Discussions
A chat-like discussion thread can be attached to any sprint-scoped entity — an agent question, a requirement, a user story, a task, the review, a general-info item, the inception prompt, or the sprint itself. Team members chat in real time, every message is durably stored as a vertex in the sprint graph, and the project's configured agent CLI can be invoked inside the thread.
Opening a thread
Look for the speech-bubble icon:
- on agent question cards (pending and answered),
- in the actions row of requirement / user story / task / general-info cards,
- on the inception "Project Description" card and the review card,
- in the Discuss tab of the activity panel, which lists every thread of the sprint with unread counts and a search box.
Threads are created lazily — one thread per entity. The first person to open the discussion creates it; everyone after that joins the same thread.
Threads open non-modally in the activity panel's Discuss tab on the right, so the rest of the app stays fully usable while a discussion is open. The back arrow in the thread header returns to the thread list.
Messages
- Markdown is supported (no raw HTML).
@opens a mention picker over the project members. Mentioned users who are currently online get an in-app notification with a jump-to-thread link. Offline users see the unread badge on their next visit — there is no email or push delivery in v1.- Messages are persisted first, then delivered: even if your tab crashes right after sending, the message reaches everyone.
- Unread badges only clear when you have actually seen the newest message (the thread is scrolled to the bottom in a visible tab) — merely opening the thread does not mark anything read. The thread opens at a "New" divider marking your first unread message.
Resolving a thread
Any project member can resolve a thread (the action is fully audited — the resolver's name is shown in the thread header). When resolving you can record a resolution summary — the durable "what did we decide" — and optionally mark one message as the accepted outcome. Resolution summaries are visible to the platform's agents as team decisions: inception/construction/review agents are instructed to consult discussions on the entities they work on.
Resolved threads can be reopened by any member.
Moderation (redaction)
Project admins and owners can redact any message from the message context
menu. Redaction replaces the content with [redacted by {name}] — the
original text is purged from the database while the audit trail (who redacted,
when) is preserved.
Known limitation: copies of the original text may persist in the memory of clients that already rendered it, and in prompts already sent to an in-flight agent assist.
Agent assist
The ✨ button in the message input (or / in an empty input) invokes the
project's configured agent CLI inside the thread:
| Command | Behavior |
|---|---|
| Suggest an answer | Question threads only. Recommends selections/free-text for the anchored question — advice only; the agent never fills in or modifies the answer itself. |
| Summarize discussion | Decisions, open points, and positions, attributed by participant. |
| Explain in detail | Explains the anchored entity in the context of the sprint graph. |
| Ask the assistant… | Free-form instruction, scoped to the thread and sprint. The agent advises humans and never mutates artifacts. |
What to expect:
- Assists run on the same agent pool as regular agent runs, so the first token can take 15–60 seconds — the thread shows an explicit "Assistant is starting…" state, then streams the reply live.
- One assist per thread at a time; a second request is rejected until the first finishes.
- The reply is posted as a regular message with a "requested by {name} · {command}" caption, and the run appears in the sprint's agent-run history. Assists never affect the sprint's phase status.
Roles at a glance
| Action | Member | Admin / Owner |
|---|---|---|
| View threads & messages, search | ✓ | ✓ |
| Post messages, mention, typing/presence | ✓ | ✓ |
| Resolve / reopen (audited) | ✓ | ✓ |
| Invoke agent assist (audited) | ✓ | ✓ |
| Redact a message | ✗ | ✓ |
Non-members get a 403 on everything, including realtime access.
Design notes
- One thread per entity (v1): busy entities may mix decision and implementation topics in one thread. The data model is ready for multiple titled threads without migration when the need materializes.
- Messages live in the sprint graph (
Discussion/DiscussionMessagevertices), so discussions are queryable collaboration context for both humans and agents — not throwaway chat history. Deleting a sprint deletes its discussions. - Search is sprint-scoped substring search (minimum 3 characters); full-text search is a named follow-up.