Bug: 400 Error (Extra inputs not permitted) when sending 'reasoning' field in chat history for kimi-k2.6

[totoDoP]

Description

I am encountering a 400 Bad Request error when using the kimi-k2.6 model via the OpenCode Go proxy. The error occurs
specifically on the second or third turn of a conversation (once history is being sent back to the model).

Error Message:
400 Error from provider: Extra inputs are not permitted, field: 'messages[X].reasoning'

Diagnosis:
My client preserves the reasoning/thinking blocks from previous turns and sends them back in the messages array
within a reasoning field to maintain context.

It appears the OpenCode Go API gateway has a strict schema validation that rejects any fields not explicitly defined
in the standard OpenAI Chat Completions schema. I have verified that connecting directly to the Moonshot AI API
(bypassing OpenCode Go) works perfectly, as the Moonshot API ignores the extra reasoning field.

Plugins

N/A

OpenCode version

OpenCode Go API

Steps to reproduce

1. Use the kimi-k2.6 model via the OpenCode Go proxy.
2. Start a conversation and let the model generate a reasoning block.
3. Send a second message to continue the conversation.
4. The request fails with a 400 error because the history now contains the reasoning field, which is rejected by the
   gateway validation.

Screenshot and/or share link

No response

Operating System

Windows 11

Terminal

pi.dev

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Bug: OpenCode Go bridge layer incompatible with kimi-k2.6 tool calls #26331: Same model (kimi-k2.6) via OpenCode Go bridge — reasoning_content missing from assistant messages in multi-turn conversations, causing 400 errors from Moonshot AI
• DeepSeek V4: reasoning_content 400 error persists on v1.14.31 — complete fix exists in unmerged PRs #25311: Same root cause with DeepSeek V4 — reasoning_content not preserved in conversation history across tool-call turns, causing 400 errors
• DeepSeek thinking mode: reasoning_content not passed back for tool call turns, causing 400 errors #24722: Nearly identical issue — reasoning_content not passed back for tool call turns when thinking mode is enabled, causing 400 errors

[mrlunchbox777]

I've seen this same issue if it doesn't meet the exact schema. It'd be helpful to optionally strip this extra metadata. Specifically, I was able to reproduce with a custom provider (@ai-sdk/openai-compatible with an api, gpt-5 and gpt-5.5, that failed with any extra fields). GET /models and direct POST /chat/completions would work, but once it started using the OpenCode request path it would fail (Invalid request format).