Bug type

Regression (worked before, now fails)

Summary

After commit 909f26a (fix(kimi-coding): normalize anthropic tool payload format), kimi-coding/k2p5 no longer returns proper tool_use content blocks. The model outputs tool calls as plain text XML instead, with stopReason: "stop" instead of stopReason: "toolUse".

Steps to reproduce

1. Configure kimi-coding/k2p5 as the model provider
2. Send a message that triggers tool usage (e.g. a coding task requiring file reads)
3. Observe model response

Expected behavior

Model returns structured tool_use content blocks with stopReason: "toolUse", as it did before v2026.3.7 (confirmed working on 2026-03-04 session logs).

Actual behavior

Model returns tool calls as plain text XML embedded in a text content block, with stopReason: "stop". The agent loop cannot parse these as real tool calls, so tools are never executed.

Root cause:
Commit 909f26a (2026-03-05, included in v2026.3.7) added compat: { requiresOpenAiAnthropicToolPayload: true } to the kimi-coding model config (src/agents/models-config.providers.static.ts:237). This triggers createAnthropicToolPayloadCompatibilityWrapper which converts tool definitions from Anthropic native format (name + input_schema) to OpenAI function format (type: "function" + function: { name, parameters }).

However, kimi-coding's anthropic-messages API endpoint expects Anthropic native tool format. Session logs from 2026-03-04 (before this commit) confirm that kimi-coding/k2p5 handled Anthropic native tool format correctly and returned proper tool_use blocks.

Suggested fix:
Remove requiresOpenAiAnthropicToolPayload: true from the kimi-coding model config, or set it to false, so tools are sent in Anthropic native format.

OpenClaw version

current main (post-v2026.3.7), regression introduced in 909f26a

Operating system

macOS 15 (Darwin 24.6.0)

Install method

pnpm dev (source)

Logs, screenshots, and evidence

Working session (2026-03-04): stopReason: "toolUse", response contains toolCall content blocks
Broken session (post 909f26a26): stopReason: "stop", response contains tool calls as XML text in text content blocks
Last known good: pre-909f26a26 (before 2026-03-05)
First known bad: 909f26a26 (2026-03-05), included in v2026.3.7

Impact and severity

Affected: All kimi-coding/k2p5 users
Severity: High (tool calling completely non-functional)
Frequency: 100% repro
Consequence: Agent cannot use any tools, making kimi-coding unusable for agentic workflows

Additional information

No response