fix(skills): 禁止自动注入完整技能内容 (#49)

## Summary - 将 dashscope-image-gen、image-to-image、markdown、tech-diagram、xiaomi-tts 改为仅按需加载，并递增 patch 版本 - 在 Market validator 与 JSON Schema 中禁止 `disable-model-invocation: false` - 将回归测试接入 `i18n Validate`，同步中英文技能编写规范 ## Root cause and impact 这 5 个技能在首次引入时即声明 `disable-model-invocation: false`，导致完整技能正文进入普通请求的 system prompt。修改后市场技能只能声明 `true` 或省略该字段，完整内容仅在显式调用 Skill 工具后加载。 ## Validation - `uv run --quiet scripts/i18n/test_validate_i18n.py` - `uv run --quiet scripts/i18n/validate-i18n.py` - `uv run --quiet scripts/i18n/translate.py --check` - `python3 -m json.tool scripts/i18n/schema/skill-frontmatter.schema.json` - `actionlint .github/workflows/i18n-validate.yml` - `git diff --check`
2026-07-23 02:03:50 +08:00 · 2026-07-14 11:55:22 +08:00
parent 5ab2994664
commit 73c94ab70e
15 changed files with 110 additions and 31 deletions
--- a/.github/workflows/i18n-validate.yml
+++ b/.github/workflows/i18n-validate.yml
@@ -58,6 +58,10 @@ jobs:
        if: steps.changes.outputs.relevant == 'true'
        uses: astral-sh/setup-uv@v3

+      - name: Run validator unit tests
+        if: steps.changes.outputs.relevant == 'true'
+        run: uv run --quiet scripts/i18n/test_validate_i18n.py
+
      - name: Validate i18n
        if: steps.changes.outputs.relevant == 'true'
        run: uv run --quiet scripts/i18n/validate-i18n.py
--- a/docs/I18N.md
+++ b/docs/I18N.md
@@ -35,6 +35,7 @@ description: >-                     # 单字段，Claude 检索用；建议含
  Use this skill when the user wants to ... 用户提到 ...、...、...
 license: MIT
 version: 1.0.0
+disable-model-invocation: true       # 必须为 true 或省略；禁止自动注入技能全文
 metadata:
  author: your-org
  updated_at: '2026-05-03'
@@ -69,6 +70,9 @@ market:

 完整 JSON Schema：[`scripts/i18n/schema/skill-frontmatter.schema.json`](../scripts/i18n/schema/skill-frontmatter.schema.json)

+市场技能只允许按需加载完整指令。`disable-model-invocation` 必须为 `true` 或省略；提交
+`false` 会被 `i18n Validate` CI 拒绝。
+
 ---

 ## 提交流程
--- a/scripts/i18n/schema/skill-frontmatter.schema.json
+++ b/scripts/i18n/schema/skill-frontmatter.schema.json
@@ -48,7 +48,10 @@
      "enum": ["enabled", "disabled", "deprecated", "experimental"]
    },
    "disable-model-invocation": {
-      "type": "boolean"
+      "description": "Must be true or omitted. Market skills may only load their full instructions on demand; automatic full-content injection is prohibited.",
+      "type": "boolean",
+      "const": true,
+      "default": true
    },
    "tags": {
      "type": "array",
--- a/scripts/i18n/test_validate_i18n.py
+++ b/scripts/i18n/test_validate_i18n.py
@@ -0,0 +1,52 @@
+#!/usr/bin/env -S uv run --script
+# /// script
+# requires-python = ">=3.10"
+# dependencies = ["pyyaml>=6.0"]
+# ///
+"""Unit tests for market validation policies."""
+
+from __future__ import annotations
+
+import importlib.util
+import sys
+import unittest
+from pathlib import Path
+
+
+VALIDATOR_PATH = Path(__file__).with_name("validate-i18n.py")
+SPEC = importlib.util.spec_from_file_location("market_validate_i18n", VALIDATOR_PATH)
+assert SPEC is not None and SPEC.loader is not None
+VALIDATOR = importlib.util.module_from_spec(SPEC)
+sys.modules[SPEC.name] = VALIDATOR
+SPEC.loader.exec_module(VALIDATOR)
+
+
+class ModelInvocationPolicyTests(unittest.TestCase):
+    def validate(self, frontmatter: dict[str, object]) -> list[object]:
+        report = VALIDATOR.Report()
+        VALIDATOR.validate_model_invocation_policy(
+            frontmatter,
+            "skills/example/SKILL.md",
+            report,
+        )
+        return report.issues
+
+    def test_allows_true(self) -> None:
+        self.assertEqual([], self.validate({"disable-model-invocation": True}))
+
+    def test_allows_omitted_field(self) -> None:
+        self.assertEqual([], self.validate({}))
+
+    def test_rejects_false(self) -> None:
+        issues = self.validate({"disable-model-invocation": False})
+        self.assertEqual(1, len(issues))
+        self.assertEqual("model-invocation-policy", issues[0].rule)
+
+    def test_rejects_non_boolean_value(self) -> None:
+        issues = self.validate({"disable-model-invocation": "true"})
+        self.assertEqual(1, len(issues))
+        self.assertEqual("model-invocation-policy", issues[0].rule)
+
+
+if __name__ == "__main__":
+    unittest.main()
--- a/scripts/i18n/validate-i18n.py
+++ b/scripts/i18n/validate-i18n.py
@@ -18,6 +18,7 @@ Checks:
  9. Skill/Agent counts and builtin skill index match the repository contents.
  10. Skill, Agent, and entry.json category references exist in categories.json.
  11. entry.json pointers have the required marketplace fields and safe source URLs.
+  12. Market Skills set `disable-model-invocation` to true or omit it; false is prohibited.

 Exit codes:
  0 = pass
@@ -108,6 +109,21 @@ def heading_count(text: str) -> int:
    return len(HEADING_PATTERN.findall(text or ""))


+def validate_model_invocation_policy(
+    frontmatter: dict[str, Any],
+    path: str,
+    report: Report,
+) -> None:
+    """Reject market skills that request full-content system-prompt injection."""
+    value = frontmatter.get("disable-model-invocation")
+    if value is not None and value is not True:
+        report.add(Issue(
+            path,
+            "model-invocation-policy",
+            "disable-model-invocation must be true or omitted; automatic full-content injection is prohibited",
+        ))
+
+
 def validate_skill(
    skill_dir: Path,
    report: Report,
@@ -128,6 +144,8 @@ def validate_skill(
        return
    assert fm is not None and body is not None

+    validate_model_invocation_policy(fm, f"{rel_dir}/SKILL.md", report)
+
    name = fm.get("name", "")
    description = fm.get("description", "")
    market = fm.get("market") or {}
--- a/skills/dashscope-image-gen/SKILL.md
+++ b/skills/dashscope-image-gen/SKILL.md
@@ -8,11 +8,11 @@ description: >-
  Use when the user mentions: generate image, draw, text-to-image, create image,
  AI painting, illustration, design picture.
 license: Complete terms in LICENSE.txt
-version: 1.4.1
+version: 1.4.2
 type: procedural
 risk_level: low
 status: enabled
-disable-model-invocation: false
+disable-model-invocation: true
 provider: auto
 tags:
  - media
@@ -25,7 +25,7 @@ requires:
    - Bash
 metadata:
  author: desirecore
-  updated_at: '2026-06-15'
+  updated_at: '2026-07-14'
  i18n:
    default_locale: en-US
    source_locale: zh-CN
--- a/skills/image-to-image/SKILL.md
+++ b/skills/image-to-image/SKILL.md
@@ -8,11 +8,11 @@ description: >-
  换背景、换风格、图片编辑、以图生图、参考图、基于这张图、
  把这张图改成、在这张图上、image edit、img2img。
 license: Complete terms in LICENSE.txt
-version: 1.0.1
+version: 1.0.2
 type: procedural
 risk_level: low
 status: enabled
-disable-model-invocation: false
+disable-model-invocation: true
 provider: auto
 tags:
  - media
@@ -24,7 +24,7 @@ requires:
    - Bash
 metadata:
  author: desirecore
-  updated_at: '2026-06-12'
+  updated_at: '2026-07-14'
  i18n:
    default_locale: en-US
    source_locale: zh-CN
--- a/skills/manage-skills/SKILL.md
+++ b/skills/manage-skills/SKILL.md
@@ -4,7 +4,7 @@ description: >-
  管理 Agent 的技能生命周期：通过 HTTP API 导入、安装、更新、删除技能，
  或通过 AgentFS 文件系统直接编写符合规范的 SKILL.md。Use when 用户要求
  安装技能、从 URL/Git 导入技能、编写新技能、或管理已有技能。
-version: 1.0.5
+version: 1.0.6
 type: meta
 risk_level: low
 status: enabled
@@ -17,7 +17,7 @@ tags:
  - agentfs
 metadata:
  author: desirecore
-  updated_at: '2026-05-05'
+  updated_at: '2026-07-14'
  i18n:
    default_locale: en-US
    source_locale: zh-CN
@@ -38,7 +38,7 @@ metadata:
      description: >-
        Manage the Skill lifecycle of an Agent: import, install, update, and delete Skills via HTTP API, or directly author standards-compliant SKILL.md files via the AgentFS filesystem. Use when the user requests to install Skills, import Skills from URL/Git, author new Skills, or manage existing Skills.
      body: ./SKILL.md
-      source_hash: sha256:eb098cb8c9bdf482
+      source_hash: sha256:57d9f0d8c351f884
      translated_by: human
 market:
  icon: >-
@@ -507,7 +507,7 @@ covering completed items, to-dos, and important decisions.
 | `risk_level`               | Recommended     | enum     | `low` / `medium` / `high`                          |
 | `status`                   | Recommended     | enum     | `enabled` / `disabled`                             |
 | `tags`                     | Optional     | string[] | List of tags, used for search and categorization                           |
-| `disable-model-invocation` | Optional     | boolean  | `false` = opt-in auto-injection of full content into the system prompt; `true` (or omitted) = no auto-injection, only loaded when explicitly invoked via the Skill tool; default `true` |
+| `disable-model-invocation` | Optional     | boolean  | Must be `true` or omitted. Full-content auto-injection is prohibited; skills are loaded only when explicitly invoked via the Skill tool. |
 | `requires`                 | Optional     | object   | Dependency declaration: `tools`, `optional_tools`, `connections` |
 | `metadata`                 | Optional     | object   | Meta information: `author`, `updated_at`                     |
 | `market`                   | Optional     | object   | Market display metadata (only required for Skills published to the Market)             |
--- a/skills/manage-skills/SKILL.zh-CN.md
+++ b/skills/manage-skills/SKILL.zh-CN.md
@@ -447,7 +447,7 @@ metadata:
 | `risk_level`               | 推荐     | enum     | `low` / `medium` / `high`                          |
 | `status`                   | 推荐     | enum     | `enabled` / `disabled`                             |
 | `tags`                     | 可选     | string[] | 标签列表，用于搜索和分类                           |
-| `disable-model-invocation` | 可选     | boolean  | `false`=opt-in 自动注入完整内容到 system prompt；`true` 或缺省=不自动注入，仅显式 Skill 工具调用时才加载；默认 `true` |
+| `disable-model-invocation` | 可选     | boolean  | 必须为 `true` 或省略。禁止把技能全文自动注入 system prompt；仅在显式调用 Skill 工具时加载。 |
 | `requires`                 | 可选     | object   | 依赖声明：`tools`、`optional_tools`、`connections` |
 | `metadata`                 | 可选     | object   | 元信息：`author`、`updated_at`                     |
 | `market`                   | 可选     | object   | 市场展示元数据（仅市场发布的技能需要）             |
--- a/skills/markdown/SKILL.md
+++ b/skills/markdown/SKILL.md
@@ -16,11 +16,11 @@ description: >-
  月报、邮件、对比分析、会议记录、散文、文章、作文、md、markdown、report、
  summary、plan、outline、memo、table、list、draft、document、notes、
  write、create、organize。
-version: 1.0.0
+version: 1.0.1
 type: procedural
 risk_level: low
 status: enabled
-disable-model-invocation: false
+disable-model-invocation: true
 tags:
  - markdown
  - md
@@ -28,7 +28,7 @@ tags:
  - text
 metadata:
  author: desirecore
-  updated_at: '2026-05-11'
+  updated_at: '2026-07-14'
  i18n:
    default_locale: en-US
    source_locale: zh-CN
--- a/skills/skill-creator/SKILL.md
+++ b/skills/skill-creator/SKILL.md
@@ -5,7 +5,7 @@ description: >-
  （frontmatter 元数据 + L0/L1/L2 分层内容 + 脚本/参考/资产）和 Claude Code
  基础格式。Use when 用户要求创建新技能、更新已有技能、或将经验封装为可复用
  的技能包。
-version: 1.0.4
+version: 1.0.5
 type: meta
 risk_level: low
 status: enabled
@@ -18,7 +18,7 @@ tags:
  - authoring
 metadata:
  author: desirecore
-  updated_at: '2026-05-05'
+  updated_at: '2026-07-14'
  i18n:
    default_locale: en-US
    source_locale: zh-CN
@@ -39,7 +39,7 @@ metadata:
      description: >-
        Guides users to create and edit standards-compliant SKILL.md skill packages. Supports the DesireCore full format (frontmatter metadata + L0/L1/L2 layered content + scripts/references/assets) and the Claude Code basic format. Use when the user requests to create a new Skill, update an existing Skill, or package experience into a reusable Skill bundle.
      body: ./SKILL.md
-      source_hash: sha256:e14a6879bb800455
+      source_hash: sha256:f6aa877d100a85e1
      translated_by: human
 market:
  icon: >-
@@ -160,7 +160,7 @@ A SKILL.md consists of two parts: **Frontmatter (YAML metadata)** and **Body (Ma

 | Field | Type | Default | Description |
 |------|------|------|------|
-| `disable-model-invocation` | boolean | `true` | `false` = opt-in auto-injection of full SKILL.md content into the system prompt (auto-loaded); `true` (or omitted) = no auto-injection, only loaded when an Agent explicitly invokes the Skill tool (aligned with Claude Skills: disabled by default, opt-in) |
+| `disable-model-invocation` | boolean | `true` | Must be `true` or omitted. Full SKILL.md auto-injection is prohibited; an Agent must explicitly invoke the Skill tool to load the instructions. |
 | `user-invocable` | boolean | `true` | `false` = does not appear in command completion; serves only as background knowledge |
 | `allowed-tools` | string[] | — | Restricts the list of tools available at execution time |
 | `requires` | object | — | Dependency declaration: `tools`, `optional_tools`, `connections` |
--- a/skills/skill-creator/SKILL.zh-CN.md
+++ b/skills/skill-creator/SKILL.zh-CN.md
@@ -100,7 +100,7 @@ SKILL.md 由两部分组成：**Frontmatter（YAML 元数据）** 和 **Body（M

 | 字段 | 类型 | 默认 | 说明 |
 |------|------|------|------|
-| `disable-model-invocation` | boolean | `true` | `false`=opt-in 自动注入完整 SKILL.md 内容到 system prompt（自动加载）；`true` 或缺省=不自动注入，仅当 Agent 显式调用 Skill 工具时才加载（与 Claude Skills 对齐：默认禁用，需显式启用） |
+| `disable-model-invocation` | boolean | `true` | 必须为 `true` 或省略。禁止把完整 SKILL.md 自动注入 system prompt；Agent 必须显式调用 Skill 工具后才加载指令。 |
 | `user-invocable` | boolean | `true` | `false`=不出现在命令补全，仅作为背景知识 |
 | `allowed-tools` | string[] | — | 限制执行时可用的工具列表 |
 | `requires` | object | — | 依赖声明：`tools`、`optional_tools`、`connections` |
--- a/skills/skill-creator/references/desirecore-format.md
+++ b/skills/skill-creator/references/desirecore-format.md
@@ -25,7 +25,7 @@

 | 字段 | 类型 | 默认 | 说明 |
 |------|------|------|------|
-| `disable-model-invocation` | boolean | `true` | `false`=opt-in 自动注入完整 SKILL.md 内容到 system prompt（自动加载）；`true` 或缺省=不自动注入，仅当 Agent 显式调用 Skill 工具时才加载（与 Claude Skills 对齐：默认禁用，需显式启用） |
+| `disable-model-invocation` | boolean | `true` | 必须为 `true` 或省略。禁止把完整 SKILL.md 自动注入 system prompt；Agent 必须显式调用 `Skill` 工具后才加载指令。 |
 | `user-invocable` | boolean | `true` | `false`=不出现在命令补全，仅作为背景知识 |
 | `allowed-tools` | string[] | 全部 | 限制执行时可用的工具列表（如 `["Edit", "Read", "Bash"]`） |
 | `model` | string | 继承 | 覆盖使用的模型 ID（如 `"claude-sonnet-4-20250514"`） |
@@ -127,10 +127,8 @@ json_output:

 L0/L1/L2 是 SKILL.md **内容组织约定**（Claude Skills 风格），帮助作者把摘要、原则、详细规范分层书写。**运行时并不在 L0/L1/L2 之间做"按需加载"**——一旦 skill 被加载，就是整篇 SKILL.md 内容（除 frontmatter）注入 system prompt。

-是否被自动加载由 `disable-model-invocation` 决定：
-
- `disable-model-invocation: false` 时：skill 被加入自动加载列表，整篇内容随 system prompt 注入
- `disable-model-invocation: true` 或缺省时：skill 不自动加载，仅当 Agent 显式调用 `Skill` 工具传入此 skill ID 时才加载
+市场技能不允许把全文自动注入 system prompt。`disable-model-invocation` 必须为 `true`
+或省略；skill 仅在 Agent 显式调用 `Skill` 工具并传入对应 skill ID 后加载。

 ## 完整示例

--- a/skills/tech-diagram/SKILL.md
+++ b/skills/tech-diagram/SKILL.md
@@ -12,11 +12,11 @@ description: >-
  Use when 用户提到 画架构图、架构图、流程图、时序图、序列图、状态图、状态机、
  ER图、类图、思维导图、出图、画图、可视化、画一张图、画个图、图表风格、
  暗色风格图、蓝图风格图、奶油风格图、draw diagram、architecture diagram、flowchart、sequence diagram。
-version: 1.1.0
+version: 1.1.1
 type: procedural
 risk_level: low
 status: enabled
-disable-model-invocation: false
+disable-model-invocation: true
 tags:
  - diagram
  - mermaid
@@ -25,7 +25,7 @@ tags:
  - visualization
 metadata:
  author: desirecore
-  updated_at: '2026-05-31'
+  updated_at: '2026-07-14'
  i18n:
    default_locale: en-US
    source_locale: zh-CN
--- a/skills/xiaomi-tts/SKILL.md
+++ b/skills/xiaomi-tts/SKILL.md
@@ -8,11 +8,11 @@ description: >-
  Use when 用户提到 语音合成、文字转语音、TTS、朗读、读出来、生成语音、
  生成音频、文本转音频、配音、念出来、小米语音、MiMo 语音、小米 TTS。
 license: Complete terms in LICENSE.txt
-version: 1.0.2
+version: 1.0.3
 type: procedural
 risk_level: low
 status: enabled
-disable-model-invocation: false
+disable-model-invocation: true
 provider: xiaomi
 tags:
  - media
@@ -26,7 +26,7 @@ requires:
    - Bash
 metadata:
  author: desirecore
-  updated_at: '2026-05-08'
+  updated_at: '2026-07-14'
  i18n:
    default_locale: en-US
    source_locale: zh-CN