agentsclimarketplace

Delegate

Skill chinayin/gox-claude-plugins/plugins/token-thrift/skills/delegate

Token-saving subagent routing. Use when facing token-heavy, read-heavy, conclusion-only work (reading Lark/Feishu in full or writing to it, wide code searches, log/long-document analysis) to decide whether to offload the work to a cheaper-model subagent and which tier to use.From its SKILL.md

Install
npx -y skills add chinayin/gox-claude-plugins --skill delegate

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

1.8 KB, 378 tokens by cl100k_base, as published. Nobody here has run it

Token-saving three-tier split

The main agent (Opus) does only orchestration and decisions with the user. Offload token-heavy work; keep raw material out of the main context.

Which tier

Kind of workDispatch toModelWhy
Read / compare / verify / search (conclusion only)cheap-readerHaikuCheapest ($1/$5); raw material is read once and discarded
Correctness-sensitive writes (Lark/Feishu XML/blocks, reliable persistence)careful-writerSonnetWrite errors are costly; Sonnet is only 40–60% of Opus
Orchestration / decisionsmain agentOpusKeep the main context lean

A read that needs subtle judgment may be promoted to Sonnet; a purely mechanical append may drop to Haiku.

When it's worth offloading

Rule of thumb: offload when the throwaway material the subagent must process is ≳ 3,000 tokens. Because:

  1. Skill loading (e.g. lark-* SKILL + references, often tens of thousands of tokens) moves from Opus pricing to the cheaper model's pricing.
  2. Raw material stays out of the main context, avoiding it being re-billed on every subsequent turn (the biggest lever).

If the task is tiny, or the main thread will need the raw material again, keep it inline instead.

How to dispatch

Use the Task/Agent tool with subagent_type set to cheap-reader or careful-writer. Have it return only a compact conclusion; if the result is large, have it write a scratchpad file and return the path. Keep the Opus main context lean.

What ships with it

Read from the repository

Just SKILL.md. No reference files, no scripts.

Keep looking

Skills are one crate of 325,949. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.