[jayleaton]

Summary

Sits on #229. The server pushes to registered phones straight through Expo's push service.

Rule (JJ, 2026-11-14): only top-level threads push, meaning no parent of any kind and not delegated from another machine. They push when they need the user (input or approval), finish, or fail. Child chats, subagents and side chats never push on their own; their questions go out as their top-level thread's.

streamDomainEvents ─filter─▶ worker(threadId)
   └ observation
        top-level thread : projectThreadAwarenessV2 phase (settled results already hidden)
        any child        : its own pending request only (never a result)
   └ decidePush(previous, next, { topLevel, snoozed })      // pure, apps/server/src/push/pushPolicy.ts
        new input / approval (new request id = new wait)
        completed / failed: top-level only, at most 2 min old; completed is quiet while snoozed
        first sighting after start: baseline only
   └ completed: hold 5 s and recheck, so a continuation stays quiet
   └ recipient thread = itself if top-level, else walk parentThreadId / lineage.parentThreadId
                        to the top-level ancestor (a delegatedFrom thread on the way stops it)
   └ send to listDeliverable() in batches of 100; DeviceNotRegistered in tickets or in
     15-minute receipts deletes that row

Payload: { to, title: <top-level thread title>, body: "<Needs your input|Needs approval|Finished|Failed> · <project>", sound, data: { environmentId, threadId, kind } }. No message content. T3CODE_EXPO_PUSH_URL overrides https://exp.host for testing. AgentAwarenessThread gains snoozedUntil and delegatedFrom. Desktop and web notifications (ThreadNotificationCoordinator) are not touched; a separate PR changes them.

Evidence

• Before: no notifier exists, so a server sends no push.
• After: tests:

  • pushPolicy.test.ts: 17 rows covering phase pairs × child, snoozed, stale, and repeat waits.

  • PushNotifier.test.ts: real store, fake Expo HTTP. It checks:

    • the exact payload;
    • ticket and receipt DeviceNotRegistered deletes the row;
    • the 5 s completion hold;
    • a subagent approval and a child-chat question both go out under the top-level thread's title and id;
    • a child chat's finish sends nothing.

    The child-chat step fails on the previous routing, which sent under the child's own title.

• Runtime transcript (before the rule change; the top-level path is unchanged). Setup: vp run dev with T3CODE_EXPO_PUSH_URL pointing at a local fake Expo, one registration, and t3 thread start --provider devReplay. The fake Expo received:

  todo-list : {"to":"ExponentPushToken[proof-iphone]","title":"go","body":"Finished · push-proof","sound":"default","data":{"environmentId":"c7353c4e-…","threadId":"mcp:f54711f8-…","kind":"completed"}}
  question  : {"to":"ExponentPushToken[proof-iphone]","title":"ask me something","body":"Needs your input · push-proof","sound":"default","data":{"environmentId":"c7353c4e-…","threadId":"mcp:adad8caf-…","kind":"input"}}
  

• vp run check:changed on this layer: all pass.

Merge Danger

Door: two-way. It adds no persisted state beyond #229's table, and a revert stops pushes.

Blast Radius: service. It reads one thread snapshot per relevant event and calls exp.host only when a push is due and a phone is registered. Thread and project titles pass through Expo and Apple. Expo's access-token requirement stays off by decision.

🤖 Generated with Claude Code

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ No successful main baseline artifact is available yet. This run establishes the initial measurement.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	—	5.0 KiB	—	6.8 KiB	✅
Codex	Thread snapshot wire	—	3.8 KiB	—	4.9 KiB	✅
Codex	Live turn WebSocket wire	—	1.2 KiB	—	2.0 KiB	✅
Codex	Live turn WebSocket decoded	—	21.5 KiB	—	24.4 KiB	✅
Codex	Live turn messages	—	1	—	8	✅
Claude	Total thread wire	—	5.0 KiB	—	6.8 KiB	✅
Claude	Thread snapshot wire	—	3.8 KiB	—	4.9 KiB	✅
Claude	Live turn WebSocket wire	—	1.2 KiB	—	2.0 KiB	✅
Claude	Live turn WebSocket decoded	—	21.9 KiB	—	24.4 KiB	✅
Claude	Live turn messages	—	1	—	8	✅

Baseline: unavailable · PR result: cc4ade5 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 108.5 KiB
• Claude decoded thread snapshot: 108.8 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.