Emoji reactions (👍❤️😂😮👎) on messages: persistence, sync, persona-voiced AI that may reply immediately to reactions

Motivation

Users want a low-friction way to react to CorpGPT messages. The AI treats each reaction as an immediate conversational input, in the Synergex corporate-satire persona. It decides for itself whether a visible reply is warranted; for example, 👎 usually means the answer missed the mark.

Scope

1. Data model and persistence

• A reaction has messageId, emoji, userId and createdAt.
• Unique key: (messageId, emoji, userId).
• Supported emoji are a single exported constant: 👍 ❤️ 😂 😮 👎. Any other value is rejected at the store/API boundary.
• Reactions are stored with the conversation using CorpGPT's existing conversation persistence mechanism (DB, file or localStorage, whichever is already used). Do not add a new storage system.
• Toggle semantics: if no reaction exists for (message, emoji, user), add one; otherwise remove it.
• Message content, order and IDs are never modified by reaction operations.
• Reactions are only accepted on eligible messages (user and assistant). Reactions targeting system or tool messages are rejected.
• Assumption: if the app has no real user accounts, userId is the existing session or user identifier. Do not build auth.

2. UI

• Each user and assistant message has a reaction control. System and tool messages have none.
• On desktop the control is revealed on hover. It is also keyboard-focusable, and on touch it is reachable via tap or the message-actions menu. There is no hover-only path.
• The picker shows exactly the 5 supported emoji.
• Reactions render beneath the message as chips showing the emoji and the count of unique users.
• Chips the current user has selected get a distinct style (e.g. highlighted border) and aria-pressed="true". Other chips get aria-pressed="false".
• Clicking or tapping a chip toggles the current user's reaction for that emoji.

3. Sync

• If CorpGPT already has a realtime multi-user channel (WebSocket, SSE, etc.), broadcast add and remove events through it to other clients in the same conversation.
• If it does not, do NOT introduce a new realtime subsystem. Provide same-session and multi-tab consistency using a lightweight browser mechanism such as BroadcastChannel or the storage event, or re-read from the store.
• The PR description states which path was taken.

4. AI awareness

• Each add or remove appends a structured event to the conversation event stream: {type: 'reaction', action: 'add'|'remove', messageId, emoji, userId, timestamp}.
• Immediate invocation: adding a reaction invokes the model right away, with the reaction event as the triggering input. The model does not wait for the next typed message.
• Optional reply: the model may choose not to reply.
  • Implement an explicit no-reply signal, e.g. a sentinel token such as [NO_REPLY], an empty response, or a tool/flag, whichever fits the existing model client.
  • When the signal is returned, no message is added to the conversation.
  • Otherwise the reply is appended as a normal assistant message.
• Assumption: removing a reaction does NOT invoke the model. It only updates the current state, so the removed reaction no longer appears in later context.
• Assumption: rapid repeated toggles on the same message are debounced or coalesced (e.g. ~1–2 s), so that at most one model call is made per burst. A toggled-off-again reaction does not trigger a call.
• Assumption: if a model response is already in flight for the conversation, the reaction is folded into context rather than starting a concurrent call.
• Context builder: includes the current reaction state (emoji and count; own vs. others where meaningful) as a compact annotation for messages in the context window. Removed reactions are absent.
• System prompt: adds reaction-interpretation guidance written in the existing CorpGPT/Synergex corporate-satire persona voice, with no fixed per-emoji response strings. Intent per emoji:
  • 👍 / ❤️: positive feedback; usually no visible reply.
  • 😂: the user enjoyed the tone; a short in-character response is sometimes fine.
  • 😮: interest; elaboration may be welcome.
  • 👎: the answer may have missed the mark; acknowledge briefly and offer a correction or clarification.
  • The guidance tells the model to use the no-reply signal when no reply is warranted.
• Reaction-driven replies use the same persona and system prompt as normal turns.

Acceptance criteria (tests unless noted)

1. A reaction control renders for user and assistant messages and not for system or tool messages.
2. Selecting a supported emoji adds a chip with count 1.
3. Repeated adds for the same (message, emoji, user) never create duplicates. This is enforced in the store/API and unit tested.
4. Toggling an existing own reaction removes it; the count decrements, or the chip disappears at 0.
5. Counts equal the number of unique users per emoji (tested with 2+ users).
6. Own reactions have a distinct style and aria-pressed="true".
7. A persistence round-trip test shows reactions survive a reload.
8. Sync, depending on which path applies:
   • With existing realtime: a second client receives add and remove without reload (integration test).
   • Without it: a second tab or instance reflects changes (test or documented manual check). The PR states which path applies.
9. Message content, order and IDs are unchanged after reaction operations.
10. The control is reachable via hover and keyboard on desktop and via tap/actions on touch; there is no hover-only path.
11. A reaction event with messageId, emoji, userId, action and timestamp is recorded on every add and remove.
12. Context builder output includes current reactions for messages in the window (snapshot or unit test).
13. After a removal, the context builder no longer includes that reaction.
14. Adding a reaction invokes the model exactly once with the reaction event in its input. Tested with a mocked model client.
15. Removing a reaction does not invoke the model (mocked client).
16. When the mocked model returns the no-reply signal, no assistant message is appended. When it returns text, exactly one assistant message is appended.
17. Rapid add/remove/add on the same message within the debounce window results in at most one model call (mocked client, fake timers).
18. The system prompt contains persona-voiced reaction guidance covering all 5 emoji and the no-reply option, with no fixed per-emoji response strings.
19. Unsupported emoji values, and reactions on system or tool messages, are rejected.

Out of scope

• Custom or uploaded emoji
• A full Unicode picker or search
• Animations
• Notifications
• Analytics or leaderboards
• Editing or removing other users' reactions
• Canned per-emoji replies
• Reactions as commands or permissions
• New auth or account systems
• A new realtime/WebSocket subsystem