[nityanandmathur]

Add tensorViz under third_party/tensorviz and register it in the marketplace manifest and README. Agents can capture a local PyTorch nn.Module, inspect measured tensor shapes, display its graph, and refresh after a saved source edit.

This entry uses the public 0.3.61 preview, compiled from current product source main d4e2865a3c43149907dd06840745eb037d55dffa. It includes the readable stdio launcher, SHA-256-pinned runtime lock, workflow skill, logo, and integration notices. The 27-tool host includes find_computation, query_computation, and analyze_computation_change. The Cursor manifest follows this repository's schema and uses ${CURSOR_PLUGIN_ROOT}.

The runtime downloads from the public release on first connection. No runtime binary or private engine source is added here. The integration is MIT licensed; runtime components retain their applicable terms and notices. This preview targets macOS Apple Silicon, CPython 3.12, CPU, with no separate Node.js or VS Code installation and no tensorViz account or provider API key required. First-use permission covers local model execution and supported environment setup; Cursor permissions still apply. The included privacy, terms, support, and installation documents disclose external services, retained local data, platform limits, and the renderer trial's 26 December 2026 expiry.

Local validation:

• This repository's plugin/schema validator and generated harness-tree check passed. Component paths, documentation links, versions, shell syntax, and copied integration hashes passed.
• Release validation passed: 72 plugin units, 1,196 shared-host tests, 124 contract tests, protocol/boundary checks, 3,411 unique frontend tests, and 5,162 backend tests with three existing expected failures. The frontend's missing compiled test fixture was built before rerunning its affected suite; the backend rerun used the documented JAX/Flax dependencies.
• All seven compiled-release checks passed, including real Codex 0.165.0 and Claude Code 2.1.308 installation, runtime download, 27-tool discovery, synthetic CPU capture, measured-shape inspection, and removal. The compiled Canvas bridge and strict Claude manifest checks passed.
• The submitted Cursor launcher was checked in an isolated profile from a path containing spaces. Its unauthenticated runtime download matched the pinned size/checksum; initialization reported 0.3.61 and 27 tools. All three computation queries refused unopened models with NO_MODEL, without model execution or a permission grant.

Actual Cursor installation, capture, rendering, and update acceptance remain to be verified. Marketplace approval is separate. Exact public release provenance and Cursor evaluation steps are in third_party/tensorviz/docs/installation.md.

Public integration: https://github.com/tensorviz-org/tensorviz-plugins

Website: https://tensorviz.dev

Publisher contact: tensorviz.app@gmail.com