fix(planning): restate only changed answers and unseen assumptions at final confirmation

已关闭 适合新手
  • 哪里不对？
#11,189 2 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
2/5
预计耗时
1-3 小时
新手友好度
68/100
Issue 类型
功能
描述清晰度
描述清楚
活跃度
活跃
技术栈
markdown
领域
ai-infra-agents

调研方向

阅读 plugins/planning/skills/interview/SKILL.md 和 context/loop.md 中的“Recap and procedure check”部分，找到添加规则的最佳位置；issue 允许放在任一位置。只添加一次规则，并将 interview-final-brief-changes-only 列为 eval 证据。运行 scripts/check-affected.sh --tests 和 scripts/check-changelog-fragments.sh；两者都通过（包括任何必需的 digest 更新）即表示完成。

由索引模型根据 Issue 内容生成。

描述

host-run priority: medium status: ready work-class: scoped

Parent: #9794 (split from #7269, now closed). Eval-first case interview-final-brief-changes-only from #7247 fails on main; #7303's eval-first page items depend on these results.

Problem

When every answer was already read back and accepted, the final restatement re-lists all of them (UTC, 50,000 rows, toolbar, gzip, audit log), burying the one changed answer and the two assumptions the user has never seen. With the plugin, runs also turned the sweep's assumptions into new questions instead of restating for confirmation.

Change

In plugins/planning/skills/interview/SKILL.md "Recap and procedure check", or context/loop.md, add: when each answer was read back as it arrived, the confirmation restatement gives the intent, each answer changed after its read-back, and each assumption the user has not yet seen, and does not re-list unchanged accepted answers (the --changed recap covers them in one counting line).

Evidence

Host run on 2026-10-30, Claude Code 2.1.314, claude plugin eval against an out-of-git copy of plugins/planning at 543166dde6, --runs 3 --judge-model opus --threshold 0.8, default with/without ablation. run-validity.py: VALID (warning: skill fired, since the slash command loads the hub without a Skill tool call; the with-arm output follows the hub's procedure). Scores: with 0.42, without 0.50, delta -0.08; with-arm changes-only-restatement 0 of 3, omits-unchanged-items 0 of 3, shows-changed-item 3 of 3, shows-new-assumptions 2 of 3. No with-arm tool call errored or was denied, including every read under skills/interview/context/, so the result is not the bare-path defect. Judge uncalibrated. Full record on #9794.

Done when

  • The rule above is in the named spoke (or hub line), stated once, with the case named as its eval evidence.
  • scripts/check-affected.sh --tests passes on the committed head, including plugins/planning/tests/interview-defenses.test.sh with any digest it pins re-pinned.
  • scripts/check-changelog-fragments.sh passes.
  • The PR body asks for a host rerun of interview-final-brief-changes-only (3 runs, opus judge); the case reaching a with-arm score of 0.8 or more in a VALID run is that rerun's check, not this slice's gate.
主要语言
Shell
星标
22
派生
2
平均合并
1 小时 38 分钟
30 天内合并 PR
911

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 没有 Pull Request 模板
  • 阅读贡献指南

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

melodic-software/claude-code-plugins 的其他 Issue

查看 melodic-software/claude-code-plugins 的全部 Issue

相似的 Issue

更多 Shell/Bash Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。