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
• 🏗️ How it works
• 🚀 Quick start
• ⚙️ Configuration
• ❓ FAQ
• ⚠️ Security notes
• 📦 Project structure
• 🙏 Credits

✨ Features

🏗️ How it works

Settings page (browser)
  └─ user picks skills + injection mode
             │ PUT /skill-injector/api/config
Host half (DSH process) ▼
  └─ Loader Config: mode + selected (volatile settings fields)
  └─ 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/created (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)

1. git clone <repo-url> (any location).
2. Add to ~/.dsh/profiles/web/package.json dependencies: "skill-injector-plugin": "link:<repo-path>", then pnpm install in the profile dir.
3. Restart DSH.

Requirements

• DeepSeek Harness 0.2.x (built and verified against 0.2.1-rc.5). On 0.1.x the settings and agent-lifecycle APIs this plugin targets do not exist.
• 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 to edit. Settings are the plugin's Loader Config (mode + selected), edited from Settings → Skill Injector; saving writes them into the profile's patch file as this entry's config: block, so they survive restarts:

Both fields are declared .volatile(), which is what lets DSH persist an edit without reloading the plugin — a Save takes effect on the next prompt and keeps the settings page and dock line live. An empty or invalid value degrades 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 profile patch file (the same place DSH stores other plugin configuration).
• 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: volatile Config (mode + selected), resolveSelection (classification + last-good bodies), section + agent/created 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

• DeepSeek Harness — the DSH plugin/dynamic runtime, Slots, theme, webServer, client-modules.
• headroom-stats-plugin — packaged-plugin pattern reference (manifest, build, README structure).
• DSH-better-sidebar — reference for the packaged client-plugin build pattern.

📄 License

MIT