review

Attack a code change and report what survives. Dispatches independent reviewers in parallel, each with a single job and no access to the implementer's reasoning: one hunts correctness bugs blind to the issue, one hunts requirement gaps against the issue text, one runs outside the process entirely. Optionally re-attacks its own findings, then synthesizes what is left. Returns findings and changes nothing — no tooling, no autofix, no edits. On request it publishes to the author, gated on demonstrated evidence. Every phase is configurable.

Post-implementation cleanup of correct code. Prunes comments aggressively: deletes what restates the code, files design rationale to the commit message and wider context to the PR or an issue, collapses one mechanism explained in three places, strips issue and PR references that make a file unreadable on its own, compacts genuine gotchas, renames unclear symbols so a comment is not needed, checks the code against the project's conventions, and simplifies what is more complicated than it needs to be. Tightens docs without gutting them. Auto-applies and reports.

Extract a behavioral specification from a source-code bundle — a handful of files up to a whole 500+ file application. Produces reproducible, stack-neutral specs another engineer or LLM could use to rebuild the same observable behavior without reading the original source.

Approach board for a problem or a decision. Dispatches independent seats — some generating candidate approaches, including the simplest thing that could work and one agent outside this process entirely, some attacking the assembled set comparatively — then verifies the surviving objections and reports a ranked recommendation with its trade-offs. An approach already on the table enters as one candidate among several. Autonomous; changes nothing.

Rule on a set of claims that already exists — review findings, a plan's assumptions, an analysis, a reviewer's objection — before anyone acts on them. Every claim gets exactly one verdict: it holds, it holds only in a narrower case, it is true but not worth the fix, it falls, or nobody can settle it. Whoever judges never sees the reasoning that produced the claim. Reads and reasons; changes nothing.

Handle a pull request from whichever side you are on. On someone else's branch it attacks the diff through changes-review, then publishes a real review with per-line comments and a verdict. On your own it works the threads people and bots left — decides its standing in each, verifies the claim before answering, runs the project's checks, then replies and resolves. Changes no code unless asked.