Reporting

The admin Reports page summarizes delivery, quality, budget, automation, and human-approval activity. The selected UTC date range applies to delivery, quality, automation, and approval reports. Budget reports always use the current UTC calendar month. Each report tab exports its displayed sections as a client-generated CSV. The Digests tab manages scheduled weekly reports.

Platform analytics time saved is the sum of each successful run’s estimate, snapshotted from the template when the run started.

Workflow runs store nullable reporting-only team and organization attribution. The team resolver uses the run connection, work-request connection, channel, requesting service, then template, in that order; the selected team’s organization wins, with the work request’s billing organization as the fallback when no team is available. This attribution does not change run visibility.

Pull requests store their template id, template version and associated run id. Writers and the historical backfill associate a PR only through an active workflow in the same repository. WorkflowRun.workflowId is unique, so the associated run is resolved by that exact workflow id, not by recency. Outcomes, per-version template analytics and report cohorts read these persisted fields instead of dynamically joining recent runs.

Delivery

Delivery cohorts are workflow runs whose startedAt falls inside the selected window. The funnel counts runs started, successful runs, and runs linked to at least one pull request. CI-passed and merged values count pull requests, not runs. Cost per merged pull request divides the cost of all cohort runs by the number of merged pull requests; it is unavailable when no pull request merged. Median time to merge measures from the cohort run’s start to each linked pull request’s merge.

Template and repository slices use the same funnel. Repository health also includes spend and run success rate from visible cohort runs grouped by repository. Delivery spend uses each run’s WorkflowRun.costUsdAccrued; when that value is zero, the matching ActiveWorkflow.costUsdAccrued is used as the legacy fallback. Totals, template and repository slices, and cost per merged pull request use that same run-level amount. Pull requests whose repository does not match their linked workflow are not attributed to a run. Delivery also groups the cohort by persisted team attribution, including an Unattributed row. The optional teamId filter applies to both the daily rollup and live slices. Delivery labels repository, template, and team slices; inaccessible repository names appear as Inaccessible repository, and templates hidden by the workflow-template visibility rule appear as Private template, without dropping their counts.

Quality

Gate metrics are grouped per (run, gate) pair. The first-attempt pass rate is the share of pairs with a PASSED row and no FAILED row; the final pass rate is the share with any PASSED row. Attempt counts sum PASSED and FAILED rows. Fix-loop attempts count terminal PASSED and FAILED rows and report both runs with a fix and average attempts per such run. CI self-heal rate is the share of runs with a CI fix that have a linked pull request with passed CI or a merge. The self-heal and merge-rate cohorts combine rollup and live rows using persisted run/PR attribution.

Review verdicts are parsed defensively from reviewer trace output. The stored trace agent key is reviewer; the individual persona is recorded in the verdict’s reviewer field (SECURITY, DOMAIN_LOGIC, or PERFORMANCE). Traces are restricted to cohort run ids, exclude records with an error and records without an approved key, and do not use a separate trace timestamp window. Remaining malformed verdict payloads are excluded from reviewer statistics and counted as nonVerdictRows. The merge comparison groups cohort runs with pull requests by whether any parsed verdict rejects the run.

Budget

Each authorized organization shows its monthly budget, alert threshold, month-to-date spend, elapsed UTC days, month length, and projected spend. Projection uses fractional elapsed days. A projection from less than one elapsed day is still shown and marked low confidence. The status prioritizes actual overspend, projected overspend, an alert threshold, and then an in-budget state. Organizations without a cap show no_cap. The report lists the five highest-cost runs started in the current month for each organization, attributed with the same organization ownership precedence as month-to-date spend. Top runs are filtered by the caller’s run visibility, so they may be a partial view of the organization’s highest-cost runs.

Automations

Automation activity counts fires and their outcomes and reasons, last fire time, starts, failed starts, and suppression categories. An automation is noisy when it has at least ten fires and more than half are suppressed. It is broken when a start fails or when at least three starts have no successful produced workflow run.

Approvals

Approval reporting counts human steps by kind and status, resolution waits for resolved steps, timeouts, currently pending steps, and the oldest pending age. Approvers are grouped by the resolving user with interactive (web, slack) and automated (api_token, mcp, assistant) counts. Resolved steps are also split by resolvedVia; null or unrecognized historical values are reported as unknown. Autonomy decisions are counted by event and risk class.

Data bounds

Report live reads are capped at 10,000 rows and responses include a truncated flag when any bounded live source exceeds that bound, alongside a per-source truncation flag. Quality workflow-step metrics use an uncapped groupBy over the bounded run cohort. Budget top-run lists intentionally show at most five runs per organization; each organization indicates when more ranked runs exist.

Weekly digests

Digest subscriptions target exactly one organization or team and deliver weekly summaries by email, Slack, or both. They compare the last seven completed UTC days with the seven days before them, including delivery metrics, top templates, the worst gate when at least ten run/gate pairs exist, pending approval age, and organization budget status. Organization subscriptions include budget data; team subscriptions do not. Send-test delivers immediately without changing the scheduled last-sent timestamp.

The weekly sweep runs Mondays at 13:00 UTC when REPORT_DIGEST_ENABLED is true. Configure REPORT_DIGEST_CRON to change the schedule. Digests require workflow-run rollups to be enabled; REPORT_DIGEST_CRON must run after WORKFLOW_RUN_ROLLUP_CRON. The sweep skips delivery and retries on its next run when any of the 14 required completed UTC days has no rollup rows. A subscription is claimed once per reporting period before delivery, so a failed email or Slack channel is recorded but is not retried in the same period. Send-test bypasses the weekly claim and does not change the scheduled last-sent timestamp; its one-minute limit is stored in Postgres. Email uses the platform SMTP configuration when available, otherwise the configured Resend transport.

Limitations

• CLI and MCP report commands are not available.
• Non-ADMIN reporting visibility is based on persisted attribution: a row is included when its teamId (WorkflowRun.attributedTeamId) is one of the caller’s member teams, or its orgId (WorkflowRun.attributedOrgId) is an organization where the caller is ORG_ADMIN. ADMIN sees all rows, including null-attribution rows; null-attribution rows are excluded for everyone else. This rule is applied to every slice in a rollup-backed response, including headline metrics, group-bys, selected runs, steps, and quality traces. ORG_ADMIN therefore sees every run attributed to their organization even when buildWorkflowRunVisibilityFilter would hide it; requester-only and shared-connection visibility do not apply to these reports. A teamId filter is allowed only for a caller’s member team or a team in an organization they administer; otherwise the route returns 403. Non-rollup reads retain the normal run-visibility filter and add the selected team filter.
• Template names hidden by the workflow-template visibility rule, channel names hidden by channel access, and repository names blocked by the repository-access gate are masked as Private template, Private channel, and Inaccessible repository; the corresponding counts remain.
• Run and merge duration percentiles are approximate: values fall into logarithmically spaced buckets and the reader returns the bucket’s geometric midpoint. CI-fix p95 and average use an exact attempt-count histogram for 0–19; its 20+ overflow bucket is represented as 20.
• Rollups are rebuilt daily and recomputed over the configured WORKFLOW_RUN_ROLLUP_RECOMPUTE_DAYS horizon to include late outcomes. Once a day is older than that horizon it is frozen; later merges and run finalizations are not reflected in that day. rollupDays and liveDays show which dates came from each source; truncation flags apply only to live reads. Global analytics daily series and duration percentiles, template per-version duration/PR statistics, and delivery metrics use this hybrid reader. Duration and merge percentiles use logarithmically spaced buckets and return a bucket midpoint; the CI-fix histogram is exact for 0–19 attempts and represents 20 or more as 20. Freshness is bounded by the daily schedule and the recompute window. Set WORKFLOW_RUN_ROLLUP_BACKFILL_DAYS to populate older history; the default of zero leaves history outside the recompute horizon on the live path.
• Approval source counts use resolvedVia; historical rows with no source are reported as unknown.
• User records do not expose a service-principal or API-token approver flag, so approval counts cannot separate people from machine identities. The human-step HTTP route requires a user principal and does not admit service-principal tokens.
• Historical pull requests without a repository-matched active workflow and run remain unattributed.
• Historical human-step and approval records do not have a resolution source and remain null.
• Historical runs whose workspace connection (runConnectionId) differs from their repository may be attributed differently by the backfill than by runtime attribution.