hermes - 💡(How to fix) Fix [Bug] Kimi provider (`kimi-for-coding`) fails with HTTP 400 temperature error — need per-model temperature override support [18 comments, 18 participants]

Bug Report

Describe the bug

When using the Kimi provider (kimi-coding) with model kimi-for-coding via https://api.kimi.com/coding/v1, all API calls fail with:

HTTP 400: invalid temperature: only 0.6 is allowed for this model

The Kimi API requires temperature: 0.6 for kimi-for-coding / kimi-k2.5 models, but there is currently no way to set a per-model or per-provider temperature override in the Hermes config. The framework sends its default temperature value, which gets rejected.

To Reproduce

1. Configure Hermes with Kimi provider:

   model:
     default: kimi-for-coding
     provider: kimi-coding

2. Start the gateway and send any message
3. Gateway logs show repeated HTTP 400 errors

Expected behavior

Hermes should support per-provider or per-model temperature configuration so that providers with strict temperature requirements (like Kimi's 0.6) can work without code modifications.

Actual behavior

• All API calls fail with HTTP 400
• No config option to override temperature per provider/model
• Attempting to patch auxiliary_client.py to force temperature=0.6 breaks other provider logic

Suggested Fix

Add a temperature field to the provider or model config:

providers:
  kimi: |
    api: https://api.kimi.com/coding/v1
    temperature: 0.6   # per-provider override
    models:
      kimi-for-coding:
        temperature: 0.6   # or per-model override

Or add a default_temperature option at the provider level that gets passed to all API calls for that provider.

Additional context

• This is a known Kimi API limitation — it hard-rejects any temperature value other than 0.6
• Filed upstream issue: MoonshotAI/Kimi-K2#124
• Other frameworks have hit the same issue: anomalyco/opencode#11134
• Current workaround: switch to a different model (e.g., Xiaomi mimo-v2-pro), but this loses Kimi's coding capabilities
• Do NOT patch auxiliary_client.py or run_agent.py — it causes cascading failures

Environment

• Hermes version: latest (as of 2026-04-18)
• Provider: kimi-coding
• Model: kimi-for-coding
• Endpoint: https://api.kimi.com/coding/v1