Foreword

The philosophy of DSH is that everything is a plugin. Once too many skills are installed, the agent is faced with a constantly growing directory. It can see the name and description of each skill, but the common behavior is “over-dumping”: loading too many skills, loading the wrong skills, or missing the workflow skill that could string the entire task together.

Here lies an asymmetric cost: once a wrong skill is loaded, the skill’s content remains in the session history indefinitely; while missing a skill only costs a single tool call. Since missing is cheaper than wrong, the reasonable strategy is to load less. skill-bartender is implemented according to this philosophy—it does not provide specific capabilities but is solely responsible for deciding “which skill this task should load, or whether to load none at all.”

What is it

skill-bartender is a task-to-skill meta-skill for DeepSeek Harness. Maintained by MJorgin, licensed under MIT, current version 0.1.2, and categorized as “Workflow” in the community directory. One-sentence positioning: load the fewest skills according to the Lazy Staircase (usually one, or zero if pure tools suffice), prioritizing workflow skills; missing skills are installed only after isolated scanning and manual approval, never automatically.

Core Features

Lazy Staircase

For each task, walk up the staircase from the bottom, stopping at the first solid rung:

  1. Pure tools are enough → load 0 skills;
  2. A match is found → load 1;
  3. Workflow skills take precedence over manually assembling multiple atomic skills;
  4. Uncertain → don’t load (better to miss than over-dump).

Routing Table

references/policy.md is a user-editable task→skill mapping that overrides default behavior. It routes by URL path patterns for skill families such as doc / drive / wiki / sheets / base / slides. Pasting this table into AGENTS.md enables always-on routing, where every task goes through the staircase.

Safe Cellar

When a skill is needed but not available locally, the process is fixed:

  1. Search for candidate skills;
  2. Download to an isolated directory, not directly into the skill root;
  3. Scan with SkillSpector;
  4. Script content is presented for human review, with a default rejection;
  5. Copy into the skill directory only after explicit consent.

The source, commit hash, and scan conclusion are recorded for every installation. Manual approval is mandatory for every installation; silent installation is never allowed. Additionally, any scripts/ inside a skill is code and will never be executed without human review.

Learning (DSH)

On DSH, loaded but unused skills are recorded; the next time a task of the same type appears, it is skipped directly. Skills that are unused for a long time are proposed for removal.

Taste Test (On-demand)

List installed skills upon request, rewrite weak descriptions into “When to use” trigger phrases, and limit to a 500-character directory cap.

Cross-platform

The same SKILL.md can be installed to DSH, Claude Code, and Codex within approximately 15 seconds.

Installation & Activation

Install via DSH bundle:

dsh plugin --profile web add github:akqwpeter-prog/skill-bartender

A note needs to be made: the installation command and the CI badge in the README point to akqwpeter-prog/skill-bartender, while the directory page and README provide the repository address and maintainer as MJorgin/skill-bartender; these are currently inconsistent, and it is unclear which is the current official repository. Please verify the source code yourself before installing.

You can also install manually to DSH:

mkdir -p ~/.dsh/skills/skill-bartender
cp skills/skill-bartender/SKILL.md ~/.dsh/skills/skill-bartender/
cp -r skills/skill-bartender/references ~/.dsh/skills/skill-bartender/

The Safe Cellar process relies on SkillSpector. If it is not installed on your machine:

uv tool install git+https://github.com/NVIDIA/skillspector.git

Or check the manual checklist item by item in references/policy.md.

To enable: say “skill-bartender” once in any session, or paste the routing table into AGENTS.md. Note that skill-bartender itself never self-triggers and never preloads; it requires a user gesture or a task match to load once.

Typical Usage

Four usage patterns, corresponding to different scenarios:

  1. Directly name it: Say “skill-bartender” in any session; suitable for one-time use or initial setup;
  2. Always-on routing: Paste the routing table from references/policy.md into AGENTS.md; after this, every task goes through the staircase;
  3. Request pairing: Ask “Which skill fits this task?” to choose among multiple skills;
  4. Cellar audit: Say “Audit my installed skills” to perform the Taste Test, rewriting weak descriptions into trigger phrases.

Applicable Scenarios & Notes

Who is it for: DSH users who have already installed a batch of skills and want the agent to load less rather than more; teams who want the installation process for missing skills to have discipline such as isolation, scanning, and manual approval; people who want to maintain explicit task→skill rules in AGENTS.md.

Relationship with the ecosystem: DshMarket, dsh-find-plugin, and dsh-plugin-autoevo are responsible for finding, searching, and auto-installing plugins, while skill-bartender fills the gap in routing strategy (staircase plus routing table) and the “isolate then approve” installation discipline. They are complementary, not replacements.

Precautions before use:

  • Skills are essentially instructions, and instructions may be adversarial (prompt injection). Passing a SkillSpector scan is only a filter, not a guarantee; the design of showing scripts for review with a default rejection is for this purpose.
  • The self-scan results of skill-bartender are SkillSpector 0 findings (score 0 / SAFE); the report is available in the docs/skillspector-report. file in the repository.
  • Plugins run with the permissions of the current dsh process. You should check the source code and license before installing (this project is MIT).

Summary

skill-bartender solves a specific problem: as the skill directory grows, the decision-making process for loading itself also requires discipline. The Lazy Staircase controls “how much,” the Routing Table controls “which,” and the Safe Cellar controls “how new skills enter the cellar.” If your agent is being dragged down by a cabinet full of skills, it is worth trying.

  • Directory Page: https://www.skillhub.cn/plugins/MJorgin/skill-bartender
  • GitHub: https://github.com/MJorgin/skill-bartender