dsh-token-skills

Cut token cost in 4 layers — without cutting model intelligence.

A DeepSeek Harness preset of four complementary token-saving skills. Each one attacks a different stage of the token pipeline, so they stack:

handoff the context → prune the input → compact the context → terse the output

Stop paying twice for tokens you already have.

flowchart LR
  A[Tokenizer] --> B[① Handoff<br>session]
  B --> C[② RTK<br>input]
  C --> D[③ Headroom<br>context]
  D --> E[④ Caveman<br>output]
  E --> F[<b>fewer tokens</b>]

Why: every token is money. A long agent session re-sends context it already paid for — past transcripts, padded logs, verbose replies. These four skills stop that leak at each stage.

---

The four layers

#	Skill	Layer	What it saves	Backs off to
①	handoff	Session	stop re-reading/rebuilding context between sessions	context-auto-handoff
②	rtk	Input	drop redundant/repeated tokens before they enter the model	RTK (Rust Token Killer)
③	headroom	Context	compress the live context so the prompt stays small	WanYanTianDe/dsh-headroom
④	caveman	Output	terse prose — fewer tokens on the way out	skill-caveman

Each layer is a thin dsh skill (skills/<layer>/SKILL.md). Where a real engine already exists, the skill wires to it instead of re-implementing compression — you get the strategy and the engine, for free.

---

Standalone repos

Each layer is also its own installable plugin, so you can add just the one you want:

Skill	Repo
handoff	WODE25500/dsh-token-handoff
rtk	WODE25500/dsh-token-rtk
headroom	WODE25500/dsh-token-headroom
caveman	WODE25500/dsh-token-caveman

# install just one
dsh plugin --profile web add dsh-token-handoff

---

When to use which

• Handoff — the highest-leverage, cheapest habit: write a HANDOFF.md note at session end so the next one starts warm. Use it on every multi-session project.
• RTK — per request, before you paste a big file/log/repo listing that's mostly noise.
• Headroom — continuous, for long-lived agents: compact before the window fills, not at the cliff edge.
• Caveman — every reply: it costs nothing and compounds. It changes style, never substance.

They stack to a single habit, not a heavy setup:

before: [past session verbatim] + [entire repo dump] + [fat context] + [verbose response]
after:  HANDOFF note      → pruned diff      → compacted headroom  → terse reply

---

Install

dsh plugin --profile web add dsh-token-skills

Or clone it and drop the skills into ~/.dsh/skills/:

git clone https://github.com/WODE25500/dsh-token-skills

Each skill is also invocable directly — handoff, rtk, headroom, caveman.

---

Validation

npm test   # runs scripts/check-skills.mjs — asserts every bundled skill's frontmatter

---

License

MIT