shrimp-skill-install-guard（虾技能安装守卫）

Skill Install Guard｜技能安装守门员

Use this skill when the user wants safer, more consistent skill installation with clear go/no-go decisions and verifiable post-install results.

What it does

This skill enforces a fixed five-step installation guardrail:

1. Source check

- Collects key source trust signals (author, activity, update freshness, public feedback when available). - Marks unavailable data explicitly so risk decisions stay transparent.

1. Local-state check

- Detects existing local installs before action. - Reduces duplicate installs and path mistakes.

1. Code review (MANDATORY)

- Enumerates all files in the target skill. - Reviews readable text files and records binary/oversize/unreadable handling with reasons.

1. Risk review

- Summarizes red flags, required permissions, risk level, and final recommendation. - Produces human-readable conclusions and machine-readable report output.

1. Install execution + post-install verification

- Runs install only when policy allows. - Verifies expected path and required key files after execution.

When to use

• Before installing any unfamiliar skill from ClawHub.
• Before adopting skills from GitHub or local directories.
• When teams need consistent installation standards.
• When security-sensitive environments require auditable evidence.

Primary command

python3 scripts/skill-install-guard.py --slug <skill-slug> [options]

Compatibility wrapper:

scripts/skill-install-guard.sh --slug <skill-slug> [options]

Required / useful inputs

• --slug <slug>: required target skill slug
• --source <source>: optional source (clawhub, local path, or URL)
• --install-cmd '<command>': real install command for execution phase (direct executable invocation only; no shell pipes/redirects/chaining)
• --expected-dir <path>: expected final install path
• --version <version>: optional version hint
• --dry-run: checks only
• --stop-before-install: end after review phase
• --allow-medium-risk: allow execution when risk is medium
• --report-json <path>: write machine-readable result

Recommended operating pattern

1. Collect slug, source, intended install command, and expected final directory.
2. Run with --stop-before-install or --dry-run for non-destructive preflight.
3. Read risk summary and recommendation.
4. If blocked, stop and report why.
5. If allowed, run with actual install command.
6. Return both risk decision and post-install verification result.

Example: non-destructive verification

python3 scripts/skill-install-guard.py \
  --slug some-skill \
  --source clawhub \
  --expected-dir skills/some-skill \
  --stop-before-install \
  --report-json tmp/skill-install-guard/some-skill-verify.json

Example: real guarded install

python3 scripts/skill-install-guard.py \
  --slug some-skill \
  --source clawhub \
  --install-cmd 'clawhub install some-skill' \
  --expected-dir skills/some-skill \
  --report-json tmp/skill-install-guard/some-skill-install.json

Output requirements

When using this skill, report at minimum:

• target skill name / slug
• source checked and source-data completeness
• all-files coverage summary
• red flags found (or explicit none)
• permissions needed (files / network / commands)
• risk level
• recommendation / verdict
• install command used or skipped
• final landed path check
• final go / no-go result

Safety rules

• Do not treat command exit status as sole success criterion.
• Do not skip post-install verification.
• If slug/source/version mismatch expectations, stop before install.
• If risk recommendation is block, do not force install.
• If risk is medium without explicit operator approval, keep blocked.
• When uncertain, prefer no-go.