Message Formatting

The platform controls how messages are formatted before being sent to the LLM. This includes assembling the system prompt, injecting the yields schema, and constructing the conversation history — all automatically based on your YAML config.

System Message Construction

The system message is automatically constructed from:

Agent prompt: The prompt field from YAML
Yields schema: Formatted as output instructions so the LLM knows what structure to produce

agents:
  solver:
    name: "solver"
    model: "gpt-4o-mini"     # required — the SDK never picks one for you
    prompt: |
      You solve math problems step by step.
    yields:
      answer: "number"
      steps: "array"

The LLM receives a system message that combines the prompt with instructions to produce output matching the yields schema.

History Formatting

Conversation history entries are formatted as alternating user/assistant messages:

System: You solve math problems step by step. [yields schema]
User: What is 2 + 2?
Assistant: {"answer": 4, "steps": ["2 + 2 = 4"]}
User: Now multiply that by 3

Recent messages from the session are included automatically. This provides multi-turn context without exceeding token limits.

Input Data Formatting

The input_data dictionary from mesh_call() is formatted as the current user message. The platform serializes the dictionary into a format the LLM can process.

Pre-Compose Influence

Pre-compose processors run before the message is formatted, modifying what gets sent to the LLM:

async def clean_input(input_data, context):
    """Remove sensitive fields before LLM sees them"""
    cleaned = {k: v for k, v in input_data.items() if k != "api_key"}
    return cleaned

Next Steps

Message System — Full message flow
LLM Agents — Agent execution pipeline

LeafMesh — Controlled message construction