caidish/cai-tools/codex

Codex Skill Guide

Requires Codex CLI 0.153+ (codex --version). --full-auto no longer exists in this line; use --approve-for-me.

Models and Reasoning Effort

Effort ladder for model_reasoning_effort: low < medium < high < xhigh < max (provider also accepts none, minimal).

• Default xhigh. max is for astra/sol on genuinely hard problems; it is slow and token-heavy.
• Suggested pairing: astra/sol xhigh or max; terra medium/high; luna low/medium.
• Astra at low roughly matches Sol at high; do not raise effort by default.
• ultra (parallel internal sub-agents) is accepted by astra/sol on some builds but is not in the provider's validated list. Use only on explicit request; on an invalid_value error fall back to max.

Running a Task

1. If the user already named a model and effort, use them. Otherwise ask once via AskUserQuestion with two questions in a single prompt: model (gpt-6-astra (Recommended), gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna) and effort (xhigh (Recommended), high, medium, max). Pre-select per the table above based on task difficulty.
2. Pick the sandbox. Default --sandbox read-only. For edits: --sandbox workspace-write --approve-for-me. Add -c sandbox_workspace_write.network_access=true when the task needs network inside workspace-write. --sandbox danger-full-access only with explicit permission.
3. Assemble:

``bash echo "<PROMPT>" | codex exec --skip-git-repo-check -m <MODEL> -c model_reasoning_effort="<EFFORT>" --sandbox <MODE> [-C <DIR>] [--add-dir <DIR>] [-i <IMAGE>] [-o <FILE>] [--ephemeral] - 2>/dev/null ``

• Always --skip-git-repo-check. Always 2>/dev/null (hides thinking tokens on stderr) unless the user asks for them or you are debugging.
• -o <FILE> writes only the final message; use it when you must parse the answer. --ephemeral for throwaway runs that need not be resumable.
• Pass timeout: 600000 or more to Bash; astra/sol at xhigh/max can run many minutes.

1. Run, capture stdout, summarize the outcome.
2. Then tell the user: "You can resume this Codex session at any time by saying 'codex resume' or asking me to continue with additional analysis or changes."

Quick Reference

Resuming

• echo "new prompt" | codex exec --skip-git-repo-check resume --last - 2>/dev/null. Global flags go between exec and resume; --last, --all, -c go after resume.
• The resumed session inherits model, effort, and sandbox. Do not pass overrides unless the user asks; resume has no -m, so change the model with -c model="<MODEL>" and effort with -c model_reasoning_effort="<EFFORT>".
• Add --all after resume if the original session started in a different directory.

Following Up

• After every codex command, use AskUserQuestion to confirm next steps, collect clarifications, or decide whether to resume.
• Restate the chosen model, effort, and sandbox when proposing follow-up actions.

Error Handling

• Stop and report when codex --version or any codex exec exits non-zero; ask for direction before retrying.
• Before using --approve-for-me, --sandbox danger-full-access, or --dangerously-bypass-approvals-and-sandbox, ask permission via AskUserQuestion unless already given.
• On invalid_value for reasoning.effort, rerun with the next lower effort in the ladder.
• When output includes warnings or partial results, summarize them and ask how to adjust.