Show HN: Plan Review – independent review for Claude Code plans

I'll pull the GitHub repo and HN thread so the paragraphs stick to real product and discussion details, not invented figures.Here is a 5–6 paragraph body you can drop into the post:

Plan Review landed as a Show HN from dachev’s open-source GitHub project at github.com/dachev/plan-review: an independent review layer for Claude Code implementation plans. The pitch is simple—plan twice, code once—by routing a finished plan through a second model before you approve execution. The HN thread sat at 1 point with 0 comments at scrape time, so the signal so far is the repo and install path rather than community debate.

Technically, the tool is a Claude Code plugin that gates plan approval with optional multi-round critique. When Claude submits a plan, you can Review the plan, pick Review with… (reviewer and model), or Skip review. The reviewer is either OpenAI’s Codex CLI or a separate Claude instance; rounds continue until blocking feedback is resolved, and every finding hits a triage dialog—Investigate (Claude evaluates and can push back) or Dismiss (dropped for good). Manual control is available via /plan-review with --policy minimal|default|strict and --max-rounds N, plus resume and ls for parked reviews. Preferences live in ~/.claude/plan-review/config.json (reviewer, model, speak, prompts). There is no plan-review-hosted backend: the chosen reviewer CLI sends plan and context under your provider account, workspace access for the reviewer is read-only, and review records stay local under .claude/reviews/ as plain markdown. Runtime needs are Bash and Node.js 18+; install is /plugin marketplace add dachev/plan-review then /plugin install plan-review@plan-review, with hooks registering on a new session. Automatic offer gates fail open so a broken gate cannot block normal plan approval.

For builders already living in plan mode, this targets the expensive class of agent mistakes: wrong architecture, missed constraints, and silent scope creep that only surface after code lands. Reviewing the plan is cheaper than reviewing a finished diff because design choices are still editable text. The Investigate path keeps a human in the loop without rubber-stamping every nit, and Superpowers workflow plans get the same offer at execution handoff even when they skip plan mode.

Market context is the crowded “second opinion for agents” niche around Claude Code and Codex, not a new foundation model. Plan Review does not replace either product; it composes them so planner and critic can be different models or vendors. That cross-provider check is the differentiator versus self-critique inside a single session. MIT licensing, a TypeScript codebase with tests (Vitest), and local-only review state position it as a developer-owned control plane rather than a SaaS review service—relevant if you care about plan and transcript sensitivity and .gitignore hygiene for .claude/reviews/ and .claude/plan-review/.

Practical next steps: install a reviewer CLI (codex and/or claude), enable the plugin, and try one real plan under default policy before tightening to strict. Watch whether automatic offers stay reliable, how long multi-round reviews take in large repos, and whether the roadmap (triage shortcuts, richer finding detail, more reviewers beyond Codex/Claude, more planner integrations) reduces dialog friction. If unattended rounds hit Claude Code’s stop-hook block cap, the project documents raising CLAUDE_CODE_STOP_HOOK_BLOCK_CAP. Track the Show HN thread and the repo for adoption signals beyond the initial 1-point, 0-comment scrape.