dsh-session-tabs

JetBrains-style session tabs for the DSH Web GUI: one tab per chat session, in a strip at the top of the conversation panel.

│ ● Add DeepSeek…  × │ ● Fix the login bug │ ● vendor sync │

A session you sent a message in keeps its tab. A session you merely opened is one disposable tab that the next open replaces, so clicking through the sidebar never leaves a trail of tabs.

The rule

This is the editor-tab behaviour from JetBrains, with sessions in the role of files:

Session	Tab
You sent a message in it	An ordinary tab that stays
You only opened it	The single "preview" tab, replaced by the next open

The stick signal is not tracked by this plugin. The session list already carries it as blank:

A blank session is the selected Workspace's provisional New Session row.

blank === false therefore means the session has real content. The host sets it on the first accepted prompt (promptAttempted, set synchronously before the first await) and its reducer deliberately refuses to re-blank a session afterwards:

handleBlank(blank) {
  if (blank === this.blankBit) return;
  if (blank && (this.promptAttempted || this.running)) return;  // sticky
  ...
}

So "sent a message" is the host's own durable notion of it, already computed and already sticky. Delegating to it means the strip cannot disagree with the rest of the UI about whether a session has been used, and there is no local bookkeeping to drift or to persist.

Two properties follow, and both are tested:

• Order is engagement order, not recency. Revisiting an old tab does not reshuffle the strip under the pointer.
• The preview is always last, because it is the one that will be replaced.

Subagent sessions are excluded: they are reached through their parent's breadcrumb rather than by switching the conversation, so a tab for one would misrepresent what the strip does. Nothing is capped or recycled: every engaged session keeps a tab in engagement order, and the row scrolls sideways when the tabs outgrow it (plain wheel included), with the active tab kept in view.

Placement — the first row of the conversation

The strip is registered in conversation.input.dock — a list seat above the composer, shared with the todo, queue, and goal docks — and drawn in a row this plugin anchors ahead of every row the conversation already has, so the tabs are the first thing in the conversation:

div.root                            ← the conversation column (flex column)
├── div[data-session-tabs-host]     ← the strip's row (inserted by this plugin)
├── header                          ← title row + Chat/Trajectory view tabs
└── div[data-conversation-content]  ← transcript + composer

Why it is not a seat

The app declares no seat above the conversation content:

• conversation.header is a single seat declared and occupied by the shipped ui-conversation plugin. A second registration at the same priority throws single slot "conversation.header" already has a registration …; registering at apply time — before the seat exists — throws slot "conversation.header" is not declared, which fails the plugin's whole fiber and aborts web boot.
• Its only unoccupied child, conversation.header.leading, sits inside the title bar (the window drag region), so it cannot hold a full-width row.
• The header's other children (…session.header.actions, .utilities) are narrow control clusters beside the title.

Placement is therefore done in the DOM: the dock registration supplies the seat and the session props, and the component inserts a host row as the conversation column's first child, then renders the strip into it with createPortal (the shell exposes react-dom as a platform seed word, so a client bundle may require it). The column is reached from the seat's own placeholder, so it is always the column that seat belongs to. The host is re-anchored whenever it leaves the document, so a re-render of the conversation column cannot strand the strip in a detached node.

The row sits outside the header's data-window-drag region, so chips click and scroll normally; window dragging keeps the title row and the sidebar.

Why the memory is the whole order, not the visible tabs

stepTabs returns the strip and the order to remember next, which is the full engagement order — never the window that happened to be rendered. That looks like a detail; it is the difference between a still strip and a rotating one:

Remembered	What 6 re-renders of 10 engaged sessions show
the visible tabs (a window)	5 distinct windows — the strip rotates on every re-render
the full engagement order	1 window, unchanged

Typing re-renders the strip constantly (the session store ticks), so a rotating strip is exactly the flicker a user sees while typing. verify.mjs replays this loop through stepTabs — the function the component calls — and fails if more than one window appears.

Why the memory is module state, not a ref

Switching sessions remounts the conversation Factory occurrence — its key carries the session generation — so a ref dies the moment you open another session. A dismissal kept in a ref therefore came back on the very next click elsewhere, and the engagement order died with it, reshuffling the tabs on every switch. dismissedTabs and engagementOrder are module state for that reason, and verify.mjs drives real remounts (fresh refs, retained module memory) to assert both survive.

A page reload would still wipe them, and the shell reloads its page whenever the app restarts, so the same memory is written to localStorage under dsh.session-tabs.v1 — the dsh.<name>.v1 convention shipped client plugins use (the conversation's transcript width, the right sidebar's layout) — and read back on load. That key lives in the dsh-app://app origin, which the desktop registers as standard + secure, so it survives restarts. A blocked store, a missing key, or a corrupt payload all read as a cold start rather than an error, and the offline suite covers the round trip, the payload validation, and a seeded "page load".

The scrollbar is the shell's

The strip sets no scrollbar-width/scrollbar-color, on purpose. ui-theme installs unscoped ::-webkit-scrollbar rules (--dsh-scrollbar-width = 5px, a rounded thumb in --dsw-alias-scrollbar-bg-l1, a hover colour from --dsw-alias-scrollbar-hover-l1), so the bar matches the transcript's and the sidebar's. Chromium ignores those rules for any element that sets the standard properties, and the strip used to set scrollbar-width: thin — which is why it had a fat native macOS bar while the rest of the app had a thin themed one. If a profile disables ui-theme, the strip's bar goes native along with every other scrollbar.

The frame is browser-style, and why the rail is paint

The strip is drawn as a browser tab strip rather than a row of chips: a bar a shade off the pane's own surface, the tabs sitting on a hairline rail along its bottom, and the active tab punched down to the pane colour so it reads as connected to the row beneath it. The chip is 30px tall with 8px shoulders and square feet (borderRadius: "8px 8px 0 0") on a 36px bar.

Three things about how that is drawn are deliberate, and each was forced by how CSS composites a scroller:

• The bar's fill is a translucent neutral (rgba(127,127,127,0.12)), not a surface token. The shell's surfaces are raised above the pane in dark mode — --dsw-alias-bg-base is the pane, and --dsw-alias-bg-layer-1…3 are lighter than it — so no single surface token can be "a bar slightly off the pane" in both themes: one theme would erase the bar, the other would paint a raised card. A 12% neutral over the pane reads as a raised bar in the dark theme and a recessed one in the light theme, from one value.
• The active chip is filled with --dsw-alias-bg-base — the pane's own colour. That, not a border or a shadow, is what merges the tab into the row below.
• The rail is the strip's bottom row of background, not a border-bottom. A child cannot paint over its parent's border, and the strip clips its own overflow because it is the horizontal scroller — so the active chip could never cover a border to break the rail under itself. Painted as the strip's own background, the chips are composited on top of it: the active chip's opaque fill hides the rail beneath it, while the gaps between tabs and the inactive tabs keep it. The strip carries no bottom padding for the same reason, so a chip's box reaches the rail.

Degradation

Conversation shape	Where the strip draws
Main column whose parent has a header child	the anchored row — the column's first row
Embedded occurrence (subagent preview — no header)	nothing; a dock strip there would be noise
Unrecognized shape (no [data-conversation-content] ancestor)	the dock it was registered in

The last row is the safety net: a future app version can move the strip, but not lose it — and none of these paths can fail web boot.

Why the registration waits for the declaration

conversation.input.dock is declared by the Conversation Factory, whose entry mounts after this plugin's two services (slots, sessions) resolve, so the registration is made inside ctx.slots.inject("conversation.input.dock", …): the callback runs when the seat is declared, now or later, and never runs if it is not. A bare ctx.slots.register({ name: "conversation.input.dock" }, …) at apply time throws slot "conversation.input.dock" is not declared (a parent entry's children table must declare it), and a throw inside apply fails the plugin's whole fiber — the desktop shell then replaces the app with its "could not start" dialog.

Layout

package.json    dsh.client declaration (platform: web) + local package metadata
lib/index.js    node half — empty apply(), gives the Loader a host-side row
lib/client.js   browser half — the strip (inlined policy) + top-row anchoring
lib/policy.js   the recycle rule, imported by verify.mjs
verify.mjs      policy + bundle contract + rendered chips (node verify.mjs)

lib/client.js cannot import lib/policy.js: a client bundle is a window.__ModuleLoader__ bundle that may require only the shell seeds react and react/jsx-runtime, and the browser half ships as-is with no bundler. The policy is therefore inlined verbatim, and verify.mjs asserts the two copies agree — on every constant, on the whole deriveTabs derivation across its fixtures, and on chipLabel, fitLabel, and graphemes across their cases — so editing one without the other fails the run.

Titles and widths

Every chip is the same width (TAB_WIDTH, 190px, border-box) regardless of its title: a strip whose tabs widen and narrow with their text reads as noise, and one busy title would shove its neighbours sideways. Inside that fixed box the label takes whatever the state dot and the reserved close slot leave (flex: 1 1 auto, min-width: 0) and ellipsizes there — so a longer title, or a larger shell font, trims the label instead of resizing the tab.

The label text is still cut first, on word boundaries when a whole word is reasonably close to the limit, so a chip reads as a phrase rather than a chopped token:

Session title	Chip
Add DeepSeek peak/off-peak session tabs	Add DeepSeek…
a_very_long_single_token_filename	a_very_long_single…
Fix the login bug	Fix the login bug

The full title stays reachable as the chip's tooltip, so truncation never hides which session a chip is.

The tab context menu

Right-click a chip for the shell's own menu material (MenuSurface + MenuItemButton from dsh-client-ui-primitives, portaled to the body and clamped into the viewport):

Row	What it does
Rename session	edits the title in place — the chip's label slot becomes an input, seeded with the full title; Enter commits, Escape and blur abandon
Close all left	dismisses every tab before this one
Close all right	dismisses every tab after it
Close all others	dismisses every other tab

Two rules keep it honest:

• Bulk closes act on tabs, never sessions. They only add ids to the dismissal set (closeRangeIds), which is the same memory the × uses.
• The active tab is never left behind. If a bulk close dismisses the session being read, the conversation follows the tab that was kept — so the strip and the transcript cannot disagree about which session is up.
• The right-clicked tab is never closed by any of the three, and a row with nothing to do is disabled rather than hidden, so the menu keeps its shape.

Rename goes through sessions.rename(sessionId, title) — the service call the sidebar's own rename dialog ends at. That dialog itself cannot be raised by a third-party plugin (its request store is plugin-local state), which is why the strip carries its own inline editor.

Interactions

• Click a chip — focus that session (ctx.uiWorkspace.openSession, looked up per click). This is not ctx.sessions.open: the sessions service has no such method, which is why an early build's chips looked dead while the sidebar navigated fine.
• Click × — drop the tab from the strip, and re-render immediately: the dismissal set is a ref (it is only read while deriving the strip), so the handler also bumps a render ticket — without it the chip lingered until an unrelated store tick and the × looked dead. The session itself is untouched; deleting a session stays the sidebar's job. Closing the active tab first selects a neighbour, so the strip is never left pointing at a session it no longer shows.
• The × is a reserved slot, not an insertion. A closable chip always carries the control's 16px (plus gap) and toggles only its visibility: always visible on the active chip, visible on hover for the rest. Appending it on hover widened that chip and pushed every chip after it sideways under the cursor. The last remaining tab and the preview carry no slot at all — neither can be closed. Hovering the × lights a small circular background under it (a tighter highlight than the chip's own), and that hover changes colours only: closeControlStyle is asserted to return identical geometry in every state.
• The last remaining tab is not closable, since an empty strip has no way back. A preview tab is not closable either — it is already disposable.
• A running session shows a green dot on its chip.

Verify

node ~/.dsh/plugins/dsh-session-tabs/verify.mjs

Covers the recycle rule (browsing leaves exactly one tab; engaging promotes it; engaged tabs survive a browse; order is stable; subagents excluded; the cap holds), the fixed-point memory (a replayed re-render loop that must not move the strip), the fitted chip labels (the pixel fit, its word-boundary and stub rules, grapheme-safe cuts, the canvas measurer's font plumbing and its no-canvas fallback, and the character budget that stands in before layout), the wheel-to-sideways mapping, the bundle contract and inline-copy parity, the dock registration (the seat and list id it claims, and that apply cannot reach register outside the declaration wait), the placement decision (anchored row, embedded silence, dock fallback), and the rendered chips including the close-control rules.

Mutation-tested rather than trusted: disabling the preview rule fails 19 assertions, breaking truncation fails 8, registering the wrong seat fails 1, reinstating a fixed tab window fails 2, a dead or vertical-only wheel mapping fails 2 and 1, a clipped (overflow: hidden) strip fails 1, and each placement guard fails its own — a row inserted below the existing ones 1, a host anchored in an embedded occurrence 1, a missing header check 1, a row that can stretch the conversation column 1, an unrecognized shape that stops falling back to the dock 1, a missing anchor marker 1, and registering without waiting for the declaration (the bug that broke web boot) 2. fails 2.

Install

The plugin goes into the desktop profile, which the DeepSeek Harness app owns. The dsh plugin command refuses that profile, so edit its files directly.

1. Quit DeepSeek Harness.

2. Add the package to dependencies in ~/.dsh/profiles/desktop/package.json:

   "dsh-session-tabs": "github:rifkyazizf/dsh-session-tabs"
   

3. Install it from the profile directory:

   cd ~/.dsh/profiles/desktop && pnpm install
   

4. Append the loader row to ~/.dsh/profiles/desktop/cordis.patch.yml:

   - insert:
       - id: ui-session-tabs
         name: 'dsh-session-tabs'
   

5. Start DeepSeek Harness. The tab strip shows at the top of the conversation panel.

To pin a version, append a tag or commit to the spec, for example github:rifkyazizf/dsh-session-tabs#a8f106b. To update, run pnpm update dsh-session-tabs in the profile directory.

For local development, clone the repo and point the dependency at the clone with "file:/path/to/dsh-session-tabs". The profile is watched live (patchReload: live) and DSH's client-HMR row polls the client bundle, so edits to lib/client.js hot-reload in an open tab. A new plugin row needs one page refresh the first time.

The strip is registered in conversation.input.dock (list seat, id session-tabs, order 10) and drawn in its own row under the conversation header.

Disable or remove

To disable the strip, comment out the ui-session-tabs insert in ~/.dsh/profiles/desktop/cordis.patch.yml. To uninstall, also remove the dependency and run pnpm install in the profile directory. Changes take effect on the next page refresh.

Known gaps

• The strip scrolls rather than recycling. With very many engaged sessions the row keeps every tab and scrolls; the × on a chip is how tabs are pruned. While it is scrolling, the shell's themed 5px horizontal bar is laid out in the strip's last row, so the row grows by 5px and the active tab's feet rest on the bar rather than on the rail. That row cannot be reclaimed — a custom ::-webkit-scrollbar consumes layout space, and the strip has to stay a scroller, since a fixed tab width is the whole point of the chip contract (see Titles and widths).
• Memory lives in localStorage, not on the host. Dismissals and the engagement order survive a reload and an app restart, but clearing site data — or reaching the GUI on a different origin — starts fresh. The engaged set itself is durable on the host (its blank bit), so only which tabs you closed and their order are local.
• A dismissal lasts until you revisit that session. Closing a tab hides it while something else is current; navigating back to that session (sidebar included) brings it back — that is the rule, not a bug.
• No drag-to-reorder, and no middle-click close. Right-click offers rename and the three bulk closes only — no pin, fork, or archive (those stay in the sidebar).
• The menu is English-only. The plugin registers no locale dictionaries, unlike the shipped client plugins, which pass their rows through ctx.locale.

License

MIT