开爪守护者

说明文档

# OpenClaw Warden

Monitors your workspace files for unauthorized modifications and prompt injection attacks. Existing security tools scan *skills* before installation — this tool watches the *workspace itself* after installation, catching tampering that other tools miss.

## Why This Matters

Your agent reads SOUL.md, AGENTS.md, IDENTITY.md, USER.md, and memory files on every session startup and **trusts them implicitly**. A compromised skill, a malicious heartbeat payload, or an unauthorized process can modify these files to:

- Inject hidden instructions that alter agent behavior
- Embed data exfiltration URLs in markdown images
- Override identity and safety boundaries
- Plant persistent backdoors in memory files

This skill detects all of these.


## Commands

### Establish Baseline

Create or reset the integrity baseline. Run this after setting up your workspace or after reviewing and accepting all current file states.

```bash
python3 {baseDir}/scripts/integrity.py baseline --workspace /path/to/workspace
```

### Verify Integrity

Check all monitored files against the stored baseline. Reports modifications, deletions, and new untracked files.

```bash
python3 {baseDir}/scripts/integrity.py verify --workspace /path/to/workspace
```

### Scan for Injections

Scan workspace files for prompt injection patterns: hidden instructions, base64 payloads, Unicode tricks, markdown image exfiltration, HTML injection, and suspicious system prompt markers.

```bash
python3 {baseDir}/scripts/integrity.py scan --workspace /path/to/workspace
```

### Full Check (Verify + Scan)

Run both integrity verification and injection scanning in one pass.

```bash
python3 {baseDir}/scripts/integrity.py full --workspace /path/to/workspace
```

### Quick Status

One-line summary of workspace health.

```bash
python3 {baseDir}/scripts/integrity.py status --workspace /path/to/workspace
```

### Accept Changes

After reviewing a legitimate change, update the baseline for a specific file.

```bash
python3 {baseDir}/scripts/integrity.py accept SOUL.md --workspace /path/to/workspace
```

## Workspace Auto-Detection

If `--workspace` is omitted, the script tries:
1. `OPENCLAW_WORKSPACE` environment variable
2. Current directory (if AGENTS.md exists)
3. `~/.openclaw/workspace` (default)

## What Gets Monitored

| Category | Files | Alert Level on Change |
|----------|-------|-----------------------|
| **Critical** | SOUL.md, AGENTS.md, IDENTITY.md, USER.md, TOOLS.md, HEARTBEAT.md | WARNING |
| **Memory** | memory/*.md, MEMORY.md | INFO (expected to change) |
| **Config** | *.json in workspace root | WARNING |
| **Skills** | skills/*/SKILL.md | WARNING |

Injection patterns trigger **CRITICAL** alerts regardless of file category.

## Injection Patterns Detected

- **Instruction override:** "ignore previous instructions", "disregard above", "you are now", "new system prompt"
- **Base64 payloads:** Suspiciously long base64 strings outside code blocks
- **Unicode manipulation:** Zero-width characters, RTL overrides, homoglyphs
- **Markdown exfiltration:** Image tags with data-encoding URLs
- **HTML injection:** script tags, iframes, hidden elements
- **System prompt markers:** `<system>`, `[SYSTEM]`, `<<SYS>>` blocks
- **Shell injection:** `$(...)` outside code blocks

## Exit Codes

- `0` — Clean, no issues
- `1` — Modifications detected (review needed)
- `2` — Injection patterns detected (action needed)

## No External Dependencies

Python standard library only. No pip install. No network calls. Everything runs locally.

## Cross-Platform

Works with OpenClaw, Claude Code, Cursor, and any tool using the Agent Skills specification.