Español · Map · Home

• The root dependency installation audited 54 packages and reported zero vulnerabilities. The full npm audit --json report is retained in local release verification output; production and development dependencies are included.
• GitHub Dependabot open alerts were queried with state=open, per_page=100 and pagination: none returned. Secret-scanning and code-scanning alert APIs likewise returned no open findings in the observed repository response. Empty provider responses are snapshots, not a security certificate or proof that every possible scanner ran.
• The repository content scanner passed, covering publishable text and retaining the exclusions documented in the security model.
• The generated OpenAI package validator passed for all 413 canonical source skills at 0.10.0. It validates local shape, references, paths, limits and reproducibility, not live listing URLs, publisher verification or public approval.
• Native exports exercised all ten layouts, conflict refusal and source-path boundaries. Full export now handles bilingual reader navigation without copying operational source or adjacent credentials. Repository-only documents become pinned release links when absent from npm.
• Filtered Angular installation uses the resolver's actual included skills, not all skills from its unfiltered source packs. Regression evidence covers Angular 14.1 exclusions; snapshot resolution shares compatibility/range checks and rejects caller path fields.

• PR #68's candidate 45eb9a1c passed all twelve reported PR checks, including CodeQL, both protected release gates and Socket. Remote release validation passed 239 repository tests and 22 script tests; Skill Lab reported 96 tests with zero failures. This is candidate evidence, not approval of a later commit.
• CodeQL initially found incomplete multi-character sanitization in documentation heading processing. The fix uses an explicit tag-state scan and an emitted-character allowlist, with malformed-tag regression coverage. The latest CodeQL run passed; the original review thread is resolved and outdated. Fresh paginated Dependabot, secret-scanning and code-scanning open-alert responses were empty.
• GitHub's separate Code scanning AI findings run failed before review because its monthly quota was exhausted. CodeQL success does not imply this optional AI review completed. Restore the owner's quota and rerun if that review is required; do not bypass a required check.
• An actual Codex 0.156.1 invocation rejected the configured gpt-6.1-sol model for the signed-in ChatGPT account. An isolated Claude Code 2.1.226 invocation produced no result within the bounded observation window and was stopped. Neither attempt proves Components/Skills discovery or successful invocation. No configured model/provider or global settings were replaced. Raw host logs remain private local evidence.

Python dependency evidence is separate from npm: an isolated Python 3.12.7 audit with pip-audit 2.10.1 resolved 38 Skill Lab dependencies and reported no known vulnerabilities on 2026-10-02. It used PyPI advisories, strict collection and wheel-only resolution without installing Skill Lab or using provider credentials. This is an unlocked, platform-specific resolution snapshot, not proof about every allowed older version or a Python runtime evaluation. Protected release gates now repeat this audit; publication inventory separately records Python source hashes and the explicit INTERNAL_NOT_PYPI state.

To repeat it, activate an isolated Python 3.12 environment, install python -m pip install --only-binary=:all: pip-audit==2.10.1, then run python scripts/audit-python-dependencies.py. The JSON report is written to dist/security/python-dependency-audit.json; nonzero audit results block the release. The protected workflow attaches this report to the GitHub release. The auditor is a validation tool, not a new Skill Lab runtime dependency.

The subsequent PR review findings were reproduced and covered by focused regressions: source-free standalone MCP resolution, compound toolchain ranges, explicit CLI approval (including install and uninstall), package-root checkpoints, owning-workspace lockfile boundaries, peer-only Angular work plans, bounded rate-limit retention and the checkpoint execution contract. semver 7.8.5 is pinned as a required runtime dependency to use npm's maintained range grammar rather than extending the intentionally restricted Angular parser. The focused tests and final current-head CI must pass before closing those review threads.

• Run the final current-head release and Skill Lab gates; record remote CI independently from local results.
• Inspect the exact-commit security packet and obtain human PR/environment approval. The packet remains NOT_APPROVED; no external Sage execution or SkillSpector audit is claimed.
• Configure the protected npm credential and rotate/revoke the historical repository-wide token. An empty release-security secret-name response is a release blocker, not permission to reuse a less protected credential.
• Verify the exact npm tarball inventory, every archive hash and the published version/latest after the protected workflow.
• Recheck third-party skill-directory audit provider, status, auditedAt and source snapshot after publication. A local source fix cannot force an external indexer to refresh.

Later regressions also cover HTTPS omitted-target inference, peer-only MCP/HTTPS snapshots, each supplied Angular lock entry, indented Markdown fences and stale generated documentation indexes. The maintainer guide now explicitly warns that protected approval of a GitHub release event authorizes npm publication. Python audit setup installs the frontmatter validator's PyYAML dependency explicitly; it does not rely on packages preinstalled in a runner's former Python runtime.

No universal real-host certification, full-catalog semantic score, external vulnerability certification, portal attestation, marketplace approval or global hook enforcement is claimed. Local fixtures and historical audit reports retain their original scope. Public update does not grant an agent permission to run instructions, replace user edits or upload private context.

See the ordered publication tasks for owner actions and each target's closing evidence.