Greptile Summary

This PR fixes Kimi and Kimi Coding being incorrectly treated as native Anthropic tool-payload providers. It adds anthropicToolSchemaMode: "openai-functions" and anthropicToolChoiceMode: "openai-string-modes" to both the fallback capability registry (PLUGIN_CAPABILITIES_FALLBACKS["kimi"]) and the kimi-coding bundled plugin, ensuring the two code paths stay aligned. Tests are updated to assert the new (correct) behaviour and a regression test is added for the kimi-coding alias.

Confidence Score: 5/5

• Safe to merge — the fix is narrow, well-tested, and the plugin/fallback paths are now aligned.
• All findings are P2 or lower. The capability change is covered by both updated unit tests and new regression tests, and the normalization logic is consistent across the fallback registry and the bundled plugin.
• No files require special attention.

_{Reviews (1): Last reviewed commit: "fix: normalize kimi anthropic tool paylo..." | Re-trigger Greptile}

Manual check: validated against a live Kimi Coding plan account.

Ran a small openclaw agent --local compare with kimi-coding/k2p5 on both main and this branch. This branch preserved the tool-call args (functions.exec:0 {"command":"pwd"}), which matches the regression covered in this PR.

Landed on main.

• Landed source commit: f7e4317
• Merge commit: b441cd2

My PR #54491 fixed an issue where Kimi tool calls could return empty parameters, and it was merged into main in release v2026.3.31 . After upgrading OpenClaw to v2026.4.2 , I’m hitting a new issue: Kimi no longer works correctly — tool calling doesn’t execute, and the agent replies with the tool call content directly instead of actually running the tool.

<|tool_calls_section_begin|>
<|tool_call_begin|> read:43
<|tool_call_argument_begin|> {"path": "..."}
<|tool_call_end|>
<|tool_calls_section_end|>

My PR #54491 fixed an issue where Kimi tool calls could return empty parameters, and it was merged into main in release v2026.3.31 . After upgrading OpenClaw to v2026.4.2 , I’m hitting a new issue: Kimi no longer works correctly — tool calling doesn’t execute, and the agent replies with the tool call content directly instead of actually running the tool.

<|tool_calls_section_begin|>
<|tool_call_begin|> read:43
<|tool_call_argument_begin|> {"path": "..."}
<|tool_call_end|>
<|tool_calls_section_end|>

the same issue with v2026.4.2 and kimi/kimi-coding api

@yuanaichi @dorbodwolf I have a follow-up PR up for this fix in #60051

Can confirm that v2026.3.13 works fine for Kimi Coding tool calling, while 2026.3.7+ and 2026.4.2 still have the regression where tool calls are printed as text instead of being executed.

It would be great if this fix could also address the response parsing side, since the current half-fix in 2026.4.2 doesn't fully resolve the issue.

This was fixed on main in follow-up Kimi patches after v2026.4.2, including the response-side parsing path that was still missing when this PR landed.

So the regression reported here is not still open on main; it just has not shipped in a release yet. It will be included in the next release.

Thanks for the quick fix and the clarification! Looking forward to the next release. 🦞