AGENT SKILL

deepseek-skill-doctor

Check Agent Skills for DSH discovery, description routing, host-path portability, and unsafe DeepSeek reasoning or thinking workarounds. Use when reviewing a SKILL.md for DeepSeek Harness, preparing a skill catalog entry, or debugging why DSH does not discover or invoke a skill.

License 许可
MIT
最近更新
2026-10-21
Stars
★ 0
Forks
0

01 前言 Front Matter

name
deepseek-skill-doctor
description
Check Agent Skills for DSH discovery, description routing, host-path portability, and unsafe DeepSeek reasoning or thinking workarounds. Use when reviewing a SKILL.md for DeepSeek Harness, preparing a skill catalog entry, or debugging why DSH does not discover or invoke a skill.

02 资源文件 1

03 SKILL.md 正文

DeepSeek Skill Doctor

Run the existing Agent Skills validators first, then add the DeepSeek/DSH profile. Keep both result sets; they answer different questions.

Check a skill

If skill-validator is installed, run its deterministic checks:

skill-validator check path/to/skill

Then run the DSH profile:

dsskill-doctor check path/to/skill --format json

Inside this repository, use uv run dsskill-doctor. Without an installation, use:

uvx --from git+https://github.com/Whning0513/deepseek-skill-doctor.git \
  dsskill-doctor check path/to/skill --format json

Pass a configured skill root instead of one bundle to check DSH's one-level discovery behavior.

Interpret the result

  1. Fix AGENT_SKILLS_SPEC_INVALID findings before DSH-specific warnings.
  2. Treat DSH_NESTED_SKILL_UNDISCOVERED as a placement error: DSH only scans direct children of a skill root.
  3. Put task and trigger words in description. DSH's model catalog does not expose whenToUse and truncates descriptions at 500 characters by default.
  4. Apply tool restrictions in DSH configuration when a skill uses allowed-tools; do not assume the filesystem provider enforces that field.
  5. Remove any instruction to manufacture reasoning_content. Preserve the provider's exact value through the tool loop.
  6. Investigate request reconstruction before presenting disabled thinking as a generic fix.

The checker reads text files only. It does not execute skill scripts, validate task quality with a model, or make API calls. Report static findings as evidence, not proof that a skill is safe or effective.

04 同仓库更多 Skills