* add remote browser support via Browser Use cloud + multi-daemon
HARNESLESS_NAME suffixes socket/pid/log — daemons are independent, no
supervisor. start_remote_daemon() creates a Browser Use cloud browser and
launches a daemon attached to it; kill_daemon() stops both. Local Chrome
path is unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* rename env vars to BU_ prefix (shorter, less noisy in tool calls)
HARNESLESS_NAME → BU_NAME
HARNESLESS_CDP_WS → BU_CDP_WS
HARNESLESS_REMOTE_BROWSER_ID → BU_BROWSER_ID
Socket/pid/log files keep the harnesless- prefix on disk so they're
recognizable in /tmp. BROWSER_USE_API_KEY unchanged (external convention).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* rename project from harnesless to bu
Socket/pid/log paths now /tmp/bu-<name>.{sock,pid,log}. pyproject package
name updated, uv.lock regenerated. Slash command now /bu.
Note: the repo directory itself is still named harnesless on disk. Rename
manually (mv harnesless bu) so the absolute paths in docs line up.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Previous commit added fill_form / mui_select_first / is_success which are
eval-harness logic, not harnesless primitives (task-specific defaults like
"dumbledore"/"Harry Potter", MUI-only shim). Per AGENTS.md — "could the LLM
rewrite this from scratch after reading it once" — the LLM should pick field
values + the submit strategy per-task, not inherit eval defaults.
Removed: fill_form, _FILL_JS, mui_select_first, is_success (~100 lines)
Kept: dispatch_key, upload_file, capture_dialogs/dialogs (universal needs)
Insights from the eval captured as gotchas in SKILL.md instead:
- React controlled inputs need native-setter + input event
- Radios/checkboxes: el.click() over el.checked=true for React
- MUI / UI-library overlays: real CDP click, not JS .click()
- CDP char event ≠ DOM keypress for special keys → dispatch_key
- Same-origin iframes: contentDocument walk, not CDP targets
- Shadow DOM: querySelector doesn't pierce, walk .shadowRoot
- Form success signals vary: element / alert / body text
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Connection management (addresses "shit ton of daemons" problem):
- Socket is the lock. Daemon refuses to start if another is already
listening (5-line check).
- PID file at /tmp/harnesless.pid written on start, removed on exit.
- ensure_daemon() / kill_daemon() / daemon_alive() helpers.
- run.py auto-calls ensure_daemon() before exec — users never manage
the daemon manually.
- kill_daemon uses PID file (robust) instead of pkill pattern matching
(was silently missing because the process command line didn't contain
the "harnesless/" prefix).
Post-task ritual added to SKILL.md: after every browser task, extract
ONE generalizable friction point and make the simplest possible
improvement (2-line helper, one-line gotcha, recipe correction).
This is how the harness sharpens itself over time.
Verified end-to-end: kill_daemon → 0 processes + files cleaned → next
run.py → auto-starts exactly one daemon. Second `uv run daemon.py`
exits with "daemon already running" message instead of racing.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Observed friction during an Azure portal task:
- Daemon's default session went stale (user closed the attached tab),
which broke every subsequent call including browser-level Target.*.
- new_tab() had been removed in the previous simplification pass but
was needed to recover.
- Azure portal renders blade panels in iframes; js() on the main page
returned nothing for picker contents.
Changes:
- daemon: browser-level Target.* calls now bypass self.session entirely
(so a stale session doesn't poison them). On "Session with given id
not found" for session-scoped calls, clear + re-attach + retry once.
Merged start() and the new recovery path into attach_first_page().
- helpers: add new_tab(); js() accepts target_id for iframe queries;
iframe_target(substr) to find blade iframes; ensure_real_tab() now
resilient to stale-session exceptions.
- SKILL.md: one-line note on iframe-site workflow.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Session of 7 real tasks (Upwork, X, Google Flights, HN, Netflix, iPhone
comparison) used only half the original helper surface — every DOM
interaction went through js() + a bespoke selector, never through the
indexed-DOM helpers. Dropped 14 that never fired: get_dom,
click_element, type_in, element_pos, save_cookies, load_cookies,
set_viewport, screenshot_full (folded into screenshot full=True),
double_click, right_click, move_mouse, new_tab, close_tab,
handle_dialog, back, reload.
Also:
- daemon: shorter identifiers, collapsed boilerplate, TOCTOU fix via
try/except on read_text, is_real_page module-level, no variable
shadowing of `url`.
- helpers: one shared INTERNAL_URL_SCHEMES tuple instead of three
subtly-different ones; current_tab() collapsed from 4 round-trips
with dead fallback to one; wait_for_load() stops draining the shared
event buffer.
- SKILL.md / AGENTS.md: cut recipes and anything the LLM already knows
(what click does, what CDP is). Kept only project-specific context.
Net: 8600 -> 4030 tokens across the project (53% reduction); the file
the LLM reads every skill invocation (helpers.py) dropped from 2193
to 1219 tokens.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three-process architecture: daemon.py holds one persistent CDP WebSocket
to the user's running Chrome (via chrome://inspect), short-lived run.py
processes talk to it over a Unix socket, helpers.py is the transparent
layer the LLM reads and edits at will.
Philosophy: no CLI, no fixed API surface. The LLM writes Python blocks
against ~13 tiny helpers (cdp, click, type_text, screenshot, get_dom,
etc.) and edits helpers.py on the fly when a pattern repeats. Coordinate
clicks default because they pass through iframes/shadow DOM/cross-origin
at the compositor level.
Uses cdp-use internally for send_raw only (ignores its 36k lines of
typed wrappers — raw CDP strings tokenize better than typed calls).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Gregor Žunič