---
name: yamtam
version: "0.40.0"
description: "Sovereign-grade safety OS for AI coding agents. 45 hooks, 3,440 skills, L1 memory, circuit breakers, and cross-engine enforcement — blocks rm -rf, force push, pipe-to-shell, and 40+ attack vectors before they reach your repo."
argument-hint: "yamtam status | yamtam audit | yamtam hooks | yamtam memory"
allowed-tools: Bash, Read, Write
homepage: https://phamlongh230-lgtm.github.io/yamtam-engine/
repository: https://github.com/phamlongh230-lgtm/yamtam-engine
author: phamlongh230-lgtm
license: Apache-2.0
user-invocable: true
metadata:
  openclaw:
    emoji: "🛡️"
    requires:
      env: []
      optionalEnv: []
      bins:
        - bash
        - python3
        - jq
        - openssl
    homepage: https://phamlongh230-lgtm.github.io/yamtam-engine/
    tags:
      - safety
      - hooks
      - memory
      - agents
      - gates
      - guards
      - circuit-breaker
      - sovereign
      - claude-code
      - codex
      - cursor
      - audit
      - ai-agent
      - agent-os
---

# YAMTAM ENGINE — Sovereign Safety OS

YAMTAM ENGINE is an agent operating system layered on top of Claude Code (and other AI coding agents). It intercepts dangerous commands, enforces safety rules, manages persistent memory, and provides 3,440+ reusable skills.

## What YAMTAM does

When loaded, YAMTAM automatically:

- **Blocks 40+ attack vectors** via 45 pre/post hooks — `rm -rf`, force push, pipe-to-shell, path traversal, supply chain attacks, and more
- **Enforces rules** across all agent actions using a 9-layer middleware pipeline (injection scan → blast radius → permission check → egress → sanitize → PII scrub → size cap → audit log)
- **Persists memory** in a 2-tier system: L1 Atomic (permanent, hash-chained) and L2 session facts
- **Provides 3,440 skills** covering frontend, backend, IaC, AI/LLM, security, multi-agent, K8s, WebAssembly, and more

## Install

```bash
# Install via skills CLI
npx skills add phamlongh230-lgtm/yamtam-engine

# Or install directly into .claude/
curl -L https://github.com/phamlongh230-lgtm/yamtam-engine/releases/latest/download/yamtam-engine-latest.zip -o yamtam.zip
unzip yamtam.zip -d .claude/
```

## Quick commands

| Command | What it does |
|---------|-------------|
| `/status` | Show hooks, memory, and gate status |
| `/audit` | Run full security audit on repo |
| `/quick-commit` | Safe commit with gate checks |
| `/session-wrap` | Persist session state to L1 memory |
| `/smart-fix` | Auto-fix with feedback loop |
| `/code-review` | Multi-agent code review |

## Safety hooks overview

YAMTAM hooks fire on every tool call:

```
PreToolUse:
  - safe-run.sh         — blocks 40+ dangerous command patterns
  - scope-guard.sh      — enforces declared file scope
  - risk-scorer.sh      — rates action risk 0–100
  - token-budget-guard  — prevents runaway token usage

PostToolUse:
  - session-checkpoint  — saves state every 5 tool calls
  - audit-logger        — appends to Merkle hash-chain log

Stop:
  - truth-gate.sh       — warns on completion claims without evidence
```

## Memory system

```bash
# Save a fact to L1 (persists across sessions)
bash core/scripts/add-fact.sh "tag" "fact content" "high"

# Search L1 memory
bash core/scripts/search-facts.sh "tag"

# View session memory
bash core/scripts/list-session-facts.sh
```

## Hard-blocked commands

YAMTAM will always block these regardless of context:

```
rm -rf          git push --force     eval "$user_input"
curl ... | bash  DROP TABLE           chmod 777 core/
```

## Cross-engine enforcement

YAMTAM enforces safety on all AI coding agents via a shared `safe-run.sh` proxy:

- **Claude Code** — native hooks
- **Codex** — `CODEX_SHELL_EXEC_HOOK` → safe-run.sh
- **Cursor** — `.cursor/rules/` enforcement
- **Aider** — `--before-exec` hook