qwen-executorlisted

Delegate the current task to the local model (via the LiteLLM proxy on port 4000, default alias `fast-general`) using the stepped execution protocol below. The target alias is configurable via the `OLLAMA_AGENT_MODEL` env var on the MCP server — see `~/Code/claude-code-config/models.yaml` for the full roster and aliases. The skill name retains "qwen" for historical continuity; the tool names (`qwen_start`/`qwen_continue`) are unchanged. Do not attempt to answer the task yourself. Do not reason about it. Drive the loop immediately and return the final result verbatim. ## Tool Names The Ollama-agent MCP server is registered under different namespaces depending on context: - **Cowork (plugin context):** `mcp__plugin_shane-config_ollama-agent__qwen_start` / `mcp__plugin_shane-config_ollama-agent__qwen_continue` - **Claude Code (standalone):** `mcp__ollama-agent__qwen_start` / `mcp__ollama-agent__qwen_continue` Check which tools are available in your current session and use whichever namespace is present. If neither is available, use the fallback below. ## Stepped Execution Protocol 1. Call `qwen_start` (with the correct namespace prefix for your context) with `task`, `skill`, and `context` parameters. 2. Parse the JSON response: - `status: "done"` → return `result` to the user. Stop. - `status: "running"` → note the `session_id` and `step`, then call `qwen_continue` with that `session_id`. - `status: "error"` → surface the `result` as an error. Stop. 3. Repeat st