mirror of
https://github.com/ChuckBuilds/LEDMatrix.git
synced 2026-10-04 14:25:08 +00:00
577f5501a698933e402cbe3c65f7b8ba8ee1bcae
164
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 ChuckandClaude Opus 5
|
dcd6e39c96 |
fix(web): report real disk usage and MemAvailable on the live status stream (#558)
The SSE status stream sent 'disk_used_percent': 0 as a literal, so every
consumer of the live view showed 0% disk no matter how full the card was.
/api/v3/system/status computed it correctly; the stream that the dashboard
actually watches did not. On a Pi with a modest SD card that is the warning a
user most needs, and it was guaranteed to never appear.
The stream also omitted memory_available_mb. /api/v3/system/status carries it
with a comment spelling out why it matters: MemAvailable accounts for
reclaimable page cache, so it is what separates a board reading 70% "used" that
is fine from one reading 70% that is about to fail fork(). A 1GB Pi 3B+ can sit
at either. The number that predicts the failure was missing from the live view.
An unreadable disk now reports None rather than 0. The UI already renders null
as '--'; a confident 0 reads as "plenty of room", which is worse than a blank.
Metric collection moves to web_interface/system_metrics.py, with no Flask or app
imports. That is not cosmetic: importing web_interface.app constructs the Flask
application and a CacheManager, and the latter claims the cache directory with a
cleanup thread. The first version of these tests imported the generator directly
and broke test_cache_cleanup_thread_ownership ("one thread per directory") plus
four starlark route tests through that side effect. Reading a CPU percentage
should not boot a web application, and testing it should not either.
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
|
||
|
ChuckandClaude Opus 5
|
ad5bc4b819 |
perf(sports): LRU-bound the decoded logo cache (#559)
SportsCore._logo_cache was a plain dict keyed by team abbreviation with no eviction. Its entries are not file bytes but decoded RGBA thumbnails sized to display*1.5 -- roughly 36KB on a 256x64 panel, more for wide wordmarks -- and assets/sports/ncaa_logos ships 307 of them. A plugin that walked a full league held the whole league resident: about 11-18MB per manager instance, and a league runs three (live/recent/upcoming) that each keep their own cache, so the same logos were duplicated across them. On the 1GB Pi 3B+ this was measured on, one board was sitting at 439MB resident with ~290MB available, so tens of megabytes of duplicated league logos is real money. Bounded to 64 entries, which holds a full "other games" cycle (on the order of 20 games, 40 teams) without thrashing while capping the cache well below a 307-team league. Eviction is LRU rather than clear-when-full, using the OrderedDict/popitem pattern the neighbouring caches in this codebase already use (_IMAGE_CACHE_MAX, _FIT_CACHE_MAX, _TEXT_WIDTH_CACHE_MAX). That ordering matters: the logos on screen right now are precisely the ones that must not be discarded, so a cache hit moves the entry to the end. Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |
||
|
ChuckandClaude Sonnet 5
|
fb3b293ace |
fix(plugins): let a plugin ask to be polled faster while it has live content (#555)
* fix(plugins): let a plugin ask to be polled faster while it has live content
Reported: "the football plugin with live games only updates the live game in
progress if I restart the display."
The data path was never the problem. NFLLiveManager fetches ESPN with no cache,
SportsLive.update() refreshes current_game in place when the game IDs are
unchanged, and the scorebug redraws from the game dict every frame -- which is
why the reporter's logs look healthy.
The problem is cadence. _get_plugin_update_interval() read only the manifest's
static update_interval, football's manifest pins that to 60, and the plugin's
own live_update_interval (15s) was invisible to the scheduler. Measured on a rig
during the fourth quarter of the game in the report:
23:21:49 23:22:50 23:23:50 23:24:50 23:25:50 <- exactly 60s apart
A clock and score up to a minute stale during a two-minute drill reads as a
frozen panel, and a restart is the one moment it is ever current.
A single static number cannot say "every 15 seconds while a game is on, every 15
minutes in July", and only the plugin knows which is true. get_update_interval()
lets it say so per tick; returning None means "no opinion" and the existing
manifest/config resolution applies, so every plugin that predates this is
unaffected.
Requests are clamped to MIN_DYNAMIC_UPDATE_INTERVAL (5s): a plugin returning 0
would otherwise be re-entered on every tick of the render loop, busy-waiting
against its own API. A hook that raises or returns a non-number is ignored
rather than propagated -- a scheduler that fails on one plugin's bug stops
updating all the others.
Deliberately NOT changed: the manifest still beats config in the static path.
That looked like the obvious fix -- user config being silently ignored -- until
checking a real rig, where football and baseball both carry update_interval 3600
in config against a manifest 60, and weather 1800 against 60. Those values are
stale precisely because nothing has been honouring them; making config win would
have slowed three plugins by 60x, turning a one-minute lag into an hour. The
dynamic hook makes the flip unnecessary. There is a test pinning the current
precedence with that reasoning attached.
Full suite: 4,283 passed, 68 skipped.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9
* test(plugins): drive the real scheduler, not just the interval resolver
test_plugin_dynamic_update_interval.py asserts that
_get_plugin_update_interval() returns the number the plugin asked for. That is
not the same claim as "the plugin gets updated more often", and the gap between
those two is exactly where the original bug lived: the plugin knew it wanted
15s, said so in live_update_interval, and nothing downstream acted on it.
So this ticks the real run_scheduled_updates() through a simulated hour and
counts dispatches. Against pre-fix core it reports "10 updates in 10 minutes of
a live game" -- the 60s manifest cadence, matching what was measured on a rig
during the reported game. Against the fix it reports ~40.
Also pins the regression that would be worse than the bug: an idle hour must
still be ~60 updates, not 240. Asking for the live interval year-round would
poll ESPN four times a minute all summer.
Scope note, since it is easy to over-read this fix: the *switch* display path
already refreshed the manager immediately before drawing, via
_try_manager_display() -> _ensure_manager_updated(), which honours the manager's
own 15s interval. So a switch-mode card was already <=15s stale at draw time
before this change. What this fixes is the background cadence, which is what
live-priority detection, Vegas content and scroll preparation all read.
Full suite: 4,288 passed, 68 skipped.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9
* fix(plugins): reject bool and -inf hook results in dynamic interval
get_update_interval() ran bool through float() (bool is an int subclass,
so True/False became 1.0/0.0) and only checked for +inf, not -inf. Both
cases landed on the MIN_DYNAMIC_UPDATE_INTERVAL floor by coincidence
instead of falling back to the static/manifest interval as invalid
input should.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017Tst9cied2ri9bH4QRWa6H
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
||
|
ChuckandClaude Opus 5
|
28bc79566f |
fix(logo): remember a missing logo instead of re-warning every rotation (#548)
* fix(logo): remember a missing logo instead of re-warning every rotation load_logo() stat'd the path and logged a WARNING on every call, and the positive cache never covered it because a miss returns None and caches nothing. A file that is simply not there therefore produced one warning per rotation for as long as the process ran -- measured on a live rig at 114 lines in 24 hours for a single missing ticker icon, for a file nobody was going to add. Misses are now remembered for 10 minutes: warn once, then return None without touching the disk. Bounded rather than permanent because logo_downloader writes logos at runtime, so a file that appears later must still be picked up without a restart. Downloads through load_logo_with_download() clear the entry outright -- load_logo() consults the miss record before it stats the disk, so without that a freshly downloaded logo would stay invisible for the whole window. This is in the core rather than in ledmatrix-stocks, where it was found, so every plugin that goes through LogoHelper gets it. _cache_order stays a list. Swapping the pair for an OrderedDict would shave an O(n) scan per cache hit, but n is capped at cache_size (100 by default) and test_logo_helper.py pins the current structure; not worth the churn. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 * fix(logo): make the miss TTL longer than the rotation it is meant to outlast Deployed the previous commit to a live rig and measured it: no change at all. "Logo not found for VOO" stayed at ~6 lines an hour, exactly the baseline. The TTL was 600s and the display rotation is ~618s, so every recheck expired just as the plugin came round again and the negative cache never once got to suppress a warning. The fix was correct in shape and useless in practice, which only measuring on the rig would show. An hour instead. That is safe because the TTL is not the main way an entry clears: load_logo_with_download() drops it the moment a download succeeds and clear_cache() drops all of them. The TTL only covers a file that appeared some other way -- someone copying one in by hand -- and waiting up to an hour for that, or restarting, is a fair trade for not re-warning about a file nobody is going to add. The general lesson is in the comment: a TTL has to be long relative to the loop that does the asking, not merely "a while". Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
12f3790994 |
fix(install): render the systemd units from their templates, not from heredocs (#547)
* fix(install): render the systemd units from their templates, not from heredocs The installers carried their own inline copies of units that also exist as templates under systemd/, and the copies drifted. install_service.sh renders ledmatrix.service from the template correctly, then wrote ledmatrix-web.service from a heredoc that predated it -- missing Wants=network-online.target, RestartSec=10, SyslogIdentifier, CacheDirectory, CacheDirectoryMode and Environment=USE_THREADING=1. install_web_service.sh had a third copy, and install_wifi_monitor.sh a fourth, that one already differing from its template (syslog where the template says journal). startup_validator.py compares the installed unit against the template, so a rig installed this way warned on every boot -- and the remedy the warning names, "re-run scripts/install/install_service.sh", reinstalled the same stale copy. The warning could never clear. Reproduced on a live rig running exactly that unit. All three installers now render systemd/*.service through the same placeholder substitution. The template gains a __USER__ placeholder rather than hardcoding User=root, because the web interface runs as whoever installed it. That last point was a second, independent cause of a permanent warning: the validator substituted a fixed "root", so any non-root install reported drift forever. It now reads User= from the installed unit -- an install-time decision, not something the template dictates -- and compares everything else strictly. first_time_install.sh already reads the installed User= the same way. Tests cover a non-root web unit not warning, a genuinely changed directive in that unit still warning, the User= fallback, and a grep-based guard that no installer under scripts/install/ contains an inline unit body. That guard is what found the install_wifi_monitor.sh copy. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 * fix(install): escape sed replacements, use mktemp, and make render failures fatal Address CodeRabbit findings on install_service.sh, install_web_service.sh and install_wifi_monitor.sh: - Values interpolated into each script's sed expression (project root path, username) were not escaped, so a value containing &, \ or the | delimiter would corrupt the rendered systemd unit. Add a shared sed_escape_replacement() helper in the new scripts/install/lib_systemd_render.sh (sourced by all three scripts) and apply it to every sed replacement. - install_service.sh rendered the main and web units to the predictable path /tmp/ledmatrix.service.tmp before installing them -- a symlink/TOCTOU race (CWE-377). Use mktemp for both, with a trap to clean up on exit. - install_service.sh treated a missing template as a mere warning and then checked only whether a unit already existed at the destination before enabling/starting it, so a render failure could silently fall back to enabling a stale, previously-installed unit. Both unit blocks now exit non-zero on a missing template or a failed render. Also rename the ambiguous loop variable `l` to `line` in test/test_systemd_unit_drift.py (Ruff E741); ruff isn't wired into any CI workflow in this repo today, so this isn't currently CI-blocking, but the rename is trivial and correct regardless. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01S3bPMESe2TfrGvbs1ef9c5 * test(install): cover sed_escape_replacement against sed-special characters CodeRabbit asked for regression coverage using a project path containing an ampersand; the earlier commits on this branch already fixed the escaping, mktemp usage, and enable/start-on-fatal-render-failure findings, and the l->line rename was already applied -- this closes the one remaining gap. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
2df273ecfc |
fix(web): default web_display_autostart to true, as the installer already does (#556)
start_web_conditionally.py read the flag with
`config_data.get("web_display_autostart", False)`, so a config that simply
lacked the key got no web interface. Both config/config.template.json and
first_time_install.sh ship the key as true, so the code default contradicted
the shipped default in two places: absence means an older or hand-edited
config, not a request to stay down.
The failure mode was silent in the worst way. The "not starting" path exits 0,
so `systemctl status ledmatrix-web` reported the unit as successfully started
while nothing was listening on the port, and the only trace was one journal
line saying the flag was "false or not set" -- which reads as a deliberate
setting rather than a missing key.
Also start the web interface when config.json is missing or unparseable,
instead of exiting. The web interface is how a config gets created and
repaired, so a broken config is exactly when the user needs it most; leaving
it down means there is no way back in. Only an explicit false/off disables
autostart now, and the disabled message says "explicitly disabled" so the
journal distinguishes a real setting from a default.
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
|
||
|
ChuckandClaude Opus 5
|
a29c84208e |
fix(scroll): advance whole pixels per frame, not per wall-clock second (#545)
* fix(scroll): advance whole pixels per frame, not per wall-clock second Smooth motion is not a frame-rate property, and measuring it as one is why this survived three rounds of fixes. odds-ticker's frame timing is excellent -- 100.0 fps, 10.00ms median, 0% stalls, worst in-scroll frame 19.95ms -- and it still visibly stuttered. What the eye judges is whether the strip advances the same number of whole pixels on every presented frame. update_scroll_position derived position from scroll_speed * delta_time and get_visible_portion truncated it with int(), so jitter in delta_time decided which side of a pixel boundary the position landed on. The live windows show why that matters: a rock-steady 100.0 fps whose individual frames still range 5.6ms to 15.2ms, which at 100 px/s is 0.57px to 1.44px of movement. Run the measured frame times through the real helper and 5.8% of frames advance 0 or 2 pixels instead of 1 -- about six hitches a second. A frame that moves nothing followed by one that jumps two is exactly what micro-stutter looks like. It is worst at a crisp speed, which is the part that stings: at 100 px/s on a 100Hz panel the accumulator sits exactly on integer boundaries, so sub-millisecond jitter flips it either way and the motion beats at around 50Hz. Snapping to the crisp ladder fixes the average and the wall clock then throws away the per-frame uniformity the ladder was bought for. So when scroll_config snaps to a crisp speed it now also puts the helper in fixed-step mode: each presented frame advances exactly pixels_per_frame and no clock is consulted. 100% of frames move by the same amount, whatever the jitter. This is only correct because SwapOnVSync blocks until the panel has taken the frame, which makes the frame count a truer clock than time.time(). Before the swap was locked to vsync it would have run at whatever speed the loop spun at. Related: frame-based mode used to step discretely and was converted to elapsed-time accumulation earlier in this series, because its threshold comparison flipped on jitter. That was right for the code as it stood -- but it treated the symptom, replacing a broken discrete step with a smooth-looking accumulator instead of asking why a wall clock was involved at all. Non-crisp speeds keep pacing off time, and set_scroll_speed() clears the fixed step so a legacy caller changing speed is not silently ignored. Trade-off worth naming: speed is now tied to the presentation rate rather than to real time. If the loop cannot keep up with the panel the scroll runs slow rather than jumping to catch up. That is the better failure -- uniform motion at a slightly wrong speed beats correct average speed with a hitch six times a second -- and a loop that cannot hit the resolved rate is a measurement problem for the crisp ladder, not something to paper over with uneven steps. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * test(scroll): make the time-based pin actually pin something Review caught that test_time_based_stepping_is_what_it_replaces could pass against perfectly uniform motion, and it was right. update_scroll_position sets last_update_time on its way through, so the very first call sees a delta_time of zero and moves nothing in time-based mode. _advances counted that synthetic frame, which put a guaranteed zero in every histogram -- enough on its own to satisfy "uneven > 0". The test asserting the defect exists would have passed after the defect was gone. The first call is now primed and discarded, and the assertion is a proportion rather than "more than zero": against these frame times the old path misses roughly one frame in twenty, so 1% is well below the real rate and far above anything a stray frame could produce. Re-measured with the artefact removed, the numbers in the PR description are unchanged: 5.85% of frames uneven before (114 zero-advance and 120 double frames in 4000), 0.00% after. Also fills in the docstrings the review flagged: everything in the new test file, plus three pre-existing one-liners in scroll_config that the diff touched. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
f8e2e89edc |
refactor(sports): put the scoreboards on the shared scroll resolver (#542)
* refactor(sports): put the scoreboards on the shared scroll resolver Eight sports scoreboards -- afl, baseball, basketball, football, hockey, lacrosse, nrl, soccer -- scrolled through this module's own pacing while the other eleven scrolling plugins went through src/common/scroll_config. Two implementations of the same job, and this one was on the losing side of every difference. It never called set_scrolling_state. Two consequences, both of which this release's work was about: - The frame hold is applied through that call, so a speed the crisp ladder could render in whole pixels still presented a new frame every refresh. - Core only runs deferred updates while nothing is scrolling. Believing nothing was, it ran blocking work in the middle of these scrolls. The default is non-crisp today: scroll_speed 50.0 with scroll_delay 0.01 is 50 px/s, which on a 100Hz panel is half a pixel per refresh. That cannot render as motion -- it alternates 0px and 1px steps and judders at a 50Hz beat, on every scoreboard, out of the box. Resolved through the ladder it stays 50 px/s and holds each frame for two refreshes: same speed, whole-pixel motion. The stepping disagreement that used to justify a separate module is gone. scroll_config avoided frame-based mode because it stepped on a wall clock at 1/scroll_delay with scroll_delay set to the frame period, so the decision sat on its own threshold and flipped on sub-millisecond jitter. That branch now accumulates elapsed time, identical arithmetic to the time-based one, so the two differ only in the units the speed arrives in. What is NOT shared, and must not be: the two modules read identically-named keys with different meanings. Here scroll_speed is px/SECOND and scroll_delay only converts to px/frame; in scroll_config scroll_speed is px per STEP, so px/s is speed/delay. Passing this module's settings dict to the resolver turns 50 px/s into 5000, clamped to 500 -- a tenfold speed-up everywhere. So _get_scroll_settings keeps sole ownership of reading sports config, including the league merging, and hands the resolver a plain px/s. A test pins that specific number, because it is the mistake the refactor invites. MIN/MAX_PIXELS_PER_FRAME are gone; the resolver bounds speed and the helper clamps FPS. _resolve_target_fps stays, re-purposed: under the old model that key was the rate frames were presented at, so it is the faithful translation into the refresh the ladder is computed against, used when no hardware refresh is configured. Speed changes for panels that are not 100Hz: 50 px/s becomes 60 at 60Hz (+20%) and 48 at 120Hz (-4%). At 100Hz it is unchanged. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(display): drop the frame hold when a scroll times out, not just when it says so set_scrolling_state(False) clears the hold. The other way a scroll ends is is_currently_scrolling() deciding, after scroll_inactivity_threshold of silence, that it is over -- which is what happens when the rotation moves on mid-scroll or a plugin is torn down. That path cleared the flag and kept the hold, so every later plugin, scrolling or static, was presented at refresh/N by whoever scrolled last, until something called the explicit stop. The method's own docstring already states the rule this breaks: the hold "must not outlive the scroll that asked for it". The timeout was the exception it did not cover. Pre-existing, but reachable by three plugins before and eleven after the sports scoreboards moved onto the shared resolver, so it belongs with that change. The test ages the activity timestamp past the threshold rather than sleeping. Also adds scripts/sports_scroll_check.py. The sports scroll path is per-league opt-in, so a rig showing static game cards never constructs a SportsScrollDisplay and none of its pacing can be observed from a normal run -- which is exactly what happened when this change was first put on hardware: 26 minutes, zero sports scroll lines. The script drives the path directly with synthetic games and asserts the three things the resolver is meant to buy: the speed lands on whole pixels, the hold is published, and it is released after. It never starts or stops the display service, matching scroll_speeds.py, so a crash here cannot leave the panel dark. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(scripts): refuse to grab the panel while the display service has it The module docstring already said to stop ledmatrix first. Nothing enforced it, and running the script against a live service is not a harmless mistake: rpi-rgb-led-matrix configures GPIO directions and the hardware PWM inside RGBMatrix(), and when the root check fails it calls exit() from C with no cleanup. The service keeps rendering and swapping onto pins that have been reconfigured underneath it, so the panel goes black while every diagnostic says the display is healthy -- fresh framebuffer, every pixel lit, "RGB Matrix initialized successfully", nothing in the log. A restart fixes it, once you work out that is what happened. Found the hard way: this is what took the panel down on the test rig, not the change the script was written to verify. --fallback skips the check, since it never opens the matrix. --force is there for anyone who means it. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * chore(scripts): annotate the subprocess call the way this repo already does Codacy fails a PR on one new issue, and bandit B404 fires on any subprocess import. scripts/run_plugin_tests.py carries the same suppression with the same justification -- list-form argv, no shell -- so this follows it rather than inventing a second convention. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
4423ec33d5 |
feat(plugins): search, filter and sort for Installed Plugins, on a shared ListFilter helper (#540)
* feat(plugins): add search, filter and sort to Installed Plugins, on a shared helper The Installed Plugins grid had no way to narrow it down: no search, no way to see only what's enabled, disabled, or out of date. On a rig with a couple dozen plugins that means scrolling the whole grid to find one. The two sections below it already solved this, twice, independently — the Plugin Store and Starlark Apps carried a copy-paste fork of the same ~600 lines (filter state, apply-filters-and-sort, page renderer, pagination strip, active-filter badge, listener wiring). Rather than add a third copy, this extracts the shared machinery and builds the new toolbar on it. New: web_interface/static/v3/js/plugins/list_filter.js — ListFilter.create() owns debounced search, filter axes, sort, the active-filter count, Clear, and optional pagination/persistence. Callers keep their own card markup via a `render` callback. Three control types cover every axis the page uses: pills (new), select (store category, starlark author) and cycle (the tri-state All -> Installed -> Not Installed button). Installed Plugins gets a compact toolbar: search box, one-click All / Enabled / Disabled / Updates pills, and a sort dropdown (A-Z, Z-A, updates first, recently updated, category). Filters reset on load, so you never come back to a mysteriously short list. No new CSS — this is the first consumer of the .filter-pill rules already sitting unused in app.css. renderInstalledPlugins() is split so it still publishes canonical state while renderInstalledCards() draws only the visible subset; the filtered list is never assigned to window.installedPlugins, which the toggle handler, isStorePluginInstalled(), runUpdateAllPlugins() and the Alpine config tabs all read as their source of truth. Toggling a plugin while filtered pins its card so it doesn't vanish from under the cursor. The Store and Starlark migrations are behaviour-preserving: same element ids, same localStorage keys (storeSort/storePerPage, starlarkSort/starlarkPerPage), same tri-state button markup, same pagination. Verified by differential tests that run the old and new implementations side by side against identical fixtures and compare every observable after each interaction. The only visible change is the pagination attribute (data-store-page/data-starlark-page -> data-list-page), which nothing outside its own click handler referenced. Net -156 lines in plugins_manager.js while adding a feature. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 * fix(plugins): keep raw search text, and stop the store search refetching Two review findings from CodeRabbit on #540. Do not write the trimmed search value back into the input. setSearch() trimmed before storing, and syncControls() then copied that trimmed value back over what the user had typed. Pausing longer than the debounce after typing a space deleted the space (and reset the caret), making multi-word terms effectively untypable. The raw text is now kept alongside the trimmed one: filtering and activeCount() still use the trimmed value, while the input keeps exactly what was typed. Remove the legacy #plugin-search / #plugin-category listeners in initializePlugins(). They bound searchPluginStore as the event handler, so the DOM event arrived as its `fetchCommitInfo` argument — always truthy, which skipped the cached-filter fast path and refetched /api/v3/plugins/store/list with commit info on every keystroke burst and category change. The store's ListFilter controller already filters the cached list, which is what those two controls should do. This double-binding predates this PR (the old code guarded with _listenerSetup and _storeFilterInit, two different flags, so both sets stayed live); it is fixed here because the refactor owns that wiring now. Both fixes are covered by tests that fail without them: the trailing-space regressions in the installed-plugins DOM suite, and a new whole-file jsdom test that counts fetches while typing (1 request at init, 0 thereafter; previously 1 -> 2 -> 3 -> 5). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 * fix(plugins): build pagination via DOM APIs, drop computed member access Addresses the five Codacy security findings, all in list_filter.js. Pagination no longer assembles an HTML string (3 findings: 2 critical + 1 high, "unsafe assignment to innerHTML"). The interpolated values were only page integers and local class constants, so there was no injection path, but concatenating markup into innerHTML is the pattern the scanners flag and createElement is no less clear. Each button now also owns its click listener directly instead of the container being re-queried afterwards, and the strip is cleared with textContent = '' rather than by assigning empty markup. No innerHTML assignment remains in the file. haystack() now walks Object.entries(item) and keeps the configured fields, instead of reading item[field] per field ("generic object injection sink"). Field order no longer drives the haystack order, which is irrelevant to the substring test. matches() iterates controls with for...of instead of an index ("variable assigned to object injection sink"). The rendered pagination is unchanged: same buttons, labels, page numbers, disabled states and classes. The old-vs-new differential tests now compare pagination structurally (tag, text, page, disabled, sorted class list) rather than as an HTML string, since building nodes legitimately serialises differently — «/» as characters rather than «/», disabled="" rather than a bare attribute. That comparison is stronger than the string one it replaces, and the real-DOM suite still drives the actual page buttons. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 * fix(plugins): keep configured field order when building the search haystack The previous commit swapped item[field] for Object.entries(item) to clear a static-analysis object-injection warning, and in doing so changed the order of the haystack: entries follow the object's own key insertion order, not the configured `fields` order. Since the values are concatenated, that order decides which values end up adjacent, so a multi-word query spanning a field boundary matched differently. For store fields [name, description, author, id, ...] and API objects keyed {id, name, description, author, ...}, "bob plugin-01" matched before and stopped matching after. That contradicted the behaviour-preservation claim for the store and starlark migrations, and the differential tests missed it because every fixture query was a single word. Values now come out of a Map built from Object.entries, iterated in `fields` order: the original haystack is restored, and there is still no computed member access for the analyser to flag. Regression coverage for the ordering itself, at both levels: - unit: phrases spanning name->id and category->tags, plus the reverse (object-key) order asserted NOT to match - differential: the same class of query compared old-vs-new, with a guard that the phrase actually matches something so a mutual zero-result cannot pass vacuously Verified both fail without this fix (3 unit, 2 differential) and pass with it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 * test(web): add JS suites for ListFilter and the plugin-manager grids No JS toolchain exists in this repo, so these are plain node scripts with no framework: each prints ok/FAIL lines and exits non-zero. `node test/js/run_all.js` runs everything, skipping the DOM suites (rather than failing) when jsdom is absent or nothing is listening, so it stays useful in a bare checkout. unit/test_list_filter.js ListFilter search/filter/sort/count/sticky, driven through the installed-plugins config eval'd verbatim out of plugins_manager.js so the test cannot drift from the real configuration unit/test_render_cards.js renderInstalledCards markup, both empty states, and escaping of hostile plugin metadata dom/test_installed_dom.js the toolbar in a real DOM, including the HTMX partial re-swap and a getComputedStyle check that .filter-pill[data-active] matches what we emit dom/test_store_dom.js store pagination, per-page, category, tri-state Installed button, persistence across a re-boot dom/test_no_double_fetch.js loads the whole plugins_manager.js and counts requests, so a keystroke cannot refetch the store The DOM suites deliberately fetch the partial and the plugin data from a running web interface instead of using fixtures, so a renamed element id or a changed payload shape fails them loudly. Point them at a rig with a full plugin set when it matters (BASE=http://host:5000); a dev box with two plugins installed passes while exercising very little. Several assertions exist to stop specific bugs recurring: trailing spaces surviving the search debounce, a query spanning two adjacent search fields (haystack field order is load-bearing), and window.installedPlugins staying at full length while the grid is filtered. Others guard against passing vacuously — counting only non-skeleton cards, and checking a search phrase matches something before comparing two result sets. The old-vs-new differential suites that verified the store and starlark migrations are not included: they compared against the pre-refactor code, which now exists only in git history. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
26769ee37f |
fix(starlark): the store authenticated with a key nothing writes (#541)
#535 restored the thirteen routes, so the store stopped answering 404 -- and still would not load. Confirmed against a running device before anything was changed: /repository/browse answers 200 with 1000 apps in 27s, so the routes are fine. Two things underneath them are not. **The store never used the token the user configured.** The three repository routes read `github_token` off config.json. Nothing writes that key -- it is not in config.template.json, no setting offers it, and it appears nowhere else in the codebase. The configured token goes to config_secrets.json as `github.api_token`, which PluginStoreManager loads and every other GitHub caller uses. So the store could never be authenticated: 60 requests/hour, on the same per-IP budget 48 installed plugins spend on update checks, while the 5000 the user had already configured sat unused. On the device, /plugins/store/github-status reported authenticated with a limit of 5000 at the same moment /starlark/repository/browse reported 60, with 18 left. The store going blank was that 60 running out. **Every failure looked identical.** list_all_apps_cached turned any listing failure -- rate limit, DNS, timeout, non-200 -- into an empty app list, and the route sent that out as `status: success`, so a rate limit and an empty repository drew the same blank grid with no error anywhere. It now returns the reason, the route answers 502 with it, and a failure is no longer cached as an empty repository for two hours. The guard for a bad response was itself a crash: _make_request catches `(json.JSONDecodeError, ValueError)` but `json` was never imported, so evaluating the tuple raises NameError and the guard written for exactly this case never ran. Reachable whenever something on the path answers with HTML -- a captive portal, a proxy page, a DNS-hijacking router. Seventeen handlers answered 5xx with no detail at all. test_no_api_v3_handler_discards_its_exception is meant to prevent that across api_v3, but it matched one exact message string, and all thirteen Starlark routes wrote their own wording. The guard now keys on the shape that matters: if it returns 5xx, it says why. The 15 pre-existing non-Starlark functions are listed as a set that may shrink, never grow. **The listing was capped at 1000 and did not say so.** The contents API truncates a directory silently; tronbyt/apps has 1075 app directories, so the store showed a truncated repository and looked complete doing it. Now listed via the git trees API, which reports `truncated`, with the contents API kept as a fallback. Not addressed: the 27-second cold load -- 1075 manifests fetched five at a time behind skeleton placeholders -- which is probably the largest part of what "does not load" feels like, and wants its own change. 25 new tests. Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
e23f1f45d3 |
feat(starlark,on-demand): the third-party fixes worth taking, plus a Home Assistant MQTT bridge (#538)
* feat(starlark,on-demand): the third-party fixes worth taking, plus an MQTT bridge Analysis of ant456/ledmatrix-fixes-repo, a third-party collection of patches and services built while running this project on Starlark apps under MQTT control. Its patches are whole-file copies taken against an older tree, so applying them as written would revert #523's frame pacing, #534's display() bool returns and the GitHub token masking in plugins_manager.js. Three of its claimed fixes are already in main, and its api_v3 Starlark routes are #535's. What follows is the rest -- verified against current code, and reimplemented where the patch's approach did not hold up. **On-demand display.** `pinned` reached the controller from the API, was stored on it and republished in the status payload, but never narrowed the rotation -- a pinned request still cycled every mode its plugin owns. Right for a sports plugin, whose modes are views of one subject; wrong for a plugin whose modes are unrelated, which is every Starlark app. Now honoured, and it survives a restart. Restarting while on-demand was active loaded *only* the on-demand plugin, so normal rotation had nothing to return to for the life of the process -- and a restart mid-session is routine, since that is how an update is applied. The panel came back cycling one plugin's modes with no way out but clearing the cache by hand. Every enabled plugin loads now; on-demand still resumes on its saved mode. Stop requests are exempt from the duplicate guards on purpose, so that a second click stops a mode a race left running -- which means consuming the mailbox is the only thing that ends one. It was never consumed, so the same stop was re-read and re-processed on every poll, forever. Both paths now share one compare-before-delete helper. **Starlark rendering.** `extract_schema` parsed the source with a regex, which can only see option lists written out literally: an app whose dropdown is filled from a live API call inside `get_schema()` came back empty, and the config form offered nothing to pick. Now runs `pixlet schema`, which executes the app, and falls back to the parser when Pixlet is absent, too old for the subcommand, or the app fails to run. The third-party patch replaced the parser outright and hardcoded /usr/local/bin/pixlet; this keeps the fallback and the binary search. A `|` in a config value was dropped by a shell-metacharacter filter, though the command is a list with no shell involved -- and apps do use it as a separator inside one value. The key went missing silently and the app rendered its own "not configured" screen with nothing to say why. And a 0-byte render was reported as success: Pixlet exits 0 and writes nothing when an app has no content, which read downstream as a working app drawing a black panel. **Starlark display.** `display()` ignored the mode it was called with, so a specific app could not be addressed. It now accepts `display_mode` -- which is the whole mechanism, since the controller inspects the signature before passing it. Found while there: `_select_next_app` ran only while `current_app` was unset, so with several apps installed the first was picked once and shown forever while the rest were rendered on schedule and never displayed. And `enable_scrolling` was missing, so multi-frame apps were called once per rotation slot and never advanced past frame one. **GET /api/v3/display/modes.** Every mode that can be requested on-demand, with the plugin that owns it. Nothing exposed this, so anything driving the display from outside the web UI read each plugin's manifest.json off disk and reimplemented PluginManager's fallbacks. It also triggers discovery, which is otherwise lazy and normally happens because a person opened the dashboard. **integrations/mqtt_bridge.** Home Assistant control over MQTT Discovery: a mode select, a stop button, power, brightness. Rewritten against the API rather than the filesystem, so it needs no read access to config.json and cannot drift from the web UI. paho-mqtt 2.x VERSION2, TLS, an availability topic that is also the last will, and secrets from the environment. **Two opt-in extras.** A DNS single-request unit, for glibc's parallel A/AAAA lookup stalling ~5s per name on routers that answer only the A query -- which makes any plugin calling an external API slow and Starlark apps, which have a render timeout, fail outright. And a Pixlet config editor: a script you run and Ctrl+C rather than the third-party version's always-on unauthenticated Flask service, since it stops the display for the length of a session. Neither is installed by default. Long Starlark app names now wrap instead of overflowing their card. 115 new tests across 5 files. Also unblocked test_starlark_display_contract.py, which was silently skipping wherever fcntl is absent. Whole suite: no new failures against main. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(mqtt_bridge): the five issues Codacy flagged on this branch All in the new bridge, all real: * requests floor was 2.31.0, which carries CVE-2024-35195, CVE-2024-47081 and CVE-2026-25645. Raised to >=2.33.0,<3.0.0, which is what the project's own requirements.txt already pins. * `import time` was never used. * `"mqtt_password": None` in DEFAULTS read as a hardcoded credential. It is the "no password configured" default; marked nosec B105, the convention used elsewhere in the repo. Also dropped an unused `build_app` from the display-modes test imports. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix: the review findings on this PR Nine of CodeRabbit's ten, plus the CodeQL alert. The tenth is wrong and is answered below. **One bad config section blanked the whole mode list.** `/display/modes` read `full_config.get(plugin_id, {}).get('enabled')`, so a non-dict under a plugin id -- a shape DisplayController already guards, so it happens -- raised AttributeError mid-loop and answered 500 with no modes at all. Every MQTT bridge entity is built from that list. Now skipped with a warning. **The DNS scripts reported success they had not earned.** Three separate paths: `resolvconf -u` failing was swallowed by `|| true`; the systemd-resolved branch exited 0 without applying anything, so the oneshot unit recorded success while the workaround was inactive; and the installer's `|| echo` turned a failed start into "installation complete." with exit 0. All three now fail loudly. `single-request` is a glibc resolv.conf option with no resolved.conf equivalent, so on those hosts the honest answer is that it cannot be applied. A NetworkManager-generated resolv.conf is regenerated on connection changes, not only at boot, and the unit is oneshot with RemainAfterExit -- so the option can vanish mid-boot with nothing to put it back. Now detected and stated plainly rather than implied to be permanent. **`Before=` does not order a manual restart.** It only orders units already in the same transaction, so `systemctl restart ledmatrix` could bypass the fix. install_dns_fix.sh now writes a ledmatrix.service drop-in with Wants= and After=. Wants=, not Requires=: a DNS workaround failing should not stop the display. **The Pixlet editor's `--lan` is gone.** `pixlet serve` has no authentication, and a printed warning is not access control. Loopback only, with the SSH port-forward in the header where the flag used to be documented -- SSH does the authenticating and nothing is left listening. **The MQTT example config now defaults to TLS** on 8883. The installer copies it verbatim, and without TLS the broker password and every command cross the network in cleartext. A plaintext broker is still supported and documented, and the bridge warns once at startup when a password is configured without TLS. **Not taken: "the upstream Pixlet CLI has no `schema` subcommand."** Upstream tidbyt/pixlet has none, but `scripts/download_pixlet.sh` installs `tronbyt/pixlet`, whose `cmd/schema.go` is `schema [PATH]` -> JSON on stdout, built on `runtime.NewAppletFromPath`, so it does execute `get_schema()`. That is exactly what extract_schema_via_pixlet calls. A binary without the subcommand exits non-zero and falls back to the source parser, which is already covered by a test. **CodeQL stack-trace exposure: not taken either.** I removed `details` first and that broke test_web_error_detail.py::test_no_api_v3_handler_discards_its_exception, which enforces `describe_exception` across all ~75 handlers -- written because a device with failing storage answered "see logs for details" from the log viewer itself. describe_exception redacts credentials; the trade-off is the project's and is already made. Restored, with the reasoning in a comment. 11 new tests. Whole suite: no new failures against main, 4127 passed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
793b988d33 |
fix(starlark): toggle the app id the list published, and store a relocatable star_file (#537)
The two review nitpicks left over from #535. Both are still on main after that merge; the five findings alongside them landed with it. **The toggle could not find what the list had just shown.** `_starlark_virtual_plugins` publishes the raw manifest key as `starlark:<key>`, and `_toggle_starlark_app` passed it back through `_validate_and_sanitize_app_id`, which lowercases and rewrites every character outside `[a-z0-9_]`. An app stored as `My-App` was listed as `starlark:My-App` and looked up as `my_app`, so toggling an app the page had drawn a moment earlier answered 404. Keys written by `_install_star_file` are already sanitised, so this only shows up for manifests written by the starlark-apps plugin itself or edited by hand. `_validate_starlark_app_path` rejects traversal without rewriting, so it is the check to use here -- listing and toggling now agree on one key. The updater also uses `setdefault` rather than indexing: the app is loaded but its on-disk entry need not exist, and `_update_manifest_safe` does not catch `KeyError`, so that escaped as a 500 rather than writing the entry. **`star_file` was stored absolute.** Readers join it to the app's own directory -- `_standalone_render_starlark_app` does `app_dir / app_data.get('star_file', f'{app_id}.star')` -- so the key's default is a bare filename and an absolute value gave it a second meaning. Since `Path.__truediv__` discards the left side when the right is absolute, the manifest was pinned to whatever PROJECT_ROOT installed it, and a moved or redeployed install could not find its own file. Storing `dest.name` matches the default and stays relocatable. Read paths are unchanged, so manifests already holding an absolute path keep working. 7 new tests. Whole suite: no new failures against main, 4013 passed against 4007. Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |
||
|
Chuck
|
50258635a8 |
fix(starlark): restore the API routes #330 dropped (#535)
The Pixlet install button reported "Pixlet install failed: Resource not found" -- Flask's 404 handler, because the route did not exist. #253 added thirteen Starlark routes; #330 rewrote api_v3.py and dropped all of them, along with the `starlark:<app_id>` entries that surface installed apps in the plugins list and the toggle branch that enables them. Restores all thirteen routes, the plugin-list entries and the toggle path, so Pixlet installs, the app store browses and installs, and an installed app can be managed like any other plugin. Not a straight revert. Three error paths stopped returning exception text to the caller; the manifest write moved off a shared temp filename that two concurrent writers could interleave; both dynamic importers stopped leaving half-initialised modules in sys.modules; the config update rolls back when the save fails; the toggle checks that persistence succeeded; and the path check returns the validated path instead of a boolean so callers stop re-joining the raw value. New tests no longer reach GitHub. Verified on a 256x64 Pi: Pixlet installs and runs (v0.53.1), the store lists 1000 apps, install/toggle/uninstall round-trip, and traversal and command-injection probes are rejected at every entry. 25 CodeQL alerts dismissed as verified false positives -- path-injection where traversal is blocked, and one list-form subprocess with no shell. Both classes already present on main. Full core suite: 3981 passed. |
||
|
ChuckandClaude Opus 5
|
c9289e3a1d |
fix(store): update_plugin silently did nothing for four installed plugins (#536)
install_plugin() deliberately renames a plugin's directory to the MANIFEST id
when it differs from the REGISTRY id, so registry `stocks` lands in
`ledmatrix-stocks/`. Every lookup in _find_plugin_path() is by directory name,
so update_plugin("stocks") found nothing, logged "Plugin not installed", and
returned False.
Nothing surfaced that to the user. Clicking update in the web UI was a no-op
with no error, and the plugin stayed on a stale version indefinitely. Four
installed plugins hit this on a real device -- leaderboard, music, stocks and
weather -- found because a scripted update of eleven plugins failed on exactly
those four.
Adds a manifest-id scan as the LAST step of the resolution chain, so the two
documented lookups above it (configured dir, then the sibling plugins/
fallback) keep their exact meaning and ordering. That ordering is pinned by
test_discovery_path_contract.py, which characterises the divergence between
the three resolvers on purpose; this extends the chain rather than reordering
it. Directories renamed aside with '.standalone-backup-' during an install or
rollback are skipped, since matching one would report a half-finished install
as a live plugin.
Also adds scripts/audit_render_path.py, which walks the call graph from
display() and reports blocking calls reachable from it. display() runs on the
render thread, so anything slow there stalls the panel; on a vsync-paced loop
a single 15ms call drops a frame and a network round trip freezes the marquee.
Two instances were already found the slow way, by reading frame-time
histograms -- odds-ticker reading the scoreboard cache per frame, and
soccer-scoreboard timing out inside update(). The audit finds that shape in
the source instead. It is a heuristic and says so: a hit behind an interval
check may be fine.
It currently flags 23 calls across six plugins. The clearest is
ledmatrix-music, whose display() falls back to an inline
requests.get(timeout=5) when album art has not been prefetched -- a deliberate
"show the art rather than go blank" tradeoff by its author, but up to five
seconds of frozen panel. Reported, not changed; that is its owner's call.
185 store tests pass. Three of the six new tests fail without the fix.
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
|
||
|
ChuckandClaude Opus 5
|
d12323e7f1 |
perf(scroll): pace frames to the panel — 44→100 fps, stalls 14% → 0.02% (#523)
* perf(scroll): pace frames to the panel, not to a fixed sleep Scrolling ran at 44-46 fps on a 2x128x64 chain and 14-17% of frames took 41-53ms, which reads as judder. Four independent causes, each measured on the hardware; details and the diagnostic recipe are in docs/SCROLL_PERFORMANCE.md. The high-FPS loop slept a flat 8ms after every render. display() has already blocked on the panel's vsync by then, so that sleep was added to a wait that had happened: ~4ms of render plus 8ms put each iteration at ~12ms against a 10ms refresh grid, so every swap missed a refresh and the loop settled at 50fps while asking for 125 -- with no headroom, so a further 14% of frames slipped again. It now sleeps only the remainder, with a 1ms floor so plugin threads still get the GIL. ScrollHelper stepped position on a wall clock at 1/scroll_delay steps per second. Plugins set scroll_delay to the frame period, so that comparison sat exactly on its own threshold: a frame arriving a hair early moved zero pixels and rendered an identical frame, dirty-tracking skipped the swap, it returned in ~2ms, and the beat repeated. No scroll_delay value tunes that out -- a shorter delay trades stalled frames for periodic double-steps. Both modes now accumulate elapsed time at the same configured speed, so position stays proportional to real time. Sub-pixel blending goes back to off by default. It renders a half-step by mixing two adjacent columns, which on a coarse panel showing pixel-font text alternates crisp and smeared frames and reads as shimmer -- visibly worse than integer stepping on the hardware. Vegas mode still opts in. disk_cache uses orjson when importable, falling back to the stdlib. Encoding a ~1MB record drops from 14.8ms to 5.4ms end-to-end, and that work holds the GIL while a marquee is on screen. display_manager also checksummed the whole framebuffer twice per frame (dirty tracking, then the preview snapshot); the snapshot now takes the checksum the caller already computed. New src/common/scroll_config.py resolves scroll settings in one place. Five ticker plugins each hand-rolled this and disagreed: odds-ticker ranked the deprecated scroll_pixels_per_second above the documented scroll_speed/delay pair, and because that key carries a schema default the documented settings were dead for every user (ChuckBuilds/ledmatrix-plugins#408), while ledmatrix-leaderboard read the same key only as a fallback. The resolver also warns when a speed will not advance a whole number of pixels per refresh, which is the property that actually determines whether a scroll looks smooth. scripts/build_rgbmatrix_nogil.sh rebuilds the rgbmatrix binding so it releases the GIL. Upstream declares SwapOnVSync without nogil, unlike SetPixel/Clear/Fill beside it, so the render thread held the GIL for the whole vsync wait and starved background threads into long uninterruptible bursts. The script patches, builds and self-verifies into a scratch tree; --install backs up the original and rolls back if the service does not come back healthy. Measured after: 100 fps locked, no stalls observed, render thread down from 51% to 19% of one core. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(display): keep the panel swap locked to vsync while scrolling Dirty tracking skipped SwapOnVSync for byte-identical frames. That is the right call for static content, but SwapOnVSync is also what paces the render loop, so skipping it skips the wait for the panel: a duplicate frame returns in ~8ms instead of ~10ms on a 100Hz panel, advances the strip only 0.8px instead of 1.0px, and so makes the next frame more likely to repeat as well. The effect sustains itself once it starts. Measured over 20 minutes on a 2x128x64 chain, both scrollers configured identically at 100 px/s: leaderboard 10ms x35, 11ms x3 (clean) odds-ticker 10ms x26, 8ms x7, 15ms x5 (~20% duplicates mid-scroll) The duplicates were not end-of-cycle idling -- 38% of fast frames fell within 90s of a scroll completion against 35% of normal frames, a null result. The trigger is per-frame work: odds does more of it, and more variably, so it is first to land a frame that advances less than a whole pixel. Pushing an identical frame costs one canvas copy. Falling out of vsync lock costs smooth motion. Static content is untouched, because is_currently_scrolling() expires on its own inactivity threshold -- covered by test_stale_scrolling_state_stops_forcing_pushes so a plugin that stops scrolling without saying so cannot pin the panel into always-push. Also de-flakes test_snapshot_still_written_on_skip, which asserted a strict mtime increase between two writes that can land in the same filesystem tick; it failed about two runs in three on Windows regardless of the code under test. The file is now backdated before the check. 156 tests pass on the Pi. Not yet confirmed by eye on the panel. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(scroll): report the frame-time tail, and stop the row-major blit Two problems, both found by looking at the panel rather than the metric. The frame-stats line reported ONE instantaneous frame every 5 seconds -- about 1 frame in 500 -- printed beside a 100-frame average. Both hide exactly the fault they are used to chase: a 2ms duplicate and a 21ms double-wait average to precisely 10ms, so a ticker stalling on half its frames still reports a healthy "Avg FPS: 100.0". That reading cost several rounds of chasing the wrong layer. The line now aggregates every frame since the last log and reports median, p95, max, min, and explicit stall and skip rates (past 1.5x the median missed a refresh; under half never reached the panel, because dirty tracking skipped the swap so the frame never waited on vsync). On the hardware this now reads: leaderboard 100.0 fps over 501 frames | median 10.00ms p95 10.05ms max 10.34ms | stalls 0 (0.0%) skips 0 (0.0%) The binding rebuild's blit patch becomes opt-in (RGB_PATCH_BLIT=1, default off). Reordering that loop to row-major changes what a torn frame looks like: column-major tearing shows as a vertical seam, row-major as a horizontal split between the panel's upper and lower halves. On a 1/32 scan panel that reads as a one-pixel fold across the middle of every panel, which is what was reported on hardware and what went away when the blit was reverted. All of the measured gain comes from the SwapOnVSync change, so the risky half is simply not worth taking; the header says so. Also fixes --install resolving its paths against $HOME, which is /root under sudo, so it looked in /root/rgbmatrix-nogil-build and died with "no built module found" on a machine where the build had just succeeded. It now resolves SUDO_USER's home. Both build paths are verified on the Pi: default yields one GIL-release site, RGB_PATCH_BLIT=1 yields two. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(scroll): let users pick a crisp speed for their own panel Whole-pixel motion was previously only available at multiples of the refresh rate -- 100, 200, 300 px/s on a 100Hz panel. 100 px/s crosses a 256px panel in 2.6s, which is brisk for reading, and everything slower had to blend (blur) or repeat frames unevenly (judder). There was no way to ask for 50 px/s and get clean motion. SwapOnVSync takes a framerate_fraction the display manager never passed. It holds each frame for N panel refreshes; the panel keeps refreshing at its full rate throughout, so holding costs nothing in flicker and only changes how often a NEW image is presented. That turns 50 px/s into one whole pixel every second refresh instead of half a pixel every refresh. The crisp speeds are therefore refresh_hz / hold * pixels_per_frame, and that ladder depends on the panel: a Pi Zero on a long chain has a different set of good speeds from a Pi 4 on a short one. crisp_ladder() enumerates them and solve_crisp() picks the best match for a requested speed. solve_crisp weights motion quality rather than picking the numerically nearest entry, which matters more than it sounds. Asked for 30 px/s, nearest-by-value answers 28.6 -- 2px jumps at 14fps -- over 33.3, which is single-pixel motion at 33fps and obviously better on the panel. The target is also clamped into the ladder's range first, because relative error saturates near 1.0 for a target far outside it and the quality penalty would otherwise answer "10000 px/s" with the slowest entry. configure() snaps to the ladder and applies the hold when given a display manager. Without one the hold silently cannot happen and motion falls back to fractional pixels, so it warns rather than failing quietly. set_frame_hold() resets to 1 when scrolling stops, so one plugin's pacing cannot leak into whatever is on screen next. scripts/scroll_speeds.py is the user-facing part: it prints the ladder for the configured rate, measures what the panel ACTUALLY manages (--measure, for hardware that cannot reach its configured limit), highlights the nearest option to a wanted speed, and demos one live. It never starts or stops the display service itself -- doing that inside a script stranded the panel twice today. Speeds below ~20 px/s remain stepped regardless. That is the pixel pitch, not a software limit. Also fixes the dirty-tracking test spy, which stubbed SwapOnVSync with a single-argument function and would have masked the new call as a failed push, and rewrites a configure() test that had started passing for the wrong reason: it asserted a judder warning, which snapping now prevents, and was matching the unrelated "hold could not be applied" warning instead. 183 tests pass on the Pi. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(scroll): tie the frame hold to the scroll, not the plugin The hold applied in configure() never reached the panel. Plugins share one display manager, and set_scrolling_state(False) -- fired whenever ANY other plugin finishes its scroll -- reset the hold to 1. A hold set once at plugin construction was therefore always gone by the time that plugin rendered. The symptom was a log line that lied. ledmatrix-stocks reported Scroll configured: 50.0 px/s (1px every 2 refreshes = 50.0 fps, smooth) while the panel measured 100.0 fps, median 10.00ms. Config, resolution and snapping were all correct; only the pacing silently was not applied. set_scrolling_state(is_scrolling, frame_hold=1) now carries it, so the hold lives exactly as long as the scroll that asked for it. configure() reports the value as ScrollSettings.frame_hold instead of applying it -- applying it behind the caller's back could never have been right on a shared display manager. Existing callers are unaffected; the default keeps one frame per refresh. Verified on hardware: stocks at 50 px/s now measures 50.0 fps over 251 frames | median 20.00ms p95 20.09ms | stalls 0 skips 0 20.00ms being exactly two refreshes, with the panel still refreshing at 100Hz underneath so flicker is unchanged. test_another_plugin_stopping_does_not_strand_a_hold pins the interaction that broke this. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(scroll,cache): resolve CodeRabbit review on #523 Eight findings, all reproduced before fixing. scroll_config.configure() read the refresh rate *after* resolve() had already used it. resolve() fills in target_fps, pixels_per_frame and the judder warning from that rate, so on a 60Hz panel every one of them described 100Hz -- and with snap_to_crisp=False nothing downstream corrected it, so set_target_fps() paced the helper to 100 FPS. The rate is now settled first, and falls back to the global config rather than straight to the default. refresh_hz_from_config() used `(cfg.get("display") or {}).get(...)`, which raises AttributeError when either level is truthy but not a mapping -- out of a function whose whole contract is a rate or a default. The frame-stats line reported the upper-middle sample as the median and the 96th sorted sample as p95 of 100. Both are also thresholds (stalls at 1.5x the median, skips at 0.5x), so the counts were biased too. The arithmetic is now in frame_stats()/format_frame_stats(), testable without a clock. configure()'s docstring and docs/SCROLL_PERFORMANCE.md still said it applies the frame hold and warns when it cannot. It deliberately does neither since "tie the frame hold to the scroll, not the plugin"; a caller following the old text would omit set_scrolling_state() and slow snapped speeds would still present every refresh. disk_cache had no policy for non-finite floats: orjson writes null, the stdlib writes NaN/Infinity, and orjson then rejects those legacy files so DiskCache.get deleted them as corrupt. One behaviour on both paths now -- write null, keep legacy records readable. allow_nan=False detects the values; the replacement walk runs only when there is one, so the ordinary write path is byte-identical and pays nothing. build_rgbmatrix_nogil.sh picked the build artifact with a glob piped to `head -1`, which sorts cpython-311 ahead of cpython-313, so a stale .so staged in from the source tree was installed as core.so while the GIL check -- which reads the generated core.cpp, not the .so -- still passed. It now requires the current interpreter's exact ABI name and fails closed. Its systemctl calls were also unchecked under `set -uo pipefail`: a failed stop left the old service running, the following start succeeded as a no-op, and the health check reported SUCCESS for a binding that was never loaded. orjson floor raised to 3.11.6 for CVE-2025-67221 (unbounded recursion in dumps); it covers the project's Python 3.10-3.13 range. Adds test/test_cache_nonfinite_floats.py (14) plus regression tests in test_scroll_config.py and test_scroll_helper.py. 9 of the cache tests and 9 of the scroll_config tests fail against the pre-fix code. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 * test(harness): keep the visual double's signature tied to production Moves set_scrolling_state's frame_hold into the test double here, where DisplayManager gains it, rather than in #534 where it arrived a PR early. CodeRabbit flagged the #534 version correctly: a double that accepts an argument production does not lets the call pass every harness run and raise TypeError on the panel, which is the one failure a safety harness exists to prevent. The drift has now gone both ways across two branches -- double behind production on this branch, double ahead of it on #534 -- so it is pinned instead of remembered. test_display_double_parity.py compares the two signatures and fails with the direction of the drift named. It reads the files with ast rather than importing them, because display_manager imports rgbmatrix at module scope and this check should hold on a laptop and in CI as well as on a Pi. Plugins begin passing frame_hold in ledmatrix-plugins#462, which is why production and the double both need it before that lands. Full suite: 3889 passed. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
a0d3e64099 |
fix: ten defects found validating the whole plugin fleet on hardware (#534)
* fix(core): register tom_thumb, accept frame_hold in the test double, wire api_v3's managers Three independent fixes found while validating every plugin on a 256x64 rig. FontManager never registered tom_thumb even though assets/fonts/tom-thumb.bdf ships with the core, so every plugin offering it logged "Font family 'tom_thumb' not found" (16 warnings per countdown render) and had to carry a private loader to use a bundled font. Closes #524. VisualTestDisplayManager.set_scrolling_state() lacked the frame_hold parameter that DisplayManager gained, so any plugin passing it died with TypeError at render time and failed every size. Nine plugins now make that call; ledmatrix-stocks and ledmatrix-leaderboard were failing outright and the other seven only passed because their scroll path was unreachable without data. Closes #525. api_v3 declared module-level config_manager/plugin_manager = None that nothing ever assigned -- app.py sets the blueprint attributes, which the other 150+ call sites use. Three sites read the decoys, so /health reported the config unreadable and the plugin system uninitialised (making "degraded" permanent and unreachable-by-design) and /display/current fell back to a hardcoded 128x64 on every rig. The decoys are removed rather than assigned, so a bare name is now a NameError at test time instead of a silent None. The same function's first-call uptime was computed from two separate clock reads and came out negative. Closes #529. Verified on the rig: both previously-failing plugins render, the tom_thumb warnings are gone, /health reports "healthy" with all three checks passing, and /display/current reports the real 256x64. * fix(core): unique snapshot temp name, honour on-demand requests, skip empty starlark The preview snapshot wrote through a fixed "<snapshot>.tmp". /tmp is world-writable and sticky, and the display service runs as a different user from the tooling, so a leftover temp owned by anyone else became unopenable even by root -- fs.protected_regular refuses O_CREAT on a foreign file in a sticky directory. The preview and the health check's liveness proxy then froze until someone deleted the file by hand; on the test rig that meant 23 hours of a healthy display reporting "hardware: stale". Now uses tempfile.mkstemp with cleanup on failure, matching the hardware-status write a few hundred lines above. Closes #528. _poll_on_demand_requests read its mailbox with max_age=3600, and get() defaults the in-memory TTL to max_age -- so the first request was pinned in memory for an hour and every later poll returned that stale copy. No second on-demand request was honoured until the service restarted, while the API kept returning 200. get() already documents memory_ttl=0 for exactly this cross-process case. The consumed request is also now deleted: leaving it on disk meant a restart replayed the previous request, activated it, and ignored the one the caller had just made. Closes #530. starlark-apps returned None from display() when it has no app to show, which is the state of every install without Pixlet and of a fresh one before any app is added. The controller only skips on a boolean False, so that held a black panel for the full display_duration instead of rotating on. Closes #456 (core side). Verified on the rig: two consecutive on-demand requests with no restart between them are both activated, where the second was previously dropped in silence. * perf(harness): share one cache across a plugin's renders _instantiate built a fresh MockCacheManager for every (size, mode), and that mock is a per-instance in-memory dict, so each render was a cold start. A plugin that fetches per game or per player re-fetched everything N times over -- baseball-scoreboard at one size took 840s for nine renders where the arithmetic said ~72s, and at eight sizes it exceeded a 900s timeout. The second and later renders also never exercised the cache-hit path, which is what a running rig executes almost all of the time, so a caching regression could not be caught here. The cache is now built once per render_plugin_matrix call and threaded down. The display manager stays per-render -- the bounds checking depends on that -- so only fetched data is shared. Measured on the rig, same render counts and same goldens: tide-display 2s -> 1s (32 renders) cricket-scoreboard 10s -> 3s (24 renders) No pass/fail change across tide-display, cricket-scoreboard, clock-simple, geochron, christmas-countdown, of-the-day, web-ui-info and incoming-packages. Closes #533. * fix(scripts): run standalone plugin tests instead of collecting nothing run_plugin_tests.py discovered every plugin test file and handed the lot to pytest. Most plugin tests are standalone scripts -- module-level main() plus an `if __name__ == "__main__"` guard, signalling through an exit code -- and pytest collects zero items from those. The run printed how many files it had *found*, then "no tests ran", and exited without executing any of them. On a rig with all 44 first-party plugins that is 151 of 248 files. Files are now classified and each kind runs under the right runner: pytest for real test modules, subprocess for scripts, honouring the 0 pass / 2 skip / 1 fail convention ledmatrix-plugins' own runner established (a script that wants a tty or an LED matrix is a skip, not a regression). Before: $ python3 scripts/run_plugin_tests.py -p countdown -d ~/LEDMatrix/plugin-repos Found 1 test file(s) collected 0 items no tests ran in 0.31s rc=0 After: Found 1 test file(s) -- 0 collectable, 1 standalone script(s) 1 passed, 0 skipped, 0 failed (scripts) rc=0 Verified across three shapes: countdown (1 script), jellyfin-now-playing and pomodoro-timer (pytest only, 16 and 42 tests), and ledmatrix-flights (11 files split 4 collectable / 7 scripts, all seven of which had never run). Closes #532. Running the flights scripts for the first time also surfaced four genuinely failing tests there, hidden by the mirror-image bug in the plugins repo's own runner -- filed as ChuckBuilds/ledmatrix-plugins#464 and #465. * fix(harness): give an empty-looking mode a few frames before warning about it check_plugin's "drew nothing but display() returned X" warning fired on a single frame, rendered with force_clear=True, under a frozen clock. All three defeat a scrolling plugin, whose first frame is legitimately its blank scroll-in buffer. Across 44 first-party plugins, 60 of 76 warnings were false -- the rate at which people stop reading a warning, which matters because the true positives are real: a mode that draws nothing and does not return False holds a blank panel for its whole display duration. An apparently-empty frame is now re-driven for up to 48 more frames with force_clear=False (force_clear means "reset the scroll", so repeating it would redraw frame 1 for ever) and with the clock advancing -- freezegun's factory where time is frozen, a real sleep where it is not, since scroll position is usually a function of elapsed time. The first frame that draws content replaces the result. The clock is moved back afterwards. It is shared by every render in the matrix, so time borrowed by the probe leaked into later modes and drifted their goldens -- f1_upcoming picked up 5 spurious drifts before this was restored. Measured on the rig: empty warns check before after f1-scoreboard 42 0 48 PASS / 0 FAIL, goldens intact ledmatrix-elections 16 0 16 PASS / 0 FAIL on-air 8 8 true positive, kept nfl-draft 8 8 true positive, kept clock-simple/geochron/ 0 0 unchanged christmas-countdown 58 false positives gone, both true positives kept, no golden regressions. Cost is confined to modes that really are blank: plugins that draw immediately are unchanged (clock-simple and tide-display still 2s), while on-air -- eight deliberately blank modes -- goes to 21s. Closes #527. * fix(harness): load nested schema defaults, and merge caller config at leaf level load_config_defaults read only top-level properties. An object property carries its defaults on its children, not on itself, so everything nested was dropped -- 2,386 defaults across 37 of 44 plugins, soccer-scoreboard alone losing 539 of 565. render_plugin_matrix's comment says the plugin then "behaves like a real install", which for most of the fleet it did not. _defaults_from_properties now recurses. merge_config deep-merges the caller's config onto the result so an override lands at the leaf: a shallow merge would let -c '{"nhl": {"enabled": true}}' replace the whole nhl subtree and discard every other nhl default, which is the same class of bug being fixed here. Measured before/after across all 49 installed plugins on the rig: **no render changed** -- identical PASS/FAIL counts, byte-identical output, goldens intact. Plugins already fall back to the same values internally via config.get(key, default), so supplying them explicitly agrees with what they were doing. The defaults really are arriving now: ufc-scoreboard 9 -> 87 defaults ledmatrix-flights 51 -> 95 masters-tournament 10 -> 51 cricket-scoreboard 22 -> 50 tide-display 12 -> 18 and hockey-scoreboard, which used to load nhl.enabled=None, now gets nhl.enabled=True with its full display_modes block. Caveat worth carrying: the eight plugins with the most nested config (soccer, baseball, basketball, hockey, lacrosse, football, afl, nrl -- 1,634 of the 2,386 dropped defaults, 68%) could not be measured. They import src.common.sports_shared, which the test rig's core branch predates, so they fail to load there identically before and after. Re-run this comparison against a core that has that module before trusting the "nothing changed" result for them; those are exactly the plugins whose renders should change most. Closes #531. * refactor: narrow the exception handlers this branch introduced Codacy flagged the new code; it passes on other recent PRs, so the finding is mine. Four of the five broad `except Exception` clauses I added were catching far more than they needed to, which is the same shape as several bugs this branch fixes -- hello-world's TypeError sat invisible for exactly this reason. freezer() / move_to() / tick() -> (AttributeError, TypeError, ValueError) cache_manager.delete() -> (OSError, AttributeError, KeyError) The fifth stays broad and now says why: it wraps a call into a plugin's own display(), which can raise anything, and the first frame has already rendered -- so a failure there must not turn a good result into an error. Verified against a checkout of main: f1-scoreboard 48 PASS / 0 FAIL with 0 empty warnings, on-air keeps its 8 true positives, clock-simple 8 PASS. geochron shows 7 golden drifts both before and after this branch, so it is not from these changes -- its committed goldens predate #521's 1-bit text rendering. * fix: resolve CodeRabbit review and Codacy findings on #534 CodeRabbit raised six; all six were real. The test double had drifted ahead of production. VisualTestDisplayManager accepted set_scrolling_state(frame_hold=...) while DisplayManager did not, so such a call passed every harness run and would raise TypeError on the panel -- the one failure a safety harness exists to prevent. frame_hold belongs to the change that adds it to DisplayManager (#523), so it moves there and the double matches main again. The harness swallowed exceptions from re-rendered frames. _settle_loop re-renders a mode that came back blank, to give a scroll time to draw; returning silently on a crash meant a mode that renders one good frame and then explodes was reported as passing. Recorded on result.error now, keeping the captured frame so the failure stays inspectable. starlark-apps display() returned True after _display_frame() failed, so the controller held a dead frame for the whole display_duration instead of rotating on. _display_frame now returns bool on all three paths. run_plugin_tests.py used env.setdefault for PYTHONPATH and LEDMATRIX_CORE, so an inherited value won and the subprocess imported a different core than the one under test -- ledmatrix-plugins#467 exactly. Prepends PROJECT_ROOT and sets LEDMATRIX_CORE unconditionally. The on-demand mailbox is polled after every frame, ~125x/second on a scrolling mode, and the read is deliberately uncached, so it was that many disk reads per second to find nothing. Floored at 250ms, which is imperceptible for a web-UI click. Consuming it also deleted whatever was present rather than what had just been processed, so a request posted while the previous one was in flight was thrown away and never ran; the delete is now keyed by request_id. That narrows the window rather than closing it -- a true atomic claim needs a primitive the cache layer does not offer, and the code says so rather than implying otherwise. Codacy's 2 criticals were bandit B404/B603 on the subprocess call added to run_plugin_tests.py. Fixed interpreter, argument list, no shell; annotated with the repo's existing nosec convention. Bandit is clean on the file. Adds test/test_on_demand_mailbox.py (8), test_starlark_display_contract.py (4) and two settle cases in test_harness_empty_claimed.py. 4, 4 and 2 of those fail against the pre-fix code. Full suite: 3961 passed. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 * chore: satisfy Codacy's subprocess checks on the new test runner Codacy runs Bandit and Opengrep (its Semgrep fork). The new subprocess.run in scripts/run_plugin_tests.py trips three patterns, on two different lines: Bandit B404 on the import, B603 on the call Opengrep dangerous-subprocess-use-audit on the run( line dangerous-subprocess-use-tainted-env-args on the argv line A nosemgrep applies only to its own line, so the call line and the argv line each need one; a single comment on the call covered neither rule fully. Suppression is the right answer here rather than a rewrite: the interpreter is sys.executable, the arguments are a list, and no shell is involved, so there is nothing to word-split or expand. Matches the pair the rest of the repo already uses for this shape -- permission_utils.py, plugin_loader.py, install_dependencies_apt.py. Codacy: 0 new issues, up to standards. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 * chore: leave visual_display_manager untouched so #523 can merge The only change this branch made to that file was a docstring, and it collided with #523's rewrite of the same method -- so #534 and #523 each merged cleanly against main but conflicted with each other. Reverted to main's text; #523 owns this method and adds frame_hold to it. The note the docstring carried ('frame_hold arrives in #523') would have been stale the moment #523 landed anyway. The parity test in #523 is what actually keeps the two signatures honest. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 * chore: add the Ruff suppression nosec/nosemgrep do not cover Ruff reports S603 on the same call Bandit and Opengrep do, and none of the three suppressions covers the others. Confirmed the precondition first: path comes from discover_plugin_tests(), which globs test files inside the repo, and the call is a fixed interpreter with a list argv and no shell. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
Chuck
|
0730d95200 |
fix(sports): let the plugin declare its own directory, don't deduce it (#519)
_plugin_dir() returned None on every device. The consequence was silent and
reached the panel:
_plugin_dir() -> None
_schema_font_size() -> None for every element
-> a configured size equal to the schema default stops looking like a
default and is treated as a deliberate user choice
-> the snap to the font's pixel grid is skipped
-> 4x6-font.ttf renders at 6 instead of 7: 3px-wide glyphs, not 4px
On a 256x64 panel that made the odds, the team records and the date row hard to
read. Both `odds` and `detail` were affected -- anything resolving a
grid-snapped schema default was a pixel narrow.
Why it was invisible here. PluginLoader._namespace_plugin_modules renames every
bare module a plugin brought in (sports, game_renderer, ...) to
"_plg_<plugin_id>_<module>" and REMOVES the bare sys.modules entry, so two
plugins owning a module of the same name cannot collide. A class defined in
sports.py still reports __module__ == "sports", but sys.modules["sports"] is
gone, so walking the MRO for a module with a __file__ finds nothing.
Every test here imported plugins directly, which leaves the bare entry in
place, so the walk succeeded. The safety harness loads plugins its own way and
never reproduced it either. It was found by a user counting pixels on the
panel.
The directory is now declared by the plugin (_PLUGIN_DIR) and only deduced as
a fallback, for hosts that declare nothing -- the plugins' own probe harnesses
build classes with type().
Verified on hardware, which is the only place the original failure appeared:
before, the live service logged plugin_dir=None and 4x6-font.ttf@6 for all six
football managers; after, plugin_dir resolves and both odds and detail are @7.
Five regression tests, including the production shape: a class whose __module__
is absent from sys.modules still resolves via its declared directory, and the
precondition that the MRO walk alone returns None is pinned so the test keeps
meaning something if the fallback changes.
|
||
|
Chuck
|
32d637a446 |
fix(store): read the core version from disk, not from a stale import (#518)
* fix(store): read the core version from disk, not from a stale import
Updating the core to 3.3.0 and then updating plugins refused all eight sports
scoreboards:
Refusing to install nrl-scoreboard: NRL Scoreboard supports LEDMatrix
>=3.3.0, but this system is running 3.2.0.
while src/__init__.py on that machine read 3.3.0. Observed on hardware, not
theorised.
The gate ran `from src import __version__ as core_version`, which binds
whatever the process loaded at start. The plugin store's gate lives in the web
UI, a long-lived service of its own, and the update route deliberately restarts
nothing -- it replaces files on disk and asks the user to restart. Its prompt
named only the *display* service, so a user who followed it left the web
process holding the previous number.
Stale by exactly one release is the case that bites: every plugin flooring on
the release you just installed is refused, blaming a core version that is
already correct on disk. It reads as a broken plugin store. 3.3.0 is the first
release where this hits a whole family at once, since all eight scoreboards
floor there.
compatibility.current_core_version() reads the version from the file instead,
falling back to the imported value on any failure -- so it can only ever be as
correct as before, never worse. All four gate call sites use it: three in
store_manager (install, the git-pull update path, install_from_url) and one in
plugin_loader's advisory warning.
The restart prompt now names both services.
Twelve tests, including the hardware failure itself: a process holding 3.2.0
while disk says 3.3.0 refuses hockey, and reading fresh allows it. The inverse
is asserted too -- a genuinely old core still refuses, so the gate has not
become permissive. One test greps both modules for the old import-bound read;
reintroducing that line fails it, which is what stops this coming back.
Not changed: web_interface/__init__.py also imports __version__, but for
display rather than gating, and the API endpoint already reports a fresh
git describe.
* fix: drop the unused os import
Left over from a first draft that joined paths by hand before this used
pathlib. Flagged by CodeRabbit on #518; confirmed dead -- no os. reference
remains in the module.
|
||
|
ChuckandClaude Opus 5
|
bc2dbf3824 |
feat(sports): share the sports.py surface that is identical in all eight scoreboards (#515)
* feat(sports): share the sports.py surface that is identical in all eight Nine scoreboards ship their own sports.py -- 41,326 lines. Comparing executable ASTs across the eight that share a lineage, 48 method bodies are byte-identical in every one: 1,007 lines carried eight times, so 8,056 lines that must be edited eight times to fix once. They are the parts with no sport in them: the selection and rotation engine (_round_robin_favorites, _favorites_first, _compose_selection, _check_ranking_coverage, _game_divisions, _normalise_quality), the font/colour/ date subsystem (_scale_headline_fonts, _scorebug_font, _resolve_font_size, _format_game_date, _font_color), and the switch-mode upcoming card (_draw_upcoming_center_switch). Nothing here knows what an inning is. Mixins rather than free functions: every one of these reads host state, so rewriting 48 bodies into free functions would be a rewrite rather than a move, and it is the move that keeps the renders identical. Three of the 48 are deliberately left in the plugins, because a byte-identical body is not automatically safe to move: - _get_timezone calls resolve_timezone, imported from a per-plugin module (hockey_timezone, soccer_timezone, ...). All eight of those differ -- each carries its own _WRITEBACK_FIXED_IN -- so hoisting the caller would silently bind every scoreboard to one plugin's copy. - _extract_game_details and _fetch_data are @abstractmethod stubs. They are the sport contract; satisfying them from a mixin would let a plugin instantiate without implementing its own sport. _resolve_font_path went the other way: a module-level function, identical in all eight, that _scale_headline_fonts needs -- so it is inlined here. _schema_font_size needed a real change rather than a move. It located the plugin's config_schema.json with __file__, which here is src/common/, so the load failed silently, the cache stayed empty and every element fell back to an unsnapped size -- measured at 81% anti-aliased edges on a panel that should be pixel-crisp. It now recovers the plugin directory from the instance. Note that type(self).__module__ alone is not enough: SportsCore is an ABC, so a subclass built with type(name, bases, ns) -- which the plugins' own tests do -- reports its module as "abc". _plugin_dir walks the MRO past those synthetic classes to the first module sitting beside a config_schema.json. Worth recording: the 176 harness renders did NOT catch that regression. The plugin's own test_fonts_are_crisp.py did. Renders alone were not a sufficient gate here. Not merged with src/common/sports_card.py despite fourteen same-named twins. Only five are provably equivalent by source comparison; the other nine differ in ways inspection cannot settle, and a wrong guess silently changes what every scoreboard draws. That merge needs differential testing and is its own change. * fix(sports): declare the constants the mixins read, and test the contract CodeRabbit found _QUALITY_CHOICES and _RANKING_COVERAGE_SECONDS read by _normalise_quality and _check_ranking_coverage but never defined on a mixin. Confirmed: both are declared by all eight scoreboards, so nothing fails today -- it would only have bitten the ninth plugin to adopt this, at runtime, mid-render. Both are identical everywhere, so they get defaults here; each plugin's own copy still shadows them. Auditing for others showed those two were the only ones, but also that the host-contract docstring was substantially incomplete: it listed 21 attributes where the mixins actually read about 40, and omitted five hooks (_is_favorite_game, _is_game_really_over, _is_ranked_game, _passes_other_filters, _get_timezone). The section is now derived from that audit rather than remembered. test_sports_shared.py covers what is genuinely new, not the moved bodies: - The contract itself. It parses the module for every ALL-CAPS `self.X` the mixins read and asserts each is defined, so the next omission fails here rather than in the field. - _plugin_dir, the only new logic in the move. Including the case that made it necessary: SportsCore is an ABC, so a subclass built with type(name, bases, ns) -- which the plugins' own tests build -- reports __module__ as "abc". The test asserts that precondition before asserting the walk steps past it. - The three SportsLive bodies. Hockey and lacrosse disable live mode in their harness fixtures, so the 176 renders never reach this path; testing the mixin directly means coverage no longer depends on which plugin happens to have a unit test. Two of those tests pin things that would otherwise be silently undone. SportsRecentSharedMixin does carry an __init__ -- SportsRecent.__init__ was one of the 48 byte-identical bodies. Its bare super() binds to where it is defined, now the mixin, so it only reaches the host because the mixin is listed first in the bases. One test proves the chain runs; the next proves that reversing the order silently skips the host constructor. * fix(sports): drop three unused imports and let the matcher narrow Codacy flagged five issues on this file. Three are unused imports: math, abc.abstractmethod and zoneinfo.ZoneInfo. Nothing in the module references any of them -- the timezone work goes through pytz, and the @abstractmethod mention in the module docstring describes the two stubs that deliberately stayed behind in each plugin, not anything declared here. pyflakes agrees; all three are removed. The other two are "team_in is not callable" on the round-robin favourite matcher. That call is already guarded by callable(), so it cannot raise at runtime, but callable() is not a narrowing construct a static analyser follows: the name still carries the None from getattr's default. Normalising a non-callable to None and branching on `is None` gives the analyser a test it does understand, and keeps the guard. Behaviour is unchanged. _round_robin_favorites has no test coverage, so I exercised it directly on both paths -- a host with _team_in (id matching, the NRL case) and one without (abbreviation matching) -- across limits 1 to 4, and the selections are identical before and after. A host whose _team_in is present but not callable still falls back to abbreviation matching rather than raising. test/test_sports_shared.py: 27 passed. The 9 collection errors under `pytest test/ -k sport` reproduce identically on the unmodified branch and are not from this change. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |
||
|
Chuck
|
300cdaa250 |
feat(sports): share the scroll-card geometry the scoreboards all duplicate (#514)
* feat(sports): share the scroll-card geometry the scoreboards all duplicate The card helpers moved to src/common/sports_card.py, which shared the eight scoreboards' settings lookups. Their *geometry* stayed duplicated: nine methods deciding how wide the centre strip is, how much room each logo gets, and where an upcoming card's date and time land. Five were byte-identical in all eight plugins; the other four were identical in seven, each with a different single outlier. That shape is why this is a mixin and not free functions. Comparing executable ASTs against the eight plugins, 67 of the 70 method bodies are inherited unchanged and 3 become ordinary overrides -- baseball keeps its own _logo_slot_width and _draw_upcoming_game_status, hockey its own _upcoming_date_and_time. No per-sport branching goes inside the base. It deliberately has no __init__ and no state. The plugins' constructors differ six ways and none of it is worth unifying, so adoption is one line on the class statement plus deleting what now comes from here. Placed in src/common/ rather than src/base_classes/sports/ on purpose: importing that package pulls core.py -> DisplayManager -> rgbmatrix, and this is pure geometry that must not drag a hardware import into every plugin that uses it. It sits next to sports_card.py, which the same plugins already use. Only _SCORE_PROBE varies between plugins, so leagues that reach three digits a side override that one ClassVar; the four gap constants are identical everywhere. The tests drive the mixin through a host that provides exactly the surface the module docstring names and nothing else, so the mixin growing a new self.* dependency the plugins do not have fails the contract test rather than shipping. * fix(sports): reject non-finite card settings before they abort the render A center_gap of inf passes `isinstance(x, (int, float)) and x >= 0` unharmed and then raises OverflowError out of int(). The surrounding guards caught only (TypeError, ValueError), so it escaped and took the whole card render with it. The same holds for center_gap_ratio, the two clamp bounds, and layout offsets, where "inf" arrives as a string and float() is happy to produce it. Four of the five paths crashed; only a NaN ratio happened to survive, by accident of min/max rather than by design. This is pre-existing behaviour -- the bodies moved here verbatim from the eight plugins and every one of them has it today. Fixing it in the mixin fixes it in all eight at once, which is the argument for the mixin. Guarded with math.isfinite() before any int()/round(), falling back to the same defaults the finite paths already use, plus OverflowError added to the except clauses as a backstop. Ordinary settings are untouched: all 192 scroll-card renders (8 plugins x 8 panel sizes x 3 game types) stay byte-identical to pristine main. Found by CodeRabbit on #514 and confirmed by running it before fixing. |
||
|
ChuckandClaude Opus 5
|
10da2f97fd |
feat(sports): share the card helpers the eight scoreboards each carried (#513)
Twenty methods were byte-identical in all eight scoreboards' game_renderer.py: the colour pickers, the scroll_card settings lookup, the date and time formatting, the favourite-team rules and the font-size grid snapping. Every fix to any of it had to be made eight times, and a new scoreboard began by copying them a ninth. src/common/sports_card.py holds them once. 245 lines leave each plugin. **Free functions, not a base class.** Every helper takes config/logger/fonts as arguments rather than reading them off an instance, so a plugin keeps its method and delegates the body -- call sites, signatures and override points are all untouched. Adoption is therefore per-function and reversible, which is what let all eight move with byte-identical renders. The bodies are the plugins' code moved, not rewritten. Two deliberate differences, both verified: - crisp_size takes the seven-plugin guard (`not desired`) rather than football's. They agree on every real input; the extra guard only stops a None size raising TypeError, so adopting it is a no-op for seven plugins and removes a crash path for the eighth. - schema_font_size caches per schema PATH. The plugins cached on their own class, which is the same distinction expressed without a class to hang it on; two plugins never share an entry. The path has to be passed in because the plugins derived it from __file__, and __file__ here is the core's. Verified before any plugin was touched: 534 differential comparisons of the helpers against afl's originals and 704 more of the font-sizing chain against all eight plugins' originals -- 1,238 comparisons, zero differences. Writing the constant tables by hand introduced two errors that check caught: the tie colour was (255,255,0) instead of the plugins' (255,200,0), and a "five_by_seven" alias that does not exist. Both are now taken from the plugins verbatim. 43 tests pin the contract, including the cases the plugins' own comments record as having bitten: a three-character string must not iterate into a colour, a shared font face must give up rather than guess an element, a font_size equal to the schema default carries no intent, and a bad timezone falls back to UTC rather than blanking the card. Full suite 3768 passed, 6 skipped. Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
92f9d06af9 |
fix(logos): stop a failed download pinning a team to a grey box forever (#512)
* fix(logos): stop a failed download pinning a team to a grey box forever When a logo download fails, create_placeholder_logo writes a 64x64 grey PNG under the *real* logo's filename. Every later call then hits `if filepath.exists(): return True` and reports success, so the real logo is never attempted again. One transient failure -- no network at boot, ESPN blipping -- permanently costs that team its logo. This is not hypothetical. Five of the eleven cached AFL logos in my checkout were 384-byte stubs written in a single bad minute, and they had stayed that way ever since; the scoreboard rendered COLL, FRE, NMFC, PORT and SYD as grey text boxes on every card. Placeholders are now stamped with a `ledmatrix_placeholder` PNG text chunk carrying their creation time, and `is_placeholder_logo` recognises them. It also matches on the placeholder's exact geometry and background colour, so the stubs already sitting on users' disks are picked up too -- without that, this fix would only help teams whose logos break in future. Verified against the real stubs: all five detected, all six real logos untouched. `download_missing_logo` now treats an existing placeholder as the failed download it is and retries, rather than as a satisfied request. The retry is rate-limited to PLACEHOLDER_RETRY_SECONDS (6h) so this does not trade a permanent grey box for an ESPN request every frame; a failed retry rewrites the placeholder, restarting the clock. The age comes from the stamp rather than mtime, so a backup restore, an rsync, or a permissions script cannot silently reset it. `download_missing_logos_for_league` gets the same treatment -- a bulk pass is exactly where a previously failed logo should get another chance -- and `LogoHelper.load_logo_with_download` no longer accepts a stale placeholder as a cache hit. That import is lazy and guarded so the module still works against a core build predating the marker. `LogoHelper._create_placeholder_logo` needs no change: it returns an in-memory image and never writes it to disk, which is the behaviour this bug argues for. Tests cover marked and legacy-unmarked detection, the two false-positive cases (a real 500x500 logo, and a 64x64 image that is merely the same size), the retry, the rate limit, and that the age survives an mtime touch. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(logos): address review — unify eligibility, invalidate cache, restart back-off Three findings from the review on #512, all confirmed against the code: 1. The three download sites each had their own idea of "already have it". download_missing_logos_for_league() retried *any* placeholder, ignoring the back-off entirely, while download_all_ncaa_football_logos() was never updated and still skipped placeholders forever. They now share one should_attempt_download(), which also covers force_download, so the sites cannot drift apart again. download_missing_logo() reads through the same helper. 2. LogoHelper.load_logo_with_download() answered from the in-memory cache before touching the disk, so after a stale placeholder was successfully replaced the *cached placeholder image* was still returned -- the real logo would not have appeared until the process restarted. The cache entry for that file (every size of it) is now dropped after a successful download. 3. A failed retry left the stale placeholder on disk with its old timestamp, so the next call saw it as stale again and retried immediately: a download attempt per call, which is precisely what the back-off exists to prevent. refresh_placeholder_timestamp() restamps it, and the helper calls that on the failure path. It refuses to touch anything that is not a placeholder. Tests cover both bulk loops in both directions (fresh placeholder skipped, stale one retried), the eligibility rule including force_download, the timestamp refresh, and the two LogoHelper paths -- including that a freshly-downloaded logo is actually what comes back rather than the cached placeholder. Two of the new bulk-loop tests initially passed for the wrong reason: the fetch_teams_data stub returned {}, which is falsy, so the loops bailed before reaching the eligibility check at all. Fixed to return a truthy payload. Re-verified end to end: with both halves in place, rendering the AFL scoreboard took FRE.png from a 362-byte stub to a 12,928-byte logo. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
a686932c7e |
fix(store): land the install_from_url gate, which never reached main (#511)
#510 shows as merged, but into fix/gate-git-pull-updates -- #508's branch -- rather than main. #508 reached main first, so the sideload gate was left behind on a branch. Same failure as plugins #350/#351, which merged into each other's bases; worth knowing the pattern, because GitHub reports these as MERGED and `gh pr list` shows nothing outstanding. main today has two of the three routes gated: install_plugin (#431/#433) and update_plugin's git branch (#508). install_from_url validates required manifest fields and then installs whatever it found, never comparing the core version. Cherry-picked unchanged from the orphaned branch -- it applies to main with no conflict. TestSideloadGate pins the three cases the other routes pin: refuses a floor above this core leaving nothing behind, still allows a compatible plugin (the guard against a gate that refuses everything), and does not block a 2.0.0 floor on a core reporting an untrustworthy version. Full suite 3725 passed, 6 skipped. Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
154525beb8 |
fix(background): release the payload after every callback, not after each one (#509)
Callers that join an in-flight fetch share one FetchResult. #499 released the payload inside the delivery loop, so the first callback got the data and every joiner got `result.data is None`. That is not a quiet degradation. Consumers read `result.data.get('events')`, so they raise AttributeError -- which the delivery loop catches and logs. The entire failure surfaced as one line: ERROR - src.background_data_service - Error in callback for request nhl_2026_...: 'NoneType' object has no attribute 'get' and a manager that silently never received its schedule. Seen on hardware: NHLRecentManager logs "Background fetch completed for 2026: 1000 events" and the very next line is the error, from NHLUpcomingManager's callback on the same request -- which had already logged "No events found in shared data." Deduplication is the normal case, not a corner. A sport's recent, upcoming and live managers all want the same season schedule, so the second and third are joiners on almost every cycle. _release_payload's own docstring said "once A callback has been handed the data", singular, which is the assumption that broke: the loop above it was written for many, and says so. Moved after the loop, and guarded on `callbacks` being non-empty. The guard matters: a request submitted without a callback must keep its payload, because polling get_result() is then the only way to collect it. The per-delivery release got that right by accident -- an empty list never entered the loop body -- and the existing test for it caught the omission. test_background_payload_release.py gains TestJoinersAllGetTheData: two submitters on one in-flight cache_key, asserting both are handed a populated payload, plus that the memory fix still happens once they have all had it. test_background_fetch_dedupe.py already proved the joiner's callback FIRES; it never checked what the callback received, which is the gap that let this through. Verified the new test bites: restoring the release inside the loop fails it with "'second' was handed a released payload". Full suite 3716 passed, 6 skipped. Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
9b522d412c |
fix(store): gate the git-pull update path (#508)
* fix(store): gate the git-pull update path `install_plugin` gates every route that re-downloads, `_reinstall_with_rollback` included. `update_plugin` has one branch that re-downloads nothing: a git checkout pulls in place, installs dependencies, and returns True. A pull could therefore deliver a manifest flooring above this core and nothing would notice until the plugin failed to load — which surfaces as one line in the journal and a display that silently stopped appearing. Checked after the pull rather than before it, for the same reason `_install_plugin_impl` checks after the download: the registry carries no compatibility field, so the incoming floor is only knowable once the new commit is on disk. Undone with `git reset --hard` to the pre-pull commit rather than by removing the directory. This is a live checkout, the old commit is still in the object store, and the reset leaves the user on the exact version they were already running — the same promise `_reinstall_with_rollback` makes, reached by the means this path actually has, with no window where the plugin directory does not exist. An unreadable manifest allows: it is not evidence of a floor. Scope, stated plainly: monorepo plugins install as archives and update through `_reinstall_with_rollback`, so they were already gated. Only registry entries with no `plugin_path` reach this branch. It is closed anyway because the sunset rule in the plugins repo's `08-shared-sports-code.md` names, as condition 3, that the core enforces the floor "at install/update time" — and B6 rests on that being true rather than merely written down. `install_from_url` is still ungated; the tests say so rather than letting the next reader assume otherwise. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 * fix(store): do not pull what the gate cannot un-pull Review of the gate found a data-loss path it had introduced, plus two smaller scope errors. All three from CodeRabbit on #508. **The stash failure was load-bearing and was not treated as one.** update_plugin stashes local changes before pulling; when that stash failed or timed out it logged a warning and pulled anyway. That was harmless while nothing ever undid a pull. It is not harmless now: the gate's rollback is `git reset --hard`, which discards uncommitted tracked edits -- exactly the edits the stash existed to protect. A pull does not refuse on a dirty tree as long as the incoming commit touches other files, so the sequence completed silently: pull succeeds, gate refuses, reset takes the user's work with it. update_plugin now returns before pulling unless the tree was already clean or was successfully stashed. Refusing costs an update in a case that had already gone wrong; the alternative costs data. That also makes `--hard` safe by construction in _gate_pulled_commit, and its comment now says so rather than observing it in passing. Pinned by test_a_failed_stash_stops_the_update_before_pulling, which writes a local edit, forces the stash to fail, and asserts both that HEAD did not move and that the edit is still on disk. Verified it bites: with the new guard removed the file comes back as `class P: pass`, the edit gone. **_HAS_GIT could take the module down instead of skipping it.** With no git on PATH, subprocess.run raises FileNotFoundError, and this runs at import time -- before skipif can act, so the whole file errors rather than skipping. Now catches OSError. **The doc overclaimed the gate's reach.** It said the floor is enforced on "every route that installs or updates" while the same passage notes install_from_url is ungated. Both spots now scope the claim to registry-managed installs and the two supported update paths, and name the sideload exception. Full suite 3720 passed, 6 skipped. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014RRtqXDCnvnY6EQwhT5CV9 --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
eae063700f |
test(sports): pin the four-case sunset matrix before B6 runs (#505)
B5 and B6 make different promises about an adopted scoreboard, and only the
second is obvious. The table:
| bundled copy present | bundled copy removed
pinned old core | loads (the fallback) | ERROR, naming the module
current core | loads, using core code | loads, using core code
The top-left cell is the one worth having. Nothing in the suite proved that an
adopted plugin still runs on a core predating src.common.sports_scroll, and
that claim is the entire basis for having shipped B5 ahead of B6's gate.
The bottom-left cell is asserted through PluginManager.load_plugin rather than
a bare import, deliberately. The manager catches the ModuleNotFoundError, so a
test written around pytest.raises would pass against a core where the module is
merely broken rather than absent, and would say nothing about what the user
meets: a plugin parked in ERROR and one log line. The assertion is the ERROR
state plus an error naming the module, which is what makes the log actionable.
Modelling notes, both of which were wrong first time and matter:
- The copy-removed shape is an UNGUARDED import, not the guarded one with the
legacy file deleted. Keeping the guard while removing its fallback only
mislabels the failure -- the plugin reports a missing scroll_display_legacy
and never mentions the core module that is actually absent.
- Only the leaf module is hidden. A pre-3.2.0 core still ships src/common/;
hiding the package would be a harsher core than any that shipped, and would
make the failure name the package instead of the module.
Ablation: disabling the old-core simulation fails three cells, and dropping the
error object from load_plugin's set_state fails the fourth, so none of it
passes vacuously. The install gate -- the other half of the guarantee -- stays
in test_plugin_compatibility_gate.py rather than being duplicated here.
Refs docs/SPORTS_UNIFICATION.md, "B6 -- why the sunset needs more than a
version floor".
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
||
|
ChuckandClaude Opus 5
|
af96bd5cb0 |
fix(memory): join an in-flight fetch instead of starting a duplicate (#503)
* fix(memory): join an in-flight fetch instead of starting a duplicate
submit_fetch_request() had no notion of "already fetching this". request_id
embeds a millisecond timestamp, so every submit looked new, and
active_requests is keyed by that id rather than by what is being fetched.
Two submits for the same cache_key therefore started two identical fetches.
It is not a rare race. _fetch_data in the sports managers branches: the Live
manager fetches only today's games, but Recent and Upcoming both pull the
full season schedule under the SAME cache_key. On a cache miss both miss,
both submit, and nothing stops the second. On a running 512x64 board:
138 background fetches in 24 hours, arriving in pairs at identical
millisecond timestamps, roughly hourly:
2 2026-08-25 11:47:26.612
2 2026-08-25 10:46:28.064
2 2026-08-25 08:01:55.962
Half of them redundant. Each duplicate costs a second download, a second
JSON parse -- the expensive part on a Pi -- and a second parsed copy
resident at the same time. Schedules on that board run 256KB to 20MB, 106MB
across all sports. Because the pairs land in the same millisecond they also
occupy two of the three executor slots with identical work, which is what
makes two large parses peak simultaneously.
A submit for a cache_key already in flight now joins that request: its
callback is added to the existing one and the existing request_id is
returned, so get_result() works for both. Different keys are untouched, and
dedupe applies only while a fetch is in flight -- a submit after completion
fetches again, because this is not a second cache layer.
Three details:
- The in-flight entry is dropped and the callback list snapshotted in the
SAME critical section as filing the result. Otherwise a submitter could
join a fetch whose callbacks had already run and never be called back.
- Cancellation is the other way a request leaves active_requests, so it
releases the key too. And the join path looks the request up rather than
trusting the id, so an entry stranded any other way cannot wedge a key
permanently -- it is dropped and a fresh fetch starts.
- One callback raising no longer prevents the others being delivered.
Previously there was only ever one.
Interaction with #499, whichever merges second: that PR releases the payload
after the callback runs. With several callbacks the release must happen
after ALL of them, and must not happen at all if a joined submitter passed
no callback, since polling get_result() would then be its only delivery
path. The callback list built here is the hook for that.
test_background_fetch_dedupe.py -- 8 tests, covering the join, callback
delivery to both submitters, one callback raising, distinct keys not being
coalesced, a post-completion submit fetching again, cancellation releasing
the key, a stranded entry not wedging one, and the reported count. Verified
non-vacuous by removing only the join branch: 3 fail. The callback tests
assert the ids coalesced, without which they would pass trivially on two
independent requests.
Full suite: 3698 passed, 60 skipped, 1 failure that reproduces identically
on unmodified main (test_install_lowmem, environment-dependent: /var/tmp is
disk-backed on this machine).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
* fix(memory): discard a cancelled fetch instead of letting it commit
Review follow-up on the dedupe.
Cancelling releases the cache_key, so a replacement fetch for that key can
start immediately. But _fetch_data_worker() had no cancellation check: the
cancelled worker still wrote its response to the cache, flipped its own
status from CANCELLED to COMPLETED, and ran its callbacks. The stale
response could therefore land on top of the replacement's fresher data.
The worker cannot abort an HTTP call in flight, so the response is discarded
on return instead: no cache write, no callbacks, status left CANCELLED. The
check sits immediately before the cache write, which is the first
side effect.
Also fixed, found by the new test rather than by reading:
request_id was f"{sport}_{year}_{milliseconds}", which is not unique.
Two submits inside the same millisecond produced the SAME id -- the
test's two sequential fetches collided on a fast mocked response, and
one request silently replaced the other in active_requests and
completed_requests. Rare before this PR; load-bearing now, because
dedupe hands that id back to every joiner as their handle for
get_result(). A per-service counter is appended.
Two test problems of my own, both fixed here rather than left to flake:
- The cancellation test synchronised with time.sleep(0.4). A slow worker
would have made it pass for the wrong reason. It now waits for the
request to be filed in completed_requests.
- The id-uniqueness test patched session.get, but submits are async: the
50 workers outlived the patch and made real DNS calls to the dummy host.
It stubs the executor instead, which is what a submit-time test should
exercise.
20 consecutive runs of the dedupe file: 0 failures. Full suite: 3700 passed,
60 skipped, 1 failure that reproduces identically on unmodified main
(test_install_lowmem, environment-dependent).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
* fix(memory): make cancellation terminal, not advisory
Three paths wrote request.status without checking whether the request had
already been cancelled, so a cancel could be silently undone and the work it
was meant to stop went ahead anyway.
- A request cancelled while queued had CANCELLED overwritten with IN_PROGRESS
the moment its worker started, defeating the discard check entirely: it
downloaded, cached and called back for work the caller had withdrawn. It
now skips the fetch outright, which is also the cheapest possible cancel.
- The cancelled-check and the cache write were separate critical sections, so
a cancel landing between them left the payload in the cache with the
callbacks suppressed -- every submitter that joined the fetch waited for a
call that never came. The worker now claims the commit in the same critical
section that reads the status, and cancel_request refuses once claimed. The
write stays outside the lock: it serialises a multi-megabyte payload to the
SD card, and holding the service lock across that would stall every submit,
status query and cancel behind it.
- A cancelled request that then failed was relabelled FAILED, which slipped
past the CANCELLED-only callback gate and delivered a spurious error
callback. The except path now leaves CANCELLED alone.
get_request_status() also reported a cancelled request as FAILED, since it
inferred status from result.success; the final status is now recorded on the
result. Both early returns assign to `result` so completed_requests files the
outcome that was reported rather than the untouched placeholder.
Tests cover cancellation before worker start, during the commit, and during an
HTTP failure; all three fail against the unfixed code.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
* test(memory): reach the exception path as a cancelled request
test_a_failure_after_cancelling_stays_cancelled cancelled the request while
its worker was still queued, so once the pre-start branch landed the worker
returned there and never reached the exception handler the test is named for.
It passed against the unfixed code only because that branch did not exist yet;
with it, the test passed for the wrong reason and reverting the except-path
guard did not fail it.
Cancel while the worker is parked inside the HTTP call instead, and assert the
fetch actually started so the test cannot silently degrade into the pre-start
case again. Reverting each of the three guards now fails exactly one test.
Also read the payload inside the callback rather than off the FetchResult
afterwards: #499 releases result.data once delivery is done, so the later read
saw the released object and not what the caller was handed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
||
|
ChuckandClaude Opus 5
|
bdced206dc |
fix(plugins): retain state history by age, with the count as a ceiling (#502)
* fix(plugins): retain state history by age, with the count as a ceiling
Follow-up to the cap in this PR. A flat entry count answers the wrong
question: what a reader wants from this history is "the last couple of
hours", and how many transitions that is depends entirely on the plugin's
update interval. On a real board those span 2s to 3600s, so 200 entries is
interval 200 entries covers
2s 3.3 minutes (flights, live)
10s 16.7 minutes (jellyfin)
60s 1.7 hours (default)
300s 8.3 hours (news)
3600s 4.2 days
-- the plugin churning hardest, the one worth looking at, keeps the least.
So transitions are now trimmed by AGE first
(STATE_HISTORY_MAX_AGE_SECONDS, two hours), which makes the retained window
comparable whatever the cadence, and the count cap becomes purely a memory
ceiling for pollers fast enough to exceed it inside that window. The
ceiling rises 200 -> 2000: at ~230 bytes an entry that is ~0.5MB per plugin
worst case, and only plugins updating faster than roughly every 4s can
reach it. Steady-state memory is unchanged for everything slower, since the
age trim binds first.
Two details worth stating:
- The trim reads time.monotonic(), stored alongside each transition,
rather than the datetime already inside it. A DST shift or an NTP step
would otherwise make every entry look ancient and flush the history in
one go. The human-readable timestamp is untouched and still what
get_state_history() returns.
- Trimming happens on append, so a plugin that goes quiet keeps its last
window until it writes again. That is deliberate: it is bounded either
way, and a lazy trim costs nothing on the hot scheduling path. The
guarantee is therefore about the SPAN of retained history, not its age
against the current clock, and the test asserts it that way.
The public shape is unchanged: get_state_history() still returns the same
list of transition dicts, and state_history_count is still the lifetime
total.
test_plugin_state_history_retention.py adds 7 tests. Verified against this
branch with only the age trim removed: 4 fail, 3 pass -- the three that
survive are testing the count ceiling and the monotonic clock, which this
commit does not change. Full suite 3753 passed, 60 skipped.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
* fix(plugins): build get_state_info() as one locked snapshot
Every field was read under its own lock, so an unload running concurrently
could be observed half-done: 'state' read before clear_state() removed it
and 'state_history_count' read after, handing PluginManager.get_plugin_info()
a plugin that is ENABLED with zero transitions.
The whole payload is now built in one critical section. _lock is an RLock,
so the helpers called inside it can still take it.
The regression test runs a reader against a thread that repeatedly fills and
clears the same plugin, and fails on the first torn snapshot. Verified by
removing only the lock: fails on 3 of 3 runs, passes on 3 of 3 with it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
||
|
Ron PierceandClaude Opus 5
|
39e7f8cbe0 |
fix(plugins): cap the per-plugin state transition history (#501)
* fix(plugins): cap the per-plugin state transition history
PluginStateManager recorded every state transition in a per-plugin list
and never trimmed it. The only code that removed entries was
clear_state(), called solely from PluginManager.unload_plugin(), so a
plugin that stays loaded -- normal operation -- never released one.
The list is written on the hot scheduling path. Every update cycle
appends twice: _reserve_for_update() sets RUNNING and _finish() sets
ENABLED back again. At the default 60s update interval that is 2,880
entries per plugin per day, and nothing reads them -- get_state_info()
only takes their len(). Pure dead weight.
Measured against the unpatched class, ten plugins on a 60s interval:
sim uptime history entries heap growth
1 day 28,810 7.7 MB
7 days 201,610 53.9 MB
30 days 864,010 230.9 MB (still climbing)
With the cap it is flat at 2,000 entries / 0.5 MB from day one.
On a 1 GB board 231 MB of garbage is fatal on its own, and the failure
is not a clean OOM: once MemAvailable falls far enough fork() starts
returning ENOMEM, so sshd accepts connections and closes them before its
banner while the kernel still answers pings. The board looks like a
hardware fault and needs a power cycle. Same family as the ceilings
added in #464.
Retain the most recent 200 transitions per plugin in a deque and let the
rest age out. state_history_count is surfaced through the web API, so
the lifetime total is tracked separately rather than plateauing at the
cap. get_state_history() now returns a copy under the lock; it was
handing out the manager's own list, which a caller could mutate.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(plugins): copy history entries out, lock clear_state
Review follow-ups on the transition history.
get_state_history() copied only the outer list, so a caller holding a
returned transition could rewrite the manager's record of what happened
-- which contradicted the defensive-copy guarantee in its own docstring.
Copy each entry too. Every value in a transition is immutable, so a
shallow copy per entry is enough. test_get_state_history_entries_are_copies
pins it; without the change it fails with 'tampered' == 'enabled'.
clear_state() mutated five shared dicts without holding _lock, while
every other mutator takes it. A concurrent set_state() could interleave
and leave a plugin with history but no state. Drop the five as one unit.
This does not close the wider unload-vs-worker race, which lives in
PluginManager.unload_plugin() and predates this change: an update worker
still in flight can call set_state() after clear_state() returns and
recreate the entry. Serialising that needs the per-plugin lock held
across worker join in unload_plugin(), which is a separate change.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
|
||
|
ChuckandClaude Opus 5
|
f90638a9ec |
feat(web): show available memory in Tools diagnostics (#500)
* feat(web): show available memory in Tools diagnostics System Diagnostics reported memory as used-percent plus used/total GB. Neither distinguishes a healthy board from one about to fail, because page cache counts as used and is reclaimable on demand -- a Pi can read 70% used and be fine, or read the same and be minutes from trouble. MemAvailable is the kernel's own estimate of what a new allocation can actually obtain, and it is the number that tracked the failure on a 1GB Pi 3B+: healthy running sat above 500MB, and the crash came at 73MB. By that point fork() was failing, so sshd could not spawn a session and systemd could not respawn the display, while the kernel carried on answering pings at 0% loss. Used-percent gave no warning at any point on the way there; available memory fell steadily for hours. /api/v3/system/status now returns memory_available_mb from psutil.virtual_memory().available, and Tools renders it as its own tile, coloured against the thresholds that failure implies: red under 150MB, amber under 300MB, green above. The existing memory tile is left alone -- used/total is still what you want when sizing a workload; this answers the different question of how much room is left right now. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(web): round available memory once, so the tile agrees with itself The colour was classified from the raw value while the label was rounded, and the API sends one decimal place. At the boundaries the two disagreed: 149.6 rendered as "150 MB" in red, and 299.6 as "300 MB" in amber -- each contradicting the threshold its own colour claims to apply ("red under 150MB"). A reader checking the tile against the documented thresholds would conclude the readout was broken. Rounding once and using that number for both restores agreement. It moves those two boundary cases up a band, which does not matter: the thresholds come from a measured failure at 73MB, so which side of the line a spare 0.4MB falls on carries no information. The tile agreeing with itself does. Null handling and the thresholds themselves are unchanged. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
333fd17d28 |
fix(memory): release fetched payloads once they have been delivered (#499)
* fix(memory): release fetched payloads once they have been delivered BackgroundDataService kept the fetched body on the FetchResult it filed in completed_requests, which is swept hourly and capped at 500 entries by count. For a status record that costs nothing; for a season schedule it costs a tenth of the board. Measured on a 1GB Pi 3B+ with a 1-second RSS profile: the display process sat at 404MB after plugin load, then stepped +21MB when NFL fetched its season and +90MB when NCAA football fetched 946 games for 2026 -- and stayed at 494MB. Not a leak; a staircase that never came down. When a later fetch landed while headroom was low, available memory reached ~70MB, fork() began failing, and the board stopped being able to start a process at all: sshd accepted connections and closed them before its banner, systemd could not respawn the display, and the panel went dark while the kernel carried on answering pings. The cache-hit path was the worse of the two. It runs once per update interval per sport, mints a fresh request_id each time, and files whatever the cache returned. The memory tier is capped at 150 entries on a 1GB board, so a miss re-parses the payload from disk into a genuinely new object -- separate copies accumulating toward the 500-entry cap, not shared references. Releasing is safe: the payload is written to the cache under the request's cache_key before the result is built, the callback is handed the object directly, and consumers read it back from the cache afterwards (the plugins' callbacks use it only in passing, to log a count, before reading the cache). Nothing is lost -- it moves from RAM to the disk cache that was already holding it. Requests submitted without a callback keep their payload, since polling get_result() is then the only way to collect it. That keeps the existing contract, and the existing tests covering it, intact. Not addressed here: max_workers=3 allows three concurrent fetches, so three large parses can peak at once, and there is no in-flight dedupe by cache_key -- a second submit for a key already being fetched starts a second fetch. Both bound the transient peak rather than what stays resident, and both are behaviour changes worth their own review. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(memory): file the cache-hit result before running its callback Restores the original ordering. Releasing the payload after the callback meant filing the result after it too, so a callback that queried get_result() or is_request_complete() for its own request would not have found it -- a behaviour change unrelated to the memory fix. The dict holds a reference to the same object, so releasing after filing still clears the payload. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * test: wait for the payload release, not just the filing The callback test waited on is_request_complete(), which goes true as soon as the worker files the result in completed_requests. The worker then runs the cleanup pass, then the callback, then releases the payload. Both of the test's assertions therefore raced the worker: `seen` is populated by the callback, and `data is None` only after the release that follows it. It passes today because a one-line callback usually finishes inside the 20ms poll interval. Confirmed by making the callback sleep 0.4s: _wait() returns with seen == {} and the payload still resident. _wait_for_release() polls for the released payload instead. Release happens strictly after the callback returns, so a released payload also means the callback has finished and one wait covers both assertions. Verified against the same 0.4s callback. _wait() stays for the other three fetch-path tests, which assert only what is already true when the result is filed -- the success flag, the error, and the cache write that happened during the fetch itself. Its docstring now says so, so the next reader picks the right one. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
a4a55a23fc |
fix(web): reject non-finite JSON numbers instead of raising (#497)
* fix(web): reject non-finite JSON numbers instead of raising
POST /api/v3/config/dim-schedule with {"dim_brightness": Infinity} answered
500. So did /api/v3/errors/clear with max_age_hours, and /api/v3/config/main
with multiplexing or row_address_type.
json.loads accepts Infinity/-Infinity/NaN by default -- they are not valid
JSON, but Python's parser emits them -- and Flask's get_json passes them
straight through. int(float('inf')) raises OverflowError, which is neither
ValueError nor TypeError, so validation blocks that carefully caught those let
it past and Flask turned it into a 500.
The status code was not the real damage. dim-schedule answered with
CONFIG_SAVE_FAILED and suggested "Check file permissions on config directory"
and "Check available disk space" for what was an invalid number. Every one of
these sites already had a correct 400 response written; they just never
reached it.
NaN already returned 400, because int(nan) raises ValueError. That is why this
only ever showed up for the infinities, and why it survived: the obvious test
case passes.
OverflowError is now caught alongside ValueError/TypeError at the 27 sites in
this file whose try block performs a numeric coercion. An AST sweep confirms
no int()/float() of request-derived data is left outside a block that catches
it.
Verified end to end through Flask's test client rather than by reasoning about
the parser: all four routes returned 500 before and 400 after.
Tests: five Infinity cases (which fail against the previous except tuples),
two NaN cases pinned so narrowing the tuple cannot quietly break them, and a
check that ordinary input is not rejected by the widened guard.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
* test(web): assert 400 exactly, and prove valid input is accepted
Both review points were right, and the first is the failure mode this file
exists to catch.
Accepting any 4xx meant a 404 would have passed. Renaming one of these routes
would have left the test green while it tested nothing -- the same "looks like
coverage, points somewhere safe" shape that hid the composer injections. Now
asserts exactly 400.
Both infinity signs are exercised for every route. int() raises OverflowError
either way, but only +Infinity was in the original report, and a guard that
special-cased the sign would have passed a one-sided test.
The valid-input test previously asserted "not a 400", which did not show what
it claimed: the mocked save path fails for any input, so that assertion held
whether or not validation had accepted the value. It now gives load_config a
real dict and stubs _save_config_atomic, so the endpoint reaches its success
response and the test can assert 200 -- which only happens if the value passed
validation.
8 of the 11 checks fail with OverflowError removed from the except tuples.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
||
|
ChuckandClaude Opus 5
|
085fb93a87 |
test(logging): stop the location assertion matching the clock (#496)
test_location_toggle asserted that ":42" -- a bare colon plus the record's
hardcoded lineno -- is absent from a line formatted with include_location=False.
But every formatted line starts with an HH:MM:SS.mmm timestamp, so ":42" also
matches the clock whenever the minute or the second is 42. The test fails for
roughly 3% of runs with nothing wrong:
2026-08-22 08:05:42.274 - INFO - test.logger - hello
^^^ matches ":42"
Assert on the whole "module.funcName:lineno" token the format string actually
emits ('%(module)s.%(funcName)s:%(lineno)d') instead of a fragment of it. That
cannot collide with a timestamp, and it checks the thing the test is named for.
Confirmed by formatting a record stamped 08:42:42 -- both minute and second
colliding: the old assertion fails, the new one passes.
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
||
|
ChuckandClaude Opus 5
|
c321b94085 |
fix(display): retry a plugin that is enabled but failed to load (#495)
* fix(display): retry a plugin that is enabled but failed to load A plugin whose validate_config() returns False is treated as a hard load failure. The API then reports enabled=true, loaded=false, error=null: the plugin is simply absent, with nothing saying why. hockey-scoreboard sat in that state on a live rig for four days. The recovery path existed but could not be reached. _reconcile_enabled_plugins computes to_add = desired - current, and a plugin that failed to load is never in current, so it stays in to_add and would be retried. But the reconcile is queued by _enabled_set_changed(), which compares only top-level `enabled` flags -- and the edit that actually fixes such a plugin (enabling a league, filling in an API key) is nested inside the plugin's own config section. No top-level flag changes, so no reconcile is queued, and the save that should have fixed it does nothing. Only toggling some unrelated plugin -- which does change a top-level flag -- queues the global reconcile that recovers it. Add a second gate: queue a reconcile when a discovered plugin is enabled in config but absent from the running set. It is deliberately narrow rather than "reconcile on any config change". Reconcile calls discover_plugins(), a ~39-manifest filesystem scan, and it runs on the render thread; doing that on every config save would trade this bug for a frame hitch. Gating on plugin_manifests also keeps non-plugin sections that carry their own `enabled` flag (schedule, display) from queueing a reconcile they can never satisfy. In the steady state -- every enabled plugin loaded -- the new check is False and costs nothing. The same valid-but-unconfigured => hard-fail shape still exists in text-display, youtube-stats, birdnet-go, ledmatrix-flights and mqtt-notifications; this makes all of them recoverable without a restart. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW * fix(display): snapshot the plugin mappings under their locks Addresses the review finding on the cross-thread reads. _enabled_plugin_not_running runs on the config-watcher thread and read two mappings the render thread mutates. Catching RuntimeError was not a fix: it turned a torn read into a coin flip between an unnecessary discovery scan and a missed retry, which is the bug this PR exists to remove. Both reads are now snapshots taken under the lock that guards their writes: - plugin_manifests via a new PluginManager.discovered_plugin_ids(), which copies the ids while holding the existing _discovery_lock. Discovery rebuilds that mapping entry by entry, so an unsynchronised reader can see it half-populated. - plugin_display_modes under a new controller lock, taken at the only two sites that mutate it (_register_loaded_plugin / _unregister_plugin). The locks are never nested -- each snapshot is taken and released before the next -- so this cannot deadlock against discovery, which holds _discovery_lock while it rebuilds. No cost on the per-frame path. Both mutation sites run during reconcile, which is rare, and every hot-path read of plugin_display_modes is on the render thread itself, same thread as the writes, so those stay lock-free. Tests: the accessor returns a snapshot rather than a live view, and actually takes the discovery lock (proved from a second thread, since an RLock is reentrant on the owning one) so a later refactor cannot quietly drop it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW * fix(display): consume the reconcile request before serving it Addresses the second review finding: a lost update on _pending_plugin_reconcile. The flag was cleared after a successful reconcile. Reconcile has already read its config by that point, so a config change arriving mid-flight set a flag that the trailing clear then erased -- a request that was never served, and the newest config never reconciled. That is the same "my save did nothing" symptom this PR exists to remove, so leaving it would have undercut the fix. Consume the request before running it instead, and re-arm only on a retryable failure. A change that lands during reconcile now stays set and is picked up on the next pass. The per-frame read stays lock-free. It is a fast path that can only produce a false negative -- the watcher setting the flag just after it is read is seen on the next iteration -- never a false positive that loses a request. The lock is taken only when a reconcile is actually pending or a config change arrives. Extracted _service_pending_reconcile() so the sequence is testable rather than buried in run()'s loop; the review asked for a regression test that invokes the subscriber during reconciliation, which is not reachable otherwise. Tests: 4 new, covering a request racing in mid-reconcile, the quiet success, the retryable-failure re-arm, and not reconciling when nothing is pending. Two of them fail against the previous clear-after-success semantics. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
5a1f121e6b |
Stop array-item secrets being wiped, and logging them (#493)
Three review findings from #485 that I missed when addressing that PR; it has since merged, so they land here. 1. Array-item secrets destroyed by any unrelated save (data loss). remove_empty_secrets recursed into dicts but let a list fall through to the scalar branch and kept it verbatim. Lists merge by *replacement*, so the blanks the masked form posts back went straight over the stored array: stored [{"name":"a","token":"REAL-A"}, {"name":"b","token":"REAL-B"}] posted [{"name":"a","token":""}, {"name":"b","token":""}] merged [{"name":"a","token":""}, {"name":"b","token":""}] -> both credentials gone Same failure as the scalar api_key case fixed earlier, one container deeper. Lists now prune element-wise, and a list with nothing real in it is dropped so the stored one is left alone. Where one entry does change, the new merge_secrets merges by index instead of replacing. Two details the first attempt got wrong, both caught by existing tests: - An emptied dict item must stay {}, not None. ConfigManager's _strip_secrets_recursive treats a secrets list as *parallel* to the regular one ({} = "item i has no secrets"); a None makes it stop looking parallel, and it then drops the whole key from the main config -- silently deleting the items' non-secret fields too. - The incoming list's length wins. The regular config's list is authoritative about how many items exist, so preserving surplus stored entries would let the two fall out of step and make deleting an entry impossible. 2. Submitted credentials written to the journal (security). save_plugin_config logged `Full config: {plugin_config}` at INFO and `Config that failed: {plugin_config}` at ERROR. Both run before separate_secrets, so plugin_config still held the values just typed into the form. Now keys only. Swept the rest of web_interface/ and src/ for the same shape -- these were the only two. 3. Restart banner kept stale wording. showRestartPending() cleared the stored custom text but left the DOM element alone, so a config save could show the previous update's message. The default is read back from the server-rendered copy rather than duplicated in JS, so the template stays the one owner of the string. Verified: 556 passed, 1 skipped across the web suite. Mutation-checked -- reverting api_v3 fails the logging guard and the array-merge test; reverting either half of the secret_helpers change fails the unit tests. New end-to-end coverage drives the real endpoint, not just the helpers. Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
6138a3cbef |
test(install): stop assuming pytest's tmp_path is on disk (#492)
test_returns_nothing_when_tmpdir_is_already_disk_backed asserted that
lm_disk_backed_tmpdir prints nothing when TMPDIR is already disk-backed,
and used pytest's tmp_path as the "disk-backed" directory:
# tmp_path is on the regular filesystem, so the default must be kept.
assert call("lm_disk_backed_tmpdir", env={"TMPDIR": str(tmp_path)}) == ""
That premise is false on the platform the helper was written for. Debian
13 mounts /tmp as tmpfs -- which is the entire reason lm_disk_backed_tmpdir
exists -- and pytest puts tmp_path under /tmp. So on the target platform
TMPDIR is memory-backed, the helper correctly answers /var/tmp, and the
test fails:
E AssertionError: assert '/var/tmp' == ''
The helper is right; the test was wrong. Reproduced on a box where
/tmp is tmpfs and / is ext4.
The test now looks for a directory whose backing store is actually disk
-- tmp_path, else a scratch dir under /var/tmp, else beside the library
-- using the same findmnt lookup the helper itself uses, and skips only
if no disk-backed directory exists anywhere. An earlier version of this
fix skipped whenever tmp_path was tmpfs, which made it skip on every
machine with a tmpfs /tmp; that is barely better than asserting the
wrong thing, so it now searches instead of giving up.
Verified: 31 passed, 0 skipped. Mutation-checked -- deleting the
"is the current TMPDIR memory-backed?" guard from lm_disk_backed_tmpdir
fails this test, so it still catches the regression it is there for.
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
||
|
ChuckandClaude Opus 5
|
568cb6d77f |
perf(vegas): report the frame rate when it is worth reporting (#487)
Vegas logged an FPS line at INFO every five seconds for the whole of
every run. Measured over two hours on a rig: 1410 samples, 98.5% of them
within 10% of target. The 1.5% that were not included a reading of
8.6fps against a target of 60 -- a real stall, completely invisible
inside 1389 lines reading "59.6". INFO is now reserved for a shortfall,
the recovery from one, and a slow heartbeat so a healthy marquee still
shows a pulse. Scroll-progress tracing drops to DEBUG for the same
reason: it runs for the whole of every scroll and is what you turn debug
on to watch.
Three review findings, all fixed here.
1. Per-frame timing used the wall clock (critical). The loop sleeps the
remainder of each frame budget:
frame_elapsed = <now> - frame_started
time.sleep(max(0.0, frame_interval - frame_elapsed))
These devices have no RTC, so the clock jumps by however wrong boot
time was when NTP first syncs. A backward step makes frame_elapsed
negative, `frame_interval - frame_elapsed` then exceeds the whole
budget, and the render loop stalls for the size of the correction. A
forward step instead inflates the p99 and worst-frame figures this
telemetry exists to report. Both per-frame timestamps are monotonic
now. start_time stays wall-clock: it is only used for the iteration
duration report, where a human-readable clock is the point.
2. FPS health state reset every iteration. last_fps_health_log and
was_degraded were locals of run_iteration(), which is called once per
cycle. Starting at 0.0 against a monotonic clock, `due` was true on
the first sample of every iteration, so the 300s heartbeat degenerated
into one report per cycle -- reintroducing the noise this change is
about. A recovery that crossed an iteration boundary was never
reported either, since was_degraded had already gone back to False.
Both now live on the coordinator and reset in start().
3. The degraded threshold read as an off-by-one. 90% of target is
deliberate -- a marquee jitters constantly, so "anything below target"
would report forever and mean nothing -- but nothing said so, leaving
55fps-against-60 looking like a missed case. The constant now states
the band and gives that exact example.
Also drops two soccer logo PNGs that a `git add -A` had swept into the
first commit. They are unreferenced, unrelated to frame-rate telemetry,
and 210KB.
Verified: each fix mutation-checked -- restoring the wall clock on either
per-frame timestamp, or making the health state local again, fails the
new tests. 566 passed across the vegas, coordinator and scroll suites.
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
||
|
ChuckandClaude Opus 5
|
6b74506695 |
fix(sports): fetch odds for the games shown, not the whole schedule window (#494)
* fix(sports): fetch odds for the games shown, not the whole window
SportsUpcoming.update() walked every upcoming game in the schedule
window and called _fetch_odds() on each one inside that collection
loop, narrowing to upcoming_games_to_show only afterwards. Each call is
a separate sequential ESPN request.
The comment sitting above it said odds were fetched "only for games that
will be displayed". The only narrowing it actually applied was
show_favorite_teams_only, which is not the default, so in the usual
configuration nothing narrowed it at all.
Measured on devpi, where the football plugin has the same shape:
467 odds requests in one 35s burst, 467 distinct events
315 NFL + 152 college-football -- roughly a whole season
plugin football-scoreboard operation timed out after 30.0s
The burst repeats each time the 1h odds TTL expires: 67 -> 327 -> 957 ->
1261 requests/hour across four consecutive hours. Between expiries the
cache works and the rate is zero, so this is a thundering herd on
expiry, not a caching failure.
The fetch now runs after selection, over team_games -- the list already
cut to upcoming_games_to_show. This mirrors the fix the football plugin
already carries; the shared base class never got it.
SportsLive is deliberately left as it is: it walks the raw event list
because it has to find which games are live, but only fetches odds for a
game that has already passed the is_live/is_halftime test, so its
fan-out is bounded by how many games are actually in progress. The test
pins that distinction rather than assuming it.
The test reads the AST rather than the source text, and asserts the full
set of call sites, so a new one has to be classified deliberately
instead of inheriting whichever behaviour it happens to land in. Writing
it that way is what turned up the SportsLive site, which I had missed.
Verified: reverting the fix fails the test with the offending iterable
named ("iterates over 'events'"). 525 passed, 9 skipped across the sports
and odds suites.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
* test(sports): check the odds guard structurally, not by its text
Review caught that _guards_above() collected an `if` test even when the
call sat in that if's `else`, so moving _fetch_odds() into the else of
the is_live/is_halftime test would still pass -- while fetching odds for
exactly the non-live games the guard exists to exclude.
Verifying that turned up a wider hole in the same assertion. It matched
substrings of the *unparsed source*, so a negated condition satisfied it
too:
if not (details["is_live"] or details["is_halftime"]):
self._fetch_odds(details) # every non-live game
Both names still appear in that text, so `"is_live" in guards` held and
the test passed on code doing the opposite of what it claims to check.
The guard test is now structural. It walks the AST for an enclosing `if`
whose *body* (never its `else`) contains the call, and whose test
references both names without either sitting under a `not`.
Verified by mutation: fetching odds for non-live games now fails with
"does not sit in the true branch of a test requiring the game to be in
progress". Moving the call into the else of the *favourites* test still
passes, which is correct -- the game there is still live, so the
in-progress contract holds and the fan-out stays bounded by how many
games are actually in play.
525 passed, 9 skipped across the sports and odds suites.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
||
|
ChuckandClaude Opus 5
|
5f29243e87 |
fix(config): make the device location the default for plugin location fields (#490)
* fix(config): make the device location the default for plugin location fields A user in Kansas City reported their radar centred on Dallas, TX with nothing in config.json to explain it. The radar is the `ledmatrix-weather` plugin's `radar` mode, and it centres on the same coordinates as every other weather mode: `forecast_data` lat/lon, geocoded from the plugin's own `location_city` / `location_state` / `location_country`. Those ship with schema defaults of Dallas / Texas / US. A user who never opened the weather plugin's config form therefore has no `location_city` on disk, and `PluginManager` merges the schema default in at load time — so the whole plugin (not just the radar) silently runs on Dallas. Radar is just the only mode that draws a recognisable map and gives the mismatch away. Meanwhile the device-wide `location` block that General settings writes was read by nothing at all, despite its own help text promising it was "used for weather, sunrise/sunset, and other location-based content". `SchemaManager.generate_default_config()` now substitutes the device `location` into the three fully-namespaced `location_*` keys before handing defaults back, so the promise holds: - Only `location_city` / `location_state` / `location_country` are substituted. A bare `state` key is left alone — `ledmatrix-elections` uses it for a two-letter code, and rewriting it would break that plugin. - A value the user saved on the plugin still wins: this replaces the schema default, and `merge_with_defaults` puts user config on top. - The substitution is applied on the way out of the defaults cache rather than into it, so changing the device location takes effect immediately. - No config manager, no `location` block, or an unreadable config all fall back to the plugin's own schema defaults. Every caller benefits: the plugin loader, the config form (which now pre-fills the user's real city), config save, and reset-to-defaults. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GNLrSZ32FNKpHRaduKEJsg * docs(web): name the exact plugin keys the device location seeds Review follow-up. The General settings help text said the device location was "the default for every plugin that asks for a city", which overstates what the code does: only the fully-namespaced `location_city` / `location_state` / `location_country` keys are substituted. A plugin with a bare `city` key gets nothing — deliberately, since `ledmatrix-elections` uses `state` for a two-letter code. The tips now name the exact keys. Worth noting for anyone editing these: `ui.help_tip(...)` takes a single-quoted Jinja string, so an apostrophe in the tip text has to be escaped or written around. The wording here avoids them. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GNLrSZ32FNKpHRaduKEJsg --------- Co-authored-by: Claude <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
1fbe244e49 |
fix(plugins): say when discovery skips a directory (#489)
* fix(plugins): say when discovery skips a directory A plugin can be enabled in config, enabled in plugin state, present on disk with a valid manifest and an importable entry point -- and simply absent from the running process, with nothing anywhere to say why. That is not hypothetical. hockey-scoreboard on a live rig is enabled in both places, imports cleanly when loaded by hand, and is listed in the Vegas plugin order, but is not among the 22 plugins the process actually holds. Establishing even that much meant comparing cache-file mtimes to find it had last run three days earlier. The journal had nothing, because discovery does not report what it declines to load. Two paths were silent. A directory with no manifest.json was skipped without comment, which is defensible until it is the thing you are trying to explain. Quieter still, a manifest that parsed but carried no "id" was read successfully and then dropped on the floor -- no warning, no trace, and the plugin simply does not exist as far as the rest of the system is concerned. Both now log a warning naming the directory and the reason. This does not explain the rig above; its manifest has an id. It makes the next occurrence diagnosable from the journal instead of from file timestamps. Reverting the change fails both tests. 65 plugin-system tests pass. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW * fix(plugins): warn once per directory, not once per scan Self-review catch. Discovery runs on every web UI page load and every config reconcile, so warning unconditionally about an unloadable directory would put a line in the journal each time someone opened a page -- the same log-volume problem this change exists to help diagnose. The skip is now reported once per directory per process. The diagnostic value is unchanged: the reason a plugin is missing still appears in the journal, once, where before it appeared nowhere. Test added covering five consecutive scans producing one warning. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW * fix(plugins): one unusable manifest no longer aborts the whole scan json.load accepts any JSON value, so a manifest.json holding null, [], "text" or 42 parses without complaint and then raises AttributeError on manifest.get('id'). Nothing catches that: the outer handler around the scan takes OSError and PermissionError only. So a single malformed manifest did not skip that one directory -- it aborted _scan_directory_for_plugins outright, and every other plugin on disk, however healthy, silently failed to register. Reproduced with three directories, the middle one holding `null`: SCAN ABORTED -> AttributeError: 'NoneType' object has no attribute 'get' the two valid plugins never registered That is the same failure this PR set out to fix, in its most severe form: a plugin enabled in config, enabled in plugin state, present on disk, and absent from the running process with nothing to say why -- except here it takes every other plugin with it. A manifest that is not a JSON object is now skipped like any other unusable directory, named once, with what it actually was: Skipping bad-null: its manifest.json is NoneType, not a JSON object Skipping bad-list: its manifest.json is list, not a JSON object scan returned: ['aaa-good', 'zzz-good'] Verified: removing the guard fails 6 of the 10 tests. Covers null, list, string, int and bool, and asserts the healthy plugins either side of the bad one still register. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
863e4a1ecd |
consolidate(perf): cut SD writes, log volume, and metrics churn (#486)
* fix(plugins): one bad metrics cache entry should not stop every plugin
Caught live on a rig: every plugin failing, once each, continuously.
ERROR - src.plugin_system.plugin_manager - plugin geochron operation failed:
ResourceMetrics.__init__() got an unexpected keyword argument
'consecutive_failures'
ERROR - ... plugin text-display operation failed: ...
ERROR - ... plugin news operation failed: ...
ERROR - ... plugin odds-ticker operation failed: ...
with /api/v3/health reporting plugin_system: not_initialized while the display
process itself kept running and updating the panel.
`consecutive_failures` is a plugin_health field, not a metrics one.
get_metrics() does ResourceMetrics(**cached), which raises TypeError on a
single unrecognised key, and that exception escapes into plugin_manager and is
reported per plugin. One malformed cache entry takes the whole plugin system
down.
How a health-shaped record came to sit under a plugin_metrics key on that
machine is not established, and I could not finish the diagnosis: the rig went
back into its EIO failure mode partway through -- SSH resetting pre-banner,
systemctl unexecutable -- while the web API kept answering from RAM. Checked
before that: the cache files on disk are correctly shaped and separate, and
CacheManager.get() returns the right record for each key, so it is not a live
key collision. A restored backup mixing two machines' caches is the likeliest
explanation, and that rig had one restored onto it.
Either way the loader should not be brittle enough for the answer to matter.
plugin_health already repairs its records field by field rather than trusting
what is on disk; this does the same. Known fields are kept, unknown ones are
dropped and named once in the log so a genuine schema change stays visible
rather than being silently discarded, and a non-mapping entry no longer raises.
Keeping the known fields matters: discarding the record wholesale would throw
away real call counts and timings because of an unrelated stray key.
Mutation-checked: restoring ResourceMetrics(**cached) fails 6 checks, dropping
the whole record fails the field-preservation check, and dropping unknown
fields silently fails the logging check. 28 tests pass across the resource
monitor and plugin health suites.
* perf(health): stop rewriting a health record on every healthy cycle
Every successful plugin update called record_success(), which persisted the
record unconditionally. In steady state the only fields that had changed were
total_successes and last_success_time -- a counter and a timestamp that
health_monitor surfaces for display and that nothing reads back after a
restart. Nothing alerts on the age of last_successful_update; it is carried in
the metrics dataclass and shown.
Measured on a rig running 24 plugins, all steady-state (0 consecutive
failures, circuit closed): a five-minute sample caught 22 health-file
rewrites, about 4.4 a minute or 6,300 a day. Each write is ~400 bytes through
cache_manager.set(), which writes a file per call, so each one costs a
filesystem block plus an ext4 journal write.
That lands on an SD card, where the unit of cost is an erase-block cycle
rather than the bytes involved, and where wear is what eventually kills the
card. Two cards have already failed on the other rig with the same
signature -- unreadable block device, EIO on exec, sshd unable to read its
host keys.
The circuit breaker still has to survive a restart, so the write is kept for
exactly the fields it is rebuilt from: consecutive_failures, circuit_state,
circuit_opened_time, half_open_start_time. A failure, a circuit opening and a
recovery are all still written the moment they happen. In-memory state is
updated every time either way, so the health API and web UI show what they
always did.
Tested: 100 healthy cycles now perform zero writes after the first, the
counters remain accurate in memory, and a failure, a recovery and a
half-open-to-closed transition each still reach disk. One test kills and
rebuilds the tracker from the cache to prove the breaker's state genuinely
survives what is no longer written.
Mutation-checked both ways: persisting unconditionally again fails the
steady-state test, and widening _DURABLE_FIELDS to include last_success_time
fails it too. The 46 existing health tests pass.
(cherry picked from commit
|
||
|
ChuckandClaude Opus 5
|
9c0c0dc851 |
consolidate(web): credential exposure, secret loss, and the update path (#485)
* fix(web): stop /config/main handing out every credential it holds
The endpoint returned the raw config to anyone who could reach the port, and
this web interface has no authentication of any kind. An unauthenticated
request against a live rig returned:
github.api_token 40 chars
incoming-packages.ha_token 183 chars
jellyfin-now-playing.api_key 32 chars
ledmatrix-weather.api_key 32 chars
on-air.mqtt_password 8 chars
youtube.api_key 20 chars
youtube-stats.api_key 39 chars
A GitHub token and a Home Assistant long-lived token among them. Anything on
that LAN could read them.
The x-secret masking the plugin config endpoints use does not reach here: this
route never consults a schema, and core keys such as github.api_token have no
schema to carry the marker. Several of the fields above *are* tagged x-secret
in their plugin's schema and were still returned in full, which is what rules
out the schema route as the fix for this endpoint.
Credential-named fields are now blanked. Matching on the name is blunt, and
for a whole-config dump that is the right default: anything named like a
credential should not leave the process, and a new plugin adding a
differently-shaped secret is covered without anyone remembering to tag it.
Blanked rather than removed, and safe to blank: POST /config/main merges into
the freshly loaded config and writes only the keys it was given, so a client
that round-trips this response cannot erase a secret it never saw. The web API
suites confirm it -- 81 passing, unchanged.
On the test that matters: the first version of this suite exercised the two
helpers and nothing else, and reverting the single line that wires the
redactor into the route passed all thirty of them. A property asserted on a
helper is not a property asserted on the endpoint, and it is the endpoint that
is exposed to the network. The added test goes through the view function, and
it does fail on that revert.
This also corrects an earlier claim of mine. I reported that GET /api/v3/config
did not expose these values; that path 404s, so the check proved nothing. The
real route is /config/main and it exposed all of them.
* fix(web): stop an unrelated config edit from erasing a plugin's secret
Saving any field on a plugin's config form destroyed that plugin's stored
credential. On a rig with a weather API key, changing the city silently
emptied the key, and the plugin stopped working at the next fetch with no
indication why.
The path had no guard at any step. The config partial masks secrets before
rendering (pages_v3.py:740), so the browser posts them back blank; _parse_value
deliberately preserves "" for optional string fields; separate_secrets routes
that "" into secrets_config, which is a truthy dict; deep_merge writes it over
the stored value; save_raw_file_content persists it.
The blank does not even need the round-trip. merge_with_defaults injects the
schema's api_key default ("") into every save, so a client that never sends
the field at all still erases it. test_secret_count_message_counts_top_level_keys
was counting exactly that injected blank as a saved secret field -- the visible
edge of the bug, pinned as expected behaviour.
remove_empty_secrets() already existed for this, with seven unit tests and a
docstring describing this precise scenario ("clients will send those empty
strings back ... so that existing stored secrets are not overwritten with
blanks"). It was never wired into a call site. This wires it into both save
paths that merge into the secrets file.
A blank now means "unchanged" rather than "delete", which is the same contract
the helper's tests already describe. The cost is that a secret can no longer be
cleared by emptying the field; clearing needs its own affordance, since a
control that erases credentials as a side effect of ordinary edits is not one.
Verified by reverting the guard: the new round-trip test then fails with the
stored key read back as ''. 262 web tests pass with it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
* fix(web): stop dumping the config and request headers to the journal
save_main_config logged its entire POST body and the full request headers at
ERROR on every save. The body is the configuration itself, and the headers
carry the session cookie, so a routine settings change wrote both to the
journal -- at a level that guarantees they survive any sane log filter.
The lines are leftover debug output: they say "DEBUG:" in the message while
calling logging.error, and they went through the root logger rather than the
module logger, bypassing the level configured for this blueprint.
Replaced with a debug-level line recording the shape of the request, which is
the part with diagnostic value. The local `import logging` went with them; it
shadowed a module-level import that was already there.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
* fix(web): stop /config/secrets handing out every credential it holds
GET /api/v3/config/secrets returned config_secrets.json in full to anyone who
could reach the port, and this interface has no authentication. Probed against
a real rig it produced six populated credential fields: a 40-character GitHub
token, a 183-character Home Assistant token, and Jellyfin and weather API keys.
This is the second door onto the same credentials; #477 closes the first.
Masking the response alone would have been worse than the leak. The only
client fetches every secret, edits one field and posts all of them back, and
save_raw_file_content replaces the file wholesale -- so a masked GET followed
by the client's own save would write the mask over every credential the user
had not touched. That is why this was left open when the leak was found; it
needs both halves.
Read side: mask_all_secret_values(), which already existed for exactly this
endpoint -- its docstring names it -- and had never been wired to a call site.
It leaves empty values and YOUR_* placeholders alone, so a client can still
tell "set" from "not set" without being told the secret.
Write side: strip the echoed mask and blanks from the submission, then merge
onto what is stored, so "unchanged" means unchanged. The cost is that a secret
can no longer be cleared by blanking it; that wants its own affordance, since
a control that erases credentials as a side effect of saving an unrelated one
is not one.
Browser side: the token field is now left empty rather than filled from the
response. Filling it with the mask would have stored eight bullet characters
as the token the next time the user pressed Save, and filling it with the real
value is the thing being fixed. It reports whether a token is saved instead.
Verified end to end through the Flask endpoints, not the helpers. Reverting
the masking fails the leak tests; reverting the merge fails the preservation
tests; both halves are independently guarded. 278 web tests pass.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
* fix(web): stop reporting "no update" when the update check could not run
check-update returned update_available=False whenever git failed. The banner
is the only route to the update button, so a checkout git refuses to touch
looked exactly like a current one -- permanently, with nothing on screen to
act on and only a log line recording why.
The common cause is an install performed as root. scripts/install/one-shot-install.sh
clones into ${HOME}/LEDMatrix, never consults SUDO_USER, and contains no chown
at all, while its own error text suggests running the whole thing under sudo.
The result is a root-owned checkout, and on a rig this is what every git
command in it does:
fatal: detected dubious ownership in repository at '...'
including the fetch this endpoint runs. Verified on real hardware rather than
assumed.
A failed check now reports check_failed with a message the user can act on --
for dubious ownership, the chown that fixes it. The banner shows that message
instead of hiding itself, with the update button suppressed since updating
cannot work until the cause is fixed. The success path is untouched.
This does not fix the installer, which is the real cause; it stops the symptom
being invisible. The installer needs SUDO_USER handling and a chown, and its
suggestion to run as root should go.
Reverting the endpoint change fails four of the five new tests; the fifth
guards the success path and correctly does not move.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
* fix(web): stop the installer chmod stripping exec bits on every update
git tracks five scripts as mode 644 that first_time_install.sh then chmods to
755 (start_display.sh, stop_display.sh, the two install_*_service.sh, and
one-shot-install.sh does the same to first_time_install.sh). With
core.fileMode true, the default on Linux, git reports all five as modified
from then on, in files the user never touched.
The update button stashes local changes before pulling, so it is not blocked
by this. But it never pops that stash -- stash pop and stash apply appear
nowhere in the update flow -- so the mode change is stashed away and left
there, and the files revert:
=== file modes after the update button's stash ===
664 first_time_install.sh <- installer had made these 755
664 start_display.sh
664 stop_display.sh
664 scripts/install/install_service.sh
So every web-UI update silently strips the executable bit from the installer's
own scripts, and leaves a stash entry holding the difference. start_display.sh
and stop_display.sh stop working from the shell afterwards.
A manual `git pull --rebase` over SSH fails outright, since nothing stashes for
it: "cannot pull with rebase: You have unstaged changes". That is the likely
source of the reports, since plenty of people update that way.
Tracking the five as 755 -- what they should always have been, as the
installer chmodding them attests -- removes the spurious mode change
entirely: nothing to stash, nothing stripped, no stash entry, and manual
pulls work.
The pull also passes --autostash, for the case the code explicitly tolerates:
when the stash fails it logs a warning and pulls anyway, and that pull is what
then fails. Autostash also pops what it stashes, which the manual stash does
not.
Note that `git add -A` after `git update-index --chmod=+x` silently reverts
the index to the on-disk mode, so the modes here were set by chmodding the
files themselves.
Regression test asserts the five stay tracked executable; reverting any one
of them fails it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
* fix(web): ask for the restart that makes an update take effect
The update button pulls new code and restarts nothing. There is no systemctl,
restart, reload or reboot anywhere in the 172-line git_pull handler -- it
stashes, pulls, installs changed requirements, re-removes plugins the user had
uninstalled, and returns "Code updated successfully."
Meanwhile both services go on running the code they loaded at boot. So the
display keeps rendering the old build, the web interface keeps serving the old
build, and the user is told the update worked. Nothing on screen suggests
otherwise, and the next reboot is what actually applies it -- whenever that is.
The affordance for this already exists: the restart-pending banner, raised
after main-config saves, with a Restart Now button wired to the display
service. A code update is a stronger reason to show it than a config save is.
The response now reports restart_required, and applyUpdate raises the banner
with wording for a code update rather than a config save. The banner's message
became a parameter and is persisted next to the flag, since it outlives the
page that raised it.
restart_required is only true when the pull actually moved HEAD. "Already up
to date" is a success too, and prompting after a no-op would train users to
dismiss the prompt unread.
This covers the display service, which is what the Restart Now button drives
and what users notice. The web interface still picks up its own new code on
its next restart; restarting it from inside a request it is serving is a
larger change than this one.
Reverting the flag fails the test that a pull which moved HEAD asks for a
restart. 290 tests pass.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
* Mask list-shaped secrets element-wise, close two vacuous tests
mask_all_secret_values treated any non-empty list as a scalar, so a
secrets file holding
"accounts": [{"name": "a", "token": "tok-a"}, {...}]
came back as a single "••••••••". The caller could not see how many
entries existed, and the raw editor was handed a string where the file
holds an array. Recurse into lists in both _mask_value and _contains_mask.
Lists merge by replacement, not key-wise, so strip_masked_values now
drops a list outright if any element still carries the mask -- storing a
half-masked list would discard the untouched entries.
Two tests could pass without exercising what they claim to check:
- test_git_pull_resolution asserted modes only for paths git ls-files
returned. A renamed or deleted installer target is simply absent from
that output, so its mode was never checked. Assert every CHMODDED path
is tracked first.
- test_config_secrets_masking never checked the POST status. A 500
leaves the old file in place, which satisfies every assertion that
follows. Assert 200 before reading the file back.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
||
|
ChuckandClaude Opus 5
|
71739d85d1 |
harden: cap malloc arenas, warn on unit drift, and grant the portal's sudo commands (#476)
* perf(systemd): cap glibc malloc arenas on the display service
Measured on a live rig 2.5 hours after start:
RSS 1030 MB
Private_Dirty 988 MB
anonymous mappings > 10 MB 23
largest 104, 79, 66, 63, 63 MB, on 64 MB-aligned addresses
threads 9
cores 3 -> glibc ceiling = 8 x 3 = 24 arenas
23 against a ceiling of 24, all 64 MB-aligned: these are glibc's per-thread
malloc arenas, not live objects. The data the process was actually holding
accounts for perhaps 15 MB -- the widest scroll strip observed was 35,746 x 64,
about 7 MB as RGB and the same again for its numpy mirror.
It is bloat rather than a leak: sampled four times over 135 seconds, RSS sat
between 990 and 1030 MB rather than climbing. glibc gives each allocating
thread its own arena, grows them to hold peak demand, and never gives them
back. A process that builds and drops large images across several threads is
exactly the shape that produces this.
The device had 59 MB free at the time, on 1845 MB total.
MALLOC_ARENA_MAX=2 trades a little allocator concurrency for that resident
memory. It is a tuning knob rather than a fix for a defect, so the rationale
and the measurements sit next to it in the unit file, and a test asserts they
stay there -- a bare environment variable invites removal by whoever meets it
next.
Two things this is NOT, both checked rather than assumed:
- Not an OOM problem today. A grep for "oom" in the service journal returned
24 matches, all of which were the radar logging zoom=9 and zoom=7. The kernel
OOM killer has not fired: dmesg has zero matches.
- Not currently capped by the unit's MemoryMax=85% either. That directive is in
this file but absent from the unit actually installed on the rig, which
reports MemoryMax=infinity, so nothing is enforcing a ceiling there.
The saving is unmeasured on hardware: applying it needs a service restart,
which blanks the panel, so that is the user's call rather than something to do
mid-audit. If p99 frame time regresses -- it sits at 18.4 ms against a 16.7 ms
budget for 60 FPS, so there is not much headroom -- raise the value rather than
remove it.
(cherry picked from commit
|
||
|
ChuckandClaude Opus 5
|
cc258aaffd |
fix(install): tag the secondary installer's journalctl grants too (#491)
Review was right on all three counts, and the first is the one that matters: scripts/install/configure_web_sudo.sh writes the same three wildcard journalctl rules as first_time_install.sh and none of them carried NOEXEC. So this PR closed the pager escape on one installer path and left it open on the other, which is close to no fix at all -- a rig configured through that script still hands out a root shell via less's "!command". The test could not have caught it, for two independent reasons. INSTALLERS did not list the file. And even listed, _grant_lines() kept the raw source line: that installer echoes its rules, so each one ends in a quote rather than the wildcard, and the trailing-* check skipped every one of them. Either alone would have hidden it. Both fixed: the file is covered, and an echoed rule is unwrapped to the sudoers line it actually emits. The selector test now covers -t ledmatrix as well. It asserted only the two -u forms, so deleting the -t rule would have passed. Verified by removing NOEXEC again from the secondary installer: four of the six tests fail, where before the suite passed with the vulnerability present. Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
Chuck
|
fe5a3aa99d |
harden(install): tag the journalctl sudo grants NOEXEC (#472)
journalctl starts a pager when its output is a terminal, and from less a "!sh"
is a shell with whatever privileges journalctl was given. That is the standard
journalctl escalation, and these rules end in a wildcard:
<user> ALL=(ALL) NOPASSWD: /usr/bin/journalctl -u ledmatrix *
Nothing this project runs needs the pager -- both call sites pass --no-pager,
in web_interface/app.py and api_v3.py. But a sudoers rule cannot require a flag
that sits in the middle of a command line, and reasoning about what a trailing
wildcard does and does not admit is exactly the kind of subtlety that produces
a hole. sudo's NOEXEC tag stops the command executing another program at all,
which closes it without depending on that reasoning.
NOEXEC works by LD_PRELOAD, so it applies to dynamically linked binaries.
Checked on the target hardware: journalctl there is dynamically linked. The
generated rules were run through `visudo -c` -- parsed OK.
Found while auditing the pre-existing wildcard grants, prompted by review
catching a far worse one I had added myself in the same area: `iptables *`,
where --modprobe runs an arbitrary path as root.
Reachability, stated plainly: on a stock Raspberry Pi image none of this
matters, because 010_pi-nopasswd already grants the default user
`ALL=(ALL) NOPASSWD: ALL`. It matters on a hardened install, or where the
service runs as a user without that blanket rule.
Two mutation checks: dropping NOEXEC from a rule fails, and deleting the rules
rather than tagging them fails too -- that second one matters, since "make the
test pass" and "remove the feature" would otherwise look the same.
|
||
|
ChuckandClaude Opus 5
|
10e75b977f |
Cover the next tier of untested modules and endpoints, and fix the 43 bugs that surfaced (#459)
* test(sync): cover the display sync protocol, and fix what that surfaced DisplaySyncManager had no tests at all — it appeared in the suite only as a MagicMock() stand-in, so none of its framing, handshake, or socket handling was ever exercised. Writing that coverage surfaced three bugs. Both receive loops caught the generic Exception and immediately retried. A socket left in a bad state raises on every call, so the thread spun at 100% CPU logging the same line; the reverted-code run of the new regression test takes 24 seconds where the fixed one takes 0.2. Both now back off briefly before retrying. The follower dispatched on `data[:8] == _RAW_MAGIC or len(data) > 512`. That size threshold is not part of either wire format: a control message over 512 bytes — a hello_ack carrying a long incompatibility error, for instance — went to the image decoder and was dropped, and a raw frame under 512 bytes went to the JSON parser. Both formats are already self-describing, so dispatch on the magic prefix and treat a JSON parse failure as the legacy unmarked PNG, with the shared frame bookkeeping factored into _handle_received_frame(). _oversized_frame_warned was created on first use through getattr(self, ..., False) rather than in __init__, alone among the instance attributes. 75 tests: role parsing, the hello compatibility matrix, watchdog timeouts, both receive loops, the TCP image server's length and dimension caps and decompression-bomb guard, status shape per role, and one end-to-end loopback handshake so the wire format is exercised for real and not only against mocks. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NohXi78cwsAKtN1sCfxjUh * test(logos): cover LogoHelper, and stop bad downloads poisoning the cache Nothing in test/ referenced logo_helper.py, so its caching, resizing and download-fallback logic was entirely unexercised. Two bugs surfaced. _download_logo wrote response.content to disk with no size limit and no check that the bytes were an image. A logo URL is remote input, so the response chose how much went into the assets directory; worse, an undecodable one stayed there, and because load_logo() only reports the decode failure and returns None, every later call re-read the same corrupt file. The download path never retried, so a single bad response made a logo permanently blank rather than falling back to the placeholder. Cap the response, verify it decodes, and delete it if not, which lets the existing fallback in load_logo_with_download do its job. get_cache_stats() divided by self.cache_size with no guard, so a helper built with cache_size=0 raised ZeroDivisionError from what is only a stats call. 37 tests: size-qualified cache keys, LRU eviction and refresh, the four load_logo_with_download paths, download permissions and timeout, placeholder generation, and the abbreviation normalizer — including a test pinning its deliberate divergence from LogoDownloader.normalize_abbreviation, since logo filenames on existing installs depend on both behaviors staying put. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NohXi78cwsAKtN1sCfxjUh * test(web): cover the error and response builders, and stop dropping empty values errors.py and error_handler.py's response builders had no direct tests, though every API response passes through them. Two bugs surfaced. WebInterfaceError set suggested_fixes with `or`, so a caller passing [] to mean "I have no suggestions for this one" got the default list instead. Only None should fall back. create_success_response gated `data` on `is not None` but `message` and `metadata` on truthiness, so an explicitly-passed "" or {} vanished from the response while 0 and False survived — the response shape depended on the value. api_helpers.success_response() then re-gated metadata the same way, which is the path every api_v3 endpoint actually calls, so fixing only the inner function would have changed nothing observable. Both now use `is not None`. That wrapper also merged request timing into the caller's own metadata dict in place. A caller reusing a dict across requests would accumulate previous responses' timings; it now copies before adding. 79 tests: category inference for every error code, mapped vs fallback suggestions, the JSON shape including which keys are omitted when empty, exception-to-code inference, and the success/error builders end to end. Two behaviours are pinned as deliberate rather than fixed: an empty context stays out of the response body, and from_exception's `message` is the fixed per-code string, never the raw exception text. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NohXi78cwsAKtN1sCfxjUh * test(web): cover the input validators, and close three holes in them validators.py had tests for dedup_unique_arrays only; the other eight functions were untested. Three bugs surfaced. validate_image_url checked for '..' only inside its relative-path branch, so http://host/../secret passed validation while /../secret was rejected — the traversal check now runs before the branch split, which is where a safety check on the whole URL belongs. validate_file_upload lowercased the uploaded filename's extension but compared it against the caller's list verbatim, so allowed_extensions of ['.TTF'] rejected every valid .ttf file. Both sides are lowercased now. The one in-tree caller passes lowercase already, so this only widens what future callers can hand it. validate_numeric_range accepted True and False, because bool subclasses int; a boolean then compared as 1 or 0 against the range and validated cleanly. Excluded explicitly, matching how base_plugin.py already handles the same trap for display_duration. 84 tests. Two behaviours are pinned rather than changed: sanitize_plugin_config deliberately does not HTML-escape strings, since escaping at this layer would store the escaped form in config.json — the docstring said "prevent injection", which read as a promise it does not keep, and now says what it actually does. validate_font_awesome_class's second 'fa-' check is unreachable behind its own regex; harmless, so characterized rather than removed. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NohXi78cwsAKtN1sCfxjUh * test(api): cover wifi and registry endpoints, and fix bodyless POSTs The /wifi/* routes drive the host's real networking and the registry routes reach GitHub, and neither had endpoint-level tests. Covering them surfaced a bug affecting six endpoints. Six handlers read their body as `request.get_json() or {}`. The `or {}` says every field is optional and a missing body should fall back to defaults — but get_json() without silent=True raises UnsupportedMediaType when there is no JSON Content-Type, and it raises before `or {}` is ever evaluated. Each handler's catch-all then reported that as a 500. So POSTing with no body — what curl sends by default, and what a fetch() without options sends — failed on /plugins/store/refresh, /display/on-demand/start, /plugins/config/reset, /plugins/of-the-day/json/delete, /plugins/{id}/limits and /plugins/authenticate/spotify. The shipped UI always sends a JSON object, which is why this stayed hidden. All six now use silent=True. test_api_v3_optional_body.py covers the affected endpoints and adds a source check, since the combination of `or <default>` with a non-silent read is self-contradictory wherever it appears and is easier to catch by inspection than by exercising each endpoint by hand. Also adds test/_api_v3_test_helpers.py: the blueprint holds its managers on a module-level singleton rather than in Flask app state, so a test that mocks them leaks into every later test unless the originals are restored. The existing _make_client() does this for unittest classes; this is the pytest-fixture equivalent, for the five suites still to come. 69 endpoint tests: connect/disconnect/AP/radio including the string-aware boolean coercion these endpoints deliberately use, the radio's lockout-refusal path, registry refresh and fetch-from-URL, and a guard that WiFiManager is never constructed for real. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NohXi78cwsAKtN1sCfxjUh * test(api): cover the music auth endpoints, and always clean up the wrapper The Spotify step-2 handler writes a Python wrapper script to a temp file with the user's redirect URL embedded in its source, then executes it. That is the most dangerous shape in the blueprint and had no tests. The wrapper was deleted in the success/failure branch and again in the TimeoutExpired handler. Any other failure from subprocess.run — no interpreter, a fork failure, an interrupted call — reached neither, and left a world-readable temp file containing the user's redirect URL on disk. Cleanup moves to a finally block, which is what "delete this whatever happens" should have been from the start. The injection tests are the point of this file. Eight adversarial redirect URLs (embedded quotes, backslashes, newlines, triple quotes, a full `"; import os; os.system("id"); "`) are each pushed through the endpoint and the generated wrapper is parsed with ast: it must still be valid Python, the URL must still be a single string literal bound to redirect_url, and no os.system call may appear anywhere in the tree. json.dumps holds up, but nothing was checking that it does. 40 tests. Also pins that the two endpoints are not symmetrical despite the matching names — only Spotify has a two-step flow and a wrapper; YTM runs its script directly — so a later change does not "restore" a parity that was never there. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NohXi78cwsAKtN1sCfxjUh * test(api): cover the credentials upload, and stop it hoarding secrets The endpoint that receives the user's Google OAuth credentials file had no tests. Two bugs surfaced. The OAuth-shape check ran inside `except Exception: pass`. A JSON document that parses but is not an object — a bare 42, true, null, a list — makes `'installed' not in creds_data` raise TypeError, which the bare except swallowed, and the file was then written out as credentials.json regardless. The check now decides the outcome instead of being advisory, so anything not credentials-shaped is refused up front rather than failing later inside the calendar plugin. Every overwrite copies the old file to credentials.json.backup.<ts> and nothing removed them, so a user who re-uploaded ten times had ten complete sets of OAuth client credentials sitting in the plugin directory, indefinitely. Keep the newest five. Pruning is housekeeping, so a backup that cannot be removed logs and leaves the upload alone. 27 tests: size and extension limits, malformed JSON, the shape check, 0600 permissions on the written file, backup-on-overwrite, and pruning including the repeated-upload case that stays bounded. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NohXi78cwsAKtN1sCfxjUh * test(api): cover the install endpoints, and make 14 dead guards reachable /plugins/install and /plugins/install-from-url were tested only at the PluginStoreManager layer, so the route logic — the queue-versus-direct branch, schema invalidation, discovery, state and history recording — was unexercised. Covering them surfaced the wider form of the body-parsing bug fixed for the `or {}` handlers in the previous commit. Fourteen handlers read `data = request.get_json()` and immediately guard with `if not data: return 400, 'No data provided'`. That guard cannot run: get_json() without silent=True raises UnsupportedMediaType for a request with no JSON body, so the catch-all answered 500 "an error occurred; see logs for details" where the handler plainly meant to answer 400 and say which field was missing. Every one of these endpoints told a caller who simply forgot the body to go read the server logs. All fourteen now use silent=True, so the guard each author already wrote is the one that runs. This covers /config/raw/main and /config/raw/secrets among them, whose own bodyless case had the same shape. The two remaining bare reads are left alone: neither declares what a missing body should do, so there is no stated intent to honour. 31 install tests plus 17 body tests. The install pair is checked against each other rather than only individually — the same install logic is written twice, once in the queue callback and once in the fallback, so the tests assert both produce identical schema, discovery, state and history effects. They agree today; the one difference is the success message wording, which is characterized rather than changed. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NohXi78cwsAKtN1sCfxjUh * test(api): cover the raw config write endpoints /config/raw/main and /config/raw/secrets write whatever JSON they are given straight to config.json and config_secrets.json, bypassing the secret-separation path the rest of the config surface goes through. Given how carefully that surface keeps secrets out of config.json, the pair that skips it was worth pinning precisely. Backed by a real ConfigManager over tmp_path, so the assertions are against files on disk. 20 tests covering both routes: what lands in which file, that a raw secrets write never touches config.json and vice versa, the GitHub token reload, the uninitialized-manager and empty-body branches, and the ConfigError path that carries config_path through to the response. The bypass itself is pinned as intentional rather than changed — these back the raw JSON editor, so writing the body verbatim is the feature. The test says so explicitly, because the failure mode is someone later routing plugin config through here as a convenience and silently losing secret separation. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NohXi78cwsAKtN1sCfxjUh * test(api): cover backup restore and path containment, and fix restore scope Restore is the most destructive thing the web interface can do — it overwrites config, secrets, WiFi settings and fonts, then reinstalls plugins — and neither it nor the file routes beside it had tests. A malformed `options` field fell back to {}. Every RestoreOptions flag defaults to True, so a caller who asked for a narrow restore and mis-serialized the request got a full one instead, secrets included, and was told it succeeded. Valid JSON that is not an object was worse: `"null"` or `"[1,2]"` reached .get() on a non-dict and raised, so the request died as a generic 500. Both are now refused with a 400 that says what was wrong, and restore_backup is never reached. The other file routes take a filename straight out of the URL and turn it into a path — one to read, one to unlink. _safe_backup_path is the only thing keeping those inside the export directory, and it was untested. No bypass was found; the thirteen traversal shapes are pinned so a later loosening of that pattern has to argue with something. The delete route's by-name enumeration is covered too, including that a directory sharing a backup's name is not removed. 84 tests. Two behaviours are pinned as intentional: a failed plugin reinstall turns the whole restore into an error even though file restoration succeeded, and omitting `options` entirely still means restore everything — that is the documented default, and it is only the mis-serialized case that was wrong. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NohXi78cwsAKtN1sCfxjUh * ci: raise coverage floor to 52% Measured 54.45% after the Tier 1 and Tier 2 suites, up from 50%. Keeping the same two points of headroom the 45 -> 48 ratchet used. The modules this branch set out to cover: sync_manager 0 -> 97%, logo_helper 0 -> 98%, errors and error_handler 0 -> 100%, validators 0 -> 97%. api_v3 moved less in percentage terms because it is 4,341 statements, but the endpoints covered are the destructive and credential-handling ones. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NohXi78cwsAKtN1sCfxjUh * test(sync): probe for a free port on loopback, not every interface CodeQL flagged the ephemeral-port probe in the handshake test for binding to all interfaces. The probe only needs a free port number, so loopback is both sufficient and correct — a test should not open a port to the network to discover one. The manager under test still binds to all interfaces, which is deliberate and already marked nosec: a follower has to receive the leader's UDP broadcast. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NohXi78cwsAKtN1sCfxjUh * fix: bound the logo download, and stop malformed input reading as a fault Review findings on the coverage branch. The download size cap I added checked len(response.content), which has already buffered the whole body -- it stopped the bytes reaching disk but not memory, which was the point. A server that omits Content-Length and never stops sending would still exhaust the process. Stream it instead, counting as it arrives, into a sibling .part file that is replaced over the target only once it decodes. A transfer that dies midway now leaves nothing behind rather than a truncated logo for load_logo() to cache. The follower's control-message handler caught three exception types, but two reachable UDP payloads raise others: a bare JSON scalar makes msg.get() raise AttributeError, and an "sx" carrying a non-numeric x raises ValueError or TypeError from float(). Those escaped to the outer handler, skipping the legacy-PNG fallback and -- since this branch added a backoff there -- charging one malformed packet a 0.1s stall on the receive path. The legacy-PNG path also decoded without the dimension cap its TCP counterpart applies, so a crafted 65KB frame could force a large allocation on the render thread; both paths now share one constant. Three repo_url handlers called .strip() on client input without checking it was a string, so {"repo_url": 12345} answered 500. The credentials upload parsed the same file twice, the second time inside a bare except that a preceding parse had already made unreachable. And both raw-config handlers kept a json.JSONDecodeError arm that get_json(silent=True) had turned into dead code, collapsing "sent something unparseable" into "sent nothing" -- they now say which. Two of the new tests were not testing what they claimed. The pruning round-trip wrote ten backups inside one second, so all ten landed on the same int(time.time()) filename and overwrote each other; it never reached the limit it asserted. And the sync clock helper patched attributes on the stdlib time module, freezing time process-wide for every daemon thread earlier tests had left running. Full suite: 3352 passed, coverage 54%. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NohXi78cwsAKtN1sCfxjUh * test(sync): probe broadcast by sending, not by listening The broadcast check added in the previous commit bound INADDR_ANY to receive its own probe datagram, and the free-port probe did the same to pick a port. CodeQL flagged both, correctly: a test suite has no reason to open a socket the whole network can reach. Sending is enough for what the probe is actually for. An environment that refuses broadcast raises on sendto, which is the case that occurs in sandboxes and is the one worth skipping over; confirming delivery would have required the listening socket. A network that accepts the send and silently drops it still reaches the assertion, exactly as it did before either commit. The port probe binds loopback -- it only needs a number, and the manager's own bind is the one that has to succeed, with the retry loop already covering a port taken elsewhere. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NohXi78cwsAKtN1sCfxjUh * fix: keep callback faults out of the frame-decode fallback Review follow-up on the previous two commits. Widening the control-message except tuple put the callback dispatch inside it, so an _on_new_cycle() that raised ValueError, TypeError or AttributeError sent a perfectly good control packet to the legacy PNG decoder -- which reported it as an image decode error and buried the real fault. Split the two: whether the payload parses as JSON decides frame vs control message, a second guard covers reading the fields of an attacker-shaped body, and the callback fires outside both. It still cannot kill the receive thread; the loop's own handler catches it, and now says what actually went wrong. The logo download's temp file was a fixed "<name>.part". Two plugins asking for the same logo at once would interleave writes into it, publish the mixture, or delete each other's partial. mkstemp gives each download its own name in the same directory, so os.replace stays atomic. Its descriptor is adopted by fdopen before the request runs, since a request that raises before the write would otherwise leak the fd -- quietly, because load_logo_with_download swallows that. Two test fixes: the oversized-frame test replaced PIL.Image.open process-wide, the same hazard the clock helper documents, and Ruff B007 on an unused loop variable. Full suite: 3355 passed, coverage 54%. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NohXi78cwsAKtN1sCfxjUh * test(sync): cover the announce loop, and reject non-finite scroll positions Three review findings from the follower receive path. Non-finite scroll x reached follower rendering. json.loads accepts the bare NaN/Infinity literals and float() accepts them as strings, so "x": NaN arrived as a real float and was stored verbatim. NaN loses every comparison the scroll code makes, so a follower given one sits on a position it can never advance past. It now raises through the existing malformed-control-message guard, which logs and drops the packet and leaves the last good position in place. _broadcast_available() only proves the host accepts sendto() for a broadcast; a network that accepts the send and drops the packet would let TestRealSocketHandshake run to its five-second deadline and fail on assertions the code did not break. The deadline now distinguishes the two: if not one packet crossed in either direction, that is the environment, and the test skips rather than reporting a protocol failure. That skip could hide a real regression in the announcing side, so TestFollowerAnnounceLoop covers it on mock sockets, where no network is involved and nothing can skip: hello carries this display's hardware config and goes to the broadcast address, heartbeats follow, an empty hardware config falls back to 32x64x1, hello is not resent before its interval, and a send failure is swallowed rather than killing the loop. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NohXi78cwsAKtN1sCfxjUh --------- Co-authored-by: Claude <noreply@anthropic.com> |
||
|
Chuck
|
cf0a551f7b |
fix(web): stop checkbox groups posting back options they cannot show (#465)
The enum that lets a checkbox group draw its options is also what validates the saved value. When an option goes away -- a league retires a team code, a schema drops a choice -- a config still holding the old value has no checkbox to render for it, but the value stayed in the hidden _data input anyway: that input is seeded from the stored array and only rebuilt by updateCheckboxGroupData() on change. So the stale value was posted back on every save the user did not happen to touch that widget for. The schema rejected it and the save endpoint returned 400 CONFIG_VALIDATION_FAILED, which blocks editing *any* field on that plugin until the user works out which invisible entry is at fault -- with nothing on screen naming it, because the offending value is precisely the one with no checkbox. Runtime was never affected: load_plugin() treats schema violations as warn/degrade, and a retired code already matched nothing. Only the web UI blocked. Values not in the enum are now dropped before the hidden input is seeded, and listed above the group so the selection is not lost silently. Only when the widget has options -- an empty enum means there is nothing to check against, and filtering on it would wipe the field. This is not hypothetical. ledmatrix-plugins #212 ("correct team abbreviations so config save no longer 400s") and #234 (removed the retired NHL code UTA from a picker across four plugins) are both this failure mode, fixed one league at a time. Nine shipped plugins use checkbox-group today; all of them get the fix. Tested by rendering the checkbox-group block lifted out of the shipped template, following test_enum_option_labels.py, so the tests exercise the production expression rather than a copy. Mutation-checked: removing the filter fails 2 tests, filtering unconditionally fails the empty-enum test, and dropping the notice fails the one asserting the value is named. |
||
|
ChuckandClaude Opus 5
|
9018fa23cd |
fix(web): verify the onboarding timezone step, don't compare it to the default (#462)
* fix(web): verify the onboarding timezone step, don't compare it to the default The Getting Started card's timezone step ticked when the saved timezone differed from the value config.template.json ships (America/New_York), OR-ed with the saved city differing from Tampa. Both halves were wrong. "Differs from the default" answers "did somebody edit this?", but what the checklist needs to know is whether the value is right. A user genuinely in America/New_York could never satisfy it, so the card nagged forever with four of five steps done -- the case that prompted this, on a panel whose timezone was correct all along. The city half was worse than useless: the saved city says nothing about whether the timezone is set, and because the two were OR-ed, saving a city ticked the step off with the timezone still wrong. That is the direction that actually breaks displays, since event times then render in the wrong zone. The browser already knows its own zone, so compare against that. No new persisted state, no network, and it catches the reverse case the old test got backwards: a panel still set to the old zone after a move now stays unticked, where before it ticked the moment the value stopped being the default. Zones are compared by the wall-clock time they produce for one instant rather than by identifier, so aliases (Asia/Calcutta vs Asia/Kolkata, Europe/Kiev vs Europe/Kyiv) don't read as a mismatch. When they genuinely differ the step names the browser's zone, so an unticked box says why. Configs with no timezone, an unparseable zone, or a browser without Intl leave the step open for the existing manual tick. The step still deep-links to the General tab, and the location value stays visible in its label -- it just no longer votes on whether the timezone is configured. Tests render the partial across configured zones and both cities: the step never pre-ticks server-side, carries the configured zone for the client to check, is unmoved by the city, and the panel-size step still resolves server-side. Reverting the template fails 9 of the 11. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW * fix(web): compare zones on fields Intl has always had dateStyle/timeStyle are late additions to Intl -- Firefox shipped them in 91 -- and an implementation that does not know them ignores them and formats the date alone. The comparison would then read New York, Chicago and Madrid as the same zone and tick the step for a timezone that is plainly wrong, which is the failure the check exists to catch. Silent, and only on older browsers. Explicit numeric fields (year/month/day/hour/minute) have been in Intl since ECMA-402 v1, so there is nothing left to degrade to. The options look like a stylistic choice, so a test pins them: it reads the comparison with comments stripped -- the comment names dateStyle to explain why it is not used -- and fails if either style option comes back or a time field is dropped. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW * fix(web): sample both sides of DST when comparing zones CodeRabbit caught this and it is right: comparing the wall clock at one instant treats zones that merely coincide right now as the same one. America/New_York and America/Lima hold the same offset all winter, so a panel set to the wrong one of the two ticked the step in January and then ran an hour off from March -- a silent false pass, which is the failure the whole check exists to prevent. Same shape as the dateStyle problem in the previous commit: a comparison coarser than it looks. Three instants now, all of which must agree: now, and mid-January and mid-July of the current year. Those sit either side of DST in both hemispheres, so only zones that agree year-round match. Toronto still matches New York, which is correct -- either renders the same times. Two tests. A static one asserts the comparison samples more than the current instant, since reverting to `[now]` looks like a simplification. And a table pinning which pairs must count as the same zone: aliases and same-rule zones equal, seasonal coincidences (New York/Lima, Phoenix/Los_Angeles, Sydney/Guadalcanal) not. That table mirrors the algorithm rather than executing the shipped JS -- there is no JS runtime here and the repo has no JS test infra -- so it records the verdicts the browser code has to reach, and the static guard keeps the two aligned. Mutation-checked: reverting to a single instant fails the static guard. Also documented what the city test compares. CodeRabbit read it as always failing, on the grounds that the label differs between Tampa and Seattle. It does, but timezone_step() returns the opening tag only, so the comparison is over data-done and data-tz and the label is not in it. The assertion is left as an equality over the whole tag, which is stronger than checking the two attributes by name; the docstring now says so. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
0c5b9c57d3 |
fix: keep low-memory boards reachable under load (#464)
* fix(service): survive corrupt health cache and clean exits Three independent failure modes that each end with a dark panel and no automatic recovery. 1. PluginHealthTracker._load_health_state returned the cached value verbatim. If that value is not a dict, every caller raises AttributeError: 'list' object has no attribute 'get' — during DisplayController.__init__, so the process dies before the display loop starts. systemd restarts it, the same bad entry is read back from disk, and it dies again: an unattended restart loop that survives reboots because the cause is persisted. Observed in the field with plugin_health:<id> holding an unrelated plugin's list payload. Now non-dict entries are discarded with a warning and the defaults are rebuilt. 2. ledmatrix.service used Restart=on-failure, so any exit with status 0 left the unit stopped and the panel dark indefinitely — systemd treats it as success and never brings it back. Restart=always. 3. ledmatrix-wifi-monitor.service used StandardOutput=syslog, which systemd has marked obsolete; it warns and rewrites it to journal on every load. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * perf(memory): size the cache to the board and stop reinstalling deps On a 1GB Pi 3B+ the display process settles around 600MB RSS of 905MB total. When the remaining headroom runs out the failure is not a clean crash: fork() starts returning ENOMEM, so sshd accepts connections and closes them before its banner, timer jobs stop running, and the panel goes dark, while already-resident processes keep serving normally. The board looks healthy from outside and cannot be logged into. Only a power cycle clears it. Three contributing causes: - MemoryCache had a fixed 1000-entry ceiling. Entries are parsed API payloads of tens of KB, so one ceiling cannot serve both a 512MB Zero 2 W and an 8GB Pi 5. Now scaled from MemTotal (150 entries at <=1GB, 1500 at >=8GB), overridable with LEDMATRIX_CACHE_MAX_ENTRIES. - requirements_are_satisfied() returned False for any requirement with extras, so a plugin depending on python-socketio[client] re-ran pip on every single start: ~8s, a network dependency, and a 100-200MB spike at the least convenient moment. During a restart loop it repeats for each restart. Extras are now resolved one level deep against installed metadata, keeping the conservative "anything unverifiable falls through to pip" contract. - ledmatrix.service had no memory ceiling. MemoryMax=85% expressed as a percentage so one unit file suits every board. Note this needs the memory cgroup controller, which Pi firmware disables by default; first_time_install.sh now adds cgroup_enable=memory to cmdline.txt, and the unit file documents how to verify it took effect. first_time_install.sh also enables persistent journald storage (capped at 64M). Default storage is volatile, so every reboot destroys the logs that would explain why the board rebooted. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * docs: guidance for 512MB and 1GB boards Documents the memory ceiling on small boards and, more usefully, what running into it actually looks like: sshd accepting connections and closing them before the banner, the web UI still responding normally, clean ping, a dark panel, and a wrong clock after the next boot. None of those read as "out of memory", which makes the failure hard to identify from the symptoms. Cross-referenced from SSH_UNAVAILABLE_AFTER_INSTALL.md, since "I can't SSH in any more" is how most people will first meet this. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix: address review findings on the low-memory work Nine CodeRabbit findings, five in code. **Health state (the one that matters).** The non-dict guard did not cover a dict missing fields the callers index directly, which is the shape actually seen in the wild: a record carrying only circuit_state produced `plugin clock-simple operation failed: 'circuit_state'` about fifty times a minute with the panel frozen. The record is now completed against the defaults per field rather than trusted or discarded wholesale. Per field matters: a first pass rejected any incomplete record outright, which reset a tripped breaker and real failure counts to healthy because one optional field was absent -- an existing test caught it. Values of the wrong type (a counter persisted as a string, an unknown circuit_state) fall back individually, valid neighbours survive, and newer fields the schema has grown since (degraded, degraded_reason) are carried through untouched. **Cache ceiling.** MemoryCache.set() accepted entries without bound between cleanup sweeps, which run every 300s by default, so a burst could take the cache far past max_size -- the unbounded growth the limit exists to stop. Eviction now runs under the same lock on every write, sharing one helper with the periodic sweep so the two cannot drift. **Installer, cgroups.** Only cgroup_enable=memory was checked, so a board carrying that without cgroup_memory=1 reported success and got no change, leaving MemoryMax= inert. Each parameter is now checked and appended independently; verified against all four combinations, single line preserved. **Installer, journald.** Persistence was inferred from /var/log/journal being non-empty, which proves neither Storage=persistent nor a size cap -- the directory survives a switch back to volatile. The effective configuration is read instead (systemd-analyze cat-config, falling back to the conf files), and an explicitly configured SystemMaxUse is preserved rather than overwritten. Verified across volatile, persistent-without-cap, persistent-with-user-cap, cap-without-storage, and commented-only configs. **Dependency extras.** _extras_are_satisfied stopped at one level, so a gated dependency that itself requests an extra (requests[socks]) passed on the base distribution's version while the extra's own dependency was missing, and pip was skipped. It now recurses, with a visited (distribution, extras) set so a cycle terminates. Docs: both kernel command-line paths documented (the installer falls back to /boot/cmdline.txt), daemon-reload and restart added after the systemd override example, memory exhaustion added to the SSH summary with its power-cycle-only recovery, and a language on the fenced block for MD040. Tests: five for the health-state repair including the exact wild shape and that record_failure/record_success no longer raise against it, and one for the cache ceiling. Both mutation-checked. Full suite 2927 passed, with the one pre-existing tmpfs failure that also fails on main. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW * fix: harden the health-state repair and confirm journald took effect Second review round; all three findings were valid and two were bugs in the repair added last commit. The repair could raise out of itself. An unhashable circuit_state (a list or dict on disk) hit `value in {...}` and raised TypeError -- from the code whose whole job is to stop a malformed record crashing the caller. It now requires a str before the membership test. bool is a subclass of int, so True passed the timestamp check and then compared as 1.0: enough to expire a cooldown the instant the breaker opened, while False would stop the elapsed check firing at all. Timestamps now exclude bool explicitly. The regression test for the original crash was seeded with a record that *contained* circuit_state, so it passed against the old raw-return behaviour too -- the counters are read with .get(), so circuit_state is the only field whose absence used to raise. Reseeded to omit it, and it now fails against raw-return as intended. journald: drop-ins apply in lexical order, so a local file sorting after ledmatrix-persistent.conf still wins and writing ours proves nothing. The effective Storage is re-read afterwards and a warning naming the diagnostic command is printed if persistence is still not active, rather than reporting a success that was not verified. Full suite 2934 passed, same single pre-existing tmpfs failure. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |
||
|
ChuckandClaude Opus 5
|
9083df9f5c |
fix(pixlet): resolve the release tag correctly when downloading (#461)
* fix(pixlet): resolve the release tag correctly when downloading
Starlark apps render through the pixlet binary, and the installer that
fetches it silently produced nothing, so every app failed with "Pixlet
not available - Starlark apps will not work".
Two compounding defects:
The version lookup parsed the wrong token. GitHub returns the release
JSON on a single line, so `grep '"tag_name"'` matches the whole document
and the greedy `sed 's/.*"([^"]+)".*/\1/'` captures the LAST quoted
string in it. That resolved to "mentions_count", giving a download URL
for a release that does not exist. The `[ -z "$PIXLET_VERSION" ]`
fallback never fired, because the value was not empty -- just wrong.
And `curl -L -o` without `-f` writes a 404 body to the file and exits 0,
so the download was reported as successful and the first sign of trouble
was tar complaining "not in gzip format" about a page of HTML:
→ Downloading linux-arm64...
Extracting...
gzip: stdin: not in gzip format
✗ Failed to extract archive: .../pixlet_mentions_count_linux-arm64.tar.gz
Download complete: 0/1 succeeded
Now the tag field is matched directly and the value taken from it, and
the result is checked for a version shape rather than merely being
non-empty -- a wrong-but-non-empty value is exactly what made this
silent. curl gets -f so an HTTP error is a failure, and the archive is
gzip-tested before extraction, since a proxy can return 200 with an
error page.
Verified on an arm64 rig: v0.53.1 resolved, 1/1 downloaded, the binary
runs, and the plugin's own detection finds it at
bin/pixlet/pixlet-linux-arm64.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Udr6MfaFLUPhX5Fgo67Jf5
* fix(pixlet): anchor the version check, and don't echo response bytes raw
Both CodeRabbit findings were valid.
The shape check accepted partial matches, so "v0.53garbage", "0.53" and
"v0.5" passed it and built a download URL for a release that cannot exist
-- the failure the check was added to stop, just one step later. Anchored
at both ends now. Every tronbyt/pixlet release to date is vX.Y.Z (all 38
verified against the API), with an optional suffix left for a future -rc.1
or +build tag.
The invalid-response diagnostic printed bytes straight from whatever
answered the request. NUL and newline were filtered but escape, carriage
return and backspace were not, so an error page could rewrite the output
or bury it in a CI log. Non-printable bytes are stripped and it goes
through printf. CodeRabbit suggested hex-encoding the lot; printable
characters are kept instead, because "<!DOCTYPE html>" is the diagnostic
-- hex would make the line safe and useless.
Also corrected the comment above the parse. It asserted GitHub returns
this JSON on a single line; the API is pretty-printed by default, and I
could not get a single-line response from two machines across five header
variants. The single-line case is real (it is what produces
"mentions_count", and the failing device's error named
pixlet_mentions_count_linux-arm64.tar.gz), but it is a shape to be robust
against, not a constant. As written the comment invites the next reader to
check by hand, see pretty JSON, and conclude the fix was unnecessary.
Tests drive the real script with a stubbed curl: the tag resolves from
both response shapes, non-release values fall back, an HTTP error is
reported as a download failure rather than surfacing later as a tar error,
a non-archive body is rejected before extraction, and the diagnostic
cannot carry control bytes. The stub honours -f the way real curl does --
without that, the HTTP-error test passed against the old script too, since
both end at 0/1 and only the reporting layer differs.
Mutation-checked: 10 of the 16 fail against the pre-fix script, and the 5
covering these two findings fail against this branch's previous state.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01STMbQE4YctTacQXfbYqKuW
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
|