Skip to Content
OpenInspect
Administration

Analytics

What the Analytics page measures, how each number, chart, and table is calculated, and how session cost is derived.

Last reviewed View as MarkdownEdit on GitHubGive feedback

The Analytics page in the sidebar shows usage across sessions, repositories, and users for a chosen window, plus outcomes for the pull requests those sessions opened. It has five tabs: Overview, Usage, Cost, Pull requests, and People.

Who can view it

Every role can open Analytics; it requires only analytics.read, which Viewers have. The page refreshes itself every 30 seconds; the header shows when the numbers were last updated.

Time range

The Time range control offers 7d, 14d, 30d, and 90d. The default is 30 days. Changing the range re-filters windowed metrics on every tab. While the new range loads, the previous numbers stay on screen, dimmed, and the header reads Loading…. If the new range fails to load, the page says so and keeps showing the previous selection until a retry succeeds.

The window is the last N days ending now. Session metrics count sessions created in the window; open-PR inventory is measured as of now regardless of the window. Daily charts plot every day in the window, so a day without sessions shows as zero, and the last day covers only today so far.

The range, scope, and tab are part of the page URL, so a link opens the same view.

Which sessions are counted

The Scope control offers Human (the default: sessions with user, slack-bot, linear-bot, or github-bot spawn sources, including bot-triggered sessions), Agents (sessions spawned by other sessions), Automations (sessions started by automations), and All (every session). The info button beside it explains each option.

Scope filters every session number, chart, and table; the Automations table on the Usage tab appears for Automations and All. Pull request numbers are different: they count every pull request created through OpenInspect regardless of session scope, including automation runs. The Overview labels them Pull requests · every source, and on the Pull requests tab the scope control is turned off.

About this data in the header lists the page's caveats. Older sessions recorded before cost, pull request, and duration values were tracked are still counted and may show zero for those values. PR counts reflect pull requests created through OpenInspect's own pull-request flow.

Overview

The headline numbers come in two groups. The first is captioned with the scope (for example Human sessions):

NumberValue
SessionsSessions created in the window, with the daily average and a sparkline of sessions per day.
Active usersThe number of distinct users with a session in the window, keyed by user ID and falling back to the source-control login for sessions without a linked user. The hint shows attributed sessions per person; sessions with no recorded user are left out of it.
SpendThe sum of every counted session's recorded cost, with the average per session.
Completion rateCompleted divided by completed + failed + cancelled. Sessions still running, never started, or archived are not in the denominator. A dash until a session finishes.

The second group, Pull requests · every source, shows PRs merged (pull requests merged during the window, with their average time to merge and a sparkline of merges per day) and Cost per merged PR (for pull requests opened in the window; defined under Pull requests).

Below them, Sessions per day and Pull requests opened and merged chart the window, and three lists show the top five repositories by sessions (with cost), models by spend (with cost per session), and people by sessions (with completion rate). Each panel links to the tab with the full detail.

Usage

Sessions per day charts the daily session count across the window.

Where sessions start ranks session sources (User / app, Slack, GitHub, Linear, Agent sub-sessions, Automations) by sessions, beside the users those sessions are attributed to. Select a source to narrow the users to it; All sources resets. Sources describe how a session was created, not later messages. User / app includes user-authenticated creation and historical sessions whose source defaulted to user; it does not mean browser-only usage. Users reflect recorded attribution: the requesting actor, the attributed user for an agent sub-session, or the automation owner or manual triggerer. How attribution works under the panel repeats these rules. Private sessions are excluded.

How sessions ended shows the completion rate and the count of sessions in each status: completed, failed, cancelled, archived, running, and never started. See Lifecycle and statuses.

Repositories lists each repository's sessions, completion rate, PRs, messages, and cost. When every repository shares one owner, the owner is shown once above the table. Sessions with no repository are grouped under "No repository".

Automations appears for the Automations and All scopes and lists each automation's sessions, cost, cost per session, completion rate, and PRs.

Cost

The numbers across the top:

NumberValue
SpendThe sum of every counted session's recorded cost, with the average per session.
Cache hit ratioCache reads divided by cache reads plus input tokens.
Billed to subscriptionsThe share of sessions billed through a connected subscription. Those sessions report model-price equivalents, not charges.
TokensInput plus output tokens, split into each.
Private sessionsThe cost of private sessions in the same window and scope, which Spend excludes. Only Owners and Administrators see it.

By model, By provider, and By harness list sessions, cost, and cost per session for each, sorted by cost. The provider table also shows how many sessions ran On subscription; the harness table shows the completion rate. Narrow screens hide secondary columns such as PRs and cache hit ratio.

Most expensive runs lists the costliest runs in the window, up to 20. A run is a root session plus every sub-session it spawned, so its cost and session count include the whole family. Each run links to its root session.

Tokens gives the totals for input, output, cache read, cache write, and reasoning tokens. Reasoning tokens are reported by OpenCode and Codex. Token totals cover sessions created since token capture was enabled; older sessions count as zero.

Pull requests

Rates on this tab are computed over pull requests, not sessions. Sessions that never open a pull request (questions, research, debugging) are out of scope here by design.

NumberValue
OpenedPull requests opened in the window.
Acceptance rateMerged divided by resolved, where resolved is merged plus closed without merging. Still-open PRs are not in the denominator. Shown as a dash until something has resolved. The hint reads "N merged · M closed unmerged".
Avg time to mergeMean time from open to merge for pull requests merged in the window; the hint counts "N merged in range". A dash when nothing merged.
Open nowPull requests open right now, not limited to the window, with the average age, or "Nothing waiting on review".
Cost per merged PRThe total cost of sessions that produced a pull request in the window, divided by the number merged. A dash until something has merged.

Pull requests over time plots daily opened versus merged counts. Opened is bucketed by the day the PR was opened; merged by the day it merged.

Where they stand splits the pull requests opened in the window into merged, open, draft, and closed unmerged, and lists By source where the sessions behind them came from (user or app, automation, Slack, and so on) with how many merged.

By repository lists, for pull requests opened in the window: Opened, Merged, Closed, Acceptance (the same merged ÷ resolved definition per repository), and Avg time to merge.

Cost per merged PR breaks that figure down by the Models or Harnesses of the sessions that produced the pull requests.

People

A sortable table of usage per person. Click any column heading to sort; click again to reverse.

ColumnMeaning
UserDisplay name or login. "Unknown user", marked "Sessions without linked user", collects sessions with no user record.
Daily sessionsA sparkline of the person's sessions per day across the window.
SessionsSession count.
CompletionCompleted divided by completed + failed + cancelled. Sessions still running are not in the denominator; a dash until one finishes.
PRsPull requests created by the person's sessions.
MessagesPrompts sent across the person's sessions.
CostSum of recorded cost across the person's sessions.
Avg durationAverage active duration per finished session; shown as a dash when there is none.
Last activeRelative time of the person's most recent session activity.

People by source is the same breakdown as Where sessions start on the Usage tab.

How cost is derived

Cost figures are sums of each session's recorded cost. A session's cost is a running total that the agent harness reports to the control plane as each turn finishes:

  • OpenCode reports the priced cost of each step, and the session total accumulates those values.
  • Claude Agent reports a cumulative cost per turn, and the session records the increase over the previous turn. When the harness reports no cost for a turn, the turn is recorded as $0 and the session timeline shows a warning saying so. The next turn's cost cannot then be separated from the missing one, so it is also recorded as $0 before accounting resumes.
  • Codex reports the cumulative cost of each turn, priced at the model's published rates using each request's context tier from the sandbox's cached price list. Under a ChatGPT subscription the figure is informational. Unavailable model prices (for example a model newer than the sandbox's price list) result in $0 reported cost. Partial or estimated costs can be nonzero, contribute to analytics totals, and still count toward spend limits. The session timeline shows a warning explaining missing, partial, or estimated pricing.

Nothing on this page estimates cost independently of what the harness reported. Amounts show cents below $1,000 and whole dollars above; amounts under one cent show as <$0.01. Per-session cost caps are covered in Spend limits.

Troubleshooting

Next steps

On this page