fix: pass mode='json' to model_dump() in PyHookRegistry emit path

[Diego Colombo (colombod)]

Summary

• Symptom: Orchestrator crashes with TypeError: Object of type Decimal is not JSON serializable on the first LLM response, in PyHookRegistry::emit(). The crash fires before any Python hook handler gets control, making it impossible to intercept from the consumer side.
• Root cause in the hook layer: try_model_dump() in helpers.rs calls bare model_dump()", which returns Python-native types including Decimal(used byUsage.cost_usd). The result is passed directly to json.dumps()with no custom encoder — which cannot handleDecimal`.
• Why the fix belongs here: The hook emitter is the caller of json.dumps(). It is the correct place to request JSON-safe output. Putting mode="json" here means no consumer or model field ever needs to compensate for the hook system's serialization contract.

What changed

One function, try_model_dump() in bindings/python/src/helpers.rs:

// Before
obj.call_method0("model_dump")

// After
let kwargs = PyDict::new(py);
kwargs.set_item("mode", "json")?;
obj.call_method("model_dump", (), Some(&kwargs))

mode="json" tells Pydantic to emit only JSON-native Python types (str, int, float, list, dict, bool, None) regardless of the field's underlying Python type or whether a @field_serializer is present on the model.

try_model_dump() is a shared helper called by both emit() and emit_and_collect() in hooks.rs — this one-function change covers the entire emit path.

Why this is architecturally correct (not a workaround)

The immediate trigger is amplifier-core 1.5.0 in the Docker worker image (amplifier-cache:python) predating the @field_serializer on Usage.cost_usd added in 1.5.1. But this fix is correct regardless of model version:

• The hook system should not depend on individual model fields having serializers to be JSON-safe.
• Any future field of any non-JSON-native type (e.g. UUID, datetime, custom type) in any hook payload would cause the same crash without this fix.
• This is defence-in-depth at the right layer: the serializer requests what it needs rather than relying on every upstream model to anticipate its requirements.

Test plan

• Build amplifier-cache:python image after this merges and bump/release amplifier-core
• Run orchestrator resolver end-to-end — confirm no TypeError: Object of type Decimal is not JSON serializable on LLM response
• Verify emit() and emit_and_collect() both work correctly (both flow through try_model_dump())
• Confirm fallback path (non-Pydantic objects) still passes the raw object through unchanged

Rebuilding amplifier-cache:python after this merges will fully resolve the crash for all consumers without any per-consumer workarounds.

Generated with Amplifier