safety-checks（安全检查）

safety-checks (安全)

Unified skill for runtime safety verification including model version pinning, fallback chain validation, cache staleness detection, and cross-session state checks. Consolidates 4 granular skills into a single safety verification suite.

Trigger: 事前検証 (pre-flight) or HEARTBEAT

Source skills: model-pinner, fallback-checker, cache-validator, cross-session-safety-check (from extensions)

Installation

openclaw install leegitw/safety-checks

Dependencies: leegitw/constraint-engine (for enforcement integration)

# Install with dependencies
openclaw install leegitw/context-verifier
openclaw install leegitw/failure-memory
openclaw install leegitw/constraint-engine
openclaw install leegitw/safety-checks

Standalone usage: Model pinning and cache checks work independently. Full integration with constraint enforcement requires constraint-engine.

Data handling: This skill performs local-only operations. All checks (model version comparison, cache age verification, file lock detection) are local file/metadata operations — no data is sent to any model, API, or external service. Results are written to output/safety/ in your workspace.

What This Solves

AI systems can silently degrade — model versions drift, caches go stale, sessions accumulate state. This skill catches these issues before they cause problems:

1. Model pinning — verify you're using the model you expect
2. Fallback validation — ensure degraded-mode paths exist and work
3. Cache checks — detect stale or corrupted cached data
4. Session hygiene — identify cross-session state contamination

The insight: Runtime verification catches what static rules miss. Check the system state, not just the configuration.

Usage

/sc <sub-command> [arguments]

Sub-Commands

Arguments

/sc model

/sc fallback

Note: This command validates that fallback configurations exist in your config file. It does NOT make network calls or test actual connectivity. It's a config file audit.

/sc cache

/sc session

Configuration

Configuration is loaded from (in order of precedence):

1. .openclaw/safety-checks.yaml (OpenClaw standard)
2. .claude/safety-checks.yaml (Claude Code compatibility)
3. Defaults (built-in)

Core Logic

Model Version Pinning

Ensures AI model version matches expected configuration.

How it works: This skill compares the model version reported in your agent's session metadata (e.g., the model name in API responses or agent headers) against the expected version in your config file. It does NOT call the model API to check — it reads the version string that your agent runtime already exposes.

Model version format: {provider}-{model}-{version}-{date}

Examples:

• anthropic-opus-4-5-20251101
• openai-gpt-4-turbo-20250301
• google-gemini-2-pro-20260101

# .openclaw/safety-checks.yaml
model:
  expected: "anthropic-opus-4-5-20251101"  # Provider-neutral format
  strict: true

Fallback Chain Validation

Config audit only — verifies fallback entries exist in your config file. This does NOT make network calls or test actual connectivity.

Checks that your safety-checks.yaml declares fallback configurations:

# Example config entries this command validates exist:
fallbacks:
  model: ["primary-model", "fallback-model", "cached"]
  storage: ["primary-path", "backup-path"]

Cache Staleness Detection

Prevents use of outdated cached data:

Cross-Session State

Detects state leakage between sessions:

Scope: Checks are limited to this skill's own files only. This skill does NOT scan directories, does NOT read other skills' config files, and does NOT access files outside the specific paths listed above.

Output

/sc model output (pinned)

[MODEL CHECK]
Status: ✓ PINNED

Expected: anthropic-opus-4-5-20251101
Actual:   anthropic-opus-4-5-20251101

Model version matches configuration.

/sc model output (drift)

[MODEL CHECK]
Status: ✗ VERSION DRIFT

Expected: anthropic-opus-4-5-20251101
Actual:   anthropic-opus-4-5-20251201

WARNING: Model version has changed.
This may affect behavior consistency.

Action: Update expected version in settings, or investigate change.

/sc fallback output

[FALLBACK CHECK]
Status: ✓ CONFIGURED

Config file: .openclaw/safety-checks.yaml

Fallback entries found:
  ✓ model.fallbacks: 3 entries defined
  ✓ storage.fallbacks: 2 entries defined

Note: This validates config entries exist, not actual connectivity.
  ✓ Tertiary (memory) - 4GB free

/sc cache output

[CACHE CHECK]
Status: ⚠ STALE ENTRIES FOUND

Cache Statistics:
  Total entries: 156
  Fresh (< 1h): 142
  Stale (> 1h): 14
  Critical stale: 0

Stale entries:
  - api_response_auth (age: 2h 15m)
  - user_preferences (age: 1h 30m)
  - ...

Action: Run /sc cache --clear to remove stale entries.

/sc session output (clean)

[SESSION CHECK]
Status: ✓ CLEAN

No cross-session interference detected.

Checks passed:
  ✓ No stale lock file (.openclaw/safety-checks.lock)
  ✓ No orphan temp files (output/safety/temp-*)
  ✓ Config file valid (.openclaw/safety-checks.yaml)

/sc session output (interference)

[SESSION CHECK]
Status: ✗ INTERFERENCE DETECTED

Issues found:

1. Stale file lock:
   File: .openclaw/safety-checks.lock
   Owner: PID 12345 (not running)
   Action: Remove lock with /sc session --clear-state

2. Orphan temp files (3):
   output/safety/temp-*.log
   Age: > 24 hours
   Action: Remove with /sc session --clear-state

Run /sc session --clear-state to remediate.

Integration

• Layer: Safety
• Depends on: constraint-engine (for enforcement integration)
• Used by: constraint-engine (for pre-action safety checks), governance (for health monitoring)

Failure Modes

Next Steps

After invoking this skill:

Workspace Files

This skill reads/writes only within declared paths:

.openclaw/
├── safety-checks.yaml     # Primary config (read)
└── cache/
    └── staleness.log      # Cache check history (read/write)

.claude/
└── safety-checks.yaml     # Claude Code config (read)

output/
└── safety/
    ├── model-checks.log   # Model version history (write)
    ├── fallback-tests.log # Fallback test results (write)
    └── session-state.log  # Session state log (write)

Note: This skill does NOT read .claude/settings.json or any other tool's configuration.

HEARTBEAT Integration

These checks should run periodically via HEARTBEAT:

## P1: Critical (Every Session)
- [ ] Model version pinned? → /sc model
- [ ] Session state clean? → /sc session

## P2: Important (Weekly)
- [ ] Fallback chains healthy? → /sc fallback
- [ ] Cache fresh? → /sc cache

Examples

Model Drift Detection

/sc model --strict
[MODEL CHECK]
Status: ⚠ VERSION DRIFT

Expected: anthropic-opus-4-5-20251101
Actual:   anthropic-opus-4-5-20260101

Model version changed. Review CHANGELOG for behavior changes.

Cache Cleanup

/sc cache --clear
[CACHE CHECK]
Status: ✓ CLEANED

Removed 14 stale entries from .openclaw/cache/
Freed: 2.3 MB

Security Considerations

Local-only processing: All safety checks are local file and metadata operations. No data is sent to any LLM, API, or external service. The "agent's model" is only used to interpret your commands — not to process or transmit your data.

Exact files accessed (read):

• .openclaw/safety-checks.yaml — your skill configuration
• .claude/safety-checks.yaml — alternate config location
• .openclaw/cache/staleness.log — cache check history (if exists)

Exact files written:

• output/safety/*.log — check results and history

This skill does NOT scan directories or read other skills' config files.

What this skill does NOT access:

• Other skills' configuration files
• System environment variables
• .claude/settings.json or other tool configs
• Any files outside the three config paths listed above
• Network resources or external APIs

What this skill does NOT do:

• Send data to any model or external service
• Modify files outside output/safety/
• Execute arbitrary code

Data handling:

• All checks are read-only except for --clear operations on its own cache/output
• Results are written to output/safety/ only
• No data leaves the local machine

Model version clarification: The /sc model command compares the expected version in your config against the model version string reported by your agent runtime (e.g., in session metadata or API headers). It does NOT make API calls to verify the model — it reads information your agent already has.

Provenance note: This skill is developed by Live Neon (https://github.com/live-neon/skills) and published to ClawHub under the leegitw account. Both refer to the same maintainer.

Acceptance Criteria

• [ ] /sc model verifies model version matches config
• [ ] /sc model warns or fails on version drift based on strict setting
• [ ] /sc fallback verifies fallback config entries exist in config file
• [ ] /sc cache detects entries older than TTL
• [ ] /sc cache --clear removes stale entries
• [ ] /sc session detects cross-session state leakage
• [ ] /sc session --clear-state remediates found issues
• [ ] All checks integrated with HEARTBEAT

*Consolidated from 4 skills as part of agentic skills consolidation (2026-02-15).*