Two consecutive Windows E2E runs hung inside dependency install until
the job timeout: 31449642122 65 minutes in uv pip install .[all],
31453853006 43 minutes in the SQLite-repair uv sync (RUST_LOG=uv=debug
made THAT run hang earlier and with more stderr - the tell).
Root cause: scripts/desktop-update.ps1 redirects both child pipes but
only pumps stdout while the child runs; stderr is ReadToEnd()'d after
exit. uv and pip write progress to stderr. Once that pipe hits the
~64KB buffer, uv blocks on write, hermes update blocks on uv, the
hand-off blocks on hermes update: deadlock. Slower stderr producers
survive by finishing before the buffer fills, which is why the linux
sandbox never sees this.
Fix both sides of the class:
- managed_uv.py candidate sync + main.py _run_install_with_heartbeat:
merge stderr into stdout (the pipe that IS drained). This arm heals
EXISTING installs, whose old hand-off script drives the NEW python
after the git reset.
- desktop-update.ps1: drain stderr concurrently via ReadToEndAsync so
future bases never block regardless of what a child writes there.
_run_logged_subprocess and _run_npm_install_deterministic already
merge or capture both pipes; the two fixed sites were the only update-
path spawns that redirect stderr without draining it live.
ethernet
e3d22b5b29