openclaw - 💡(How to fix) Fix Runtime safety: break out of repeated identical failed tool calls

Bug: Agent enters no-op edit loop when tool rejects identical patch

Description

The agent enters an infinite loop retrying the same failed edit/patch tool call. The tool correctly rejects it with "No changes made" (oldText === newText), but the model retries the identical action repeatedly.

Log Excerpt

edit: replace "const response = await notion.databases.query({"
with "const response = await notion.databases.query({"
→ "No changes made"
[retries same edit 20+ times]

Expected Behavior

The runtime should detect repeated identical failed tool calls and:

1. Break out of the loop after N attempts (e.g., 3)
2. Surface the failure to the user
3. Allow the agent to try a different approach (re-read file, compute different patch, or explain blocker)

Environment

• OpenClaw version: 2026.5.12 (f066dd2)
• Model: stepfun/step-3.6
• Session type: Normal chat session (Discord)
• Tool: edit (patch/replace)

Mitigation (prompt-level)

Added to workspace AGENTS.md:

No-op edit guard: If an edit/patch tool call fails with "No changes made" or the oldText equals newText, do NOT retry the same edit. Re-read the file, compute a different patch, or stop and explain the blocker.

Feature Request

Add a runtime config knob: agents.defaults.maxIdenticalToolFailures (default: 3)

When the same tool call with identical parameters fails N times in a row, automatically:

• Stop the current agent turn
• Return a system event noting the repeated failure
• Include the tool name and parameters in the event

Related

• Heartbeat watchdog for stuck background processes (HEARTBEAT.md step 14)
• Mechanical stops for batch scripts (MECHANICAL_STOPS.md)