No, rebooting your machine or restarting the session does not resolve the literal exec(...) output issue.

Are other models affected?

No, this is specific to `kimi-coding/k2p5`. Other models like `google/gemini-3-flash-preview` continue to work normally.

Is there a way to fix the code without rolling back?

Not currently. This is a deeper issue within the gateway's model-specific tool-calling logic that requires an OpenClaw update.

Fix Kimi exec(...) text output in OpenClaw 2026.3.7

TL;DR — Quick Fix

OpenClaw 2026.3.7 broke tool calling for Kimi, causing it to print raw text instead of events; roll back to 2026.3.2 to restore functionality.

Run Diagnostics

Next Step

Fix now, then reduce repeat incidents

If this issue keeps coming back, validate your setup in Doctor first, then harden your config.

Open Doctor Harden Config

Error Signal

exec({"command":literal tool call text in chat

What's Happening

After updating to OpenClaw 2026.3.7, the kimi-coding/k2p5 model is failing to trigger tool calls correctly. Instead of the expected Exec card appearing in your Control UI, you are seeing literal text like exec({"command": "..."}) printed directly into the chat. The agent acts like it is working but fails to actually execute the requested filesystem or system tasks.

The Fix

This is a confirmed regression in the 2026.3.7 release. The most reliable fix right now is to roll back to the previous stable version where tool-calling compatibility was functional.

Run these commands in your terminal to revert:

npm install -g [email protected]
openclaw gateway restart

If you cannot roll back, switch your session model to google/gemini-3-flash-preview. Our tests confirm Gemini handles the Exec tool events correctly in the current version, so it acts as an effective drop-in replacement while we work on a permanent patch for the Kimi integration.

Why This Occurs

The issue is specific to the tool-calling compatibility path for Kimi within the 2026.3.7 runtime. The model is attempting to trigger commands, but the gateway is failing to intercept those intent patterns and render them as native tool cards. Instead, the raw output from the model is being treated as plain text, rendering the agent effectively useless for automated coding tasks.

Prevention

Until a patch is released, avoid updating to 2026.3.7 if you rely heavily on kimi-coding/k2p5. Keep an eye on GitHub issue #39907 for updates regarding the fix. We are actively investigating the parsing logic changes that caused this regression in the latest build.

Last Updated: March 2026

Did this guide solve your problem?