Schemas

Structured outputs keep planner interactions predictable. Schema files live in src/schemas/ and are passed directly to llama.cpp during inference. The Bash layer still normalizes and validates outputs to guard against malformed responses.

Available schemas

• planner_plan.schema.json: base template for the tool-plan array. At runtime planner_build_plan_schema injects per-tool argument schemas and emits tuple variants that require final_answer as the terminal step. Non-terminal steps can only use non-final_answer tools, and the maximum plan length follows PLANNER_MAX_PLAN_STEPS.
• intent.schema.json: intent classification payload (intents, rationale) used to filter tools and decide whether to run pre-planner web search.
• pre_planner_search_terms.schema.json: array of one to three concise search terms (5–80 characters each) used for pre-planning web lookups.
• executor_action.schema.json: template for dynamically generated per-tool schemas used during the executor loop. Tool names and argument shapes are injected at runtime before calls to llama.cpp.
• final_answer_evaluation.schema.json: evaluator output with evaluation_type (FINAL/REPLAN), reasoning, and output fields emitted by the final-answer evaluation helper.

Free-form text arguments always appear under args.input in planner payloads, keeping prompt templates and registry-driven schemas aligned on the same canonical field name.