Forge playbook — Games (browser · Godot)

A game is an interactive-loop domain — "done" means a human (or a driven browser session) can actually play it start-to-finish, not that a screen renders. HUD/menus and responsive layout defer to forge-website / ui-boss; this file governs the playable loop, stuck-state safety, input/viewport coverage, and honest playtest evidence. The core honesty rule: never call a game playable without real evidence of a completed loop (win and lose and restart reached). Opt-in limit: browser games can be driven + screenshotted directly with the browser tool; Godot needs the Godot toolchain (editor/export), which is opt-in and owner-side — if it isn't installed, mark export/clean-run steps not-run and report only what was actually driven.

Hard rules (non-negotiable)

• A genuinely playable loop. The core loop runs end-to-end: input → state update → render → feedback, reaching a win and a lose condition and a working restart. Not a static mockup, not a title screen with a dead "Play" button.
• No stuck / soft-lock states. No dead-end where the player can't progress or restart; game-over and win are always reachable and recoverable; pause/resume works; input never stops responding. A soft-lock is a shipping-blocker, not a polish item.
• Cross-viewport + declared inputs. The canvas scales/letterboxes correctly at mobile and desktop sizes with no cut-off UI — or the game explicitly declares desktop-only. Every declared input method actually works (keyboard, mouse, touch, gamepad as appropriate; Godot: multiple resolutions/aspect ratios via a stretch mode).
• Honor the dependency constraint. If the brief says vanilla / no-framework / zero-dependency, honor it — no CDN pulls, self-contained assets. If the game ships as a Forge artifact, it must be self-contained under the artifact CSP (inline JS/CSS, assets as data URIs — no external hosts).
• Runtime health. Zero console errors during a full play session; a stable frame rate using a delta-time loop (requestAnimationFrame with elapsed time; a fixed timestep for physics) rather than assuming 60fps; and cleanup on restart/scene change — listeners, intervals, timers, and audio nodes are removed so replaying doesn't leak or double-fire.
• Persistence works if promised. If save/load or high-scores are in scope, they actually persist and restore correctly across reloads.
• Real playtest evidence. The loop is proven by actually playing it (or driving it via browser automation) with screenshots/recording showing win + lose + restart — evidence, never an assertion.

Team (conditional)

Lead: build-boss (a game is implementation-heavy). HUD / menus / responsive canvas + visual quality: ui-boss (with forge-website for layout). QA: test-boss — drive the loop via browser automation (win / lose / restart reachable, no stuck states, cross-viewport, no console errors). Code review: typescript-reviewer for browser JS/TS (honest gap: no dedicated Godot/GDScript reviewer agent exists — Godot script review is build-boss + general review). Runtime efficiency (frame budget, asset weight): seo-boss (its non-web remit is runtime performance). ml-engineer / data-scientist only if the game has procedural-generation or ML-driven content (opt-in). Optional advisor: codex-reviewer on the core loop/state code.

Skills / commands / MCP

browser is the key evidence tool — drive and screenshot the running browser game to produce real playtest proof (win/lose/restart, no console errors). systematic-debugging for stuck states, collision/physics bugs, and frame drops. forge-website for menus/HUD and a responsive canvas. Engine/API specifics via Context7 (Phaser, Godot/GDScript). Opt-in: the Godot editor/export toolchain is owner-side and not provisioned by Forge — browser games need no such dependency and can be driven directly.

Fan-out & flow

L2 for a single-mechanic browser game; L3 for a multi-level / multi-system game (physics + AI + save/load + audio). Serial: core loop (input→update→render) → win/lose/restart states → feedback/juice → cross-viewport + input coverage → playtest evidence. Parallel: independent systems (rendering ∥ input ∥ audio ∥ level data) and independent levels/assets — independent once the game-state contract is fixed (the state model is the contract).

Domain gates

• Core loop runs and completes: start → play → win AND lose → restart, all reachable; no soft-lock; pause/resume works.
• Canvas scales/letterboxes correctly at mobile + desktop sizes (or desktop-only is declared); every declared input method works.
• Zero console errors across a full session; stable frame rate via a delta-time loop (fixed timestep for physics); listeners/timers/audio cleaned up on restart (no leak, no double-fire).
• If a zero-dependency constraint was set, no external/CDN deps; self-contained (and CSP-safe if delivered as an artifact).
• Save/load / high-scores (if promised) persist and restore across reloads.
• Real playtest evidence attached (screenshots/recording of win + lose + restart) — not asserted.

Ship-readiness (unique)

Playable loop proven end-to-end with real playtest evidence (win + lose + restart actually reached); no stuck states; works at mobile + desktop viewports and for every declared input; zero console errors; stable frame rate + cleaned-up listeners; the zero-dependency constraint honored if set; save/load verified if promised. If the game is a Godot project and the Godot toolchain wasn't available, mark export + clean-run as not-run and report only what was actually driven — don't imply an export that didn't happen. Advisory checklist; optionally run codex-reviewer on the core loop/state code — not a blocker.