dsh-postmortem

When a DeepSeek Harness agent run fails, you are left staring at raw session events and an ambiguous error string. dsh-postmortem reads those existing events locally and, in seconds, answers three practical questions: what failed, what evidence supports that conclusion, and what should be verified before the next attempt. The entire process is read-only, requires no model call, and sends no data over the network.

Its local rules identify failed tools, detect three or more unchanged retries via a one-way argument fingerprint, separate terminal states from earlier causal evidence, and report user cancellations honestly without inventing a repairable failure. Beyond the text diagnosis, /postmortem-repair emits a bounded recovery prompt that instructs the next agent to verify the missing resource first and forbids repeating the identical failed call; /postmortem-plan exports strict schema-v1 JSON with evidence categories, advisory actions, and verification steps for external runners. Every output is redacted: user messages, tool arguments, tool outputs, file contents, prompts, credentials, and raw traces are never retained or exported.

It is built for DSH users who want a safe, reproducible next action after a failure, and for teams that need redacted reports to file issues or run paired evaluations. It is deliberately not an autonomous retry system, a trace-upload service, or a replacement for task-level observability. You keep full control; it simply makes the failure legible and points to the right next check.