Evidence (host log, production 0.7.547, forge probe-dnlf).
18:23:10Z forge is up … restarts=5 unclean=5
18:23:41Z child forge stopped name=probe-dnlf why="used 2.1 cores for 31s, over its limit of 1; the busiest process is bwrap (pid 2996149, 35s of CPU)"
18:23:42Z crew is up …
18:23:42Z forge is up … restarts=6 unclean=6
The person saw: builder "working now" at 49 s, then "stopped twice by its forge restarting; not tried again"; no pull was opened; the app stayed the template. Credit on home fell from $2 to $1.02 in three minutes while the app page said "Spent $0.00".
So on 0.7.547, with #11628 merged ("a stop only over twice its cores for five minutes"), a hosted forge was still stopped at 2.1× after 31 seconds. Either the path that stops a hosted forge (child forge stopped) does not use #11628's function, or its window is not the five minutes the title says. Six unclean restarts in the forge's first minutes, and the busiest process is the crew's own sandbox (bwrap): the build is what gets it killed.
What.
Find which code produced that stop and make it the one rule from #11628: no stop before five minutes of sustained use above twice the cores, never in a forge's first ten minutes, and never while the only thing running is the forge's own crew. Add the test that would have caught this: a 1-CPU cell whose sandboxed job uses 2.1 cores for two minutes is not stopped.
Actually hold a cell to its cores without killing it: a cgroup CPU quota on the cell's process tree where the host allows it, or CPU affinity to its cores; then a 1-CPU forge builds slowly instead of dying. Say on the forge's Settings what it has ("1 CPU: builds run, slower").
A run interrupted by the host stopping or restarting the forge is not the run's fault: it is queued again and tried (bounded, say five times with backoff), and the rule from #11585 ("interrupted twice in a row is failed") counts only failures of the run itself. When it does give up, the person sees one line saying why and a Try again button; "not tried again" with no action is a dead end.
The credit figure: home said $1.02 of $2 spent in three minutes with nothing built, while the app page said $0.00. An interrupted run's spend is counted once and shown the same everywhere (#11591's work; check it covers this case).
Why did the forge restart five times before the build even began (restarts=5 unclean=5 at 18:23:10, seconds after it was made)? Find and fix that too.
Acceptance. The stranger journey passes on a small forge: first ask → pull → landed → the app's own address shows the change, with zero child forge stopped lines and unclean=0 for that forge in the host log; paste the lines. Then the free plan's size goes back to small. Do not close this until that production probe passes.
Evidence (host log, production 0.7.547, forge probe-dnlf).
The person saw: builder "working now" at 49 s, then "stopped twice by its forge restarting; not tried again"; no pull was opened; the app stayed the template. Credit on home fell from $2 to $1.02 in three minutes while the app page said "Spent $0.00".
So on 0.7.547, with #11628 merged ("a stop only over twice its cores for five minutes"), a hosted forge was still stopped at 2.1× after 31 seconds. Either the path that stops a hosted forge (
child forge stopped) does not use #11628's function, or its window is not the five minutes the title says. Six unclean restarts in the forge's first minutes, and the busiest process is the crew's own sandbox (bwrap): the build is what gets it killed.What.
restarts=5 unclean=5at 18:23:10, seconds after it was made)? Find and fix that too.Acceptance. The stranger journey passes on a small forge: first ask → pull → landed → the app's own address shows the change, with zero
child forge stoppedlines andunclean=0for that forge in the host log; paste the lines. Then the free plan's size goes back to small. Do not close this until that production probe passes.