Workspace Manager

One panel for DeepSeek Harness workspaces: rename, delete, rebuild onto another directory, merge, and move single sessions between them — without losing anything. A session's working directory is baked into its stored header and into the session-store directory name, so the UI cannot change it after creation; this plugin does that through the supported persistence seam.

What it does

The manager lives in its own central panel, opened from its icon in the sidebar's panel row (工作区管理 / Workspace manager):

Action	What happens
重建… / Rebuild…	Pick a new directory for a whole workspace. Every session it holds moves with it, the record at the new path inherits the title and the sidebar position, and the old record is removed once it is empty.
合并… / Merge…	Pick other workspaces. Their sessions move into this one's directory, and a source record is dropped once nothing is left in it.
重命名 / Rename	Change only the sidebar title. The directory is untouched.
删除 / Delete	Remove the workspace record. Its sessions are archived first, and the directory with all session files stays on disk — the workspace can be added back at the same path later.

Every row also reports, before anything is clicked:

• how many sessions the directory really holds;
• 占用 N / N held — how many of those cannot move right now, because a turn is running (running) or because the app still holds the session log (open, which makes the store refuse to publish a successor artifact). Rebuild, merge and delete stay disabled on such a row, and the held sessions are named in the row, so a refusal is visible up front instead of arriving after the fact.

Single sessions still have their own entry: 移动到工作区… / Move to workspace… in the session row's ... menu moves just that session to a directory. If the target is already a workspace, the session joins it; if not, the workspace is created.

The move keeps the session id, title, every committed event (history, tool calls, approvals) and the format version, then regroups the session under the workspace that owns the target directory. If that workspace does not exist yet, it is created.

How the move works

The session store is append-only and refuses rewritten history, so the move materializes a successor artifact through the supported persistence seam instead of editing bytes:

1. sessionPersistence.stat(id) — read the header (id, version, cwd, lineage).
2. sessionPersistence.open(id, 'read') — copy every committed event. A read handle never takes write ownership, so this works while the UI holds the session open.
3. Park the old session directory (atomic rename, same volume), so the id leaves the store before the successor claims it.
4. sessionPersistence.create({ ...header, cwd }, { inheritedEventCount }) + append(events) + flush() — the backend writes a correct header frame with the new cwd in the current format.
5. Read back stat(id) and verify the stored cwd; then update the workspace registry: resolveByPath(target) ?? create(target), attachSession(id), and detachSession(id) on the old workspace.

Failure handling: if step 4 fails, the parked directory is renamed back and the result says the move was rolled back. A rebuild rolls every session it already moved back to the old directory, and removes the record it created at the target.

Two different states are refused up front, and the difference matters:

• running — a turn is in flight. A live agent keeps the old directory as its process cwd and permission root, so moving under it would be wrong. The signal is the turn boundary in the session log (turn/start with no turn/end after it), not agents.get(id): an idle session that the app merely has open also has a live agent, which is why the sidebar can say 「空闲」 while a naive check says "running".
• session-open — the session is open in the app, so its log is in memory and the store owns that id. create then refuses the successor ("already exists"), so the move cannot work until the tab is closed; refusing here names the real reason instead of failing halfway through a rebuild.

Two more guards keep a half-done move from eating history:

• A header-only session (no committed events) is refused, because create writes nothing and only the first append publishes — moving it would delete the original and produce no successor.
• If the store still reports the id right after the directory was parked (an index-based resolution), the parked copy is restored and the move reports store-cached instead of failing with a misleading write error.

After a move the registry is settled in the same process. WorkspaceRegistry builds its canonical-cwd header index once, in [Service.init], and attachSession validates against that index (readSessionHeader prefers a live Session, then the cached header, and only re-reads the store when neither exists). The plugin therefore drops the moved session's stale index entry before attaching — the fields are plain Maps on the service instance — so the next validation re-reads the rewritten header from the store. That is the state a restart would produce, without a restart. The Client then asks sessions.refreshProjections(id) to rebuild the derived view, and any other workspace that still lists the session is detached.

If the attach is still refused (for example while the session is open in a tab, where a live Session still answers with the old header), the move is recorded as pending and settles automatically on the next start.

Workspace membership is verified, and repairable

attachSession is judged by the canonical-cwd index, which the harness builds at startup. A session whose header was rewritten during this run may therefore be rejected until the harness restarts. The Host never reports that as a plain success:

• moved — header rewritten, read-back matched, workspace reported the session.
• moved-unattached — the file moved but the workspace did not accept it yet.
• attached — the session already lived in the target directory (an earlier move left the membership behind); this run only completed the membership.

Repeating the same move is the repair action: when the stored cwd already equals the target, the plugin skips the file work and only settles the workspace membership, so a partial move is finished by the same command that made it. workspace-unattached means even that failed; restart DSH and repeat once the index is fresh.

The unattached case also heals itself on startup: the Host records the pending (sessionId, target) pair in .remaker-pending.json, and two seconds into the next boot — after the registry has rebuilt its index — it settles the membership and clears the entry. So a move that landed its file but not its membership needs one restart, not a second click.

Both halves also refuse to break application start: a throwing Client factory degrades to an absent plugin instead of failing the Web boot gate (which fails the whole app with 1 entry did not activate), and a throwing Host apply is logged to .workspace-manager.log instead of failing the entry.

The workspace title follows its directory, unless you renamed it

create(path) derives a title from the last path segment, so a title that still equals its own basename was never typed by the user: a rebuild carries that kind of title over to the new directory name (ce rebuilt onto D:\Github\workspace test becomes workspace test). A title that differs from its basename is treated as a deliberate rename and is preserved untouched.

Records created by an earlier build ignored that rule and kept the pre-rebuild title. The workspace row in the manager offers a one-click repair for exactly that shape — the button only appears when the stored title differs from the folder name — and it calls retitleWorkspace(workspaceId, title), which uses the supported Workspace#setTitle and needs no restart.

Install

From npm — the published package is dsh-workspace-control; the repository is named after the panel it ships:

dsh plugin --profile web add dsh-workspace-control

Or from a local checkout:

plugin_manager install_bundle target=D:\Github\dsh-workspace-manager

Then verify the loader row exists:

cordis_inspect_query Config.listConfigs name=dsh-workspace-control

Replacing an older install: plugin_manager remove_bundle target=dsh-workspace-remaker drops the previous bundle and its sidebar entry.

The panel reports the same activity the sidebar does

A workspace row says N running only when a session really is mid-turn, and it gets that answer from the same two sources the shipped sidebar uses:

useSessionStatus().get(sessionId)?.running ?? useSessions().byId[sessionId]?.running

A session the app merely keeps open is shown as N open · idle, its buttons stay enabled, and the host releases it before touching any file. When neither source knows the session, the panel falls back to the host's own summary instead of inventing an answer.

The panel

The panel is a card grid rather than a table: each workspace is an article.wsm-card carrying a status dot, its name, the canvas path, activity badges and four actions (rebuild / merge / rename / delete). Actions disable themselves only while a session is running — a workspace that is merely open still operates, because an idle holder is released before the move.

Styling rides on the host theme instead of hard-coded colours, so the panel follows the Harness light/dark switch:

.wsm-card {
  border: .5px solid var(--dsw-alias-settings-card-stroke);
  border-radius: var(--dsw-radius-xl);
  background: var(--dsw-alias-settings-card-fill);
}
.wsm-card:hover { background: var(--dsw-alias-interactive-bg-hover); }

While the first read is in flight the grid shows three pulsing skeletons, so the panel never flashes an empty state. prefers-reduced-motion turns the pulse off.

Tests

node --test

The suite runs without the harness: test/fakes.js stands in for the storage, the workspace registry, sessionController and the agent registry, publishing real files into a throwaway DSH_HOME.

• paths.test.js — samePath / titleFollowsPath / encodeId edge cases.
• move.test.js — the whole move transaction: success, a held session, a turn in flight, an empty log, a failing publish, a successor that reads back short, a store that still lists the parked id, and the repair path when a move is repeated.
• contract.test.js — loads client.js in a vm and asserts the Client mounts exactly the descriptors the Host registers (ids, namespaces, parameter wires) and that both locale dictionaries carry the same keys.

Notes and limits

• Moving a session does not move the files on disk. Only the session's recorded working directory changes.
• Nested workspaces are legal in current DSH builds, so nesting is not refused; only a target that cannot host a workspace (missing, not a directory, not writable) is.
• Sessions stored in an older format report migration-required; open them once so the harness upgrades them, then move them.
• A forked session keeps its parentSession link and its inherited event offset; the parent stays where it is, so move parents first if you want the whole tree in one directory.
• The move is refused while the session runs. Stop it, move it, re-open it.
• A session the app has open but idle used to be refused: the store keeps its write ownership, so no successor artifact can be published for it. The plugin now releases an idle holder first (#releaseHeld disposes the agent's own Cordis scope and re-reads the registry to confirm) and only refuses when the release did not happen. A session mid-turn is still refused: it has real work to finish.
• Every seat draws the same occupancy chips from listWorkspaces, so a blocked workspace is visible before a button is pressed.
• 删除 / Delete removes the record only; the directory and its session files are left on disk, and the sessions are archived so they are easy to find again.
• The plugin never edits committed bytes: it publishes a successor through sessionPersistence, so the store's no-rewrite rule and framing checks stay intact.

Verification

# 1. the loader row exists and the bundle applied
cordis_inspect_query Config.listConfigs name=dsh-workspace-control
# 2. the Client half registered its three seats (live Slots tree)
cordis_inspect_query Slots.listSubTree root=sidebar.panellist
cordis_inspect_query Slots.listSubTree root=main
cordis_inspect_query Slots.listSubTree root=sidebar.workspaces.session.menu.item

sidebar.panellist and main both carry the id workspace-manager; the first is the sidebar button, the second the panel it dispatches. The same panel is also reachable from settings.section (id workspace-manager-settings).

Finished results retire on their own after a few seconds, so an operation that succeeded cannot leave a banner sitting in the panel.

The offline checks (no harness needed) cover the refusal paths, inspect, and the boot self-heal against a fake registry:

node .scratch/smoke.mjs

The same operation is reachable by the agent as the workspace_manager tool, so the move can be exercised without the UI:

workspace_manager { "action": "inspect", "sessionId": "session-<uuid>" }
workspace_manager { "action": "move", "sessionId": "session-<uuid>", "target": "D:\\target" }

action=listWorkspaces reports every workspace with its real session count and its blockers (running / open), which is the same data the panel draws its occupancy chips from. action=inspect is read-only and reports the recorded directory, the owning workspace, whether the session runs, and the storage format. action=move performs the move. Every step appends a line to .workspace-manager.log next to this package, which is the first place to look when a move refuses.

The cached title follows the header

The Session list builds its projection column from sessionProjectionCache.cachedSnapshot(header), and that read only matches a stored record whose lifecycle identity (formatVersion, createdAt, cwd, isSeeded) equals the header it is handed. A move rewrites cwd in the header, so the record written before the move stops matching and the sidebar renders the row as “未命名” (unnamed) even though the cached title row is intact.

A move therefore ends with a cold projection read off the complete log (sessionProjectionCache.coldSnapshot(header, inheritedEventCount, events)), which folds the log against the new identity and writes a fresh checkpoint back. action=move reports the outcome as projections: { asOfSeq, keys }, and repeating a move to the directory a session already lives in re-runs only this repair step.

Reloading the plugin without restarting the harness

The host half is reloaded by toggling its entry through the plugin manager — plugin_manager set_plugin with target: "include:workspace-manager", enabled: false then enabled: true — which re-imports index.js in place. The client half is refreshed with the page. Neither needs a DSH restart.