zzh-newlearner/dsh-postmortem
Local-first failure postmortems for DeepSeek Harness sessions.
Turns a failed DeepSeek Harness run into a redacted recovery plan in seconds: reads the DSH session events you already have and answers what failed, what evidence supports it, and what should be checked before the next attempt. /postmortem prints detected findings (e.g. 'step 1: Tool shell failed'); /postmortem-repair returns a copy-only recovery prompt telling the next attempt to verify the missing resource first; /postmortem-plan gives strict redacted JSON advice (evidence category, advisory action, verification step) for external runners. Output contains no user messages, tool arguments, tool output, files, prompts, credentials, or raw traces.
Install
dsh plugin --profile web add @huichangzz/dsh-postmortemdsh plugin --profile <profile-name> add @huichangzz/dsh-postmortem (npm 0.9.2 verified on registry) — installs into the profile and auto-registers its bundle layer; restart that profile, then run /postmortem after a failed turn. No manual Loader entry needed. Optional model review: add a postmortem config override (model.enabled/provider/model/timeoutMs) to the profile's cordis.patch.yml; autoOnFailure defaults to 'detected' (only actionable detected failures). No key is configured or stored.
Compatibility
DSH profiles (bundle layer auto-registered). Reads existing session events only — read-only, local-first, useful without a model.
Details
- Repo: zzh-newlearner/dsh-postmortem
- Category: Agent Capabilities
- Stars: 2
- Version: npm 0.9.2 (registry-verified 2026-08-29)
- Last push: 2026-09-06
- First seen: 2026-08-13
Recent updates
postmortem/postmortem-repair/postmortem-plan commands; read-only session-event analysis; redacted deterministic reports; optional LLM review layer.
FAQ
- What does it read?
- Only the DSH session events you already have — it is read-only and never executes recovery actions.
- Is output redacted?
- Yes — no user messages, tool arguments/output, files, prompts, credentials, or raw traces appear in reports.
- Does it need a model?
- No — the deterministic report works without one; an optional model layer (reusing DSH's llm service) can add review.
Alternatives
LeslieWylie/dsh-task-relay · dongsheng123132/dsh-benchmark · LeslieWylie/dsh-ops-kit