Auto commit in skill /implement

[dzinhtuong]

Why should we add an automatic commit instruction after code review? I think the decision to commit should be made by the user, not by an AI agent. While an AI agent might be able to suggest the next step after implementation, it shouldn't be the sole decision-maker for committing.

[YevheniiVolosiuk]

Because it was created for an agent loop using SandCastle, you should review PRs rather than reviewing code before it is committed.

[jackiectl2]

That concern makes sense for an interactive local workflow. The current implement skill ends with the unconditional line “Commit your work to the current branch.” As the earlier reply notes, this fits an unattended agent loop where the human reviews a PR afterward; it is a workflow choice, not a requirement for /code-review itself.

If you want the commit to remain a human decision, I would keep a local copy/override of implement and replace only that final instruction with something explicit, for example:

After code review and tests, show git status and a concise diff summary. Do not commit or push. Ask the user whether to commit this exact change set; if they agree, use a descriptive commit message.

That keeps the useful parts of the skill (ticket, TDD seams, checks, review) while making the handoff point clear. A separate commit policy in AGENTS.md could also constrain the agent, but an override of the instruction that actually says “Commit” is less ambiguous. If you later switch to the unattended PR loop, restore the original final line. I would also review the override when updating this repo, since a copied skill does not inherit upstream edits automatically.