sol-loop

Your agent checks each Sol finding against the code, fixes the real problems with a small change, and answers disagreements with evidence. Findings outside the intended behavior become follow-ups instead of expanding the diff.

Install the skill with the command above, then say "sol loop" and identify the change and its purpose. The agent selects it for that phrase. For example:

sol loop on this branch; it only fixes the retry backoff.

Use it before opening a PR or when an existing change needs another review. Your host must be able to run OpenAI's GPT Sol; the skill reports unavailable Sol access instead of substituting another model.

Full-send uses it to review the code and PR description together.

The agent verifies the change range, excludes unrelated work, and runs covering checks before sending the diff to one Sol reviewer. Findings have three severities:

The agent verifies every finding itself. Agreed findings get the smallest fix; disputed findings get an evidence-based response. Nits and out-of-scope items are declined with a reason and listed as follow-ups. After fixes, the agent reruns checks and reduces any unnecessary diff growth.

The default budget is three rounds. It can extend to five for security, permissions, authentication, data loss, or state and concurrency work while an evidenced blocker or should remains open, or when you ask. A round with only nits ends the loop early.

SKILL.md has the exact scope, round, and approval rules.

• Agreement when Sol approves the current behavior diff and no blocker or should remains, or both sides of a disagreement for you to decide.
• Rounds, the model that ran, changes prompted by review, before-and-after diff size, and fixed or declined findings with reasons.

Exhausting the budget is not approval: you get both sides of any open disagreement to decide. The agent never pushes during the loop. After agreement, it pushes only if you authorized a push or PR update. Use babysit-pr for CI and bot review after that.