Typed Outputs

PyFlue validates structured results with Pydantic v2.

from pydantic import BaseModel


class TriageResult(BaseModel):
    fix_applied: bool
    summary: str

Pass the model to prompt or skill:

result = await session.skill(
    "triage",
    args={"issue_number": 123},
    result=TriageResult,
)

Result Delimiters

When a result model is provided, PyFlue asks the backend to return JSON between these delimiters:

---RESULT_START---
{"fix_applied": true, "summary": "Fixed one typo."}
---RESULT_END---

PyFlue extracts the JSON and validates it with pydantic.TypeAdapter.

Validation Errors

If the model returns invalid data, PyFlue asks the backend to repair the response. The default retry count is 3 and can be configured with typed_retries in pyflue.toml.

[agent]
typed_retries = 3

You can override retries for one call:

result = await session.prompt(
    "Return a valid triage result",
    result=TriageResult,
    retries=1,
)