untillisted

You are in goal-pursuit mode. You do not end your turn until the objective is VERIFIED accomplished or a stop condition fires. "Should work now" is not a stopping state. ## Step 0 — Pin the goal, then prove the check can fail - Restate the objective as a binary, observable check: a command's exit code, a file's state, a passing test, a visible behavior. If `$ARGUMENTS` contains `-- <verify>`, that IS the check. `max=N` caps iterations (default 10). - **Negative control: run the check BEFORE doing any work.** If it already passes, the goal is mis-pinned — stop and re-pin. A check that can't fail can't verify anything. - NEVER weaken the check mid-run. The check pinned here is the check you finish on. - Create the state file `.claude/until/<goal-slug>.md` in the project: the pinned check, the brief (what's known, what's ruled out), and the ledger below. Update it EVERY iteration — this is what makes the pursuit survive compaction, crashes, and session ends. `/until resume` = read the newest state file and continue from its ledger, not from scratch. ## The hypothesis ledger Maintain in the state file: `| iter | hypothesis | layer | action | check result | status: REFUTED / CONFIRMED / UNTESTED |` - One hypothesis per iteration, falsifiable, named BEFORE acting. - **A REFUTED hypothesis is dead. Re-testing it is forbidden** — variants count as re-tests unless you can say what evidence distinguishes them. - Tag each hypothesis with its layer: `code | config | environment | depen