Improving a Forge game until done is READY

Overview

Read the shared game production guide from the mubithack checkout. Apply its priority bands, relevant genre/capability checks and verified-outcome rules. All AI decisions and generated fixes/checks must use Mubit-backed tasks.

done(slug) is the implemented readiness bar. Read its current parts rather than assuming this list is exhaustive. They include:

• checks: every executable check passes;
• features: every agent-added feature is verified;
• open: the last fix cycle left nothing high or medium;
• critic: the latest score is ≥ 8 (n/a for seeded games).

Choose the lever for the missing evidence. Repeating improve_game blindly can leave feature and coverage gaps.
The current bar also includes experience and interaction evidence. A green bar does not waive additional requirements for the user's target stage in the shared guide.

The loop

1. done(slug): write down the missing parts. That is your baseline.

2. Per missing part:

   Part	Lever
   checks: "no check run yet"	mubit-tester test_feature("<core loop>", game=slug, playtest=false) runs every saved check and records the results that done reads
   checks: failing ids	improve_game(slug), or fix that file. Then test_feature(..., game=slug) again
   checks: no checks exist	use the Mubit-backed scenario planner: .venv/bin/forge checks <slug> --plan --run --studio <bound-studio>, then inspect coverage and results
   features: not verified	test_feature("<that feature>", game=slug) until PASS
   open	improve_game(slug, studio=...). It follows the job and ends with the done bar
   critic < 8	iterate(slug, iterations=2, studio=...), or feature(slug, request) for what the critic says is missing

3. After each lever, done(slug) again. Compare it with your baseline.

4. Stop when:

   • READY;
   • the [ ] lines of done (part + detail) are word-for-word the same twice in a row (a
     plateau): report what's stuck and the evidence;
   • you've done 6 rounds.

   "Don't stop to ask" means don't ask. It doesn't mean loop forever.

5. If a round made things worse (a check went PASS → FAIL): diff_since_last_run (mubit-tester),
   then bisect_versions(slug, check), then rollback(slug, version).

Rules

• Pass studio="<your local place>" to improve_game, iterate and feature. Never use a
  placeId window.
• Run one job per Studio window at a time.
• Human-owned files get patches, not edits (runs/<slug>/improve/patches/): show them, don't
  claim they're applied.

Report format

<slug>: READY after <n> rounds | NOT READY (plateau | budget): <parts missing + details>
Baseline → now: checks a/b → c/d, open x → y, critic s → t
Rounds: 1) <lever> → <effect> 2) ...
Evidence: job ids, test ids, versions kept

Common mistakes