This repository is now the transition workspace for a globally installed two-agent orchestration system.
The target model is:
- shared workflow contracts in
workflow/core/ - Claude-specific runtime/control behavior in
workflow/claude-control/ - Codex-specific worker/verifier behavior in
workflow/codex-control/ - integrated end-to-end behavior and examples in
orchestration/
The ownership contract does not live only in this root README anymore.
The canonical files are:
workflow/core/contracts/agent-ownership.mdworkflow/core/contracts/routing-matrix.md
Those files define:
- who owns which phase
- who owns which task types
- when dual-pass review is justified
- what must stay shared versus tool-specific
README.md is now the high-level map for the repo.
workflow/
core/ shared contracts, schemas, migration inventory
claude-control/ Claude-specific control/runtime layer
codex-control/ Codex-specific worker/verifier layer
orchestration/ integration docs, examples, smoke tests, rollout notes
scripts/ root compatibility wrappers
At a high level:
- Claude owns ambiguity, planning, orchestration, UX, MCP/external-tool work, and final user communication.
- Codex owns constrained backend execution, refactors, debugging, adversarial review, and verification.
The detailed matrix lives in workflow/core/contracts/routing-matrix.md.
You currently have two globally installed setup repos on this machine:
~/Projects/claude-code-setup~/Projects/codex-setup
They overlap conceptually but differ correctly in runtime assumptions. This repo is the staging area where the shared source of truth gets extracted without forcing a premature merge of the two global systems.
Owns shared contracts:
- agent ownership
- routing matrix
- shared migration inventory
- future shared schemas, checklists, and reusable fragments
Owns Claude-specific behavior:
- always-on orchestration rules
- runtime integration into Claude Code
- future prompts/hooks/settings that make Claude follow the shared contract by default
- future installable plugin/package landing zone under
workflow/claude-control/plugin/ - current staged copy of the existing Claude setup under
workflow/claude-control/plugin/source/
Owns Codex-specific behavior:
- worker/verifier consumption of the shared contract
- Codex-native prompts, wrappers, and install surfaces
- future migration from the current
codex-setuplayout - future installable plugin/package landing zone under
workflow/codex-control/plugin/ - current staged copy of the existing Codex setup under
workflow/codex-control/plugin/source/
Owns integration-level material:
- examples
- smoke tests
- end-to-end behavior notes
- rollout and adoption docs
- canonical orchestration scripts under
orchestration/scripts/ - canonical orchestration agent definitions under
orchestration/agents/
It is not the source of truth for ownership policy.
The canonical implementations now live under orchestration/scripts/.
The root wrappers remain as compatibility entrypoints:
scripts/codex-preflight.shscripts/codex-worker.shscripts/codex-verifier.sh
These delegate to the orchestration-owned sources so the repo remains directly usable during the transition.
Codex is called directly via shell scripts — no agent .md wrappers.
Claude builds a self-contained prompt and calls the script via Bash:
# Worker (implementation, analysis, refactors)
scripts/codex-worker.sh --prompt-file /tmp/prompt.txt --output /tmp/result.txt
# Verifier (independent verification)
scripts/codex-verifier.sh --requirements-file /tmp/requirements.txt --output /tmp/result.txtFor parallel Codex tasks, dispatch multiple Bash calls simultaneously.
Use subscription auth for both tools.
- Requires an active Claude subscription
- Prefer ChatGPT subscription sign-in
- Check status with
codex login status - Start login with
codex login - Treat
OPENAI_API_KEYas an optional fallback only
These checks are currently expected to pass:
claude --help
codex login status
bash -n scripts/codex-preflight.sh scripts/codex-worker.sh scripts/codex-verifier.sh
scripts/codex-worker.sh --help
scripts/codex-verifier.sh --help