skill-injector-plugin · Auto-Inject Skills into DeepSeek Harness (DSH)
Auto-inject user-chosen skills (e.g. /caveman, /ponytail) into every DSH session — every prompt or once at session start — with a settings page and a composer indicator. 在 DSH 会话中自动注入所选技能（如 /caveman、/ponytail）：每轮提示或仅在会话开始时，含设置页与输入区指示。
中文文档: README.zh.md · LLM index: llms.txt · Agent guide: AGENTS.md
Keywords: dsh-plugin · deepseek-harness-plugin · skills · skill-injection · caveman · ponytail
📑 Table of Contents
✨ Features
| Feature | Description |
|---|
| ⚙️ Settings page | Settings → Skill Injector: checkbox list of available skills, injection-mode radio, save; unresolved skills are listed with the reason and a Retry button (plus Remove once a session has established scope) |
| 🪧 Composer indicator | One line under the chat input: Injected Skills: caveman, ponytail · Mode: each-prompt, refreshed every 5s (with a · N unresolved suffix while skills are unresolved) |
| 🔁 Two injection modes | each-prompt (system-prompt section, re-rendered every request) vs start-only (one durable skill-invocation message stamped at session start) |
| 📚 Live skill registry | Reads skill bodies from ctx.skills at load time; no copied files; each selected skill is classified (ready / unknown / unavailable / load-failed / no-scope) and a transient failure keeps injecting the last-good body |
| 🧩 Subagents included | Injection applies to subagents too (consistent voice; no filtering) |
| 🌗 Theme-aware | All colors use --dsw-alias-* design tokens |
| ♨️ Survives restarts | Real profile-bundled plugin: install once with dsh plugin add, auto-loads on every DSH boot — no per-session define, no cordis_define |
🏗️ How it works
Settings page (browser)
└─ user picks skills + injection mode
│ PUT /skill-injector/api/config
Host half (DSH process) ▼
└─ skill-injector settings namespace (mode + selected)
└─ ctx.skills.get(name) → live skill bodies from the skill registry
└─ each-prompt: systemPrompt.section (re-rendered every request)
OR start-only: agent.inject() on agent/session-start (one stamp per skill)
└─ webServer route GET /skill-injector/api → JSON snapshot
│
Client bundle (browser) ▼
└─ single 5s poller → fetch(/skill-injector/api)
├─ settings.section (id skill-injector) → checkbox list + mode radio + save
└─ conversation.composer.dock (id skill-injector-dock) → active-skills line
- Pure pull model: no push, no events driving the UI — the client polls
/skill-injector/api every 5s and the host re-reads ctx.skills on demand; the snapshot's unresolved array carries a reason per skill, the poll is the automatic retry, and the settings page offers a manual Retry plus Remove.
- Transient failures keep the last-good body: a provider error, an incomplete discovery, or a not-yet-scoped read keeps injecting the body loaded last time. Only a complete, scoped read that does not advertise the name marks it
unknown, and only then is it dropped.
- Live registry: injection always reads the skill bodies from
ctx.skills at load time — nothing is copied into this package, so skill edits are picked up and deleted skills degrade gracefully.
- Persistence: ships
dsh.bundle (cordis.patch.yml) + dsh.client (exports["./client"], bundled) so it installs as a real profile plugin that DSH loads on every boot.
🚀 Quick start
Standard install: dsh plugin add (persists across restarts)
Install the package from a local checkout (or from GitHub once published):
# local directory (from the parent of this repo):
dsh plugin --profile web add ./skill-injector-plugin
# or directly from GitHub (any DSH machine):
dsh plugin --profile web add git+https://github.com/<org>/skill-injector-plugin.git
dsh plugin add is a pnpm add into the profile plus a dsh.profile.bundles reconcile: seeing this package's dsh.bundle declaration, it appends skill-injector-plugin to the bundle stack. Restart DSH (or hard-refresh). On boot the client-modules scanner resolves exports["./client"] and the settings page + dock line appear. No per-session define, survives restarts.
Manual profile mount (alternative)
git clone <repo-url> (any location).- Add to
~/.dsh/profiles/web/package.json dependencies: "skill-injector-plugin": "link:<repo-path>", then pnpm install in the profile dir.
- Restart DSH.
Requirements
- Skills exist in
~/.agents/skills (or another configured skill root the ctx.skills registry reads).
- No skill selected → nothing is injected; the dock line reads
Injected Skills: none · Mode: each-prompt and the settings page still works.
⚙️ Configuration
No config file. Settings live in the skill-injector settings namespace, edited from Settings → Skill Injector and persisted in the user settings document:
| Key | Default | Meaning |
|---|
mode | 'each-prompt' | Injection mode: 'each-prompt' (system-prompt section, every request) or 'start-only' (one stamped message at session start) |
selected | [] | Kebab-case skill names to inject (max 16, validated; duplicates deduped) |
The namespace is registered via ctx.settings.register('skill-injector', schema, { base }) — the stored user section layers over the base automatically, so an empty/invalid stored value falls back to { mode: 'each-prompt', selected: [] }.
❓ FAQ
Q: Why don't my settings apply to the session I already have open?
A: In each-prompt mode the section is re-rendered per request, so the next request picks it up. In start-only mode the stamp is written once at session start — it applies to sessions started after the change, not the one already open.
Q: What happens when a skill file is deleted?
A: The next complete, scoped catalog read marks it unknown (not in the skill catalog), it stops injecting, and the settings page lists it under Unresolved with a Remove button — so a stale name can always be cleared. Every other selected skill keeps working. Restore the file and the next refresh picks it up again.
Q: The settings page says "provider discovery incomplete" or "waiting for a session to establish scope" — is my skill gone?
A: No. The first means provider discovery did not complete, and the second means no session has established a lookup scope yet (that is normal before the first message of a process). Both are transient: the skill keeps injecting its last-good body and the next refresh clears it. Retry re-runs that refresh immediately, and Remove only appears once a scoped read has actually run.
Q: Do subagents follow the injected skills?
A: Yes — injection applies to subagents too, by design, with no origin filtering. The chosen voice is consistent across the whole agent tree.
Q: What is the token cost?
A: The skill content is sent every request either way: each-prompt keeps it in the system prompt, start-only in one stamped message; both are visible to the model on every turn. Select only the skills you actually want active.
Q: How do I remove it?
A: dsh plugin --profile web rm skill-injector-plugin (or delete the profile dependency + bundle entry) and restart DSH.
⚠️ Security notes
- The plugin is read-only over the skill registry — it never writes, renames, or deletes skill files.
- The selection is stored in the user settings document (same store as other DSH settings).
- Injected content is the skill's own trusted local markdown; nothing is fetched from the network.
- In the
each-prompt (system-prompt section) path, {{ in skill bodies is escaped to keep strict variable interpolation from failing; start-only messages are not interpolated, so they keep the raw content.
📦 Project structure
skill-injector-plugin/
├── src/
│ ├── index.ts # host half: settings namespace, resolveSelection (classification + last-good bodies), section + session-start injection, routes
│ ├── helpers.ts # pure helpers: validateSelection, escapePromptBraces, buildInjectionMessage, classifySelection, applyBodies
│ └── client/index.tsx # client bundle: 5s poller, settings form, dock line
├── cordis.patch.yml # dsh.bundle patch (inserts the plugin row on boot)
├── tsdown.config.ts # builds host (node ESM) + helpers + client (CJS ModuleLoader)
├── package.json # name, exports["./client"], dsh.client + dsh.bundle
├── lib/ # build output (index.js, helpers.js, client.js)
├── tests/
│ ├── helpers.test.mjs # node:test unit tests against lib/helpers.js
│ ├── host.test.mjs # node:test smoke test driving apply() with a stub ctx
│ └── fixtures/ # real-shape skill definition sample
├── AGENTS.md # repository guide for AI agents
├── llms.txt / llms-full.txt
├── README.md / README.zh.md
└── LICENSE
🙏 Credits
📄 License
MIT
