tmux compact interrupt=trueStatus, 2026-11-21: implemented as
interrupt=auto|true|falseonfix/compact-interrupt(hive-tmux.claude.interrupt). The consent gate below was replaced by a screen gate: auto cuts in only on a /goal pane afterinterrupt_after_mswhile a read-only wait or no tool runs, never mid-write or over a draft. The defaultinterrupt_via=force-send(queued /compact + C-x C-s) backgrounds the running tool instead of killing it. Measured: docs/assay/compact-interrupt-2026-11-21.md. The text below is the original note, kept for its reasoning.
Card 20261120021807-1a6075cf, follow-up of 2026-11-20 02:35. Not implemented in feat/tmux-job-cancel.
tmux compact waits for rest before it types: a settled idle, or on a
/goal pane the single turn boundary between two turns. A pane whose one
/goal turn never ends gives it neither. %5 ran one turn for 2h15m, up to
509k tokens. The job then times out after :idle-timeout-ms (30 min) and the
context keeps growing. Cancel and supersede (this branch) free the slot, but
they do not make a boundary.
swarm tmux compact pane=<ref> interrupt=true, explicit pane only (never in
a sweep, the same rule as supersede).
interrupt_after_ms.:vim-mode on the row).
Record it as the step :interrupt. It is a keystroke, so it is claimed
through claim-typing and the job can no longer be cancelled.poll/wait-turn-boundary), then run
the normal compact, re-seed and goal restore.:start-timeout-ms of the interrupt: fail :timeout :during :interrupt. Never send a second interrupt.An interrupt cuts a running tool call: a half-written file, a commit in flight, a migration. So it is refused unless consent is evidenced:
A refusal names the gate and how to satisfy it. The decision is a pure function over the row, the hand-off evidence and the grant. Only the boundary reads the hive-mind and sends the key.
This branch only takes slots and never types more than before: cancel and
supersede act only on jobs that have not typed. Interrupting is a new kind of
keystroke into a WORKING turn, which every job here is built never to do. It
needs its own consent model, a live-pane verification plan and review, and it
touches poll and the composer. That is not small, so it belongs on its own
card.
Measured on Claude Code 2.1.330: a message submitted while a turn runs is
drawn above the spinner as ❯ text (background 237) over
⎿ Queued · ctrl+x ctrl+s to send now, and the empty composer shows
Press up to edit queued messages. The message was SUBMITTED: it is not a
draft. screen/read-screen reports it under :queued, never :input.
The force-send interrupt sends the queue first with C-x C-s (step
:human-sent, job key :human-sent), so the human's intent is delivered,
never dropped, and then types /compact. A /compact already in the queue
goes out with that flush and is not typed twice. With a message queued, an
interrupt_via=escape job force-sends instead: Esc would cut the turn that
takes the message in. An unsent draft at the prompt still holds the job
(:prompt-holds-text).
A goal can be set without a <command-name>/goal transcript entry (session
15c7760e). Claude Code still writes {"type":"goal_status","condition":...}
each time it checks the goal, so panes/goal-text reads the newest entry of
either kind (a met goal is none). With neither, the job records
:goal-unreadable (step and job key): the goal will not be restored, and the
pane is compacted anyway.
Assay, queued-human scenario, n=20 per arm: force-send 20/20, escape 20/20, old job 0/20, 0 writes cut.
Can you improve this documentation?Edit on GitHub
cljdoc builds & hosts documentation for Clojure/Script libraries
| Ctrl+k | Jump to recent docs |
| ← | Move to previous article |
| → | Move to next article |
| Ctrl+/ | Jump to the search field |