Skip to content

A local video studio for your coding agent. Describe a video in one sentence: your agent directs, your machine renders. Claude Code, Codex, Cursor, Devin and more. No API keys, no uploads.

ci: the HTML export's real-time checks skip on starved CI runners; two files at a time in every shard

Favio Vázquez106473292b84
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index d7ae0d7..d225973 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -2,7 +2,7 @@ name: ci
# Two tiers (the quick release checks are checks.yml):
# every push / pull request : the fast test suite, split into 3 parallel shards per OS image
-# (run_all.py --shard I/3, -j auto inside each), after a core-tier setup;
+# (run_all.py --shard I/3, -j 2 inside each), after a core-tier setup;
# shard 1 also runs a 2 s 360p render and the path-with-spaces shims
# nightly / manual : the full test suite (voice, ASR, longer renders)
# Changes that only touch docs, Markdown, README art or the site do not run it.
@@ -129,13 +129,15 @@ jobs:
Pop-Location
python -c "import json,sys; d = json.load(open(sys.argv[1], encoding='utf-8-sig')); assert d['ok'], d['counts']" (Join-Path $dest 'showtime\doctor.json')
- - name: Fast test suite, shard ${{ matrix.shard }}/3 (-j auto)
+ # -j 2: two files at a time on every image (auto would pick 1 on the 3-core Apple Silicon runner,
+ # 2 on the 4-core Linux, Windows and Intel Mac runners); the shards are balanced for 2
+ - name: Fast test suite, shard ${{ matrix.shard }}/3 (-j 2)
shell: bash
env:
ONLY: ${{ inputs.tests }}
run: |
only=(); for t in $ONLY; do only+=(-k "$t"); done
- python skills/showtime/tests/run_all.py --fast -j auto --shard "${{ matrix.shard }}/3" "${only[@]}"
+ python skills/showtime/tests/run_all.py --fast -j 2 --shard "${{ matrix.shard }}/3" "${only[@]}"
- name: 2 s 360p render smoke
if: matrix.shard == 1
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 95b50e0..9641846 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -91,14 +91,19 @@ All notable changes to showtime. Each entry says what changed and why, so this f
- **CI runs the fast suite in three shards per OS.** `run_all.py --shard I/N` runs one of N parts of the
test files, split by fixed per-file weights (`SHARD_WEIGHTS`, fast-suite seconds from a CI run; longest
first into the lightest part), so every machine computes the same split and each file runs exactly
- once; `-j auto` still runs files in parallel inside a shard. `ci.yml` runs shards 1/3, 2/3 and 3/3 as
+ once; inside a shard two files run at a time (`-j 2`, also on the 3-core Apple Silicon runner, where
+ `auto` would pick 1). `ci.yml` runs shards 1/3, 2/3 and 3/3 as
separate jobs on Ubuntu, Windows, Apple Silicon and, on the public repository, the Intel Mac
(`macos-15-intel`: the `macos-13` image is retired); shard 1 also runs the render smoke and the
path-with-spaces shims. On one 2-core runner the suite took 23 to 31 minutes per OS.
- **The HTML export test runs alone** (`SERIAL` in `run_all.py`). Its player checks watch playback in real
time; on the public 4-core runners, sharing the machine with another file's Chrome left the player
drawing 1 to 4 frames in the 1.5 s window on Windows and the Intel Mac. A file in `SERIAL` counts double
- when the shards are balanced.
+ when the shards are balanced. Running alone was not enough on the Windows image (1 frame drawn while the
+ clock ran on), so on a CI runner (`CI` set) those real-time checks skip with the reason when the browser
+ draws fewer than 5 frames in the window or the sound is still loading, as they already did on 2 cores;
+ everything before them (requests, errors, duration, soundtrack) still runs, and on a desktop they
+ always run.
### Changed: the examples have their own repository
diff --git a/skills/showtime/tests/test_export.py b/skills/showtime/tests/test_export.py
index 5a562b9..6c15309 100644
--- a/skills/showtime/tests/test_export.py
+++ b/skills/showtime/tests/test_export.py
@@ -557,19 +557,23 @@ class ExportTest(unittest.TestCase):
def too_slow(r):
"""Why this browser cannot be watched in real time, or None. A 2-core runner drawing 1080p with a
software GPU can take 10 s to get through the driver's 1.5 s window (the video plays on in real
- time meanwhile), or draw only a few frames in it; a player bug shows up on a normal machine."""
+ time meanwhile), or draw only a few frames in it; so can a 4-core CI runner (the Windows image drew
+ 1 frame in the window while the clock ran on). A player bug shows up on a normal machine: outside
+ CI, on more than 2 cores, these checks always run."""
ms = r.get("windowMs") or 0
if ms > 3000:
return "the click and the driver's 1.5 s playback window took %.1f s" % (ms / 1000)
cores = os.cpu_count() or 1
- if cores > 2:
+ ci = bool(os.environ.get("CI"))
+ if cores > 2 and not ci:
return None
+ where = "a %d-core %s" % (cores, "CI runner" if ci else "machine")
n = len(r.get("startFrames") or [])
if n < 5:
- return "the player drew %d frames in 1.5 s on a %d-core machine" % (n, cores)
+ return "the player drew %d frames in 1.5 s on %s" % (n, where)
a = (r.get("after") or {}).get("audio") or {}
if a and (a.get("readyState") or 0) < 2:
- return "the sound was still loading (readyState %s) after 1.5 s on a %d-core machine" % (a.get("readyState"), cores)
+ return "the sound was still loading (readyState %s) after 1.5 s on %s" % (a.get("readyState"), where)
return None
@staticmethod