dual-review-loop

dual-review-loop

Claude Code plugin. Auto-iterate work under dual-review verification, auto-apply high-confidence findings, atomic-commit. Two modes:

• plan mode (/dual-review-loop:dual-review-loop) — iterate a plan file's - [ ] checkbox tasks.
• task mode (/dual-review-loop:dual-review-task) — iterate a free-form inline task; you decompose one sub-step per iteration.

Companion to dual-review.

⚠️ Currently requires Korean-language dual-review output. The injected iteration prompt instructs the LLM to apply Accept findings using Korean headings (## ✅ Accept — 양쪽 독립 합치 etc.); English support is not yet implemented — see Prerequisites.

Prerequisites

• Claude Code
• jq, git on PATH
• dual-review skill installed at ~/.claude/skills/dual-review/
• At least one reviewer backend that dual-review can dispatch:
  • superpowers:code-reviewer + codex:adversarial-review (preferred), or
  • oh-my-claudecode:critic (fallback)
• Korean-language dual-review output. The injected iteration prompt names Accept buckets by their Korean headings (## ✅ Accept — 양쪽 독립 합치, ## ✅ Accept — 단일 리뷰어, 기술적으로 타당, ## Open Questions). The hook itself only parses ## Open Questions (English heading) for the early-stop gate; the rest live inside the prompt body. If you fork dual-review to emit English Accept headings, update the prompt strings in commands/dual-review-loop.md and commands/dual-review-task.md accordingly.

Install

Inside Claude Code:

/plugin marketplace add hcooch2ch3/dual-review-loop-plugin
/plugin install dual-review-loop@dual-review-loop-plugin

If /plugin install reports "source type your Claude Code version does not support", fall back to the CLI form (same operation, different code path):

claude plugin marketplace add hcooch2ch3/dual-review-loop-plugin
claude plugin install dual-review-loop@dual-review-loop-plugin

For local development, clone the repo and add the local path as a marketplace instead:

/plugin marketplace add /absolute/path/to/dual-review-loop-plugin

See the Claude Code plugin docs for details.

Use

Plan mode — iterate - [ ] checkboxes in a plan file:

/dual-review-loop:dual-review-loop /abs/path/plan.md [--max-iters N] [--max-minutes M]

Task mode — iterate a free-form task description (no plan file):

/dual-review-loop:dual-review-task "<task description>" [--max-iters N] [--max-minutes M] [--max-files N] [--max-loc N] [--max-reviews N]

Cancel either mode:

/dual-review-loop:cancel-loop

Defaults: --max-iters 20, --max-minutes 30. Both hook-enforced (Gates 10, 10b). --max-minutes 0 disables wall-clock cap (24h idle gate still applies). Task mode also enforces cumulative caps: --max-files 30, --max-loc 1500, --max-reviews 15 (Gates 10c–f). Plan mode defaults these to "Infinity" so they don't fire unless explicitly lowered.

Cancel manually: rm .claude/dual-review-loop.state.json in project root.

Which mode?

	plan	task
Input	plan file with - [ ] checkboxes	inline "<task description>"
Iteration unit	one checkbox at a time, executed in order	one sub-step you decide each iter
State of progress	checkbox flip on the plan file	append-only iteration log at .claude/dual-review-loop/task-<session>.log.md
Use when	requirements are pre-decomposed into steps	task is one paragraph; LLM should figure out the decomposition

Cancel preserves the task log (post-mortem artifact). Add .claude/dual-review-loop/ to .gitignore to keep logs out of git history.

Known issues