A community fork of zhengjy01/dsh-skill-studio. Not an official DeepSeek AI product. Upstream provides skill browsing, editing, and manual extraction; this fork adds the four things unattended use needs: per-turn judgement, auto-accept, semantic merging, and models that follow DSH.

Features

Leave it running. It decides what this turn taught, which existing skill it belongs to, and whether merging would create duplicates.

• Per-turn worth-check: three seconds after a turn ends, it sends a tiny YES/NO prompt (10 digests from the most recent session). Only YES triggers a full extraction. Judgement is throttled to 90s, extraction to 5 minutes.
• Straight into the library: accepted candidates are written to ~/.agents/skills and load on the next turn. The review queue only fills up when a write fails (name collision and similar).
• Merge instead of piling up files: each candidate is compared against the whole library. Only genuinely new topics create a new skill. Real-world check: the same workflow re-worded scores just 0.07–0.24 on token similarity, so local similarity only pre-filters — the model sees the full catalog and decides.
• Models come from DSH: the panel offers a dropdown of providers/models configured in DSH; leave it empty to follow whichever session triggered the run. The plugin stores no API key and asks for no gateway URL.
• "Recently added" in the panel: the Skill Studio panel lists recently touched skills with name, description, and timestamp.
• Clean writes: duplicated sections are truncated, protocol residue like MERGE: is stripped, and any "merge" that loses a large chunk of the original is rejected.

How merging works

A candidate passes four gates before it lands. Any gate that says no turns it into a new skill (or drops it, per rule).

new candidate
  │
  ├─ ① Tombstone: this name was merged away before (scan of ~/.dsh/skill-archive*/)
  │     → force-merge into the same-category skill with the largest body;
  │       if that fails, drop it — never write a duplicate again
  │
  ├─ ② Category tags: catalog reads "- [decompile] name: description"
  │     → the model sees the category before the prose
  │
  ├─ ③ Local re-check: a FOLD_INTO verdict is verified once more
  │     → same category: accept; cross-category: require similarity ≥ 0.5, else SAVE
  │
  └─ ④ The merge itself: model folds new experience into existing sections
        → duplicate sections truncated, protocol residue stripped,
          a result shorter than 60% of the original is refused

Categories are derived from the skill name (gui-* → gui, decompiled-* / java-decompile-* → decompile, portable-* / offline-* → portable, …). CATEGORY_ALIASES lives in lib/extractor.js; one line extends it.

Installation

The plugin is installed into the DSH profile with a link:, so code edits need no republish.

1. Clone anywhere:

git clone <this-repo> D:/dsh-plugins/dsh-skill-studio

2. Register it in the profile package.json:

{
  "dependencies": { "dsh-skill-studio": "link:D:/dsh-plugins/dsh-skill-studio" },
  "dsh": { "profile": { "bundles": ["dsh-skill-studio"] } }
}

3. Junction it into node_modules:

New-Item -ItemType Junction `
  -Path "$env:DSH_PROFILE_DIR\node_modules\dsh-skill-studio" `
  -Target "D:\dsh-plugins\dsh-skill-studio"

4. Restart DSH and open Settings → Skill Studio.

Configuration

Config file: ~/.dsh/dsh-skill-studio/extractor.json (mode 0600).

{
  "enabled": true,
  "autoPerTurn": true,
  "intervalMinutes": 60,
  "windowDays": 7,
  "maxCandidatesPerRun": 3,
  "autoAccept": true,
  "mergeThreshold": 0.2,
  "skillLanguage": "auto",
  "agentSkillDir": "C:\\Users\\<you>\\.agents\\skills",
  "provider": "",
  "model": ""
}

Key	Purpose
autoPerTurn	Run the YES/NO judgement at the end of every turn
intervalMinutes	Fallback sweep interval; 0 disables it and relies on per-turn only
autoAccept	Write straight into the library instead of the review queue
mergeThreshold	Local similarity pre-filter (the merge decision itself is the model's)
skillLanguage	auto / zh / en; code, commands, and paths are never translated
provider / model	Empty = follow the session; or pick one in the panel

The legacy llmBaseUrl / llmApiKey / llmModel fields still work as an HTTP fallback for hosts without ctx.llm. Leave them empty in normal use.

Tools

Tool	Purpose
skillmgr_list	List every skill with source, state, and group
skillmgr_get	Read one skill in full
skillmgr_save	Write a whole SKILL.md back
skillmgr_policy	Toggle model/user invocation
skillmgr_extract_run	Run one extraction now
skillmgr_extract_status	Switches, interval, queue depth, last run
skillmgr_extract_config	Change configuration
skillmgr_extract_list / _accept / _reject	Work the queue (usually empty with auto-accept on)

HTTP routes live under /api/dsh-skill-studio/* — that is what the panel talks to.

Debugging

Every model response is dumped to:

~/.dsh/dsh-skill-studio/last-llm-raw.txt

When parsing fails, read that file first: truncated JSON, markdown fences, or an off-topic answer are immediately distinguishable.

Two regression scripts, neither touches the real library:

node scripts/e2e-preview.mjs    # extract → judge → merge → accept, 6 assertions
node scripts/e2e-tombstone.mjs  # tombstone beats the model's verdict (it says SAVE, then KEEP)

Differences from upstream

The full list is in FORK-NOTES.md (15 entries). Highlights:

1. Model calls go through ctx.llm.prepareCall(); no stored credentials
2. Per-turn YES/NO judgement, auto-accept into the library
3. Semantic merging: category tags, model verdict, local re-check, tombstones
4. Merge hygiene: dedupe, residue stripping, short-body rejection
5. Panel: "recently added" list and a model dropdown
6. Truncation salvage for extraction output, maxTokens raised to 8000

License

MIT, same as upstream. Upstream copyright belongs to zhengjy01.