Kimi k2.5 has issues with tool calling

[kjj9]

Description

Kimi k2.5 has issues with tool calling:

⚙ invalid [tool=bash, error=Invalid input for tool bash: JSON parsing failed: Text: {"command":":1.
Error message: JSON Parse error: Unterminated string]

~ Searching content...
Model tried to call unavailable tool 'invalid'. Available tools: bash, read, glob, grep, webfetch.
Invalid request Error

Provider: moonshot sub

I cant tell if it's issue with model of provider or OC.

Plugins

No response

OpenCode version

1.3.13

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

macOS 26.3.1 (a)

Terminal

No response

[github-actions]

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

• Regression after 1.3.12: tool-call failures (PowerShell + malformed JSON behavior) on official Kimi provider #20577: Regression after 1.3.12: tool-call failures (malformed JSON behavior) on official Kimi provider — very similar symptoms with Kimi tool-call failures and JSON parsing errors

[fernandodamaso]

i"m having the same problem, #20658
And in the previous issue i've opened regarding this (and that was closed cause it was badly worded) two more people said they were having the same problem as well.

[rekram1-node]

I havent managed to reproduce

[fernandodamaso]

@rekram1-node From the tests i'm doing here, the problem is regarding the "reasoning_content" being lost, and that is causing of sorts of tool call errors, goose was having this problem and they fixed it aaif-goose/goose#6962

GPT 5.4 explaining:
"Kimi thinking + tools can work properly if the client/proxy preserves and replays reasoning_content on later tool rounds. Moonshot explicitly requires that for multi-step tool use.
The problem is that OpenCode currently has a bug where reasoning_content is handled on the immediate turn but then stripped when that assistant message is replayed in history.
This is not unique to OpenCode; multiple tool stacks hit the same failure, and the common fix is always the same: preserve reasoning_content in assistant messages with tool calls."

[rekram1-node]

@fernandodamaso that literally cannot be the issue if you are using the kimi for coding plan :)

We use the messages api for kimi for coding plan actually and reasoning is being persisted and sent back and forth (reasoning_content isnt a thing for anthropic compat endpoints, thats for openai compatible ones).

You can verify this by running opencode through an http proxy and sniffing all requests and responses.

[rekram1-node]

Examples:

[caffeinated92]

Sometimes I got this too when I use kimi-for-coding:
⚙ invalid [tool=read, error=Invalid input for tool read: JSON parsing failed: Text: {"filePath":": [.
Error message: JSON Parse error: Unterminated string]

→ Skill ": 0.028655, "
Skill ": 0.028655, " not found. Available skills: bmad-testarch-test-design, bmad-create-prd, bmad-shard-doc, bmad-testarch-trace, bmad-code-review, bmad-market-research, bmad-create-story, bmad-quick-dev, bmad-edit-prd, bmad-review-adversarial-general, bmad-generate-project-context, bmad-architect, bmad-testarch-framework, bmad-advanced-elicitation, bmad-analyst, bmad-review-edge-case-hunter, bmad-testarch-automate, bmad-sm, bmad-master, bmad-retrospective, bmad-domain-research, bmad-ux-designer, bmad-create-product-brief, bmad-testarch-atdd, bmad-testarch-ci, bmad-dev, bmad-editorial-review-prose, bmad-pm, bmad-validate-prd, bmad-qa-generate-e2e-tests, bmad-sprint-status, bmad-dev-story, bmad-quick-dev-new-preview, bmad-technical-research, bmad-party-mode, bmad-index-docs, bmad-quick-flow-solo-dev, bmad-testarch-nfr, bmad-create-epics-and-stories, bmad-tech-writer, bmad-create-architecture, bmad-correct-course, bmad-teach-me-testing, bmad-brainstorming, bmad-sprint-planning, bmad-quick-spec, bmad-document-project, bmad-create-ux-design, bmad-check-implementation-readiness, bmad-help, bmad-editorial-review-structure, bmad-qa, bmad-testarch-test-review

[srberard]

I'm having the same issue with Opencode 1.3.13 using Gemma 4 26B.

⚙ invalid [tool=glob:glob, error=Model tried to call unavailable tool 'glob:glob'. Available tools: invalid, question, bash, read, glob, grep, edit, write, task, webfetch, todowrite, skill.]
I apologize for that error. It seems I had a momentary glitch in my tool calling mechanism. I'm ready to assist you now.

[fernandodamaso]

Also i dont know if this helps, but in kilo code CLI (the fork of opencode) kimi 2.5 from the same provider, with the same api key, is working fine. Unlike Opencode.

[theolm]

Same here.
Kimi k2.5 via MoonshotAi