When to invoke
Use this skill when you need to understand what permissions a skill requires before granting trust. Invoke it when:
- Reviewing a new skill before install — a skill has been downloaded or proposed and you need to know what it accesses before allowing execution.
- Checking skill permissions — you want a summary of every permission-relevant signal (binaries, network, env vars, file paths, shell commands, config files, package managers, risky capabilities).
- Generating a permission manifest — you need a human-readable markdown report and a machine-parseable JSON manifest for audit or policy decisions.
- Deciding whether to sandbox or reject a skill — the skill's observed behaviour may exceed its declared capabilities and you need a scored recommendation.
- Auditing a skill's declared vs observed behaviour — you want to compare what SKILL.md claims against what the source code actually does.
If the task does not involve evaluating skill trust, permissions, or install safety, this skill is not the right tool.
How to call — CLI
Run the skill from the command line, passing the skill directory as the sole argument:
npx permission-manifest-guard /path/to/skill
Stdout: The markdown permission manifest (human-readable report). Pipe or redirect as needed.
Stderr: Diagnostic warnings from analysis stages, plus the path to the written JSON file.
Side effect: Writes permission-manifest.json into the skill directory.
This file contains the structured manifest for machine consumption.
Exit code: 0 on success, 1 on invalid arguments or analysis failure.
How to call — programmatic
Import analyzeSkill and call it with the skill directory path:
import { analyzeSkill } from 'permission-manifest-guard';
const result = await analyzeSkill('/path/to/skill');
The returned AnalysisResult has three fields:
| Field | Type | Description |
|---|---|---|
markdownManifest |
string |
Complete markdown permission report. |
jsonManifest |
object |
Structured manifest matching the markdown data. |
diagnostics |
DiagnosticWarning[] |
Warnings from discovery, extraction, or classification stages. Each entry has file, stage, message, and error fields. |
What to do with the output
The analysis produces a disposition recommendation based on how well the skill's declared metadata matches its observed source-code behaviour. Act on the recommendation as follows:
- allow — Declared capabilities match observed behaviour. No undeclared access detected. Install the skill normally.
- review — Minor mismatches found (e.g. an undeclared env var or a low-severity capability gap). Inspect the mismatches in the manifest before proceeding. A human or policy check should confirm intent.
- sandbox — Significant undeclared capabilities detected. Install only within an isolated execution environment with restricted filesystem, network, and process access.
- reject — Severe mismatches or dangerous undeclared capabilities. Do not install. Flag the skill for further investigation or removal from the registry.
The markdown manifest is the artifact for humans — include it in review threads, audit logs, or approval workflows. The JSON manifest is for automated pipelines — feed it to policy engines, dashboards, or gating checks.
If diagnostics is non-empty, some files could not be fully analysed.
Treat a skill with analysis warnings more cautiously — incomplete analysis
means the true permission surface may be larger than reported.
Limitations
- Static analysis only. The skill inspects source code and metadata files. It does not execute the skill or trace runtime behaviour. Dynamically constructed commands, lazy imports, or runtime code generation will not be detected.
- No transitive dependency scanning. Only files within the skill directory are analysed. If the skill depends on external packages that perform privileged operations, those will not appear in the manifest.
- Heuristic extraction. Binary, domain, env-var, and shell-command extraction uses pattern matching. Unusual code patterns may produce false positives or false negatives.
- Single-directory scope. The tool analyses one skill directory per invocation. Multi-skill repositories require separate runs.