26 KiB

Raw Blame History

Full repository audit - 2026-05-23

Status: complete

This audit tracks the current deep pass over the repository for bugs, inconsistencies, drift, documentation issues, and release/process risks. The working tree was checked from /Users/nicco/Projects/antigravity-awesome-skills on main.

Patch follow-up:

Resolved in follow-up patches: unsafe archive fallback extraction, Telegram Node vulnerable dependency stack, installer symlink target migration, stale web SEO counts/social card, stale web canonical URL docs, malformed WhatsApp HMAC signature handling, Telegram token-in-URL webhook guidance, .disabled web asset publishing, Junta TLS bypasses, legacy manifest verification drift, Telegram HTML escaping, Remotion chart typo, nested skill ID collision coverage, Chinese/localized docs staleness, path-aware internal markdown link repair, and deterministic link/glossary validation reports.
Still open after these patches: none from this audit report. The legacy strict skill-quality warnings remain tracked as explicit backlog debt behind tools/config/audit-skills-strict-budget.json.

Validation evidence

Passed:

npm run validate
npm run validate:strict
npm run validate:references
npm run plugin-compat:check
npm run check:warning-budget
npm run audit:skills
npm run audit:consistency
npm run audit:maintainer reported version/count/warning/consistency clean, then exited non-zero because this audit report itself is an untracked changed file
npm run check:stale-claims
npm run bundles:check
npm run check:readme-credits
npm run security:docs
npm test
npm run test:local
npm run pr:preflight
npm run app:test
npm run app:test:coverage
npm run app:build
cd apps/web-app && npm run lint
cd apps/web-app && npm run verify:seo
npm audit --audit-level=high --omit=dev
npm audit --audit-level=moderate
cd apps/web-app && npm audit --audit-level=moderate
cd data && npm audit --audit-level=moderate
cd skills/loki-mode/examples/todo-app-generated/backend && npm audit --audit-level=moderate
cd skills/loki-mode/examples/todo-app-generated/frontend && npm audit --audit-level=moderate
cd skills/loki-mode/examples/todo-app-generated/backend && npm run build
cd skills/loki-mode/examples/todo-app-generated/frontend && npm run build
Temp-copy npm run build for skills/telegram/assets/boilerplate/nodejs
Temp-copy npm run build for skills/whatsapp-cloud-api/assets/boilerplate/nodejs
Temp-copy npm audit --audit-level=moderate for skills/whatsapp-cloud-api/assets/boilerplate/nodejs
npm pack --dry-run --json
bash scripts/validate-links.sh
python3 -m py_compile equivalent pass over tracked Python sources, excluding generated mirrors/output
bash -n pass over tracked shell scripts, excluding generated mirrors/output
node --check over tracked JavaScript/CommonJS/ESM source files, excluding generated web output and plugin mirrors
JSON/JSONL parse over tracked JSON-like files
Ruby/Psych YAML parse over .github/workflows/*.yml, discussion templates, and issue templates
Ruby/Psych YAML parse over all tracked *.yml and *.yaml
Python XML parser over tracked XML/XSD/SVG files
actionlint v1.7.12 run from /tmp over .github/workflows/*.yml
Global frontmatter parse over all tracked *SKILL.md files
Full-plugin and bundle-plugin file drift checks against canonical skills/ sources
Plugin manifest parse/check over .codex-plugin/plugin.json and .claude-plugin/plugin.json
Canonical skills_index.json ID/path duplicate check
Targeted typo/refuso scan for common English misspellings across active repo sources, excluding generated output and plugin mirrors
Tracked temporary/bytecode artifact scan for __pycache__, *.pyc, .DS_Store, *.tmp, *.bak, and *.orig

Observed counts:

Root visible skills: 1465
Tracked *SKILL.md files: 4559
Canonical indexed skills: 1465 files, 0 duplicate index IDs, 0 duplicate index paths
Canonical skill frontmatter: 0 duplicate names, 0 missing risk/source, 229 missing date_added
Nested indexed skills: 20 entries under second-level category paths such as skills/game-development/2d-games, skills/libreoffice/base, and skills/security/aws-iam-best-practices
Root skills_index.json and data/skills_index.json: equal
audit:skills: 1465 scanned, 0 errors, 740 warning-only
App coverage: 12 test files / 72 tests passed, 89.58% statements, 76.84% branches
Web app lint: passed with --max-warnings 0
Web app SEO verifier: passed, with the manifest coverage caveat noted below
Python syntax pass: 1077 files checked, 0 errors
Shell syntax pass: 61 scripts checked, 0 errors
JavaScript syntax pass: 74 source files checked, 0 errors
YAML parse pass: 81 files checked, 0 errors
XML/XSD/SVG parse pass: 342 files checked, 0 errors
Tracked symlinks: 7 checked, all targets exist
Plugin mirror SKILL.md files: 3094 matched canonical sources byte-for-byte, 0 drift
Global tracked skill frontmatter: 4559 SKILL.md files parsed, 0 frontmatter errors, 0 missing frontmatter, 0 missing name
Disabled source skills: 34 skills/.disabled/*/SKILL.md files present and included in the global frontmatter parse
Full plugin file drift: 11561 files checked in plugins/antigravity-awesome-skills*, 0 missing canonical files, 0 drift
Bundle plugin file drift: 950 files checked in plugins/antigravity-bundle-*, 0 missing canonical files, 0 drift
Plugin manifests: 76 manifests parsed, 0 missing name/version/reference errors
Maintainer audit: repository sickn33/antigravity-awesome-skills, version 11.5.0, skills 1,465+, warning budget 0/135, consistency clean; tracked working tree clean, but the untracked audit report makes the full gate fail until added/removed
Stale-claims checker: no stale claims detected in active docs
Editorial bundles: manifest and generated doc are in sync
README credits check: no changed skill files detected
Telegram and WhatsApp TypeScript boilerplates build successfully in temp copies after npm install --ignore-scripts
WhatsApp TypeScript boilerplate temp-copy moderate audit: 0 vulnerabilities

Failed / issue-producing checks:

Temp-copy npm audit --audit-level=moderate for skills/telegram/assets/boilerplate/nodejs: 5 vulnerabilities, including 1 critical (form-data <2.5.4) through node-telegram-bot-api -> @cypress/request-promise -> request
npm run audit:skills:strict: 1465 scanned, 0 errors, but 740 warning-only skills / 865 warnings made the strict audit exit non-zero
Strict JSON parsing found 5 JSONC-style files with comments: apps/web-app/tsconfig.json plus the canonical and mirrored typescript-expert/references/tsconfig-strict.json files
Targeted typo scan found one confirmed source typo: skills/remotion-best-practices/rules/charts.md:20

Findings

High - archive refresh fallback extracts untrusted archives without path validation

File: apps/web-app/refresh-skills-plugin.js
Lines: 220-261
Evidence: syncWithArchive() downloads GitHub tar/zip archives and extracts them with tar -xzf, Expand-Archive, or unzip -o directly into update_temp.
Risk: an archive containing path traversal entries or malicious symlinks can write outside the intended extraction root before the code checks for antigravity-awesome-skills-main/skills.
Current tests: apps/web-app/src/__tests__/refresh-skills-plugin.security.test.js covers loopback/auth and fast-forward behavior but does not cover archive extraction safety.
Suggested fix: replace shell extraction with a safe extraction helper that rejects absolute paths, .. segments, unsafe symlinks, and post-extraction realpaths outside update_temp; add tar/zip regression tests.

High - Telegram Node boilerplate resolves a vulnerable request stack

File: skills/telegram/assets/boilerplate/nodejs/package.json
Lines: 11-21
Evidence: a temp-copy install/audit of this package reported 5 vulnerabilities: critical form-data <2.5.4, moderate request, qs, tough-cookie, and uuid. npm ls traces the critical path through node-telegram-bot-api@0.66.0 -> @cypress/request-promise@5.0.0 -> request@2.88.2 -> form-data@2.3.3.
Risk: users following this boilerplate get a vulnerable dependency graph even though the boilerplate TypeScript itself compiles.
Current tests: the root dependency audits do not cover this package because it has no committed lockfile and is not part of the root workspace install.
Suggested fix: replace node-telegram-bot-api with a maintained Telegram client that does not depend on deprecated request, or commit a tested lockfile/override strategy that removes the vulnerable transitive versions.

Medium - full-repo install migration preserves target symlinks

File: tools/bin/install.js
Lines: 474-493
Evidence: when target.path/.git exists and target.path itself is a symlink, the migration renames the symlink to a backup and then recreates the same symlink at target.path.
Risk: this diverges from the non-symlink migration path, where a fresh directory is created. Subsequent writes continue through the symlink target, which is surprising during a "skills-only layout" migration and weaker as a trust boundary.
Suggested fix: either reject symlinked full-repo migration targets with a clear error, or materialize a real directory at target.path after backing up the symlink.

Medium - prerendered web app SEO count is stale

File: apps/web-app/scripts/prerender-routes.js
Lines: 12, 156-159
Evidence: HOME_CATALOG_COUNT is 1273 and the generated title is hardcoded to 1,273+, while the current registry has 1465 skills. The checked apps/web-app/dist/index.html contains 1,273+ installable AI skills catalog.
Risk: search/social metadata undersells the catalog and drifts from README/CATALOG counts.
Suggested fix: derive the title count from the actual skills.json count used by prerendering, or update the constant/title together with generated assets.

Files: apps/web-app/index.html, apps/web-app/public/social-card.svg
Evidence: index.html still contains 1,326+ in title, description, Open Graph, and Twitter metadata. social-card.svg also displays 1,326+ Agentic Skills and describes a 1,326 plus headline.
Risk: the built homepage is partially corrected by prerendering, but source metadata, fallback HTML, and social preview assets remain stale and can leak into previews or static hosting paths.
Suggested fix: derive counts from the same generated registry source used by the app, or update these assets during the web asset sync step.

Medium - app environment docs name an unused canonical URL variable

File: apps/web-app/README.md
Line: 59
Evidence: the README documents VITE_SITE_URL, but apps/web-app/scripts/generate-sitemap.js reads SEO_SITE_URL || WEBSITE_BASE_URL and apps/web-app/scripts/prerender-routes.js reads SEO_SITE_URL.
Risk: maintainers testing non-default hosts may set an environment variable that has no effect on sitemap/prerender canonical URLs.
Suggested fix: document SEO_SITE_URL and optionally WEBSITE_BASE_URL; remove or implement VITE_SITE_URL.

Medium - WhatsApp webhook HMAC validation throws on malformed signature length

File: skills/whatsapp-cloud-api/assets/boilerplate/nodejs/src/webhook-handler.ts
Lines: 38-41
Evidence: crypto.timingSafeEqual(Buffer.from(signature), Buffer.from(expectedSignature)) is called without first checking that the two buffers have the same byte length. A direct Node check with a short signature throws RangeError: Input buffers must have the same byte length.
Risk: malformed x-hub-signature-256 headers can turn an invalid webhook request into an exception/500 path instead of a clean 401, creating avoidable error noise and a small DoS surface.
Suggested fix: validate the sha256=<64 hex chars> shape before comparison, or compare only after checking equal buffer lengths; return 401 for malformed signatures.

Medium - Telegram webhook URL embeds the bot token

File: skills/telegram/assets/boilerplate/nodejs/src/bot-client.ts
Lines: 24-45
Evidence: the Express route and registered webhook URL both use /webhook/${this.token} while Telegram already supports secret_token.
Risk: bot tokens placed in URL paths are likely to appear in reverse-proxy logs, request logs, telemetry, browser history, and webhook provider dashboards. The optional secret header loses much of its value if the bearer token itself is the path secret.
Suggested fix: use a random non-secret route ID or fixed /webhook route plus x-telegram-bot-api-secret-token; never include the bot token in URLs.

Medium - strict skill audit is not clean even though standard gates pass

Resolved in follow-up patch: npm run audit:skills:strict now enforces a versioned warning budget in tools/config/audit-skills-strict-budget.json, with explicit limits for total warnings, warning-only skills, and the top strict finding codes. The current legacy baseline is documented and the command fails on errors or warning regressions above that budget.

Command: npm run audit:skills:strict
Evidence: strict mode scanned 1465 skills and found 0 errors, but still exited non-zero because 740 skills are warning-only with 865 total warnings. Top findings were risk_suggestion (1013), missing_examples (387), and skill_too_long (193).
Risk: regular validation and warning-budget gates can look clean while the stricter skill-quality bar is materially unmet.
Suggested fix: either wire a consciously scoped subset of strict findings into CI/release gating, or document strict mode as an advisory backlog with explicit thresholds and owners.

Medium - web asset setup publishes disabled skills as static files

File: tools/scripts/setup_web.js
Lines: 70-80
Evidence: app:setup copies the entire skills/ directory into apps/web-app/public/skills without excluding dot directories. Current generated assets contain apps/web-app/public/skills/.disabled and apps/web-app/dist/skills/.disabled, each around 17 MiB, with 1294 files copied from 34 disabled skill directories. skills_index.json correctly excludes .disabled and reports 0 disabled entries.
Risk: disabled skills are not shown in the catalog but are still shipped as static web assets and can be fetched directly by path, which conflicts with the validator/indexer semantics that hidden/disabled directories are excluded.
Suggested fix: make copyFolderSync() skip dot-prefixed directories, or explicitly skip .disabled when copying web assets; add a web setup test that asserts public/skills/.disabled is absent after app:setup.

Low - Junta scrapers bypass the shared TLS verification switch

Files: skills/junta-leiloeiros/scripts/scraper/jucisrs.py, skills/junta-leiloeiros/scripts/scraper/jucesp.py, skills/junta-leiloeiros/scripts/scraper/jucema.py
Lines: jucisrs.py:178-182, jucisrs.py:223-229, jucisrs.py:245-257, jucesp.py:115-121, jucema.py:161-169
Evidence: base_scraper.py provides should_verify_tls() and tests assert TLS verification is enabled by default, but these state-specific scrapers instantiate httpx.AsyncClient(verify=False) or Playwright with ignore_https_errors=True directly.
Risk: the test suite proves the shared base helper default, but not the actual transport behavior of these scraper implementations; network interception can still tamper with scraper input by default.
Suggested fix: replace direct verify=False / ignore_https_errors=True with should_verify_tls() in state-specific scrapers, and extend test_junta_tls_security.py to scan or monkeypatch concrete scraper clients.

Low - nested skill IDs lose their category namespace in generated routes

Files: tools/scripts/generate_index.py, apps/web-app/src/pages/SkillDetail.tsx, apps/web-app/src/components/SkillCard.tsx
Lines: generate_index.py:879-887, SkillDetail.tsx:61-79, SkillCard.tsx:16
Evidence: generate_index.py sets each skill id to os.path.basename(root) while keeping the full nested path separately. The current index has 20 nested skills, including base => skills/libreoffice/base, templates => skills/app-builder/templates, and 2d-games => skills/game-development/2d-games. The web app routes to /skill/${skill.id} and resolves detail pages with skills.find(s => s.id === id).
Risk: there are currently 0 duplicate IDs, so this is not breaking today. A future top-level base, another nested base, or similarly generic nested name would create an ambiguous route/star/canonical-ID collision because the public ID omits the category namespace.
Suggested fix: either generate route-safe IDs from the relative path for nested skills, or add a uniqueness test that fails when any future nested basename collides with another indexed skill.

Low - SEO verifier checks a legacy manifest filename, not the linked manifest

Files: apps/web-app/index.html, apps/web-app/scripts/verify-seo-assets.js, .github/workflows/pages.yml
Evidence: index.html links %BASE_URL%site.webmanifest, while verify-seo-assets.js and the Pages workflow check dist/manifest.webmanifest. Both files currently exist, so the gate passes, but it does not prove the linked manifest is present.
Risk: a future build could break site.webmanifest while the verifier still passes because manifest.webmanifest remains.
Suggested fix: make the verifier and Pages workflow check site.webmanifest, or explicitly verify every manifest referenced by index.html.

Low - Telegram HTML messages interpolate display names without escaping

File: skills/telegram/assets/boilerplate/nodejs/src/handlers.ts
Lines: 8-14, 31-37
Evidence: /start injects msg.from?.first_name into a message sent with parse_mode: 'HTML'; /about does the same with bot profile fields.
Risk: Telegram display names containing <, >, or & can break formatting, cause send failures, or create unintended markup in generated bot replies.
Suggested fix: HTML-escape interpolated values whenever parse_mode: 'HTML' is used, or switch these messages to plain text/Markdown with appropriate escaping.

Low - Chinese documentation is release/count stale

Files: docs_zh-CN/README.md, docs_zh-CN/users/getting-started.md, docs_zh-CN/final-validation-report.md
Evidence: English docs and package metadata are at 11.5.0 / 1465, while docs_zh-CN/README.md is synced to 10.7.0 / 1436, contains an internal mention of version 8.10.0, and docs_zh-CN/users/getting-started.md still says V10.7.0. docs_zh-CN/final-validation-report.md is also for V8.10.0.
Risk: localized entry points give outdated release and catalog information.
Suggested fix: either refresh translated docs after canonical English updates or mark old translation reports as historical snapshots so they are not read as current documentation.

Low - internal markdown links are broken in integration and localized docs

Files: docs/integrations/jetski-gemini-loader/README.md, docs_zh-CN/README.md, docs/vietnamese/*.vi.md, docs_zh-CN/integrations/jetski-gemini-loader/README.md
Evidence: a relative-path-aware local scan over README.md, docs/, and docs_zh-CN/ found 105 broken internal links. Confirmed examples:
- docs/integrations/jetski-gemini-loader/README.md links to ../../docs/integrations/jetski-cortex.md, which resolves to nonexistent docs/docs/integrations/jetski-cortex.md; the real target is docs/integrations/jetski-cortex.md.
- docs_zh-CN/README.md links to paths such as docs/users/bundles.md, which resolve under docs_zh-CN/docs/... instead of the repo root.
Risk: GitHub users following localized or integration docs hit dead links even though the current validate:references gate passes.
Suggested fix: add a path-aware markdown link checker to CI, then repair localized relative paths and the Jetski integration link.

Low - Remotion chart rule has a visible typo

File: skills/remotion-best-practices/rules/charts.md
Line: 20
Evidence: the sentence reads See Bar Chart Example for a basic example implmentation.
Risk: low editorial quality issue in a user-facing skill rule.
Suggested fix: change implmentation to implementation.

Low - link validation script writes a stale, machine-specific translated report

Files: scripts/validate-links.sh, docs_zh-CN/link-validation-report.txt
Evidence: scripts/validate-links.sh writes docs_zh-CN/link-validation-report.txt with a timestamp and absolute local paths. The script labels this as a translated report but scans docs/, not docs_zh-CN/, and resolves internal links by basename only. The script itself notes that links such as ../../CATALOG.md can be false positives.
Risk: rerunning the script creates non-portable tracked drift and produces a report that can mix real broken links with known false positives.
Suggested fix: either make the report deterministic and unambiguous, or stop tracking it. Resolve links relative to each source file and run separate checks for canonical docs and translated docs.

Low - deliberate pipe-to-shell examples remain allowlisted

Resolved in follow-up patch: executable curl|bash, curl|sh, and irm|iex install examples in canonical skill sources were replaced with package-manager or download-inspect-execute flows, and obsolete pipe-to-shell allowlists were removed. The remaining security-allowlist comments cover non-pipe rule families or sandbox notes.

Files include skills/bun-development/SKILL.md, skills/linkerd-patterns/SKILL.md, skills/cloud-penetration-testing/SKILL.md, skills/varlock/SKILL.md, skills/evolution/SKILL.md, and plugin mirrors.
Evidence: npm run security:docs passes because these examples are allowlisted, but the repo still contains executable curl|bash, curl|sh, and irm|iex examples.
Risk: this is accepted by current policy, but remains risky guidance for automated agents.
Suggested fix: where practical, replace executable pipe-to-shell examples with package-manager or checksum-verified alternatives; keep allowlists only where no safe practical alternative exists.

Low - documentation examples use realistic-looking secret placeholders

Resolved in follow-up patch: canonical examples now use <AWS_ACCESS_KEY_ID>, <AWS_SECRET_ACCESS_KEY>, and <BASE64_PRIVATE_KEY>, and npm run security:docs blocks those realistic placeholder patterns from returning.

Files include skills/comfyui-gateway/references/integration.md, skills/security/aws-iam-best-practices/SKILL.md, skills/k8s-manifest-generator/resources/implementation-playbook.md, and plugin mirrors.
Evidence: secret-pattern scanning found AKIAIOSFODNN7EXAMPLE, wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY, and -----BEGIN PRIVATE KEY----- examples. These are not confirmed live secrets: the AWS values are standard examples and the private key block is an ellipsis placeholder.
Risk: realistic secret-looking examples can create noisy secret-scan alerts and make it harder to distinguish documentation placeholders from real leaks.
Suggested fix: replace realistic credential examples with unmistakable placeholders such as <AWS_ACCESS_KEY_ID>, <AWS_SECRET_ACCESS_KEY>, and <BASE64_PRIVATE_KEY>.

Non-findings / bounded checks

The apparent --- frontmatter prefix in eight skills is accepted by the repo parser (^---\s*\n) and those skills are indexed, so this is not currently a validator bug.
Duplicate names under skills/.disabled are intentionally skipped by validator/index generation.
Root npm production audit reported 0 vulnerabilities.
Root, web app, and data package moderate audits reported 0 vulnerabilities.
Loki todo-app frontend and backend moderate audits reported 0 vulnerabilities.
Loki todo-app frontend and backend builds completed successfully.
Telegram and WhatsApp Node boilerplates compile successfully when installed in temp copies.
WhatsApp Node boilerplate moderate audit reported 0 vulnerabilities in a temp copy.
Python syntax checking found 0 errors across 1077 source files.
Shell syntax checking found 0 errors across 61 shell scripts.
JavaScript syntax checking found 0 errors across 74 source files.
YAML and XML-family parse checks found 0 syntax errors in tracked files.
The 7 tracked symlinks all resolve to in-repo targets.
The strict JSON parse failures are JSONC/commented config files rather than malformed runtime JSON; TypeScript accepts tsconfig.json comments, and the commented tsconfig-strict.json is a reference artifact mirrored byte-for-byte.
Plugin mirror skill files are byte-for-byte synchronized with canonical skills/ sources: 3094 matched, 0 drift.
Full-library plugin skill counts differ from the root visible catalog intentionally: the Codex plugin advertises 1,429 plugin-safe skills, and omitted root skills are compatibility-filtered rather than missing files.
Duplicate name values across all tracked SKILL.md files are expected because canonical skills are copied into full plugins and bundle plugins; canonical-vs-plugin byte checks found no content drift.
All 76 plugin manifests parsed with required identity/version fields present.
npm run audit:maintainer did not reveal a repo health issue beyond the intentional untracked audit report; git status --porcelain --untracked-files=no was clean.
npm pack --dry-run --json produced the expected installer-only package contents: LICENSE, README.md, package.json, tools/bin/install.js, and tools/lib/*.
No tracked temporary/bytecode/macOS artifact files were found for __pycache__, *.pyc, *.pyo, .DS_Store, *.tmp, *.bak, or *.orig.
The targeted common-typo scan produced two Teh matches that are a person's surname in the Geoffrey Hinton skill, not typos.

26 KiB Raw Blame History