rjv-codex-ollama-subagentslisted
Install: claude install-skill rjvim/ai-skills
# Codex + local Ollama
Two ways to put a local Ollama model to work:
```text
(a) Separate Ollama-backed Codex processes WITH repo tools
main OpenAI Codex = orchestrator / reviewer
gemma-explorer = local read-only repo scout
gemma-worker = local scoped coding worker
(b) One-shot local text function, NO repo tools → see the last section
qwen2.5-coder:14b-instruct = fast checker when Codex supplies exact context
```
**Which mode:** if the model must read files or edit through the harness, use (a),
a local Codex process. If you can hand it a complete self-contained prompt and it needs
nothing else from the repo, use (b), the one-shot runner. If you find yourself wishing
a (b) run could "check the other file", the job needs repo tools: route it to (a).
Mode (a) runs as separate processes because native subagents spawned by an OpenAI
parent inherit its provider in Codex 0.155.1, so those roles cannot switch to Ollama.
## Setup
When the user asks to configure this machine (install the `~/.codex-ollama` profile),
or wants to run local Gemma agents now and Ollama is not yet serving with the tuned
flags, read [`SETUP.md`](SETUP.md). It holds the installer, the `ollama serve`
flags and the expected model-load evidence.
## Recommended model routing
| Work | Agent |
|---|---|
| Planning, review, final decision | main OpenAI model |
| Bounded repository exploration | `gemma-explorer` (`gemma4:26b`) |
| Scoped edits and tests | `gemma-worker` (`gemma4:26b`) |
| Fa