think
16px
820px
← all files

Daily-Driver Build Plan (synthesis of 11 confirmed designs)

Full per-feature specs: designs-2026-07-06.json (11 designs, all feasibility confirmed, produced by parallel design agents reading the real code; two live-verified against claude 2.1.201).

Reconciled event vocabulary (all carry projectId)

  • activity — extended with optional detail (stdout/stderr excerpt, Edit/Write line-diff, error), capped payload; same kind feeds tool rows AND ultracode subagent view.
  • turn_interrupted — shared by Stop button and queue's "interrupt & send now"; panel suppresses the following session_done{failed} banner.
  • question {text, options[]?} — emitted at turn end when the model asked (ASK convention + trailing-? heuristic); panel renders an answer card; reply goes through the normal continue path (context preserved via resume).
  • queue_update {queued[]}, conversation_updated/selected, turn_auto_resumed / turn_resume_abandoned, plan_proposed; session_started gains {ultracode, permissionMode}.

Endpoints (no collisions)

POST /api/interrupt {projectId?} · GET /api/commands?projectId · GET /api/notify/config · conversation CRUD + POST /api/conversations/current · send bodies gain permissionMode?, ultracode?; messages endpoint returns {queued:true,position} when busy instead of 409.

Key verified facts

  • Ultracode: CLAUDE_CODE_EFFORT_LEVEL=ultracode env on the spawned process (live-verified). Composes with existing subagent activity view. Org-policy may restrict xhigh — surface the error.
  • Answer-questions: persistent-stdin single-process exists but is fragile; chosen design detects questions at turn boundaries and answers via resume — full context, no process gymnastics (live-verified flow).
  • Auto-resume: in-flight marker file per project, cleared in .finally(); SIGTERM exit(130) skips clearing → marker present ⇔ ungraceful death. One resume attempt max (poison-turn guard), turn_resume_abandoned otherwise.
  • projects.json v2: conversations[] + currentConversationId with v1 migration (lastSessionId → first conversation). Land BEFORE queue/auto-resume (both reference conversation state).

Build phases (each independently deployable)

  1. See + control ← start here: activity detail + expandable rows; Stop button (turn_interrupted). Biggest pain first.
  2. Threads: projects.json v2 multi-conversation (sidebar nesting, titles from first prompt) + per-project queue with cancel / interrupt-and-send.
  3. Ask + tell: question cards; browser notifications (+Telegram config probe).
  4. Power: ultracode toggle; permission-mode selector + retry; slash commands (/clear UI-local; /compact passthrough; custom commands via GET /api/commands).
  5. Polish + resilience: copy buttons, mobile hamburger layout; auto-resume-on-restart.

Cross-feature risks to respect while building

  • History reload drops tool-detail rows (transcript has no activity) — acceptable; live turns keep them.
  • Notification replay-dup guard (arm after SSE replay settles).
  • Queue drains through the same launch path as auto-resume — serialize per project (run map already guarantees one turn per project).