Why it exists
Agents say "done" when it isn't
A step passes only when an independent check proves it: tests run, page loads, source says so. The check is written before the work, and a fresh reviewer decides, not the worker.
There are too many tools
Thousands of skills, MCP servers and repos, some malicious. Stillwright picks by evidence, reads the code before trusting it, asks you once, and pins it by hash so it can't change silently.
Agents talk too much
One inbox with four card types: ask, review, notify, idea. Silence is the default.
What you see
Every run leaves a plan with the check per step and a log with each check's output, as plain files you can read without the tool. When it needs you, you get a card like this:
Measured, not assumed
Three small tasks (code, writing, research), each done by a fresh Claude Sonnet agent with and without the Stillwright skills, graded by scripts and a read-only judge. Then the same tasks priced in dollars from Claude Code's own receipts.
On tasks this small it does not make the agent more capable or cheaper. It makes the work checkable and keeps the agent from claiming things it didn't do. The raw runs, the grading script and the judge's report are in the dataset, so anyone can re-grade or dispute them.
