a current run_id is known, prefer the matching .agents/rpi/runs/<run-id>/execution-packet.json archive over the latest alias.

Gate Detail

Validation has multiple blocking conditions. Validation cannot fix code — it can only report and fail closeout when the lifecycle contract is not met.

• Blocking FAIL conditions: vibe FAIL, code-surface failure in STEP 1.5, --strict-surfaces failure on any closure surface, CVSS >= 9.0 dependency findings in STEP 1.7b unless --allow-critical-deps, and post-mortem FAIL in STEP 2.
• PASS/WARN: Log verdicts, continue through the remaining steps.
• FAIL: Extract findings from the latest evaluator output, write phase summary with FAIL status, output <promise>FAIL</promise> with findings attached. Suggest: "Validation FAIL. Fix findings, then re-run /validation [epic-id]".

Why no internal retry: Retries require re-implementation (/crank). The caller (/rpi or human) decides whether to loop back.

Phase Summary Format

Write to .agents/rpi/phase-3-summary-YYYY-MM-DD-<slug>.md:

# Phase 3 Summary: Validation

- **Epic:** <epic-id or "standalone">
- **Vibe verdict:** <PASS|WARN|FAIL>
- **Post-mortem verdict:** <verdict or "skipped">
- **Retro:** <captured|skipped>
- **Forge:** <mined|skipped>
- **Complexity:** <fast|standard|full>
- **Status:** <DONE|FAIL>
- **Timestamp:** <ISO-8601>

Phase Budgets

On budget expiry: allow in-flight calls to complete, write [TIME-BOXED] marker, proceed.

Flags

Quick Start

/validation ag-5k2                        # validate epic with full close-out
/validation                               # validate recent work (no epic)
/validation --complexity=full ag-5k2      # force full council ceremony
/validation --no-retro ag-5k2             # skip retro only
/validation --no-forge ag-5k2             # skip forge only

Completion Markers

<promise>DONE</promise>    # Validation passed, learnings captured
<promise>FAIL</promise>    # Vibe failed, re-implementation needed (findings attached)

Troubleshooting

Reference Documents

• references/four-surface-closure.md — four-surface closure validation (code + docs + examples + proof)
• references/forge-scope.md — forge session scoping and deduplication
• references/idempotency-and-resume.md — re-run behavior and standalone mode

See Also

Core phases: vibe, post-mortem, retro, forge, crank, discovery, rpi. Lifecycle Step 1.7: test, deps, review, perf.