agentHost: rotate stale OTel session anchors

[Austin Traver (austintraver)]

Summary

• replay each synthetic Agent Host session anchor immutably while its generation is current
• rotate the anchor span ID and timestamp after 12 hours while retaining the native trace ID
• rematerialize Claude's resumable subprocess when its propagated traceparent or tracestate changes
• document the downstream trace-segmentation trade-off and cover replay, rotation, and Claude rebind behavior with tests

Problem

Agent Host currently creates a zero-duration vscode.agent_host.session parent and can reference that same span for the entire lifetime of a persisted session. Native-OTLP consumers may retain their trace/span-to-run parent-resolution mappings only temporarily. After that state expires, a consumer can acknowledge a late child with HTTP 200 but never materialize it.

I reproduced this with valid late spans still present in Agent Host's local SQLite store while the downstream LangSmith project omitted them. In a controlled comparison, a fresh root materialized but a fresh child referencing an old session anchor did not, despite successful ingestion responses for both.

Re-exporting the identical anchor restores parent-resolution state for shorter idle periods and preserves immutable span identity. It cannot work indefinitely, though: sufficiently old spans can fall outside a downstream ingestion window. Updating the timestamp on the same span ID would mutate history and can produce invalid parent/child chronology.

This change therefore bounds each anchor generation to 12 hours. At the boundary it emits a fresh zero-duration parent with a new span ID and current timestamp, but keeps the session's native trace ID. The interval matches LangSmith's documented self-hosted parent-mapping default and remains inside its observed 24-hour ingestion window; this does not assume LangSmith Cloud uses the same TTL.

Copilot and Codex already resolve the current session context at turn boundaries. Claude receives W3C context through its long-lived subprocess environment, so it uses the existing resumable pipeline rebind path when the context changes. The SDK conversation is resumed rather than replaced.

If a backend's native trace mapping has already expired, it may display a rotated anchor as the beginning of a new backend trace even though the native trace ID is stable. That segmentation is preferable to silently losing every subsequent span.

This is separate from github/copilot-cli#4726, where a post-reload resumed root can omit captured input content.

Validation

• NODE_OPTIONS=--max-old-space-size=8192 ./node_modules/.bin/tsc6 --project ./src/tsconfig.json --noEmit --skipLibCheck
• NODE_OPTIONS=--max-old-space-size=8192 npm run gulp compile-client (compile-src completed with 0 errors; unrelated generated monaco.d.ts drift from current main was excluded from this commit)
• node test/unit/node/index.js --run src/vs/platform/agentHost/test/node/otel/agentHostOTelService.integrationTest.ts --run src/vs/platform/agentHost/test/node/claudeAgent.test.ts (307 passing)
• git diff --check origin/main...HEAD

[vs-code-engineering]

📬 CODENOTIFY

The following users are being notified based on files changed in this PR:

TylerLeonhardt

Matched files:

• src/vs/platform/agentHost/node/claude/claudeAgentSession.ts

[Copilot]

🟡 Changes recommended

Peer chats use inconsistent anchor keys, and anchor delivery is not ordered before provider spans.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Updates Agent Host OTel session anchors to remain resolvable for long-lived sessions.

Changes:

• Replays anchors and rotates them after 12 hours while preserving trace IDs.
• Rebuilds Claude subprocesses when propagated trace context changes.
• Adds documentation and tests for replay, rotation, and rebinding.

File summaries

File	Description
src/vs/platform/agentHost/OTEL.md	Documents anchor rotation behavior.
src/vs/platform/agentHost/common/otel/agentHostOTelService.ts	Updates the trace-context contract.
src/vs/platform/agentHost/node/otel/agentHostOTelService.ts	Implements anchor replay and rotation.
src/vs/platform/agentHost/node/claude/claudeAgentSession.ts	Rebinds Claude when context changes.
src/vs/platform/agentHost/test/node/otel/agentHostOTelService.integrationTest.ts	Tests anchor replay and rotation.
src/vs/platform/agentHost/test/node/claudeAgent.test.ts	Tests Claude trace-context rebinding.

Review details

• Files reviewed: 6/6 changed files
• Comments generated: 2
• Review effort level: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.