llm-cost-observability

Use when adding a new LLM call to a service, or wiring an existing service to the company LiteLLM proxy so its calls get logged to ClickHouse. Triggers when the user mentions calling Claude/GPT/an LLM, "instrument" or "add observability" to LLM calls, switching from a direct provider SDK to the proxy, or asks how to send tags. Don't use this for general LLM-coding questions unrelated to our proxy.

Use when the user asks a cost, usage, latency, or attribution question about LLM calls — e.g. "what did feature X cost last month", "spend by model", "p95 latency for our summarization endpoint", "which prompt version is most expensive", "how much did customer Y spend on Claude". Also use for debugging a specific past LLM call by request ID, or for offline evals against historical traffic. Queries the `litellm_logs` table in ClickHouse Cloud via the ClickHouse MCP.