feature: measure the payoff — agent pass rate with vs without the pack, graded by the existing headless harness

未關閉
#373 1 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視

還沒有人認領這個 Issue。

評估

難度
5/5
預估耗時
一週以上
新手友好度
25/100
Issue 類型
功能
描述清晰度
描述清楚
活躍度
活躍
技術堆疊
blender, python

研究方向

Start by reading the existing headless grader at tests/smoke/run_example.py and the issue’s listed Blender API examples and CI versions. Then define reproducible evaluation tasks, run them with and without the plugin across versions and trials, and publish results on the landing page and README. Done means the reported pass-rate comparison includes task count, model, and date and can be reproduced from the repo.

由索引模型根據 Issue 內容生成。

描述

feature marketing priority: high stale testing

Value

This would be the project's biggest differentiator. Nothing in the README, site or ROADMAP measures whether the pack changes agent output. The landing page claims skills cover "the mistakes AI assistants actually make", with no data. A published number turns a content pack into evidence, and answers the skeptical visitor's first question.

Why this repo is unusually well placed

It already has what most eval efforts lack:

  • deterministic headless graders (tests/smoke/run_example.py, exit-code contracts, falsifiers);
  • 4.5, 5.1 and 5.2 Blender CI;
  • a catalog of known-hard API contracts (slotted actions, temp_override, GN interface, VSE .strips, export axes).

Claude Code also ships claude plugin eval (plugin eval suites).

Suggested approach

  1. Write about 20 task prompts with graders. Examples:
    • "keyframe this object's rotation using 5.2 slotted actions"
    • "export to Unreal-ready glTF"
    • "bevel these edges with bmesh and free it"
    • "add a COLOR strip to the VSE on 5.2"
  2. Each run: the agent writes a script, it runs headless against a checker, and you record pass or fail.
  3. Run with and without the plugin, per Blender version, for a named model, N trials each.
  4. Publish a table on the landing page and README, and refresh it on the weekly cron (or manually, for cost).
  5. Use the results to prioritise skill improvements, e.g. the skill trigger-description issue.

Done when

The landing page shows "agent pass rate X% → Y% with the pack (N tasks, model, date)", and it is reproducible from the repo.

主要語言
Python
星號
17
分支
3
平均合併
23 小時 48 分鐘
30 天內合併 PR
3

環境準備

從這裡開始

  1. 先讀完整個 Issue,再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫,在一個分支上完成修改。
  4. 送出 Pull Request,並在描述裡引用這個 Issue 編號。

TMHSDigital/Blender-Developer-Tools 的其他 Issue

查看 TMHSDigital/Blender-Developer-Tools 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。