Decisions
How an agent asks you for a call with request_input, how you answer on the Decisions page, and how the answer gets back to the agent.
In Phean, a decision means one thing: an agent's request for a person's call. A question asked only in chat is lost when the chat scrolls away. A request recorded with request_input waits on your Decisions page with the agent's context, its links, and a one-click way back into its conversation, until you answer.
Question or decision
| The agent needs | kind | Example |
|---|---|---|
| A choice, approval or sign-off | "decision", with options | "Offer Globex 5% or escalate to their CFO?" |
| A fact or detail only you have | "question" (the default) | "Which cost center do the offsite flights go to?" |
A request belongs to a task. If no task fits, the agent captures one first. A task holds one open request at a time: asking again on the same task replaces the agent's earlier request. For several independent decisions, the agent uses several tasks or subtasks.
The request
request_input takes these fields:
| Field | What goes in it |
|---|---|
taskId | The task the request unblocks. |
kind | "decision" or "question". |
question | One concrete question, answerable in a word or a line. Up to 1,000 characters. |
context | The facts, the trade-offs, and the agent's recommendation, for someone reading it cold days later. Up to 5,000 characters. |
options | For decisions: short, mutually exclusive answers. The guidance asks for 2 to 5; Phean accepts up to 8, each up to 200 characters. You can always write your own answer instead. |
multiple | true when several options may apply together. Needs at least two options. |
links | Up to 10 references, as [{ url, title, kind }]. |
session | Where you can resume with the agent: { url, app, brief }, or sessionId or command when there is no link. |
Multiple answers
With multiple: true, you tick any number of options and send them together, and the agent gets every one in chosenOptions. The options need not be exclusive then. The question should read that way: "Which channels should announce the launch? Pick any that apply."
Links
Each link has a kind, which sets how it shows on the request:
conversation: back to the chat or thread with you. Agents add one to every request when their client exposes the conversation's URL.source: the permalink of what the question is about: the email, Slack message, calendar event, ticket or PR.doc: a draft, spec or file. When the decision is about a draft with no URL of its own, the agent attaches it withattach_documentfirst and links theurlit returns (see Documents).other: anything else.
Links use https:// or an app scheme (slack://, cursor://, vscode://, claude:// and a few others). Phean refuses other schemes. Agents are told never to invent a URL; if they have none, they say in context where to look.
Session and the way back
The session is how you get back to the agent. On the overview and the Decisions page, Resume in Claude (or the agent's app, from session.app) leads the request's entry when the agent left a link. session.app names where you reach the conversation (Slack, Claude, Cursor), never the bot itself. Without a link, Resume lists the session id and resume command, ready to copy.
A complete example
{
"taskId": "t_8k2m4q9w1z3x",
"kind": "decision",
"question": "Globex disputes invoice #4471 ($2,400) and asks for 10% off. Offer 5% to settle this week, or escalate to their CFO?",
"context": "They have paid on time for two years and the dispute is about a late delivery we caused. I recommend offering 5%: it costs $120 and keeps the account.",
"options": ["Offer 5% this week", "Escalate to their CFO"],
"links": [
{ "url": "https://claude.ai/chat/5d0c2e7a-91b4-4a8e-b1f3-2c6e8a4d7f10", "title": "This chat", "kind": "conversation" },
{ "url": "https://mail.google.com/mail/u/0/#inbox/globex-4471", "title": "Globex's dispute email", "kind": "source" }
],
"session": { "url": "https://claude.ai/chat/5d0c2e7a-91b4-4a8e-b1f3-2c6e8a4d7f10", "app": "Claude", "brief": "Two of three invoices are paid; Globex disputes #4471." }
}
request_input returns the task and a decisionsUrl. The agent tells you in one line what is waiting and links it, batching several requests into one message.
Answering
The Decisions page (/w/<slug>/decisions) shows what waits, with context, options and links, and what was answered, with whether the asking agent has picked the answer up. Requests also appear on the overview.
- A decision with options: click an option. On a request that takes several, tick them and Send.
- Your own answer: Other answer (or Answer here for a question, Decide here for a decision without options) opens a short reply.
- Answered in the conversation: in the entry's ⋯ menu. Use it only when you already answered the agent in its own conversation. It asks you to confirm.
After you answer, the status region says what was sent and offers Undo. Undo takes the answer back and reopens the question exactly as it was asked, as long as the agent has not collected the answer yet. Once the agent has it, tell it in its conversation if your answer changed.
Keyboard
On the Decisions page and the overview, one request is always selected.
| Key | Does |
|---|---|
1 to 9 | Pick the selected request's option (or tick it, where several may be picked). |
Enter | Send the ticked options, where a request takes several. On the selected entry itself, open its task. |
J | Next request. |
K | Previous request. |
? | Show all keyboard shortcuts. |
Keys never fire while you are typing in a field.
How the agent gets the answer
Each answer is delivered to the asking agent once, through either of two tools:
wait_for_answersblocks until you answer one of its requests (only ontaskIds, if given).timeoutSecondsdefaults to 55, because some clients give up on a tool call after 60 seconds, and can go up to 240. If nobody answered in time it returnstimedOut: truewithstillOpen; the agent says what it is still waiting on, then waits again or does other work.collect_answersreturns the answers since it last collected, oldest first. Agents call it at the start of every session and before acting on work that had an open request.
A bot with an HTTPS endpoint can be told the moment you answer, through a webhook (topic decisions) or the realtime stream. It still takes the answer through collect_answers, which marks it delivered.
Outcomes
Two outcomes reach the agent:
answered:answerholds what you wrote or picked.chosenOptionnames the option when the answer matches exactly one, andchosenOptionslists every option picked. A custom answer is authoritative even when it matches no option.answered_elsewhere: you marked it answered in the agent's own conversation; the agent reads it there.
A request can also end without reaching the agent as an answer: the agent closes its own with resolve_input once it has the answer some other way (resolved), and completing, releasing, or moving the task out of progress closes it (closed). Every closed request stays on the task's history. On the Decisions page, settled requests sit under Answered, newest first, each as one line: the question, what was picked, when, and whether the agent has it yet. Open a line for the full record (every option offered, the written answer, the links), or open its task in the side peek. Past the latest eight, the rest fold under Show earlier answers.
If you take an answer back before the agent collects it, a listening bot gets a task.input_retracted event, the question is open again, and collect_answers returns nothing for it until you answer again.
After the answer
The agent acts on it: it updates or captures follow-up tasks, does the work it can, and says what it did. A choice that will matter again goes into memory with remember (kind decision, shown as "Choice"), so other agents do not ask you the same thing.
The phean-decisions skill teaches agents all of this with the same example.