Skip to content

VS Code (GitHub Copilot)

GitHub Copilot's agent compacts context internally and does not expose a user-settable compaction model. There is no compaction_model equivalent to Zed's. So the only way RustySqueezer helps VS Code is gateway-side: put LiteLLM in front of the model and let the router compact before dispatch.

The gateway route

flowchart LR
  A[VS Code agent] -->|OpenAI-compatible| B[LiteLLM]
  B -->|over budget? compact first| C[RustySqueezer]
  B -->|trimmed thread| D[Real model]
  1. Configure LiteLLM as in the LiteLLM guide: a router entry with an honest declared window, and context_compaction on the auto-router with rusty-compactor pointing at RustySqueezer.
  2. Point the VS Code agent at LiteLLM's base URL as its OpenAI-compatible endpoint.
  3. Copilot sends full threads to LiteLLM; LiteLLM compacts anything over the trigger and forwards a trimmed thread to the real model.

Which clients can be re-pointed

This works for VS Code agents that let you set a custom OpenAI-compatible base URL (Continue, Cline, and custom scripts). Copilot's own "bring your own key" flow lets you choose a provider, but not an arbitrary base URL, so it cannot be routed through your gateway directly.

RustySqueezer as an Ollama model

RustySqueezer also implements the Ollama surfaces (/api/chat, /api/generate, GET /api/tags), so it shows up in tooling that discovers models from an Ollama-style server. Copilot's local-model support can list it this way - but Copilot would then use it as the chat model, not as a summariser, so this is only useful for agents that have a separate summariser slot.

Rule of thumb

If a client has a compaction setting, use it (Zed, OpenCode). If it does not, put the gateway in front and let LiteLLM do the compaction - that is the whole reason the gateway-side path exists.