openclaw - 💡(How to fix) Fix Streaming repetition loops bypass idle timeout; no semantic circuit breaker or context sanitization [1 participants]

Streaming repetition loops bypass idle timeout; no semantic circuit breaker or context sanitization

Incident Overview

On 2026-04-08, an architectural gap was identified when google/gemini-3-flash-preview entered a repetition loop within its <think> block. Despite the loop continuing for an extended period, the OpenClaw idle timeout (60s) failed to trigger because the stream remained technically active with continuous token generation.

Issues Identified

1. Idle Timeout Evasion: The current idle timeout monitors stream activity (the presence of incoming data) rather than progress. Continuous repetitive output satisfies the "active stream" condition, allowing runaway sessions to persist until maxTokens is reached.
2. Session Poisoning: Repetitive or aborted turns are currently retained in the session history. This "poisons" the context for subsequent turns, often leading the model back into the same loop or degraded reasoning states.
3. Lack of Semantic Circuit Breakers: There is no mechanism to detect n-gram repetition or structural loops within the streaming content in real-time.

Proposed Requirements

1. Semantic Repetition Breaker: Implement an n-gram detector (or similar heuristic) to identify and kill runaway streams that exhibit high repetition, even if the stream is technically active.
2. Context Sanitization: Implement a mechanism to sanitize or "roll back" the session history when a turn is aborted or detected as looping, preventing the corrupted output from poisoning future reasoning.
3. Enhanced Defense-in-Depth: While maxTokens prevents total cost runaway, it is a blunt instrument that does not address the degraded state of the persistent session.

Reference

This issue refers to the incident report dated 2026-04-08.