feat: Add auto and wrapper instrumentation for @github/copilot-sdk

[Stephen Belanger (Qard)]

Summary

• Adds auto-instrumentation (via --import braintrust/hook.mjs and bundler plugins) and a wrapCopilotClient(client) manual wrapper for @github/copilot-sdk v0.3+
• Span tree: Copilot Session (TASK) → Copilot Turn (TASK) → github.copilot.llm (LLM) / tool: <name> (TOOL) / Agent: <name> (TASK for sub-agents)
• Token metrics on LLM spans use the shared Anthropic cache-token helpers: prompt_tokens, completion_tokens, prompt_cached_tokens, prompt_cache_creation_tokens, completion_reasoning_tokens, reasoning_tokens, tokens
• Copilot-specific billing fields (cost, quotaSnapshots, copilotUsage) flow into namespaced github_copilot.* metadata, not metrics
• Orchestrion configs target CopilotClient.{createSession,resumeSession} in dist/client.js and CopilotSession.sendAndWait in dist/session.js (both ESM and CJS)
• Also fixes imports.test.ts to skip node_modules and .d.ts files during the dynamic-import scan (those directories can be populated by test fixture installs)

Architecture notes

The Copilot SDK is an agent-runtime client (JSON-RPC over a CLI subprocess), not a chat-completions wrapper — so instrumentation follows the lifecycle/processor pattern used for @anthropic-ai/claude-agent-sdk rather than the per-method pattern used for Groq/OpenAI.

Span lifecycle is driven by the assistant.turn_start/turn_end event stream and assistant.usage (per-LLM-call metrics). Tool and sub-agent spans use tool.execution_start/complete and subagent.started/completed/failed events. Session close is triggered by the injected onSessionEnd hook.

Parent span IDs are resolved by awaiting span.export() (same pattern as claude-agent-sdk-plugin.ts). Async event handlers are chained through state.processing: Promise<void> to preserve ordering.

Test plan

• tsc --noEmit clean (no type errors)
• pnpm run build succeeds
• All 961 unit tests pass
• pnpm run lint — warnings only (same pattern as rest of codebase)
• pnpm run formatting clean
• Unit tests cover: extractMetricsFromUsage (all token/cache/reasoning/billing fields), plugin lifecycle, orchestrion config shape, BraintrustPlugin wiring
• E2e scenario at e2e/scenarios/github-copilot-instrumentation/ requires a GitHub Copilot subscription or a BYOK provider endpoint (BRAINTRUST_E2E_MODEL_BASE_URL)

🤖 Generated with Claude Code