Shared server-side model-call pipeline (Ollama client, role->model routing, prompt loader + canon-log digest, §10 call logging, retry->typed 502) wired through /dm/narrate as the first role. Non-streaming; server returns raw prose, client extracts tags. Two phases: pipeline foundation, Narrator online. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>