agent-trace-debuggerlisted
Install: claude install-skill alexzhu0/agent-ready-skills
# Agent Trace Debugger
## Purpose
Make noisy agent traces readable enough to debug quickly.
Use this skill for JSONL traces, tool logs, run transcripts, orchestration logs, and retry histories.
## Inputs
- Trace file, JSONL events, pasted logs, or tool-call transcript.
- Expected task outcome if known.
- Error messages, timestamps, run IDs, and environment notes.
## Workflow
1. Build a chronological timeline of user input, agent decisions, tool calls, tool results, errors, and retries.
2. Mark the first observed failure separately from later cascading failures.
3. Classify failures: missing input, tool error, auth, network, parsing, policy, timeout, bad plan, or bad assumption.
4. Identify evidence for each root-cause candidate.
5. Recommend the smallest retry or fix that would prove or disprove the top candidate.
6. Preserve raw error strings exactly when they are useful for search.
## Output
Produce Markdown with these sections:
- Run Snapshot
- Failure Timeline
- First Failure
- Root-cause Candidates
- Retry Or Fix Plan
- Evidence To Preserve
Keep the timeline concise and omit routine success events unless they explain the failure.
## Validation
- The first failure is not confused with a later symptom.
- Every root-cause candidate has trace evidence.
- Retry steps are ordered from lowest-risk to highest-risk.
- Unknowns remain explicit.
- Raw error strings are not paraphrased away.