Bug: OpenCode Go bridge layer incompatible with kimi-k2.6 tool calls #26331

Bug Report: OpenCode Go bridge layer incompatible with kimi-k2.6 tool calls

Environment

• Client: MyAgents (Claude Agent SDK based)
• Provider: OpenCode Go
• Model: kimi-k2.6 (via Moonshot AI)
• API Protocols tested: OpenAI-compatible (/zen/go/v1) and Anthropic (/zen/go)

Problem Description

When using kimi-k2.6 through OpenCode Go's bridge layer, tool calls fail due to incompatible message formatting. The issue manifests differently depending on which protocol is used:

1. OpenAI Protocol (/zen/go/v1) - Original Issue

Error:

API Error: 400 {"type":"error","error":{"type":"invalid_request_error",
"message":"Error from provider: Provider returned error"}}

Root cause (from provider logs):

thinking is enabled but reasoning_content is missing in assistant tool call message at index 2
provider_name: Moonshot AI

Analysis:

• kimi-k2.6 has thinking enabled by default
• When the AI makes tool calls in subsequent conversation turns, the bridge layer fails to properly forward the reasoning_content field in the message history
• Moonshot AI rejects the request because the thinking content is missing from previous assistant messages

2. Anthropic Protocol (/zen/go) - Alternative Test

Error:

Invalid request: function name is invalid, must start with a letter 
and can contain letters, numbers, underscores, and dashes
provider_name: Moonshot AI

Analysis:

• The Anthropic protocol bridge has a different issue with tool name formatting
• Tool names passed through the bridge don't meet Moonshot's validation requirements

Steps to Reproduce

For OpenAI Protocol Issue:

1. Configure OpenCode Go with OpenAI protocol endpoint: https://opencode.ai/zen/go/v1
2. Use model kimi-k2.6
3. Start a conversation that triggers tool use (e.g., file operations, bash commands)
4. After the first tool call response, send a follow-up message
5. The 400 error occurs on the second turn when the conversation history includes tool results

For Anthropic Protocol Issue:

1. Configure OpenCode Go with Anthropic protocol endpoint: https://opencode.ai/zen/go
2. Use model kimi-k2.6
3. Attempt any tool call
4. The request fails immediately with function name validation error

Expected Behavior

• kimi-k2.6 should work seamlessly through OpenCode Go bridge layer
• Tool calls should succeed across multiple conversation turns
• Thinking/reasoning content should be properly preserved and forwarded

Actual Behavior

• Both protocol bridges fail with kimi-k2.6 tool calls
• OpenAI bridge: Missing reasoning_content in message history
• Anthropic bridge: Invalid tool name format

Workaround

Using Moonshot AI's native API endpoint (https://api.moonshot.cn/anthropic) works correctly with kimi-k2.6 and tool calls, confirming this is specifically an OpenCode Go bridge layer issue.

Additional Context

• kimi-k2.5 (without thinking) works fine through OpenCode Go
• The issue is specific to kimi-k2.6's thinking capability combined with tool use
• This appears to be a bridge layer message format conversion issue

Request

Please investigate and fix the bridge layer to properly handle:

1. reasoning_content field preservation in OpenAI protocol bridge
2. Tool name formatting in Anthropic protocol bridge

Thank you!