| name | overnight |
| description | Long-running mode with file checkpointing + session resumption. `cmax overnight <spec> --budget-credits N` runs until the spec is done or the budget is exhausted, snapshotting on every turn so a SIGTERM or crash doesn't lose work. |
/overnight — long-running mode
For multi-hour or multi-day autonomous runs. Combines /goal with file checkpointing and session resumption.
Mechanics
- Runs /goal loop with
enableFileCheckpointing: true.
- Snapshots session ID + cumulative credit usage to
.claudemax/state/overnight/<slug>.checkpoint.json after every meaningful turn.
- On crash / SIGTERM / max-turns hit, re-running
cmax overnight <spec> picks up from the last checkpoint.
- Hard budget cap via
--budget-credits N (Agent SDK credit, USD). Run terminates when budget hit.
CLI
cmax overnight SPEC.md --budget-credits 50
cmax overnight SPEC.md --budget-credits 50 --max-turns 500
When to use
- A multi-hour migration you want to start before bed.
- A big test suite that takes hours to debug.
- Any /goal run where you'd rather come back to results than babysit turns.
When NOT to use
- Quick tasks — overnight overhead (checkpointing) costs you turn latency.
- Tasks where you genuinely need to babysit (security-sensitive, irreversible) — use plain /goal instead.
What's preserved across restarts
- Session ID (Claude Code resume).
- Cumulative credit usage.
- File checkpoints (SDK feature; per-file rewind possible).
What's NOT preserved
- In-memory orchestrator state. /overnight assumes the SPEC is the source of truth — re-reading SPEC.md + the session is enough to resume.
