# ๐ธโโ๏ธ dsh-skill-router
### *Rule-first pre-step skill routing for DeepSeek Harness: pours matched skills, stays silent when unsure.*
[](LICENSE)
[](https://github.com/akqwpeter-prog/dsh-skill-router/actions/workflows/test.yml)
[](https://github.com/topics/dsh-plugin)
[](README.md#-how-it-works)
[](README.md#-how-it-works)
[](https://github.com/akqwpeter-prog/skill-bartender)
[](docs/lang/README_ZH.md)
Companion executor to [skill-bartender](https://github.com/akqwpeter-prog/skill-bartender):
the skill carries the policy *judgment*, this plugin carries the *execution*.
Deterministic, **zero LLM calls, zero token cost** until a rule actually pours:
- โก **Pre-step hook** โ reads the latest user message before every step.
- ๐งญ **Rule-first matching** โ user-editable YAML policy (`~/.dsh/skill-router.yaml`,
bundled defaults in `default-policy.yaml`), first match wins.
- ๐ **Silent miss** โ no hit โ zero intervention; the model keeps its normal
catalog flow.
- โป๏ธ **Once per session** โ each skill pours at most once.
- ๐ก๏ธ **Broken YAML never breaks the session** โ falls back to bundled defaults.
[Why](#-why) ยท [How it works](#-how-it-works) ยท [What you get](#-what-you-get) ยท [Quick start](#-quick-start) ยท [See it in action](#-see-it-in-action) ยท [Policy](#-policy) ยท [Tested](#-tested) ยท [Scope & non-goals](#-scope--non-goals) ยท [FAQ](#-faq) ยท [Layout](#-layout) ยท [License](#-license)
[**English**](README.md) ยท [**็ฎไฝไธญๆ**](docs/lang/README_ZH.md)
---
## ๐ค Why
Most skill loading is left to the model's judgment: it sees the catalog every
step, re-decides every time, and often loads late, wrong, or not at all. A
router that runs *before* the model answers fixes that:
| | dsh-skill-router | LLM-judge router | Manual loading |
|---|---|---|---|
| Decision maker | **rules** (deterministic) | LLM / embeddings | the model, per step |
| Token cost | zero until a rule pours | every step | every step |
| Latency added | ~0 ms | model round-trip | n/a |
| Reproducible | โ
same message โ same pour | โ varies | โ varies |
| User control | edit YAML, done | prompt it | hope it remembers |
**Why rules and not an LLM judge?** Speed, cost, and predictability. A
URL-path rule routes `feishu.cn/x/docx/` to `lark-doc` in microseconds, for
free, every time โ and `skill-bartender`'s routing table is where the policy
*judgment* lives. This plugin is the muscle, not the brain.
## โ๏ธ How it works
- Hooks `agent/pre-step`, reads the latest user message.
- Matches it against user-editable rules (`~/.dsh/skill-router.yaml`, bundled
defaults in `default-policy.yaml`). First match wins.
- On a hit: pours the matched skill bodies into the step as
`skill-invocation` messages โ the catalog's "already loaded, don't
re-load" rule applies automatically.
- No hit: zero intervention. The model keeps its normal catalog flow.
- Each skill pours at most once per session.
## โจ What you get
| Capability | What it does |
|---|---|
| โก Pre-step hook | `agent/pre-step` โ the pour happens *before* the model starts thinking |
| ๐งญ YAML policy | User-editable `~/.dsh/skill-router.yaml`; broken YAML falls back to bundled defaults |
| ๐ `whenToUse` triggers | Installed skills' `whenToUse` frontmatter acts as a secondary trigger (literal phrase match, appended after YAML rules) |
| ๐ Zero cost | No LLM judge, no embeddings โ rules only (fast, free, deterministic) |
| โป๏ธ Once per session | Dedupes pours per session; no skill body floods the context |
| ๐ Companion | Works with skill-bartender's routing table and taste test |
## โก Quick start
```sh
dsh plugin --profile web add github:akqwpeter-prog/dsh-skill-router
```
Then restart the running instance (profile bundles load at boot).
Verify: say "็ๆไธๅผ ๆตทๆฅ" โ `media-tools` pours automatically; say "่ฟไธชๆชๅพๅธฎๆๆฃๆฅไธไธ" โ `vision-review` pours. No rules matched? The model just works as usual.
## ๐ธ See it in action
*One picture: a rule hits โ the skill pours before the model answers; no hit โ total silence.*