Pause and Resume
Keep this workflow in the current session. Store its state in the pause
response; do not create files, commits, or new tasks solely to checkpoint it.
Pause
- Stop expanding the task. Do not begin another subtask, retry, or optional
check.
- Let an already-started atomic action finish only when interruption would
leave inconsistent state. Otherwise stop at the nearest safe boundary. Use
only the smallest necessary read-only check or already-authorized cleanup to
record an accurate, safe state.
- Preserve facts, not guesses. Distinguish completed, in progress, blocked,
and unknown work. Record exact non-sensitive paths, parameters, job IDs, and
receipts when they are needed to continue. Never expose secrets.
- If work continues independently, record its handle, observed status, and
how to reconcile it. Do not cancel, restart, or dispatch it again merely to
pause.
- End the turn with the compact checkpoint below, in the user's language.
Keep the marker exact, omit irrelevant detail, and state explicitly when
there is no known non-repeatable effect.
PAUSE_CHECKPOINT: ACTIVE
- Goal and done condition:
- Completed:
- Current stopping point:
- Remaining work:
- First action on resume:
- Do not repeat:
- Constraints and settled decisions:
- Live state, running work, and facts to recheck:
Finish with the equivalent of: Paused. When you return, say "continue". Then
stop. Do not perform the first resume action in the pause turn.
If there is no unfinished task, say so and stop without emitting an active
checkpoint.
Resume
When the user later asks to continue the paused work:
- Use the newest unconsumed
PAUSE_CHECKPOINT: ACTIVE in this session. A
checkpoint is consumed once substantive work has resumed from it. If none
exists, interpret the request from the immediate conversation normally.
- Apply later user instructions over the checkpoint. Recheck only mutable or
high-risk state that could have changed while paused, especially files,
worktree state, running jobs, and external side effects.
- Reconcile recorded in-flight work before starting replacements. Never repeat
an item under
Do not repeat unless current evidence proves it did not occur
and repeating it is authorized.
- Resume directly from
First action on resume, then continue the recorded
remaining work. Do not ask the user to restate known context or produce a new
plan unless material drift makes a decision necessary.
This workflow cannot recover a killed process, missing conversation, or a
different session; use a handoff or persistent checkpoint workflow for those
cases.
1---2name: pause-and-resume3description: Cooperatively pause unfinished work in the current session, leave a precise continuation checkpoint, and resume from it when the user later says continue. Use only when the user explicitly requests this pause workflow. Do not use for crashes, app shutdown, new-session handoffs, or ordinary task summaries.4---5
6# Pause and Resume
7
8Keep this workflow in the current session. Store its state in the pause
9response; do not create files, commits, or new tasks solely to checkpoint it.
10
11## Pause
12
131. Stop expanding the task. Do not begin another subtask, retry, or optional
14 check.
152. Let an already-started atomic action finish only when interruption would
16 leave inconsistent state. Otherwise stop at the nearest safe boundary. Use
17 only the smallest necessary read-only check or already-authorized cleanup to
18 record an accurate, safe state.
193. Preserve facts, not guesses. Distinguish completed, in progress, blocked,
20 and unknown work. Record exact non-sensitive paths, parameters, job IDs, and
21 receipts when they are needed to continue. Never expose secrets.
224. If work continues independently, record its handle, observed status, and
23 how to reconcile it. Do not cancel, restart, or dispatch it again merely to
24 pause.
255. End the turn with the compact checkpoint below, in the user's language.
26 Keep the marker exact, omit irrelevant detail, and state explicitly when
27 there is no known non-repeatable effect.
28
29```text
30PAUSE_CHECKPOINT: ACTIVE
31- Goal and done condition:
32- Completed:
33- Current stopping point:
34- Remaining work:
35- First action on resume:
36- Do not repeat:
37- Constraints and settled decisions:
38- Live state, running work, and facts to recheck:
39```
40
41Finish with the equivalent of: `Paused. When you return, say "continue".` Then
42stop. Do not perform the first resume action in the pause turn.
43
44If there is no unfinished task, say so and stop without emitting an active
45checkpoint.
46
47## Resume
48
49When the user later asks to continue the paused work:
50
511. Use the newest unconsumed `PAUSE_CHECKPOINT: ACTIVE` in this session. A
52 checkpoint is consumed once substantive work has resumed from it. If none
53 exists, interpret the request from the immediate conversation normally.
542. Apply later user instructions over the checkpoint. Recheck only mutable or
55 high-risk state that could have changed while paused, especially files,
56 worktree state, running jobs, and external side effects.
573. Reconcile recorded in-flight work before starting replacements. Never repeat
58 an item under `Do not repeat` unless current evidence proves it did not occur
59 and repeating it is authorized.
604. Resume directly from `First action on resume`, then continue the recorded
61 remaining work. Do not ask the user to restate known context or produce a new
62 plan unless material drift makes a decision necessary.
63
64This workflow cannot recover a killed process, missing conversation, or a
65different session; use a handoff or persistent checkpoint workflow for those
66cases.
majiayu000