Pimp My IDE / garage dispatchBack to the garage
October 9, 2026 / telemetry / agent operations

A falling chart can be a broken witness.

GitHub found that some IDE agent sessions lost their IDE identity after moving through the Copilot SDK. The work continued. The report changed.

Treat every adoption chart as a measurement system with versions, routes, and failure modes.
ATTRIBUTION SCOPE / 28 DAY REPORTDIAGNOSE
Agent editsDown
CLI bucketUp
Active usersSteady
Pattern is a clue. It is not a verdict on developer behavior.

The missing work was not recoverable.

GitHub reported that several IDEs moved Copilot agent sessions to the Copilot SDK without identifying the source IDE. The usage system omitted most of that activity and counted some of it as Copilot CLI activity. Billing did not change.[1]

The fix reached Visual Studio Code in version 1.139.0. GitHub listed later fixes for Visual Studio, JetBrains IDEs, Eclipse, and Xcode. Old records cannot be backfilled because the missing IDE identity was never sent. Reports recover as clients update.

A dashboard can be internally consistent and still describe the measurement route more than the work.

One chart mixes several witnesses.

GitHub's metrics documentation says most detailed IDE metrics come from client telemetry. Server records can recover active-user counts that the client misses. They cannot recover every feature, language, model, or line-of-code detail.[2]

That split matters during an incident. A steady active-user count beside falling agent edits does not prove that developers stopped editing with agents. It can mean that the server still saw requests while the client stopped attaching enough detail. A rising CLI bucket can also be an attribution error rather than a sudden change in tools.

Version belongs beside the metric.

Do not compare a client-derived series across a client migration without marking the change. Record the IDE version, extension version, telemetry policy, network route, and SDK boundary for each cohort. Keep deployment dates on the chart.

GitHub exposes the last known IDE and plugin versions in per-user reports. That gives operators a concrete place to look before blaming training, adoption, or developer behavior. It does not prove that every event arrived. It helps separate old clients from new ones.

Give users custody of the raw route.

The OpenTelemetry project recommends OTLP export for logs, traces, and metrics when products need to deliver telemetry to a user's own system. Its design guide separates self-hosted software, where the process exports, from cloud platforms, where the provider forwards data.[3]

That does not repair GitHub's historical data. It gives product builders a better contract. Document which process emits each signal, which identity fields travel with it, which Collector or service forwards it, and what cannot be reconstructed later.

Diagnose the witness before the workforce.

When a coding-agent chart moves, compare independent evidence before changing policy. Check server activity, pull requests, client versions, network delivery, and known release changes. Preserve the mismatch in the incident note. Do not average it away.

The practical rule is short. No adoption verdict crosses a client migration without a version split and an independent witness.

Interactive makeover / dashboard verdict replaced

Run the attribution dyno.

A normal dashboard asks you to trust one line. This dyno combines a symptom selector, witness checklist, connected comparison manifold, and copyable incident card. It drafts a diagnosis. It does not inspect your telemetry or prove a cause.

Name symptomSplit witnessesMark versionsWrite decision

Signal controls

Choose the observed pattern. Then select the evidence that the incident card must request.

Observed pattern
Required witnesses
Comparison manifold / evidence required

Incident card

Witness load0 / 4 selected
Witnesses missingThe selected symptom is a clue only. No version, server, delivery, or repository evidence is attached.
Sources read / October 9, 2026

Source log

Open the evidence boundary
  1. GitHub Changelog, "Update your IDE to restore agent activity in Copilot usage metrics", published October 6 and read October 9, 2026. GitHub describes the missing IDE attribution, affected report fields, temporary CLI inflation, client-version fixes, and the lack of backfill.
  2. GitHub Docs, "GitHub Copilot usage metrics", read October 9, 2026. The documentation separates client and server telemetry, supported IDE versions, report freshness, metric meanings, and expected differences across report scopes.
  3. OpenTelemetry, "OTel-Native by Design", published October 8 and read October 9, 2026. The article compares self-hosted and platform export, describes OTLP delivery for logs, traces, and metrics, and documents endpoint and routing choices.
  4. Hacker News discussion, "OTel-Native by Design", item 50016974, read October 9, 2026. We resolved the exact item URL through the Hacker News API. The discussion led to the OpenTelemetry article. It does not verify the GitHub incident.

Scope: the dyno changes a diagnosis template only. It does not read an organization, call a metrics API, inspect an IDE, or prove why a chart moved. Attach real report windows, client builds, route checks, outcomes, and an owner before making an operational decision.