Every push to a pull-request branch starts the CI matrix and both review
bots, and runs from several open pull requests queue behind one another.
Move the fix rounds off GitHub: a branch is reviewed with the Codex and
Greptile CLIs until both are clean, then pushed once.
- AGENTS.md: the rule, linked to the procedure
- agent-workflow.md: commands, loop, stop conditions and exemptions
- agent-context-map.md: route the topic to the workflow document
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
* perf(electron): let hidden and minimized windows report themselves hidden
The main window was created with backgroundThrottling: false (since #1123,
without a stated reason). Electron then keeps document.visibilityState at
"visible" for a hidden, minimized or fully covered window and never lets
Chromium throttle it, so every renderer timer, rAF and CSS transition ran at
full rate in the background, and the playback keep-awake gate, which
releases the display for a minimized window, could never see one.
Use Chromium's default. Audible media and picture-in-picture are exempt
from background throttling in Chromium, and a local check confirmed HLS
playback continues unchanged through more than six minutes minimized,
audible and muted.
Playwright's focus emulation pins every page it attaches to as visible, so
the new window-visibility E2E launches the app without Playwright and
drives it over raw CDP (electron-unautomated-launch.ts).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(e2e): harden the unautomated Electron launch
Review follow-ups for the window-visibility E2E:
- Resolve `electron` in the main process through a require created from
the `node:module` builtin instead of `process.mainModule`, which only
exists when the app entry is CommonJS.
- Bound teardown like closeElectronApplicationAndConfirmExit: SIGTERM,
then SIGKILL, 5 s each, then fail instead of waiting forever.
- Surface CDP protocol errors from Runtime.evaluate instead of returning
undefined.
- Wait until the window is actually shown before hiding it. The app shows
its window on ready-to-show, and a hide() that lands earlier is undone
by that show(); this was the first-attempt failure on the macOS shard.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(e2e): check the playback display lock is released while hidden
Review follow-ups for #1724:
- Add an E2E that plays the webm fixture in the unautomated launch,
records the main process's prevent-display-sleep blockers, and asserts
the keep-awake lock is taken while visible, released when the window is
hidden, and taken again when it is shown. The visibility tests alone
would still pass if the renderer gate or the bridge stopped updating
powerSaveBlocker.
- Validate CDP replies before dispatch (CodeQL
js/unvalidated-dynamic-method-call): only a numeric id with a pending
settle function is called.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(e2e): skip killing an Electron that already exited
stopElectron now checks the recorded exit state before each signal and
tolerates a kill that races the exit: on Windows taskkill throws for a PID
that no longer exists. Startup cleanup can no longer replace the startup
error that explains the failure; a cleanup failure there is logged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(stalker): keep the watchdog cadence while the window is hidden
With background throttling on, Chromium wakes a hidden, silent page's
timers at most once per minute after five minutes, so a portal that asks
for get_events every 30 s would see pings at half its cadence while the
window is minimized. Tick the watchdog from a dedicated worker
(createBackgroundInterval, an inline blob worker allowed by the renderer
CSP), whose timers are not subject to page throttling. It falls back to a
page setInterval where no worker is available or the worker fails to load.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(stalker): keep a stopped watchdog interval stopped
A worker error that arrived after stop() started the page fallback
interval, which nothing cleared, so pings continued for an inactive
playlist. stop() now marks the interval stopped, detaches the worker
handlers, and the fallback refuses to start afterwards.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: 4gray <fourgray@proton.me>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* test(e2e): explain black canvases in the packaged frame-copy smoke
When the smoke's canvas stays black, print and attach the session
snapshot (with mpv's drop counter), every slot of the helper's
shared-memory frame rings read from /dev/shm, and the session's verbose
mpv log (log-file). Together they separate these cases: the helper never
published a frame, mpv rendered black, or the preload pump did not draw
a real frame.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(playback): stop random black frames on software GL in frame-copy
mpv's LUT scalers (the default lanczos/spline family) build their weight
texture from an uninitialized talloc_array: mp_compute_lut fills only
kernel->size of every stride floats, so each row's padding is heap
garbage (reinit_scaler in video/out/gpu/video.c, mpv 0.41 and master).
The LUT is an rgba16f texture sampled with linear filtering. When the
padding holds NaN, or a value large enough to become Inf as half-float,
Mesa's llvmpipe carries it through the zero-weight neighbour texel, the
weights become NaN and the frame clamps to pure black. Whether it happens
depends on heap contents, so the packaged smoke's paused frame was black
in most attempts on the CI runner.
Evidence from CI (8 repeats each, no retries): baseline 2/8 passed;
LUT-free bilinear scalers 8/8; gpu-dumb-mode 8/8; LP_NUM_THREADS=1 6/8.
A synchronous glReadPixels right after mpv_render_context_render()
showed the black frame already in the helper's FBO. The readback, the
shared-memory ring and the preload pump were correct.
On a CPU rasterizer the helper now selects mpv's LUT-free bilinear
scalers and disables sigmoid upscaling. Software GL cannot afford the
LUT scalers anyway. A session option for the same key still wins.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix(playback): report software scaler results and scope smoke ring diagnostics
The software-renderer fallback now reports which scaler options mpv
accepted and which it rejected (with mpv's error), instead of always
printing success. The smoke's black-canvas diagnostics read only the
failed session's rings (<sessionId>-g<N>), not every impv-fc ring in
/dev/shm.
Addresses Greptile review feedback.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: 4gray <fourgray@proton.me>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* test(performance): count startup phases and SQL statements for the J1 launch journey
Implements plan item A2. With IPTVNATOR_PERF_CAPTURE=1 the main process
keeps named counters and registers a main-only performance:read-counters
IPC handler; without the flag nothing is counted and the handler does not
exist.
- debug-trace.ts owns the registry; traceStartupPhase replaces the
trace('startup', ...) sites and counts main.startupPhases.
- The database worker counts executed statements through better-sqlite3's
Statement prototype (the verbose callback expands every statement and
made bulk inserts 2-4x slower) and posts the count over its message
port, flushed before every other worker message. The main-thread shared
connection is counted through a new connection observer in the shared
database library.
- The first main window freezes main.modulesRegisteredBeforeWindow at
creation and main.sqlStatementsBeforeReadyToShow at ready-to-show.
- The journey gate drops the ready-to-show that Electron emits for the
about:blank detour, so the app sees the real document's first paint,
and taps the counters handler; the J1 record reads both counters after
the renderer probe completes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(database): require one SQL statement per exec during initialization
The performance capture counts one exec call as one statement, because
SQL cannot be split reliably in the counter (trigger bodies contain
semicolons). The historical-upgrade driver now wraps exec on every
connection initDatabase opens and fails on a batch, so that counting
assumption holds for the fresh profile and all historical schemas.
Documents the definition in the counter and the architecture docs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* test(performance): count SQL statements only for the launch journey
Codex review: the M3U import, refresh-cancellation and Xtream benchmarks
also run with IPTVNATOR_PERF_CAPTURE=1, so the statement hook wrapped
every row of their bulk inserts and changed what they measure.
SQL counting now also needs IPTVNATOR_PERF_COUNT_SQL=1, which only the
launch journey sets; a harness test fails if another source sets it.
Startup phases, the window snapshot and the read handler stay on the
capture flag. Without SQL counting no ready-to-show listener is attached,
so a zero is never reported for statements nobody counted.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: 4gray <fourgray@proton.me>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Registering IPC handlers costs 0.4 ms; evaluating the modules behind them (axios, drizzle-orm, better-sqlite3, electron-updater, fix-path) before the window could load was the real cost. main.ts now keeps only the pre-paint wiring and loads the rest as the deferred-events.js chunk inside the main window's did-start-loading listener, where the import and its registrations complete before any renderer invoke can arrive.
Interleaved A/B on the performance build: app.whenReady 323 -> 265 ms, did-finish-load 499 -> 447 ms; J1 journey spawnToDidFinishLoad ~405 -> ~365 ms with identical counters. Packaging ships the chunk explicitly, verify:package-layout requires it, and the benchmark build identity hashes it.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
main.js is now a small entry that enables Node's on-disk V8 compile cache under userData/v8-compile-cache and then requires the application bundle main.app.js. Warm launches reach app.whenReady about 13 ms sooner at the median; IPTVNATOR_DISABLE_COMPILE_CACHE=1 turns it off and IPTVNATOR_COMPILE_CACHE_DIR relocates it.
Packaging now ships main.app.js explicitly and verify:package-layout requires the main-process entry files in app.asar, because nx-electron copies the backend through an allowlist. The Xtream benchmark build identity hashes both the launcher and the bundle.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* test(electron): reap Windows process trees and retain smoke diagnostics
* test(electron): require confirmed process exit before relaunch
* test(electron): retain process handle after application exit
* test(electron): bind captured processes to application instances
