The multi-line git commit message inside the run block dropped below the
literal block's indentation, ending the block mid-command and failing the
whole workflow file at parse time (0s failure, no jobs). Use two -m flags on
one shell line instead. YAML now validated locally before push.
Token included so this push performs the regeneration.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01LHH3wPBsoruwCaJo8Rpf6a
workflow_dispatch only works for workflows on the default branch, which this
file will not reach until the next promotion. Gate the job on an explicit
[update-visual-baselines] commit-message token instead so it is dispatchable
from unstable now; the workflow's own commit omits the token, so it cannot
retrigger itself.
This commit carries the token, so pushing it performs the first regeneration.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01LHH3wPBsoruwCaJo8Rpf6a
The visual-regression job pixel-compares tests/visual/baselines.json on
ubuntu-latest, but the committed baselines were captured on developer machines.
Chromium builds and font fallbacks differ between environments — the token
stack's "Inter"/"Fira Code" are not shipped or installed anywhere, so every
machine renders its own system fallback — and the kitchen-sink default-theme
diff drifted past MAX_DIFF_RATIO on CI while passing 13/13 locally. ci.yml
documents exactly this mismatch as the reason the job is quarantined from
blocking deploys.
The CI diff artifact confirms it: baseline and actual render identically except
for ghosted/offset text in the high-contrast lower half — font-metric drift,
not a design change. (The dark upper half diffs too little to trip the 0.2
pixelmatch threshold, which is why only the default theme fails.)
Fix the class of problem instead of the instance: capture baselines in the SAME
runner image, browser install and font stack that the comparison job uses.
workflow_dispatch only — regenerating baselines is an explicit decision made
when a visual change is intentional, never a side effect of a push. The
workflow re-runs the comparison against its own output before committing, so a
bad capture cannot land.
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01LHH3wPBsoruwCaJo8Rpf6a
