summerscar's DSH plugins

Sync backup

@summerscar/dsh-mnemon-sync
Plugin Tools includes a Web UI

Back up and restore dsh-mnemon memory and user-level skills to S3-compatible storage or WebDAV: memory spaces, project documents, runtime memory and skill files, with scheduled backups and selective restore.

Install

 dsh plugin --profile web add @summerscar/dsh-mnemon-sync

Replace web with your profile name. Restart the profile after installing; the plugin then shows up under Settings → Plugins.

About

From README.md in 0.1.15.

dsh-mnemon-sync

Back up and restore dsh-mnemon memory — archive it to S3-compatible storage or WebDAV, then pick a snapshot to bring back.

npm version npm downloads license DeepSeek Harness 开放插件市场

dsh-mnemon keeps long-term memory in local SQLite databases and Markdown files, and DSH keeps your skills as SKILL.md folders on disk. This plugin adds off-site backup and on-demand restore for both: once installed, the memory system gains a "Sync backup" tab.

Sync backup page: choose components, target and cadence

Install

dsh plugin --profile web add @summerscar/dsh-mnemon-sync

Restart DSH, then open the memory system to find "Sync backup". The plugin still runs without dsh-mnemon installed — it just cannot show its tab.

Desktop: install it from the desktop app's plugin page. The desktop profile is managed exclusively by the Electron app; dsh plugin --profile desktop is rejected by DSH itself.

What it does

Pick any subset of five components. The original three are selected by default; both skill components start unselected, because carrying a machine's skills is a deliberate act — a restore writes into the skill roots, so it should not happen as a side effect of the first backup:

Component Contents Default
Memory spaces <MNEMON_DATA_DIR>/data/, one mnemon.db per space plus the registry selected
Project documents <MNEMON_DATA_DIR>/documents/, index.json and every Markdown file selected
Runtime memory <MNEMON_DATA_DIR>/runtime/, memories.json and both projections selected
Skill files The real skill trees under $DSH_HOME/skills and $DSH_AGENTS_HOME/skills not selected
Skill data <MNEMON_DATA_DIR>/skills/: ledger.json, evals.json, versions/, trajectories/, market/, trash/ not selected

Skills are split into two components because they behave differently: a SKILL.md is what makes the model capable, so it must travel to a new machine, while the ledger and trajectories are local evolution history that is large and recomputable — leave it off if you would rather not upload it.

Project-level skills are out of scope. <project>/.dsh/skills and <project>/.agents/skills belong to a repository and travel with git. Copying them into a memory backup would give the same content a second source of truth, and a restore could overwrite your checkout, so only the user's own layer is backed up.

Archives go to S3-compatible storage (Cloudflare R2, MinIO, Backblaze B2, Aliyun OSS …) or a WebDAV server, on a timer from hourly to weekly. You can list the archives already on the target, inspect one, and restore only the components you choose. Restore previews by default; every file is SHA-256 verified before upload and again before restore, and live databases are copied through node:sqlite's online backup().

OS bookkeeping files (.DS_Store, Thumbs.db, desktop.ini, ._* resource forks, editor .swp / ~ backups) are never packed. An older archive that carries them still restores — those entries are skipped rather than failing the run.

Credentials live in a 0600 file outside the DSH profile and are never packed into an archive. No runtime dependencies outside Node.

Configuration

Everything is set on the "Sync backup" page. Saved secrets are shown only as a mask; leaving the mask (or the field) unchanged keeps the stored value.

The memory directory is detected automatically (MNEMON_DATA_DIR, else ~/.mnemon) and the page always shows what is actually being read. dsh-mnemon's storage root is not always the default one — its storageScope may also be workspace, custom or workspaces — and a third-party Source is never told which, so the backup scope offers a custom data directory field. Leave it blank to auto-detect; fill it in to read exactly that directory.

See the 中文说明 for field-by-field examples for S3 / Cloudflare R2 and WebDAV.

Development

This repository is a pnpm monorepo; the plugin lives in plugins/dsh-mnemon-sync/.

git clone https://github.com/summerscar/dsh-plugins.git
cd dsh-plugins
pnpm install
pnpm --filter @summerscar/dsh-mnemon-sync test
pnpm check

The code is the developer documentation.

License

SATA 2.1. Versions up to and including 0.1.11 published on npm remain MIT.