claude-code - 💡(How to fix) Fix OTEL: hook_execution_start telemetry events dropped before event logger initialized

Summary

With OTEL telemetry enabled (CLAUDE_CODE_ENABLE_TELEMETRY=1, OTEL_LOGS_EXPORTER=otlp), the extension fires hook_execution_start events ~37 ms BEFORE getOtlpLogExporters() finishes and the event logger is registered. The harness logs 3 [WARN] entries per session start:

[WARN] [3P telemetry] Event dropped (no event logger initialized): hook_execution_start
[WARN] [3P telemetry] Event dropped (no event logger initialized): hook_execution_start
[WARN] [3P telemetry] Event dropped (no event logger initialized): hook_execution_start

The dropped events are real hook lifecycle data (which hook fired, with what input, when) — losing them blinds session-start observability dashboards.

Environment

• VS Code Insiders, extension Anthropic.claude-code v2.1.139
• OTEL config:
  • CLAUDE_CODE_ENABLE_TELEMETRY=1
  • OTEL_LOGS_EXPORTER=otlp
  • OTEL_METRICS_EXPORTER=otlp
  • OTEL_TRACES_EXPORTER=otlp
  • OTEL_EXPORTER_OTLP_PROTOCOL=http/json
  • OTEL_EXPORTER_OTLP_ENDPOINT=https://<our-collector>/api/otel
  • Bearer token via OTEL_EXPORTER_OTLP_HEADERS
• Hooks: 24 registered across SessionStart, InstructionsLoaded, PreToolUse, PostToolUse, ConfigChange, SessionEnd, PostCompactLog. Several async: true.
• Observed 3 dropped events per session, consistently, across 5 sessions on 2026-05-12.

Evidence (from extension log timestamps)

2026-05-12T16:45:54.656Z [DEBUG] [Telemetry] enabled
2026-05-12T16:45:54.693Z [WARN] [3P telemetry] Event dropped (no event logger initialized): hook_execution_start
2026-05-12T16:45:54.710Z [DEBUG] [Telemetry] getOtlpLogExporters() called
2026-05-12T16:45:54.732Z [DEBUG] [Telemetry] Event logger registered

hook_execution_start at 54.693 ms fires in a 76 ms gap between "telemetry enabled" (54.656) and "logger registered" (54.732).

Reproducer

1. Configure OTLP logs exporter with any reachable collector (or even an unreachable one — the drop happens client-side before the export attempt).
2. Configure any hook bound to SessionStart or InstructionsLoaded in hooks.json.
3. Open a new session; check the extension log:

   grep "Event dropped (no event logger initialized)" \
     ~/.vscode-server*/data/logs/<latest>/exthost*/Anthropic.claude-code/Claude\ VSCode.log

4. Expect 0 dropped events; observe 3.

Suggested fix

Move getOtlpLogExporters() + event-logger setup BEFORE the first PostHookExecutionEvent fires. Alternatively, buffer events emitted in the telemetry-enabled → logger-ready window and replay them when the logger registers. The buffer would only need to hold a handful of events (the gap is ~76 ms in practice) so memory cost is negligible.

If the gap is fundamentally hard to close (e.g. the exporter factory does I/O), a smaller stop-gap is to emit the drop count as a counter metric so consumers know how many events were lost rather than silently discarding them.

Why this matters for us

We're building a session-start latency dashboard that needs hook_execution_start to bin hook invocations against the SessionStart timeline. Losing the first 3 means we under-count hook frequency and miss the very hooks that run earliest (which are often the slowest, e.g. memory hydration, sync staleness check).

Happy to test any candidate fix.