Add a Release dsh plugin workflow (dsh-plugin-v* tags or workflow_dispatch)
that reads NPM_TOKEN from the GitHub Environment of the same name. The
Cordis plugin id stays dsh-plugin-browserskill; only the npm package name
changes.
Add the missing blank line before the actions comment, and hoist
.tool-button:focus-visible above the wrap-hover rules so specificity
no longer descends.
- @browser-skill/ui and @remixicon/react move to devDependencies (the client
bundle inlines both; the host face never imports them — verified).
- Drop private, add license: MIT, ship the repo-root MIT LICENSE in files.
- peerDependencies: add the runtime-required @deepseek-ai/schemastery, and
the previously dangling meta entries @deepseek-ai/dsh-attachment and
@deepseek-ai/dsh-llm (optional, matching actual type-level use).
- prepack runs the full build; npm pack now yields exactly 7 files
(lib/index.mjs, lib/index.d.mts, lib/client.cjs, cordis.patch.yml,
README.md, LICENSE, package.json).
- publint clean: the client face ships as lib/client.cjs (CJS extension in
a type:module package); add lint:publint script. Built declarations pass a
NodeNext consumer typecheck (strict, skipLibCheck false).
Cordis contexts are fiber-owned: assigning onto an Object.create() overlay of
the apply-time context fails with 'cannot set property tools in multiple
fibers' when the reveal fires from the tools/result listener fiber (caught by
the e2e: the skill invocation matched but the suite never registered).
registerTools now tracks disposers through a plain closure over the captured
context — no context mutation anywhere. Also documents lazyTools in the
README and hardens the reveal against a failed registration (log and stay
hidden instead of latching).
Verified end to end against the real stack: first request carries 0 browser_*
schemas (skill catalog entry only); after the scripted skill invocation the
next request carries exactly the 11 browser_* tools and browser_session_start
dispatches for real.
Progressive disclosure, final stage: with lazyTools on (default), the eleven
browser_* tool schemas stay out of the system prompt — the browser-skill
catalog entry is the sole advertisement. One successful skill invocation
(model tool call via tools/result, or a /browser-skill user gesture via the
session append feed) registers the suite for the process lifetime, idempotent
on repeats; session resume is covered by scanning durable events for a past
tool/call+tool/result pair (or skill-invocation message) on session entry and
at apply time. lazyTools: false keeps the legacy always-on registration.
Investigation (recorded in the ticket): tools/result carries normalized
name/arguments with an isError discriminant; dsh 0.1 has no official global
visibility switch (ctx.tools.restrict is agent-scoped and throws from a plain
context), making conditional registration the intended pattern; per-step
assembly picks the new suite up on the next step, and visibility changes are
first-class inputs of tool-skill's catalog digest. registerTools now returns
the combined disposer via a prototype-preserving tracking overlay; the skill
prelude tells the model the browser_* tools unlock on skill load.
Tests: +8 (hidden-before/reveal/idempotent/gesture/history hit+miss/dispose/
scanner pairing/two-state apply wiring). 124/124 green.
The plugin exposed only tools; the BrowserSkill agent skill never reached
the harness's <available_skills> catalog. Register it as an embedded runtime
skill (progressive disclosure: catalog entry resident, body on skill-tool
invocation) via the official ctx.skills seam — silent no-op when the
composition lacks it.
Content is build-time assembled so there is one source of truth:
skill/prelude.md (committed, dsh-specific: tool<->CLI map, owned-session
semantics, plugin-only overrides) + the canonical repo-root skill/SKILL.md
verbatim (the same file crates/bsk-cli/build.rs mirrors; frontmatter
stripped, name/description supplied by the registration). Generated into
src/skill-content.generated.ts (gitignored) — registration and every
pre-step snapshot are pure in-memory reads: no disk, no process, no daemon.
Scripts: build:skill chained into build/test/typecheck (same pattern as
build:css); biome excludes the generated artifacts. Tests: +3 (catalog
content, in-memory weight, silent degradation). 116/116 green.
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.
The client entry imports the generated bsk-ui.nomodule.css; the file is
gitignored, so a fresh CI checkout could not resolve it. Generate it as part
of the test script.
The previous pass aligned the observation overlay with the dsh shell; the
product call is the opposite direction — the floating card should read as a
BrowserSkill surface, not a shell-native widget.
- Reuse @browser-skill/ui directly (Button, cn); status dots spec'd after the
extension popup's ConnectionStatusIndicator; Remix icons for
interrupt/pin/warn/pop-out.
- Compile the BSK tailwind utility sheet scoped under the .bsk-obs root class
(scripts/build-client-css.mjs + postcss-prefix-selector) and ship the oklch
design tokens on the same scope, so nothing leaks into the host shell and
the shell theme cannot bleed back. tsdown injects .nomodule.css verbatim
(minified, unhashed) alongside the hashed CSS modules.
- BSK ships no shared tooltip, so the one-time interrupt semantics hint is a
card-spec bubble (bg-card/border/12px radius), retired after first use.
- Tests: pin react/react-dom to this package's 18.x copies in vitest — the ui
package sources peer on react ^19 and would otherwise emit react-19
(transitional) elements that the react-18 renderer rejects. Production is
unaffected (react stays external, provided by the shell).
- stylelint: ignore the generated bsk-ui.nomodule.css.
The plain top:16 default overlapped the shell's Session log action;
dock the card and capsule at the content area's top-right (64px) so
neither collides with the header controls.
Review follow-ups: the card/capsule now default to the top-right (clear
of the composer), and the overlay reads as part of the shell instead of
a foreign widget:
- shared primitives everywhere: Button (outline/ghost sm) for
Interrupt/Pop out with the shell's danger hover token, Tooltip for
the one-time interrupt semantics hint (native span anchor — Button
does not forward refs), StateDot for status, and outline icons
(Stop/RightUp/ChevronDown/Warning) replacing text glyphs;
- every color/surface/shadow/label now uses --dsw-alias-* semantic
tokens (bg-overlay, border-l1/l2, state-*-primary, label-*,
interactive-bg-hover[-danger], brand-primary focus rings,
dsw-shadow-lv1/lv2) — no invented tokens, no literal colors (dsh 0.1
ships no spacing/radius scale; those stay px like dsh's own code);
- strip item focus ring uses brand-primary; focus-visible and
reduced-motion preserved.
A child killed via killFor (overlay interrupt) exits with a null code
and empty output, which used to render as a duplicated 'bsk x failed:
bsk x failed'. Report it as interrupted instead.
The session-authorized client RPC (readAttachment) refuses images that
no session log references — observation frames are plugin-owned runtime
data, so every overlay thumbnail failed with ATTACHMENT_NOT_REFERENCED.
The host now keeps the full attachment ref per frame and serves bytes
through GET /bsk-observation/thumbnail/<id> (verified readImage path);
the client loader is a plain fetch of that route.
The daemon accepts only one unfinished command per session; the
thumbnail loop racing a model tool call produced 'session already has
an unfinished command' failures on both sides (found by the Phase 1-C
e2e: navigate/snapshot/click all errored while captures kept running).
Add a KeyedExecutor: every command for a session — tool calls and
observation captures alike — runs FIFO; queued tasks reject early on
abort, running tasks keep the runner's signal-driven kill.
ctx.get('webServer') at apply time raced the web composition's service
registration (the fallback SPA then answered /bsk-observation/* with
index.html). Ride ctx.inject(['webServer'], …) instead: the callback
runs when the service is provided and never runs in headless
compositions, keeping route registration timing-safe and optional.
- Strip (meeting-style multi-session layout): one item per session
(mini frame + id + status dot), horizontal row with scroll overflow;
hover reveals a per-item interrupt button that acts without
refocusing; click pins the focus view (pin badge, click again to
release); auto-follow picks the most recently active session but
never steals focus for errored or dead sessions (red edge / greyed
item instead).
- Edge states: the host flips an availability flag after repeated
global capture failures (client shows 'browser unavailable', keeps
the last frame, greys interrupt, and recovers on the next success);
a session_not_found envelope marks the session dead (no more frame
requests, grey strip item, removable as usual); lastError now clears
when the next action starts or succeeds.
- State/SSE wire gains { available } and an 'availability' event;
SessionObservation gains dead?.
Tests (+8, 97 total): host availability flip, dead marking +
instrumentation drop + clean removal, lastError lifecycle; client
strip render/pin/unpin/auto-follow, error no-steal, unavailable strip,
dead grey, strip-item interrupt without refocus.
Client half of the PiP observation window, wired to the host
ObservationService over the /bsk-observation HTTP+SSE seam:
- ObservationClientStore: initial state fetch + SSE increments +
on-demand thumbnail blob loading (session-authorized readAttachment,
same path as the toolview) + interrupt POST. All I/O injectable.
- ObservationOverlay registered into the shell.overlay list seat (the
sanctioned frame-wide surface; root is off-limits). Hidden with no
owned sessions; appears on first start; vanishes when all stop.
- Focus view: status row (green/grey/red dot + session + action + mm:ss
ticking), breathing thumbnail (new frames fade in, failures keep the
last frame with a warning badge, pre-navigate placeholder), and the
action area.
- Interrupt: single click, no confirm; greys into 'Interrupting…' and
back; disabled when nothing is in flight; one-time tooltip explains
the semantics (stops the current action only; the run continues).
- Card: drag-move via the header, corner-handle resize (min 240x180,
max 80% viewport), both clamped and kept for the page lifetime;
collapsible status capsule.
- PiP: Pop out (user gesture) opens Document PiP with the card's
current size, portals the same content in, and clones the document's
style nodes; pagehide falls back to the card with state intact;
unsupported browsers hide the button; nothing auto-pops.
Tests (+17, 89 total): store (fetch/SSE apply/malformed frames/
thumbnail lifecycle/interrupt wire/stop cleanup) and overlay (hidden->
visible lifecycle, status row, thumbnail via loader, interrupt
disabled/active/hint-once, capsule, resize clamps, move clamps, PiP
unsupported, PiP pop/fallback). Styling follows web-styling (dsw-alias
tokens, CSS Modules, focus-visible, reduced-motion).
Per-session live observation for owned sessions only, feeding the
upcoming client overlay:
- State model: SessionObservation (sessionId/url/action/since/
thumbnailAttachmentId/lastError), one entry per owned session;
added on start, removed on stop, cleared on dispose.
- Instrumentation: the shared runBsk wrapper begins/ends an action
(tool name mapped to a verb) around every model-facing call and tags
the child with its session id; action end triggers an immediate
thumbnail refresh.
- Thumbnail loop: 1.5s cadence while active, 8s idle downclock, x3
failure backoff, silent frame retention on errors; captures run
OUTSIDE the instrumentation (no action events, no current-pointer
movement) and land in the attachment store by reference.
- Interrupt: runner.killFor(tag) kills exactly the tagged in-flight bsk
children — chat-Stop-equivalent semantics for the current or a named
owned session; foreign ids and the no-in-flight case return false.
- Remote seam: dsh 0.1's Typert Remote pipeline and forwarded-event
allowlist are closed to out-of-tree packages, so state/events/
interrupt are served over the documented webServer route seam:
GET /bsk-observation/state, GET /bsk-observation/events (SSE),
POST /bsk-observation/interrupt; routes mount only when a webServer
service exists (headless compositions skip them).
- Config: observationEnabled (true), thumbnailIntervalMs (1500),
idleIntervalMs (8000).
Tests (+13, 72 total): state machine + event stream, cadence fast/idle/
backoff, headless no-store silence, observation-traffic isolation,
interrupt routing (default/specified/none/foreign), HTTP route
handlers, disabled-webServer no-op.
Tighten the shared-daemon boundary from 'cleanup only touches owned
sessions' to 'the plugin only ever sees and operates on sessions it
created':
- browser_session_list no longer queries the daemon at all; it returns
the plugin's own session table (with the current marker), so foreign
sessions cannot even be enumerated through the plugin.
- Every tool's optional session argument must name a plugin-created
session; foreign/unknown ids are rejected with a clear error before
any command reaches the daemon (the reference-adoption channel is
gone).
- Unchanged: dispose stops exactly the owned set, stop requires
ownership, the start reservation protocol, and handle-precise child
kills.
Tests updated (59): explicit-foreign rejection without daemon contact,
registry-only listing, reservation race, stale-handle-tolerant dispose.
e2e: a manually started foreign session was rejected on reference,
absent from session_list, refused on stop, and survived dispose while
both plugin sessions were cleaned up.
Review hardening for daemon-sharing deployments (other agents,
terminals, or dsh instances on the same bsk daemon):
- Ownership split: only sessions created by browser_session_start are
'owned'. Explicitly referenced foreign sessions are still tracked for
the current-session pointer, but browser_session_stop refuses them
and unload cleanup stops exactly the owned set — no path can stop a
session this plugin did not create.
- Start race: the capacity check is now a synchronous
reserveStart/completeStart/abandonStart protocol, so two concurrent
starts can never both pass the cap (check-and-reserve is atomic on
the event loop); a rejected or failed start never leaks a session.
- Stale handles: dispose already swallowed per-stop errors; the owned
set is computed at dispose time so an externally stopped session is a
no-op, not a failure.
Tests (+7): reservation race (two concurrent starts, cap 1 -> exactly
one spawn), foreign-stop refusal (no stop command reaches the daemon),
reference adoption never becoming owned, resolveForStop pointer
stability, and an apply-level dispose test proving only owned sessions
are stopped while stale stops are tolerated. e2e: a manually started
foreign session survived the plugin's full lifecycle while both plugin
sessions were cleaned up.
Make the package dual-face: a dsh.client declaration (platform 'web')
plus an exports["./client"] bundle built in dsh's closure-factory shape
(window.__ModuleLoader__.load handoff, platform modules external,
everything else inlined, CSS Modules via lightningcss).
The client registers a keyed 'tool.call.toolview' view for
browser_screenshot that keeps the terminal block (command + output) and,
when the settled result carries an image block, resolves the durable
attachment through the client session's authorized readAttachment RPC
and renders it with the shared MessageImage thumbnail/lightbox atoms.
Path-only results render unchanged; every other browser_* tool keeps
the stock terminal card.
Tests: view-model derivation (running/ok/error, image vs path-only),
component rendering through @testing-library/react + happy-dom, and the
registration key. biome.json now excludes the package's lib/ build
output (same treatment as dist/).
browser_screenshot on an image-capable route returns text + image
blocks; the shared result presenter required exactly one text block and
bailed to the generic card (tool name + raw args). Project the first
text block instead so the terminal card shows in both modes.
The daemon's JSON error envelope carries an actionable hint (e.g.
'choose an input, textarea, or contenteditable element from the latest
snapshot'), but the thrown BskError only embedded the message field, so
the model-facing error text never showed the hint. Append it.
The concurrency cap was checked in registry.add() AFTER
`bsk session start` had already created the session, so a rejected
start leaked a live session outside the plugin's tracking (found by the
real-browser e2e: the rejected third session stayed in
`bsk session list` after dispose). Check capacity before spawning;
registry.add() keeps the same check as a backstop.
Add a new monorepo package that registers BrowserSkill (bsk) browser
automation as model-visible tools in DeepSeek Harness (dsh):
- 11 tools mapping to bsk CLI --json commands: session start/stop/list,
navigate, snapshot, observe, click, fill, press, screenshot, emulate
- multi-session support: optional session arg with a current-session
pointer, session id echoed in every result, configurable concurrency
cap (default 5), and dispose-time cleanup of plugin-started sessions
- cancellation: exec.signal aborts kill the underlying bsk child process
- terminal-style UI cards (command line as the call card, output as the
result card); screenshots inline the PNG through the host attachment
store when the model route accepts image input, else return the path
- config: bskPath / defaultTimeoutMs / maxSessions via Schemastery schema
- install guidance when the bsk binary is missing; bsk JSON error
envelopes (code/message/hint) surfaced to the model
- unit tests mock the bsk runner (no real browser required); the root
lint script now also runs the package's typecheck + tests so the
existing CI frontend job covers the new package (a dedicated CI job
diff is provided in the PR; the pushing token lacks workflow scope)
Add a per-browser "Control hints" switch to the popup that hides the in-page control overlay (status pill + orange glow, along with its input blocker) while the Agent controls a tab. Preference lives in chrome.storage.local (bsk_control_hints_hidden, default: shown) and is applied live in content scripts via storage.onChanged.
Both popup settings rows share a single Switch component; the hint copy lives in an info-icon tooltip.
Refs #62
Add a popup launcher + action-recording panel that builds a copyable
`bsk record start` prompt, add `--browser` targeting, and default `--url`
to https://example.com/ when omitted. Drop the popup browser-alias field.
Capture user actions in the Agent Window and export a record-only
trace (pages[] + navigate/click/fill/select/press steps with target
descriptors). Adds the record overlay UI aligned with the control
overlay, and wires record start/stop/await through CLI, daemon,
protocol, and extension.
BB-fat