The runner side of the Anvil runner relay — fangorn/anvil#298 (epic fangorn/anvil#145).

What

When a runner carries the inference label, anvil runner start now also polls for inference requests and serves them from its own local model, exactly like the CI job loop runs over the runner’s outbound connection. The Anvil server never makes an outbound model call — the inference has the same trust boundary as a CI job.

• src/runner/inference.rs — an inference poller (mirrors loop_runner’s CI poller): claims POST /api/v1/runners/:id/inference/claim, does the Anvil↔Ollama wire translation, calls local Ollama /api/chat, and reports the result to POST /api/v1/runners/:id/inference/:request_id/result.
  • Anvil internal → Ollama: messages (string and typed-block content, incl. tool_use / tool_result round-trips), tool defs → function tools, system/temperature/max_tokens options.
  • Ollama → Anvil: content / tool_uses / stop_reason / usage (mirrors Anvil.Agents.Providers.Ollama).
• loop_runner — spawns the inference poller when labels include inference (a runner can be CI-only, inference-only, or both).
• config — ollama_url (default http://localhost:11434).
• runner configure — --ollama-url flag; args grouped into ConfigureOpts (clears clippy too_many_arguments).

Verified end-to-end

Against a dockerized Anvil (this branch’s server code) + a local llama3.2:1b:

• runner configure --labels self-hosted,inference registers; runner start logs [inference] polling ….
• A routed agent run (AgentProviderConfig transport: routed → resolver → Routed provider) enqueues a request; the runner claims it, calls local Ollama, posts the result; the caller gets a real model response. DB shows the request completed, claimed by this runner.

Notes

• Non-streaming v1 (the request runs to completion; token streaming from the runner is a later refinement).
• cargo fmt / clippy -D warnings / cargo test (86) all clean.