INSPECT-ONLY — returns the plan WITHOUT running it. For a real multi-step DC Hub question call `execute_plan(intent="...")` instead: it uses the SAME deterministic no-LLM planner and then RUNS the sequence server-side, returning the answers in one envelope. Reach for plan_query only to review, log, diff or audit a plan before executing it yourself. Deterministic keyword routing over the tool registry — no LLM, no network, same intent always returns the same plan (free). Returns _entity=query_plan {best_tool, intent_confidence + workflow_confidence (dual 0-1: question-read vs executability), reason, planner_rationale, recommended_sequence:[{step, tool, depends_on, estimated_calls, why, args_hint}], execution_waves (steps grouped into concurrency waves), execution_strategy.parallel_groups, execution_estimate {estimated_calls, estimated_latency_ms, parallelizable}, alternatives (each with when + rejected_because), coverage_notes, matched_classes} plus a versioned `replay` (schema_version 1): planner_version, decisions:[{id, step, kind, status, decision, rationale, decision_confidence, depends_on}], rejected:[{id, tool, reason}], execution_graph:{waves, parallel_groups} — auditable and machine-readable, safe to log and diff across versions. args_hint values in <angle brackets> come from the named earlier step — substitute them, never invent them. Pass structured hints via context (lat/lon, iso, market, capacity_mw, candidate_id, state, since) to sharpen the plan. For a family-level browse use discover_tools. This tool plans — it never executes; tools/list stays canonical for schemas.
