Miscoped Failure Cache Caused Autonomous Pentesting Agent to Silently Skip Tools

A developer building Halo, an open-source autonomous pentesting agent powered by the Gemma 4 12B local LLM, discovered a state-management bug that caused the agent to permanently skip tools without any error or warning. The agent's failure cache in agent_cache.py used SHA-256 fingerprints keyed only by tool name and target, with no engagement-level scoping. This meant a single failed tool run against one target would globally blacklist that tool across all future, unrelated engagements. The fix involved adding an engagement_id field to the cache key, so failure records are now isolated per session and tools start fresh on new engagements. The incident highlights a broader design risk in agentic systems: caches scoped too broadly can silently degrade an agent's capabilities over time without triggering any obvious errors.