Detailed Analysis
A Windows 11 user running Google Antigravity — a fork of Visual Studio Code that bundles multiple AI models — has encountered a persistent chat history failure when using the Claude Code extension with an Anthropic Pro subscription. The user modified Antigravity's extension marketplace endpoints to point to the standard VS Code extension store, thereby circumventing the fork's curated model limitations and enabling installation of third-party tools including Claude Code. While core Claude Code functionality appears intact in this configuration, session data does not persist across restarts, forcing the user into manual workarounds: either keeping the editor open indefinitely, exporting conversation state to Markdown files, or prompting Claude to re-analyze the entire project codebase at the start of each new session — a process that consumes significant token budget.
The technical root cause almost certainly lies in how Claude Code stores session data relative to VS Code's native extension host environment. Claude Code was designed and tested against official VS Code builds, and its chat persistence mechanism likely writes to or reads from storage paths, APIs, or context providers that Antigravity either remaps, restricts, or does not fully implement. When a VS Code fork diverges from upstream in its extension host, workspace storage, or state management APIs — even subtly — extensions that depend on those systems for persistence can silently fail to write or retrieve session records. The user's approach of swapping marketplace URLs is a common workaround in the VS Code fork ecosystem, but it does not guarantee behavioral parity for extensions that interact deeply with editor internals.
This situation reflects a broader tension emerging in the AI developer tooling space: as VS Code forks proliferate — driven by enterprise, privacy, and multi-model integration use cases — extension developers like Anthropic face an expanding matrix of partially compatible runtime environments. Claude Code, being a sophisticated agentic tool rather than a simple syntax highlighter, relies on stateful session management that is far more sensitive to runtime environment differences than traditional extensions. The gap between "the extension loads" and "the extension functions fully" is widening as these tools grow more complex.
From a user experience standpoint, the token-cost implications of losing chat context are non-trivial. Agentic coding assistants like Claude Code derive much of their value from accumulated conversational context — understanding project structure, prior decisions, and established conventions. Re-analyzing an entire codebase from scratch at the start of every session not only wastes tokens but degrades the quality and continuity of AI assistance, partially negating the productivity gains the tool is meant to deliver. The user's Markdown export workaround is a pragmatic stopgap, but it underscores how dependent modern AI coding workflows have become on robust session persistence infrastructure.
The incident points to a gap in Anthropic's current compatibility documentation and testing surface for Claude Code. As the extension gains adoption beyond vanilla VS Code — including in forks, remote development environments, and enterprise-modified distributions — Anthropic will likely face increasing pressure to either formally document supported runtime environments, provide diagnostic tooling for persistence failures, or architect Claude Code's storage layer with greater resilience to non-standard extension host implementations. Until then, users running Claude Code in derivative editors should anticipate undocumented behavioral deviations, particularly in stateful features like conversation history.
Read original article →