How it works

From a source record
to a saved response.

A small set of explicit steps connects your application to Codex CLI agents. Each step has a clear owner.

Prepare the input

Your preparation script selects source records and adds reusable tasks. A task names a prompt file, supplies JSON arguments, and can carry a string unit identifier.

store.add_task("summarize.md", {"text": source_text}, unit_id="note:001")

Create a run

Select the task IDs, model, and reasoning effort. Creating a run also creates its initial pending attempts in one transaction. Task rows alone are not a runnable batch.

store.create_run(model, reasoning, task_ids)

Process with Codex CLI

The coordinator claims pending attempts and renders their prompts. Independent workers invoke Codex CLI with bounded output, a deadline, and mandatory MCP and additional-integration isolation.

pending → in_progress → completed / error / timeout

Inspect what happened

The local dashboard distinguishes task progress from attempt history. A soft stop drains active work. Failures can be selected for an explicit retry without erasing the original attempt.

A retry creates a new attempt; it does not replace history.

Validate and export

The framework saves successful response text. Your exporter applies domain rules and writes to the application destination. Saving a response does not automatically deliver it.

store.iter_results(run_id)