Context Token Efficiency and Reuse in Assistant Panel

[basilex]

Hi Zed Team,

I've recently switched from VS Code to Zed and I'm absolutely loving the performance. However, I’ve encountered a significant issue regarding how the Assistant Panel handles context in large-scale projects (specifically working on a massive CRM module).

The Problem:
Currently, the context limit (around 200k tokens) is being consumed very rapidly. When I open a new chat session or move between windows, the engine (whether it's Sonnet 4.6 or Zed’s internal one) seems to re-scan and re-ingest the entire project structure/context from scratch.

In my case, just "indexing" the module context often eats up 50% of the 200k limit before I even ask a question. This leads to:

High latency while the context is being gathered.

Significant financial waste due to redundant token usage.

Rapidly hitting context window limits, making it impossible to hold long architectural discussions.

Suggestions:

Context Caching: Implement better caching mechanisms so that if the project structure hasn't changed, the model doesn't need to "re-learn" it for every new prompt or session.

Granular Context Control: Allow us to pin certain parts of the context globally across sessions or define "Context Profiles" for specific directories.

Persistent Context: A way to keep the indexed state of a module active across multiple chat threads without re-billing for the same tokens.

I really want to stick with Zed, but the current token consumption for large projects is becoming a deal-breaker for professional use.

Best regards, Oleksandr
alexander.vasilenko@gmail.com