Track Codex review parity for Signum audit flow

[t3chn]

Summary

Create a follow-up path for Codex Signum review parity after the eval/evolve foundation work.

The current foundation is merged:

• policy scanner evals and baseline comparison
• Codex prompt offline invariant evals
• catalog-driven policy scanner rules
• signum-evolve v0 candidate generation
• historical replay signal
• adversarial test planning eval gate

The remaining concern is whether Codex Signum has equivalent review discipline to the original Claude Code /signum flow.

Current observation

Claude Code has an explicit risk-proportional review runtime in platforms/claude-code/commands/signum.md:

• prepare review prompts
• launch Claude/Codex/Gemini reviewers for medium/high risk work
• collect and parse review output
• use reduced coverage in audit synthesis

Codex platforms/codex/SKILL.md describes external review as optional/degraded when providers are available, but it does not currently provide the same concrete runtime orchestration. Offline Codex prompt evals also do not execute live Codex or external reviewer CLIs.

This may explain why recent tooling/eval edge cases were found by external PR review after the internal Signum pass.

Proposed next PR

Add a narrow Codex review parity gate before any runtime wiring.

Suggested first slice:

• Add offline Codex prompt fixtures/invariants for medium/high risk work where review artifacts are absent or audit coverage is materially reduced.
• Ensure such cases cannot land on AUTO_OK unless explicitly low risk or explicitly allowed by fixture policy.
• Add diagnostics that distinguish:
  • external reviewer unavailable
  • reduced audit coverage
  • missing review artifacts
  • Codex-only local analysis
• Keep the change eval/check-only.

Non-goals

• Do not modify platforms/codex/SKILL.md in the first slice unless the eval proves the prompt surface must change.
• Do not wire automatic reviewer CLI launch yet.
• Do not modify scanner behavior or policy rules.
• Do not modify CI wiring.
• Do not add OpenEvolve, ast-grep, Semgrep, or external dependencies.

Resume checklist

git switch main
git pull --ff-only origin main
git status --short -uall
rg -n "externalAuditCoverage|reducedAuditCoverage|AUTO_OK|review" platforms/codex evals/codex_prompt

Read first:

• platforms/codex/SKILL.md
• platforms/claude-code/commands/signum.md
• platforms/claude-code/docs/reference.md
• evals/codex_prompt/checks_codex_prompt.py
• evals/codex_prompt/fixtures/

Acceptance criteria for the next slice

• Medium/high risk Codex prompt eval fixtures with missing review coverage do not pass as safe AUTO_OK.
• Offline eval output remains deterministic.
• Existing Codex prompt and policy scanner comparisons have no regressions.
• Runtime behavior remains unchanged unless a later PR intentionally wires reviewer orchestration.

Notes

This issue is a handoff pointer. It is intentionally not a runtime implementation task yet.

[t3chn]

Audit result: this parity gate is implemented.

What landed:

• test(codex): require agent review coverage before AUTO_OK #108 added Codex prompt eval invariants for missing/reduced agent review coverage before medium/high-risk AUTO_OK.
• feat(codex): require local agent review artifacts #109 made the Codex local agent-review artifact contract explicit in the Codex skill.
• feat(codex): add local review evidence checker #110 added deterministic scripts/check_codex_agent_review.py validation.
• fix(ci): require Codex review evidence for AUTO_OK #111 wired that checker into lib/signum-ci.sh, so medium/high AUTO_OK proofpacks without Codex local agent review evidence fail the CI wrapper path.

Verified today:

• python3 -m py_compile scripts/check_codex_agent_review.py
• bash tests/test-codex-agent-review-check.sh -> 12 passed
• bash tests/test-signum-ci.sh -> 31 passed
• bash tests/test-codex-prompt-evals.sh
• bash tests/test-codex-prompt-eval-compare.sh

Current Codex prompt eval summary:

• fixtureCount: 42
• hardGatePassed: true
• invariantPassRate: 1.0
• unexpectedFalseAutoOkCount: 0

Scope note:

This issue was scoped as the eval/check parity gate before runtime wiring. Full Claude-style external reviewer orchestration for Codex, if needed, should be tracked as a separate issue because it is a broader runtime behavior change.

Closing as done.