zzh-newlearner/dsh-postmortem ↗★ 2
@huichangzz/dsh-postmortem
Local-first failure postmortems for DeepSeek Harness sessions.
安装
此插件尚未提供可验证的 bundle,或兼容性检查未通过。请先阅读仓库说明。 阅读完整 README ↗
说明文档
阅读完整 README ↗DSH Postmortem / DSH 事后复盘
面向 DeepSeek Harness 的本地优先、只读失败复盘插件。
Local-first, read-only failure postmortems for DeepSeek Harness.
dsh-postmortem 从既有 DSH session event log 生成简短、脱敏的故障报告。它不会改变 agent loop、重试工具、注入提示词或上传完整轨迹。可选模型复盘始终是第二层:确定性规则结论优先,模型输出无效或不可用时会被丢弃。
dsh-postmortem turns the existing DSH session event log into a short, redacted incident report. It does not alter the agent loop, retry a tool, inject a prompt, or upload a transcript. Optional model review is deliberately a second layer: deterministic findings are authoritative, and invalid or unavailable model output is discarded.
安装 / Install
npm install @huichangzz/dsh-postmortem
在 DSH profile 的 cordis.patch.yml 中加入:
Add this to the DSH profile's cordis.patch.yml:
- id: postmortem
name: '@huichangzz/dsh-postmortem'
config:
autoOnFailure: true
model:
enabled: false
provider: your-provider
model: your-model
timeoutMs: 10000
插件使用 DSH 已配置的 llm 服务。请在 DSH 中配置模型 provider 和密钥;本包没有 API key 配置,也不会写入密钥。
The plugin uses DSH's existing llm service. Configure model providers and keys in DSH, rather than this plugin; the package has no API-key setting and never writes a key.
命令 / Commands
| 命令 / Command | 结果 / Result |
|---|---|
/postmortem [turn] | 查看最近或指定轮次的可读报告。 / Readable report for the latest or selected turn. |
/postmortem-export [turn] | 导出脱敏的 schema-v2 JSON 报告。 / Redacted schema-v2 JSON report. |
/postmortem-repair [turn] | 为已检测失败生成仅可复制的修复提示。 / Copy-only repair prompt for a detected failure. |
命令使用 recordInput: false,因此轮次选择不会写入 command/run 事件。修复命令只返回文本:不会调用工具、提交 follow-up、重试操作,或成为模型上下文。
Commands use recordInput: false, so turn selection is not copied into command/run events. The repair command returns text only: it never calls a tool, submits a follow-up, retries an action, or makes itself model context.
检测与隐私边界 / Detection And Privacy Boundary
本地规则检测工具错误结果、已结束轮次中缺失的工具结果、三次及以上未变更的失败重试,以及非 completed 的轮次结束。仍在执行的轮次会保持 inconclusive,不会仅因结果尚未到达就被误判为失败。