Theme toggle sits next to the settings icon. The choice starts as 'system' and
follows the OS until you press it, after which it's explicit and persisted.
An inline script in <head> stamps the resolved theme on <html> before the
first paint - resolving it from JS after load means a visible flash of dark on
the way to light. Because that attribute is always present, the stylesheet
drops its prefers-color-scheme query entirely rather than having a media query
and an explicit override fighting over the same tokens.
The voice picker was effectively empty: Chrome reports zero voices
synchronously and only fills the list when voiceschanged fires, and while the
narrator did reload them, nothing told preact to re-render - so the dropdown
kept whatever existed at construction, which was nothing. It now notifies, and
the list arrives (181 voices here).
That many voices needs shape, so they're sorted with your own language first
and offline voices ahead of network ones, then grouped into optgroups by
language. Network voices are marked as such since they're useless offline,
which for an app built to work in a field matters.
Speed and pitch are adjustable too, both fed through to every utterance.
The settings button gains a class of its own: the header now has two icon
buttons, so identifying it by .icon-button alone hits the theme toggle.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QK7PKpRVoy69Fpa32R8abb
Home screen gains shutter, aperture and ISO as dropdowns plus a focus stepper
(three nudge sizes each way, and AF) where the body drives focus over PTP.
Those three drop out of the read-only strip rather than being shown twice.
Everything is built from what the camera actually reports, so a body without
one of these controls doesn't get it rather than showing something that
silently fails. The 450D marks shutter and aperture readonly unless the mode
dial is somewhere they apply, so the dropdowns disable themselves and say why.
Focus uses the EOS manualfocusdrive steps and autofocusdrive, and warns when
live view is off, which Canon bodies generally require for focus commands.
The settings drawer is now tabbed: app settings stay on the first tab, and
each config section the camera reports gets its own. Empty sections are
dropped, and a selected section that disappears falls back to the first tab.
Tabs wrap rather than scroll - a scrolled-off tab is an undiscoverable one.
Reverts the automatic camera search added in the previous commit, restoring
the Select camera button. Auto-connect could park with no way to reach the
device chooser: WebUSB permissions are per-origin, so a grant on localhost
doesn't carry to the deployed copy, and a camera that's asleep or held by
another app never arrives.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QK7PKpRVoy69Fpa32R8abb
Fullscreen: a button on the preview pane fullscreens the live view, carrying
the countdown over as an overlay - fullscreen hides the entire control column,
so without it you're left staring at a picture with no idea when the next
frame lands.
Wake lock: held from app start until you switch it off, rather than only for
the duration of a sequence. You're usually mid-setup when the display would
otherwise sleep.
PWA: manifest, icons and a service worker, so it installs to a standalone
window and runs with no network. That matters more here than for most web
apps - the camera is on a USB cable, so this is fully functional in a field
with no signal.
The worker precaches the pinned unpkg dependencies at install rather than
leaving them to the runtime cache. On a first visit the worker isn't
controlling the page yet, so the app's own imports go straight to the network
and never reach the fetch handler; without the precache it looked cached but
died offline on its imports. Our own files are network-first so a redeploy
always wins, and the version-pinned CDN files are cache-first.
Connecting: no connect button. The app reaches for the camera on load, retries
every 1.5s, and listens for USB connect events, so switching the camera on
mid-wait attaches it with no click or reload. The exception is a browser that
has never been granted access to the device: Chrome will not open its WebUSB
chooser outside a user gesture, so that case still shows a one-off prompt.
After that the permission is remembered and getDevices() finds the camera with
no interaction.
tsconfig excludes sw.js, which runs in ServiceWorkerGlobalScope and reports
every worker global as undefined when checked against the DOM lib.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QK7PKpRVoy69Fpa32R8abb
A 🔊 toggle on the home screen counts you into each frame - "five, four,
three, two, one" - plus start and finish announcements. Useful when you're in
front of the camera rather than at the laptop. Uses the Web Speech API, so
it's local and needs no configuration.
The narrator runs off the intervalometer's state updates rather than a timer
of its own, so it can't drift away from what the sequence is doing; each
second is spoken at most once even though state arrives ~5x/sec.
The count is capped at one second under the interval, so a 3s interval says
"two, one" instead of talking over the previous frame, and a 1s interval stays
silent. A start delay isn't clamped, since that gap is whatever you set.
Countdown length, voice choice and frame-number announcements live in the
settings drawer; only the toggle is on the home screen.
Also:
- The toggle sits in the status card rather than the button row: three buttons
don't fit a 400px column, and it wrapped onto its own flex line where it
stretched to the wrong height.
- tsconfig excludes dist/, which build-dist.sh fills with copies of these same
files and which otherwise gets type-checked twice.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QK7PKpRVoy69Fpa32R8abb
Replaces the gPhoto2 demo UI (live view beside the raw config tree) with a
timelapse intervalometer. The home screen carries only what changes between
runs - interval, frame count, start delay, start/stop, countdown, progress
and a log - while every camera setting moves into a settings drawer.
- intervalometer.js: schedules frames on an absolute grid (start + n *
interval) so transfer time doesn't accumulate as drift over a long run. An
overrun logs and fires as soon as the camera is free rather than dropping a
frame; three consecutive failures abort.
- storage.js: frames stream into a folder via the File System Access API,
named for capture time (20260801-172713_00001.JPG) so sorting by name is
sorting by time. Falls back to downloads. Also holds the screen wake lock,
since background tabs get their timers throttled.
- config-utils.js: config tree lookups, shutter speed parsing, and bulb
capability detection (bulb toggle or Canon eosremoterelease).
- home.js: sequence controls plus a read-only exposure readout and pre-flight
warnings when the interval can't fit the exposure and transfer.
- settings.js / index.js: app prefs and the full config tree behind a drawer.
Config polling now only runs while that drawer is open, leaving the USB link
to the captures during a sequence.
Live view stays up between frames and steps aside only while the shutter
fires, which is what the EOS driver requires; it recovers afterwards with
backoff instead of hammering a busy camera.
Deployed as an assets-only Cloudflare Worker. The WASM is built with pthreads
and allocates a shared WebAssembly.Memory, so _headers reproduces the COOP/COEP
pair from serve.json - without cross-origin isolation the app fails to start.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QK7PKpRVoy69Fpa32R8abb
