Updated just now

Share this article

Related News

VS Code 1.141 adds agent-session grids and guarded worktree cleanup

Oct 12, 2026

VS Code 1.141 adds agent-session grids and guarded worktree cleanup

VS Code 1.141 can display agent sessions side by side and reclaim inactive worktree storage. Cleanup protects active or unpreserved work and keeps session history, so teams should distinguish session state from disk deletion.

Visual Studio CodeGitHub CopilotAI agents

GitHub Copilot CLI can discover Ollama models without making the session offline

Oct 12, 2026

GitHub Copilot CLI can discover Ollama models without making the session offline

Copilot CLI 1.0.94-0 adds local Ollama discovery to the model picker. It does not install models, guarantee tool support or enable offline mode, so developers still need to verify the runtime, endpoint and network path.

GitHubGitHub Copilot CLIOllama

GitHub Copilot local sandboxing is GA, but network access stays on by default

Oct 12, 2026

GitHub Copilot local sandboxing is GA, but network access stays on by default

Copilot local sandboxing now covers CLI, the Copilot app and VS Code Agent Host sessions. It restricts tools through operating-system controls, while outbound internet, local-network access and credential proxies still need explicit review.

GitHubGitHub Copilotsandboxing

GitHub Copilot telemetry can name users without capturing prompts — Source: OpenTools Team. Source-based editorial diagram: Copilot telemetry identity and content control map.Source image

Enterprise AI

GitHub Copilot telemetry can name users without capturing prompts

By OpenTools Team

GitHub now separates user attribution from five content‑capture switches for Copilot agent telemetry. All six controls default off, but managed‑setting precedence and client support determine what an enterprise actually records.

Identity and content are now separate telemetry decisions

GitHub's October 12 announcement adds a telemetry managed setting for Copilot agent clients. An administrator can enable identity attribution without recording prompts or responses, or enable selected content classes without attaching the three human‑readable identity fields.

The identity switch, capture.identity, can add the GitHub login as user.name, the local operating‑system account as process.user.name, and the machine name as host.name. It is off by default and separate from the pseudonymous identifier already used by telemetry. That distinction matters for incident review: a stable pseudonymous identifier can group activity, while the new fields can connect it to a person and device.

The first deployment choice is therefore purpose, not transport. Decide whether the question is capacity, reliability, individual attribution or content‑level debugging, then enable only the fields needed for that question.

Five content switches expose different kinds of sensitive material

GitHub splits content capture into prompts, responses, toolArguments, toolOutput and policyDetail. All five default to false. The managed‑settings reference warns that captured data can include source code, file contents, terminal commands, tool results, workspace paths and user identifiers.

The categories are not interchangeable. Prompts and responses reveal the conversation; tool arguments can expose filenames, URLs and command inputs; tool output can contain secrets or production data; policy detail can show why an action was allowed or blocked. A broad troubleshooting toggle therefore has a larger privacy and retention surface than identity‑only attribution.

Use a field‑level approval record with an owner, purpose, destination, retention period and stop date. Send telemetry only to an approved HTTPS endpoint and test the backend's access controls before enabling content in a managed fleet.

Managed precedence can replace the entire telemetry object

GitHub says the highest‑precedence managed source supplies the whole telemetry object; lower‑precedence sources are not merged into it. An enterprise setting can therefore replace organization or local values rather than adding one field. A reviewer who checks only the local file can misread the effective configuration.

The legacy captureContent shorthand still enables all five content classes, but it never enables identity. That makes it easy to create the opposite of an intended identity‑only setup if an old template is copied forward. Prefer the fine‑grained object and remove the shorthand after verifying equivalent behavior.

For acceptance, record the source that won precedence and export the resolved settings from a representative client. Change one field at a time, generate a harmless test span, and verify both the fields present and the fields absent at the collector.

Client support is uneven, so policy text is not delivery proof

The announcement covers Copilot CLI, the Copilot app and Visual Studio Code Agent Host sessions. GitHub's reference also notes that support for the fine‑grained controls varies by client. In particular, local VS Code agent‑harness sessions support identity and the legacy content switch while the Agent Host supports the separate content classes.

The VS Code monitoring guide describes the OpenTelemetry path and the kinds of events an operator can observe. Use it to verify an actual client/version pair rather than assuming every Copilot surface implements the same schema.

A small matrix should list client, version, effective setting source, protocol, collector, identity result and each content class. Mark unsupported fields explicitly. An empty field can mean disabled, unsupported or no qualifying event, and those states should not be combined.

A privacy‑first rollout proves absence as carefully as presence

Start with identity and all content capture off, verify ordinary operational spans, then enable the smallest required field for a test cohort. Inspect raw collector output for a synthetic repository that contains recognizable non‑secret markers. Confirm that prohibited markers do not appear before expanding the cohort.

Repeat after configuration precedence, client version or collector processing changes. Downstream processors can enrich or retain data even when the client configuration looks narrow, so the acceptance record should include the final stored event and deletion test.

Identity‑aware telemetry can improve accountability, but it also changes who can be observed. Publish the purpose and access policy to affected developers, limit access to reviewers with a defined job, and disable fields when the diagnostic window ends. Follow GitHub's OpenTools organization page for related agent‑platform changes.

Figure: The separate control paths for Copilot telemetry identity and content capture. Sources: GitHub's announcement, the managed‑settings reference and the VS Code monitoring guide. Figure by OpenTools Team; no third‑party expressive material used.

Sources

1.GitHub's announcement (github.blog)
2.VS Code monitoring guide (code.visualstudio.com)
3.docs.github.com (docs.github.com)

Tags

GitHub CopilotOpenTelemetryAI agentsenterprise governancedeveloper tools