LGTM Skill - Query Observability Backends

Why subagents matter here

lgtm commands return raw JSON — sometimes thousands of lines. If you run queries directly in the main conversation, you'll flood the context window and make it harder to reason about what actually matters. Haiku subagents are the right tool: they run the queries, distill the results, and hand you back just the signal you need.

The pattern is: you orchestrate, haiku executes.

Orchestrator Pattern

• You (orchestrator): Coordinate the discovery → investigation flow. Evaluate summaries returned by subagents, decide what to query next, synthesize findings for the user. Don't run lgtm commands yourself.
• Haiku subagent: All query execution — discovery, investigation, aggregation, analysis. Fast and sufficient for the vast majority of tasks.

Run independent queries in parallel — spawn multiple Task calls in one message when queries don't depend on each other (e.g., check logs AND metrics AND traces simultaneously).

Two-Phase Approach

Phase 1: Discovery