Skip to content

Session outcomes

A finished agent session is mostly noise to a reader who wasn’t there: hundreds of tool calls wrapping a handful of decisions that actually mattered. Capacitor now leads with the decisions. When a session ends, its detail panel opens on a new outcome-first Overview tab — and, when a summary model is configured, the server generates the outcome that leads it: a short, structured account of what the work accomplished.

Open any ended session and the detail panel lands on Overview rather than the transcript. It’s the outcome-first read: the summary at the top, then the supporting context, then the links you’d reach for next.

  • The outcome — a written account of the session in the CLI’s knowledge-base shape: the context it worked in, the key decisions it made, and anything unfinished or risky left behind. It captures the resolution, not just the opening move — long sessions keep their tail so the “how it ended” isn’t truncated away.
  • Plan & design — when the session produced a plan, spec, or design artifact, it’s promoted into its own card here rather than buried in the transcript.
  • Try next time — the retrospective’s most actionable suggestions, with deep links into the Evaluation tab for the evidence behind each one.
  • Links — an elevated strip of the PR, work item, and related sessions this one connects to, so Overview is also the jumping-off point to the rest of the work.

If a session’s outcome is still being written when you open it, the tab fills in live — no need to reselect or refresh. And a session that has no outcome yet (an older session, or one whose generation failed) shows a Generate outcome button for its owner or an admin to run it on demand.

The session detail panel’s tabs are now:

TabQuestion it answers
OverviewWhat did this session accomplish? — the default for ended sessions.
TranscriptWhat did the agent and I actually say? — the read-only replay (formerly “Chat”).
EventsWhat happened, in order, with what it cost?
TraceHow is the session structured? — turns and subagents rolled up.
EvaluationHow good was it? — the LLM-judge score and retrospective.
DetailsMetadata — repo, branch, PR, ownership, visibility, tokens.

A live session opens on Transcript (there’s no outcome to show until it ends), and hosted-agent sessions keep the live terminal as their landing tab. Every tab is deep-linkable with ?tab=… on the session URL.

Outcomes are generated server-side when a session ends, using the summary model your admin has configured. On a tenant with no summary model set, outcome generation cleanly no-ops rather than erroring — the session is still fully captured; it just won’t carry a written outcome until a model is configured. Admins tune this under What’s Done & classification in the console.

The same outcome text is what kcap recap prints and what session recall hands an agent as a session’s summary — Overview is simply its home in the dashboard.