kimi (K3/K3-256K) delegation fails: wrong endpoint path + temperature must be 1

[Demon673]

Environment

• pi coding agent, @getpipher/vision 0.5.2
• Primary model: deepseek-v4-flash (text-only) → DELEGATE mode
• Vision model: kimi k3-256k (via the built-in kimi-coding provider, OAuth subscription)

Problem 1: wrong endpoint path (404)

callVisionModel requests ${visionModel.baseUrl}/chat/completions.

The built-in kimi-coding provider has baseUrl: https://api.kimi.com/coding (it uses the Anthropic Messages API, /v1/messages). Concatenating gives https://api.kimi.com/coding/chat/completions → 404.

Kimi's OpenAI-compatible endpoint is https://api.kimi.com/coding/v1/chat/completions (with the OAuth access token as Bearer).

Workaround: users must register a separate provider in models.json with baseUrl: https://api.kimi.com/coding/v1 + api: openai-completions, which is non-obvious and breaks the "just point at kimi-coding/k3" expectation.

Problem 2: hardcoded temperature: 0 is rejected by kimi (400)

lib/delegate.ts hardcodes temperature: 0. Kimi's OpenAI-compatible endpoint rejects it for k3/k3-256k:

400 invalid temperature: only 1 is allowed for this model

This makes kimi vision models completely unusable through the extension. Local fix: drop the temperature field (kimi defaults to 1).

Suggested fixes

1. Omit temperature (or make it configurable) in callVisionModel — kimi requires 1, and temperature is irrelevant for image description anyway.
2. For the endpoint: either document the baseUrl-with-/v1 workaround for kimi, or make the path join smarter (e.g. respect a baseUrl that already contains the API prefix).

Happy to submit a PR if useful.