nightshift/docs/devlog/phase17.md

Phase 17 Devlog: Local Model Backend

Implemented

• Added first-class backend: ollama agent config support.
• Required model for Ollama agents.
• Kept backend: command unchanged.
• Reused the existing prompt bundle for Ollama.
• Invoked Ollama as ollama run <model> with prompt input on stdin.
• Persisted Ollama responses through the same agent artifact format.
• Added tests with mocked subprocess calls so Ollama is not required.

Decisions Made

• Ollama is implemented as a local subprocess backend instead of an HTTP API wrapper.
• Missing Ollama executable returns a failed agent invocation artifact rather than crashing.
• Backend artifacts remain comparable across command and Ollama agents.

Notes

• Real model quality and model availability are user environment concerns; tests do not require a running Ollama daemon.