Skill Safe Install (L0 Strict)

Enforce a conservative, auditable install workflow.

Purpose

Use this skill to reduce accidental or risky third-party skill installs:

• - Force risk review before installation.
• Require sandbox verification before formal install.
• Require explicit user confirmation before sensitive actions.
• Avoid hidden trust escalation (no author-based bypass, no implicit allowBundled writes).

Non-negotiable rules

1. 1. Never skip steps.
2. Never auto-trust by author, popularity, or “official-looking” name.
3. Never modify persistent config (openclaw.json) without explicit user consent in the current conversation.
4. If risk cannot be evaluated, treat as high risk and pause.

Workflow (Step 0 → Step 6)

Step 0 — Confirm target

• - Resolve exact skill slug and (if available) version.
• If input is ambiguous, ask for confirmation before install.

Suggested checks:

• - INLINECODE1
• Verify exact slug/version from results.

Step 1 — Duplicate/state check

• - Check whether the skill is already installed.
• Check current trust state (whether already in skills.allowBundled).

Suggested checks:

• - INLINECODE3
• Read ~/.openclaw/openclaw.json (or platform-equivalent config path)

Step 2 — Mandatory security review (no whitelist bypass)

Run inspect and summarize at least:

1. 1. Maintainer/source and recent update signal
2. Required secrets/credentials (API keys, OAuth, tokens)
3. Network/system access scope
4. Command execution or file-system mutation risk
5. Persistence behavior (config edits, auto-run, always-on behavior)

Suggested check:

• - INLINECODE5

Risk rating rubric

• - LOW: Text/process guidance only, no credentials, no system mutation.
• MEDIUM: Requires limited credentials or external API access with clear scope.
• HIGH: Broad command execution, config mutation, or multi-system OAuth.
• CRITICAL: Destructive capability, privilege escalation, stealth persistence, or unclear behavior.

Gate policy

• - LOW / MEDIUM: Continue to sandbox.
• HIGH: Continue only after explicit confirmation.
• CRITICAL: Do not install by default; require explicit override and warn strongly.

Step 3 — Sandbox install (isolated workdir)

Install in a temporary isolated directory first.

• - Use isolated workdir (do not install to primary skill directory yet).
• Confirm install result and basic behavior.
• If sandbox fails, stop.

Example pattern:

• - INLINECODE6

Step 4 — User confirmation checkpoint

Before formal install, present:

• - Chosen skill slug/version
• Risk rating + top risks
• Sandbox result
• Exact next action

Proceed only after explicit “yes/install/继续”.

Step 5 — Formal install

Run formal install only after Step 4 consent.

Example:

• - INLINECODE7

If install fails, stop and report error + rollback advice.

Step 6 — Optional trust persistence (allowBundled)

Default is do not write trust list.

Only perform this step when user explicitly asks to persist trust.

Required safeguards:

1. 1. Backup config with timestamp.
2. Show exactly what key will change (skills.allowBundled).
3. Append skill slug only if absent (idempotent).
4. Confirm backup path and rollback command.

Do not use hidden or implicit trust writes.

Output format (required)

• - INLINECODE10
• INLINECODE11
• INLINECODE12
• INLINECODE13
• INLINECODE14
• INLINECODE15
• INLINECODE16

Refusal conditions

Stop and ask for confirmation/override when any condition is met:

• - Skill identity is ambiguous.
• Inspect output is unavailable or incomplete.
• Risk is HIGH/CRITICAL and user has not explicitly approved.
• Requested config mutation lacks explicit consent.