Herdr tasks

Read the master first if needed: ls-herdr-task-routing

Use localsetup herdr task --help. Load layout only when allocating or changing the work area: ls-herdr-pane-layout

Run a command and finish the job

For a finite command in a new owned pane:

localsetup herdr --workspace "$HERDR_WORKSPACE_ID" --pane "$HERDR_PANE_ID" \
  --state-dir "$task_state" task run --cwd "$PWD" --wait 5 -- \
  python3 -c 'print("ready")'

Replace the example argv with the requested command. task_state is the existing task's absolute private state directory. Continue with returned result.task_id using task status <task-id> or task wait <task-id>, then task cleanup <task-id> after a confirmed terminal receipt. Cleanup preserves the actual exit outcome; it is different from cancellation.

1. Resolve the explicit target and working directory. Use an owned, available shell pane; never type into a busy shell, editor, approval dialog, or agent.
2. Submit the literal argument vector through the task runner. Preserve the operation ID; do not replace it when a wait times out.
3. Use status/wait to obtain the operation-specific exit receipt and bounded output. A timeout means unfinished or unknown work, not permission to rerun.
4. Report the command's actual outcome. Clean up the owned pane when the work is complete unless a continuing process or visible result is intended.

Native pane.run acknowledges queue admission. Output matching can match old text. An apparent shell prompt does not prove the submitted command finished. The shared runner's receipt binds the operation and its exit status; use it. Never replay after connection loss until read-only reconciliation proves what happened.

Logs, servers, and long operations

Keep a requested server or log follower in an owned pane, report its handle, and leave it running when appropriate. Prefer bounded recent unwrapped output for readable logs. Wait on native events or operation records rather than writing an ad hoc busy polling loop. Refresh a snapshot after event loss.

Cancel only the exact owned operation when interruption is requested or already covered by the task. Do not send Ctrl+C to a guessed pane, kill an entire Herdr server, or treat disconnect as cancellation.

Requested agent work

Pane layout and recognized-agent lifecycle are different. Native agent start uses an existing available pane; it does not create one. Use returned agent names/IDs, supported native agent operations, and the user's chosen client. Do not select a model, expand delegation authority, bypass Plan Mode, or answer another agent's approval/question UI on the user's behalf.

Native agent prompt --wait observes semantic states. idle, done, and blocked are not shell exit codes or proof that every background task finished. Read the relevant result and validate the requested deliverable. A stalled or timed-out prompt may have been delivered; inspect before resubmitting.