01 / InputsPlain prompts. Explicit arguments.
Store templates in a prompts directory and supply task arguments as JSON objects. Values are inserted literally; they are not executed as code.
- Reusable task inputs across multiple runs
- String application identifiers separate from task IDs
- Prompt paths contained within the configured directory
02 / ExecutionCodex CLI, coordinated.
The framework launches independent Codex CLI workers. Each stored run determines the model and reasoning effort used by its tasks.
- Configurable concurrency and attempt deadlines
- A shared rolling limit on task runner starts
- Soft-stop draining and forced process cleanup
03 / ResultsAn outcome you can follow.
A successful response and its terminal attempt state are committed together. A failed attempt stays in history when an explicit retry is created.
- One saved successful response per run and task
- Distinct task progress and historical attempt totals
- JSONL export with application unit identifiers
04 / MonitoringA local view of the work.
The operator dashboard shows the selected run, active workers, recent completions, and effective session limits.
- Processing rate and estimated remaining time
- Reported token usage, when available
- Account-wide quota observations kept separate
05 / StorageA small, deliberate data model.
Four framework-owned SQLite tables hold tasks, runs, attempts, and results. Application data stays in a separate database.
- WAL mode and foreign-key enforcement
- One local coordinator per service database
- Explicit migrations preceded by a verified backup
06 / IntegrationYour application at both ends.
Preparation and export scripts are invoked separately. Another domain can connect through the Store API without adding application SQL to the core.
- Prepare tasks and create a runnable batch
- Validate saved text using application rules
- Own your conflict checks and export policy