Continue without reconstructing everything.
Keep the next session’s starting point outside the conversation. Recover synchronization separately from native execution.
When a session is getting too large
Do not make reuse of a long conversation the automatic default. Decide whether the next outcome belongs in this session or a fresh bounded successor. Compaction can lose useful context, so retain the essentials in durable project records first. Automatic rotation and compaction detection are not implemented product features.
A practical handoff
Before an authorized successor starts, save:
- The goal, acceptance criteria and accountable owner.
- Decisions, constraints, permission limits and dependencies.
- The current source revision and exact files or worktree.
- What is done, failed or unknown, with tests and evidence.
- The single next action and any precise human/native approval wait.
- Predecessor and successor session identities, including device and provider.
Use short human-readable names in routine notes and retain exact identities in private evidence. Independently started successors can be linked by explicit or corroborated lineage; weak matches stay suggestions.
Manual workflow today. Versioned product handoffs, automatic task splitting and successor discovery policies remain implementation work. Do not force-stop or replace a session solely because of age, idle status, file timestamps or compaction.
Resolve before replacing
- Use the existing session’s supported stop/control if stopping is authorized and necessary.
- Check whether the original turn ended, remains active or has uncertain delivery.
- Preserve the original request and any receipts; do not blindly replay a launch.
- Start the authorized successor with the durable handoff and a bounded outcome.
An offline device is not proof that its session stopped. Session replacement does not mark the task accepted.
If your client disconnects
Restore the connection, inspect the last accepted report and check partial/history outcomes. Reconnect synchronizes state; it does not restart AI work. Use the owning collector’s Stop/Start or Retry controls. Do not kill a process by a saved PID or launch another watcher for the same config.
History and backups
Opted-in Claude history has bounded private capture with complete, partial and pending states. A completed transfer means its captured generation arrived, not that every original byte, attachment or branch was retained. Full requested history coverage remains open.
Early disposable experiments demonstrated native continuation, selected forks and compressed-history recovery. A general online vault, safe removal and cross-device restore workflow is not accepted as a complete product. Do not remove native history on the assumption that a catalog entry is a recoverable backup.
Provider credentials never travel with a handoff or backup. Native sign-in on a destination device remains local. Snapshot upload, same-device recovery, cross-device restore and Desktop visibility need separate proof.