Skip to content

[Bug]: kimi-coding/kimi-k2.5 tool calling broken on all versions after 3.13 — parameters never reach tools #54688

Description

@ccyisafool

Bug type

Regression (worked before, now fails)

Summary

Summary
kimi-coding/kimi-k2.5 tool calling works correctly on v2026.3.13 but is broken on v2026.3.22, v2026.3.23-2, and v2026.3.24. Tools receive empty parameters — the model generates tool calls, but path, oldText, newText, and other required fields are not passed through. This causes infinite retry loops as the agent repeatedly calls tools with missing parameters.
The v2026.3.12 fix (PR #40008, #38669, #39907, #40552) that restored native Anthropic tool format for kimi-coding works in 3.12–3.13, but a later change re-broke the response parsing path in subsequent versions.

Environment

OpenClaw (broken): 2026.3.22, 2026.3.23-2, 2026.3.24
OpenClaw (working): 2026.3.13
Model: kimi-coding/kimi-k2.5
Provider: kimi-coding (anthropic-messages API)
OS: macOS 26.3.1 (Apple Silicon)
Node.js: 22.22.0
Channel: Telegram

Steps to reproduce

Install OpenClaw v2026.3.23-2 or later
Configure kimi-coding/kimi-k2.5 as the model for an agent
Ask the agent to perform any file operation (read, write, edit)
Observe tool call failures in logs

Expected behavior

Agent executes tool calls with correct parameters. Session logs show successful read, write, edit tool invocations with populated path, oldText, newText fields. This is the behavior on v2026.3.13.

Actual behavior

Tools receive empty/missing parameters. The agent enters an infinite retry loop. Gateway logs show:
error [tools] read failed: Missing required parameter: path (path or file_path). Supply correct parameters before retrying.
warn agent/embedded read tool called without path: toolCallId=tool_xxx argsType=object
error [tools] edit failed: Missing required parameters: path (path or file_path), oldText (oldText or old_string), newText (newText or new_string). Supply correct parameters before retrying.
These errors repeat every 3–5 seconds in a loop until the session is manually reset. The model is generating tool calls, but the parameters are not reaching the tool execution layer.
Note: GPT-5.4 via openai-codex is unaffected on the same OpenClaw version — this is specific to the kimi-coding provider path.

OpenClaw version

2026.3.23 and 2026.3.24

Operating system

macOS Tahoe

Install method

No response

Model

Kimi-coding/k2.5

Provider / routing chain

OpenClaw -> kimi k2.5

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't workingregressionBehavior that previously worked and now fails

    Type

    No type

    Fields

    Priority

    None yet

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions