A run died with "'charmap' codec can't encode character '▶'" at step 1.
Python picks the console code page for stdout, and under PM2 - where stdout is
a pipe rather than a console - that is cp1252 on Windows. cp1252 handles the em
dashes in these files but not the run markers, so the first one raised
UnicodeEncodeError from inside run_steps and the runner reported it as a step
failure rather than an output problem.
Reconfigure stdout and stderr to UTF-8 at the top of each entry point, with
errors="replace" as a backstop for streams that cannot be reconfigured. Fixing
the encoding once beats stripping the glyphs from ~30 call sites, and covers
manual runs in a plain console too.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
scrollToLoad walks a progressively-loading list to the bottom before the steps
that act on its items run. Stopping is two-part: no new matches appeared AND the
container was already pinned to the bottom — counting alone stops early on a slow
fetch. Hitting the scroll cap is reported rather than passed off as done, so a
later step never works quietly on a partial list.
The scrolling element is usually not the window. Lists like this live in a div
with its own overflow, and scrolling the document does nothing at all, so the
step walks up from a matched item to the ancestor that actually scrolls —
overflow allows it and there is more content than fits — with containerSelector
to name one outright when the guess is wrong. Verified against a page whose
document also scrolls, which is the case that tells the two apart: it found the
inner div and pulled 12 items up to 60 in 7 scrolls.
waitFor gains `absent`, for waiting on something to go rather than arrive — a
modal closing after a reset. It only accepts a genuine "selector matched
nothing"; an unreachable extension looks the same from a distance and would
otherwise satisfy the gate for the wrong reason, sending the next iteration into
a page that still has the modal open.
The locate queue carries a free-form options blob now, so a new kind of request
stops meaning a new column each time.
Also fixes a missing comma in the reset flow that broke the build.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
focus.py raised the first browser in its list that happened to be running. On a
machine with both Chrome and Edge installed that is a coin flip, and losing it is
silent: coordinates measured from a tab in one browser, the click delivered into
a window of the other. It presents as selectors failing for no reason. A VPS with
both installed hit exactly this.
The extension now reports which browser is hosting it, and that travels with the
measurement, so the clicker raises the browser the coordinates actually came
from. Asking for a browser that is not running now fails honestly instead of
quietly raising a different one, and the verification step rejects the wrong
browser coming forward. Chromium, Opera and Vivaldi are recognised alongside
Chrome, Edge and Brave, on both platforms.
diagnose.py answers the question a remote desktop makes hard: whether the mouse
is really moving or the viewer simply is not drawing it. It moves the cursor and
reads the position back from the OS, so the answer does not depend on anything
being rendered, and it reports DPI mode, screen size, whether this is an RDP
session, and whether the browser can be raised at all. Nothing is clicked.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Next's dev server intermittently answers a 500 while recompiling a route: it
reads a build manifest mid-write and cannot parse it. A single one of those
during the locate poll was fatal, so a blip in the pipeline killed a run partway
through an auth flow on Windows.
5xx responses and dropped connections are now a distinct TransientError, retried
until the step's own timeout. A 4xx still fails immediately — those are verdicts
about the request, not blips. If the errors persist all the way to the timeout,
the message says so rather than blaming a missing extension.
Error bodies are also summarised. A dev-server 500 replies with a full HTML page,
and printing it raw buried the one line that said what went wrong under kilobytes
of script tags.
This makes the client tolerant of the fault, which is not the same as fixing it:
the real answer on an automation host is to run a production build rather than
`next dev`, so those manifests are written once instead of continuously.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The DPI awareness call added for Windows went into both entry points, but the
import only went into clicker.py, so starting the runner died immediately with
NameError: name 'focus' is not defined.
py_compile does not catch this — a missing import is a runtime error, not a
syntax one — and the tests around it stub the modules rather than starting the
process, so nothing exercised the real startup path. Verified this time by
booting the runner against a dead port, which reaches the claim loop.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
focus.py now verifies on Windows rather than assuming activation worked. Windows
declines to raise a window for a process that doesn't own the foreground — it
flashes the taskbar and the call returns as if it succeeded — so the runner reads
the foreground window's process back and reports a failure instead of clicking
into a background window. Same gap that was fixed on macOS earlier.
The runner also declares itself DPI-aware at startup. Without it Windows reports
a virtualised screen size and rescales the coordinates it accepts, while the
browser keeps reporting CSS pixels; on a display at 125% or 150% the two disagree
and clicks drift further off the further they are from the top-left.
Browser windows are matched on the owning process rather than the window title, so
an editor with chrome.js open is no longer mistaken for the browser. Linux now says
window management is unsupported there, rather than reporting no browser found —
pygetwindow has no X11 backend, and "no browser window" reads like Chrome is shut.
The main README gained a section on the AutoBuyer: what the three pieces are, how
to load the extension, and that neither the extension nor the runner reloads
itself when the source changes. That last point has been the cause of most of the
confusing failures so far, so it is called out in Updating too.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Repeat. A `repeat` block runs its steps several times, with the count either
fixed in the config or taken from an input the user sets on the dashboard. The
block is unrolled in resolveSteps before the runner sees it, so the runner needs
no loop, the run's total step count stays honest, and every iteration appears in
the log as its own line — a failure on the third purchase reads as "(3/5)"
rather than as an indistinguishable repeat of the first.
Counts are clamped server-side against the automation's declared min/max, and
expansion is capped at 400 steps and three levels of nesting. Each iteration can
be a purchase, so the number is not taken on trust from the client, and the
confirmation dialog names it before anything runs.
skipIfNotFound on a click or type step tolerates an element that is not on the
page — a cookie banner, a modal that only sometimes appears. Only absence is
tolerated. That distinction needed a new NotFoundError: previously a missing
element, an unreachable dashboard, a missing tab and a covered button all
surfaced as the same DashboardError, and skipping that whole class would mean a
step quietly passing while the extension was down.
Orphaned runs are now reaped. Only one run executes at a time, so a run left in
'running' when its runner went away blocked every future run — restarting the
daemon mid-run deadlocked the queue, which is exactly what happened. The
heartbeat decides: a runner that is gone, or up and reporting idle, is not
driving that run whatever the status column says. Gated on the busy flag rather
than elapsed time alone, since a run sitting in a waitFor gate or a sign-in wait
can legitimately go minutes without progress.
Lucid Trading is scaffolded with no automations yet. One match pattern covers
both its hosts — `*.` matches the apex as well as subdomains, confirmed against
a live tab. Its signed-out pattern is `//lucidtrading.com/` rather than
`lucidtrading.com/dashboard`: the leading slashes anchor it to the start of the
host, and without them the substring also matches dash.lucidtrading.com, which
would abort every step while properly signed in.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Step vocabulary gains `waitFor`: block until a selector exists, then continue.
Nothing is clicked or typed — it is a gate for conditions something outside the
run has to satisfy. Unlike every other step it carries no signed-out guard,
because the things worth gating on often sit on the login page, where that guard
would abort the run at exactly the wrong moment. Honours the dashboard's Stop
button, since a two-minute gate that ignored it would be worse than no gate.
Session handling. A run that lands on the login page must not continue: once
redirected, every selector resolves against a login form, so a click aimed at
"Add Account" hits whatever that form renders in the same place. Runs now detect
the redirect and stop before sending any input, with a distinct SignedOutError
rather than a generic failure.
Two ways out of that state, in order: a firm's `authSteps` run and the failed
step is retried, or — when none are defined — the run pauses for
signedOutWaitSeconds so a human can sign in, then resumes. Auth steps are
verified rather than trusted: they can all "succeed" while the site still
rejects the sign-in, so the session is re-checked before the retry, and the run
stops with "auth steps ran but the session is still signed out" if it did not
take.
That check polls for up to 20s instead of reading once. Submitting a login form
starts a network round trip and then a redirect, so the tab still shows the
login URL for a second or two afterwards; checking immediately failed a sign-in
that was merely in flight, killing run #15 nine seconds after it had actually
worked. Third instance of the same mistake in this system — reading page state
immediately after an action that triggers async navigation.
The runner reports its version in the heartbeat and the dashboard blocks the
buttons when it is behind. A running Python process does not reload when the
source changes, so a stale runner fails on step types it predates; that cost a
debugging round when a navigate step reached a runner that had never heard of
one.
lib/automations.ts carries the Tradeify buy-accounts flow: navigate to the
dashboard, open Add Account, pick the account type and size, enter the account
name, and work through the challenge widget before submitting. Selectors are
authored by hand against the live page.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Turns the autobuyer from a page scraper into something that acts. A dashboard
button queues a run; a desktop process executes it against the real browser.
lib/automations.ts — automations are declarative step lists nested inside the
firm whose site they drive. Steps are click / type / wait / navigate, and they
inherit the firm's tab pattern and URL, so one firm's automation can't act on
another's tab. Adding a button means adding an entry here; the page renders
buttons from the API and the runner receives steps from the server, so neither
needs editing. Runs key on firm:automation — every firm will plausibly have its
own "buy-accounts", and a bare id would resolve to the wrong one.
clicker/runner.py — the daemon behind the buttons. Claims a queued run, works
through the steps, reports each one back for the page's live log. Only one run
executes at a time: two processes driving one physical mouse would interleave
clicks. Heartbeats on its own thread, because a step can block for tens of
seconds and folding the beat into the main loop would show the runner as offline
in the middle of the run it was executing.
clicker/actions.py — one implementation of the safety checks, shared by the CLI
and the runner. Refuses to act when the element is covered by an overlay, when
coordinates fall off-screen, when the browser can't be confirmed frontmost, or
(for type) when the target isn't an editable field.
Typing: uneven human cadence, and the field is read back afterwards and compared
against what was typed — a field that never took focus fails silently and looks
identical to success otherwise. Non-ASCII is rejected because pyautogui skips
those characters without complaint, and newlines because Enter may submit the
form. Typos are deliberately not simulated: a mistyped digit in a trading form
is a real loss, and the correction is the part that can go wrong.
Extension: opens the firm's page when no tab matches, navigates to a specific
page for a navigate step (skipped when already there, so page state survives),
and retries the locate while a freshly loaded React app mounts — `complete` only
means the document loaded.
Staleness reporting, after it cost three debugging rounds: Chrome doesn't reload
an unpacked extension and Python doesn't reload a running process, so both now
report their version. A stale runner gets a red banner naming both versions and
the automation buttons are disabled, rather than failing mid-run on a step type
it predates.
Scale detection is now conservative: a raw OS/browser width ratio is only
trusted when it lands on a real scaling factor. On this multi-monitor desktop
the previous logic would have silently halved every coordinate.
Verified end to end against the live browser: navigate, locate, and a real
click (run #12, all three steps). API round-trips, claim-once semantics, run
cancellation, the heartbeat online/offline lifecycle, motion geometry and
timing, focus activation, and typing verification all pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Builds the pipeline the autobuyer needs: see the page, find an element,
click it.
extension/ — MV3 Chromium extension. Polls /api/autobuyer/status and,
while on, scrapes the target tab's HTML and posts it back. Also serves
locate requests: focuses the window, scrolls the element into view, and
reports its position. host_permissions is scoped to tradeify plus
localhost so it cannot read other sites — an empty target pattern would
otherwise capture whatever tab happened to be active, including banking
or mail.
app/api/autobuyer/ — status toggle, capture store, and the locate request
queue. CORS is open because the extension's origin changes every time an
unpacked extension is reloaded.
app/autobuyer/page.tsx — ON switch, source view (default) and a rendered
view. The render uses sandbox="allow-scripts" without allow-same-origin:
the page's own JS is needed because sites ship content at opacity:0 and
fade it in, but the frame must not reach the dashboard's same-origin API
routes, which serve firm credentials.
clicker/ — Python CLI. Asks the extension where a selector is, adds the
element rect to the window's screen position and the browser chrome
height to get desktop coordinates, then clicks with a human motion model
(curved path, eased velocity, occasional overshoot, dwell before press).
Raises the browser application first, since macOS consumes a click on an
unfocused window rather than delivering it.
Refuses to click when the element is covered by an overlay, when the
coordinates fall off-screen, or when the browser cannot be confirmed
frontmost.
Verified: API round-trips, capture pruning, locate claim-once semantics,
motion geometry and timing, and focus activation — the last two against
stubs, since pyautogui and pyobjc are not installed here. NOT verified
end to end: Chrome is still running a stale build of the extension, so a
locate request has never completed against a real page and no real click
has been sent.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
