Semantic Task Clustering
Groups changed files into logical tasks such as auth refactors, UI updates, and database schema work with task-oriented labels and file impact metrics.
Open Source VS Code Extension
ADV is a local-first VS Code extension that helps review large AI-generated git diffs by grouping changes semantically, surfacing risk, and reconstructing agent intent for human-in-the-loop review.
Install From MarketplaceADV focuses on a hard review problem: AI agents can produce broad, fast diffs that are difficult to validate safely at file level. Instead of forcing reviewers through raw patches alone, ADV turns change sets into task-oriented review units enriched with intent, trust, and architectural context.
The project is open-source under the MIT License, and it is built to support workflows around Cline, Aider, OpenDevin, Copilot Agent mode, Cursor, and related agentic tooling.
Groups changed files into logical tasks such as auth refactors, UI updates, and database schema work with task-oriented labels and file impact metrics.
Aggregates risk signals such as secrets, complexity, drift, and architectural smell into a single trust metric.
Extracts user prompt and agent reasoning from local Cline and Aider logs so reviewers understand why changes happened.
Detects hardcoded secrets, sensitive path changes, and complexity spikes before generated edits are accepted.
Compares prompt scope against changed files and alerts when edits extend beyond declared intent, including severity, evidence, and affected files.
Visualizes file relationships and architectural risks with impact edges and smell edges for cross-layer issues or circular dependencies.
Lets reviewers scrub through agent execution flow and roll back from any step onward while keeping earlier approved work.
Persists task-level review state per workspace with pending, approved, rejected, and read markers.
Search for adv.* in VS Code settings. Ollama is optional and only used for local labeling when explicitly configured.
{
"adv.maxGroups": 8,
"adv.ollamaModel": "llama2"
}ADV scans the workspace root and subdirectories for local agent logs on startup to recover prompt and reasoning context.
From v0.1.7 onward, ADV also watches VS Code OutputChannels in real time and extracts working intent without extra setup.