groundworklisted

# groundwork Write down what you believe about the repo, with the command that proves each belief, before any production code exists. Then hand the record to a fresh agent whose only job is to find what is absent from it. ## Why this exists Most bugs a coding agent writes are not reasoning errors. They are two things: 1. **A claim about the repo that was never checked.** "The handler already validates this." It did, in the file the agent read forty minutes ago, on a path that is not this one. 2. **A second path that was never enumerated.** The change is correct for the caller in the diff and wrong for the cron job, the admin action, the migration, or the test double that also calls it. Both are absences. Neither is visible in the code the agent produces, which is why they survive a code review of that code — the reviewer reads what is there. They are visible in the *record*, because a record is a finite list and a list can be checked for missing rows. Measured on one repo's 51 review-bot findings: about a fifth were unverified claims about the repo, and about a third were an unenumerated second path. Both classes are cheaper to catch before the code exists than after. **The economics: reviewing a record costs one agent call. Finding the same defect after the push costs a review cycle, a red gate and a fix commit.** ## When to run this Run it when the change is non-trivial: it touches a shared function, a schema, a public contract, money, auth, or anything