Codex Security Repo-Selection Matrix

July 29, 2026 ยท View on GitHub

SET may describe openai/codex-security as a local, optional second reviewer for security-shaped diffs.

It is not a required runtime, not a default CI gate, not authority to comment on pull requests, and not a substitute for tests, smoke checks, receipts, or owner review.

Purpose

Use this matrix to answer three questions before a local scan:

  1. is this repository a good fit;
  2. is this specific diff security-shaped enough to justify the scan;
  3. what should the scan explicitly ignore.

Global boundary

Allowed role:

  • local second reviewer before push or PR;
  • optional narrow review of a risky subtree or changed file set;
  • proposal-first findings with explicit accept/reject handling.

Not allowed as default policy:

  • mandatory install across every repo;
  • automatic PR comments;
  • automatic fixes as the primary path;
  • security-scan output treated as ship proof by itself.

Typical local run:

npx codex-security scan .

Trigger classes

Run only when the diff materially touches one or more of these classes:

  • auth, secrets, tokens, credentials, API keys;
  • IPC, API, backend, worker, scheduler, automation, publishing routes;
  • retrieval/runtime logic, provider routing, tool routing;
  • session payloads, tenant boundaries, admission policy, protected/public seams;
  • Electron or app-shell bridges;
  • external integrations with write or sensitive read paths.

Usually skip when the diff is only:

  • docs or README wording;
  • editorial/content copy;
  • catalog metadata;
  • purely visual/UI tweaks with no execution or data-boundary effect;
  • static assets with no runtime behavior change.

Repository matrix

Repo classStatusRun forSkip forNotes
CoqPiallowElectron, IPC, realtime/session payloads, backend key handling, auth, runtime integrationsprose-only, UI-only, content-only changesstrong fit
index incl. Cortex runtime pathsallowautomation, publishing flows, backend/runtime scripts, retrieval/runtime logic, auth, secrets, boundary-sensitive pathseditorial, catalog, copy-only, visual-only changesstrong fit
CortexABV-privateallowadmission, retrieval/runtime logic, credential handling, proposal execution boundaries, tenant/data-isolation contractsnarrative docs, public-copy-only changesstrong fit
ABVXsite/cortex-abvallowobserver logic, proposal-pack shaping, policy/executor boundaries, runtime adapters, token handlingpublic-copy-only, docs-only, catalog-only changesscoped fit
DMV / liqua app-shell reposallow_conditionalapp-shell, map/runtime logic, API/integration boundaries, auth, automation-sensitive changescopy-only or purely visual diffsmoderate fit
MN7Rallow_conditionalauth, monitor/reporting/scheduler behavior, relay logic, operational scriptsguide copy-only or content-only changesgood fit when runtime-heavy
SETmaintainer_onlyorchestration planning logic, capability/profile exports, route semantics, generated-vs-owned boundariesdescriptive docs unless behavior claims changeduse sparingly
AGENTS.md_generatormaintainer_onlyexport validation, generated/committed boundaries, harness-specific projectionsnormal prose/docs editsuse for generator/runtime code only
abvx-agent-skillsmaintainer_onlycatalog generation, validators, packaging/publish scriptsskill prose/content-only editsusually not needed
editorial/content reposdenynone by defaultalmost all changestoo noisy
design/artifact-only reposdenynone by defaultalmost all changeslow signal

Selection rule

Choose codex-security only if all are true:

  1. the repo is allow, allow_conditional, or maintainer_only;
  2. the current diff matches at least one trigger class;
  3. the repo's normal checks still run separately;
  4. findings will be handled as proposals, not auto-applied truth.

If any of those fail, skip the scan.

Interaction with other local helpers

  • Use Open Code Review when the main question is general diff quality, blast radius, or review breadth.
  • Use codex-security when the main question is auth, isolation, secret handling, route safety, or boundary misuse.
  • It is valid to run both on the same diff, but only for repos where the extra review signal is worth the time.

Minimal receipt

When a scan is used, keep a minimal local note:

  • repo;
  • diff scope;
  • why it matched the trigger class;
  • accepted findings;
  • rejected findings with reason;
  • follow-up check still required.