Repository navigation
test(web): 400 ms latency gate for send, thread switch and launch - #117
Conversation
Adds a latency vitest browser project run on production React, timed from the input event to the frame after the UI shows the result, median of five samples. Composer send is gated. Thread switch and mount to an open thread are measured and reported but not gated yet because they currently miss the budget. ARC CI installs headless Chromium unprivileged and runs the gate. Model: claude-opus-5-5 Task: implementation Agent: Claude Code (CLO-868 feature map lane) Lane: magpie Harness: t3/claude-code Model-family: Claude Model-generation: 5.5 Skills-SHA: df60fa7be83ccc7a8e3c64e3c58ba9ed7a84cfc4 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… own list The dry-run parse found nothing on the ARC runner, so the gate launched Chromium without libnspr4. The step now resolves Playwright's Ubuntu 24.04 Chromium packages directly and fails with the missing names if ldd still reports any. Model: claude-opus-5-5 Task: implementation Agent: Claude Code (CLO-868 feature map lane) Lane: magpie Harness: t3/claude-code Model-family: Claude Model-generation: 5.5 Skills-SHA: df60fa7be83ccc7a8e3c64e3c58ba9ed7a84cfc4 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…imports it Model: claude-opus-5-5 Task: implementation Agent: Claude Code (CLO-868 feature map lane) Lane: magpie Harness: t3/claude-code Model-family: Claude Model-generation: 5.5 Skills-SHA: df60fa7be83ccc7a8e3c64e3c58ba9ed7a84cfc4 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review found the done check ran a full-page text scan on every mutation inside the timed window, the end time was taken before the final paint, a missing input event hung until the test timeout, and the transcribe test name would also run in the dev-React browser project. Each frame now checks once and the end is read in a task after that frame paints; the result must not already show before the input, and an action without input fails at once. Latency tests are named *.latency.ts so only the latency project runs them, and the report goes to stdout through a browser command so CI keeps every sample. Measured on an idle ASUS: composer send about 66 ms and thread switch about 150 ms, both gated now; mount to an open thread about 350 ms, reported but not gated. Model: claude-opus-5-5 Task: implementation Agent: Claude Code (CLO-868 feature map lane) Lane: magpie Harness: t3/claude-code Model-family: Claude Model-generation: 5.5 Skills-SHA: df60fa7be83ccc7a8e3c64e3c58ba9ed7a84cfc4 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rst mount On the ARC runner the first mount's connection failed transiently and the thread never showed. The mount stayed in the page, so later tests found two copies of every thread. A thread that fails to open now unmounts, and the warm-up's first mount may fail before the measured mounts start. Model: claude-opus-5-5 Task: implementation Agent: Claude Code (CLO-868 feature map lane) Lane: magpie Harness: t3/claude-code Model-family: Claude Model-generation: 5.5 Skills-SHA: df60fa7be83ccc7a8e3c64e3c58ba9ed7a84cfc4 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The runner image has no fonts, so text laid out with no size and every visibility wait in the latency gate timed out, while the text-only app launch check passed. The step now extracts DejaVu and fontconfig, points Chromium at them with FONTCONFIG_FILE, and fails early when no font is found. Model: claude-opus-5-5 Task: implementation Agent: Claude Code (CLO-868 feature map lane) Lane: magpie Harness: t3/claude-code Model-family: Claude Model-generation: 5.5 Skills-SHA: df60fa7be83ccc7a8e3c64e3c58ba9ed7a84cfc4 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars.
|
The simulation wrote its log where the runner cannot, so the 400 ms latency gate died before running. Download-only is the path the voice gate already uses. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
fonts-dejavu-core depends on fonts-dejavu-mono, and resolving that dependency makes apt clean the root archive dir, which this runner cannot write. The library list now matches the voice step that already passes. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
{"reviewed_head":"3d4e72b99ed340d11709ba4b13daed4a484a664a","reviewer":"grok-4.7-build (claudeAgent, read-only)","findings":[]}The review of the CI step at 3d4e72b was clean. CI then failed again in "Gate interaction latency at 400 ms" with |
libfontconfig1 depends on fontconfig-config, which depends on fonts-dejavu-core and fonts-dejavu-mono. Resolving those sends apt back to the root archive dir, which this runner cannot write. The package list now matches the voice step that already passes. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
{"reviewed_head":"59067a6a6233061bc84b996cd9daf5fd853466fd","reviewer":"grok-4.7-build (claudeAgent, read-only)","findings":[]}Quick re-check of |
apt-get update runs APT::Update::Post-Invoke from the runner image, and that hook removes /var/cache/apt/archives/partial/*.deb, which this runner cannot write. Job 136517003866 died there before the 400 ms gate ran. Pointing Dir::Etc::Main and Dir::Etc::Parts at empty files in the runner temp dir drops those hooks while the download still uses the system source lists. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
apt applies -o Dir::Etc only after it has already read /etc/apt/apt.conf.d, so the runner image Post-Invoke still removes /var/cache/apt/archives/partial/*.deb and fails the 400 ms gate. APT_CONFIG is read first. Point it at an empty config under the job temp dir, keep the image dpkg status, and download only the Chromium libraries that image does not already have. Co-Authored-By: Claude Code <noreply@anthropic.com>
Kameron wants interactions under 400 ms. Nothing measured that, so slow paths such as the transcribe button's 1–3 s "starting" shipped unnoticed.
latencyvitest browser project (apps/web/vite.config.ts). Run it withvp run test:latencyinapps/web; it uses production React, and a warm-up test fails if it finds the development build.apps/web/test/interactionLatency.tssets a 400 ms budget and gates on the median of five samples.apps/web/test/app.latency.tsxmounts the whole app on the in-process fake environment. Medians on an idle ASUS:*.latency.ts(x), so only this project runs them. The transcribe press-to-recording test onfix/clo-919-transcribe-latencyjoins the gate once it is renamed tosrc/speech/voiceInput.latency.ts.lddstill finds a missing one. The package list is copied from Playwright 1.64.1's Ubuntu 24.04 table, so update it if the Playwright version or runner image changes.Not measured: desktop Electron cold launch. Earlier numbers from a heavily loaded ASUS (thread switch about 510 ms, launch about 1.5 s) came from host load and from a timing check that scanned the whole page on every DOM change. The check now runs once per frame.
Deploy: none
Model: claude-opus-5-5
Task: CLO-868 feature map lane (perf), CLO-919 item 4 gate
🤖 Generated with Claude Code