claude-foundations/memory/script-review-loop-run.md

review-loop-run

Purpose: Launch and track detached MiniMax M3 reviews of /review-loop artefacts (verify / plan / spec / tests / code) from a fixed prompt template per kind, so sessions stop hand-writing prompt files, launcher scripts and (fatally) pkills.

Location: small-scripts/scripts/review-loop-run → ~/sbin/review-loop-run; spec small-scripts/specs/review-loop-run.spec.md; test tests/test-review-loop-run.sh.

Usage

review-loop-run launch --kind spec --name spec-BC --round 2 --file /tmp/x/spec-BC.diff --context "…"
review-loop-run status --name spec-BC
review-loop-run counts <result.md>        # exit 4 when CRITICAL+HIGH > 0

Files: ${XDG_RUNTIME_DIR:-/tmp}/review-loop/<repo>/<tag>-r<N>-{prompt.txt,launch.sh,stdout.txt,result.md}. Round ≥2 auto-adds the previous round's result and the "do not re-raise adjudicated findings" paragraph.

How it works

• One task template per kind (the spec template excludes the "symbol does not exist yet" class for spec-ahead specs; tests adds Bash(python3 -m pytest *); verify adds git show).
• Default reading set per kind from best-practices/ + the project's memory/gotchas-*.md; absent files are skipped and reported in --dryrun.
• Launch is a generated <tag>-launch.sh run under setsid nohup; the stdout file ends with exit=N. Kill by PID from the launch script's process, never by pattern.
• counts parses table rows, - **SEV:** n lists, or heading counts as a last resort.

Gotchas

• pkill -f 'MiniMax-M3' kills the invoking shell — the reason this script exists. Never reintroduce it.
• MiniMax reviews take 5–10 min for ~15 input files; schedule a wakeup rather than polling.
• counts heuristics depend on the reviewer honouring "end with a count by severity"; the prompt asks for it, but spot-check the result on the first round of a new kind.

Used by: /review-loop (custom-claude-skills).