Roadmap

What's next, and how far along each thing actually is — "coming soon" on its own tells you nothing about whether something is a week away or waiting on a third party.

For what already works, see Agent capabilities and the Changelog.

In beta

MCP server. whodunit-mcp exposes the dashboards' database to any MCP-speaking agent over stdio, read-only. list_metrics, run_metric and describe_schema are implemented and tested against a live instance; known-empty detection and the release gate are not built yet. Details are in the whodunit-mcp repository.

Full attribution for Cursor and Copilot

Both are read today at declared, from the trailers they write themselves. Both can reach intersected — the same evidence level as Claude Code — because each records the lines it wrote to a local store whodunit can read. The data is there; the adapter isn't written yet.

Cursor. Its session store keeps each edit's literal text, the file path, and the working directory that scopes it to a repository. It also records which model ran, whether the edit succeeded, and whether the work came from an agent run or inline autocomplete.

GitHub Copilot. Its edit records live in VS Code's workspace storage, each carrying the inserted text and an absolute file path. Per-turn timing comes with them — among supported agents, only Codex reports that today.

Neither records token or cache figures, so both will be excluded from cost denominators rather than counted as free, the same treatment Antigravity gets.