mirror of
https://github.com/software-mansion/argent.git
synced 2026-09-14 19:27:14 +08:00
Ignacy Łątka
b165bc39ee
## Problem
The **Boot + drive AVD on headless Weston** job's *"Screenshot returns
real pixels"* step takes a single screenshot 15s after `boot_completed`
and asserts it is > 20 KB. That single shot races **two** independent
warm-up delays:
1. SurfaceFlinger painting the lockscreen UI, and
2. the simulator-server's screenshot stream, which only starts warming
on the **first capture request**.
On slow/contended runners the one-and-only capture comes back as either:
- `{"error":"...no image to export"}` — a cold stream, response has no
`.data` envelope, so the step's `json.load(...)['data']` raises
`KeyError: 'data'`; or
- an all-zero framebuffer that PNG-compresses to ~3–7 KB and fails the
`> 20 KB` check.
Either way the step fails even though the boot was healthy. This is
intermittent — the same commit passes most runs and fails some.
## Fix
Replace the single fixed `sleep 15` + one capture with a **bounded
poll** (15 attempts, 4s apart). Each request also warms the stream, and
the loop exits as soon as a frame exceeds 20 KB. A cold-stream error
response is tolerated and retried rather than aborting the step.
Genuinely broken framebuffers still fail, now with per-attempt
diagnostics in the log.
- Healthy boots pass in the first attempt or two (no slower than before
in the common case).
- Worst-case added headroom is ~60s, only spent when the framebuffer
never warms — which is a real failure worth surfacing.
## Verification
The workflow file is unchanged elsewhere; only this step's `run:` script
changed.
- `prettier --check` passes on the workflow file.
- Extracted the exact `run:` script and ran it against a mock `curl`
covering: cold-error → all-zero frame → real frame (exits 0 on the real
frame); always-cold (exits 1 after 15 attempts); always all-zero
framebuffer (exits 1). Exit codes confirmed 0 / 1 respectively.