Vibe Coding
qwen-code Bans Its Review Agents From Claiming Mutations They Never Ran
PR #9923, merged 2026-08-25, adds witness discipline to two reader-agent briefs: an unrun mutation must be phrased as a hypothesis, the exact strings "ships N/N green" and "verified N/N green" are named and banned, and any finding whose weight depends on an unperformed run must carry `witness: not run — <why>`. The post-mortem in the PR body is the useful part: across multi-round dogfood reviews, code-behaviour Criticals were well-evidenced with real reproductions, but the coverage findings that made up the bulk of the tail asserted "mutant verified 27/27 green" while the same review's gap notice admitted the worktree had no `node_modules` and the suite could not execute.
Source
↳ Follow the thread