cursor's row is graded live-run-partial against 3.17.8 — someone watched it run, but
not across every event the matrix claims. The row is honest about that; it is just incomplete.
Cursor matters more than most here: it is the only agent in the matrix graded enforceable
at pre_tool, which makes it the strongest claim the project makes about anyone. A claim that
strong deserves a complete run behind it.
Completing it means running the enforcement probe rather than only the capture probe, so the row
can say what happens when a hook denies and what happens when a hook dies:
python3 tools/experiment.py run --agent cursor --report \
--agent-version <version> --reporter @yourhandle > report.json
python3 tools/verify_report.py report.jsonverify_report.py shows you exactly what a maintainer will see before you send it.
Read CONTRIBUTING.md's warning about this probe first: unlike the capture probe, the
experiment probe denies, crashes and stalls on purpose. It only ever runs in a throwaway
directory the harness creates and removes. Never point it at a config you work in.
Requires the agent installed. No code required beyond running it.
cursor's row is graded live-run-partial against3.17.8— someone watched it run, butnot across every event the matrix claims. The row is honest about that; it is just incomplete.
Cursor matters more than most here: it is the only agent in the matrix graded
enforceableat
pre_tool, which makes it the strongest claim the project makes about anyone. A claim thatstrong deserves a complete run behind it.
Completing it means running the enforcement probe rather than only the capture probe, so the row
can say what happens when a hook denies and what happens when a hook dies:
python3 tools/experiment.py run --agent cursor --report \ --agent-version <version> --reporter @yourhandle > report.json python3 tools/verify_report.py report.jsonverify_report.pyshows you exactly what a maintainer will see before you send it.Read
CONTRIBUTING.md's warning about this probe first: unlike the capture probe, theexperiment probe denies, crashes and stalls on purpose. It only ever runs in a throwaway
directory the harness creates and removes. Never point it at a config you work in.
Requires the agent installed. No code required beyond running it.