feat(ce-work): add Codex delegation mode#328
Conversation
Add an optional "Codex Delegation" section to ce:work, parallel to the existing Swarm Mode section. When triggered by saying "use codex for this work", ce:work delegates implementation tasks to codex exec while keeping planning, review, and git operations in the current agent. Uses stdin piping for long prompts, per-task fallback on failure, and graceful degradation when Codex CLI is not installed.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f5b9194975
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
- Remove GH token injection from prompt (not needed for code writing, avoids credential exposure in model input) - Replace shell chaining (cd && echo | codex) with discrete steps using a temp file, compliant with AGENTS.md shell rules
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: a70a23fd7b
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
|
@mvanhorn love this. Is it possible to refactor into a work-beta skill for time being while we test it? |
|
@tmchow funny enough I had it in beta, but to simplify my PR I took it out. let me get it in there |
|
Another interesting thing I wonder is how we might add a config for this so it would use Claude vs codex by default for this phase. But it would have to be careful which coding env you're in and respect the flag. Just thinking out loud. |
… and enforcement to ce:work - ce:plan-beta Phase 1.1b: detect Codex delegation intent from user signals - ce:plan: add Execution Posture step for Codex delegation detection - ce:work: recognize plan-level Execution target signals as delegation trigger - ce:work: add anti-rationalization enforcement to Codex Delegation Workflow Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…ction - Use mktemp for unique prompt files (prevents cross-task races) - Pass prompt as argument instead of stdin redirection (AGENTS.md compliance) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 71bce5cc15
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
interesting! I've only used CE in CC. |
Even just in CC, it would make sense possibly for a setting you set with the |
|
Sounds good - I'll keep it on-demand for now and we can add a |
|
@mvanhorn some thoughts after spending time reviewing and thinkign about this. Environment DetectionThe skill runs across harnesses (Claude Code, Codex, Gemini CLI, etc.). If someone runs Strategy IntegrationThe Codex delegation section is bolted on after Swarm Mode as a standalone workflow, but Key questions the skill needs to answer:
Suggestion: wire delegation into the existing strategy selection as a task-level modifier rather than a separate section. The strategy (inline/serial/parallel) still applies, but the implementation step within each task delegates to Codex instead of executing directly. Plan AbstractionThe Hardcoded Modelyou use Mixed AttributionPhase 4 requires attribution like Error Threshold"One failure does not disable Codex for remaining tasks" is fine for transient errors, but not for systemic issues (API down, auth expired, rate-limited). Consider a consecutive-failure threshold (e.g., 3 failures in a row → disable delegation for remaining tasks)? Beta SkillPer our earlier discussion, refactor this into a Already flagged by automated review (still open)
|
…, env guard, error threshold - Rename "Codex Delegation" to "External Delegate Mode" for tool-agnostic plans - Replace `Execution target: codex` with `Execution target: external-delegate` in plans - Add environment guard to detect sandbox recursion (CODEX_SANDBOX, CODEX_SESSION_ID) - Wire delegation into existing strategy selection as task-level modifier - Remove hardcoded model flag - use delegate's default - Add consecutive failure threshold (3 failures disables delegation) - Add mixed-model attribution guidance for Phase 4 - Pipe prompt via stdin instead of argv expansion (ARG_MAX safety) - Remove shell chaining/redirection from examples (AGENTS.md compliance) - Add ce:work-beta skill wrapper with disable-model-invocation for safe testing - Remove credential embedding from prompt (pass auth via env vars) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 289d14658b
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
|
Took a stab at this - I like your thoughts. Heads up the description for Compound Engineering on GH says a Claude Code skill. Addressed all points in 289d146: Environment Detection - Added sandbox guard that checks Strategy Integration - Rewrote the section to frame delegation as a task-level modifier on the existing strategy (inline/serial/parallel), not a parallel workflow. The strategy still applies - delegation just changes how individual tagged tasks execute. Plan Abstraction - Replaced Hardcoded Model - Removed Mixed Attribution - Added guidance for single-delegate, single-standard, and mixed-model scenarios in Phase 4. Error Threshold - Added 3-consecutive-failure circuit breaker that disables delegation for remaining tasks. Beta Skill - Created Shell compliance + stdin - Removed all shell chaining/redirection from examples. Prompt goes via stdin, not argv. Credentials via env vars, not prompt text. |
|
@codex provide a final review, focused on the latest changes. also check if the repo and various READMEs need to be updated. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 289d14658b
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| This is the beta variant of `ce:work` that includes the External Delegate Mode feature for testing. Once stable, the delegation section will be promoted to `ce:work` and this beta skill will be removed. | ||
|
|
||
| **All behavior is identical to `ce:work`** - this skill simply re-includes the full `ce:work` content with the External Delegate Mode section active. | ||
|
|
||
| To use: run `/ce:work-beta [plan file]` instead of `/ce:work [plan file]`. |
There was a problem hiding this comment.
All of our loaders/targets only register and copy each skill’s own skills/<name>/SKILL.md (src/parsers/claude.ts, src/targets/codex.ts), so invoking /ce:work-beta will load only this 20-line wrapper. Because this file does not actually contain or link in the ce:work instructions in a supported way, anyone testing the beta command will miss the real execution workflow and the new delegation steps entirely.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
@mvanhorn in order to make the beta skill work, you have to duplicate the entire skill which incorporates your changes. You can't just do what you did. Also by having an entire encapsulated skill, it makes it easier to promote the skill to stable.
| External delegation activates when any of these conditions are met: | ||
| - The user says "use codex for this work", "delegate to codex", or "delegate mode" | ||
| - A plan implementation unit contains `Execution target: external-delegate` in its Execution note (set by ce:plan-beta or ce:plan) |
There was a problem hiding this comment.
The beta rollout framework in docs/solutions/skill-design/beta-skills-framework.md only works if the stable skills stay unchanged while users opt into the -beta variant, but these trigger phrases now enable delegation from normal /ce:work sessions, and stable ce:plan also stamps Execution target: external-delegate into plans (plugins/compound-engineering/skills/ce-plan/SKILL.md:154-156). That means a regular /ce:plan -> /ce:work flow can hit the experimental delegate path without ever choosing /ce:work-beta, which defeats the safe side-by-side trial and can surprise users who merely mention Codex or token conservation.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
@mvanhorn address this one too. see docs/solutions/skill-design/beta-skills-framework.md if you want to see how i've structured how beta skills work.
2 participants
Summary
Adds an optional "Codex Delegation" section to
ce:work, parallel to the existing Swarm Mode section. When the user says "use codex for this work", implementation tasks are delegated tocodex execwhile planning, review, and git operations stay in the current agent.Why
Users on Max20 plans burn significant context on code implementation. Codex Pro plans include free compute that can handle the writing work. This keeps Claude focused on planning and review.
Changes
One file changed:
plugins/compound-engineering/skills/ce-work/SKILL.md(+53 lines)The new section follows the same structure as Swarm Mode:
Technical details from a battle-tested implementation (open-source-contributor runs this pattern in production):
command -v codexfor health check (graceful fallback if not installed)echo "$PROMPT" | codex exec ... -to pipe long prompts via stdin (avoids shell quoting issues)model_reasoning_effortconfig key (verified from~/.codex/config.toml)Relation to PR #223
PR #223 (lfg-ext) replaces the entire work pipeline with external delegates in parallel worktrees. This PR adds a mode within the existing
ce:workflow - same relationship as Swarm Mode has to standard mode. The two are complementary.Testing
bun run release:validatepasses (29 agents, 45 skills, 1 MCP server - no count changes)codex execinvocation pattern works in production via open-source-contributor commit 588cf99This contribution was developed with AI assistance (Claude Code).