Automatic Context Compression with Configurable Threshold

[J4GL]

I’d like to propose a new feature for zed: automatic context compression. This mechanism would dynamically shorten or summarize the editor’s context (e.g., file buffers, chat history, code snippets) whenever it exceeds a user-defined threshold, ensuring the language model remains performant and within token limits.

---

Motivation

• Token Limits: As projects grow, the total context fed into the AI can surpass model limits, leading to truncated or failed completions.
• Performance: Reducing unnecessary or low-priority context speeds up prompt construction and API calls.
• Relevance: An intelligent compressor can prioritize the most pertinent code sections, comments, and conversation snippets.

---

Proposed Behavior

1. Threshold Setting: In the IDE settings, allow users to define a maximum context size (e.g., 4 000 tokens).

2. Compression Trigger: Whenever the aggregated context approaches the threshold, invoke the compressor.

3. Compression Strategy:

   • Least-Recently-Used: Demote or remove code regions least recently edited.
   • Semantic Summarization: Generate concise summaries of large code blocks or chat threads.
   • User Annotations: Honor “keep” or “summarize” comments to guide compression.

4. Preview & Rollback: Display a summary preview before sending; allow users to revert or expand individual sections on demand.

---

[jabcross]

This is what Github Copilot does, right?

[J4GL]

i don't use github copilot, claude code automaticly compress the context but there is no configurable threshold (yet)