Vestige

Understand this change and what it will break, so you can push it with confidence.

Vestige makes AI-generated diffs reviewable before they're pushed. When an LLM changes a lot of code at once, Vestige answers three questions fast:

1. Where do I start? — entry points and a guided reading order for the change.
2. What could this affect? — the blast radius: everything the change can reach or break.
3. What did this code do before? — the surrounding context, without re-deriving it.

How a review works

1. Open a folder. Vestige scans and parses it locally — no LLM calls — to build a call graph and a file-dependency graph.
2. Read the graphs. Click the Vestige icon in the activity bar, or run Vestige: Open Graph, for two tabs on one panel: Repo Graph (function calls) and Feature (feature clusters). Click any node to jump to that code.
3. Get notified when something changes outside your own typing — an AI agent, a git pull, a bulk edit. Vestige detects it and offers to open the Splash Zone, or opens it automatically if you've turned on vestige.splashZone.autoOpen.
4. Review in the Splash Zone. The change is broken into an ordered set of review cards, each with its own blast radius, an inline or split diff, an LLM-generated one-line summary, an Explain action for deeper context, and accept/revert controls.

New to Vestige? Run Vestige: Getting Started for a five-step walkthrough (map your project, read the graphs, review a change, judge the blast radius, explain & decide) that links straight to the relevant commands.

Design principles

• Impact analysis is computed, not guessed. Blast radius comes from the parsed call/dependency graph — deterministic and local, never an LLM hallucinating what a change touches.
• The LLM only labels and summarizes. Snippet explanations, change-review summaries, and feature-cluster names are generated text on top of the analysis; they never decide what's affected.
• Analyses are cached. Re-opening a graph or a review doesn't re-run or re-bill work that's already been done.

Data & privacy

Call-graph, dependency-graph, and blast-radius analysis run 100% locally — nothing about your code's structure leaves your machine for these.

Two features send code to the Vestige backend (hosted on AWS, which calls OpenAI to generate text):

• Explain Selection sends the selected code snippet.
• Change-review summaries (the one-line summaries on Splash Zone cards) send the changed-code snippets for that chunk.

Feature-cluster analysis sends pre-computed summaries of your code, not raw source.

Nothing is sent unless you use one of these features. There is no telemetry and no analytics.

First-launch limitations

• TypeScript/JavaScript, Python, Go, and Java are supported. The parsers ship with the extension (tree-sitter compiled to WebAssembly), so nothing to install — open a folder in any of these languages and the graphs, blast radius, and reviews work the same. Other file types are treated as plain dependency-graph nodes.
• Single-root workspaces only. Vestige scans the first folder in your workspace.
• Re-parsing on file change isn't incremental at the AST level yet. Fine for small/medium repos; large monorepos will feel the rescan.
• Feature View's feature map is a whole-repo analysis, generated on demand and cached, not something that updates live.
• This is a first release. Expect rough edges, and remember the blast radius is only as good as the static call graph — dynamic dispatch and reflection can escape it.

Commands

v0.1.0 — first public launch.