Skip to main content
Run a native Prefect flow that retrieves an existing Magic Hour image project. It makes one GET request, checks that the response belongs to the saved ID, and returns only its ID and status. It does not create a generation, call a language model, download media, or require a Prefect Cloud account. Download recover-image.py and requirements.txt into the same private working directory.

​
Install

Use Python 3.12 and a dedicated virtual environment. The example pins Prefect 3.9.0 and HTTPX 0.28.1. Install the normal declared dependencies:
Use the full prefect package for the local server. prefect-client alone does not include that server. No model provider credentials or optional integrations are needed.

​
Start a private local Prefect server

Choose a private Prefect state directory and keep it separate from existing Prefect profiles or deployments. The directory contains local orchestration metadata, including the saved project ID. Use the same directory in both terminals. Set these supported options in the server terminal:
Keep the server bound to loopback. Use an available port and update the API URL if 4200 is already in use. Do not expose this local server publicly. The server process does not need the Magic Hour key. Starting it separately also avoids passing that key to Prefect’s automatic ephemeral server subprocess, which normally inherits its parent’s environment.

​
Run the recovery flow

In a separate terminal, activate the same virtual environment and set the same PREFECT_HOME, PREFECT_API_URL, telemetry, and logging options. Remove any inherited PREFECT_API_KEY when using this local server. Provide MAGIC_HOUR_API_KEY through your normal private environment configuration in this flow terminal only. Keep it out of scripts, command arguments, Prefect parameters, screenshots, and logs. Select an image project owned by that account:
The task reads the key inside its function. The key is not a flow or task parameter, so it is not part of the saved parameter metadata. The production example uses native @flow and @task decorators with retries=0, persist_result=False, and log_prints=False. Result persistence is disabled for both; Prefect does not use a task result cache without persistence. API log delivery is also disabled. Only safe project ID/status metadata leaves the task, and errors use a static message without a raw API response or download URL.

​
Interpret the result

The final output contains only id and status. A successful GET confirms retrieval, not necessarily a completed generation. If the provider reports a pending status, check the same ID again later. Do not submit a replacement generation to recover existing work. The example does not configure schedules, deployments, retries, or automatic polling. Stop your own local Prefect server when finished. For disposable runs, remove only the private state directory you created after stopping the server; preserve existing Prefect state.

​
Verification scope

The exact example ran with Prefect 3.9.0 and HTTPX 0.28.1 against a stock local Prefect server. Its flow and task both reached COMPLETED after one authenticated GET of an existing Magic Hour project. The result contained only the requested ID and complete status. Both native run records had retries set to zero and no persisted state data. The server environment was verified free of the Magic Hour key before and after the flow. The task received the key through its own process environment; it was not a flow or task parameter. Normal Prefect lifecycle messages appeared on standard error; disabling API log delivery does not silence terminal logs. The owned server processes, private state and listener were cleaned up after the run. This verifies native local orchestration and saved-project status recovery. It does not prove image creation, media download, Prefect Cloud deployment or an upstream integration listing.