SDK 0.4.39: follow managed jobs

0.4.39 - 2026-10-11

• Follow managed jobs with held reads instead of polling (gateway 0.9.72
  serves job reads from the node without entering the guest):
  • get_job(sandbox_id, job_id, *, wait_seconds=None, after=None) waits (at
    most 25 s) until the record's cursor differs from after or the job ends
    (?wait_ms=&after=).
  • read_job_logs(..., wait_seconds=None) waits until the stream has bytes
    beyond offset or the job ends (&wait_ms=). SandboxJobLogChunk.job is
    the job's record as of the read (None from older gateways).
  • SandboxJobRecord.cursor: "<sequence>.<stdout_bytes>.<stderr_bytes>",
    the server's value or computed locally; cursor_from_server says which.
  • SandboxJobRecord.retry_after and SandboxJobLogChunk.retry_after: the
    answer's Retry-After in seconds (a parked sandbox), else None. Neither
    these nor cursor_from_server take part in equality.
  • New job_read_delay(read, *, requested_seconds, elapsed_seconds, poll_seconds=1.0, after=None, record=None) -> float | None paces a custom
    follower (for example one driving stdout and stderr together): None when
    the job is terminal and the stream read, Retry-After when set,
    poll_seconds after an early answer without progress (or a status answer
    without a server cursor), else 0. Raises JobOutputTruncatedError for a
    log read whose record shows truncation. wait() and follow() use it.
  • JobHandle.wait() / AsyncJobHandle.wait() long-poll the status with
    after=<cursor>, paced by job_read_delay.
  • New JobHandle.follow(stream="stdout", offset=0, *, limit=1 MiB, poll_seconds=1.0) (a generator; an async iterator on AsyncJobHandle)
    yields output through log long-polls. It ends when the carried record is
    terminal and the offset equals the stream's byte count, and raises the new
    JobOutputTruncatedError (a SandboxApiError) when either stream was
    truncated.
  • A waiting read's request timeout is the client timeout plus the wait.
    Older gateways ignore wait_ms and are polled every poll_seconds, as
    before.