What problem does it solve?
It prevents Codex CLI runs from silently failing or losing evidence when prompts are large, quoting breaks, network timeouts occur, or outer schedulers terminate early.
Core Features & Use Cases
- Prompt file first: Forces the prompt to be written to a file and piped via stdin to avoid shell escaping and quoting issues in long commands.
- Deterministic execution + retry loop: Runs
codex exec with per-attempt console/stderr logs and retries, then verifies exit code and non-empty raw output.
- Failure classification + recoverability: Classifies failures into timeout/no-stream, timeout/partial-stream, network instability after retries, empty output, and generic execution errors; extracts
session_id when available and provides a resume suggestion.
Use case: You need Codex to execute a long, non-interactive planning or coding task across projects while always producing retriable, auditable artifacts under docs/codex/.
Quick Start
Ask the AI to run Codex using the prompt file docs/codex/_inputs/codex-task.prompt.md with scope codex-task and to archive raw and failures under docs/codex/.