Until

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 | dependency | data | assumption. You'll need this for escalation.

Each iteration

1. Checkpoint — before any risky change: a WIP commit, or git stash push ONLY in a tree where every piece of dirt is your own — never in a tree carrying a peer's/foreign dirt (a stash there silently swallows their in-progress work; the S38 parallel-build near-miss). A failed fix must be rolled back, not layered under the next fix. Compounding half-fixes is how loops corrupt the workspace.
2. Act — smallest change that tests the hypothesis.
3. Verify — run the pinned check for real; capture output. A pass must be reproduced once (re-run, clean state if cheap) before you trust it — flaky passes are how loops end in false victory.
4. Record — update the ledger + state file. Failed experiment? Roll back to checkpoint before iterating.

The escalation ladder — strategy shifts are forced, not optional

• 3 consecutive REFUTED in the same layer → that layer is probably innocent. Next hypothesis MUST come from a different layer.
• 5 consecutive REFUTED → your context is now part of the problem: you're anchored on dead hypotheses. Write a clean brief (symptom, check, ledger summary, what's ruled out — NOT your theories) and spawn a fresh-context subagent on it. Its naivety is the asset; do not contaminate the brief with your favorite suspect.
• 7 consecutive REFUTED → attack the goal itself: spawn an adversarial subagent to argue the check is testing the wrong thing, the bug is upstream of where you're looking, or the objective hides a false assumption. If it lands, re-pin WITH THE USER (this is the one allowed check change — explicit, never silent).
• Cap reached → write the handoff: state file finalized, brief sharpened, "what I'd try next" listed. A future /until resume — yours or another agent's — starts from iteration N, not zero.

Stop conditions — never declare false success

• ✅ Check passes, reproduced → report what worked + verifying output.
• 🛑 Thrashing: no new falsifiable hypothesis exists in any layer.
• 🛑 Cap (default 10).
• 🛑 User-blocked: needs credentials, a decision, money, or hardware only the user has — name it precisely.
• 🛑 Diminishing returns on a subjective/scored check — when the verifier is a graded judge/panel, not a binary pass, its score is ASYMPTOTIC: it plateaus and the all-high bar may never cleanly fire. Track the score delta per iteration; when it flattens (~2 iterations of marginal gain), STOP and surface the plateau + best result + any anchor-conflict as a fork for the user. Cap the expensive judge cadence to every 2-3 iterations, not every one. (Banked from a /until loop that ran ~10 panels / ~3.7M tokens against an asymptotic bar.)

Final report

✅ or 🛑 · the verified evidence (real output, not paraphrase) · iterations used · the ledger · state-file path. On 🛑: the exact resume command.