Send a one-shot prompt to a local Ollama model and return the response.

Use for any handoff where the cloud model's full reasoning isn't needed: drafts, boilerplate, simple extractions, formatting, quick lookups. Runs on the user's own GPU and consumes no cloud-LLM usage.

Args: prompt: The task / question. model: Override the default model. system: Optional system prompt to shape behavior.

Multi-turn chat against a local Ollama model.

Use when the handoff needs more than one turn of context. messages is a list of {"role": "user"|"assistant"|"system", "content": str}.

Summarize a block of text using the local model.

Cheap offload for long files, logs, transcripts, or docs the cloud model doesn't need to fully ingest. Returns a concise structured summary.

Args: text: The content to summarize. Can be very long (context window is configurable). focus: Optional focus hint, e.g. "errors and stack traces" or "API surface only".

Quick first-pass code review using the local coder model.

Catches obvious bugs, style issues, and risky patterns. Use as a cheap pre-filter before asking the cloud model for a deeper review.

Args: diff_or_code: A unified diff or a code block to review.

Draft a conventional-style commit message from a diff using the local model.

Cheap and fast — good for routine commits where the cloud model's analysis isn't needed.

Args: diff: The output of git diff --staged or similar.

Extract specific information from a text block using the local model.

Good for pulling structured data out of unstructured text — function names, URLs, error messages, TODO comments, etc.

Args: text: The source text. what_to_extract: What to pull out, e.g. "all function definitions" or "every URL in the file".

List the Ollama models available locally.

Return the server's effective configuration (model, context size, etc.).