README.md

Flow Fundamentals (basic-flows)

Python port of js/testapps/flow-sample1. No model is used; these flows exercise the framework itself — traced steps, streaming, context propagation, error handling (caught and uncaught), and a long-running flow you can stare at in Dev UI to confirm spans appear live.

Flows

Flow	What it shows
basic	Two ai.run() traced steps
parent	One flow calling another
withInputSchema	Typed object input via Pydantic
withContext	Reading the request context inside a flow
streamy	Streaming count chunks at 1s intervals
streamyThrowy	Stream a few chunks, then raise mid-stream
throwy	Run a step, then raise from the flow body
throwy2	Raise from inside a traced step
flowMultiStepCaughtError	Catch an error from a middle step and keep going
multiSteps	Several traced steps with reused span names
largeSteps	~1MB string outputs per step (stress the trace pipe)
test-long-broadcast	Multi-minute flow with nested spans (broadcast test)

Run once

uv sync
uv run src/main.py

This runs basic, parent, withInputSchema, multiSteps, and flowMultiStepCaughtError and prints their results. The streaming, throwing, and long-broadcast flows are skipped here — pick them from Dev UI.

Run in Dev UI

genkit start -- uv run src/main.py

Open http://localhost:4000 and pick a flow from the sidebar.

Suggested manual checks

1. basic — "hello" → returns foo: subject: hello. Two spans show.
2. streamy — 5 with streaming on → five {count: N} chunks at 1s intervals, then done: 5, streamed: 5 times.
3. streamyThrowy — 5 with streaming on → three chunks, then a RuntimeError: whoops that surfaces in the trace.
4. throwy / throwy2 — "hello" → flow errors out; the failing span is highlighted in Dev UI.
5. multiSteps — "world" → returns 42; check that the reused step1 name appears twice in the trace.
6. test-long-broadcast — {"steps": 5, "step_delay_ms": 5000} → ~25s flow with nested fetch/process/save spans you can watch arrive live.