| name | swarm:continue |
| description | Resume the most recent paused swarm run. Reconciles the saved state.yml against disk + PM ground truth, surfaces any drift for confirmation, then re-enters the orchestration loop. Resilient to clean TERMINAL_PAUSE and to crashes/reboots/network loss. |
| user-invocable | true |
| argument-hint | [no args — resumes the run named by .agent-tools/swarm/active-run] |
Resume Swarm (/swarm:continue)
/swarm:continue resumes a paused or interrupted /swarm <goal> run. It is resilient to all
exit reasons — not just a clean TERMINAL_PAUSE, but also machine reboot, network loss,
terminal crash, or power outage, all of which leave state.yml at whatever was last
atomically written.
The governing principle: never trust state.yml as ground truth. It is a hint about
intended state at last write. Actual truth is on disk (worktrees, branches, commits,
session-state.md files) and in the PM tool. /swarm:continue always reconciles before
resuming.
User Input
$ARGUMENTS
/swarm:continue takes no arguments; it resumes the run named by active-run.
Procedure
1. Locate the run
Read .agent-tools/swarm/active-run.
- Absent → inform the user there is no active run; suggest
/swarm <goal>. Exit.
- Present → load
.agent-tools/swarm/sessions/<run-id>/state.yml.
- If
state.yml is missing or unreadable → surface it; offer to clear the active-run
pointer. Exit.
- If an orphan
state.yml.tmp exists (crash mid-write) → remove it; state.yml retains the
last good state.
2. Announce recovery
Recovering — last known state from <state.yml.last_updated>. Re-classifying from project
state…
3. Reconcile against ground truth
Re-classify every item not in stage merged from disk + PM, using
references/classification-rules.md (the same classification logic the orchestrator uses).
Do not assume state.yml stages are correct.
4. Surface drift
Compare the reconciled classification against state.yml and present every difference, e.g.:
- "AER-101: state.yml had
implementing; re-classified as implemented (commits past plan
present, no review yet)."
- "AER-110: state.yml had
reviewing; worktree missing → re-classified as planned (possible
manual cleanup)."
Special cases:
- Goal already complete (user merged everything manually) → report GOAL_COMPLETE, clear
active-run, exit.
- Goal no longer achievable (issues closed externally without merge) → surface and ask the
user how to proceed.
5. Confirm
Display the reconciliation summary and ask the user to confirm or correct before resuming. Do
not silently overwrite their intent.
6. Persist & resume
Write the reconciled classification to state.yml (atomically — temp + rename). Clear any
stale last_handoff.
Explicit charter load (resuming swarm run): Read the charter files before re-entering:
- Read
.agent-tools/charter/charter.md
- Read
.agent-tools/charter/project.md
- Read
.agent-tools/charter/engineering.md
- Read
.agent-tools/charter/workflow.md
Then re-enter the orchestrator main loop at the between-wave merge
sweep (Phase 3(A) of /swarm), which is the safe re-entry point.
Notes
/swarm:continue shares all orchestration mechanics with /swarm <goal> — see the swarm
SKILL.md and references/dispatch-mechanics.md. It differs only in the reconciliation
preamble above.- It never pushes to remote and never trusts cached state. Same safety constraints as the
orchestrator.
