duncanmcqueen opens Mike to self-hosted AI models

The fork makes it far easier to use locally run or third-party AI models without rebuilding the product.

duncanmcqueen is proposing a configurable model directory for Mike. A firm or product team can declare an OpenAI-compatible AI endpoint, whether it runs on its own hardware or is supplied by another provider, and retain the platform's existing reliability and response-handling behaviour.

Configurable models: add an approved compatible endpoint through deployment settings rather than a product change.
Tool-call recovery: recognise common cases where local models describe an action in text instead of returning it cleanly.
Reasoning cleanup: keep internal reasoning markup out of the visible answer.
Safer responses: use a complete response where a partial stream would make an action ambiguous.

The proposal also checks configurations and availability, with documentation for operators.