Version 3.0
Build
Data Collections
Recent activity

Recent activity and intent history

Every question someone asks against a Data Collection lands here. Use it to read the most recent runs, drill into one, and spot recurring questions worth promoting to a Skill or a Verified action.

Audience: builders

Where Recent activity shows

The most recent six intent runs surface in Recent activity on the Dashboard's left column. Click a row to expand it. The Dashboard preview is the always-on view; you do not need to leave the page to read the latest summaries.

For the full list, scroll back through expanded rows. A deeper paginated history surface is not yet a separate page; see the product gap log.

Read the row shape

Each row carries a fixed payload:

  • Type icon. What lane the agent picked: Question, Definition, Write preview, Write applied, Catalog edit, Agent fallback.
  • Title. A short label our AI synthesizes from the request.
  • Status. Succeeded, running, queued, failed, cancelled. The dot color follows.
  • Updated at. When the run ended (or last advanced).
  • Badges. Optional chips that surface the source, the lane, or other run metadata.
  • Summary. A markdown summary of what ran. Click the row to expand.

Drill into one run

Click a row's chevron to expand. The expanded panel shows:

  • The full type label and timestamp.
  • Any badges from the run.
  • The full markdown summary (the same summary_markdown the analyst view reads).
  • An Open task button when the run is linked to a background task flow.

Open task is gated. The button appears only when the run carries a task_flow_id. Today most rows do not (backend brief: data-collection-recent-activity-task-link-brief.md). When the field is missing, the summary is your record of the run.

Filter recent activity

Filtering by question text, source, or time window is not yet wired into the Recent activity surface. You can scan by type icon (Question, Definition, Write preview) and by status tint, but query-shaped filters require the future history page.

Until then, two practical paths help:

  • Drill in. Expand a row to read the summary and decide whether the question deserves a follow-up.
  • Audit. For workspace-wide search across every collection, the audit ledger at /audit filters by actor, time, and group kind. Each intent run writes ledger entries.

Privacy and visibility

Recent activity is a workspace surface, not a personal one. Every signed-in workspace member who can see the collection sees its recent activity. The list is not filtered to the current user.

For audit-grade evidence (who asked what, with what credentials, against which tables), use the audit ledger. The Recent activity card is the operator's quick read; the audit ledger is the controller's receipt.

Spot work worth promoting

The list is the cheapest way to notice patterns. A few signals to watch for:

  • The same question, different phrasings. Three rows asking variations of "what is our AP exception volume this quarter" is a Skill candidate. Save it once and let your team call it by name.
  • A definition that keeps coming back. If analysts keep asking what vendor_status means, the catalog entity's description is too thin. Open Catalog and tighten it.
  • A write preview that gets repeated. A change someone keeps previewing but never applies is either a Verified action waiting to be authored, or a guardrail you should encode explicitly. See Verified logic.
  • A failing pattern. Failed runs that share a source or a phrasing point at a scope gap or a missing entity. Open Setup → Selection and add the table; rebuild.

Related

DashboardDocument intelligence