English | 中文

dsh-compressor

A slim port of Headroom. Compresses tool output and cuts up to 20% of context, without affecting the model's context cache or agent performance.

Install

dsh plugin --profile web add dsh-compressor

How it works

A tool run leaves a long log / JSON / diff. Only a small part of it is useful. This plugin shortens that output before the next turn, keeps the original on disk, leaves the already-sent prefix alone so the cache stays valid, and gives the agent a tool to pull the full context back.

flowchart LR
  A[Tool finishes] --> B{Output long?}
  B -->|no| C[Full text stays]
  B -->|yes| D[Shorten and save original]
  D --> E[Model sees the useful part]

Compressor code comes from Headroom. Wired in today: Log / Smart / Text / Search / Diff. The only crushers not wired in yet:

• Code-aware (tree-sitter, source code)
• Kompress (ONNX, long prose)

Develop

cd plugins/dsh-compressor
pnpm install
pnpm test

Contributing

Open PRs on this repo, not on Headroom.

Plugin code is in plugins/dsh-compressor (Node 22+, pnpm):

cd plugins/dsh-compressor
pnpm install
pnpm test
pnpm typecheck

Crushers live in crates/:

cargo test --workspace
make test-parity

One change per PR. Plugin changes need tests; crusher changes need cargo test. Details: CONTRIBUTING.md.

This plugin is MIT. crates/headroom-* remains Apache-2.0.