Agent Tool Trace for Debug

The problem

Agent runs fail in ways that are hard to see afterward. A task drifts. A file goes missing. A model says “done” without proof. Without a trace, nobody can tell where the run broke.

What it does

• Logs agent actions, decisions, discoveries, and artifacts as JSONL.
• Labels failures with 10 modes, including silent-scope-drift and artifact-evaporation.
• Generates pass/fail/unknown release receipts backed by evidence.
• Creates replay prompts so another agent or reviewer can retrace the run.
• Exports SFT-shaped JSONL, including negative eval records when failures exist.

Frameworks & tools

Python CLI using local JSON and JSONL artifacts. Core file: afr/cli.py.

Why this beats prompting it yourself

Asking an agent what happened after the fact is weak. This gives the run a paper trail: logs, failure labels, receipts, replay prompts, and exportable training/eval records.

Use cases

• Audit a skill, prompt pack, or workflow before release.
• Debug multi-step agent failures.
• Create receipts for sensitive file or release operations.
• Turn real agent runs into structured training and evaluation data.

Known limitations

No passive background recording. Requires CLI use, scripted integration, or post-hoc transcript analysis.