Inbox
The board answers “is anything waiting on me?” The Inbox answers the next question: “what exactly do I need to do?” Agents work on many issues across many projects at once. What they need you to decide, what they want you to know, and what they just finished all land in the Inbox in the sidebar, with a count. There’s also an Inbox button at the top of the page for a quick look at the most urgent items.
The Inbox is what AI hands to you. The other direction, what you hand to AI, such as quick thoughts you haven’t sorted yet, is Capture, backed by your space’s 0_Inbox/.
Three groups
| Group | What’s in it | What you do |
|---|---|---|
| Need you | Decisions waiting in an issue; failed runs; issues you decided on whose next step couldn’t be dispatched | Answer, or dispatch again |
| You should know | What an agent wrote under an issue’s “You should know”; changes AI wrote into your notes | Read it, mark it seen |
| Completed | Issues that reached a closed state, with whether distill wrote the notes back | Glance, mark it seen |
You should know and Completed each have Mark all seen. Marking something seen only records that you read it: nothing runs and no note changes, and no work waits on it.
Deciding in the Inbox
When an agent needs your call, it writes it into the issue’s ## Decisions: one heading per decision, a few options, the recommended one marked. The Inbox lists each decision as an item next to what the issue is about.
- Pick an option and click Submit choice. Your choice is written back into the issue and that option is checked.
- If none fits, write a sentence in “Write your own option” at the bottom; it’s written back as “Other”.
- To talk it over first, click Discuss the whole thing. That opens the issue’s chat with the decision and its options quoted. Opening the chat isn’t an answer.
Once every decision on an issue is made, it moves to its type’s first Agent’s turn state and Automation dispatches the next run. After you decide, the item keeps telling you what happened next: queued, being worked on, finished, or not dispatched and why. If other decisions are still open, it says how many.
When a run fails
When a run fails or is stopped, Bonsai doesn’t retry by itself. It puts a Run failed card under Need you: which role, which round, on which runtime, and the reason the runtime reported. Dispatch again sends the same Automation row once more; the earlier run stays on record. If the cause is the runtime or the role (offline, no role note), fix that first.
When AI changed your notes
Every change AI writes into your notes over MCP stays pending until you’ve seen it. The Inbox groups them by note: who changed it, how many passages, each passage before and after, or the whole note. Marking it seen records that you read it; the note itself doesn’t change.
Explain
Every item has Explain. It opens a temporary chat on the right, running on your own runtime with the issue attached. It starts by explaining what matters in the item; you can follow up with “Why this recommendation?”, “What happens next?” or anything else.
The chat is read-only: it can read the issue and related notes but can’t write anything. It isn’t listed with your chats, it’s cleared when you close the panel, and reopening starts a fresh one. If your runtime asks permission before a tool call, the request shows up in the Explain panel and you answer it there.
With no runtime available, Explain answers from the item and its sources only, says so, and can’t take follow-up questions.