[skill] add skill-journal & tilelang-skill-review feedback loop#998
[skill] add skill-journal & tilelang-skill-review feedback loop#998hedi515 wants to merge 8 commits into
Conversation
hedi515
commented
May 11, 2026
hedi515
commented
- skill-journal: schema for collecting per-op skill feedback entries
- tilelang-skill-review: aggregates pending entries and applies approved changes
- op-generate: add section 6 to emit feedback into skill-journal after each run
- skill-journal: schema for collecting per-op skill feedback entries - tilelang-skill-review: aggregates pending entries and applies approved changes - op-generate: add section 6 to emit feedback into skill-journal after each run
|
👋 Hi! Thank you for contributing to the TileLang project. Please remember to run We appreciate you taking this step! Our team will review your contribution, and we look forward to your awesome work! 🚀 |
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request introduces a 'Skill Journal' system for capturing feedback during operator development, adding a 'Skill Review' skill and updating the 'Operator Generation' skill to mandate feedback collection. The review feedback identifies an accidental deletion of essential development steps and coding standards in the operator generation skill. Suggestions were also provided to improve the feedback coverage requirement by adding a 'no_issue' type and to enhance the readability of the review table by using HTML line breaks for evidence entries.
| --- | ||
|
|
||
| ## 4. 关键编码规范 |
There was a problem hiding this comment.
此处删除了大量核心步骤（步骤 5-7）以及 GEMM/Autotune 的关键编码规范。这似乎是在添加新章节时的误操作，这些内容对于指导算子开发至关重要，建议恢复。
| | # | 检查项 | 必须通过 | | ||
| |---|--------|---------| | ||
| | 1 | `skills_consulted` 包含本次查阅的所有 skill | ✅ | | ||
| | 2 | 至少 50% 的 `skills_consulted` 在 entries 中至少出现一次（避免只反思 op-generate 自己）| ✅ | |
There was a problem hiding this comment.
“至少 50% 的反馈覆盖率”要求可能过于严格。如果查阅的 Skill 质量很高且确实没有发现问题，强制要求 50% 的覆盖率可能会导致 Agent 产生虚假反馈。建议在 skill-journal/README.md 的类型词表中增加一个如 no_issue 的类型，以便明确记录“已查阅且无问题”的情况，从而在不产生噪音的前提下满足此项自检。
| | `unclear_workflow` | 工作流步骤模糊或漏检查 | 没说"先搜 examples/" 的强制顺序 | | ||
| | `mode_misjudgment` | 编程模式选型描述误导 | 把混合算子说成可用 Developer 单模式 | | ||
| | `pass_config_gap` | pass_configs 配置说明不全 | 没提 `AUTO_CV_SYNC` 必须配 `AUTO_CV_COMBINE` | | ||
| | `other` | 不属于以上 | | |
There was a problem hiding this comment.
建议增加一个用于记录“查阅后确认无误”的类型，以支持 op-generate 中的反馈覆盖率自检要求，避免在没有问题时被迫填写错误分类。
| | `other` | 不属于以上 | | | |
| | `no_issue` | 查阅后确认无误 | 用于满足反馈覆盖率要求，记录正面反馈 | | |
| | `other` | 不属于以上 | | |
| - 频次 = entry 数量 | ||
| - 严重度 = 取最高（high > medium > low） | ||
| - 来源 = 同桶内含 developer 来源就标 `👤+🤖`（混合）；纯 developer 标 `👤`；纯 agent 标 `🤖` | ||
| - 证据 = 合并所有 evidence（用 `; ` 分隔，仅保留前 3 条） |
There was a problem hiding this comment.
由于 evidence 字段通常包含多行文本（如报错堆栈或代码片段），在 Markdown 表格中使用分号 ; 进行合并会导致显示格式混乱。建议改用 <br> 标签进行分隔，以便在表格中保持可读性。
| - 证据 = 合并所有 evidence（用 `; ` 分隔，仅保留前 3 条） | |
| - 证据 = 合并所有 evidence（用 <br> 分隔，仅保留前 3 条） |
hedi515
force-pushed
the
ascendc_pto
branch
3 times, most recently
from
576c707 to
cd8d3a8
Compare
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… pass Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…st shapes Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2 participants
