ReAct Planner

Complex multi-step queries are automatically routed through a Reason + Act loop (src/lib/react-planner.ts), replacing the flat single-pass function-call pattern with a deliberate planning cycle.

The loop

1. THINK — the Conductor analyzes the request and generates a multi-step plan
2. ACT — each step dispatches to Virtuosos via function calls
3. OBSERVE — the result of each action is recorded
4. REFLECT — the plan adapts based on intermediate results
5. Repeat until the plan completes or max iterations are reached

Plan steps

export interface PlanStep {
  stepNumber: number               // 1-indexed
  thought: string                  // what the Conductor plans to do
  action: { name: string; args: any } | null  // null = pure reasoning step
  observation: string | null      // result observed from the action
  status: 'planned' | 'executing' | 'completed' | 'failed' | 'skipped'
}

Every step is broadcast over the Symphony Bus, so the Orchestra Visualizer shows the plan executing live, and startTimer() telemetry captures per-step latency.

Conductor function declarations

The planner acts through the same function surface as single-pass commands (src/lib/conductor-functions.ts):

Every call is validated against Zod schemas (conductor-schemas.ts) before execution; validation failures are logged via logValidationFailure and surfaced back to the planner as observations so it can self-correct.

References: Yao et al. (2023), ReAct: Synergizing Reasoning and Acting in LLMs; Wei et al. (2022), Chain-of-Thought Prompting.