Automatically commit your changes with this commit template:
subsystem: short one-line description; semicolon if multiple changes
Model-output: model name e.g. Claude Fable 5
<prompt>
user prompt, verbatim
AskUserQuestion question-answers, if any
</prompt>
<slop>
model's response at the end of the dialogue, verbatim, in markdown format
</slop>
"(mid-turn)" if I added something mid-turn; multiple ... if the conversation had several real turns.
Commit often; okay to commit more than once per turn!
After each commit you make, get it reviewed by Codex (GPT-5.6-Sol at xhigh reasoning):
codex review --commit <sha> -c model="gpt-5.6-sol" -c model_reasoning_effort="xhigh"
Notes:
- Codex is configured globally in
~/.codex/config.toml(approval_policy = "never",sandbox_mode = "danger-full-access") to never ask for permission and run unsandboxed, so reviews andcodex execruns never block on prompts. If codex ever stalls waiting for approval, check that file. - A review can take several minutes; run it in the background and continue if you have other work.
- Sol often nitpicks, or cares about bizarre, irrelevant edge cases. Ignore those findings; they should not stop you from making progress.
- For oversights that are true and interesting, fix them and make another commit (using the usual commit template). If you fixed nothing, say briefly in your reply why the findings didn't warrant changes.
- Do NOT send that follow-up fix commit through another Codex review — the review cycle for a change ends after one round of findings and fixes. (Exception: the follow-up grew into something substantial beyond addressing the findings.)
- If you made several commits in a row, make sure the reviews cover all of them: either review
each commit, or run one ranged review of the whole batch with
codex review --base <sha before your first commit>plus the same-coptions.