VS Code Copilot Chat Improvements from Initial Testing

[danielrobbins]

First off, thank you for adding the new VS Code Copilot Chat integration to llama-vscode. This is a great feature, and after spending some time testing it, I think it opens up a very promising workflow for local and hybrid model use inside VS Code.

While doing my initial testing, I ran into a handful of issues around tool-call handling, VS Code chat context reporting, build compatibility, Windows ARM install behavior, and settings/schema consistency. After developing fixes for all the issues I encountered, I split the follow-up work into focused PRs so each fix is easier to review independently.

Key Copilot Chat fixes

These are the two higher-impact fixes that came directly out of exercising the new VS Code Copilot Chat integration in real use:

• Fix VS Code Chat context window #191 Fix VS Code Chat context window
  Fixes token-limit reporting for VS Code Copilot Chat models so llama.cpp-backed models no longer advertise the wrong 12K fallback budget to VS Code. This fixes an aggressive history compaction/summarization bug when auto-summarization is enabled which often causes the session to quietly abort.

• Enhance VS Code Chat Provider Bridge to Fix Create-File (and other) Tool Calls #190 Enhance VS Code Chat Provider Bridge to Fix Create-File (and other) Tool Calls
  Fixes the VS Code native chat bridge so tool metadata is preserved instead of flattened into plain text. This improves reliability of tool calling and in particular allows Gemma4 to create files which were consistently hallucinated before, even when some other tool calls worked.

Supporting maintenance and compatibility fixes

These are other fixes that improve compatibility, setup, and configuration correctness:

• Remove the obsolete VS Code language model chat shim #187 Remove the obsolete VS Code language model chat shim
  Removes the old local LM chat-provider shim that now conflicts with current @types/vscode and unblocks clean compilation against the current VS Code typing surface. This is required for some of my PRs.

• Fix Windows 11 ARM64 llama.cpp Download #188 Fix Windows 11 ARM64 llama.cpp Download
  Fixes the Windows on Arm install path so the extension requests the correct ARM64 llama.cpp package via winget. Without this, the sub-optimal x86-64 version is installed.

• Fix agent list settings schema #189 Fix agent list settings schema
  Fixes a schema/runtime mismatch in llama-vscode.agents_list, normalizes systemInstruction handling, preserves compatibility with legacy system_instruction, and corrects the subagentEnabled schema type. This fixes an inconsistency in the codebase that could impact users.

Notes

A few of the PRs currently overlap with #187 because I propagated the shim cleanup while validating that each branch still compiled cleanly in my local environment. If #187 merges first, I am happy to rebase or otherwise help reduce overlap as needed.

I wanted to post this issue mainly as a project-level summary so the fixes are easy to track in one place rather than scattered across separate PR threads.

Thanks again for adding the VS Code Copilot Chat integration. It fills a critical gap in the VS Code ecosystem.