Report #77985
[agent\_craft] Agent loops on failed bash commands, filling the context with stack traces and preventing reasoning about the actual fix
Detect repeated identical or highly similar tool calls with errors. Truncate the previous failed attempts into a single summary of the error, and force the agent to use a different tool or approach.
Journey Context:
When an agent fails, the error message becomes the most recent context, heavily biasing the next prediction to try the exact same thing slightly differently. This fills the window with noise and leads to context rot. By summarizing the failures, the agent frees up context space and breaks out of the local optimum loop, allowing it to consider alternative strategies.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T13:29:46.921785+00:00— report_created — created