Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Generate and verify integrity manifests for AI agent plugins and tools. Detect tampering, enforce version pinning, and establish supply chain provenance.
.claude/skills/agent-supply-chain/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-09 | ✗→✓ | ▲ Improved | — | — |
| case-14 | ✗→✓ | ▲ Improved | — | — |
| case-05 | ✗→✓ | ▲ Improved | — | — |
| case-03 | ✗→✓ | ▲ Improved | — | — |
| case-16 | ✗→✓ | ▲ Improved | — | — |
Generate and verify integrity manifests for AI agent plugins and tools. Detect tampering, enforce version pinning, and establish supply chain provenance.
Agent plugins and MCP servers have the same supply chain risks as npm packages or container images — except the ecosystem has no equivalent of npm provenance, Sigstore, or SLSA. This skill fills that gap.
Plugin Directory → Hash All Files (SHA-256) → Generate INTEGRITY.json
↓
Later: Plugin Directory → Re-Hash Files → Compare Against INTEGRITY.json
↓
Match? VERIFIED : TAMPEREDCreate a deterministic INTEGRITY.json with SHA-256 hashes of all plugin files.
pythonimport hashlib import json from datetime import datetime, timezone from pathlib import Path EXCLUDE_DIRS = {".git", "__pycache__", "node_modules", ".venv", ".pytest_cache"} EXCLUDE_FILES = {".DS_Store", "Thumbs.db", "INTEGRITY.json"} def hash_file(path: Path) -> str: """Compute SHA-256 hex digest of a file.""" h = hashlib.sha256() with open(path, "rb") as f: for chunk in iter(lambda: f.read(8192), b""): h.update(chunk) return h.hexdigest() def generate_manifest(plugin_dir: str) -> dict: """Generate an integrity manifest for a plugin directory.""" root = Path(plugin_dir) files = {} for path in sorted(root.rglob("*")): if not path.is_file(): continue if path.name in EXCLUDE_FILES: continue if any(part in EXCLUDE_DIRS for part in path.relative_to(root).parts): continue rel = path.relative_to(root).as_posix() files[rel] = hash_file(path) # Chain hash: SHA-256 of all file hashes concatenated in sorted order chain = hashlib.sha256() for key in sorted(files.keys()): chain.update(files[key].encode("ascii")) manifest = { "plugin_name": root.name, "generated_at": datetime.now(timezone.utc).isoformat(), "algorithm": "sha256", "file_count": len(files), "files": files, "manifest_hash": chain.hexdigest(), } return manifest # Generate and save manifest = generate_manifest("my-plugin/") Path("my-plugin/INTEGRITY.json").write_text( json.dumps(manifest, indent=2) + "\n" ) print(f"Generated manifest: {manifest['file_count']} files, " f"hash: {manifest['manifest_hash'][:16]}...")
Output (INTEGRITY.json):
json{ "plugin_name": "my-plugin", "generated_at": "2026-04-01T03:00:00+00:00", "algorithm": "sha256", "file_count": 12, "files": { ".claude-plugin/plugin.json": "a1b2c3d4...", "README.md": "e5f6a7b8...", "skills/search/SKILL.md": "c9d0e1f2...", "agency.json": "3a4b5c6d..." }, "manifest_hash": "7e8f9a0b1c2d3e4f..." }
Check that current files match the manifest.
python# Requires: hash_file() and generate_manifest() from Pattern 1 above import json from pathlib import Path def verify_manifest(plugin_dir: str) -> tuple[bool, list[str]]: """Verify plugin files against INTEGRITY.json.""" root = Path(plugin_dir) manifest_path = root / "INTEGRITY.json" if not manifest_path.exists(): return False, ["INTEGRITY.json not found"] manifest = json.loads(manifest_path.read_text()) recorded = manifest.get("files", {}) errors = [] # Check recorded files for rel_path, expected_hash in recorded.items(): full = root / rel_path if not full.exists(): errors.append(f"MISSING: {rel_path}") continue actual = hash_file(full) if actual != expected_hash: errors.append(f"MODIFIED: {rel_path}") # Check for new untracked files current = generate_manifest(plugin_dir) for rel_path in current["files"]: if rel_path not in recorded: errors.append(f"UNTRACKED: {rel_path}") return len(errors) == 0, errors # Verify passed, errors = verify_manifest("my-plugin/") if passed: print("VERIFIED: All files match manifest") else: print(f"FAILED: {len(errors)} issue(s)") for e in errors: print(f" {e}")
Output on tampered plugin:
FAILED: 3 issue(s)
MODIFIED: skills/search/SKILL.md
MISSING: agency.json
UNTRACKED: backdoor.pyCheck that agent dependencies use pinned versions.
pythonimport re def audit_versions(config_path: str) -> list[dict]: """Audit dependency version pinning in a config file.""" findings = [] path = Path(config_path) content = path.read_text() if path.name == "package.json": data = json.loads(content) for section in ("dependencies", "devDependencies"): for pkg, ver in data.get(section, {}).items(): if ver.startswith("^") or ver.startswith("~") or ver == "*" or ver == "latest": findings.append({ "package": pkg, "version": ver, "severity": "HIGH" if ver in ("*", "latest") else "MEDIUM", "fix": f'Pin to exact: "{pkg}": "{ver.lstrip("^~")}"' }) elif path.name in ("requirements.txt", "pyproject.toml"): for line in content.splitlines(): line = line.strip() if ">=" in line and "<" not in line: findings.append({ "package": line.split(">=")[0].strip(), "version": line, "severity": "MEDIUM", "fix": f"Add upper bound: {line},<next_major" }) return findings
Use integrity verification as a gate before promoting plugins.
pythondef promotion_check(plugin_dir: str) -> dict: """Check if a plugin is ready for production promotion.""" checks = {} # 1. Integrity manifest exists and verifies passed, errors = verify_manifest(plugin_dir) checks["integrity"] = { "passed": passed, "errors": errors } # 2. Required files exist root = Path(plugin_dir) required = ["README.md"] missing = [f for f in required if not (root / f).exists()] # Require at least one plugin manifest (supports both layouts) manifest_paths = [ root / ".github/plugin/plugin.json", root / ".claude-plugin/plugin.json", ] if not any(p.exists() for p in manifest_paths): missing.append(".github/plugin/plugin.json (or .claude-plugin/plugin.json)") checks["required_files"] = { "passed": len(missing) == 0, "missing": missing } # 3. No unpinned dependencies mcp_path = root / ".mcp.json" if mcp_path.exists(): config = json.loads(mcp_path.read_text()) unpinned = [] for server in config.get("mcpServers", {}).values(): if isinstance(server, dict): for arg in server.get("args", []): if isinstance(arg, str) and "@latest" in arg: unpinned.append(arg) checks["pinned_deps"] = { "passed": len(unpinned) == 0, "unpinned": unpinned } # Overall all_passed = all(c["passed"] for c in checks.values()) return {"ready": all_passed, "checks": checks} result = promotion_check("my-plugin/") if result["ready"]: print("Plugin is ready for production promotion") else: print("Plugin NOT ready:") for name, check in result["checks"].items(): if not check["passed"]: print(f" FAILED: {name}")
Add to your GitHub Actions workflow:
yaml- name: Verify plugin integrity run: | PLUGIN_DIR="${{ matrix.plugin || '.' }}" cd "$PLUGIN_DIR" python -c " from pathlib import Path import json, hashlib, sys def hash_file(p): h = hashlib.sha256() with open(p, 'rb') as f: for c in iter(lambda: f.read(8192), b''): h.update(c) return h.hexdigest() manifest = json.loads(Path('INTEGRITY.json').read_text()) errors = [] for rel, expected in manifest['files'].items(): p = Path(rel) if not p.exists(): errors.append(f'MISSING: {rel}') elif hash_file(p) != expected: errors.append(f'MODIFIED: {rel}') if errors: for e in errors: print(f'::error::{e}') sys.exit(1) print(f'Verified {len(manifest[\"files\"])} files') "
| Practice | Rationale | |----------|-----------| | Generate manifest after code review | Ensures reviewed code matches production code | | Include manifest in the PR | Reviewers can verify what was hashed | | Verify in CI before deploy | Catches post-review modifications | | Chain hash for tamper evidence | Single hash represents entire plugin state | | Exclude build artifacts | Only hash source files — .git, __pycache__, node_modules excluded | | Pin all dependency versions | Unpinned deps = different code on every install |
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-10 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-09 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-14 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-21 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-05 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-13 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-19 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-17 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-03 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-16 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-12 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-08 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-11 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-20 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-02 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-01 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-07 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-04 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-06 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-18 | fail→fail | — | — | — | — | — | — | — | — | — | — | — | — |
case-15 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
case-22 | fail→pass | — | — | — | — | — | — | — | — | — | — | — | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 21 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +50 percentage points is the difference between those two pass rates over the 21 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
The per-case answers from this run were removed by the retention sweep, so the case table below shows the verdicts without the text either arm produced. The counts above were recorded at the time and are unaffected. Answers are now kept for 180 days.
Other measured skills in the registry, with their headline benchmark lift.