guardian-ai

name	guardian-ai
description	Codex-inspired Guardian AI — NEURO-powered safety supervisor that reviews agent actions on EVERY prompt before execution. Uses NEURO API as its brain. Validates code changes, catches security issues, prevents destructive operations. MANDATORY — cannot be skipped.
category	META
chains_with	["neuro","code-hardener","security"]

Guardian AI — NEURO-Powered Safety Supervisor

Inspired by: OpenAI Codex's Guardian AI pattern. Powered by: NEURO API (neurometric/clawpack model). Purpose: Reviews every prompt for safety BEFORE any work begins.

---

Architecture

User Prompt
    ↓
SENSOR GATE (Stage 0-2) — classify intent, select skills
    ↓
GUARDIAN AI (via NEURO API) — safety review ← CANNOT BE SKIPPED
    ↓
  ├─ APPROVED → proceed to skill chain
  ├─ REJECTED → STOP, report to user
  └─ HUMAN_REQUIRED → ask user, wait for approval
    ↓
SKILL CHAIN (Stage 4) — execute selected skills
    ↓
Response

---

How It Works

1. Hard-Block Check (instant, no API)

Patterns that are instantly rejected without NEURO:

• rm -rf / — root filesystem delete
• DROP TABLE — database destruction
• git push --force main — force push to main
• sudo rm — superuser delete

2. Safe Pattern Check (instant, no API)

Patterns that are instantly approved:

• pytest — run tests
• ruff check — lint
• mypy — type check
• git status — check status

3. NEURO API Review (the brain)

For everything else, NEURO evaluates:

• Destructive operations
• Security risks (auth, secrets, credentials)
• Data loss risk
• Architecture violations
• Convention compliance

---

Usage

# Via sensor gate (automatic on every prompt)
python .opencode/skills/chain-orchestrator/scripts/sensor_gate.py --prompt "Fix the bug"

# Direct call
python .opencode/skills/guardian-ai/scripts/guardian_ai.py --prompt "delete all users table"

# JSON output
python .opencode/skills/guardian-ai/scripts/guardian_ai.py --prompt "add rate limiting" --json

---

Risk Levels

Level	Auto-Approve?	Example
low	Yes	Lint fix, comment update, import reorder
medium	Yes (with logging)	Feature addition, refactor, test addition
high	No → Human required	Auth changes, data contract changes, CI config
critical	No → Human required	Secret handling, destructive ops, prod deployment

---

Output Format

[GUARDIAN] Safety Review
- decision: APPROVED | REJECTED | HUMAN_REQUIRED
- risk_level: low | medium | high | critical
- source: neuro | rule_based | hard_block | safe_pattern
- reason: <1-2 sentence explanation>

---

Integration with Skill Chain

exhaustive-crosscheck → neuro → guardian-ai → code-hardener → implementation → lint-fixer

Guardian AI runs before code-hardener, acting as a pre-flight safety check.

---

Configuration

In .opencode/config/opencode.yaml:

guardian_ai:
  enabled: true
  review_before_apply: true
  max_auto_approve_risk: "low"
  require_human_approval:
    - "high"
    - "critical"

---

Logging

All Guardian AI decisions are logged to evolution/guardian_ai.jsonl:

{
  "timestamp": "2026-06-11T16:00:00Z",
  "prompt_excerpt": "fix the login bug",
  "decision": "APPROVED",
  "risk_level": "medium",
  "reason": "Bug fix, non-destructive",
  "source": "neuro"
}