If you code daily with Claude 3.5 Sonnet, ChatGPT Plus, or GitHub Copilot, you’re familiar with the midday lockout:
"You've reached your message limit until 4:15 PM."
Most developers assume they simply sent too many messages. But under the hood, provider rate limits aren't just counting turns—they're tracking token consumption velocity over a sliding window.
When you ask an AI assistant to analyze a bug across four files, the assistant often slaps 4,000 lines of raw code into the prompt. That's 30,000+ tokens burned on boilerplate, imports, and irrelevant helper functions for a one-line bug fix.
The Solution: AST Slicing vs. Naive Summarization
Summarizing code with an LLM before sending it to another LLM is slow, expensive, and frequently strips critical types.
Instead, we built Tokonomics (tokonomics-ai on VS Code) using local Tree-sitter WebAssembly parsers. It performs structural code pruning:
- Edit targets remain 100% byte-exact.
- Dependencies & callers are sliced down to interfaces, exported contracts, and type definitions.
- Results: 70% to 75% reduction in context payload size, 3x–4x more prompt headroom within your subscription cap, and faster Time-To-First-Token (TTFT).
Check out the full open-source portfolio and extension here:
- VS Code Marketplace: Tokonomics
- GitHub: Tokonomics-Team/tokonomics
Top comments (0)