gpt-workflow for Python

gpt-workflow is a synchronous, typed Python wrapper around the deterministic gpt-workflow CLI. It blocks until each command finishes, captures the CLI's NDJSON and diagnostics, and does not print or retry work automatically.

Prerequisites

• Python 3.12 or newer
• Bun 1.3 or newer
• an authenticated Codex CLI for workflows that call agents

The Python and npm distributions are synchronized at 0.3.4. This release always executes bunx --bun gpt-workflow@0.3.4; it never resolves @latest or a separately installed gpt-workflow executable.

Install the Python package after it is published:

python -m pip install gpt-workflow==0.3.4

Run and inspect workflows

Set gpt_workflow.cwd once to the repository that owns the workflow. It must be a pathlib.Path; relative script paths and durable run storage are resolved from it.

from pathlib import Path

import gpt_workflow

gpt_workflow.cwd = Path("/absolute/path/to/repository")

execution = gpt_workflow.run(
    ".codex/workflows/summarize.js",
    {"topic": "deterministic orchestration"},
    default_model="your-codex-model",
)

print(execution.result)
print(execution.status.run_id)
print(execution.run_directory)

for summary in gpt_workflow.runs():
    print(summary.run_id, summary.status)

status = gpt_workflow.status(execution.status.run_id)
available_models = gpt_workflow.models()

Omitting the second argument leaves workflow args undefined. Passing None sends explicit JSON null. WorkflowResult contains the JSON result, detailed WorkflowStatus, and durable run directory. runs() returns lightweight summaries; status() loads detailed phase, agent, token, failure, and fallback journal state; models() returns unique canonical model names in CLI order. Unknown compatible JSON fields are retained in each value's extra mapping.

This no-agent workflow is deterministic and spends no model tokens:

export const meta = {
  name: "python-smoke",
  description: "Python SDK smoke test"
}
return { answer: 42 }

Errors and interruption

Configuration and validation fail before work starts with idiomatic exceptions: WorkflowDirectoryUnset, FileNotFoundError, TypeError, ValueError, BunError, or CLIProtocolError. Once a trustworthy run exists, failures use WorkflowError subclasses whose status and run_directory are always set. These distinguish invalid workflows and arguments, unavailable models, Codex App Server failures, budgets and limits, Git failures, JSON boundaries, cancellation, and unclassified workflow execution.

Ctrl-C before run.started remains an ordinary KeyboardInterrupt. After the run starts, the wrapper forwards SIGINT, gives the CLI a bounded opportunity to flush persisted failure state, then raises WorkflowInterrupted. That exception is also a KeyboardInterrupt and carries status and run_directory.

v1 scope

This release intentionally keeps Bun as a system prerequisite and exposes only synchronous APIs. CI, PyPI publication, bundled Bun support such as a gpt-workflow[bun] extra, and async APIs are deferred to later work.