AI coding agents now read files, run commands, retry failures, and produce final summaries. But the final answer does not always show what happened along the way.
That is why I am experimenting with agent-postmortem-kit.
This is a private v0.1 technical preview, not a finished product and not a production-ready audit platform.
The current loop is small:
sample agent log -> postmortem findings -> HTML / JSON report -> skill-candidate draft
The tool currently works with a synthetic sample log. It can flag repeated failures, dangerous commands, secret exposure risk, evidence gaps, unfinished work, and human approval points.
It can also export a Markdown skill-candidate draft with –skill-out. The goal is not only to inspect a trace, but to turn failures into next-run rules.
This is different from AgentTrace. AgentTrace is stronger for history, token use, cost, timing, tool failures, and health. agent-postmortem-kit is focused on the postmortem layer: what failed, what looked risky, what evidence exists, and what rule should prevent the same failure next time.
What it cannot do yet:
– dedicated OpenClaw / Hermes / Codex / Claude Code / AgentTrace adapters
– complete secret redaction
– safe automatic processing of private real logs
– production-grade policy packs
Real logs should not be published. Generated reports need human review before sharing.
For now, I am keeping the repository private and using articles or NotebookLM-style explanations to see whether the idea is understandable.
Read the log. Find the failure. Keep the evidence. Turn the lesson into a rule.
superdoccimo/agent-postmortem-kit
