DSH Recovery Center / DSH 恢复中心
Local plugin recovery points for DeepSeek Harness, with a rescue interface that works even when the DSH frontend is broken.
Preview: tested with DSH 0.1.2-rc.1, Node 22, macOS. Linux is implemented but not yet verified. Windows and supervised/Desktop restarts are not supported in this first release. Save a point before changing plugins. This is not a full DSH or workspace backup.
中文使用说明
Languages
Chinese and English are included. Inside DSH, the navigation label and recovery panel follow Settings → General → Language immediately. The standalone rescue page follows the browser language, offers a 中文 / English selector, and remembers the choice. User-written recovery point names are preserved. CLI output follows LANG/LC_ALL or --lang en / --lang zh.
Features
- Settings → 恢复中心: save a named recovery point and restore it.
- Copies installed plugin files, not just package names: restore does not need npm or network access.
- SHA-256 verification before restore; refuse changed/corrupt points or different DSH versions.
- Automatically save a before-restore point so a restore can itself be undone.
- Staged file replacements and an on-disk journal for interruption recovery.
- Independent CLI/browser rescue copied outside the plugin directory on activation.
- No model-visible tools, telemetry, downloads, external scripts, or third-party API calls.
Installation
Install from npm using the DSH CLI (then restart DSH):
dsh plugin --profile web add dsh-recovery-center@0.1.2 --config.auto-install-peers=false
Ensure dsh-recovery-center is in package.json → dsh.profile.bundles for this profile. Open Settings → 恢复中心 and save a first recovery point while the environment is healthy.
This release deliberately declares the tested DSH API versions. Do not force-install it into a different core version without retesting.
What is restored
Only these paths inside the selected profile:
package.json, pnpm-lock.yaml, pnpm-workspace.yamlcordis.yml, cordis.patch.ymlnode_modules/ including installed plugin bytes.dsh-market/state.json
Chat/session databases, workspaces, provider credentials, global settings, the DSH executable/core packages, and files outside the profile are not restored. Configuration embedded in the tracked files is captured; points are private local directories and must not be published.
Snapshots are profile-specific and machine-local. External dependency symlinks are rejected instead of producing an incomplete backup. Dangling internal links are preserved exactly, including old .bin shims. Installed copies of file: dependencies are captured, but their source directories are not. Do not install/update packages while taking a recovery point. Points are retained until manually removed while DSH is stopped; they consume roughly the installed dependency size (APFS clones can share disk blocks).
Emergency recovery
On first activation, the plugin generates a runnable rescue launcher using that installation’s Node executable and data directory. On a local macOS installation, it also creates an executable .command on the current user’s Desktop (when the Desktop exists and is writable). Names include a per-installation identifier so multiple DSH profiles do not overwrite each other. Existing unrelated files are never overwritten. Set desktopShortcut: false to opt out. SSH launches do not create a desktop shortcut.
The recovery panel shows the actual launcher location and a Create / repair launcher button. On Linux, or if macOS Desktop access is unavailable, a .sh / .command is still saved in the independent rescue directory. This is on the DSH host computer, not necessarily the browser’s computer. Windows remains unsupported. Save at least one recovery point before experimenting with plugins.
The plugin copies its rescue runtime into:
$DSH_HOME/recovery-center/<profile>/rescue/
DSH_HOME defaults to ~/.dsh. This directory is outside the plugin tree and is preserved during restore/removal. Use Node 22+:
node "$DSH_HOME/recovery-center/web/rescue/cli.js" serve --store "$DSH_HOME/recovery-center/web" --open
The rescue server binds a random port on 127.0.0.1 only. A random per-run token lives in the browser URL fragment and is sent in a custom request header. Leave the terminal open while using rescue; close it to stop the rescue server.
CLI alternatives:
dsh-recovery list --profile web
dsh-recovery capture --profile web --label "Before a new theme"
# Stop DSH first for these offline operations:
dsh-recovery restore <point-id> --profile web --yes
dsh-recovery repair --profile web
repair clears a lock only when its owning PID no longer exists and reverses a journaled interrupted restore. It refuses while the recorded DSH process exists. If a filesystem error prevents journal repair, keep the recovery directory intact for manual recovery.
Restart behavior
The UI explicitly confirms task interruption. A detached worker validates the selected point, checks the recorded host process identity, asks that process to exit gracefully, restores the files, and relaunches its original command. It never kills a different PID owner and never force-kills an unresponsive host.
For launchd/systemd/pm2/Desktop or custom supervisors, set plugin config allowRestart: false and use the supervisor plus offline CLI restore. Detection is best-effort; do not assume arbitrary supervisors can be detected. Do not enable auto-restart for an unknown launch method.
Development
No build tool or bundled third-party runtime is needed. The client registers a settings iframe using the DSH module factory contract, with React supplied by DSH. Host HTTP routes use the DSH authentication and origin fence. Mutations additionally require a same-origin JSON POST.
npm test
npm run check
npm pack
Publication in a plugin catalog is separate from npm publication. See docs/PUBLISHING.md in the source repository for the checklist.
