openclaw - 💡(How to fix) Fix Bug: Circuit breaker blocks tool but not agent turn; exec loop only warns, never blocks [1 comments, 2 participants]

### Evidence from logs

#### Phase 1: `tdai_conversation_search` loop (08:24 - 08:29)
08:24:48 tools loop: tools=tdai_dialogue_search_level=warning action=warning detector=generic_repeat count=10
08:25:24 tools loop: tools=tdai_dialogue_search_level=warning action=warning detector=generic_repeat count=20
08:26:23 Tools loop: tools=tdai_dialogue_search_level=critical action=block detector=global_circuit_breaker_count=30
08:26:27 Tools loop: tools=tdai_dialogue_search_level=critical action=block detector=global_circuit_breaker_count=30
08:26:30 Tools loop: tools=tdai_dialogue_search_level=critical action=block detector=global_circuit_breaker_count=30
... (continues indefinitely, every ~3 - 4 seconds)

The circuit breaker correctly blocked the tool after 30 calls. However, the agent turn was NOT terminated. The agent kept retrying, and each retry was blocked, producing an endless stream of error logs.
#### Phase 2: `exec` loop (08:38 - 08:42+)
08:38:58 Loop warning: exec is called 10 times with the same parameters 08:39:06 Loop warning: exec is called 11 times with the same parameters 08:39:12 Loop warning: exec is called 12 times with the same parameters... (Lasts more than 30 calls) 08:41:18 Loop warning: exec is called 30 times with the same parameters 08:41:23 Loop warning: exec is called 30 times with the same parameters... (Infinite duration)
The `exec` tool loop never reached critical/block status. The detection logic saw different outputs on each call (system command returns with different timestamps) and classified them as "making progress", so only warnings were emitted. The agent continued executing the same command 30+ times.
### Root Cause Analysis
1. **Circuit breaker only blocks tool execution, not agent turn**: When `blocked: true` is returned, the tool call fails but the agent continues its reasoning loop. The agent does not receive a clear signal to abandon the task.
2. **`exec` detection checks output, not just parameters**: The loop detection compares tool name + parameter hash + result hash. For `exec`, the result changes on every call (e.g., `date` command returns different timestamps), so the detection never sees "identical no-progress outcomes" and never triggers critical/block.
3. **Warnings are ineffective**: The agent receives warning messages like "WARNING: You have called exec 10 times with identical arguments. If this is not making progress, stop retrying and report the task as failed." but completely ignores them and continues the loop.
### Proposed Solutions
1. **Terminate agent turn on circuit breaker**: When `global_circuit_breaker` triggers, the agent turn should be terminated immediately. The agent should receive a system message indicating the task cannot be completed and should report failure to the user.
2. **Parameter-only detection for `exec` and similar tools**: For tools where the output naturally changes (like `exec`, `read`, `browser`), the loop detection should only compare tool name + parameter hash, ignoring the result hash. If the same command is executed multiple times, it is a loop regardless of output.
3. **Escalating warnings with hard stop**: After N warnings (e.g., 20), the system should force-terminate the turn rather than continuing to warn. The agent should not be allowed to continue after the critical threshold.
4. **Better agent guidance**: The warning message should be more explicit: "This is warning #20 of 20. The task will be terminated. You must report failure to the user now."
### Impact
- **Token waste**: Each loop iteration consumes input tokens for the agent to process the blocked result and decide to retry again
- **Log pollution**: Endless error logs fill up the log file, making it hard to diagnose other issues
- **User experience**: User sees the agent "stuck" with no clear failure message
### Alternatives Considered
- **Lowering thresholds**: Would reduce the number of wasted calls but does not solve the fundamental issue of the agent not terminating
- **Disabling loop detection**: Would allow the circuit breaker to work but leaves the agent in an infinite loop
- **Manual intervention**: Requires user to manually stop the agent, which is not practical for autonomous operation