Dive into my thoughts on coding, tech trends, and developer life. Explore my latest posts below.
55 DeepInfra models had tool_choice support but were missing the function_calling flag, causing upstream frameworks to silently skip tool use. A data-layer fix with regression coverage.
PydanticAI's custom exceptions lost attributes during pickle round-trips, breaking error propagation in Celery, Ray, and multiprocessing. Fixed with __reduce__ methods and 16 parameterized tests.
Gradio's BrowserState component converted Pydantic models to Python repr strings instead of JSON dicts, breaking the frontend data contract. Fixed with model_dump() and 8 regression tests.
Gradio's App.create_app() hardcoded debug=True, ignoring the user's launch(debug=False) setting. A clean one-line fix with tests.
How a one-line serialization regression broke Notion sync for all Dify v1.13.0 self-hosted users, and how I traced it through a masking test fixture to deliver a clean fix with regression coverage.
Added runner-level regression tests to verify MCP tool failures produce proper tool_call_output_item events in both sync and streaming paths.