Recipe Pattern

The core innovation of HART OS: learn a task once, then replay it without repeated LLM calls.

Two Modes

CREATE Mode

1. User input is decomposed into hierarchical flows and actions.
2. Each action is executed against the appropriate tool or LLM.
3. The full execution trace is saved as a recipe for future reuse.

REUSE Mode

1. A matching recipe is loaded from disk.
2. Each step is replayed deterministically -- no LLM calls required.
3. Achieves approximately 90% faster execution compared to CREATE mode.

Hierarchical Task Decomposition

User Prompt
+-- Flow 1 (Persona A)
|   +-- Action 1
|   +-- Action 2
|   +-- Action 3
+-- Flow 2 (Persona B)
    +-- Action 1
    +-- Action 2

ActionState Machine

ASSIGNED --> IN_PROGRESS --> STATUS_VERIFICATION_REQUESTED --> COMPLETED
                                                          \--> ERROR --> TERMINATED

States auto-sync to the SmartLedger for persistence across sessions.

Recipe Storage

Path Pattern	Contents
prompts/{prompt_id}.json	Prompt definition
prompts/{prompt_id}_{flow_id}_recipe.json	Trained recipe for a flow
prompts/{prompt_id}_{flow_id}_{action_id}.json	Individual action recipe

Autonomous Fallback

When an action enters the ERROR state, the StatusVerifier LLM auto-generates a context-aware fallback strategy. No user prompts are required for fallback, enabling fully autonomous agents.

Source Files

• create_recipe.py -- Agent creation, action execution, recipe generation.
• reuse_recipe.py -- Recipe loading and trained agent execution.
• helper.py -- Action class, JSON utilities, tool handlers.
• lifecycle_hooks.py -- ActionState machine, ledger sync.