The next milestone for packages/app is a design-partner / private beta: real users hitting it daily. The bar shifts from "demo well" to "hold up under real workflows." This roadmap is built on a feature audit (below) plus explicit user priorities: fix the two highest-pain mocks, ship Google Drive + GitHub connections first, and build out agent autonomy and sharing.

Status: REAL = wired to packages/server; MOCK = local simulator; HYBRID = part real, part placeholder.

• Chat (send/receive, history, WS push) — App.tsx:30
• Roomy artifact gallery
• Tasks board + calendar — modeled as messages with kind=task — App.tsx:187
• Context Library (full CRUD: upload, folder, link, move, delete)
• Global command palette (/api/search, scoped)
• Settings: profile, password, agents, workspaces, models
• Auth (bearer tokens, /api/me boot check)
• Real-time WS at /ws with reconnect

• Today/Inbox reply — list is real, reply UX is useMockChat — TodayInbox.tsx:380
• Artifact Conversation tab — useMockChat, doesn't reach the model — ConversationPanel.tsx
• Connections — only Claude + ChatGPT keys are real; Drive/Notion/GitHub/Slack/Figma/Linear/Web Clipper are stub catalog entries — connections.ts

Ordered by priority for the private-beta milestone.

Goal: every visible reply field actually answers.

• R1.1 Inbox reply uses real chat backend. Swap useMockChat in TodayInbox.tsx and TodayDetailPanel.tsx for usePostChatMessageMutation against the inbox item's chat. Keep optimistic UI.
• R1.2 Artifact Conversation tab uses real chat. Same swap in ConversationPanel.tsx; decide whether the conversation lives in the originating chat or its own thread (likely originating chat, scoped by artifact reference).
• R1.3 Delete use-mock-chat.ts once unreferenced.

Goal: two real third-party integrations the beta cohort will actually use.

• C2.1 Google Drive — OAuth flow, server-side token storage alongside /me/providers, file picker that ingests selected files into the existing Library entity (reuse /api/library rather than inventing a parallel store), and "drive://" reference type so chats can attach without copying.
• C2.2 GitHub — App install or PAT, repo/issue picker, surface as Library entries, plus a kind=issue message variant so issues can flow through Today/Inbox like tasks already do.
• C2.3 Connections shell — promote /me/providers into a generic connection store so adding the next 5 doesn't require schema changes.

Goal: agents that do work without a user prompt.

• A3.1 Scheduled triggers — let an agent run on cron (existing packages/server/scheduler is a likely home; verify before assuming) and post results as a regular chat message + Today/Inbox entry.
• A3.2 Event triggers — run on incoming events (e.g. a new GitHub issue, a new Drive doc). Reuses the WS event spine.
• A3.3 Run history & output review — every autonomous run produces a reviewable artifact + an awaiting-user state when human input is needed.
• A3.4 Tasks-as-first-class prerequisite check — recurrence, dependencies, and "this task is owned by an agent" all rub against the current messages.kind=task modeling. Decide early whether to lift Tasks into a real table or extend message metadata; the answer shapes A3.

Goal: a beta user can show their work to a colleague.

• S4.1 Read-only artifact share links — signed URLs, expirable, no login required for the recipient.
• S4.2 Workspace invites — email invite → joiner gets scoped access to a workspace's chats/artifacts/library.
• S4.3 Per-resource permissions — chat-level and artifact-level visibility (private / workspace / shared link).
• S4.4 Comments on artifacts — async feedback without polluting the chat thread.

• N5.1 Notifications & digest — beta users will not keep the tab open. Email digest for awaiting-user messages; web-push for live ones. Today/Inbox is the natural source-of-truth.
• N5.2 Onboarding & empty states — first-run flow that walks a partner through: pick a model, connect Drive, upload one library file, run their first agent. Today the empty workspace gives no scaffolding.
• N5.3 Workspace memory — long-lived facts the agent retains across chats (project context, preferences, glossary). Surfaced and editable in Settings.
• N5.4 Usage visibility — token + cost per workspace/agent/chat; design partners ask within week one when they're using their own API keys.
• N5.5 Templates / prompt library — reusable starting prompts and agent recipes; lowers cold-start cost for new partners.

• Error states + retry UX for every RTK Query surface (today many fail silently)
• Server-side rate-limit / quota signaling surfaced in the UI
• Audit log of what an agent did on the user's behalf (especially with connections)
• Basic observability for the server (latency + error rates per endpoint) so we know what beta users hit
• Test coverage parity per AGENTS.md: each new real surface gets an integration test against real Postgres + real AI (free opencode/big-pickle), not just unit tests

1. Sprint 1–2: Theme 1 (Make real) + N5.2 (Onboarding) + start C2.3 (Connections shell).
2. Sprint 3–5: C2.1 (Drive) and C2.2 (GitHub) in parallel; N5.1 (Notifications) lands alongside since both want a server-side trigger story.
3. Sprint 6–8: A3.1/A3.2 (Scheduled + event triggers) building on the trigger infra from N5.1; decide on A3.4 (tasks model lift) before this work starts.
4. Sprint 9–10: S4 (Sharing) + N5.4 (Usage) + N5.3 (Workspace memory).
5. Throughout: Theme 6 polish — done as the price of admission for each shipped feature, not a separate phase.

• packages/app/src/hooks/use-mock-chat.ts — to be deleted (Theme 1)
• packages/app/src/components/today/ — Inbox reply real-ification
• packages/app/src/components/artifact/ConversationPanel.tsx
• packages/app/src/data/connections.ts — connection catalog grows, plus per-kind config
• packages/server/api/src/routes/ — new routes for Drive, GitHub, triggers, sharing
• packages/server/scheduler — verify shape before placing scheduled-run work
• packages/server/api/src/ws/ — extend for trigger events

The private-beta milestone is met when:

• A new design partner can sign in, connect Drive or GitHub, run an agent on real data, receive an emailed digest of awaiting-user items, and share an artifact link with a colleague — without anyone on the team intervening.
• No surface in the app uses useMockChat or stub connection data.
• The team has dashboards showing per-endpoint error rate and per-workspace token usage so we can see what beta users hit before they tell us.