Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

After ~1 hour of conversation with multiple exec/process calls, the AI agent
started sending empty JSON objects {} instead of properly formatted parameters.

Examples:

• exec: sends {} instead of {"command": "ls -la"}
• process: sends {} instead of {"action": "poll", "sessionId": "..."}
• read: sends {} instead of {"file_path": "/path/to/file"}

Restarting server, killing sessions, and even restarting the gateway
does NOT fix the issue. The problem persists across session resets.

Expected: AI should always include required parameters
Actual: AI sends empty objects causing "Validation failed" errors

Steps to reproduce

<KIMI_REF type="file" path="/root/openclaw/kimi/downloads/19d2ff9c-83a2-8a35-8000-0000456b3df0_image.png" name="image.png" id="19d2ff9c-83a2-8a35-8000-0000456b3df0" />

root@server:~$ openclaw sessions kill -all
error: unknown option '-all'

Expected behavior

AI should always include required parameters

Actual behavior

AI sends empty objects causing "Validation failed" errors

OpenClaw version

2026.3.23-2 (7ffe7e4)

Operating system

Linux

Install method

/root/openclaw/kimi/downloads/19d30021-4732-8851-8000-00005a538672_image.png" name="image.png" id="19d30021-4732-8851-8000-00005a538672

Model

kimi/k2p5

Provider / routing chain

/root/openclaw/kimi/downloads/19d30021-4732-8851-8000-00005a538672_image.png" name="image.png" id="19d30021-4732-8851-8000-00005a538672

Additional provider/model setup details

• Provider: kimi-claw
• Model: kimi/k2p5 (Kimi K2.5)
• Channel: Telegram direct
• Runtime: OpenClaw main agent on Linux x64

Logs, screenshots, and evidence

Impact and severity

Affected users/systems/channels: Single user (telegram:280847218) using kimi-claw channel. Potentially affects any user with long-running sessions (>1 hour) with multiple exec/process calls.

Severity: blocks workflow - AI agent completely unable to execute shell commands, process management, or file operations. Core automation functionality broken.

Frequency: intermittent/edge case - occurs after extended session duration (~1 hour) with multiple tool calls. Not reproducible immediately in fresh sessions.

Consequence: Failed automation tasks (transcription job interrupted), manual workarounds required, user must restart session or use alternative interfaces. Background processes killed when session reset attempted.

Additional information

NOT_A_VERSION_REGRESSION - issue appears to be runtime/cognitive degradation rather than code change.

Last known good: 2026-03-26 09:01 (successful transcription for another user with full exec/process functionality)
First known bad: 2026-03-27 ~19:51 (during current session with user art_tom_two_words)

Additional context:

• AI can still respond textually and provide reasoning
• Issue isolated to tool call parameter generation (exec/process/read/web_fetch)
• Pattern resembles "cognitive lock" where AI defaults to empty {} regardless of intent
• Server/sessions/gateway restarts do NOT clear the condition
• Suggests state persistence at model/prompt level rather than infrastructure
• Potential trigger: extended session (>1 hour) with repeated exec/process calls + background task