source-command-sr-loop

Invoke the sr-loop skill with a PR URL, branch name, or blank for the current branch's diff. It repeats simplify passes, then code-review --fix passes, testing, committing, and pushing after each pass until both phases report zero new findings.

Each phase converges when a full pass returns zero new findings. A pass that finds only nits is applied, tested, committed, and pushed, then treated as clean without running another pass unless the user requests it.

Yes. The skill runs `claude -p "/code-review low --fix" --max-turns 40` as a headless subprocess with a 10-minute timeout. A -p prompt arrives as user input, so the disable-model-invocation flag does not block it.

The skill falls back to invoking the e-code-review skill with `low --fix`, which follows the same review-and-fix contract without the invocation restriction, then continues the loop normally.

Scoped tests covering only files beside the touched code keep each pass within a single foreground tool call's lifetime. A full-suite run can outlive one tool call, and a hung or newly slow suite is treated as a finding to investigate.

The loop must run in the main conversation. A wrapper subagent cannot see the review skills, and a nested agent's own subagent reports route to the main conversation instead of back to it, breaking the loop.