| name | super-audit |
| description | Super audit council — three Cursor readonly subagents plus two OpenRouter API legs (Fusion + premium) and one free local Ollama cross-check when available (6-model default). Falls back to 5-model (3 Cursor + 2 OpenRouter) without local. Operator may override routing. Use for super audit, /super-audit, council super audit, pre-ship multi-opinion review. |
| license | MIT |
| metadata.author | cemini23 |
| metadata.version | 1.6.0 |
| disable-model-invocation | true |
Super audit (5–6 model council)
Default stack: 3 Cursor + 2 OpenRouter + 1 local Ollama when OpenRouter key and Ollama are both up. Without local → 3 Cursor + 2 OpenRouter. Operator may override (custom auditors.json, quick mode, SUPER_AUDIT_SKIP_LOCAL=1).
Tier 1 only — all auditors are readonly; they report, they do not edit.
Default auditor layout
| Slot | Channel | Role | Default model |
|---|
| 1–3 | Cursor Task | mode → roles (cursor-audit matrix) | premium Cursor slugs |
| 4 | OpenRouter API | api-adversarial | openrouter/fusion |
| 5 | OpenRouter API | api-deep-reasoning | OPENROUTER_PREMIUM_MODEL → z-ai/glm-5.2 |
| 6 | Ollama (local, free) | api-local-cross-check | best qwen*, else largest *b |
Slot 6 is additive — not a replacement for slots 4–5. Skipped when Ollama is down or SUPER_AUDIT_SKIP_LOCAL=1.
Fallback tiers (discover_api_keys.py)
| Tier | When | Auditors |
|---|
| default-6-model | OpenRouter + Ollama | 3 Cursor + Fusion + premium + local |
| default-5-model | OpenRouter, no Ollama | 3 Cursor + Fusion + premium |
| fallback-local-only | Ollama only (no OR key) | 3 Cursor + 2× local |
| degraded | neither | 3 Cursor only |
Setup
OpenRouter (slots 4–5) — ~/.cemini/llm-routing.env:
OPENROUTER_API_KEY=sk-or-v1-...
OPENROUTER_BASE_URL=https://openrouter.ai/api/v1
OPENROUTER_PREMIUM_MODEL=z-ai/glm-5.2
Ollama (slot 6) — auto-discovered:
ollama serve && ollama pull qwen3.5:32b
Verify:
python3 .cursor/skills/super-audit/scripts/discover_api_keys.py
Operator overrides:
SUPER_AUDIT_SKIP_LOCAL=1 — force 5-model even when Ollama is running- Custom
--auditors path.json — full manual routing
quick mode — 3 Cursor + 1 API only (see mode matrix)- User says "local only" / "no OpenRouter" — tailor per request
Never paste API keys into chat.
When to run
| Signal | Run? |
|---|
| GO/NO-GO with engineering + money on the line | Yes |
| cursor-audit ran but user wants more coverage | Yes |
| Pre-prod deploy | Yes |
| Trivial fix | No — cursor-audit or direct edit |
| User says "quick" | Yes — narrowed scope |
vs cursor-audit
| cursor-audit | super-audit |
|---|
| Auditors | 3 Cursor | 5–6 (3 Cursor + 2–3 API) |
| Cost | 3× subagent | 3× subagent + 2× OR + 0–1 free local |
| Best for | Triage | Pre-ship council |
Read cursor-audit for shared synthesis rules.
Workflow (parent agent)
Super audit progress:
- [ ] 0. Tailor — domain prompt, mission, artifacts
- [ ] 1. Scope — target, question, constraints
- [ ] 2. Mode — Cursor roles + run discover_api_keys.py (expect 5 or 6 auditors)
- [ ] 3. Pack — build_audit_pack.py
- [ ] 4a. Cursor leg — 3× Task in ONE message, readonly
- [ ] 4b. API leg — run_api_auditors.py --discover (2–3 parallel HTTP calls)
- [ ] 5. Synthesize — N-model consensus / conflicts (N = 5 or 6)
- [ ] 6. Act — report only unless user says fix
Announce before dispatch:
Super audit — mode: {mode} · {N} auditors · Cursor: {r1,r2,r3} · API: {slots 4–6 from discover} · pack: {path}
Step 2 — Mode + delegation
- Classify mode (reference.md).
- Run
discover_api_keys.py --json — read auditor_count and tier.
- Respect operator overrides stated in chat before building auditors.
- Cursor slots 1–3: cursor-audit role matrix. API slots 4+: use discovery output unless tailored.
Step 4b — API leg
python3 .cursor/skills/super-audit/scripts/discover_api_keys.py
python3 .cursor/skills/super-audit/scripts/run_api_auditors.py \
--pack reports/audit/pack-{slug} \
--out reports/audit/premium-{slug} \
--discover
Runs all discovered API slots in sequence (Fusion, premium, local when present).
Step 5 — Synthesize
Use auditor_count from discovery (5 or 6). Template:
# Super audit — {target}
**Mode:** {mode} · **Auditors:** {N} · **Tier:** {tier}
| Slot | Channel | Role | Model | Verdict |
|------|---------|------|-------|---------|
| 1–3 | cursor | … | … | … |
| 4 | api | api-adversarial | fusion | … |
| 5 | api | api-deep-reasoning | premium | … |
| 6 | local | api-local-cross-check | qwen… | … | *(if present)* |
## Consensus (≥3) · Strong (≥4, ≥5 if 6-auditor run)
## Unique · Conflicts · Patch backlog · Overall SHIP/REWORK/REJECT
Invocation phrases
/super-audit · super audit on … · council super audit- Optional:
quick, skip local, 5-model only, custom auditors
Cost discipline
- Default 6-model: 3× Cursor + 2× OpenRouter + 1× free local.
- Use quick or SUPER_AUDIT_SKIP_LOCAL=1 to trim cost/time.
- Do not re-run full council on every tiny edit.
Related
