openclaw - 💡(How to fix) Fix [Bug]: Kimi Code (Moonshot) Provider enters an infinite loop during tasks [1 participants]

Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

When using kimi-code (Kimi API) as the LLM provider in OpenClaw, the agent frequently gets stuck in an infinite loop. It repeatedly generates the same response text followed by a tool call with empty arguments (e.g., read: {}), which fails validation.

Steps to reproduce

1. Configure OpenClaw 2026.3.24 with the kimi-coding:default provider.
2. Start a task requiring file system access (e.g., "Analyze app.js").
3. Observe the agent repeating the same analysis text and calling read with empty arguments in the session logs (~/.openclaw/agents/main/sessions/).

Expected behavior

The agent should preserve its thinkingSignature across turns and correctly provide the path argument in the tool call, as observed with other Anthropic-compatible models.

Actual behavior

The agent repeats the previous turn's text and issues an empty toolCall for read, resulting in a Missing required parameter: path alias. Supply correct parameters before retrying. error in the logs.

OpenClaw version

2026.3.24

Operating system

Linux (Ubuntu/Debian via NVM)

Install method

npm global (nvm)

Model

kimi-coding/k2p5

Provider / routing chain

openclaw -> kimi-coding (direct)

Additional provider/model setup details

Additional provider/model setup details Using kimi-coding:default auth profile. The issue occurs because preserveAnthropicThinkingSignatures defaults to false and openAiPayloadNormalizationMode is set to "moonshot-thinking", which appears to interfere with the Anthropic-compatible tool protocol.

Logs, screenshots, and evidence

JSONL log evidence:

{"role":"assistant","content":[{"type":"text","text":"... (repeated analysis) ..."},{"type":"toolCall","id":"tool_...","name":"read","arguments":{}}]}
{"role":"toolResult","toolCallId":"tool_...","toolName":"read","content":[{"type":"text","text":"{\n  \"status\": \"error\",\n  \"tool\": \"read\",\n  \"error\": \"Missing required parameter: path alias...\"\n}"}]}

Impact and severity

Affected users/systems/channels: Users of the Kimi Coding provider extension. Severity: High (blocks all multi-step coding/analysis tasks). Frequency: Always observed during multi-step workflows. Consequence: Agent becomes unusable for tasks requiring tool interaction.

Additional information

Local fix confirmed:

1. Setting preserveAnthropicThinkingSignatures: true and openAiPayloadNormalizationMode: "default" in extensions/kimi-coding/index.ts.
2. Adding anthropicToolSchemaMode: "openai-functions" and anthropicToolChoiceMode: "openai-string-modes" to Kimi capabilities.
3. Simplifying tool parameter labels from "path alias" to "path" in pi-tools.params.ts.