lockin

# /lockin Lock in on one goal and keep working until it is provably true. No stopping for check-ins, no claiming done without evidence. ## 1. Lock the goal Turn the request into a done-condition with four parts: | Part | Example | | --------------- | --------------------------------------------------------------------------- | | **End state** | All tests in `test/auth` pass and lint is clean | | **Proof** | `npm test test/auth` exits 0; `npm run lint` reports 0 errors | | **Constraints** | Don't touch the public API in `src/auth/index.ts`; don't edit or skip tests | | **Cap** | Stop after 25 iterations (default if the user gives none) | ### Examples **Fix failing tests** > "get the auth tests passing" ``` LOCKED IN End state: All tests in test/auth pass and lint is clean Proof: `npm test test/auth` exits 0; `npm run lint` reports 0 errors Constraints: Don't touch the public API in src/auth/index.ts; don't edit or skip tests Cap: 15 iterations ``` **Fix a bug** > "checkout crashes when the cart is empty" ``` LOCKED IN End state: An empty cart shows "Your cart is empty" instead of throwing Proof: New test `checkout.empty-cart.test.ts` fails before the fix and passes after; full suite green Constraints: No changes to payment flow files in src/payments/ Cap: 15 iterations ``` **Type errors** > "turn on strict mode" ``` LOCKED IN End state: tsconfig has "strict": true and the project type-checks Proof: `npx tsc --noEmit` exits 0 with 0 errors Constraints: No `any`, `@ts-ignore`, or `@ts-expect-error` added; no runtime behavior changes Cap: 50 iterations ``` ## 2. Stay locked in Each iteration: 1. **Plan:** name the single next step that moves toward the end state. 2. **Act:** edit, run, investigate. 3. **Prove:** run the proof commands. Paste the real output (exit codes, test summary, error counts) into the transcript. Unshown evidence doesn't count. 4. **Judge:** one line: `[n/cap] NOT MET — <reason>` or `[n/cap] MET — <evidence>`. Rules: - **Don't break focus to ask** unless you're blocked on a decision that can't be undone, or on missing access or info the codebase can't supply. - **Never weaken the goal.** Deleting, skipping or loosening tests or checks to make them pass is a failure, not progress. - **If the same failure shows up 3 iterations in a row,** change approach. Re-read the error, question your assumption, look wider in the codebase. - **Respect the constraints.** Run `git diff` before judging MET to confirm nothing out of scope changed. ## 3. Verify independently Before declaring MET, if subagents are available, spawn a fresh one that hasn't seen your work. Give it only the `LOCKED IN` block, then have it re-run the proof commands and check the diff against the constraints. If it disagrees, go back to the loop with its reason. ## 4. Unlock - **MET:** report the goal, the final proof output, files changed, and the iteration count. - **Cap reached or blocked:** report what's done, what's still failing (with output), what's blocking, and a suggested next `/lockin` goal. Never unlock on MET without proof from the final iteration.