Chat: OTel enrichment for Claude agent sessions

[zhichli]

Overview

Follow-up to #318031, which introduced the canonical github.copilot.* OTel attribute namespace on the foreground Copilot Chat agent. This PR extends the same enrichment to Claude Code agent sessions so traces carry parity attributes across agent surfaces.

What

• invoke_agent claude (claudeOTelTracker) — now emits:
  • github.copilot.agent.type: 'builtin'
  • github.copilot.git.{repository,branch,commit_sha} and github.copilot.github.org (via workspaceMetadataToOTelAttributes(resolveWorkspaceOTelMetadata(gitService)), mirroring toolCallingLoop).
• execute_tool for Claude tool calls (claudeMessageDispatch.handleAssistantMessage) — now uses an injected tool-parameter extractor (node-layer provided) to emit:
  • Ungated: github.copilot.tool.parameters.edit_type, skill_name, mcp_server_name_hash, mcp_tool_name.
  • Content-gated (captureContent): command, file_path, mcp_server_name.
• execute_hook (claudeMessageDispatch.handleHookResponse) — now emits:
  • github.copilot.hook.decision — pass (outcome success) / block (exit code 2) / non_blocking_error (other failure).
  • github.copilot.hook.duration — seconds, measured between hookStarted and hookResponse via a new hookStartTimes map on MessageHandlerState.
• extractToolParameters / genAiAttributes — added Claude tool-name aliases so the classifier recognises Bash, Read, Edit, MultiEdit, Write, NotebookEdit.
• Truncation fix — gen_ai.tool.call.arguments now truncates with otelService.config.maxAttributeSizeChars to avoid oversized OTel attributes being dropped.

Note: github.copilot.hook.tool_names is intentionally not emitted for Claude — the Claude SDK's SDKHookStartedMessage does not surface a tool_name to the extension. The user-facing doc already qualifies this attribute as "When tool-scoped".

Why

After #318031, dashboards built on the canonical github.copilot.* namespace were missing rows for the Claude agent path — its invoke_agent / execute_tool / execute_hook spans were only partially queryable. This brings Claude to parity with the built-in agent's enrichment while preserving extension layering constraints.

Backwards compatibility

All existing attributes (gen_ai.*, copilot_chat.*) remain unchanged. Only additive attribute writes; no behavioural change to dispatch or hook handling.

Docs

Updated extensions/copilot/docs/monitoring/:

• agent_monitoring.md — Claude invoke_agent claude attribute table now lists the new github.copilot.{agent.type,git.*,github.org} keys; added a proper attribute table for Claude execute_tool listing tool.parameters.*.
• agent_monitoring_arch.md — corrected the github.copilot.* namespace description to reflect that it's the canonical Copilot namespace (extension-emitted enrichment + CLI SDK internal metrics), not just SDK-internal.

Tests

• extractToolParameters.spec.ts — adds Claude-name classification (Bash, Read, Edit, MultiEdit, Write, NotebookEdit).
• claudeMessageDispatch.spec.ts — adds hook.decision / hook.duration cases and tool.parameters.* emission.
• claudeCodeAgentOTel.spec.ts — exercised via existing end-to-end coverage.
• All 89 tests pass across the three affected spec files; typecheck clean.

[Copilot]

Pull request overview

Extends the Copilot extension’s canonical github.copilot.* OpenTelemetry enrichment to the Claude agent session flow so Claude invoke_agent / execute_tool / execute_hook spans become queryable on the same dimensions as built-in Copilot Chat spans.

Changes:

• Enriches Claude invoke_agent spans with github.copilot.agent.type and workspace git/GitHub org attributes via workspace metadata resolution.
• Emits structured github.copilot.tool.parameters.* attributes for Claude tool calls using extractToolParameters, with content-sensitive fields gated behind captureContent.
• Records hook completion attributes (github.copilot.hook.decision, github.copilot.hook.duration) and extends tool-name classification to Claude’s capitalized tool names.

Show a summary per file

File	Description
extensions/copilot/src/platform/otel/node/test/extractToolParameters.spec.ts	Adds coverage for Claude-capitalized tool-name classification.
extensions/copilot/src/platform/otel/node/extractToolParameters.ts	Extends edit-type classification to Claude tool names.
extensions/copilot/src/platform/otel/common/genAiAttributes.ts	Adds Claude tool names to shell/file tool-name sets used by OTel extraction.
extensions/copilot/src/extension/chatSessions/claude/node/claudeOTelTracker.ts	Adds canonical Copilot git/org/agent-type attributes to Claude root spans.
extensions/copilot/src/extension/chatSessions/claude/node/claudeCodeAgent.ts	Wires additional per-request timing state for hook duration tracking.
extensions/copilot/src/extension/chatSessions/claude/common/test/claudeMessageDispatch.spec.ts	Adds tests for tool parameter emission and hook decision/duration attributes.
extensions/copilot/src/extension/chatSessions/claude/common/claudeMessageDispatch.ts	Emits tool parameter attributes and hook decision/duration attributes on spans.

Copilot's findings

• Files reviewed: 9/9 changed files
• Comments generated: 1

[zhichli]

Pushed 55224ab to fix CI:

• Layering violation (root cause of the 3 CI failures): common/claudeMessageDispatch.ts was importing extractToolParameters from platform/otel/node/, which violates the common→node layering rule. Fixed by injecting the function through MessageHandlerState from the node-layer caller (claudeCodeAgent.ts); the test file uses a small local stub. Verified with eslint on the affected directories.
• Inline review feedback: addressed @copilot-pull-request-reviewer's comment — truncateForOTel(...) for gen_ai.tool.call.arguments now passes maxAttributeSizeChars.

89 tests pass; typecheck clean (the unrelated AgentTask / @vscode/copilot-api errors exist on main and are not from this PR).