hermes - 💡(How to fix) Fix [Bug] opencode-go + kimi models: cannot specify both 'thinking' and 'reasoning_effort' (HTTP 400)

Bug Description

When using a Kimi/Moonshot model (e.g. kimi-k2.6) through the opencode-go provider with agent.reasoning_effort set to a non-empty value, the API request includes both thinking and reasoning_effort parameters. The Moonshot API rejects this with:

HTTP 400: Error from provider: cannot specify both 'thinking' and 'reasoning_effort'

These two parameters are mutually exclusive on the provider side, but Hermes sends both simultaneously.

Steps to Reproduce

1. Configure opencode-go as the provider with a Kimi model:

model:
  default: kimi-k2.6
  provider: opencode-go
  base_url: https://opencode.ai/zen/go/v1
  api_mode: chat_completions

1. Set a global reasoning effort:

agent:
  reasoning_effort: medium

1. Send any message to the agent.

2. The API call fails immediately:

BadRequestError [HTTP 400]
Provider: opencode-go  Model: kimi-k2.6
Endpoint: https://opencode.ai/zen/go/v1
Error: HTTP 400: Error from provider: cannot specify both 'thinking' and 'reasoning_effort'

Expected Behavior

Hermes should send either thinking or reasoning_effort, never both. When reasoning_effort is set, Kimi/Moonshot models should receive it as a top-level reasoning_effort kwarg and the thinking parameter should be omitted, or vice versa.

Actual Behavior

Both thinking and reasoning_effort are included in the same request body, causing a non-retryable 400 error that aborts the session.

Root Cause

In agent/transports/chat_completions.py, the build_kwargs() method has separate code paths that add reasoning_effort (top-level kwarg for Kimi/Moonshot models, per #21577) and thinking (for thinking-capable models when reasoning_effort is configured). These two paths are not mutually exclusive, so both parameters end up in the request.

The fix should ensure that when reasoning_effort is sent as a top-level kwarg for Kimi/Moonshot, the thinking parameter is not also added.

Workaround

Set reasoning_effort to empty in config and use the /reasoning slash command per-session for non-Kimi models:

agent:
  reasoning_effort: ''  # disable global default

Environment

• Provider: opencode-go (https://opencode.ai/zen/go/v1)
• Model: kimi-k2.6
• Config: agent.reasoning_effort: medium

Related Issues

• #21577 — Added top-level reasoning_effort support for Kimi/Moonshot and opencode-go (likely introduced or exposed this conflict)
• #21256 — Feature request for per-model reasoning_effort defaults (would also help avoid this)