Point README-windows.md at the published v2.0-beta.1 release so users
download Clawdmeter.exe directly from Gitea instead of building it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Package the tray+daemon as a single PyInstaller onefile exe so it runs on a
fresh Windows 11 box with no Python and no pip install. Verified on hardware:
the frozen exe connects over BLE, reads the Windows media session, and pushes
now-playing (incl. Cyrillic) at the 3s / 60s cadence.
- clawdmeter.spec: onefile, windowed (no console; logs to daemon.log). collect_all
for winrt + bleak is the key bit — bleak loads its WinRT backend dynamically and
the winrt.windows.* projections are split distributions that PyInstaller's static
analysis misses, so BLE and the media session would otherwise fail at runtime.
upx off (lowers AV false-positive rate).
- build-exe.ps1: one-command build (venv + deps + pyinstaller + spec).
- tray_windows.py: resolve the brand-logo asset from sys._MEIPASS when frozen.
- autostart_windows.py: when frozen, the HKCU\Run "Start at login" value points at
the exe itself (sys.executable), not pythonw + script.
- .gitignore: ignore /build, /dist, /.venv — the exe ships via a Gitea release.
- README-windows.md: document the standalone-exe path (download / run / build).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Decouple the two data sources that share the BLE link:
- Anthropic usage / rate-limit: still polled every 60s.
- Windows media session: read every 3s and pushed the moment the track or
play/pause state changes, so a song change reaches the watch in seconds
instead of at the next 60s poll. The last usage payload is cached and merged
into each now-playing write, so the firmware always gets one complete JSON
object and the usage screens never blank between polls.
Fix: long media titles (a 44-char Cyrillic title -> ~256 B once json escapes
each char to \uXXXX) overflowed the ATT MTU, so every write-without-response
failed with E_INVALIDARG and tripped the zombie-link break in a reconnect loop.
Switch the RX write to response=True (WinRT does a reliable long write; the RX
char already advertises WRITE and NimBLE reassembles into its 512 B buffer) and
serialize with ensure_ascii=False so Cyrillic goes as 2-byte UTF-8 instead of
6-byte escapes. Verified on hardware: stable link, writes succeed across the
60s heartbeat.
Firmware: only re-set the Now Playing labels when the track / state actually
changed, so the faster cadence does not restart the circular title-scroll
animation on every (often identical) payload.
requirements-windows.txt: add winrt-Windows.Media[.Control]. Phase 5 imports
them but they were never declared, so now-playing silently degraded to
"nothing playing" on a fresh machine.
tests: fix two stale poll_api assertions that expected an "ok" key poll_api has
not emitted since that flag moved to the caller (connect_and_run).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The daemon reads the Windows "now playing" media session (WinRT SMTC) and
adds np/nt/na to the BLE payload; the watch shows a play/pause glyph, a
scrolling title, artist and status. Best-effort and token-independent, so
it works even when the rate-limit data is unavailable.
Cyrillic titles: Styrene B has no Cyrillic glyphs, so the title/artist use
new composite fonts (font_styrene_cyr_{28,20}) — Latin from Styrene B,
Cyrillic + typographic punctuation from Montserrat, merged by lv_font_conv.
Latin titles stay on-brand; only Cyrillic falls back to Montserrat.
- daemon: read_now_playing() via winrt-Windows.Media.Control (lazy import,
never raises; PLAYING->1, PAUSED->2, else 0; title/artist truncated,
empty fields omitted)
- firmware: UsageData np_state/np_title/np_artist + parser; the real Now
Playing screen replaces the shared stub (SCREEN_NOWPLAYING)
- fonts: assets/Montserrat-Medium.ttf, font_styrene_cyr_{28,20}.c
- tools: patch_lvgl9_font.py (automates the 4 LVGL 9 font patches),
screenshot_win.py (Windows serial screenshot via pyserial + Pillow)
- README: document Cyrillic composite font generation
Verified on hardware (waveshare_amoled_206): idle state, a live Latin
title (scrolling), and Cyrillic (Prohozhdenie / Dunduk) all render.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
In a managed Agent-SDK environment `claude login` is unavailable, so once
the stored access token expired the 5h/7d rate-limit % went dark forever.
The daemon now renews its own access token from the stored refreshToken via
the standard Claude Code OAuth refresh grant and writes the rotated tokens
back atomically (temp + os.replace, so a crash can't corrupt the file that
both the daemon and Claude Code read).
- proactive: refresh ~300s before expiry, checked each poll cycle
- reactive: on a genuine API 401, force one refresh + retry the poll once
- refresh tokens rotate on every use -> always persisted
- never raises: any refresh failure is logged and the loop continues
- secrets never logged (only lengths / expiry); error bodies are {"error":..}
Endpoint + client_id verified empirically against a live refresh
(HTTP 200, expires_in=28800).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The Session screen's tokens/cost are computed locally from transcripts and need
no network, but they were gated behind a successful rate-limit API call — so an
expired OAuth token or API outage froze the whole watch.
Daemon: always compute and send local usage every cycle; rate-limit utilization
is now best-effort and merged on top when available, with ok=true only when it's
fresh (ok=false on expired token / API down). poll_api returns just the
rate-limit fields; connect_and_run owns the payload + ok flag.
Firmware: ui_update always refreshes the Session screen; the rate-limit bars and
the freshness clock update only when data->ok, so the usage view falls back to
its existing idle "Zzz" state instead of showing stale or zeroed percentages.
Verified: with the token expired (HTTP 401) the daemon still sends
tk/tc/tn/to + ok:false and the Session screen keeps updating.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Pulling the USB Bluetooth dongle made BleakScanner raise from inside the main
loop; the exception propagated out of daemon_main, the bg thread died, and the
tray froze on its last state with no recovery (reported: "daemon crashed and
doesn't come back").
Two layers of hardening:
1. main() loop now wraps scan+connect in try/except — adapter-gone errors are
logged, the tray shows Scanning, and it backs off and retries, so replugging
the adapter recovers automatically. CancelledError is re-raised.
2. The tray supervises the asyncio loop: an unexpected crash is logged and the
loop is RESTARTED after a backoff, gated by a quit_evt so Quit still stops
cleanly (and call_soon_threadsafe is guarded against a closed loop).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two bugs made the Session numbers wildly inflated (~$300/122M when the real
figures were ~$46/57M):
1. No de-duplication. Claude Code copies prior history into a new transcript
file on every compaction/resume, so the same API turn appears in several
.jsonl files. Summing all lines counted each turn multiple times — ~2.7x
inflation today (a heavily-compacted session). De-dup by (messageId,
requestId), matching ccusage.
2. Wrong Opus pricing. Used the older $15/$75 in/out rates; claude-opus-4-x is
actually $5/$25 (cache-write $10, cache-read $0.50) — 3x too high. Rates
verified against ccusage/LiteLLM, which they now reproduce to 100%.
Verified end to end against a fresh ccusage run: total tokens, output tokens and
cost all match exactly (57,256,809 tok / $45.63). Daemon-only change; no firmware
reflash needed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The total is ~97% cache reads, which reads as implausibly large on its own.
Show the cache-inclusive total as the headline token number and the output
tokens ("what Claude actually generated", ~1.1M) as a small line beneath it —
impressive total, believable sub-number. Daemon sends a new `to` field; firmware
parses output_today and renders it. Re-spaced the Session layout for the 5th row.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The count is assistant API turns (each tool-use round is one), not user chat
messages — "messages" misled into reading it as conversation count / all-time.
"requests" reads correctly.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Daemon: compute_today_usage() sums today's Claude Code token usage across the
local project transcripts (~/.claude/projects/**/*.jsonl), reading only files
modified today so the scan stays cheap, and prices it with Anthropic list rates
to show the equivalent API cost (the ccusage-style flex for subscription users).
The compact fields tk (total tokens), tc (cost in cents) and tn (message count)
piggyback onto the existing 60s BLE payload.
Firmware: UsageData gains tokens_today (64-bit) / cost_cents_today /
messages_today; parse_json reads tk/tc/tn; a new Session screen shows the cost
as the hero with total tokens and message count below, refreshed from ui_update
so it's current whenever opened. Wire SCREEN_SESSION to the real screen (was a
stub). Verified end to end: daemon sends e.g. tk=105374109 tc=26838 tn=663.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add a Soundpad app screen: a 2x3 grid of colored pads that each emit one of
F13..F18 over the existing BLE HID keyboard, for mapping to a host soundboard.
Key release is deferred via a one-shot LVGL timer so the UI thread never blocks.
Raise the HID report-map keycode ceiling from 101 (0x65) to 115 (0x73) in
ble.cpp so F13..F24 (0x68..0x73) are within the declared key-array range — they
were silently undeliverable before. Wire SCREEN_SOUNDPAD to the real screen
(was a stub).
NOTE: the HID report descriptor changed, so the host must re-pair to refresh its
cached map before F13..F18 are recognized.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The tile grid used track_cross_place=CENTER, which vertically centers the
overflowing block of rows — hiding the top row under the title and springing
back to it after an elastic pull. Switch to START so rows stack from the top:
the first row sits under the title and the rest is reached by scrolling down.
Nudge the back chevron down/right so it clears the rounded top-left corner.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replace the splash<->usage tap-toggle with a home + launcher model. Tapping the
Claude logo on the home (usage) screen opens a scrollable 2-column launcher
driven by an app registry — adding a screen is one registry line plus a builder,
no edits to shared navigation. A shared back chevron returns launcher->home and
app->launcher; tapping the splash returns to home.
Screens wired: Usage (home) and Animations (splash) are real; Bluetooth shows
live connection state + device name + MAC; Soundpad/Session/Now Playing/Home
point at a shared "coming soon" stub until their phases land. Enable Montserrat
20/28 for the LVGL symbol glyphs used by tile icons and the back chevron (the
brand Styrene/Tiempos fonts are Latin-only). Flash 20.3% -> 20.9%.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
On-hardware calibration (5-point corner+center tap) confirms the FT3168 on the
2.06 panel reports directly in panel pixel space (~0..409 x ~0..501), correctly
oriented with no axis swap or mirror — a center tap reads ~(209,228) against the
410x502 panel center (205,251), a 1:1 mapping within finger error. The existing
raw passthrough is correct; record the verified result in the driver comment.
The temporary on-screen TOUCH_DEBUG overlay used to gather this has been removed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Freeze the classic line on main (tagged v1.0-classic) and open the v2
development branch. Add a firmware version constant (2.0.0-dev) reported in
the boot banner so the daemon can tell classic from v2, and document the
phased v2 plan in docs/v2-roadmap.md.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
