Token compression for AI coding assistants

The coding assistant token problem

A typical coding assistant request includes: the current file being edited (500-2,000 tokens), relevant project files (3-10 files, 1,000-5,000 tokens), git diff context (200-1,000 tokens), conversation history (500-2,000 tokens), and system prompts for language and framework rules (300-500 tokens). Total: 2,500-10,500 tokens per request.

Most of this context is irrelevant to the specific coding task. The file the developer is working on contains 80% of the needed context. The rest is noise.

Compression strategy for code context

SuperCompress scores each included file's content against the developer's query or cursor position. Files containing relevant imports, function calls, and type definitions are kept. Files included because of project structure heuristics (but irrelevant to the task) are removed.

from supercompress import Compressor
comp = Compressor()

# Before sending context to the coding assistant
code_context = '''File: app.py
from flask import Flask
app = Flask(__name__)
...'''

query = "Add a new route for user login"
result = comp.compress(code_context, query)
# Result keeps only the file sections relevant to the login route task

Estimated savings

Frequently asked questions

Does compression work with Cursor's agent mode?

Yes. Compress the gathered context before the agent processes it. The agent behavior is unchanged.

Will compression remove important imports?

No. Query-aware compression scores each import line against the request. Imports needed for the task are preserved.