dsh-session-surgeon

dsh-session-surgeon

Copy a session ID from the sidebar ⋯ menu, then paste it into a new chat so the new session can learn from the old one — the Codex-style “continue from this thread” move that stock DSH does not expose.

Also repairs DeepSeek Harness sessions that refuse to load — seq gap, torn zstd, lone surrogates, and events missing their message id. 官方以后修加载器，也救不回已经坏掉的 session.jsonl.zstd。

Compatible with @deepseek-ai/dsh@0.1.0-rc.6.

Why copy the session ID

Stock DSH session ⋯ only has rename / fork / archive. There is no “copy ID”.

After this plugin:

1. Left sidebar, click ⋯ on a session → 复制会话 ID

2. Open a new chat and paste something like:

   Continue from session session-1e66cda9-a046-4893-8f4b-b817080acbea.
   Read that log if you need prior context.
   

   or, for an agent with tools:

   session_inspect id=session-1e66cda9-a046-4893-8f4b-b817080acbea
   Then keep going from where that conversation left off.
   

The ID is the durable handle (with or without the session- prefix — same session). Fork duplicates a file; copying the ID lets a new session refer to the old one, the way people pass a Codex thread id around.

Install

One command, same as other DSH plugins:

dsh plugin --profile web add "github:xiaoshenming/dsh-session-surgeon#main"

Restart dsh web. Then:

• Session ⋯ menu: 复制会话 ID (the everyday action) / inspect / dry-run repair
• Sidebar 会话医生 / Session surgeon: browse conversations, copy id, dry-run repair, apply repair, compact preview, export JSONL

Developers working in a checkout can still use dsh plugin --profile web add link:"$(pwd)".

Also: repair unloadable sessions

Real crash families from official Discussions:

• #317 stack overflow on huge history
• #1497 / #1586 seq gap in committed region
• #436 lone UTF-16 surrogate → permanent HTTP 400
• #674 leftover .tmp plaintext

Default repair is dry-run. --apply writes .bak.<utc> first, never invents missing seqs, and fills missing message ids without dropping events.

CLI

No dsh web required after the plugin (or this repo) is on disk:

npx --yes github:xiaoshenming/dsh-session-surgeon scan
npx --yes github:xiaoshenming/dsh-session-surgeon inspect <session-id>
npx --yes github:xiaoshenming/dsh-session-surgeon repair <session-id>          # dry-run
npx --yes github:xiaoshenming/dsh-session-surgeon repair <session-id> --apply  # writes .bak.<utc> first

Or from a clone: node bin/dsh-session-surgeon.mjs scan.

Agent tools after install: session_scan / session_inspect / session_repair (apply defaults to false).

Commands

command	meaning
scan [root]	list sessions + header health + orphan .tmp
inspect <id>	decode every zstd frame, expand packed rows, report seq gaps
repair <id>	default --dry-run; --apply rewrites after .bak.<utc>
compact <id> --keep-last-turns N	keep the last N complete turns, renumber seq from 0
export <id>	JSONL dump; redacts secrets unless --no-redact
index [root]	session / parent / goal / health table

--format text prints a human table. Exit 0 on success, 2 on usage error, 1 on not-found / refuse.

What this is not

Not a marketplace, not a token heatmap, not a memory plugin, not a Codex task store.

DSH has session id + optional same-session goal id, not a Codex-style resumable task id. Copying the session ID is the closest everyday equivalent. See docs/LEARNING-TASKS.md.

Docs

• docs/PLAN.md — milestones, safety, release
• docs/SESSION-FORMAT.md — zstd frames, header, packed rows, when the official loader refuses
• docs/REPAIR-SPEC.md — repair steps aligned with the official loader
• docs/LEARNING-TASKS.md — why there is no Codex task id
• docs/IMPLEMENTATION-CONTRACT.md — multi-agent implementation contract

Safety

• Default is read-only. Write paths require --apply and write .bak.<utc> first.
• Never commit raw files from ~/.dsh/sessions (they contain user text and secrets).
• Do not put @deepseek-ai/dsh-tools in dependencies.
• Export redacts sk-*, PEM blocks, and home paths unless --no-redact.

License

MIT