This file is created once for this project by the main agent. Edit the choices below before delegation; the main agent and controller must preserve your edits and use them before falling back to the bundled skill defaults.
| Role |
Preferred model |
Fallback |
Reasoning effort |
Project notes |
| Controller subagent (outer runtime) |
<choose an available medium-capability model> |
<another available medium-capability model> |
high |
Do not use a cheap/lite/flash tier for controller orchestration. |
| Shard role |
Preferred cmd --model |
Fallback |
Use for |
| Cheap read-only |
deepseek/deepseek-v4-flash |
<account default> |
Small investigation, summaries, log triage |
| Research / synthesis |
<choose an available strong model> |
<account default> |
Ambiguous investigation, long-context evidence |
| Code implementation |
gpt-5.3-codex |
<account default> |
Multi-file edits, tests, refactors |
| Code review |
<choose an available strong coding model> |
<account default> |
Correctness, dependency tracing, diff review |
| Architecture / planning |
<choose an available strong reasoning model> |
<account default> |
Broad design or risky decisions |
- Allowed providers or model IDs:
- Disallowed providers or model IDs:
- Maximum cost / latency preference:
- Prefer account default when:
- Additional routing notes:
- The controller must use the Controller row's preferred or fallback medium-capability model.
- Before selecting a Command Code model, run
cmd --list-models when availability is uncertain.
- If a selected model is unavailable or denied, use this file's fallback; otherwise omit
--model to use the account default.
- Keep shard prompts narrow. A low-cost model that repeatedly retries is not low cost.