Blocking: - B1: delete screenshot scratch PNGs — observation frames reuse one fixed per-session path and unlink after every read (finally); browser_screenshot unlinks once the bytes are in the attachment store (kept only when the file itself is the model-facing artifact). - B2: replicate dsh's browser-trust fence on /bsk-observation/*: loopback Host only, Origin must match Host, sec-fetch-site: cross-site refused, POST requires application/json; README documents the loopback trust premise and the 0.0.0.0 warning. Major: - M1: KeyedExecutor tail chains previous+task (allSettled) so a task aborted while queued cannot release the next one into the running session. - M2: client tracks the live frame per session — replacing/removing/reset revokes the old blob URL at once; loads settling after replacement never resurrect it. Minor: drop replaced thumbRefs (+on remove); beginAction fires inside the queue (no label overwrite while queued, no idle flash); half-initialized session cleanup stops through the queue with one retry; SSE (re)open refetches /state (+ snapshot flag renamed subscribed); install probe uses --version (no daemon spawn); emulate mobile documents the width+height requirement (the daemon refuses it alone — verified). Nit: single BskRunOptions declaration; tails map entry dropped on drain; SSE cleanup on res close; popOut catches requestWindow rejections; resize handle gains keyboard control (arrows, 16px steps, aria value attrs). Tests: +14 (fence rules, scratch lifecycle, queue race, blob revocation, instrumentation timing, SSE resync). 113/113 green.
9.4 KiB
dsh-plugin-browserskill
A DeepSeek Harness (dsh) tool plugin that exposes
BrowserSkill (bsk) browser automation to the model.
Each tool maps to one bsk <cmd> --json invocation: the plugin spawns the bsk CLI, parses its
structured JSON output, and returns a canonical typed value. The bsk daemon, browser, and Chrome
extension keep owning the actual browser control — this package is a thin, well-typed bridge.
Tools
| Tool | bsk command | Purpose |
|---|---|---|
browser_session_start |
bsk session start |
Open an Agent Window session; optional initial URL, window size, and device emulation preset. Returns the session id and makes it the current session. |
browser_session_stop |
bsk session stop |
Stop a session (the current one by default) and close its Agent Window. Only plugin-created sessions can be stopped. |
browser_session_list |
— (registry only) | List the sessions this plugin created, marking the current one. Foreign daemon sessions are never shown. |
browser_navigate |
bsk navigate |
Navigate the active tab, with waitUntil / timeout control. |
browser_snapshot |
bsk snapshot |
Indented aria-tree snapshot with @eN refs for interaction tools. |
browser_observe |
bsk observe |
Semantic VOM observation (read-only) with @eN refs. |
browser_click |
bsk click |
Click a snapshot ref or CSS selector (button / click-count options). |
browser_fill |
bsk fill |
Fill an input / textarea / contenteditable (clears first by default). |
browser_press |
bsk press |
Dispatch a key or combo, optionally focusing a target first. |
browser_screenshot |
bsk screenshot |
PNG capture of the tab or a ref-cropped element; inlines the image when the deployment supports image input, otherwise returns a file path. |
browser_emulate |
bsk emulate |
Apply or clear mobile device emulation on the active tab. |
Multi-session model
One agent conversation can drive several browser sessions at once:
browser_session_startreturns the session id and makes it the current session.- Every operation tool accepts an optional
sessionargument. When omitted, the call acts on the current session (the one most recently started or used); when given, that session becomes current. - Every tool result echoes the session it actually acted on, so the model never has to guess.
- The number of concurrent sessions started through the plugin is capped (
maxSessions, default 5). - Unloading the plugin stops every session it started and kills in-flight bsk processes.
Ownership boundary: the bsk daemon may be shared with other agents, terminals, or dsh
instances. The plugin therefore only ever sees and operates on sessions it created itself —
an explicit session argument naming a foreign or unknown id is rejected, browser_session_list
shows plugin-created sessions only (no daemon-wide view), and stop/unload cleanup can never touch
a session owned by another program.
Installation
The plugin follows the standard dsh bundle layout (dsh.bundle manifest + cordis.patch.yml):
dsh plugin --profile <name> add <path-or-spec-of-this-package>
dsh --profile <name>
Prerequisite: the bsk CLI must be installed and on PATH, and the BrowserSkill Chrome extension
must be connected — see the BrowserSkill README. When bsk
is missing, tool calls fail with install guidance instead of a bare spawn error.
Configuration
All fields are optional and validated through the plugin's Schemastery Config:
# cordis.patch.yml override example
- insert:
- id: browserskill
name: dsh-plugin-browserskill
config:
bskPath: bsk # path to the bsk binary (default: resolve from PATH)
defaultTimeoutMs: 120000
maxSessions: 5
# observationEnabled: true # live PiP/overlay observation (below)
# thumbnailIntervalMs: 1500 # frame cadence while a session is active
# idleIntervalMs: 8000 # idle cadence / recent-activity window
Observation overlay (PiP mini-window)
When the plugin runs inside the dsh Web UI, an observation overlay floats over the app
(registered into the shell.overlay seat): a breathing thumbnail per owned session plus its
current action and elapsed time. The card docks at the top-right of the content area (clear of
the composer and the shell's header controls) and wears the BrowserSkill product family's own
look: the overlay reuses @browser-skill/ui components (Button, cn) and its oklch design
tokens (--card, --primary, --destructive, --ring, …), status dots spec'd after the
extension popup's connection indicator, and Remix icons. The BSK utility sheet is compiled
scoped under the .bsk-obs root class (scripts/build-client-css.mjs), so the overlay looks
like the BrowserSkill extension without leaking a single selector into the host shell — and the
shell's theme cannot bleed back in.
- Lifecycle: hidden while the plugin owns no sessions; appears on the first
browser_session_start; disappears when all sessions stop (or the plugin unloads). - Focus view: status row (green/idle/red dot + session + action + mm:ss), the latest page frame (refreshes every ~1.5s while active, ~8s when idle, keeps the last frame on errors), and the action row (Interrupt + Pop out).
- Interrupt: one click kills the in-flight bsk command of the focus session (same semantics as the chat Stop button — the current action fails, the agent run may continue). Strip items carry their own hover interrupt button.
- Multi-session strip: every session gets a tile (mini frame + id + status dot); focus auto-follows the most recently active session; clicking a tile pins focus (pin badge, click again to release); errored sessions get a red edge without stealing focus; sessions the daemon lost are greyed out; prolonged daemon/browser outage shows "browser unavailable" and greys the interrupt button until captures recover.
- Drag & resize: drag the header to move the card, drag the corner handle to resize (min 240×180, max 80% of the viewport); both are remembered for the page lifetime.
- Pop out (PiP): upgrades the card into a native Document PiP window (requires a user gesture, per browser rules), sized from the current card; closing the PiP falls back to the in-page card with state intact. Browsers without Document PiP simply hide the button.
- Wire: the host serves
GET /bsk-observation/state,GET /bsk-observation/events(SSE),POST /bsk-observation/interrupt, andGET /bsk-observation/thumbnail/<attachmentId>over the dshwebServerroute seam (dsh 0.1's Typert Remote pipeline is closed to out-of-tree packages). All commands for one session — tool calls and frame captures alike — run through a per-session FIFO, because the daemon accepts only one unfinished command per session. - Trust model: these routes expose live screenshots (and an interrupt write), so they
replicate the browser-trust fence dsh applies to its own
/apiroutes: the request Host must be a loopback authority (localhost,127.0.0.0/8,[::1]), a present Origin must match the Host,sec-fetch-site: cross-siteis refused, and POST requires anapplication/jsonbody (cross-site simple requests can never satisfy that). The channel is therefore built for loopback-only serving — binding the dsh web server to0.0.0.0and reaching it through a LAN address will (deliberately) fail the fence; do not put these routes behind a non-loopback reverse proxy without adding your own authentication. - Configure with
observationEnabled/thumbnailIntervalMs/idleIntervalMs.
Behavior notes
- Cancellation: aborting a tool call (
exec.signal) kills the underlying bsk child process, matching BrowserSkill's cooperative tool-cancellation model. - UI cards: calls render as terminal cards (command line as title, output as the completed card). Screenshots additionally attach the image itself when the host mounts an attachment store and the active model route declares image input; otherwise the PNG path is returned.
- Web UI toolview (browser half): the package is dual-face.
dsh.client(platformweb) shipslib/client.js, which registers a keyedtool.call.toolviewview forbrowser_screenshot. The custom view keeps the terminal block (command + output) and, when the settled result carries an image block, resolves the durable attachment through the client session's authorizedreadAttachmentRPC and renders it with the sharedMessageImagethumbnail/lightbox atoms. Every otherbrowser_*tool keeps the stock terminal card. The bundle follows the dsh client contract: a CJS closure factory handed towindow.__ModuleLoader__.load, platform modules (react,dsh-client-ui-*) external, everything else inlined, CSS Modules compiled by lightningcss. - Errors: non-zero bsk exits surface the CLI's JSON error envelope (
code,message,hint) so the model gets the daemon's actionable guidance. - Long-running work (e.g.
bsk record) is not backgrounded viactx.jobsyet — tracked as a follow-up.
Development
pnpm install
pnpm --filter dsh-plugin-browserskill typecheck
pnpm --filter dsh-plugin-browserskill test # unit tests mock bsk; no browser needed
pnpm --filter dsh-plugin-browserskill build # tsdown -> lib/
License
MIT