agent-plan-reviewer

A read-only adversarial plan-review subagent for Claude Code, Codex, and OpenCode. Hand it a plan; it tries to break the plan before you commit to it.

What it does

Given a plan or design plus context, it distrusts every claim and verifies it against the real code (does the file, symbol, or API actually exist and behave as assumed?), then hunts for gaps, blind spots, edge cases, infeasibilities, intent divergence, and missed steps. It returns findings tagged [BLOCKER] / [SHOULD-FIX] / [NIT], each with a concrete scenario, file:line evidence, and a suggested fix. It never edits anything; when it needs to confirm an external library or API, it uses whatever lookup capability is available rather than guessing.

Its instructions are tool-agnostic: it uses whatever capabilities the environment exposes (the built-in file, search, and shell tools, plus any servers the project provides) and treats them as best-effort. If a capability is missing, unhealthy, or errors, it notes that and falls back rather than halting, so it works across very different setups and never stalls a fan-out because one tool is down. It stays read-only by rule, even where write-capable tools are present.

When to use it

At a planning gate, before confirming a non-trivial plan (on Claude Code, before ExitPlanMode). Run two or three in parallel with disjoint lenses (correctness, edge cases, feasibility) and fold the findings back into the plan.

Install

npx @ctxr/kit@latest install @ctxr/agent-plan-reviewer

@ctxr/kit installs the bundle and mirrors it into the host's agent directory; the agent then appears as agent-plan-reviewer.

Releasing

Releases are PR-gated and the publish step is a manual dispatch; the CI workflows (ci, release, tag-on-main, publish) carry the details.

License

MIT