TRAJDEBUG: Tracing Error Lifecycle to Identify Critical Failures in Long-Horizon Agent Trajectories
TrajDebug tracks whether errors persist, resolve, or cause terminal failure, offering a sharper way to debug long coding-agent runs than flagging every local mistake.
**TrajDebug** compresses long histories at multiple granularities, identifies errors from evidence, then tracks resolution status and terminal impact. **TrajErrBench** contains **486 manually annotated failed trajectories** from Tau2Bench and SWE-Bench Pro.
For coding-agent observability, preserve enough trajectory context to distinguish a harmless local mistake from the earliest unresolved error that caused the final failure. That attribution can make remediation more targeted.
**TrajDebug** compresses long histories at multiple granularities, identifies errors from evidence, then tracks resolution status and terminal impact. **TrajErrBench** contains **486 manually annotated failed trajectories** from Tau2Bench and SWE-Bench Pro. For coding-agent observability, preserve enough trajectory context to distinguish a harmless local mistake from the earliest unresolved error that caused the final failure. That attribution can make remediation more targeted. The abstract claims the best overall performance and actionable downstream feedback but gives no scores or baseline details. Code and data are promised for release, so reproducibility cannot be assessed from this material.
This turns long-trajectory inspection into lifecycle attribution: identify errors from evidence, track whether they were resolved, and isolate the earliest unresolved mistake linked to terminal failure. It operationalizes prior calls for queryable trajectories and flexible analysis, while going beyond pass/fail verification toward targeted remediation. The absent scores, baseline details, code, and data keep the claimed advantage provisional.