ContextLens: Verified Tests

A VS Code extension for Python, JavaScript, and TypeScript that adds two CodeLens actions above every function, method, and nested function/method it can find:

🧪 Generate & Verify Tests

1. Extracts the function/method's signature and used imports (regex-based, no external parser).
2. Retrieves relevant context from your workspace via a lightweight, embeddings-free RAG index (per-function/class chunking, term-frequency scoring, scoped to the file's language).
3. Asks an LLM (OpenAI-compatible, Anthropic, or Google Gemini) to write tests — unittest for Python, Node's built-in node:test for JavaScript/TypeScript (no extra test-framework install required either way).
4. Actually runs those tests locally in a sandboxed temp directory (python3 -m unittest, or node --test / node --experimental-strip-types --test for TS).
5. If they fail, feeds the failure back to the model and retries (up to contextLens.maxRetries, default 3).
6. Once tests pass, asks the model for a short analysis of the code — bugs, uncovered edge cases, duplicate logic elsewhere in your codebase, and convention mismatches — using a second, source-biased RAG retrieval pass.
7. Opens the verified tests (with the analysis as a leading comment block) in a new editor tab. Nothing is ever written back into your source tree automatically.

🔎 Explain Issues

A lighter-weight companion action for when you don't need generated tests, just a second opinion: asks the model to review the function/method for concrete bugs and edge cases (same RAG context as above), and surfaces findings as real diagnostics in the Problems panel — with squiggly underlines at the relevant lines — instead of a wall of text in a separate file.

Install

Search "ContextLens: Verified Tests" in the VS Code Extensions view, or install directly: marketplace.visualstudio.com/items?itemName=eyobelg.context-lens-python

Setup

You need your own API key. ContextLens does not ship with, embed, or proxy any API key — nothing is bundled into the extension. You must supply your own key from OpenAI, Anthropic, or Google (Gemini has a free tier — get one at aistudio.google.com/apikey). Requests go straight from your machine to the provider you configure; ContextLens never sees or forwards your key anywhere else.

Configure under Settings → ContextLens:

The progress notification shown while generating is cancellable — click the ✕ to stop mid-run; it aborts the in-flight model request and kills the sandboxed test process.

TypeScript note: native type-stripping requires Node 22.6+ (behind --experimental-strip-types, always passed) or Node 23.6+ (unflagged, the flag is simply ignored). Earlier Node versions will fail to verify TypeScript tests — Python and JavaScript are unaffected.

Contributing

npm install
npm run compile
npm test    # compile + unit tests
npm run lint

Press F5 in VS Code to launch an Extension Development Host. See CLAUDE.md for architecture details and AGENTS.md for the full build/lint/test/package workflow.

Known limitations / roadmap

• Nested functions (a def/function inside another one) get a CodeLens, but since they aren't independently callable, the model is prompted to test them indirectly through the enclosing function — results are best-effort and depend on the nesting.
• JS/TS parsing is regex/brace-based, not a real parser. Common patterns (function declarations, arrow functions assigned to a variable or class field, class methods) are handled; exotic syntax (decorators changing method shape, functions defined inside template-literal interpolations) may not be detected. Same philosophy as the Python side — no bundled parser.
• Internal package name still says context-lens-python. The user-facing display name no longer does, but the underlying Marketplace identity (eyobelg.context-lens-python) does — changing that would abandon the current listing and publish a brand-new one, so it's left alone.
• No CI. Compile/lint/test aren't enforced automatically on push.
• Lexical RAG only. Retrieval is term-frequency based, not embeddings — fine for small/medium repos, weaker on very large ones.
• Explain Issues diagnostics are replaced, not merged, per file — running it again on a different function in the same file clears the previous findings.