DSH chat migration

English | 中文

Capture conversations from Edge on chat.deepseek.com (one or many), export them as JSON, then rebuild them as resumable DeepSeek Harness sessions.

Status: 0.x, pre-release — data formats and CLI flags may change. Unofficial: not affiliated with or endorsed by DeepSeek or DeepSeek Harness. Licensed MIT; see CONTRIBUTING.md and SECURITY.md.

Edge extension  ──export──▶  dsh-deepseek-*.json  ──dsh-import.cmd──▶  ~/.dsh/sessions + workspace.json

Install

Requires Node.js 22+ (session compression uses built-in zstd).

# from npm (once published)
npm i -g dsh-deepseek-import
dsh-import --list

# or run straight from source
git clone https://github.com/Hungry-M/dsh-deepseek-import.git
cd dsh-deepseek-import
node dsh-import.mjs --list

The Edge extension must be loaded separately (next section) — a browser extension cannot ship inside an npm package.

1. Install the extension (once)

Open edge://extensions/ in Edge
Turn on Developer mode
Click Load unpacked and pick the edge-extension folder
Pin it to the toolbar (optional)

2. Capture a conversation

Open any conversation on chat.deepseek.com and let it finish loading
Open the extension; the conversation appears once it has been captured
To capture more, open them one by one — the list accumulates
Tick what you need → Export selected

If a conversation never shows up, open it and click Capture chat. It tries three sources in order: the browser's local cache (IndexedDB), a screen-by-screen backfill, then only the visible screen.

Capturing very long conversations

DeepSeek loads long conversations page by page. The extension does two things:

Merge by message identity — each message carries message_id; the extension unions them by id, so scrolling up or capturing twice only ever adds messages, never overwrites what was already captured.
Capture chat reads deepseek-chat / history-message straight from IndexedDB first. For most conversations that is the whole history in one read, with no scrolling at all.

3. Import into DSH

Double-click dsh-import.cmd in the workspace (or drop JSON files onto it). With no arguments it uses the newest dsh-deepseek-*.json found in your download folder.

Lists your DSH workspaces; pick one by number
Each conversation becomes one session: title, order and body (including code blocks and tables) are preserved
Attachments (images, txt, pdf, docx…): saved under <workspace>/.dsh-imported/<sessionId>/ with their original names only when the data source carries them; anything that cannot be downloaded degrades to [附件未下载：name (size) 原链接]. DeepSeek's IndexedDB cache copy usually carries no attachments
Thinking goes to thinking.md in the same folder (never injected into the context)
Edited / regenerated versions: only the selected branch stays in the body, all versions go to revisions.md
Interrupted outputs keep only a numbered placeholder ([中断输出 #27：未产生完整正文…]); add --keep-interrupted to keep the partial text
workspace.json is backed up before it is modified
Everything is mirrored to import.log

Important: restart DSH after importing

The sidebar list comes from ~/.dsh/storages/workspace.json, which DSH only reads at startup. Quit DSH completely and reopen it.

If the session still does not show up, use this order instead: quit DSH → run dsh-import.cmd → start DSH. A running DSH can overwrite workspace.json with its in-memory list.

Configuration (shared by CLI / TUI / extension)

import.config.json, next to this file:

{ "downloadDir": "D:\\Downloads", "subdir": "dsh-import", "lang": "zh" }

Key	Meaning
`downloadDir`	Where to look for exports (defaults to the real system download folder)
`subdir`	Subfolder convention: the extension saves to `<downloadDir>/<subdir>/`, the importer scans it too
`lang`	`zh` / `en`, affects the TUI and the extension UI
`autoTrim`	Default `true`: auto-trim conversations above the threshold
`autoTrimChars`	Trigger threshold, default `120000` chars
`autoTrimTurns`	Max turns kept, default `20`
`autoTrimBudget`	Char budget for the kept part, default `60000` (walks backwards from the end)

The same four auto-trim values live in the extension's ⚙ settings (checkbox + a collapsed "Advanced" section). They travel inside the export's options, and the importer resolves them in this order:

CLI flags  >  export options (extension settings)  >  import.config.json  >  built-in defaults

Sidecar location is also selectable in the extension (Write sidecars to DSH home), which equals --no-sidecar.

CLI (for AI / scripts)

node dsh-import.mjs --list                      # list workspaces
node dsh-import.mjs                             # newest export in the download folder
node dsh-import.mjs export.json -w "D:\proj"     # explicit workspace (auto-registers an unknown dir)
node dsh-import.mjs export.json --dry-run        # parse only
node dsh-import.mjs export.json --thinking       # also inline thinking (default: thinking.md only)
node dsh-import.mjs export.json --json           # machine-readable: {ok, workspace, created[]}
node dsh-import.mjs export.json --keep-versions   # keep pre-edit versions (default: current branch only)
node dsh-import.mjs export.json --keep-interrupted # keep partial interrupted output
node dsh-import.mjs export.json --tail 20        # only the last 20 turns
node dsh-import.mjs export.json --no-auto-trim   # disable auto trimming
node dsh-import.mjs export.json --no-sidecar     # sidecars to ~/.dsh/import-archive/
node dsh-import.mjs export.json --sidecar-dir "E:\arch"

dsh-import.mjs is also a module:

import { runImport, getWorkspaces, listExports, peekExport } from './dsh-import.mjs';
await runImport({ files: ['x.json'], workspacePath: 'D:\\proj', only: ['<conversation id>'] });

Interactive UI

node dsh-import-tui.mjs

Pick an export (size/time shown) → preview conversations (messages/chars/thinking/attachments) → select (all, 1,3, 1-2) → pick a branch point → pick workspace → confirm.

Branching from a specific message

To start a session at a chosen point instead of "the last N turns":

node dsh-import.mjs export.json --from-match "some text you typed"
node dsh-import.mjs export.json --from-index 157

Auditing imported sessions

DSH session format v4 requires the first model-visible node to be a system/message. Imports from 0.1.0 predate that fix and would break the moment you kept chatting in them. 0.1.1 writes the head on import, and this command repairs sessions created earlier:

node dsh-import.mjs --audit     # report only
node dsh-import.mjs --repair    # insert/move the head, resequence, drop the projection cache
node dsh-import.mjs --doctor    # environment / workspaces / sessions / permissions

--from-match does a case-insensitive substring match over user messages, first hit wins
If the same text was edited and appears several times, the TUI lists them chronologically and you pick by number
Choosing a branch point disables auto-trim (your choice wins); the full text still goes to transcript.md

How sidecar files are found later

The first user message ends with an index line:

[导入存档（需要时可直接读取）：思考 2 段 → .dsh-imported/<sessionId>/thinking.md；附件 → …]

Those files sit inside the workspace, so the model can read them on demand with ordinary file tools (nothing is pushed into the context automatically).

Don't import very long conversations in one piece

A 227-message / 520k-char conversation is roughly 250–350k tokens. DSH compacts automatically (dsh-compaction-basic), but after compaction only ~16% of recent history stays verbatim and the rest becomes a summary — the import is effectively wasted.

node dsh-import.mjs export.json --tail 20

Only the last 20 turns are imported; the full text goes to .dsh-imported/<sessionId>/transcript.md
The first message carries a note plus the sidecar index, so the model can grep the transcript when it needs old details
Conversations over 150k chars always get a transcript.md, even without --tail
transcript.md starts with an index (#N role chars — first line), and each section is headed ## #N, so grep "## #123" jumps straight to a message

FAQ

Double-clicking the cmd closes instantly? Older builds had LF line endings, which cmd.exe mis-parses. Current builds use CRLF, always stop at pause, and write import.log.

"Node.js not found"? Lookup order: node-path.txt → PATH → Program Files\nodejs → Program Files (x86)\nodejs → %LOCALAPPDATA%\Programs\nodejs → %APPDATA%\npm → node.exe next to the script.

Which file does the cmd import? Without arguments: the newest dsh-deepseek-*.json in the download folder. You can also drop files/folders onto it.

Body is only a fragment? Open the conversation and click Capture chat; then check the "来源" (source) tag in the list.

Known limitations

Citations and attachments depend on the data source: DeepSeek's IndexedDB cache copy keeps the body and thinking only; the references table and files field are usually absent. references.md and the attachment folder are therefore created only when the source carries them — their absence is a normal result, not a failure.
Imported sessions show "Untitled" in the sidebar first: the session list reads only the header and the projection cache — it never opens a cold log just to display a title. A freshly imported session has no cache yet, so the title is blank; opening it once restores the title permanently. The session/title event is in the log; nothing was lost. --doctor reports how many sessions are in this state.
Thinking is archived, not inlined: DSH reasoning blocks carry provider signatures; faking them would poison later requests. Read thinking.md, or force inline with --thinking.
Attachments are not added to DSH's attachment store: they are written to disk and referenced by path. When no download URL is available only the name and size survive.
Web-search citations: the cache record has no references table, so [reference:N] markers stay as plain text with no links.
Only text and attachments are migrated; tool calls and search results have no structured export.
Session format is DSH internals (v4 JSONL + zstd, multi-frame). A future DSH change may require an update here.
The extension relies on DeepSeek's own responses; if a redesign breaks capture, use Capture chat as the fallback.

4. Moving an imported session to another workspace

node move-session.mjs --list                                    # sessions per workspace
node move-session.mjs <session-id> --to "D:\other"               # move
node move-session.mjs <session-id> --to "D:\other" --dry-run     # preview

It updates four things: cwd in the session log, the project folder that holds the session, sessionIds in workspace.json, and the .dsh-imported/<id> folder. Restart DSH afterwards.

Rollback

Delete a session folder: ~/.dsh/sessions/--<workspace>--/session-<uuid>/
Restore the list: copy the matching workspace.json.bak-* over workspace.json
Single session: node remove-session.mjs <session-id>

Layout

edge-extension/
  manifest.json      MV3 manifest
  merge.js           window merging by message id / overlap
  inject.js          main world: read-only sniffing of fetch/XHR responses
  stream-parser.js   streaming patch protocol parser
  content.js         isolated world: normalize, cache reader, DOM fallback
  background.js      capture storage, attachment fetching
  popup.html/js/css  selection and export UI
dsh-import.mjs       export JSON → DSH sessions
dsh-import-tui.mjs   interactive importer
dsh-import.cmd       double-click entry

dsh-deepseek-import

Установка

README