llm-observabilitylisted

**CONSUMES**: User observability/LLMOps task + production LLM-or-agent system description + optional existing tracing/cost/registry configs **PRODUCES**: Applied observability judgment rules + tracing platform recommendation + cost-attribution & budget-governance design + OTel GenAI semconv audit + latency/TTFT profiling plan + prompt-registry versioning policy + online-eval & drift-detection pipeline # LLM Observability & LLMOps Capability Pack **Version**: 0.1.0 **Compatibility**: Claude Code (Phase 1); Codex / Cursor / Gemini in Phase 3 **License**: Apache 2.0 --- ## What This Pack Does AI agents wire up LLM observability by copying a tracing SDK's quickstart. They log end-to-end latency and call it monitoring — missing TTFT, the metric that actually governs perceived responsiveness in streaming. They store a single pre-computed dollar cost per call, which silently goes wrong the moment a provider changes list prices. They aggregate spend globally by API key, so no one can say which tenant drove the bill. They hardcode prompts in app code, coupling every typo fix to a redeploy. They run a Kolmogorov–Smirnov test on raw embeddings to "detect drift" and find nothing, because univariate tests are blind to high-dimensional shift. This pack embeds the judgment rules that LLMOps and reliability engineers apply automatically — rules from production tracing platforms, cost-attribution gateways, the OpenTelemetry GenAI semantic conventions, vLLM inference monitoring, prompt-r