Idempotency key normalization

When implementing caching/idempotency for tools that generate external media, make cache keys stable and semantically accurate.

Rules

1. Normalize before hashing: Resolve optional fields through input_schema defaults; treat an explicit None as “unset”, so omitted and defaulted inputs produce the same key.
2. Hash only what changes output: idempotency_key_fields should include only fields that are actually consumed by the request/output builders. Do not include fields the provider ignores, or you’ll get false cache misses.
3. Test key behavior: Add tests for both:
   • Stability: inputs that differ only by omitted/defaulted values produce identical keys.
   • Sensitivity: inputs that truly change the generated result produce different keys.

Example pattern

class MyTool(BaseTool):
    input_schema = {
        "type": "object",
        "required": ["prompt"],
        "properties": {
            "prompt": {"type": "string"},
            "prompt_influence": {"type": "number", "default": 0.3},
            "loop": {"type": "boolean", "default": False},
        },
    }

    # Include only fields that affect the request/output
    idempotency_key_fields = ["prompt", "prompt_influence", "loop"]

This ensures cache/idempotency behaves predictably: no accidental misses from default handling, and no accidental hits from hashing irrelevant fields.