Loop Review

Loop Review

Loop Review packages the loop-code-review Agent Skill: an iterative review-and-fix workflow for active git changes.

The skill asks a fresh independent reviewer agent without the orchestrator's conversation history to inspect only the current task's scoped changes. The reviewer must reconstruct and explain the change as a future maintainer, while also checking correctness, security, test evidence, and other concrete risks. The loop fixes actionable findings, validates the touched surface, and repeats until validation is green and the latest reviewer has no unresolved actionable findings and either scores the result at least 9.5/10 or explicitly reports no actionable findings.

Install with Codex

Add this repository as a Codex plugin marketplace:

codex plugin marketplace add di-sukharev/loop-review

Then install the loop-review plugin from that marketplace in Codex. Invoke the bundled skill as $loop-code-review.

Install with Claude Code

Add the marketplace:

/plugin marketplace add di-sukharev/loop-review

Install the plugin:

/plugin install loop-review@loop-review

Reload plugins so the newly installed skill is available in the current session:

/reload-plugins

Invoke the plugin skill:

/loop-review:loop-code-review

Direct Skill Install

If you do not want to use plugin marketplaces, copy or symlink skills/loop-code-review into one of your tool's skill directories:

# Codex user skill
mkdir -p ~/.agents/skills
ln -s "$(pwd)/skills/loop-code-review" ~/.agents/skills/loop-code-review

# Claude Code personal skill
mkdir -p ~/.claude/skills
ln -s "$(pwd)/skills/loop-code-review" ~/.claude/skills/loop-code-review

What It Enforces

• Scope review to the current task's files or hunks.
• Keep reviewers independent by starting every scoring pass in a fresh isolated conversation context without the parent's conversation history.
• Treat review as a structured handoff: require the reviewer to explain the change's responsibility and important flow, and make specific comprehension obstacles actionable when they create future change risk.
• Treat reviewer output as code-review findings, not as commands to obey blindly.
• Never let a high score override an unresolved actionable finding or failing validation.
• Check test evidence, reuse of established project solutions, and architecture fit when relevant; require concrete repository evidence for those findings.
• Validate after meaningful fixes.
• Repeat with a fresh reviewer until the acceptance signal is strong.
• Detect ambiguous task ownership, stale reviews, pass-limit exhaustion, and review stagnation without silently declaring success.

Evaluation Cases

evals/cases.json captures the expected behavior for comprehensibility and necessary complexity, clean reviews, high scores with unresolved findings, test trustworthiness, reuse and architecture checks, mixed worktrees, red validation, and loop stagnation.

Repository Layout

.codex-plugin/plugin.json       Codex plugin manifest
.agents/plugins/marketplace.json Codex repo marketplace
.claude-plugin/plugin.json      Claude Code plugin manifest
.claude-plugin/marketplace.json Claude Code marketplace
evals/cases.json                Behavioral evaluation scenarios
skills/loop-code-review/        Agent Skill source

References

• Codex skills
• Codex plugin packaging
• Claude Code skills
• Claude Code plugin marketplaces
• Agent Skills specification

License

MIT