Queued/scheduled sessions and auto-continue on usage reset

[mattmoran56]

Summary

Two related affordances around Claude usage limits:

1. Schedule a session for later — pick a time (the dialog pre-fills from your next 5 h reset), enter a custom prompt, and it sits in a Scheduled sidebar panel until fire time. At fire time the worktree, branch, and agent are spun up automatically, with the prompt piped into claude via heredoc so it's bulletproof; claude's existing auto-restart then drops you back into claude --resume so you can keep chatting. Doesn't yank your view to the new session.
2. Auto-continue on usage reset — when an active session's 5 h window crosses 95 %, surface a toast (or, with a setting toggled on, queue silently) that lines up continue to fire ~1 minute after reset. The in-session chip shows the queued message + cancel.

Both queue items persist via electron-store, so a closed app picks them up on next launch (with a 2.5 s renderer-ready buffer for past-due items).

Test plan

• Unit: 39 tests across scheduleTime, schedulerStore, scheduler.service, useSchedulerBootstrap, writeWhenReady
• E2E: 5 Playwright tests covering dialog flow, fire path, prompt injection, and cancel
• Manual: scheduled a 2-min session with a custom prompt; worktree + branch were created at fire time, prompt appeared in claude's input, claude responded, and the session was reachable via the sidebar afterward
• Screenshots — the PR touches UI (Scheduled panel, schedule dialog, usage-limit toast, queued-message chip, settings section). I'll attach them in a follow-up commit unless you'd like to grab them from your dev run.

Notes

• The original >-detection-then-write injection path (used by manually-created sessions with a preset startup prompt) is unchanged — it's still the path startup prompts use. The queued-fire path uses a new IPC scheduler:spawn-agent-with-prompt that pipes the prompt via heredoc, which is more reliable in the queued context where the renderer subscribe-to-data timing is racy.
• Cross-project firing has explicit unit-test coverage to make sure the target project's session list never gets corrupted.

🤖 Generated with Claude Code