VibeThrift for VS Code and Cursor

VibeThrift is the editor-native interface for focused AI coding. It maps the open workspace locally, compares direct GPT-5.6 and Mercury 2 + focused GPT-5.6 paths, and requires native diff review before it applies verified edits.

Install from source

Requirements: Node.js 20+, VS Code or Cursor compatible with VS Code API 1.90+, and a desktop extension host.

From the repository root:

npm run check --prefix extension
npm run package --prefix extension

This creates extension/vibethrift.vsix. In VS Code or Cursor:

Open Extensions.
Choose Install from VSIX….
Select extension/vibethrift.vsix.
Reload the editor if prompted.
Open the VibeThrift icon in the activity bar.

For local development, open this extension directory in VS Code or Cursor and press F5 to launch an Extension Development Host.

First paid run

Use the small gear beside Map context & forecast or run VibeThrift: Configure Provider Keys from the Command Palette.

The direct GPT-5.6 route requires an OpenAI key.
The hybrid route requires OpenAI and Inception keys.
Keys are stored in VS Code SecretStorage. They are never stored in workspace settings, source files, or localStorage.
The gear menu and VibeThrift: Remove Provider Keys command permanently delete saved keys after confirmation.

Workflow

Enter a coding task and select a context mode.
Select Map context & forecast. This scans the workspace locally and makes no paid provider call.
Review the direct and hybrid list-price forecasts. The hybrid total includes both Mercury and GPT-5.6.
Choose the route you want to run.
Review the returned plan and open each native diff.
Select Apply verified edits only after reviewing every change.
Optionally run the suggested allowlisted test and inspect its result.

Applying an already-reviewed edit is local: it does not trigger another model call. The extension revalidates the exact expected content before it writes.

Context modes

Mode	Broad investigator cap	Focused GPT evidence cap	Best for
Cheap	35,000 estimated tokens	800	Contained fixes and fast iteration
Balanced	110,000 estimated tokens	1,200	Default repository work
Thorough	220,000 estimated tokens	2,500	Cross-cutting changes with more unknowns

The selected mode is saved in workspace settings. Advanced users can set vibethrift.broadContextLimit or vibethrift.contextBudget; use 0 to defer to the selected mode.

When broad context exceeds its cap, VibeThrift prioritizes task-relevant files and tests, then tells you how many files and estimated tokens were omitted. It does not imply that omitted files were inspected. Common build artifacts, dependency lockfiles, and obvious credential-bearing files are excluded before provider calls.

Usage and safety

The VibeThrift Runs Explorer view records provider-reported usage and the extension's corresponding list-price estimate. Forecasts are not billing receipts: they exclude caching, discounts, retries, and network overhead.

Suggested tests require an explicit click, run only from a small shell-free allowlist, and write output to the VibeThrift Tests output channel. Run any other command manually after review.

Use persistent chats for related work. VibeThrift carries only a bounded summary of recent task/outcome turns into a continuation request. Select New chat for a fresh context, or use Recent chats to return to an earlier thread.

Verify a build

Run the automated checks from the repository root:

npm run test:all

For a manual end-to-end rehearsal, follow TESTING.md.

VibeThrift

haranprithvii