Copilot Agent Metrics Missing? Check Your IDE Version
Your team edits with Copilot agents every day, but the usage dashboard shows agent activity falling. Where did those sessions go?
- Check client versions before interpreting a decline in agent activity.
- Use the agent-edit fields for direct file changes, rather than completion suggestions.
- Allow for reporting delay and verify telemetry coverage after updating.
Missing activity can reflect a client problem
GitHub's October 6 notice identifies an attribution failure in some IDE versions that moved agent sessions to the Copilot SDK. Sessions omitted their IDE identity, leaving activity unreported or misclassified as Copilot CLI use. Billing was unaffected. VS Code 1.139.0 and later include the fix; missing activity and past CLI misattribution cannot be reconstructed.
Inspect versions before planning the rollout
The per-user report fields offer a practical starting point. Within totals_by_ide, inspect last_known_ide_version and last_known_plugin_version to identify environments that need attention.
The general supported-IDE minimums describe eligibility for metrics, not necessarily a fix for this later regression. For other IDEs, verify the corrective release in that product's release information. The dates in the October 6 announcement were rollout forecasts, not proof that every client now has the patch.
Direct edits do not follow the completion model
A second source of confusion is expecting agent-written files to appear as accepted inline suggestions. GitHub's LoC reference treats direct agent edits separately:
| Question | Field or feature |
|---|---|
| Did this person use IDE agent mode? | used_agent, a usage flag rather than a volume |
| How much did the agent change in files? | agent_edit, with loc_added_sum and loc_deleted_sum |
| How much code was suggested? | loc_suggested_to_add_sum, excluding direct agent edits |
One prompt can trigger edits across several files, and each edit contributes to the totals. Prompt counts, displayed suggestions and applied changes are different measurements. A zero in a suggestion field is not enough to establish that no file changes occurred.
Allow time for fresh data, then check coverage
GitHub says reports become available within two full UTC days after the reported day closes. An immediate dashboard refresh is therefore a poor test of an IDE update.
Server-side signals can also identify an active user while feature breakdowns and code counts remain empty. Richer details depend on client telemetry. Check IDE collection settings and proxy or firewall delivery under your organization's telemetry policy.
Mark the update date in internal trend reports. A subsequent rise may include restored coverage as well as a real change in adoption. Treating missing historical activity as zero usage, or treating the entire rebound as increased output, would hide the change in measurement conditions.