Error Writing Large Files (at least with Kimi K2.5)

[marcoschwartz]

Description

Got an error while using the new Kimi K2.5 model - Invalid input for tool write: JSON parsing failed

Plugins

No response

OpenCode version

1.1.39

Steps to reproduce

Just use Kimi K2.5 with any provider and ask to edit/write a large file

Screenshot and/or share link

No response

Operating System

macOS 26

Terminal

iTerm2

[github-actions]

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

• Qwen3 Coder on OpenCode Zen: Tool call metadata being printed to chat instead of actually being called #10855: Qwen3 Coder on OpenCode Zen: Tool call metadata being printed to chat instead of actually being called (similar tool execution failures with specific models)
• Added Kimi API key invocation and the new model Kimi-k2.5 #10850: Added Kimi API key invocation and the new model Kimi-k2.5 (existing issues with the Kimi K2.5 model)
• Google Models cannot call tools anymore #10953: Google Models cannot call tools anymore (broader pattern of tool calling failures across models)
• Add metadata awareness and data type validation/blacklists for LLM tool execution #10820: Add metadata awareness and data type validation/blacklists for LLM tool execution (addresses root cause of data type validation issues)

Feel free to ignore if none of these address your specific case.

[ideallove]

I also encountered the same issue. I was working on Agent Skill yesterday and today. Yesterday it was completely correct, but today it failed.
The SKILL I was using involved generating JSON, HTML, and escape characters, and the same problem occurred with the write component.
I was using the Windows platform, with LLM being Claude Opus 4.5 and Minimax M2.1.
Just now, I removed OpenCode and Desktop for version 1.1.39. After reverting to version 1.1.36, there were no abnormalities.

Therefore, the temporary solution is to revert to version 1.1.36.
I am currently attempting to analyze and review the updates from 1.1.37 to 1.1.39.
Thank you to the developers for their work.

[ideallove]

@marcoschwartz I have fixed the problem I encountered. Maybe your problem has also been resolved.

[Mithrillion]

I think equally frustrating is that when you encounter this issue with Kimi K2.5, the agent just stops instead of retrying or continuing the task with a different approach, like in typical tool failures. And if it is in a subagent and the subagent fails, you often have to scrap the whole run, because you cannot simply ask it to continue.

[ideallove]

Why This Causes invalid Tool Errors?

• Without maxOutputTokens: The AI SDK doesn't send a max_tokens parameter to the provider
• Provider uses default limit: Many providers default to a low value (e.g., kimi K2.5 default 32768 from https://platform.moonshot.ai/docs/guide/kimi-k2-5-quickstart)
• Large tool calls get truncated: When writing files with substantial content, the JSON output exceeds the default limit
• JSON becomes incomplete: Output is cut mid-string, e.g., {"content": "# -*- coding... (missing closing "})
• Parse error occurs: JSON Parse error: Unterminated string
• Fallback to invalid tool: The experimental_repairToolCall handler marks it as invalid

It is more likely to trigger in the subagent. In conclusion, hardcoding will definitely cause problems.

Apart from reverting to the previous version, adding a detection for JSON parsing might make the system more stable.