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

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

ao metrics cite "<learning-path>" --type applied 2>/dev/null || true

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/ directory AND no .agents/specs/ directory
            Skip if: --no-behavioral flag set
            
            Sub-steps:
              a) List active scenarios and agent-built specs:
                   ao scenario list --status active 2>/dev/null
                   find .agents/specs -name "*.json" -type f 2>/dev/null
              a.5) For each agent-built spec in .agents/specs/, treat as a scenario
                   with source="agent". Validate against scenario schema (auto-* id
                   pattern). Add to evaluation set alongside holdout scenarios.
              b) If 0 scenarios AND 0 specs → skip with note "No behavioral validation artifacts found"
              c) Spawn evaluator council with AGENTOPS_HOLDOUT_EVALUATOR=1
                 Pass scenarios + implementation diff as judge context
              d) Each judge evaluates: "Does the implementation satisfy the scenario's
                 expected_outcome? Score each acceptance_vector dimension 0.0-1.0."
              e) Compute satisfaction_score per scenario (mean of dimension scores)
              f) Aggregate: mean satisfaction across all scenarios
              g) Gate:
                   mean >= scenario.satisfaction_threshold → PASS
                   mean >= 0.5 → WARN ("Partial satisfaction — review scenarios")
                   mean < 0.5 → FAIL ("Implementation does not satisfy holdout scenarios")
              h) Write results to .agents/rpi/scenario-results.json
              i) Include satisfaction_score in validation_state
            
            PASS/WARN? → continue to STEP 2
            FAIL? → write summary, output <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 ensure-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>

validation_state = {
  epic_id: "<epic-id or null>",
  complexity: <fast|standard|full>,
  no_retro: <true if --no-retro>,
  no_forge: <true if --no-forge>,
  strict_surfaces: <true if --strict-surfaces>,
  vibe_verdict: null,
  post_mortem_verdict: null
}

# 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>

$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

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