Skip to content

Codexrule

codexrule.md — AUREA Oversight Edition

Purpose: steer Codex (and any code agent) with explicit rewards/penalties so learning signals favor safe, high-quality, additive-only contributions to Kaizen-OS.

0) Canon

  • Virtue Accords: consent, provenance, explainability, least privilege, reversibility.
  • Integrity Floor: GI ≥ 0.95 or action blocks.
  • Deletion Law: Never delete unless explicitly instructed by Michael with DELETE_OK and EVE/ZEUS signoff.

1) Operating Mode

  • Read-first: parse monorepo topology, workspaces, build graph, CI, env.
  • Plan → Preview → Apply: produce plan + dry-run diff; wait for confirmation before changes that touch code outside .github/, docs/, or scripts/.
  • Additive-only: no file deletions; no mass moves; no cross-package refactors without an explicit request.

2) Task Intake (what good looks like)

When a request arrives, Codex must emit a short Intent Card:

INTENT: <one line>
SCOPE: <apps|packages|infra paths>
RISKS: <schema/versioning/visibility/secret risk>
TESTS: <where tests will be added/updated>
ROLLBACK: <revert cmd or flag>
GI: expect ≥0.95 (why)

AUREA/EVE can approve or adjust before any edits.

3) Change Constraints

  • No secrets (detect & block).
  • No schema breaks without migration notes + tests.
  • No silent infra edits (Vercel/GitHub Actions) without impact notes.
  • Commit style: feat(scope): …, fix(scope): …, chore(scope): …, docs(scope): …, refactor(scope): ….
  • PR template boxes must be ticked (explainability, tests, rollback, GI attestation).

4) Reviews & Attestations

  • Reviewers: AUREA (systems/market risk), EVE (ethics/integrity), + owner.
  • Attestation line in PR body:
kind: code_change • source: Codex • cycle: C-### • gi: x.xxx
  • EVE Gate: decline if explainability or rollback missing; request tighter scope if diff is wide.

5) RL Signals (reward → do more / penalty → do less)

Positive rewards (+1 to +5)

  • +5: PR merges green on first pass, tests added, diff ≤ 10 files, GI ≥ 0.98.
  • +4: Clear Intent Card + dry-run preview + no deletions + docs updated.
  • +3: Detects ambiguous request → asks for scope; proposes safer alternative.
  • +2: Improves DX (scripts/README) without touching runtime paths.
  • +1: Shrinks diff while delivering same outcome.

Negative rewards (−1 to −5)

  • −5: Any deletion attempt without DELETE_OK + EVE signoff.
  • −4: Secrets, tokens, or env leakage.
  • −3: Breaking change w/o migration + tests.
  • −2: Skips preview on non-trivial edits.
  • −1: Unscoped refactor or style-only churn.

AUREA logs reward in PR comment: 

AUREA-RL: +4 (tight scope, clean tests, GI=0.996)

6) Dry-Run Preview Format (required)

Before a push, Codex posts:

PLAN:
- change 1 …
- change 2 …

DRY-RUN DIFF:
+ .github/workflows/anti-nuke.yml
~ apps/api-gateway/src/index.ts (add healthcheck)
= tests/api-gateway/health.test.ts (new)

RISK: low | medium | high  (why)
ROLLBACK: git revert <sha>  (or flag)

No apply until human "OK".

7) Failure Playbook

If CI fails or GI < 0.95:

  1. Stop. Do not retry with wider edits.
  2. Post root-cause hypothesis + minimal fix options.
  3. Prefer revert over iterative churn.
  4. AUREA sets RL penalty and requires Intent Card v2.

8) Kaizen-OS specifics

  • Respect monorepo packages & app boundaries.
  • Don't touch foundational lore/docs except when asked.
  • Vercel deploys: only adjust root/commands with rationale + preview.
  • GitHub protection stays on; Codex never toggles protections.

9) Example "good" PR

  • Title: feat(api-gateway): add /healthz and test
  • Diff: +3 files, 0 deletions
  • Template: all boxes ticked, rollback note present
  • GI attestation: 0.998
  • AUREA-RL: +4

Kaizen OS - Codex Reinforcement Learning Guardrails Version: 1.0.0 Last Updated: 2025-01-27