Single-agent baseline that gives one Codex CLI session the complete search loop inside the task's own image: Codex edits and evaluates, while Galapagos independently scores every observed checkpoint.
# Codex — delegate the whole loop to one `codex exec --json` session. Docker-mode tasks install the
# CLI into the task's own image; local-mode tasks use the host CLI. `general.max_iterations` is the
# prompt-level improvement budget. Since Codex has no --max-turns flag, Galapagos also enforces a
# safety cap over completed actionable JSONL items. Auth is ChatGPT-managed: CODEX_ACCESS_TOKEN wins,
# then a file-backed `codex login` cache; API keys are deliberately ignored.
seed: 0
general:
max_iterations: 100
checkpoint_interval: 10
proposer:
codex_model: gpt-5.6-sol # `codex exec --model`; null = isolated CLI default
codex_cli_max_steps: null # actionable JSONL item cap; null = 2x general.max_iterations
codex_wall_timeout_seconds: null # whole-session timeout; null = derived from step/eval caps
reasoning_effort: max # minimal|low|medium|high|xhigh|max|ultra (model-dependent)
reasoning_summary: auto # auto|concise|detailed|none; null = CLI default
codex_web_search: disabled # disabled|cached|indexed|live; disabled keeps benchmarks fair
codex_docker_image: null # task image override; null = resolve from task card
codex_docker_memory: null # Docker memory cap, e.g. "4g"; null = unbounded
codex_docker_cpus: null # Docker CPU cap; null = unbounded
codex_docker_network: null # null = task setting or Docker's default bridge
codex_docker_cli_version: null # pin @openai/codex@X; null = latest during image build