Cadeo Chat

Cadeo Chat is the open conversation surface inside Cadeo — a place to ask Cadeo anything outside the structured rituals (morning brief, Plan My Day, AI Inbox triage).

How to open

• Keyboard: Cmd+J (anywhere in Cadeo).
• The small ⚡ Ask Cadeo button next to Plan My Day in the planner header.
• The command bar (Cmd+K) — "Chat with Cadeo" appears at the bottom of the list, along with the assistant's verbs as "Ask Cadeo: …" entries.

Press Esc to close — the conversation is preserved.

What you can ask

Chat is for the questions that don't fit a ritual:

• "What did I commit to last week?"
• "Summarise my last 5 1:1 notes with Sarah"
• "What am I forgetting before Friday?"
• "Help me think through the Q3 OKR"

Cadeo can now also do things — find, create, update, complete, reschedule and break down tasks; create projects. When the AI proposes a change, an inline tool card appears with Accept · Edit · Reject. After Accept, you have 8 seconds to Undo before the change is locked in.

Available tools in this release:

• Reads (auto-run, appear as rows in the activity log above the answer): Looked up tasks · notes · calendar events · areas · projects · tags, Searched your workspace, Summarised, Looked up your recordings, Read the meeting transcript
• Writes (need your Accept): Create task, Update task, Reschedule task, Complete task, Break into subtasks, Delete task, Block time, Create project, Set up project

Block time puts a real event on your calendar — "block 90 minutes tomorrow morning for the deck" creates an event in your connected Google or Apple calendar (whichever you've set as primary for new events). Like every change it's reviewable before it lands and reversible with Undo — undoing it removes the calendar event again.

Set up project goes further than Create project: ask Cadeo to set up a project for something — "set up a project for the office move" — and it proposes the project, a set of sections, and a list of starter tasks all at once. The starter tasks come up as a checklist (the project and its sections are always created); un-tick any you don't want before accepting.

Deleting a task is a destructive action, so it gets extra care: the card is red, names exactly which task (and how many subtasks) will go, and the confirm button reads Delete rather than Accept. It can never be added to "Always allow" — a delete always asks. In a batch, delete rows arrive unchecked so you opt into them deliberately. As with any change, you have 8 seconds to Undo — and Undo fully restores the task and its subtasks.

Cadeo can run several reads in one turn — so "what's on my plate this week?" can check your tasks and your calendar together.

Asking about your meeting recordings

Cadeo can find and summarise the meetings you've recorded. Try things like:

• "Can you sum up my last recorded meeting?"
• "Summarise all my meeting transcriptions this week"
• "How many meetings did I record this week?"

The look-up happens in two steps the activity log will show you: first Cadeo lists which recordings exist (metadata only — cheap), then it pulls the assembled transcripts (with speaker labels) for the ones it actually needs to read. Counting questions stop after the list step, so they're fast and don't waste tokens.

Tool cards use friendly labels ("Created task" instead of task.create). The underlying schema name still appears in the tool's data attributes for routing — what changed is just the copy users see.

Opening a ritual from chat

If you ask Cadeo something a guided flow handles better — "help me plan my day", "triage my inbox" — it runs that instead of improvising in prose. Plan My Day and Break down happen right in the chat (you approve the result there); Inbox Triage, Morning Brief and the End-of-Day wrap-up open their own screen, with a breadcrumb left in the chat so you know where you went.

Reviewing a batch of changes

When an action fans out into several changes — Break into subtasks proposing a whole list of subtasks, or Cadeo proposing several edits at once in answer to one request — you get a checklist card instead of separate cards. Every proposed change is its own row with a checkbox. Un-tick the ones you don't want, then Accept Selected applies only the checked ones (the count on the button updates as you toggle), or Reject All drops the lot. It stays one card with one Undo — accepting the batch and undoing it are each a single action.

For tools you trust, tick "Always allow" on any tool card — the next time the AI proposes that same tool, it'll run automatically (still with the 8-second Undo). Revoke from AI Settings → Auto-actions. Destructive actions never auto-run, regardless of allow-list.

What Cadeo did (the activity log)

When Cadeo is working through a complex request, a calm activity log appears above the answer — a short vertical list showing the steps Cadeo took, one row at a time:

• Thinking it through — Cadeo is reasoning
• Plan · Checking your tasks and calendar — Cadeo's stated plan for the turn, in plain language
• Looking up tasks · 7 found — Cadeo checked your tasks
• Checking your calendar · 3 events — Cadeo checked your calendar
• Searching your workspace · 12 of 87 matches — Cadeo ran a workspace search

Each row has a small status dot: a breathing blue dot while a step is running, a quiet tick once it's finished, an amber ✕ if it failed. Only one row is active at a time. Finished rows stay put so the log doesn't jump around as the next step starts.

Once the answer arrives, the whole log collapses to a one-line summary — "3 steps · Show" — so it stays out of the way. Click Show to inspect what Cadeo did; click Hide to put it away again. Your choice persists with the thread.

The thinking row inside the expanded log is special — click it to expand the raw reasoning text, with Copy for sharing. Reasoning is hidden by default and never affects the visible answer — it's there if you want to look, ignorable if you don't.

A plain conversational reply ("hello!", "thanks") shows no activity log at all — there were no steps to log, so the strip stays away. Showing the log only when something happened is itself a quality signal.

Why? — the reasoning inspector

Every AI answer carries a quiet Why? button at the bottom-right. Click it to open the reasoning inspector — the audit-grade detail behind the answer:

• The model used (Qwen 3, Gemini, etc.)
• The classifier verdict (trivial / complex — see the section above)
• Cadeo's stated plan for that turn
• Every tool call, with arguments and result summary
• Sources / context Cadeo referenced
• The full answer
• The model's raw reasoning text, verbatim

It's the place to go when an answer looks off, or when you want to confirm what Cadeo actually used. Closes on Esc, click outside, or the × button.

Threads

Cadeo Chat opens with a fresh thread by default. Closing the sheet preserves it for 24 hours — reopen with Cmd+J to continue.

To keep a thread permanently, click the bookmark icon in the header (or tooltip "Keep this"). Pinned threads have no expiry.

To start fresh without losing the current thread, click the new thread icon. (Pinned threads are kept; ephemeral ones drop on the next opening if they're older than 24h.)

Privacy & active model

The composer footer shows which model will answer your next turn:

• ● Cloud (Gemini Flash) — Plus users with cloud chat enabled. Faster answers, larger context, multi-step reasoning across many recordings.
• ● Local (Ollama) — free tier, or Plus users who've chosen local. Your messages and Cadeo's responses run entirely on your Mac. Nothing leaves the device.
• ● Local-only (Ollama) — strict privacy mode is on. Cloud-AI chat tools (semantic search, plan day, web fetch) are also hidden.
• ● Ollama not running — desktop only, with no cloud option active. Start Ollama from Settings → AI.

Two toggles in Settings → Recording & AI → Data & Privacy control where chat runs:

• Local-only AI mode — strict gate. Your data NEVER goes to a third-party AI provider (Gemini, OpenAI). Cloud-AI tools are hidden. Use this if your privacy posture demands it. Works for everyone.
• Use cloud AI for chat (Plus-only) — defaults on. Routes your chat through Gemini Flash for richer, faster answers and the multi-step agent.followup chain (see "Multi-step reasoning" below). Disable to keep chat on the local Ollama model even with Plus — quality vs. privacy preference within Plus.

Local-only mode always wins over the cloud chat toggle: if Local-only is on, cloud chat is off regardless of the per-user preference. Web users (no Ollama) and iOS get cloud chat unconditionally when they're on Plus.

Multi-step reasoning (Plus, cloud)

Some questions need a chain of look-ups to answer well — "summarise my last meeting" needs to (1) find the most recent recording, then (2) fetch its transcript, then (3) synthesise. With cloud chat, Cadeo now automatically chains these for you: the activity log shows each step, intermediate bubbles are suppressed so you see one consolidated log + final answer, and the chain is capped at 3 hops to keep costs predictable. If a question genuinely needs more than 3 steps, Cadeo stops and says so — usually a sign the question can be split or narrowed.

"Thinking…"

If you ask Cadeo to do something that involves looking up tasks, notes, or calendar events (or anything more complex than small talk), you'll see Thinking… appear under CADEO for a few seconds before the answer streams in. That's the local model reasoning through your request and deciding which tool to use. On heavier prompts this can take 30–60 seconds — it's working, not stuck.

Small talk skips the wait. Greetings, thanks, quick acknowledgments, and simple factual questions ("hi", "thanks", "what year is it") get an answer in a second or two — Cadeo recognises that there's nothing to reason about and skips the reasoning step entirely. If the activity log appears above the answer, Cadeo thought; if it doesn't, Cadeo answered straight away. That's the at-a-glance signal.

Keyboard

Key	Action
Cmd+J	Toggle chat
Cmd+Enter	Send message
Esc	Close (thread preserved)
Cmd+.	Stop streaming response
Shift+Enter	New line in composer

Slash commands — the verb menu

Start a message with / to see everything Cadeo can do — its verbs — and pick one:

Command	What it does	Where it opens
/plan	Plan my day — proposes a schedule you approve block by block	In the assistant
/breakdown	Break a task down into subtasks	In the assistant
/week	Plan my week	Weekly planner
/triage	Process my inbox — sort tasks into areas and days	Inbox triage
/pulse	Project health	Project pulse
/wrapup	End-of-day wrap-up	Wrap-up
/brief	Morning brief	Morning brief

Plan My Day and Break down open right here in the assistant, where you approve the result. The others open their own screen for now (a breadcrumb is left in the chat so you know where you went).

The same verbs are reachable from the Cmd+K command bar as "Ask Cadeo: …".

In local-only mode, actions that would use cloud AI (Plan My Day, weekly planning, project health, the morning brief) are hidden from the / menu and ⌘K — only the local-capable verbs (like Break down) appear, keeping the local-only promise.

Arrow keys to navigate, Enter to select, Esc to dismiss the picker without sending.

@-mentions

Type @ to reference a specific task, project, note, or date. The picker shows matches as you type. Selecting one drops in a structured pill — Cadeo sees the actual entity ID, not just the title, so when you ask "reschedule @Email Priya to Thursday" it can hit the right task without ambiguity.

Available @ mentions:

• Tasks — your open tasks (most recent first)
• Projects — by name
• Notes — by title or first-line content
• Dates — @today, @tomorrow, @yesterday, @last week

What Cadeo sees automatically

Every time you send a message, Cadeo silently receives a short context block — today's date and weekday, your active view, your timezone, and counts of your open tasks (scheduled today, overdue, this week, in inbox) and today's calendar events. You don't have to remind it what day it is or how many things are on your plate.

When you use @-mentions, Cadeo also gets the full content of each referenced entity — title, schedule, duration, area, notes excerpt. So "@Email Priya — actually move this to Thursday morning" lands on the right task with the right notes context, not just an ID.

What's coming next

• Implicit retrieval: even without an @-mention, Cadeo will pull in relevant tasks/notes when you reference them by name ("Sarah's notes from last week").
• Inline retrieval transparency: see which tasks/notes Cadeo used to answer your question.
• iOS cloud chat: the same Gemini-backed chat surface on iPhone and iPad.

Have feedback on the chat? Let us know — we want to hear it.