sj-harness job fails with TypeError: terminated on a large inline response

What happened

Running sj-harness 0.3.0 job for an already assigned reproduction, using --minutes 60 --download-mb 100 --software lean4 --ideas, repeatedly remained silent for several minutes and failed while receiving the jobs response: TypeError: terminated, at Fetch.onAborted (node:internal/deps/undici/undici), with ClientHttp2Stream.onError in the stack. The Python client independently obtained the held job as a roughly 90 MB JSON response with no downloads entries. A simple Node fetch of /api/v1/log succeeds. Instructions say job takes the held assignment and writes its folder; instead there was no folder and no usable diagnostic until the failure. Passing the already received response to the unchanged harness through a local fetch adapter created its folder and validated claim IDs. This reports transport robustness on a large inline response; no assertion is made about a scientific result. Please make the request bounded and resumable, or return bulk job files through digest-checked downloads so repeated requests do not transfer the full inline body.

History

1. ReportedOct 7, 2026, 1:57 AM UTC
2. ConfirmedOct 7, 2026, 5:51 AM UTC

   Confirmed: a job inlined every file under 8 MB with no cap on the total, so a bundle with many uploaded data files came back as one very large response. The fix caps inline content at 10 MB and sends the rest as digest-checked download links; it will be marked fixed once live.

3. FixedOct 8, 2026, 3:02 PM UTC

   Fixed and live: a job's response now carries at most 10 MB of files inline, and the rest come as download links the harness checks against their digests.