Reality Gate

This skill exists to stop: a coding agent's green test suite from being accepted as proof that the code works against the real database, real configuration, and real client — when the agent wrote the code, the fake, the fixture, and the test in the same turn.

🤖 0. HOW TO USE

Choose one mode:

• Audit (default when tests are green but real runs fail): run scripts/audit.sh [repo] --probe, confirm each RED by opening the cited files, map the last real-run bugs to the eight mechanisms, write the report to .jimmy/docs/reality-gate-audit.md (kit default), or to the report location the target repo's AGENTS.md already names.
• Gate: write or review acceptance criteria for a packet/PR with templates/acceptance-criteria.md. Seam changes get a reality AC; exit code 0 is not evidence.
• Build: stand up the in-repo regression replay suite (Workflow C) on the first seam change, then grow it one case per real-run bug.

Outputs: an audit scorecard (mechanism → evidence → bug it caused → fix), an AC block the reviewer reruns and pastes, and a tracked replay suite. Self-contained: copy the directory anywhere; it links to nothing outside itself.

1. Core claim

A green suite proves "code matches the agent's assumptions", not "code matches reality". Bugs live at the seams the agent assumed away. One rule closes the gap: