implement-with-sol

implement-with-sol

Codex implements in a worktree; you keep planning, review, and synthesis. Ends with a PR ready for human review.

Steps

1. Worktree: create one with git wt <branch> origin/<default-branch> and run everything from inside it.

2. Spec file: save the task spec verbatim to a temp file (e.g. gh issue view <n> --json title,body -q '"# " + .title + "\n\n" + .body' > /tmp/<slug>.md) so the prompt references a path instead of inlining it.

3. Launch from the worktree, in the background:

   codex exec --sandbox workspace-write "<prompt>"
   

   Global flags precede the subcommand: follow-ups are codex exec --sandbox workspace-write resume --last "<prompt>" — flags placed after resume fail. The model comes from ~/.codex/config.toml; override with -m.

   The prompt names: the spec path with "follow every decision in it"; the existing code and test patterns to study; the definition of done (build, lint, and tests pass; required cases covered); the review instruction "when done, run the review-fix skill on the working tree with its Default Review Policy — supply no review brief"; "Do NOT commit — leave changes in the working tree"; and "print files changed and each verification command with its result".

Word the review instruction exactly that way: any prose describing what to review becomes a Review Brief, and a brief replaces the Default Review Policy with a single ad-hoc reviewer instead of the full panel.

1. Verify: rerun the claimed checks yourself, review the diff against the spec decision by decision, and smoke-test edge cases the spec calls out. Codex's report is a claim, not evidence.

2. Fix rounds: send each finding via resume --last as in step 3 and re-verify, until a review pass finds nothing. Trivial one-file tweaks may be fixed directly instead.

3. Ship: commit per the logical-commits skill (git-hunk skills get core logical-commits), then invoke sending-pull-request.

Done when a review pass finds nothing, the checks pass locally, and the PR URL is returned.