hive.terminal-tools-job-control

Start it with terminal_job_start to get a job_id, then poll terminal_job_logs with since_offset set to the previous call's next_offset. Repeat until status is exited, or block once with wait_until_exit=True and a timeout.

Send signal_int first via terminal_job_manage, wait 2-5 seconds, then escalate to signal_term, and finally signal_kill only if the process is unresponsive. Check exit status between steps with terminal_job_logs using wait_until_exit.

The process moves into the JobManager with output flowing into the ring buffer and you receive a job_id. Take over by calling terminal_job_logs with tail=True for recent output, or since_offset=0 with wait_until_exit to capture everything.

Job output lives in a 4 MB ring buffer per stream, so slow polling lets old bytes get evicted. When truncated_bytes_dropped is non-zero, poll more frequently or accept the gap and continue reading from next_offset.

No, jobs die when the terminal-tools server restarts and there is no re-attach mechanism. For durability, launch processes with nohup via terminal_exec to detach them into the system process tree and track the PID yourself.

Use merge_stderr=True when logs are meant to be read together with ordering preserved, which suits most servers and build tools. Keep streams separate when stderr is purely errors and stdout is data you process differently.