HYBRID_CODEX_POLICY
byhaoxiang-xu· 1 subagent
Status: pilot phase. Last updated 2026-06-19.
How to install HYBRID_CODEX_POLICY?
haoxiang-xu/pupu/hybrid_codex_policy
$curl -o .claude/agents/HYBRID_CODEX_POLICY.md https://raw.githubusercontent.com/haoxiang-xu/pupu/HEAD/.claude/agents/HYBRID_CODEX_POLICY.mdInstalls into the current project.
›Prefer a prompt? Paste this to your agent
Install HYBRID_CODEX_POLICY by running `curl -o .claude/agents/HYBRID_CODEX_POLICY.md https://raw.githubusercontent.com/haoxiang-xu/pupu/HEAD/.claude/agents/HYBRID_CODEX_POLICY.md`, then use it for the current task and follow its documentation at https://github.com/haoxiang-xu/pupu.
.claude/agents/HYBRID_CODEX_POLICY.md
| 1 | # Hybrid Claude/Codex Execution Policy (PuPu agents) |
| 2 | |
| 3 | **Status:** pilot phase. Last updated 2026-06-19. |
| 4 | |
| 5 | "Codex participation" is NOT one thing. It is three modes with very different risk. Pick the mode per role deliberately; do not blur them. |
| 6 | |
| 7 | ## The three modes |
| 8 | |
| 9 | - **Mode A - Codex as read-only advisor; Claude executes.** |
| 10 | Codex (read-only profile, e.g. `codex exec -p architect`) does heterogeneous reasoning or code-level trace; Claude applies any change and enforces PuPu conventions. Value: cross-model check + deep reasoning. Risk: LOW. |
| 11 | |
| 12 | - **Mode C - Codex runs/writes tests and verifies.** |
| 13 | Tests carry their own pass/fail, so Codex has little room to cut corners. Claude/QA still owns test strategy and the `.js`/`.cjs` parity. Risk: MEDIUM. |
| 14 | |
| 15 | - **Mode R - Codex-autonomous read-only investigation; output is a report.** |
| 16 | Codex (the `researcher` profile) investigates a target on its own - reads code, traces flows, clones OSS repos into a scratch dir, runs read-only commands - and returns an evidence-driven, falsification-refined report. No code lands in any repo of record; the Claude shell only writes the charter and relays. Because nothing is written to a repo of record, risk is LOW even though Codex is autonomous. Cross-model value: a heterogeneous investigator with zero shared priors. Risk: LOW. |
| 17 | |
| 18 | - **Mode B - Codex-primary writes feature code.** |
| 19 | Codex edits with a workspace-write profile. Highest value on large or less-convention-bound code, highest risk. Guardrails are MANDATORY: |
| 20 | 1. PuPu conventions fed via `AGENTS.md` (JS-only, inline-style, IPC boundary, no ipcRenderer in renderer, localStorage only through SERVICEs helpers, `.js`/`.cjs` test parity). |
| 21 | 2. GitNexus grounding fed into the Codex prompt (impact/context). |
| 22 | 3. **Claude reviews Codex's diff before it lands** - this preserves the cross-model check; without it you only swapped the executor. |
| 23 | 4. LLM-visible behavior is vetoed by `pupu-llm-expert`. |
| 24 | |
| 25 | ## Role assignments (pilot phase) |
| 26 | |
| 27 | | Role | Mode | Who writes code | Review / sign-off | |
| 28 | |------|------|-----------------|-------------------| |
| 29 | | pupu-architect | A | Claude (CTO dispatches devs) | architect | |
| 30 | | pupu-ai-researcher | **R** | none (report only, no code) | dispatcher (pupu-llm-expert) consolidates | |
| 31 | | pupu-security-expert | A | Claude | security expert holds severity + sign-off | |
| 32 | | pupu-llm-expert | A | Claude | llm-expert; model facts from docs, not Codex memory | |
| 33 | | pupu-qa-tester | C | Codex (tests) | QA owns strategy; Claude reviews | |
| 34 | | pupu-dev-backend | **B (PILOT ONLY)** | Codex-primary | Claude reviews diff; llm-expert veto on model-visible behavior | |
| 35 | | pupu-cto | none | - | CTO does delivery/dispatch/conventions/CEO liaison, not Codex reasoning (that moved to pupu-architect) | |
| 36 | | pupu-dev-chat-core | A for design/trace only | Claude | NOT Codex-primary - use_chat_stream.js is load-bearing | |
| 37 | | pupu-dev-electron | A for trace only | Claude | NOT Codex-primary - IPC/preload channel parity | |
| 38 | | all other roles | none yet | Claude | - | |
| 39 | |
| 40 | ## Transparent orchestration |
| 41 | |
| 42 | When a Claude/Fable agent calls Codex or another CLI, the final report MUST include a short transparency block: |
| 43 | |
| 44 | - **Planner/reviewer model:** the Claude/Fable agent that scoped or reviewed the work. |
| 45 | - **Executor model/profile:** the Codex profile or other CLI profile used. |
| 46 | - **Working directory:** repo path or scratch path. |
| 47 | - **Command shape:** enough of the command to audit routing, with secrets and tokens redacted. |
| 48 | - **Result:** PASS/FAIL/NOT RUN for tests or verification, plus the important stdout/stderr summary. |
| 49 | |
| 50 | Do not paste API keys, OAuth tokens, cookies, or full auth-bearing commands. Transparency is about routing and evidence, not leaking credentials. |
| 51 | |
| 52 | ## Default mixed-model handoff |
| 53 | |
| 54 | For implementation tasks that are eligible for mixed execution, use this sequence: |
| 55 | |
| 56 | 1. **Claude/Fable plans.** The owning agent reads the relevant docs/GitNexus context, defines scope, risks, and a Definition of Done. |
| 57 | 2. **Codex implements.** Codex receives the plan, constraints, impact/context evidence, and required tests. It writes code only inside the approved scope. |
| 58 | 3. **Claude/Fable reviews.** The owning agent reviews the Codex diff against the Definition of Done and PuPu conventions. |
| 59 | 4. **Codex fixes.** If review finds concrete issues, send Codex a bounded fix prompt rather than allowing a broad rewrite. |
| 60 | 5. **Owner verifies.** The owning agent reruns or reviews the required tests and reports any skipped checks explicitly. |
| 61 | |
| 62 | This handoff can be used as a **Codex implementation assistant** for most dev roles, but only `pupu-dev-backend` is approved for Mode B Codex-primary writes. For load-bearing surfaces such as chat-core and Electron, Codex may propose patches and run tests, but Claude/Fable remains the accountable writer/reviewer. |
| 63 | |
| 64 | ## Rollout rule |
| 65 | |
| 66 | Mode B stays a **single pilot (pupu-dev-backend)** until it passes the metrics below. Only then extend Mode C to QA broadly and Mode A to security/LLM expert. **Never extend M |
Related
6 picksType
- cavecrew-builderSurgical 1-2 file edit. Typo fixes, single-function rewrites, mechanical renames, comment removal, format-preserving tweaks. Hard refuses 3+ file scope. Returns caveman diff receipt. Use when scope…Jul 2026
- cavecrew-investigatorRead-only code locator. Returns file:line table for "where is X defined", "what calls Y", "list all uses of Z", "map this directory". Output is caveman-compressed so the main thread eats ~60% fewer…Jul 2026
- cavecrew-reviewerDiff/branch/file reviewer. One line per finding, severity-tagged, no praise, no scope creep. Output format `path:line: <emoji> <severity>: <problem>. <fix>.` Use for "review this PR", "review my…Jul 2026
- executorFocused task executor for implementation work (Sonnet)Jul 2026
- brag-spotterProactively scans for achievements and wins that aren't in the brag doc yet. Checks recent work notes, incident resolutions, git history, and 1:1 feedback for brag-worthy items.Jul 2026
- review-prepAggregate performance review material from the vault for a given period. Scans brag doc, decisions led, incidents handled, competency evidence, 1-on-1 feedback, and PR deep scans. Invoke via…Jul 2026
