Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 

Repository files navigation

compressor

Filters noisy command output down to what a coding agent actually needs, before it spends context.

The Problem

The biggest source of token waste in an agentic coding loop isn't the conversation — it's the command output. Every test run, package install, and build spits out hundreds of lines an agent has to read: pass/fail spam, dependency chatter, environment metadata. Somewhere in there are the two or three lines that actually matter — the failing assertion, the exit code, the one warning that's load-bearing. The agent pays full price in tokens to find them, every single time.

Blind truncation isn't the fix: cutting output for length alone risks dropping the actual failure or next clue, which sends an agent into repeated, uninformed retries instead of fixing the real problem.

Direction

A CLI you wrap your commands with — command-aware filters rather than one generic summarizer, recognizing common command output (test runners, package managers, build tools, git) and reducing each to its own shape of signal: failures, summaries, counts, and next clues survive; everything else is safe to drop. A tunable aggressiveness dial, and a way to verify agents don't regress once it's on.

Status

Planning. Design and benchmarking approach are being worked out — nothing shipped yet.

Stack

  • Language: TypeScript
  • Structure: Effect, for typed errors, structured concurrency, and composable pipelines through the compression stages
  • Distribution: compiled to a single native binary via bun build --compile — no runtime install required on the target machine

Tools in this space commonly ship as standalone binaries written in Go. This takes the same distribution model — install one binary, wrap your commands, done — while getting there through TypeScript and Effect instead.

About

Compress context before LLM inference to reduce token consumption, lower inference costs, and maximize effective context window utilization

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors