App — Sidekick

The AI is the product. Every other app feeds the Sidekick, and the Sidekick can act on every other app.

Field	Value
Status	🟢 GA
Owner	@aphaiboon
Tech lead	@aphaiboon
Customer-facing name	The user names theirs

What it is

A conversational interface (chat) that has full read/write access to the user’s workspace and a persistent memory. Lives in:

• Sidekick section of the sidebar (always accessible)
• Inline inside other apps (compose a draft, summarize a thread, plan a project)
• Slash-trigger from anywhere — /think, /quick, /code

The four modes

Mode	Model	Tools	Latency	When
Quick	FAST_MODEL	disabled by default	3–5s	Short factual answers
Autopilot (default)	CAPABLE_MODEL	full toolset	5–15s	The everyday mode
Deep Think	FLAGSHIP_MODEL	full + extended reasoning	15–30s	Reports, deep analysis
Code	CAPABLE_MODEL	sandbox tools, 32k tokens	5–15s	Code pairing

Auto-elevation: Quick → Autopilot when a tool is needed. Transparent to the user.

→ Full pipeline: Engineering → Architecture → Sidekick Routing Pipeline

Threads & chats

• Chat — one conversation. Standalone, in a Thread, or linked to a Project.
• Thread — long-lived container for related chats. Like a Slack channel for one topic.
• Project — container for Threads + Tasks + Brain docs (lives in Projects app).

UI hierarchy: Projects ⊇ Threads ⊇ Chats

Copy rules:

• ✅ Always say “chats” (lowercase)
• 🚫 Never say “conversations” (banned in user-facing copy)
• ✅ Action is “Make it a Thread” (not “Upgrade to Thread”)

Context tags

Users can attach durable context to a chat with @ or + mentions:

Tag	What it injects
@person	Person’s recent activity, notes, deal status
@project	Project overview, recent tasks, recent threads
+brain-document	Full doc content (up to 8k chars for DB docs, summary for file-backed)
+brain-folder	Up to 5 docs × 1500 chars from the folder

Important: when explicit Brain tags are present, auto Brain semantic search is skipped — pinned context wins.

Search scope (knowledge mode)

Scope	Behavior
Smart (default)	Brain + web + memory
Internal	Brain + memory — skip web
External	Web — skip Brain / memory / live queries

Tool catalog (excerpt)

Sidekick can invoke ~40 tools. By category:

• Tasks — create_task, update_task, complete_task, query_tasks
• Calendar — schedule_event, find_time, update_event, query_events
• Comms — send_message, draft_email, summarize_thread, find_message
• Brain — create_brain_doc, search_brain, update_brain_doc, summarize_folder
• Projects — create_project, link_chat_to_project, update_project
• CRM — create_contact, log_interaction, find_company
• People — recall_person, update_person_memory
• Integrations — call_composio_tool (Slack, GitHub, Gmail, Drive, Stripe, …)
• Web — search_web, fetch_url, fetch_news
• Meta — set_sidekick_mode, switch_thread, archive_chat

Personalisation

Setting	Where
Sidekick name	First-run flow + Settings → Sidekick
Avatar	SidekickAvatar component, /api/ai/preferences avatar style
Voice / persona	Settings → Sidekick (planned)
Default mode	Per-workspace preference

Engineering rule: never hardcode “Sidekick” / “Ruby” / “Assistant” in UI. Always use useSidekickName() from @/hooks/use-sidekick-name.

Memory & persistence

• Per-chat memory — recent N turns in the chat
• Working memory — short-term relevant snippets across chats
• Brain — long-term knowledge (user-curated documents)
• Person memory — facts about specific people in the user’s life
• Ghost Mode — skips persistence, memory extraction, usage tracking

Ambient context (always injected)

Every Sidekick turn includes:

• Explicit day name + full date
• Timezone
• User name + role
• Active integrations
• Current app context (which screen the user was on)

Analytics events

• sidekick_message_sent — every user message
• sidekick_response_received — every AI response (with mode + tier + latency)
• sidekick_tool_invoked — per tool call
• sidekick_mode_switched — explicit mode change
• sidekick_context_tag_attached — @ or + mention
• sidekick_thread_created — thread upgrade from chat

Performance budget

Tier	Target latency	Hard ceiling
Trivial	<2s	3s
Simple	3–5s	7s
Standard	5–15s	25s
Complex	15–30s	45s

Latency over the hard ceiling pages on-call.

Related

• Engineering → Sidekick Routing Pipeline
• App — Brain — primary knowledge source
• Model Card — Apex Family — models powering Sidekick