Insights & fleet benchmarks
The Insights panel surfaces analytics about your agent sessions and your repositories — so you see where time and tokens go.
Agent insights
Section titled “Agent insights”The Agents tab shows session-level analytics:
- Scorecards — turns, tokens, cost per session.
- Outcome scorecards — merged PRs in the window, and cost, turns, and sessions per merged PR.
- Charts — turn count and cost over time.
- Tool breakdowns — which tools each session used and how often.
- AI recommendations — suggestions for improving session efficiency, generated progressively from session data.
What a merged PR costs
Section titled “What a merged PR costs”The outcome scorecards answer which harness ships for less. Merged pull requests come from your local git history — the merge or squash commit subject carries the number — so the figures work offline, need no git-host credentials, and cover history that already exists.
A merged PR is matched to the sessions that produced it through the pull request a chat linked. Sessions with no linked merged PR are left out of the numerator rather than spread across PRs, so the cost is what those PRs actually cost. Merged PRs with no linked session are counted and shown separately — a low cost per PR should read as an incomplete link set, not as cheap work.
Set model rates to see currency; without them the cost cards show —, as the
session cost card does.
The Waste tab explains the Agents tab’s waste count. Instead of one number, it names the cause, how many analyzed sessions carry it, and which sessions to open:
- Context thrash — 20 or more turns without a context reset.
- Cache churn — three consecutive turns rewriting the prompt cache.
- Retry loop — one tool failed three or more times in the session.
- Low yield — 10 or more turns averaging under 200 output tokens each.
Every cause is a fixed rule over session data, not a model judgement, and each row states the rule that fired it. A session can carry several causes, so the per-cause counts sum to more than the session total — the tab says so rather than presenting them as a split of the whole.
The Agents and Waste tabs share one session analysis, so switching between them does not re-analyze anything.
Attention
Section titled “Attention”The Attention tab measures what the fleet costs you in interruptions:
- Interruptions in the window, and how many are still waiting on you.
- Median and p90 time to unblock — how long work sat waiting.
- Per merged PR — interruptions divided by merged pull requests, so the number reads as a cost of shipping rather than a raw count.
- Breakdowns by cause (needs input, error, work finished, marked unread) and by workspace.
An interruption is recorded on the same transition that lights the attention dot, so the history and the badge always agree. Durations cover resolved interruptions only — one still waiting on you has no duration yet, and counting it as zero would flatter the figure.
Recording starts when this version first runs. Until an interruption happens the tab says so, rather than showing zero and letting it read as a calm fleet. The record lives in the daemon’s local store and holds no message text.
Git insights
Section titled “Git insights”The Git tab shows repository metrics from local git history:
- Commit frequency over time.
- Contributor stats — who committed, how much.
- Stale branch detection — branches with no recent activity.
All git metrics are computed locally from your repository. No external calls, no model cost.