dispatchlisted

# Dispatch The craft layer for handing a scoped task to a non-Claude model and getting a result back. Pick the right model tier for the task type, write a tight spec, run the dispatch tool, and escalate taste/client-facing/irreversible calls to the project owner before anything ships. ## Lanes - **Lane 1 — Claude work:** spawn Agent Teams teammates (Claude Sonnet / Opus / Haiku). Do NOT use this skill for Claude-to-Claude dispatch. - **Lane 2 — non-Claude work:** your dispatch script (see Setup below). ## Setup 1. Copy `SKILL.md` to `~/.claude/skills/dispatch/SKILL.md`. 2. Place `dispatch.py` (and optionally `dispatch_web.py`) in a stable directory on your PATH or reference by full path. 3. Open `dispatch.py` and register your models in `HTTP_MODELS` (or `CLI_MODELS` for a CLI-based coding tool). The keys are what you pass to `--model`; the values are the provider's model IDs. 4. Set two environment variables in your shell profile or agent config: - `DISPATCH_BASE_URL` — your provider's OpenAI-compatible base URL (e.g. `https://api.openai.com/v1`, a local Ollama URL, or a LiteLLM proxy). - `DISPATCH_API_KEY` — API key for that endpoint. 5. Verify: `echo "Write hello world in Python." | path/to/dispatch.py --model <your-model> --spec -` 6. For web-browsing mode (`dispatch_web.py`), register a tool-call-capable model in `CHAT_MODELS` and wire up your search provider in `_web_search()` — see the TODO comment there for Serper, Tavily, and SerpAPI options. ## Model Mat