dsh-sidebar-reader
In a curated listMaintenance: Activelafantasque520/dsh-sidebar-reader
Let the model read the current session's sidebar: the active tab, open tabs, terminal screens and browser page text (page text is desktop only; the terminal channel needs DSH >= 0.2.1-rc.2).
$ dsh plugin add dsh-sidebar-readerInstall
dsh has no central install command — add this plugin’s entry (documented in its README) to your profile or patch config, then restart.
How installs work0
stars
0
forks
JavaScript
Language
MIT
License
2026-11-03
Created
2026-11-16
Last push
README
dsh-sidebar-reader
Let your model read the current session's sidebar on demand: the active tab, official right-sidebar tabs, terminal screens, and browser page text.
Once installed, just tell the model "read my sidebar" or "show me my terminal" — the model triggers a capture through the plugin's command channel and receives a snapshot within seconds.
How it works
model ──POST /sidebar-reader/command──▶ host route (lib/index.js)
client polls every 2s
▼
client capture (lib/client.js, runs in the Web UI)
├── active tab (sidebarRight.active(), active pane's active tab)
├── right-sidebar tabs (sidebarRight.openTabs, current session only)
├── terminal screens (remote.terminal.attach, first snapshot → ANSI-free,
│ incremental vs. the previous read)
└── web page text (webview innerText, desktop only)
│
client ──POST /sidebar-reader/latest──▶ host keeps the latest snapshot only
model ◀──GET /sidebar-reader/latest─── host
Agent tool (auto-registered)
The plugin's host side registers a native agent tool sidebar_read with the DSH tool registry: it appears in every new session's tool list automatically, and one call returns the full snapshot — no HTTP, no configuration. The "prefer this over pwsh/web_fetch for terminal or page questions" preference is encoded in the tool description itself.
The only parameter is delta (default true): terminal screens are incremental after the first read. Pass delta: false when you need every terminal's complete screen again.
The HTTP endpoints below remain as a fallback channel.
Features
| Channel | Source | Notes |
|---|---|---|
| Active tab | sidebarRight.active() |
The active pane's active tab (id / kind / title / contentId), plus expanded telling whether the column is showing |
| Right-sidebar tabs | sidebarRight.openTabs live snapshot |
Every open tab of the current session (tabId / kind / contentId) |
| Terminal screens | remote.terminal.attach first snapshot frame |
Full screen including scrollback, ANSI-stripped, then incremental (see below); input control is released right after reading |
| Web page text | Sidebar browser webview innerText |
Desktop (Electron) only; strict mode, see below |
Three passes on terminal text
A raw xterm screen is a poor thing to hand a model: it is full of color/cursor escapes, it repeats the whole scrollback on every read, and it can be arbitrarily long. So every terminal screen goes through three passes in lib/client.js:
| Pass | What it does | Why |
|---|---|---|
stripAnsi |
Drops ANSI/OSC escape sequences and invisible control characters | Escapes are pure token cost and unreadable |
screenDelta |
Remembers each terminal's last reported screen (per terminal id) and returns only the new part; falls back to the common line prefix when the screen was redrawn; returns the whole screen when nothing matches, so text is never lost | Repeated reads cost a fraction of the context |
capScreen |
Keeps only the tail, max 500 lines / 20 000 chars per terminal | Scrollback can be huge; the newest output is at the bottom |
Each terminal entry therefore carries screen (clean text), delta (false = this is a complete screen, i.e. the first read of that terminal or an explicit delta:false) and truncated (true = the caps kicked in). delta:false also refreshes the delta baseline, so the next incremental read stays consistent.
Strict mode for web page text
The page is read only when the active tab is a browser and the column is expanded. Otherwise page is null.
Why: when a sidebar browser tab moves to the background, its webview stays in the DOM with zero width, and the fallback picker would still return it — the model would receive a page the user is not looking at. null is better than the wrong page.
What you get
| What you are looking at | Readable | Notes |
|---|---|---|
| A page in the sidebar browser | ✅ title + URL + text | Must be the active browser tab, with the column expanded |
| A file / document tab | ⚠️ address only | You get its contentId (e.g. dsh-resource://file/...); read the file by that path |
| A terminal | ✅ screen text | ANSI-free; first read is the full screen, later reads add only new lines (max 500 lines / 20k chars per terminal) |
| Plain web build (browser access) | ❌ page text | No webview; tabs and terminals still work |
Snapshot example
{
"page": { "title": "…", "url": "https://…", "text": "…", "chars": 1235 },
"right": {
"sessionId": "session-…",
"expanded": true,
"active": { "id": "tab5", "type": "browser", "title": "Browser", "path": "sidebar://browser/…" },
"tabs": [
{ "id": "tab3", "type": "files", "path": "sidebar://files" },
{ "id": "tab5", "type": "browser", "path": "sidebar://browser/…" }
]
},
"terminals": {
"sessionId": "session-…",
"terminals": [
{ "id": "…", "title": "…", "state": "running", "screen": "…", "delta": false, "truncated": false }
]
},
"commandId": "cmd-…",
"receivedAt": "2026-11-08T13:11:42.350Z"
}
Design disciplines:
- Scope — everything is filtered to the session on screen before leaving the browser; other sessions are unreadable;
- Never throws — each channel degrades independently; one failing channel doesn't break the others;
- Transparent — a toast is shown on every read;
- Revocable — all side effects live in
ctx.effectand are cleaned up when the plugin is disabled.
Install
More in CLI & Terminal
deepseek-harness-docker
by runzhliu
Community Docker and Kubernetes packaging for DeepSeek Harness (@deepseek-ai/dsh), with a hardened image, Compose stack, Helm chart, Web UI, and headless CLI.
deepseek-harness-android
by dphmoblie
Android app for managing a local DeepSeek Harness Ubuntu user space: runtime install/reset, Ubuntu terminal, optional Shizuku shell, and an embedded loopback-only Harness web UI.
dsh-subagent-codex
by mjylfz
Adds a subagent_codex tool that delegates tasks to the local OpenAI Codex CLI, running the work in an isolated external process.
mnemon
by mnemon-dev
LLM-supervised persistent memory for AI agents — graph-based recall, cross-session knowledge, single binary. Works with DeepSeek Harness, Claude Code, OpenClaw, and any agent runtime.
