[theCodeDrift]

Adds Cloudflare's security-audit skill (MIT) through dotagents.

• agents.toml: trusts cloudflare/security-audit-skill and declares the security-audit skill (installed gitignored like the other managed skills; agents.lock is gitignored here).
• dotagents.md: lists the skill as managed, with its origin and the updated trust snippet.
• eslint.config.js: ignores .agents/skills/security-audit/. The skill ships top-level validate-*.cjs files that crashed local pnpm lint with a typed-linting error. CI never installs skills, so only local runs were affected. Open question: this could instead ignore every managed skill.

No changeset: repo-internal agent tooling, nothing ships to package consumers.

Checks: pnpm lint, prettier, pnpm dotagents doctor (4 skills, symlinks intact).

Commits are unsigned: the signing key couldn't be unlocked remotely, and the owner chose to ship unsigned.

Closes #567

[theCodeDrift]

@claude /review

[github-actions]

Claude finished @theCodeDrift's task in 1m 3s —— View job

---

Review mode: incremental — read 1 prior review thread(s) before reviewing.

.prior-review.json contained no review threads and no prior review summaries — the only top-level comment on file is the @claude /review trigger itself. There is nothing previously raised to classify as addressed or still-open, so this is effectively a first full review of the diff.

• Read .prior-review.json (empty — no prior findings to carry forward)
• Fetch PR diff and metadata
• Review agents.toml
• Review dotagents.md
• Review eslint.config.js
• Post final summary

Assessment

This is a small, well-scoped, well-documented chore. Checked each file against the whole-file context (not just the hunks):

• agents.toml — Adds cloudflare/security-audit-skill to trust.github_repos and declares the [[skills]] entry with a matching source. Consistent with the existing entries' shape; the trusted repo matches the skill's source exactly, so there's no mismatch that would let the skill resolve from an untrusted source.
• dotagents.md — The "managed by dotagents" skill list, the count ("three" → "four"), the credit table (with MIT noted, consistent with how iterate-pr's Apache-2.0 and openspec-*'s MIT are annotated), and the duplicated trust snippet are all updated together and stay consistent with the real agents.toml. No drift between the doc and the config.
• eslint.config.js — Adds .agents/skills/security-audit/ to ignores, with a comment explaining that the skill ships top-level zero-dependency CommonJS validators the type-aware rules can't parse. This is consistent with the file's existing pattern of ignoring specific CommonJS/DATA directories rather than broad globs (e.g. .agents/skills/*/scripts/, .github/scripts/). On the PR's own open question — whether to instead ignore every managed skill wholesale — the targeted approach taken here is the better default: the other managed skills (dotagents, code-review, pr-writer) are prose/SKILL.md-shaped and apparently have nothing that trips the linter today, so ignoring them preemptively would just hide future problems in those directories. Scoping the ignore to the one skill that actually needs it is right; it can be widened later if another managed skill ships lintable JS.

No correctness, security, or test-adequacy issues found. Trusting a new GitHub source is inherently a trust decision rather than a code one — the doc already frames that tradeoff accurately ("a skill is instructions an agent will follow... adding a trusted source has more in common with adding a dependency than with adding a bookmark"), and Cloudflare's security-audit-skill is a named, attributable upstream, so there's nothing further to flag there.

No inline comments filed — nothing rose to that bar.

[theCodeDrift]

Re: @github-actions — "Claude finished @theCodeDrift's task in 1m 3s"…
#684 (comment)

Acknowledged: the review found no issues and endorsed keeping the ESLint ignore scoped to .agents/skills/security-audit/, so no changes were made.

— AI Coding Agent