dsh-skill-lazy

Lazy skill catalog loading for DeepSeek Harness (DSH). Save tokens on every session by injecting only skill names + one-line summaries instead of the full catalog, and look up full descriptions on demand via skill_search.

Why

By default, DSH injects the full skill catalog — the name plus up to a 500-character description of every installed skill — into the context of every new session. With dozens of skills installed this costs roughly 3K tokens of persistent context per session, even though the model rarely touches most of those skills.

dsh-skill-lazy replaces that full dump with a compact catalog: just the skill name and a one-line summary (≤ 40 characters by default). Full descriptions stay one skill_search call away and are only fetched when actually needed.

How it works

First pass: compact catalog injection

• On every agent pre-step, the plugin rewrites the official skill-catalog message into a <system-reminder> block where each skill appears as `name`: one-line summary.
• Entries are sorted by call frequency (most-used first), falling back to alphabetical order for ties.
• The underlying source.entries of the catalog message is left untouched, so the harness's own digest/deduplication logic is unaffected.

On demand: skill_search

• Registers a skill_search tool that queries the full catalog by keywords and returns hits grouped by MECE domains (mutually exclusive groups, collectively exhaustive coverage), each with its full description, usage count, and a domain-coverage report that flags empty or weak domains.
• Query scoring covers exact name matches, description substring matches, tokenized keywords, and Chinese 2-gram overlap.

Usage tracking

• A tools/post-execute hook records every skill tool invocation into usage.json (runtime data — see .gitignore). The accumulated counts drive catalog ordering.
• An optional proactive match hint (top-3 skills with relevance ≥ threshold) can be injected right after a user message.

skill_mece_check

• A skill_mece_check tool runs before creating or evaluating a skill: it checks mutual exclusivity (keyword overlap against every existing skill, with overlap ratio) and collective exhaustiveness (domain coverage matrix with empty/weak domain flags), then recommends whether to create a new skill or extend an existing one.

cost_audit

• A cost_audit token-cost watchdog (added 2026-08-21) tallies the character/token cost of every plugin tool description and every skill directory description, compares against budgets (tool descriptions default 1500 tokens; skill directory default 800 tokens), and flags anything over budget for slimming or collapsing.
• Single tool descriptions longer than 700 characters are marked for slimming — the detail belongs in the skill's own docs, not in the injected description.

Installation

dsh plugin --profile web add github:boomzikazita/dsh-skill-lazy

The plugin mounts via its bundle patch (cordis.patch.yml), which adds the skill-lazy bundle to the profile's dsh.profile.bundles.

Configuration

Option	Default	Description
catalogDescriptionMaxLength	40	Max characters of the one-line summary per skill in the compact catalog.
topK	6	Max hits per domain group returned by skill_search (clamped to 1–20).
sortByUsage	true	Sort catalog entries by recorded call frequency (fallback: alphabetical).
matchThreshold	6	Minimum relevance score for the proactive match hint.

Example bundle patch with custom config:

- insert:
    - id: skill-lazy
      name: 'dsh-skill-lazy'
      config:
        catalogDescriptionMaxLength: 40
        topK: 6

Requirements

• DeepSeek Harness with @deepseek-ai/dsh-tools ≥ 0.1.0-rc.6 (peer dependency — it is not bundled).

License

MIT © 2026 boomzikazita. See LICENSE.