Restore the lean style from PR #167 while integrating essential file
transfer documentation for upload/download. Reduces the skill from
362 lines / 26k chars to 180 lines / 9.2k chars.
A/B eval (24 browser runs, Composer 2.5 Fast, core suite) confirms
no pass-rate regression: candidate 12/12 vs baseline 10/12 (baseline
failures were a network blip and a behavioral variance, not skill
quality).
Co-authored-by: Cursor <cursoragent@cursor.com>
Keep feat/expand-upload-download (drop mode superset) over the base
branch version for all 12 conflicting files. The drop-to-upload branch
already includes all upload/download functionality plus the new --mode
drop mechanism, overlay suppression, OOPIF coordinate handling, and
expanded error reasons.
Co-authored-by: Cursor <cursoragent@cursor.com>
Keep the expanded command-reference tables from feat/upload-download
(including the new File Transfer section) over main's compact two-line
format.
Co-authored-by: Cursor <cursoragent@cursor.com>
main rewrote SKILL.md into a leaner core (#167) while this branch documented
the opt-in --probe-hover flag against the old structure. Keep main's rewritten
prose and re-apply the probe-hover guidance on top of it: the [has-submenu] /
[expanded] markers, when reaching for --probe-hover is warranted, and its slot
in the page-reading escalation ladder ahead of get-html/screenshot.
Co-authored-by: Cursor <cursoragent@cursor.com>
The previous wording asked the agent to notice when a page "looks like"
it hides content behind hover. That is not something an agent reading a
text tree can judge: hover-only content is invisible by definition, and
the marker that used to point at it came from the probing that is now off
by default.
Replaces it with a condition the agent can actually evaluate — an
expected control is missing and no [has-submenu] marker points at a
trigger — and places it on the escalation ladder ahead of get-html and
screenshot, which is where that dead end is currently reached.
Also documents [has-submenu]/[expanded] as the primary hover signal. It
comes from aria-haspopup in the accessibility tree, costs nothing, and is
unaffected by probing being off, so it covers menus that are marked up
correctly.
Every tool.observe moved the real cursor across the page to look for
hover-only menus and tooltips. Measured against real Chrome, that cost a
median of 2.5s and up to 4.7s per observation — 73% to 99% of total
observe wall clock — and returned 3 hits across 12 pages. A positive
control confirms the probes work; the yield on real pages simply does not
justify putting them on the default path.
observe now probes only when asked, via `probe_hover` (`--probe-hover`).
Agents that know which element hides a menu should keep using
`bsk hover <ref>`, which is both cheaper and precise. A held hover latch
still suppresses probing, since probing would move the cursor off the
element the caller is deliberately holding.
Correctness fixes that apply whether or not probing is requested:
- Hover probing ran between the DOM snapshot and the accessibility tree
fetch, so a hover that opened a menu left one half of an observation
describing the page before the change and the other half after it. Both
chains now run after all capture is complete.
- The overlay bypass toggled once per chain and swallowed restore
failures, which could leave the agent overlay hidden for the rest of
the tab's life. It is now reference counted across both chains, and a
failed restore is reported instead of discarded.
- Observations report whether they hovered the page and whether that
revealed content the static tree lacks, so callers can judge how far
the live page may have drifted from the returned snapshot.
Probe budgets were checked at the top of the loop, so a candidate could
start with 1ms remaining and still run a full settle window, overshooting
by up to 634ms. Candidates are now admitted only if they fit. Raising
MAX_HOVER_PROBE_MS to 2600 is part of that fix rather than a relaxation:
2000 was never the real ceiling, and enforcing it literally would have
silently dropped a candidate the positive control depends on.
Tooltip misses are cached too, with a shorter TTL than hits, so repeat
observations of the same page no longer re-probe every unlabelled control
from scratch.
The CLI skill had grown to 20.8 KB of command tables and flag catalogs that
`bsk --help` already answers at runtime, so every agent paid for them on every
task. Rewrite both skills around the durable rules an agent cannot rediscover
-- session lifecycle, the observe/act/observe loop, the Agent Window boundary,
and the stop-on-success contract -- and delegate flag details to the CLI's own
help output. The CLI skill drops to 6.9 KB (-67%) and the DSH skill to 5.1 KB.
Keep the parts that help discovery has no answer for: the literal
`[hover first: ...]` observation marker, the `request-help` outcome values, the
complete command-name inventory, and the required flags agents guess wrong
(`fill --value`, `select --value`, `screenshot --out`, `emulate --device`).
Also bind `--no-focus` explicitly to `bsk session start`, since a bare mention
led agents to pass it to `bsk navigate`.
Validated with 36 agent runs of the `core` eval suite against a real browser,
comparing the old and new instructions with everything else held constant.
Pass rate did not regress (the lean version was better on `form-controls`) and
token cost per task came out even. The inventory and flag signatures cut
`--help` round-trips from 1.7 to 0.7 per task and removed the invented
`bsk back` / `bsk forward` calls the first draft produced.
Tighten the DSH size assertion to a range so the budget is bounded from both
sides and accidental truncation fails the test.
Co-authored-by: Cursor <cursoragent@cursor.com>
The exit code reflects RPC success only, not JS success. Document that programmatic callers must use --json and check the .ok field to detect JS errors. Addresses the discoverability gap raised in issue #133.
Agents were stopping on popup and status copy that looked like a hard incompatibility. Keep the connection usable, reframe exit 5 as an unknown method, and ask users to upgrade instead of aborting.
Co-authored-by: Cursor <cursoragent@cursor.com>
Reject legacy JSON --output before recording starts, persist the completed Trace before writing the bundle so `bsk record stop --output <dir>` can retry after an export failure, and tell users to pass --url when https://example.com/ does not load.
RecordedTrace validation now follows classify_value — numeric version
selects v3, otherwise v2 — so mixed pages/states and v2 envelopes with
a version field fail Schema checks before Recall import. Keep standalone
trace_step as a v2|v3 union for tools that validate a single step.
Split standalone v2/v3 schemas and drop TraceV2 deny_unknown_fields so legacy traces with extra keys still parse; mix and version checks stay in RecordedTrace.
Require Trace v3 version const 3, reject mixed pages/states payloads, and
align TypeScript selection optionality with the generated schema. Trace v2
on main already records hover steps, so the frozen v2 schema must accept
them instead of dropping them during reduction.
Co-authored-by: Cursor <cursoragent@cursor.com>
Define Trace v3 wire types and generated schemas while keeping protocol
1.1 peers compatible with 1.0. Default record start still produces Trace
v2, and only an explicit trace_version=3 request enables the new shape.
Refs #90
Co-authored-by: Cursor <cursoragent@cursor.com>
After folding the IPC connection count and activity timestamp into a
single mutex-protected IpcActivityState, the idle task no longer
references the connection count directly; should_exit_for_idle already
returns false while a connection is open or within the idle window.
Update the comment to describe the post-refactor control flow so it
matches the code.
The extension is now published to Edge Add-ons, but every install path
still pointed Edge users at the Chrome Web Store. Both READMEs and
AGENT_INSTALL.md now name a store per browser, and `bsk doctor` offers
both listings when no browser is connected.
Two Edge-specific gaps went with it: Edge blocks content-script injection
on its own storefront the same way Chrome does, so the borrow-confirmation
overlay must treat microsoftedge.microsoft.com/addons as non-injectable;
and the Chinese README still described the pre-install.ps1 manual Windows
setup.
Also give the DeepSeek Harness section its npm package link and the
profile-start step, so plugin users have a complete path.
Co-authored-by: Cursor <cursoragent@cursor.com>
Tool cancellation is now cooperative end-to-end: the queue stays busy until compensation completes, cancellation is checked cooperatively across the pipeline, the fallback window is compensated too, and cleanup_failed is surfaced explicitly. Includes a gated delayed-compensation E2E test.
chrome.tabs.captureVisibleTab reads back the window surface, which fails
outright on some Windows/Chrome combinations (Chromium
FAILURE_REASON_READBACK_FAILED: "Failed to capture tab: image readback
failed"), leaving full-tab screenshots without any working path.
Capture the full tab via CDP Page.captureScreenshot (fromSurface) when
the visible-tab capture rejects; the CDP path goes through the
renderer BeginFrame pipeline instead of the surface readback. When both
paths fail, return cdp_failed with data.reason=screenshot_capture_failed
carrying both underlying messages, and render a CLI hint that points at
the browser-side readback cause.
Refs #71
The per-user pipe name embedded %USERNAME% verbatim
(\\.\pipe\bsk-daemon-{user}-{hash}). When the username contains
apostrophes (e.g. z'z'f'l'g'y), NPFS behaviour diverges between create
and open: the daemon's CreateNamedPipeW reports success and logs that it
is listening, yet every client CreateFileW on the same name fails with
ERROR_FILE_NOT_FOUND / ERROR_PATH_NOT_FOUND, so the CLI can never reach
the daemon (issue #75).
The name is now hash-only (\\.\pipe\bsk-daemon-{hash}); the hash
already covers user + home, so uniqueness and per-user/BSK_HOME
isolation are unchanged, and no raw username characters can reach NPFS.
The pipe name is exchanged exclusively through daemon.json (the CLI
never computes it), so old/new CLI/daemon combinations stay compatible.
Unit tests pin the safe character set for apostrophe/space/Unicode
usernames and the per-(user, home) uniqueness.
Fixes#75
`bsk session start --no-focus` opens the Agent Window in the background
without stealing focus. The flag threads focused=false through
CLI -> daemon IPC -> tool.session_start -> chrome.windows.create; an
omitted `focused` keeps the extension default (true) so older clients
are unaffected.
Reworked onto current main so it composes with #52 window sizing:
AgentWindowApi.create / SessionManager.start (extension), start_session
(CLI) and the daemon's session start now take options objects/structs
({ size, focused }) instead of colliding positional params.
Co-authored-by: alectimison-maker <alec.timison@gmail.com>
The daemon's periodic update check now installs a newer release instead
of only refreshing the hint cache: it reuses the update.rs install path
(download, sha256 verify, extract, replace), skips the round while any
agent session is live, and otherwise restarts itself into the new binary
(spawns a detached replacement that waits for the old pid, then exits).
On by default; BSK_AUTO_UPDATE=off keeps the old cache/hint-only
behaviour. Windows can only stage the replacement, so it just logs that
a restart is needed. The CLI hint wording reflects the switch.
