Home / anionex / agent-vision-toolkit · GitHub

DESIGN.md@research/task-evalA

DESIGN.md@research/task-eval from anionex/agent-vision-toolkit is a DESIGN.md file: the design rules coding agents follow in that repo.

Install it, pinned
mdr add anionex/agent-vision-toolkit/research/task-eval/DESIGN.md@git:20260818.025eeac

Pins git:20260818.025eeac in mdr.lock. Pin the fingerprint instead.

Audit grade
A, 14 of 14 checks passed
Kind
DESIGN.md file
Repository
anionex/agent-vision-toolkit
Path
research/task-eval/DESIGN.md
Versions recorded
5
Latest version
git:20260818.025eeac, committed Aug 18, 2026
First version recorded
Aug 4, 2026
Size
13,527 bytes
Repository stars
1,243
License
MIT

Collected from public GitHub. Every version is saved exactly as it was and never changes. Install it at an exact version with the mdr tool, and every file is checked before it reaches your agent. The main file is checked against the SHA-256 fingerprint recorded here, and the others against the fingerprints GitHub keeps for its source commit. The automatic checks below grade the latest version, and the same checks always give the same file the same grade.

What the file says

# 任务级评测 — 设计与预注册

2026-08-04 · 在产出任何数据之前写完并提交。预测写在下面，跑完不改。

机制层的结果见 `research/2026-08-04-focus-hint-eval.md`：230 道单图问答，
五个条件，全部 p ≥ 0.227，零结果。那一层的结论**不外推到本层**，本层也不复用它的题库。

## 为什么换一层测

机制层测的是「同一张图、同一个问题下，换 prompt 能让描述好多少」。它有一个内建的
让步：**每道题都自带一个明确的问题**，问题本身就替代了 hint 和转写条款的一部分
作用——描述侧被告知要找什么之前，答题侧已经知道了。

真实使用不长这样。用户粘一张 CI 失败截图说「修一下」，没人说「请注意第 87 行的
Expected: 4500」。**agent 必须自己决定看什么、看几次、什么时候承认没看清。**
本层测的就是这个，且它能测到机制层测不到的三件事：

1. 描述漏掉的东西，agent 会不会**察觉**并回查（决策日志 #2）。
2. 逐字转写条款防的**静默遗漏**，在没有明确问题时是否真的发生（决策日志 #4）。
3. 任务成败本身。

## 五个条件

Codex 用 `codex exec` 非交互驱动，每个条件一个独立 `CODEX_HOME`
（`~/cdv-task-eval/homes/<condition>/`），**生产配置 `~/.codex/config.toml` 全程只读，
不修改**。除代理端点与 skill 外，配置逐字复制生产设置（模型、reasoning effort、
`disable_response_storage`），因为要测的是用户真实在用的系统。代理一律以默认
flag 启动（`--codex-header-compat` 与 `--inject-reasoning-summary` 保持关闭）。
密钥经 `env_key` 由环境传入，不落盘。

| 条件 | Codex 指向 | 隔离出的变量 |
|---|---|---|
| `blind` | 代理变体，图片一律替换为固定占位文字 | 视觉通道整体的价值（下限） |
| `naive` | 代理变体，`DEFAULT_PROMPT`（"describe this image in detail"） | 普通转接方案 |
| `hint` | 生产代理（`_ROLE_PROMPT` + hint + 转写条款） | **线上实际行为** |
| `hint_tools` | 生产代理 + 已发布的 `vision-skills` skill | 回查工具的价值 |
| `native` | Codex 直连原生多模态模型 | 文字中转的代价（上限） |

四个判断：
…

Read the whole file at its exact version.

How to install

The exact contents, by fingerprint
mdr add anionex/agent-vision-toolkit/research/task-eval/DESIGN.md@sha256:6f41edd229187cb5

The line at the top of this page pins the label git:20260818.025eeac, which is what most people want. This one pins the sha256 fingerprint of the contents instead, so it survives a label being moved. Either way the exact version is written to mdr.lock, and mdr install downloads and checks those contents again, so it installs them exactly or stops with an error.

Badge

mdr badge

[![mdr](https://markdownregistry.com/badge/art_keyvs7wfxksmu5g7.svg)](https://markdownregistry.com/a/art_keyvs7wfxksmu5g7)

2 badge image fetches in 30 days. What the badge shows.

Versions

The registry has recorded 5 versions of this file, between Aug 4, 2026 and Aug 18, 2026, about one every 3 days.

versioncommittedcommitsizeaudit
git:20260818.025eeac latest2026-08-18 025eeac 13,527 BA view · diff
git:20260805.c3ce9042026-08-05 c3ce904 13,526 BA view · diff
git:20260804.43e68c52026-08-04 43e68c5 13,532 BA view · diff
git:20260804.4bf5d7a2026-08-04 4bf5d7a 11,476 BA view · diff
git:20260804.852ee112026-08-04 852ee11 9,574 BA view

Audit of the latest version

A  14 of 14 checks passed. Fixed automatic checks with no AI involved, so the same checks give the same answer every run.
  • pass: Size between 200 bytes and 200 KB (13527 bytes)
  • pass: No zero-width or bidi control characters
  • pass: No instruction hidden inside an HTML comment
  • pass: No link to an exfiltration or paste host
  • pass: No credential-shaped string
  • pass: No instruction to send local credentials anywhere
  • pass: No text hidden with inline styles
  • pass: No prompt-injection phrasing
  • pass: No curl or wget piped into a shell
  • pass: No recursive delete of root, home or parent
  • pass: No instruction to read or print local credentials
  • pass: No base64 blob over 200 characters
  • pass: No link to a raw IP address
  • pass: No script tag

Source

GitHub

anionex/agent-vision-toolkit · 1,243 stars · license MIT · pushed 2026-10-13 · branch main

API

GET https://markdownregistry.com/api/v1/artifacts/art_keyvs7wfxksmu5g7
GET https://markdownregistry.com/api/v1/resolve?ref=anionex/agent-vision-toolkit/research/task-eval/DESIGN.md
GET https://markdownregistry.com/api/v1/blob/6f41edd229187cb544330984174f546a8a7f3540356f19e4829148a25b8755b1

Your AI assistant does the legwork. You hear about the deals worth your word. Hand yours the standing instructions at modelranch.com and it joins the network that reads files like this one.

Guides: what DESIGN.md is and how agents use it · which agent file does what.

More from anionex/agent-vision-toolkit

DESIGN.md@research/hint-eval DESIGN.md file
anionex/agent-vision-toolkit · research/hint-eval/DESIGN.md
git:20260805.c3ce904 · audit A · 1,243 stars
vision-skills SKILL.md skill
anionex/agent-vision-toolkit · skills/vision-skills/SKILL.md · Local vision CLIs: glance (describe/ask/OCR an image), ground (locate a target, pixel box), detect (element inventory)…
git:20261002.888e689 · audit A · 1,243 stars

Every file in anionex/agent-vision-toolkit

Browse by kind, by grade A, or by owner.