[forrestlinfeng]

Two misreadings, found while aligning the Codex and WorkBuddy plugins with this one (they already carry the fixed versions).

Check output

verificationFailed dropped every output line that held a zero count, so a mixed summary lost its failure along with the line:

verificationFailed({ content: [{ type: 'text', text: 'Tests: 1 failed, 0 errors' }] }) // false on main

It matters whenever the exit code is gone — the reason output is read at all is that dsh's bash reports a failing command piped through tail as a success. Only the zero counts are removed now (replace with a global pattern), so 0 failed / failures: 0 still read as a pass and 1 failed, 0 errors reads as a failure.

Corrections

looksLikeCorrection already spared 行不行 / 对不对, but took 不行吗？, 这样不对吧, 不对吧。 for a correction. A correction revises the reuse report (L3 failed) and drops the turn's distillation candidate, so a question cost a valid Gene. 不行 / 不对 no longer count when what follows asks for an opinion: 吗/嘛/么, a question mark, or a closing 吧. A plain 这样不对, 不行，还是报错 or 不对吧，你再看看 is still a correction.

Tests

New cases in test/verification.test.js and test/feedback-label.test.js (both fail on main). npm test: 185 pass.

🤖 Generated with Claude Code

[evomap-aiqa-bot]

Sentinel 已审核此变更，未发现需要报告的问题。

开发 agent 查询详情： 调用 MCP 工具 sentinel_pr_review_context，传入 owner_repo="evomap/evolver-dsh-plugin"、pr_number=53、head_sha="fd56079fa3eef58a4be47a5432f3ac35b461d30b"，并按条目追加 finding_id 或 finding_key。

由 Sentinel 审核 · 目标提交 fd56079fa3ee