Status: active · Task: mobile-chat · Approach: C — Hybrid Seams

Mobile Chat Port — Detailed Design

Altitude note. Kept deliberately high-level per the project owner's request: this is a map, not a line-by-line spec. Each PR phase (see 05-pr-roadmap.md) gets its own detailed-analysis session that grills the owner before coding. Treat file lists and shapes here as the intended structure, to be confirmed/refined per phase.

⚠️ DECISION OVERRIDE (2026-06-29) — chat logic is NOT shared. See the matching callout in 05-pr-roadmap.md. The owner reversed Approach C's shared-seam plan for chat code: nothing chat-related is extracted into @onyx-ai/shared. Everywhere this document lists web/lib/shared/src/contracts/* or web/lib/shared/src/utils/* (e.g. ndjson.ts, messageTree.ts, chatHistory.ts, streaming.ts, chat.ts, files.ts, agents.ts, projects.ts, fileDescriptors.ts), read it as mobile/src/chat/* — mobile-owned, no web re-points, no shims. The "Shared message-tree fns", "Shared createNdjsonBuffer()", and the shared-dist/watch.mjs integration notes below are superseded: mobile owns these copies and web is untouched. @onyx-ai/shared keeps receiving only cross-platform design primitives. Implemented in PR 2 under mobile/src/chat/ (streamingModels.ts, interfaces.ts, ndjson.ts, messageTree.ts, __tests__/).

Database design

N/A — no backend or database changes. The mobile client talks to the existing Onyx backend and its existing schema. All "data model" work is client-side (in-memory zustand + TanStack Query cache, plus MMKV persistence already configured). The relevant backend tables (chat_session.project_id, user_file, persona, project, Project__UserFile) already exist and are documented in 01-research.md.

Client data model

Ephemeral chat state — chatSessionStore (zustand, not persisted)

Mirrors web's useChatSessionStore shape, trimmed to the locked core (no multi-model, no regeneration, no queued-messages, no doc-sidebar fields):

Actions (ported, trimmed): setCurrentSession, createSession, updateSessionAndMessageTree, updateChatState, setAbortController, abortSession.

Server-state lists — TanStack Query (persisted to MMKV, PII-excluded)

Keys extend mobile/src/api/query-keys.ts, all keyed by serverUrl: chatSessions(serverUrl), chatSession(serverUrl, id), agents(serverUrl), projects(serverUrl), projectFiles(serverUrl, projectId).

Attachment progress — uploadStore (zustand, ephemeral, separate)

Keyed by client-side temp file id → { uri, name, mimeType, size, status, bytesSent, totalBytes }. Consumed by the input bar and the project file screen.

Class / interface design

No classes. The new surface is functions + types + hooks. Key seams:

• Mobile-native createNdjsonBuffer<T>() (mobile/src/chat/ndjson.ts) — pushChunk(text: string) => T[] and flush() => T[], generic over the caller's packet type. Holds the partial-line string; splits on \n, keeps the trailing partial, JSON.parses each complete line with web's brace-recovery fallback. No fetch/TextDecoder — those stay in the transport. Mirrors web's handleSSEStream parsing core (reader removed); not shared (PR 2 Decision (2026-06-26) in 05-pr-roadmap.md).
• Mobile streamChatMessage(body, signal): AsyncGenerator<Packet> (mobile/src/api/chat/stream.ts) — owns the expo/fetch POST + getReader() + TextDecoder loop; feeds the mobile-native buffer; filters chat_heartbeat; reader.cancel() on abort/finally.
• Mobile packet-renderer registry (mirrors web's renderMessageComponent.tsx + MessageRenderer<TPacket,TState> contract) — the extensibility foundation, built in PR 3 even though only one renderer ships then:
  • MessageRenderer<TPacket, TState> (mobile contract) — { matches(packetType): boolean; reduce(state, packet): TState; render(state): ReactNode }. RN-coupled (returns RN nodes) → mobile-owned, not shared. The packet-grouping step can later be shared if reuse is proven.
  • findRenderer(packetType): MessageRenderer | null — the dispatch (mobile equivalent of web's findRenderer).
  • Core (PR 3) registers exactly one renderer: MessageTextRenderer (MESSAGE_START/DELTA/END → markdown string + isComplete/error). This is ~the same code as a flat concatenator, so it does not grow core scope.
  • usePacketDisplay(node) groups the node's packets and walks the registry to produce render output. Deferred rich-chat PRs (9a–9e) add a renderer to the registry (and, for agentic steps, a timeline composition layer) — no core rewrite. (Web's React-coupled usePacketProcessor stays web-only.)
• Mobile-native message-tree fns — upsertMessages, getLatestMessageChain, getMessageByMessageId, buildImmediateMessages, buildEmptyMessage, SYSTEM_NODE_ID, MessageTreeState. Ported from web's messageTree.ts against the minimal mobile Message (per the PR 2 minimal-sharing decision — not shared; web keeps its own copy untouched). processRawChatHistory likewise lives in mobile/src/chat/chatHistory.ts.

New files

Shared package (web/lib/shared/src/)

No new shared files. Per the PR 2 Decision the chat pure layer is mobile-native; nothing chat-related enters @onyx-ai/shared.

Mobile-native chat layer (mobile/src/chat/) — per the PR 2 Decision the entire chat pure layer lives in mobile, not shared

Mobile (mobile/src/)

File structure (tree)

web/lib/shared/src/   (NO chat changes — nothing chat-related enters shared)
web/src/              (UNTOUCHED — streamingUtils.ts / messageTree.ts / fileUtils.ts stay web-owned)

mobile/src/
├── app/
│   ├── _layout.tsx         (modified: mount (app) group)
│   └── (app)/              (new)  _layout · index · chat/[id] · history · projects/*
├── chat/                   (new, mobile-native pure layer)
│   ├── ndjson.ts (new)     contracts/ (new)  streaming · chat · files · agents · projects
│   ├── messageTree.ts (new)  chatHistory.ts (new)  fileDescriptors.ts (new)
│   └── *.test.ts           (new)  ndjson + tree + history unit tests
├── state/                  (new)  chatSessionStore.ts · uploadStore.ts
├── api/
│   ├── query-keys.ts       (modified: add chat/agents/projects keys)
│   ├── chat/               (new)  stream.ts · sessions.ts · agents.ts · projects.ts
│   └── files/              (new)  upload.ts
├── hooks/                  (new)  useChatController · useChatSessionController · usePacketDisplay
└── components/chat/        (new)  MessageList · MessageRow · StreamingMarkdown · InputBar · AgentPicker · AttachmentChips

Integration points

• Auth / HTTP — list calls reuse mobile/src/api/client.ts apiFetch (bearer + ApiError). The streaming call (api/chat/stream.ts) reuses getBaseUrl() (config.ts) + the token from tokenStore.ts, but uses expo/fetch directly.
• Navigation — (app) group mounts under the existing AuthGate in mobile/src/app/_layout.tsx; sidebar (mobile/src/components/sidebar) surfaces sessions/projects.
• Query cache — extend mobile/src/api/query-keys.ts + reuse the persisted client in mobile/src/query/client.ts. Required (not conditional): chat content is sensitive by nature, so the chat session-list + session-detail/message query keys must be added to the dehydrateOptions PII-exclusion list (alongside the existing me exclusion) in PR 1, before any chat history is persisted to MMKV. Consequence: chat history is not cached to disk and refetches on launch — the correct PII-safe default (mirrors the me-exclusion pattern).
• No web changes — the mobile chat port touches no web files. web/src/lib/search/streamingUtils.ts, .../messageTree.ts, and .../fileUtils.ts stay web-owned; mobile reimplements the parser/tree/file-descriptor logic natively in mobile/src/chat/.
• No shared-build coupling — nothing chat-related enters @onyx-ai/shared, so there's no dist-rebuild/file:-relink dependency for the chat port. (@onyx-ai/shared keeps its existing design-token + interactive/typography + numbers/format surface.)

Important notes before implementation

• De-risk spikes first (Phase 0/early): (1) confirm expo/fetch exposes response.body.getReader() on a device dev build (RN 0.85 / SDK 56) — fallback is XHR progress feeding the same NDJSON buffer; (2) confirm react-native-streamdown builds on RN 0.85 / Reanimated v4 — fallback react-native-marked (both keep block-memoization). Both require a dev client (already have expo-dev-client), not Expo Go.
• Never persist chatSessionStore — it holds AbortControllers and live streams. Keep it strictly separate from the persisted TanStack cache; on relaunch, rehydrate a session via GET get-chat-session + the mobile-native processRawChatHistory.
• Batch flushes ~50ms and memoize rows on packetCount (not array identity) — the single biggest streaming-perf lever; port web's stillCurrent/abort guards so a backgrounded stream can't write into the wrong session.
• Send gating — block send until attached files are indexed (token_count != null); surface a FAILED/stuck status instead of blocking forever (3s status polling, mirror ProjectsContext).
• Keep the mobile contracts/streaming.ts minimal — only the core packet types now; add rich types in their deferred phases. Avoid dragging web's full OnyxDocument shape across until citations land.
• Renderer foundation in PR 3, renderers as follow-ups — build the MessageRenderer contract + findRenderer dispatch in PR 3 (registering only MessageTextRenderer) so rich-chat features are additive registrations, not core rewrites. The agentic-timeline composition layer (AgentTimeline) is itself deferred — the first timeline renderer PR (9b) builds it; the dispatch seam it plugs into already exists from PR 3. Do not build the rich renderers in core — only the seam.
• Agent/project selection is implicit — carried by the session's persona_id/project_id at create-time; there is no per-message agent param. Selecting an agent for an existing session is not supported by the backend (matches web) — start a new session.
• Error handling — ERROR/PacketError packets set the assistant node to an error state; raise OnyxError-style messages only on the backend (client surfaces ApiError messages). No HTTPException concerns here (client-only).
• expires= / Celery — N/A (no backend tasks introduced).