One agent by default. Premium reasoning only where failure is expensive.
Codex Lean is a small, project-scoped routing policy for Codex. It keeps routine work on GPT-5.6 Luna, escalates coherent difficult work deliberately, reserves Sol High for costly failure modes, and prevents recursive or overlapping agent teams.
It is designed to improve accepted work per credit without weakening review or testing.
Codex Lean does not add another orchestrator, mandatory advisor, provider bridge, daemon, credential, or fixed multi-agent graph.
YOUR TASK
|
Luna High root
classifies risk
|
+-----------------+-----------------+
| | |
routine work deeper work costly failure
root owns Luna XHigh owns Sol High owns
| | |
+-----------------+-----------------+
|
frozen independent review
Most tasks stay in the root. Delegation is an exception with a measurable reason.
| Work | Model | Behavior |
|---|---|---|
| Routine deterministic work | Luna High root | no delegation |
| Non-trivial coherent implementation | Luna XHigh worker | one implementation owner |
| Financial, security, permissions, destructive data, migration, deployment, or concurrency risk | Sol High worker | fail closed if unavailable |
| Bounded exploration | Luna High explorer | read-only |
| Routine frozen review | Luna XHigh reviewer | independent and read-only |
| Critical frozen review | Sol High reviewer | independent and read-only |
The project cap is two child threads with depth one. Every routed child receives a self-contained brief with fork_turns = "none". Children never create descendants.
Add the marketplace and plugin:
codex plugin marketplace add HenkDz/Codex-Lean
codex plugin add codex-lean@codex-leanStart a new Codex task, then run:
/codex-lean setup
Codex previews the project changes before creating them. Existing model defaults, repository governance, permissions, and unrelated configuration are preserved unless you explicitly approve a conflict.
After setup, work normally. Repository guidance and custom-agent descriptions provide the routing policy.
Useful commands:
/codex-lean status
/codex-lean optimize this project
/codex-lean update
/codex-lean remove
/codex-lean no subagents — fix this isolated warning
Setup is project-scoped and version-controlled:
.codex/config.toml
.codex/agents/lean-explorer.toml
.codex/agents/lean-worker.toml
.codex/agents/critical-worker.toml
.codex/agents/lean-reviewer.toml
.codex/agents/critical-reviewer.toml
AGENTS.md
- Optimize accepted-change cost, including retries and review corrections—not raw token count alone.
- Do not pay an orchestration tax on a task one agent can finish.
- Route by cost of error, not by file count or impressive model names.
- Keep one implementation owner.
- Use fresh child context instead of cloning the full conversation.
- Freeze before review; any later edit invalidates the review.
- Never select Max or Ultra automatically.
- Never claim a child route worked unless runtime evidence confirms it.
- Subagents can increase raw token use even when cheaper models reduce weighted credits.
- Custom-agent configuration does not prove which model a live child actually used.
- Parent runtime permission choices can override a child sandbox request.
- Model pricing, availability, identifiers, and Codex routing behavior can change.
- A repository's stricter engineering or security rules always win.
Codex Lean fails closed when a critical route is unavailable. It never silently performs high-risk work with a cheaper inherited model.
The bundled checker is read-only:
python3 plugins/codex-lean/scripts/check_project.py /path/to/projectIt validates the routing ladder, thread limits, role boundaries, and project marker. It explicitly does not claim live model identity.
python3 -m unittest discover -s tests -vCodex Lean uses no network service, provider bridge, API key, or credential store. See SECURITY.md.
MIT