Practical use.
Watch Mode is most useful before an interruption, while edits are still changing. Its job is to keep local recovery material fresher than a handoff created only after the session has already disappeared.
Autosave / Watch Mode.
Watch Mode keeps a local recovery state fresh while you work, so crashes, rate limits, context overflow, or provider failures do not erase the thread.
What it carries.
- Local snapshots while work changes
- Fresh changed-file awareness
- Recovery state before lockout
- No cloud sync or telemetry
How it fits ShardStitch.
Watch mode focuses on the changing project state: it helps capture meaningful file and verification changes before a session becomes difficult to trust.
Before the failure happens.
Watch Mode is useful during long edits, provider instability, and work that is difficult to reconstruct later. It keeps a local recovery state close to the changing files so a sudden lockout does not leave the next session with only a vague summary.
Snapshots are recovery evidence, not automatic commits or backups of every external service. Check the local storage boundary and keep normal version-control and backup practices in place.
Use Watch Mode during a risky work session.
Before a multi-file refactor, enable the local watch or autosave workflow and confirm the project being watched is the repository the agent is editing.
- Check the watched project path and repository branch before starting; a watcher pointed at a parent or stale checkout records the wrong working state.
- After a meaningful change, confirm the local recovery record advances and that the changed-file list includes the files you expect.
- If the agent stops responding, compare the latest record with git status and the diff; resume only from facts that still match disk.
Worked example.
For a refactor that touches several modules, start watching the correct checkout before edits begin. After a stable checkpoint, confirm the recovery record reflects the new files; after an interruption, compare that record with the current diff. If the editor still had unsaved buffers, treat those as a separate recovery problem rather than assuming the watcher captured them.
Limits and verification.
A watch record is a recovery aid, not a transactional backup of every keystroke and not a substitute for Git commits or separate backups. Do not assume it captured an unsaved editor buffer, a file outside the watched project, or work performed after the last observed update.