vibe（vibe 搜索）

Vibe Skill

Purpose: Is this code ready to ship?

Three steps:

1. Complexity analysis — Find hotspots (radon, gocyclo)
2. Bug hunt audit — Systematic sweep for concrete bugs
3. Council validation — Multi-model judgment

Quick Start

/vibe                                    # validates recent changes
/vibe recent                             # same as above
/vibe src/auth/                          # validates specific path
/vibe --quick recent                     # fast inline check, no agent spawning
/vibe --structured recent                # 6-phase verification report (build→types→lint→tests→security→diff)
/vibe --deep recent                      # 3 judges instead of 2
/vibe --sweep recent                     # deep audit: per-file explorers + council
/vibe --mixed recent                     # cross-vendor (Claude + Codex)
/vibe --preset=security-audit src/auth/  # security-focused review
/vibe --explorers=2 recent               # judges with explorer sub-agents
/vibe --debate recent                    # two-round adversarial review
/vibe --tier=quality recent              # use quality tier for council calls

Execution Steps

Step 0: Load Prior Review Context

Before reviewing, pull relevant learnings from prior code reviews and known patterns:

if command -v ao &>/dev/null; then
    ao lookup --query "<target-scope> code review patterns" --limit 3 2>/dev/null || true
fi

Apply retrieved knowledge (mandatory when results returned):

If learnings or patterns are returned, do NOT just load them as passive context. For each returned item:

1. Check: does this learning apply to the code under review? (answer yes/no)
2. If yes: include it as a known_risk in your review — state the pattern, what to look for, and whether the code exhibits it
3. Cite the learning by filename in your review output when it influences a finding

After applying, record the citation:

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

Skip silently if ao is unavailable or returns no results.

Project reviewer config: If .agents/reviewer-config.md exists, its full config (reviewers, plan_reviewers, skip_reviewers) is passed to council for judge selection. See skills/council/SKILL.md Step 1b.

Crank Checkpoint Detection

Before scanning for changed files via git diff, check if a crank checkpoint exists:

if [ -f .agents/vibe-context/latest-crank-wave.json ]; then
    echo "Crank checkpoint found — using files_changed from checkpoint"
    FILES_CHANGED=$(jq -r '.files_changed[]' .agents/vibe-context/latest-crank-wave.json 2>/dev/null)
    WAVE_COUNT=$(jq -r '.wave' .agents/vibe-context/latest-crank-wave.json 2>/dev/null)
    echo "Wave $WAVE_COUNT checkpoint: $(echo "$FILES_CHANGED" | wc -l | tr -d ' ') files changed"
fi

When a crank checkpoint is available, use its files_changed list instead of re-detecting via git diff. This ensures vibe validates exactly the files that crank modified.

Step 1: Determine Target

If target provided: Use it directly.

If no target or "recent": Auto-detect from git:

# Check recent commits
git diff --name-only HEAD~3 2>/dev/null | head -20

If nothing found, ask user.

Pre-flight: If no files found: Return immediately with: "PASS (no changes to review) — no modified files detected." Do NOT spawn agents for empty file lists.

Step 1.5a: Structured Verification Path (--structured mode)

If --structured flag is set, run a 6-phase mechanical verification pipeline instead of the council flow. This produces a machine-readable verification report suitable for PR gates and CI integration.

Phases: Build → Types → Lint → Tests → Security → Diff Review.

Read references/verification-report.md for the full report template and per-phase commands. Each phase is fail-fast — if Build fails, skip remaining phases and report NOT READY.

After all phases complete, write the structured report to .agents/council/YYYY-MM-DD-verification-<target>.md and output the summary table to the user.

When to use: Pre-PR gate, CI integration, when you need a mechanical pass/fail rather than judgment-based review.

Step 1.5: Fast Path (--quick mode)

If --quick flag is set, skip Steps 2a through 2e as heavy pre-processing, plus 2.5 and 2f, and jump to Step 4 with inline council after Steps 2.3, 2.4, 2g, and Step 3. Domain checklists, compiled-prevention loading, test-pyramid inventory, and inline product context are cheap and high-value, so they still run in quick mode. Complexity analysis (Step 2) still runs — it's cheap and informative.

Why: Steps 2.5 and 2a–2f add 30–90 seconds of pre-processing that mainly feed multi-judge council packets. In --quick mode (single inline agent), those inputs are not worth the cost, but test-pyramid and product-context checks still shape the inline review meaningfully.

Step 2: Run Complexity Analysis

Filter by language present in the change set first. Run only the analyzers whose language actually appears in the diff. A docs/shell/BATS-only epic must NOT trigger gocyclo against the entire cli/ tree (it has hung in past runs); a Python-free epic must NOT trigger radon.

# Detect which languages are present in the diff (or in <path> for full audits).
# Use `git diff --name-only <base>...HEAD` for a PR; fall back to listing
# files under <path> when no diff base is available.
mkdir -p .agents/council
HAS_GO=false; HAS_PY=false
DIFF_FILES="$(git diff --name-only "${BASE:-HEAD~1}"...HEAD 2>/dev/null || find <path> -type f)"
echo "$DIFF_FILES" | grep -q '\.go$'  && HAS_GO=true
echo "$DIFF_FILES" | grep -q '\.py$'  && HAS_PY=true
echo "$(date -Iseconds) preflight: HAS_GO=$HAS_GO HAS_PY=$HAS_PY" >> .agents/council/preflight.log

For Python (only when HAS_PY=true):

if [ "$HAS_PY" = "true" ]; then
  echo "$(date -Iseconds) preflight: checking radon" >> .agents/council/preflight.log
  if ! which radon >> .agents/council/preflight.log 2>&1; then
    echo "⚠️ COMPLEXITY SKIPPED: radon not installed (pip install radon)"
  else
    radon cc <path> -a -s 2>/dev/null | head -30
    radon mi <path> -s 2>/dev/null | head -30
  fi
else
  echo "ℹ️ COMPLEXITY SKIPPED: no .py files in diff"
fi

For Go (only when HAS_GO=true):

if [ "$HAS_GO" = "true" ]; then
  echo "$(date -Iseconds) preflight: checking gocyclo" >> .agents/council/preflight.log
  if ! which gocyclo >> .agents/council/preflight.log 2>&1; then
    echo "⚠️ COMPLEXITY SKIPPED: gocyclo not installed (go install github.com/fzipp/gocyclo/cmd/gocyclo@latest)"
  else
    gocyclo -over 10 <path> 2>/dev/null | head -30
  fi
else
  echo "ℹ️ COMPLEXITY SKIPPED: no .go files in diff"
fi

For other languages: Skip complexity with explicit note: "⚠️ COMPLEXITY SKIPPED: No analyzer for "

Interpret results:

Include complexity findings in council context.

Step 2.3: Load Domain-Specific Checklists

Detect code patterns in the target files and load matching domain-specific checklists from standards/references/:

For each matched checklist, load it via the Read tool and include relevant items in the council packet as context.domain_checklists. Multiple checklists can be loaded simultaneously.

Skip silently if no patterns match. This step runs in both --quick and full modes (domain checklists are cheap to load and high-value).

Steps 2.4-2f, 2h, 3-3.6 (Deep Checks & Pre-Council Prep): Read references/deep-checks.md for compiled prevention, prior findings, pre-council deep analysis checks, product context, spec loading, suppressions, pre-mortem correlation, and model cost tiers. Loaded automatically unless --quick mode is set. In --quick mode, skip directly to Step 2g.

Compiled prevention inputs: Load .agents/pre-mortem-checks/ and .agents/planning-rules/ when available. These compiled artifacts contain known_risks from prior findings that inform the review — carry matched finding IDs into council context so judges can assess whether the flywheel prevented rediscovery.

Step 2a: Prior Findings Check

Skip if --quick. Load prior findings from .agents/findings/registry.jsonl.

Step 2b: Constraint Tests

Skip if --quick. Run compiled constraint tests from .agents/constraints/.

Step 2c: Metadata Checks

Skip if --quick. Verify file metadata consistency.

Step 2.5: OL Validation

Skip if --quick. Run organizational-lint checks.

Step 2d: Knowledge Search

Skip if --quick. Search for relevant prior learnings via ao lookup.

Step 2e: Bug Hunt

Skip if --quick. Run proactive bug-hunt audit on target files.

Step 2f: Codex Review

Skip if --quick. When --mixed is passed and Codex CLI is available, send the first 2000 chars of the diff to Codex for a parallel review. Cap input at 2000 chars to stay within Codex context budgets.

Step 3: Product Context

Skip if --quick as a separate judge-fanout step. When PRODUCT.md exists and the user did not pass an explicit --preset override, quick mode still loads DX expectations inline in the single-agent review. In non-quick modes, add a DX (developer experience) judge: 2 independent + 1 DX judge (3 judges total). The DX judge evaluates whether the code aligns with the product's stated personas and value propositions.

Step 2g: Test Pyramid Inventory (MANDATORY)

Assess test coverage against the test pyramid standard (the test pyramid standard (loaded via /standards)).

Read skills/vibe/references/test-pyramid-weighting.md for test pyramid weighting — L3+ tests found all production bugs, weight them 5x.

Test Pyramid Weighting: Weight test coverage by level: L0–L1 at 1x, L2 at 3x, L3+ at 5x. Unit-only coverage is a WARN signal, not a PASS. See references/test-pyramid-weighting.md.

Run even in --quick mode — this is cheap (file existence checks) and high-signal.

1. Identify changed modules from git diff or target scope
2. For each changed module, check coverage pyramid (L0–L3):

- L0: Does a contract/spec enforcement test cover this module? - L1: Does a unit test file exist for this module? - L2: If module crosses boundaries, does an integration test exist?

1. For boundary-touching code, check bug-finding pyramid (BF1–BF5):

- BF4 (Chaos): Do external call sites have failure injection tests? - BF1 (Property): Do data transformations have property tests? - BF2 (Golden): Do output generators have golden file tests?

1. Compute weighted pyramid score for changed code paths: Formula: weighted_score = (L0_count x 1 + L1_count x 1 + L2_count x 3 + L3_count x 5 + L4_count x 5) / max_possible Where max_possible = total_test_count x 5 (the score if every test were L3+). Count tests at each level for changed code paths: Interpretation: Satisfaction exposure: The weighted_score is also exposed as satisfaction_score (with source "test-pyramid-weighted") in the test_pyramid output block AND promoted to the top-level verdict JSON as 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). Optional — absent means not computed."). Downstream consumers (e.g., /validation STEP 1.8 holdout evaluation) can use satisfaction_score as a normalized quality signal. Include in council packet and vibe report output: ## Test Pyramid Score | Level | Count | Weight | Contribution | |-------|-------|--------|--------------| | L0 | 2 | 1x | 2 | | L1 | 8 | 1x | 8 | | L2 | 0 | 3x | 0 | | L3 | 0 | 5x | 0 | | L4 | 0 | 5x | 0 | | **Total** | **10** | | **10 / 50 = 0.20** | WARN: weighted_score 0.20 < 0.3 and all tests are L0-L1 only

- L0: Build/compile checks (weight 1) - L1: Unit tests (weight 1) - L2: Integration tests (weight 3) - L3: E2E/system tests (weight 5) - L4: Smoke/fresh-context tests (weight 5) - weighted_score >= 0.6 — strong pyramid, L2+ tests present - 0.3 <= weighted_score < 0.6 — acceptable, but recommend more integration tests - weighted_score < 0.3 AND all tests are L0-L1 only — WARN: unit-only test coverage (feeds into vibe verdict as a WARN signal, not a separate gate)

1. Build coverage table and include in council packet as context.test_pyramid:

"test_pyramid": {
  "coverage": {
    "L0": {"status": "pass", "files": ["test_spec_enforcement.py"]},
    "L1": {"status": "pass", "files": ["test_module.py"]},
    "L2": {"status": "gap", "reason": "crosses subsystem boundary, no integration test"}
  },
  "bug_finding": {
    "BF4_chaos": {"status": "gap", "reason": "external API calls without failure injection"},
    "BF1_property": {"status": "na", "reason": "no data transformations in scope"}
  },
  "weighted_score": 0.20,
  "satisfaction_score": 0.20,
  "satisfaction_source": "test-pyramid-weighted",
  "score_breakdown": {"L0": 2, "L1": 8, "L2": 0, "L3": 0, "L4": 0},
  "max_possible": 50,
  "warn_unit_only": true,
  "verdict": "WARN: weighted_score 0.20 < 0.3, all tests L0-L1 only"
}

Verdict rules:

• weighted_score < 0.3 AND all tests L0-L1 only — WARN: unit-only coverage (include in council findings)
• Missing L1 on feature code — WARN (include in council findings)
• Missing L0 on spec-changing code — WARN
• Missing BF4 on boundary code — WARN (advisory, not blocking)
• All levels covered with weighted_score >= 0.6 — no mention needed

When coverage gaps are found, run /test <module> to generate test candidates for uncovered code.

Step 4: Run Council Validation

With spec found — use code-review preset:

/council --preset=code-review validate <target>

• error-paths: Trace every error handling path. What's uncaught? What fails silently?
• api-surface: Review every public interface. Is the contract clear? Breaking changes?
• spec-compliance: Compare implementation against the spec. What's missing? What diverges?

The spec content is injected into the council packet context so the spec-compliance judge can compare implementation against it.

Without spec — 2 independent judges (no perspectives):

/council validate <target>

2 independent judges (no perspective labels). Use --deep for 3 judges on high-stakes reviews. Override with --quick (inline single-agent check) or --mixed (cross-vendor with Codex).

Council receives:

• Files to review
• Complexity hotspots (from Step 2)
• Git diff context
• Spec content (when found, in context.spec)
• Sweep manifest (when --deep or --sweep, in context.sweep_manifest — judges shift to adjudication mode, see references/deep-audit-protocol.md)

All council flags pass through: --quick (inline), --mixed (cross-vendor), --preset=<name> (override perspectives), --explorers=N, --debate (adversarial 2-round), --tier=<name> (model cost tier: quality/balanced/budget). See Quick Start examples and /council docs.

Step 5: Council Checks

Each judge reviews for:

Step 6: Interpret Verdict

Council Verdict:

Step 7: Write Vibe Report

Write to: .agents/council/YYYY-MM-DD-vibe-<target>.md (use date +%Y-%m-%d)

Read references/report-format.md for the full vibe report markdown template. The report includes: complexity analysis, council verdict table, shared/critical/informational findings, all findings (when --deep/--sweep), recommendation, and decision checkboxes.

Step 8: Report to User

Tell the user:

1. Complexity hotspots (if any)
2. Council verdict (PASS/WARN/FAIL)
3. Key concerns
4. Location of vibe report

Step 9: Record Ratchet Progress

After council verdict:

1. If verdict is PASS or WARN:

- Run: ao ratchet record vibe --output "<report-path>" 2>/dev/null || true - Suggest: "Run /post-mortem to capture learnings and complete the cycle."

1. If verdict is FAIL:

- Do NOT record ratchet progress. - Extract ALL findings from the council report for structured retry context (group by category if >20): Read the council report. For each finding, format as: FINDING: <description> | FIX: <fix or recommendation> | REF: <ref or location> Fallback for v1 findings (no fix/why/ref fields): fix = finding.fix || finding.recommendation || "No fix specified" ref = finding.ref || finding.location || "No reference" - Tell user to fix issues and re-run /vibe, including the formatted findings as actionable guidance.

Step 9.5: Feed Findings to Flywheel

If verdict is WARN or FAIL, persist reusable findings to .agents/findings/registry.jsonl and optionally mirror the broader narrative to a learning file.

Registry write rules:

• persist only reusable issues that should change future review or implementation behavior
• require dedup_key, provenance, pattern, detection_question, checklist_item, applicable_when, and confidence
• applicable_when must use the controlled vocabulary from the finding-registry contract
• append or merge by dedup_key
• use the contract's temp-file-plus-rename atomic write rule

If a broader prose summary still helps, also write the existing anti-pattern learning file to .agents/learnings/YYYY-MM-DD-vibe-<target>.md. Skip both if verdict is PASS.

After the registry update, if hooks/finding-compiler.sh exists, run:

bash hooks/finding-compiler.sh --quiet 2>/dev/null || true

This keeps the same-session post-mortem path synchronized with the latest reusable findings. session-end-maintenance.sh remains the idempotent backstop.

Step 10: Test Bead Cleanup

After validation completes, clean up stale test beads (bd list --status=open | grep -iE "test bead|test quest") via bd close to prevent bead pollution. Skip if bd unavailable.

Integration with Workflow

/implement issue-123
    │
    ▼
(coding, quick lint/test as you go)
    │
    ▼
/vibe                      ← You are here
    │
    ├── Complexity analysis (find hotspots)
    ├── Bug hunt audit (find concrete bugs)
    └── Council validation (multi-model judgment)
    │
    ├── PASS → ship it
    ├── WARN → review, then ship or fix
    └── FAIL → fix, re-run /vibe

Examples

User says: "Run a quick validation on the latest changes."

Do:

/vibe recent

Validate Recent Changes

/vibe recent

Runs complexity on recent changes, then council reviews.

Validate Specific Directory

/vibe src/auth/

Complexity + council on auth directory.

Deep Review

/vibe --deep recent

Complexity + 3 judges for thorough review.

Cross-Vendor Consensus

/vibe --mixed recent

Complexity + Claude + Codex judges.

See references/examples.md for additional examples: security audit with spec compliance, developer-experience code review with PRODUCT.md, and fast inline checks.

Troubleshooting

Write-Time Quality Hook

The hooks/write-time-quality.sh PostToolUse hook runs automatically after every Write/Edit tool call, catching common anti-patterns at edit time rather than review time. It checks:

• Go: unchecked errors, fmt.Print in library code
• Python: bare except:, eval/exec, missing type hints on public functions
• Shell: missing set -euo pipefail, unquoted variables

The hook is non-blocking (always exits 0) and outputs warnings via JSON. See references/write-time-quality.md for the full design.

See Also

• skills/council/SKILL.md — Multi-model validation council
• skills/complexity/SKILL.md — Standalone complexity analysis
• skills/bug-hunt/SKILL.md — Proactive code audit and bug investigation
• .agents/specs/conflict-resolution-algorithm.md — Conflict resolution between agent findings
• test — Test generation and coverage analysis
• perf — Performance profiling and benchmarking

Reference Documents

• references/deep-checks.md
• references/verification-report.md
• references/write-time-quality.md
• references/deep-audit-protocol.md
• references/examples.md
• references/go-patterns.md
• references/go-standards.md
• references/json-standards.md
• references/markdown-standards.md
• references/patterns.md
• references/python-standards.md
• references/report-format.md
• references/rust-standards.md
• references/shell-standards.md
• references/typescript-standards.md
• references/vibe-coding.md
• references/vibe-suppressions.md
• references/test-pyramid-weighting.md
• references/yaml-standards.md