Team analytics

CodeLotse reports how reviews, triage, assignment suggestions and fixes work out for a team, and what the model calls cost. Analytics are team-level: there is no per-person view for anyone except the person themself, no ranking, no leaderboard and no per-person export.

Where to find it

Screen	Who	Content
Analytics → Team metrics	Tenant admins	Whole tenant, a team (group of repositories) or one repository
Analytics → Costs	Tenant admins	Tokens, cost and prompt-cache hits for the tenant
Analytics → Teams	Tenant admins	Teams are groups of repositories; people are never assigned to teams
Analytics → Settings	Tenant admins	Minimum group size, personal metrics switch, budget alert thresholds
Analytics → My metrics	The person themself	Only when allowed (see below)
Repository Analytics tab	Everyone who can view the repository	Team metrics of that repository; costs for repository admins
Settings → Evaluation runs	Instance admins	Runs recorded with quavon eval full --record-db

Every chart has a table alternative ("Show as table"). Values can be downloaded as CSV (format=csv).

Privacy

• Minimum group size (k). A point (period, and dimension value when split) is shown only when at least k distinct people contributed to it (default 5, configurable 5 to 50). Otherwise the point is hidden and never exported. Bots never count. The population is the authors of the pull requests or issues the events belong to; it is computed from the source tables at query time and never returned (no head counts are shown).
• Admin-only wide scopes. Tenant and team scopes need tenant:manage. Members see repositories they can view.
• Differencing. Comparing overlapping scopes (tenant minus team, team A minus team B) can narrow a group. Each scope applies k on its own, and the wide scopes are admin-only; the residual risk remains and should be covered by the works agreement.
• Personal metrics (/api/me/tenants/:tenantId/analytics) need the tenant switch Allow people to see their own metrics, the person's own opt-in in their assignment settings (people.individual_metrics_opt_in) and no team-only metrics (works-council mode or the tenant setting). They exclude repositories whose assignment policy sets team-only metrics. Nobody else, including tenant and instance admins, can read them. Users signed in through OIDC have no linked person and get 404.
• Costs are not about people and are exempt from k; they are only available to tenant or repository admins.

Metric dictionary

The period of an event is the UTC day of the timestamp named below. Weeks are ISO weeks starting on Monday; months are calendar months (UTC). Rates are numerator / denominator summed over the period.

Metric	Definition	Day of	Population	Split by
acted_on_rate	Findings whose feedback outcome is acted / findings with outcome acted or dismissed. The latest event per finding and day counts.	feedback_events.occurred_at	PR authors	category, severity (from the event)
dismissed_rate	dismissed / (acted + dismissed), same scope as above	feedback_events.occurred_at	PR authors	category, severity
comments_per_pr	Posted comments created that day / distinct pull requests with a completed or silent review run that day	posted_comments.created_at, review_runs.finished_at	PR authors	–
silent_pr_rate	Review runs with decision silent / review runs with status completed or silent	review_runs.finished_at	PR authors	–
findings	Findings with status published	findings.created_at	PR authors	category, severity
time_to_first_review	First finished_at of a completed or silent review run minus pull_requests.opened_at (fallback created_at); mean, median and 90th percentile	first review	PR authors	–
time_to_merge	merged_at minus opened_at (fallback created_at) of merged pull requests; mean, median and 90th percentile	pull_requests.merged_at	PR authors	–
triage_label_acceptance	label_kept / (label_kept + label_removed); latest event per triage run, label and day	feedback_events.occurred_at	Issue authors	–
duplicate_hits	Triage runs with at least one entry in result.duplicates	triage_runs.created_at	Issue authors	–
assignment_kept_rate	Assignments with outcome kept / assignments with an outcome; only suggested or applied rows, never shadow	assignments.outcome_at	Authors of the assigned PRs and issues	–
fix_pr_merge_rate	Published fix runs whose pull request is merged / published fix runs	fix_runs.finished_at	Issue authors	–
llm_tokens	Input plus output tokens; split by token kind gives input, output, cached and cache-write tokens	llm_calls.created_at	– (cost)	feature, model, tier, token kind
llm_cost_eur	Sum of cost_eur	llm_calls.created_at	– (cost)	feature, model, tier
cache_hit_rate	Cached tokens / input tokens	llm_calls.created_at	– (cost)	feature, model

Percentiles are interpolated from duration buckets (DURATION_BUCKETS_SECONDS); values above the largest bucket are reported as that bound. A split a metric does not support is reported as dimension_not_supported. Model calls without a repository count towards the tenant scope only.

Rollups and retention

The worker task analytics_rollup runs hourly. Per tenant it recomputes the last 3 days (ANALYTICS_ROLLUP_LOOKBACK_DAYS) and, on the first run, the last 90 days (ANALYTICS_BACKFILL_DAYS) into analytics_daily (additive numerators and denominators per repository, day, metric and dimension value; no person columns). Re-runs replace the same days. Rows older than the tenant's data retention are deleted. The screens show when the data was last updated.

Pull requests store opened_at, closed_at and merged_at from GitHub from this release on; older pull requests fall back to created_at and are not backfilled.

Budget alerts

The worker task budget_alerts runs every 15 minutes and compares model spend (llm_calls.cost_eur) in the current budget period with the tenant budget and each repository budget (model policy Budget). The period starts at floor(epochDay / periodDays) * periodDays (UTC). Each threshold (50, 80 and 100 % by default, chosen in Analytics → Settings) raises one alert per scope and period. An alert appears as a banner on the analytics and overview pages, is written to the audit log (budget.alert, actor system) and logged as a warning. Acknowledging it is audited. Each new alert is also sent to the notification channels that subscribe to budget alerts (Slack, Microsoft Teams, e-mail or a signed webhook; see notifications.md).

CSV export

GET …/analytics?…&format=csv (and /costs) returns metric,dimension,dimension_value,period,value,numerator,denominator. Hidden points are left out. Cells starting with =, +, -, @, tab or carriage return are escaped against formula injection. Responses are sent with cache-control: no-store. Personal metrics cannot be exported.