dsh-sidebar-control

Agent-facing control plane for the DSH right Sidebar. One tool (sidebar) plus a compact system-prompt overview, over a registry of tab kinds rather than a second browser.

The DSH right Sidebar holds tabs of several kinds: the built-in browser, the terminal, file previews, plugins' own pages. This plugin gives the model a hierarchical overview of that workspace and the navigation verbs — open / focus / close / expand — while the per-kind content actions stay with the tools that own them (browser pages with browser_*, files with the file tools, terminals with terminal_*).

sidebar status                     → tree: tabId, kind, title, url/path, active, mounted
sidebar open {kind:"browser", url}        → opens the user-visible in-app browser tab
sidebar open {kind:"browser", url, tabId} → navigates that existing browser tab in place
sidebar open {kind:"file", path, line?}   → document-preview tab (relative paths resolve against the Session directory)
sidebar open {kind:"terminal"}            → opens a visible terminal tab
sidebar expand {tabId, mode?}             → content + the exact tools for it (mode: text/elements/html on browser tabs)
sidebar focus / close {tabId}      → navigation verbs

Why this exists (and what it does not do)

DSH's built-in sidebar browser ships user-facing only — it registers no model tools. The ecosystem now has solid MIT plugins that bolt agent control onto it; this plugin is designed to compose with them instead of competing:

• browser tab content is read/clicked by dsh-agent-browser (browser_read, browser_act, browser_screenshot, browser_navigate, …). sidebar expand on a browser tab asks that plugin's Client provider for the page text and hands the tabId over.
• file tabs are opened through the document-preview resource address (dsh-resource://file/session/<sessionId>/<path>); expand returns a bounded text excerpt and the real path, and the file tools do the rest.
• terminal tabs are opened through the Sidebar terminal type; expand names the terminal_* tools. v0.1 does not claim identity between a visible terminal tab and a terminal_* tool session — it says so instead of pretending.
• unknown kinds (other plugins' pages) are inventory-only: listed, focusable, closable, but not readable.

What this plugin does not do: it never creates a webview, never starts a loopback bridge, and never re-implements page reading. It talks to the renderer through the Cordis Inspect channel (one Client provider, SidebarControl) and to the Sidebar through the public ctx.sidebarRight service.

Platform facts

• The Sidebar is per Session. Commands act on the Session whose Sidebar is on screen; a Session whose sidebar store was never adopted has nothing to act on, and the tool reports that ("not on screen") instead of writing into a surface nobody is drawing.
• The desktop shell isolates the in-app browser from the DSH Host. sidebar open {kind:"browser"} mirrors that policy before the browser plugin could surface a bare ERR_FAILED (-2): any loopback hostname on the DSH Host port is refused with an explicit policy error. file:// is unavailable by design; file tabs go through the resource address.
• Progressive disclosure. The prompt section lists at most a handful of tabs as an overview; full content and the exact follow-up tools appear only in sidebar expand results, so the context budget follows actual need.
• Another Session is reachable (0.2.x). Pass sessionId to status/open/focus/close/expand. Read-only status needs no switch. Navigation commands write through the target Session's live sidebar store when it has one (status marks that Session live-store) and therefore do not touch the view at all; a Session that has not been displayed since the app started has no live store, so the app is asked to show it (the same path a session-row click takes) and then restores the previous Session — receipts say which Sessions were involved and how long the switch took (measured ~130–160 ms on 0.2.1-rc.3, so roughly 200–300 ms of another conversation). The first operation on such a Session warms its store: every later operation on it is seamless for the rest of the app run. switchView: false refuses the visible switch instead of flashing the view (useful for Agents that must not disturb the user). expand of a file tab works without a switch (the Host reads the file directly) and relative paths resolve against the target Session's working directory. Two further limits stay, both from the adopted browser plugin's session-scoped contract: the content of another Session's browser tab cannot be read (only its metadata is returned), and an existing browser tab of another Session cannot be navigated by tabId (open a new tab there instead). Why the first-touch switch cannot be removed from this side, and the upstream API that would remove it, are documented in docs/upstream-sidebar-store.md.

Install

dsh plugin --profile desktop add github:Dax233/dsh-sidebar-control

Or through the Plugins page (github:Dax233/dsh-sidebar-control). The bundle patch mounts the Host half; package.json's dsh.client loads the renderer half. The plugin manager can apply the bundle live; otherwise restart DSH Desktop.

Local development install (a file: copy, so reinstall after edits):

dsh plugin --profile desktop add file:F:/github/dsh-sidebar-control

Recommended companion (browser content): github:guanhaisen/dsh-agent-browser.

Configuration

Optional, in the profile cordis.patch.yml:

- id: dsh-sidebar-control
  name: dsh-sidebar-control
  config:
    maxPromptTabs: 8      # tabs rendered in the system-prompt overview
    maxTreeTabs: 40       # tabs rendered by sidebar status
    maxExpandChars: 4000  # expand content budget (200..20000)
    maxFileBytes: 524288  # files larger than this are not excerpted

Model experience

One tool (sidebar), plus a short prompt section that names the workflow and the last seen tab tree. Results are plain text trees/receipts; nothing is injected into the conversation beyond what the model asks for. The tool never touches page content itself, so prompt-injection surface stays with the tools that do (and those already wrap page content as untrusted data).

License

MIT.