DiffGate Review (VS Code Extension)
Catches the guardrails your coding agent deletes while editing — inline, on the lines you actually changed. We measured frontier models: 0% textbook OWASP bugs written from scratch, but security footguns reintroduced in 13% of edits (the measurement). DiffGate is the deterministic tripwire for that residue: it reviews only your changed lines (vs the committed baseline) and tells you which edits are safe, which need a second look, and which are high-impact enough to gate — before you ever open a pull request, before any review bot.
Tiers
| Tier |
Meaning |
Examples |
| 🟢 Green |
Safe / self-contained |
comments, local logging |
| 🟡 Yellow |
Review — soft dependency |
deprecated APIs, raw queries, network calls |
| 🟠 Orange |
High-impact — gate it |
schema changes, secrets, auth/crypto, injection sinks, public-API & signature changes |
Features
- Inline diagnostics on changed lines only (configurable to whole-file mode).
- Hover cards: why it's risky, who owns the baseline line (git blame), and quick links to AI explain or Deep Review.
- Deep Review: for orange findings, an agentic loop uses real repo tools (grep, read_file, git_blame, find_references) to investigate blast radius and shows a verdict badge directly in the hover card.
- Quick-fixes: replace a deprecated call with its successor (⌘. / Ctrl+.).
- Dismiss false positives (⌘. / Ctrl+. → "Dismiss as noise", or the hover link): suppresses that exact flagged pattern in a committed
.diffgate/learnings.json, so it's gone for every teammate and in CI — no inline // disable comments. Mark a finding "Confirm as a real risk" to feed diffgate stats signal-vs-noise. CLI dismissals reflect live in the editor.
- Risk Review tree: every pending change across the workspace, grouped by file and tier.
- Status bar risk summary for the active file.
- Verification gate: run your project's
testCommand for high-impact changes.
- Signature-drift detection: warns when you change an exported function's parameter list (callers may break).
- Hybrid AI (optional, provider-agnostic): works with Anthropic, OpenAI, OpenRouter, Groq, Together, LM Studio, Ollama, or any OpenAI-compatible endpoint.
Configuration
Project rules live in a .diffgate.json at the repo root (run diffgate init from the CLI, or DiffGate: Open Config). Editor settings:
| Setting |
Default |
Description |
diffgate.scanMode |
diff |
diff = changed lines only; file = whole file |
diffgate.diffMode |
working |
working = all uncommitted changes; staged = staged only |
diffgate.ai.enabled |
false |
Enable AI explanations (needs provider API key, except for local models) |
diffgate.ai.provider |
(from config) |
LLM provider override |
diffgate.ai.model |
(from config) |
Model id override |
diffgate.ai.deepReview.model |
(from config) |
Stronger model for Deep Review (e.g. claude-opus-4-8, gpt-4o) |
diffgate.runGateOnSave |
false |
Auto-run the gate on save when orange findings are present |
Develop / run
From the repo root: press F5 ("Run DiffGate Extension") to launch an Extension Development Host. Open a file (e.g. in mock_project/) to see findings; open a git repo to see diff-aware mode and the Risk Review tree.
Build a .vsix to install permanently:
npm install --prefix extension
npm run package --prefix extension # produces extension/diffgate-review-<version>.vsix
code --install-extension extension/diffgate-review-<version>.vsix
License
Apache 2.0
| |
