SkillTrustOps

Review AI agent skills before an agent trusts them.

An agent skill is more than a Markdown file. Its package can contain instructions, scripts, dependencies, assets, archives and permission assumptions. SkillTrustOps treats that package as untrusted input and produces a policy-bound report for local review or CI.

It does not run the skill. It returns stable rule IDs, redacted evidence, clear remediation, and a result that keeps findings separate from scanner errors.

SkillTrustOps supports Python 3.11 and newer.

python -m pip install skilltrustops

skilltrustops policy init --profile recommended-v2
skilltrustops scan .

scan discovers every SKILL.md below the target, inspects each adjacent skill package and applies one repository policy. policy init creates skilltrustops.yaml and never overwrites an existing file.

Use JSON for automation or SARIF for code-scanning platforms:

skilltrustops scan . --format json > skilltrustops.json
skilltrustops scan . --format sarif > skilltrustops.sarif

The default scan runs structure, security and privacy checks. Findings include a severity, source location, redacted evidence and a suggested fix. Reports also record the tool version, ruleset version, effective policy and policy hash.

See the security scan reference for the complete rule list and detection limits.

SkillTrustOps keeps three outcomes separate:

Scanner errors never become passes. Default JSON and SARIF reports omit wall-clock timing so the same input and policy produce stable evidence. Add --metrics when you need timings, or use --benchmark to replay the scan and verify that the evidence is identical.

skilltrustops scan . --benchmark
skilltrustops scan . --metrics

The two modes are intentionally separate: timing is nondeterministic; trust evidence should not be.

The CLI supports the rest of the review workflow without introducing a second control plane:

# Explain one stable rule and include evidence from a report
skilltrustops explain STO-SEC-103 --report skilltrustops.json

# Produce a prioritized Markdown backlog from the same findings
skilltrustops scan . --debt-report engineering-debt.md

# Show which controls have evidence and which were not assessed
skilltrustops certify .

certify produces an evidence matrix, not a blanket certificate. Unsupported controls stay NOT ASSESSED.

For accepted risk, generate a review-required baseline and apply it explicitly:

skilltrustops scan . --write-baseline skilltrustops-suppressions.yaml
skilltrustops scan . --suppressions skilltrustops-suppressions.yaml

Suppressions require a rule ID, path, justification and expiry. Fingerprinted entries apply only to the exact finding that was reviewed.

Static scanning answers what is present in a package. Behavioral testing asks what a selected model attempts when an attack tries to make it reveal data, cross an authorization boundary, skip confirmation or misuse a tool.

Add a reviewed skilltrust-package.yaml beside each skill, then run the offline reference workflow:

skilltrustops scan path/to/skills --redteam

The reference target is a deterministic fixture. It validates manifests, attack assertions, simulated tools, evidence capture and CI wiring without an API key. It does not establish how an unqueried production model will behave.

For a live assessment, use the advanced red-team command with OpenAI or a generic HTTPS model provider:

skilltrustops redteam run path/to/SKILL.md \
  --provider openai \
  --model <approved-model-id>

Live runs send the skill and attack context to the selected provider. Tool calls remain in-memory simulations over synthetic data, so the harness cannot perform the proposed filesystem, network, communication, destructive or financial action.

Every run writes a human-readable report, structured report, event log and integrity manifest. Its decision has deliberately narrow semantics:

passed_scope is evidence for a recorded scope, not a universal safety claim. Never convert inconclusive into a pass.

Read Red-team testing for manifest review, live providers, sandbox choices and evidence handling.

The same deterministic batch report is available as a library:

from skilltrustops import scan

report = scan("path/to/skills", policy_path="skilltrustops.yaml")

for skill in report.skills:
    print(skill.relative_path, skill.status)

Install skilltrustops[observability] to emit a skilltrustops.scan OpenTelemetry span. The library does not configure an exporter or record skill contents, findings, credentials or provider payloads in span attributes.

Package inspection is bounded to 2,000 regular files and 32 MiB of input, with a 1 MiB decoded-text limit per file. Archives are inspected as metadata only; they are never extracted. Links are reported but never followed. Crossing a bound is an error, never a pass.

SkillTrustOps is a pre-trust review gate. It is not a runtime proxy, a malware sandbox, or a replacement for agent permissions and isolation.

The committed benchmark locks 605 public skills from eight repositories by commit and file hash. Whole-package structure, security and privacy scanning completed for all 605 with zero scanner errors and no model or network access.

On the 1 CPU / 512 MiB profile, the median run took 26.031 seconds (23.242 skills/s) with 127.3 MB peak memory. These numbers measure compatibility, throughput and resource use. The corpus has no adjudicated security labels, so it does not establish real-world detection accuracy.

Read the benchmark summary or reproduce the run.

The library's own release path is evidence-backed:

• tests run on Linux, macOS and Windows with Python 3.11–3.13, with an 80% branch-coverage floor;
• Ruff, strict mypy, Bandit, pip-audit, Snyk and CodeQL cover code and locked dependencies;
• wheel and source distributions are built and installed in isolated smoke tests; and
• releases use OIDC Trusted Publishing, an SPDX SBOM and build-provenance attestations.

Each badge applies to a specific commit and workflow run. It does not prove that unknown vulnerabilities do not exist. See Project assurance for the exact gates, versions and evidence locations.

uv sync --locked --extra dev
uv run pytest --cov=skilltrustops --cov-report=term-missing
uv run ruff check .
uv run mypy src

See CONTRIBUTING.md before opening a pull request. Report security issues through SECURITY.md, not a public issue.

MIT. See LICENSE.