Codex + local Ollama

Two ways to put a local Ollama model to work — pick by whether the job needs the repo:

(a) Codex native subagents WITH repo tools  → most work; the bulk of this skill
    main Codex GPT-5.5 = orchestrator / reviewer
    qwen-explorer / gemma-explorer = local read-only repo scouts
    qwen-worker / gemma-worker     = local scoped mechanical edit workers
    optional OpenAI mini           = fast cheap explorer when latency matters

(b) One-shot local text function, NO repo tools  → see the last section
    complete prompt in → local Ollama text out (draft-from-spec, tests, classify)