dsh-skill-on-demand

English | 中文

DeepSeek Harness publishes the session skill catalog once per session, and every later request carries it again as part of the conversation. With dozens of installed skills that is a few thousand tokens of description text re-sent on every turn, most of it about skills the current task will never touch.

This plugin removes that catalog and replaces it with three small tools.

What it does

Before	With this plugin
One message lists every skill name and its description, sent in every request	The catalog message is never published
The model reads all descriptions to decide what it needs	skill_list returns skill names only; descriptions are fetched only for a search (skill_list with query)
Nothing records what the task actually uses	skill_pick records the skills kept for this session; everything else is ignored
Every skill body is reachable but unmeasured	skill loads exactly one full skill body, by name

The plugin injects one short protocol message per session: list names, judge which skills the task needs (five or fewer), record them with skill_pick, then continue. It never repeats, and subagent sessions get none — they keep the tools, they just do not pay for the prompt.

How the catalog is removed

@deepseek-ai/dsh-tool-skill publishes its catalog only while ctx.tools.get('skill', agent) still resolves to its own tool registration. This plugin registers a same-name skill tool in each agent's own scope (the nearest scope wins), so that check fails and the catalog stops being published. The replacement tool does the same job — load one skill's instructions by exact name.

Presets that never mounted a skill tool are left untouched: there is nothing to replace, and a deliberately minimal preset should not silently gain tools.

Install

dsh plugin --profile web add dsh-skill-on-demand

Restart the harness afterwards; the bundle is loaded at boot.

Verify

Start a new session and ask anything. The first turn should show skill_list → skill_pick and no <available_skills> block. skill_list with no argument returns names only.

Limitations

• Soft, not enforced. If the model ignores the protocol, nothing breaks and nothing is saved — the catalog is still gone, so the saving holds either way.
• Old sessions keep their history. A session that already had a catalog published before the plugin was installed still carries that message, plus one "no skills available" correction from the shipped plugin. Start a fresh session for a clean slate.
• Selection is per session. A new session re-runs the judgement; that is intentional, since the right skills depend on the task.
• Duplicates. If a preset already registered skill in the agent's own layer, registration fails and the plugin degrades to "picker tools available, catalog unchanged" instead of breaking the session.

Development

index.js is a single Host plugin, no build step. scripts/selftest.mjs runs it against a mock context. Run it from inside the installed package directory, so module resolution still reaches the harness's own @deepseek-ai/* packages:

node <profile>/node_modules/dsh-skill-on-demand/scripts/selftest.mjs

License

MIT