magi: Chat-only CLI chains claude and codex to release

magi is built around a magi.toml configuration file split into sections such as [graph], [verify], [merge] and [release]. Before implementation, three advisors by default each produce a read-only design proposal, and a synthesizer merges the good parts into one policy rather than picking a winner; this stage can be turned on or off with [graph] advise. Implementation can run as a competition, where candidates are labeled A, B or C, judges rank them, and a split leads to discussion and a vote. The default for [graph] implementers is 1, but the magi.toml that magi init writes sets it to 3, so using it as-is runs three competitors.

During the review loop, each reviewer gets a dedicated detached worktree fixed to the commit under review, so the working tree does not change and does not clash with the fixer. Commands written in [verify] e2e run in that worktree and their output goes to the fixer, which fixes blocking findings or can reject them with evidence. Finding IDs in a format like R2-1-3 are assigned by magi rather than the agents, which the author says allows reviewers' accuracy to be aggregated, and a round where a reviewer fails to answer due to a timeout is not counted as "no findings".

The author writes that the biggest personal benefit is no longer having to fill the gap between asking and approving: notifications reach a phone only where a judgement is needed, and everything else keeps moving. Judging mechanisms and other parts are collected in the README.