V3 deterministic local-inference capability pack. By-value connector composition (service_kind=cli, no separate service ingredient): the local ollama CLI runs a prompt against a locally-installed model and returns the model's reply as captured text, via the model.run catalog operation. Nothing leaves the machine — the prompt and the reply are processed by a model running on the paired recued-server, so this is the no-egress path for summarizing / rewriting / classifying private content that must never reach a hosted provider. Read-tier and approval=never: a local inference has no durable side effect (it writes nothing, sends nothing) — the reply is captured to the op result's `stdout` field (a string the next step consumes directly). Requires the ollama binary on PATH and the target model already pulled (ollama pull <model>); a missing model fails the run rather than triggering a network download mid-recipe. Distinct from the BYOK / free-pool AI substrate (ai-* functions): this is a raw local subprocess, deliberately chosen when the value proposition is "the bytes stay on this machine," not structured-JSON contracts. Part of the local developer / privacy toolkit.