codex — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited codex (Plugin) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
A Claude Code plugin that invokes the local Codex CLI as an independent analysis partner.
Gives Claude Code a structured way to delegate analysis tasks to Codex — brainstorming alternatives, red-teaming plans, debugging from a fresh perspective, reviewing diffs adversarially, and more. Codex runs locally, reads the codebase, and returns its analysis for Claude to synthesize.
Brainstorm, Red-team, Debug, Plan Review, Diff Review, Spec Extraction, Rollout/Rollback, Compare/Decide, Test Gaps, Explain, Post-mortem, Attack Surface, Exhausted Hypotheses.
Across my own Claude Code transcripts (one operator, so an internal audit, not a benchmark): ~450 Codex reviews over 9 projects, ~436 judged in full.
Across that set: ~3 findings each, a ~2.1% false-finding rate, and ~32% that pushed past a local edit into a plan or direction change. In convergence mode (below), ~62% of review chains reached an affirmative verdict in-session. Weakest on subjective style review. Full breakdown and caveats: the agent-tools README.
Multi-round iterations on the same artifact (review plan v1 → v2, iterative debugging) can preserve prior Codex context via named sessions:
/codex --new-session review-auth # create a named session
/codex --session review-auth # resume it later
/codex list # see all sessions
/codex delete review-auth # clean upSessions are stored per-worktree under .claude/.codex-sessions/. See the Session Management section in skills/codex/SKILL.md for the full workflow.
For reviewing artifacts that evolve across multiple revisions — specs, plans, designs — Codex can run in a convergence loop: review → fix → re-review until the reviewer gives an affirmative verdict or you stop. Claude orchestrates the loop with user gates after each round, cites prior findings on each pass so the reviewer can detect drift, and watches for the scope-drift failure mode where each round's "real" findings pull the artifact into a design the user never asked for.
See the Convergence Mode section in skills/codex/SKILL.md for the loop shape (two user decisions per round — which fixes to apply, then whether to continue), per-round prompt construction, and the anti-pattern guidance that tells Claude when to stop and re-confirm scope instead of continuing.
source on a bash script — skills/codex/session-mgr.sh)reviewer subagentThe skill runs a mandatory QA pass using a reviewer subagent after summarizing high-stakes modes (plan-review, red-team, diff-review, exhausted-hypotheses, attack-surface). It expects the subagent from koenvdheide/claude-reviewer. Without it, the skill falls back to self-review against the same fidelity rules — workflow still works, just less rigorous.
Via the agent-tools marketplace:
/plugin marketplace add koenvdheide/agent-tools
/plugin install codex@agent-tools
/reload-pluginsRefresh later with /plugin marketplace update agent-tools, then /reload-plugins.
Claude invokes the skill automatically when a task matches, or you can invoke it directly:
/codex red-team my authentication refactor planMIT
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.