厨房战争原型 (Kitchen War RTS)

Overview

Deliver and iteratively improve a complete, single-file HTML5 RTS game that
mimics the Red Alert (红警) core loop with a kitchen-war theme. The game lives in
assets/index.html — one self-contained file (Canvas + a side HTML/CSS panel,
zero external dependencies). To give the user the game, copy assets/index.html
into their workspace and open it in the built-in preview via present_files.

This skill also encodes the hard-won quality bar: the game must be verified
with a real browser end-to-end test that drives the actual game loop — not a
transient/one-frame check — because multi-frame interaction bugs (e.g. the
"moved a few steps then froze" deadlock) are invisible to instant tests.

When to Use

• User requests a Red Alert-style / playable browser RTS game.
• User says "做个能玩的战略游戏 / 红色警戒那样的游戏 / 超越红警".
• User wants to create, fix, extend, or re-verify the Kitchen War game.
• After ANY edit to the game, run the verification gate before declaring done.

Delivering the Game

1. Copy assets/index.html to the user's workspace (e.g. 桌面 or current project).
2. Present it with present_files so it opens in the live preview panel.
3. Tell the user how to play: left-click select / drag box-select, right-click to
   move or attack, B/I/V switch build tabs, S stop, M music, Esc cancel; click
   "烤鸭空袭" (superweapon) when charged, then click a target.

Game Feature Set (what "done" means)

The game is a faithful Red Alert-shape RTS. Confirm these exist after changes:

• Base building: 中央厨房(yard) → 配电箱(power) → 冰箱仓库(refinery) →
  备菜台(barracks) → 灶台工厂(warfactory) → 科技烤箱(tech) → 微波塔(turret).
  Placement uses a green ghost with snap-to-valid-tile; buildings must touch an
  existing completed building.
• Economy: 采集车(harvester) mines resource tiles → returns to refinery → cash.
• Production: build queue per building + rally points; veterancy (promote on kills).
• Tech tree: 科技烤箱 unlocks advanced units.
• Factions: 灶火盟 (player) vs 冰锋营 (enemy, tougher) — asymmetric stats.
• Fog of war + reveal-on-sight; radar minimap.
• Selection/commands: single select, drag box-select, right-click move /
  attack / attack-move; group commands apply to all selected units.
• Combat: projectiles, hit flash, death particles, veterancy damage/speed multipliers.
• Enemy AI: builds economy, produces army, and launches attack waves at the
  player's base (verified: units switch to attackMove toward player buildings).
• Superweapon: 烤鸭空袭 charges over time → button → click target → 2.4s
  incoming warning → area damage (700 units / 1100 buildings).
• Win/Lose: destroying the enemy 中央厨房 = victory; losing yours = defeat.
• Tutorial: 5-step onboarding; sound: WebAudio blips + simple music (M).
• Engagement & retention layer (the hooks that keep a new player around):
• First-visit help overlay (H or ? button) — explains goal, controls, hotkeys,
  and tips. Dismissed once, accessible any time. Shown automatically on first load.
• Top toolbar — always-visible buttons for 帮助 / 速度(1x·2x·3x) / 暂停 /
  音效 / 配乐 / 截图 / 成就. All toggles persist to localStorage.
• SFX module (procedural WebAudio on the existing beep engine) —
  build / produce / explode / promote / capture / superweapon / select / ach /
  victory / defeat. Toggle with X or the toolbar button. Persisted.
• 10 achievements persisted to localStorage with toast + in-game panel
  (🏆 成就): first win, speedrun (<5 min), low-HP survival, 10000 cash,
  30 units produced, 10 buildings built, unit → vet-2, engineer capture,
  superweapon kill, win on all 3 maps.
• Screenshot/share (📷 button or F) — exports the canvas as a PNG with
  map / difficulty / cash / time watermark. Drop straight into a share.
• Hotkeys — 1-6 quick-select first 6 build items, Space / P pause,
  H help, X SFX, M music, F screenshot, Esc cancel, plus the
  legacy B / I / V tabs and S stop.
• Game speed (1x / 2x / 3x, scales loop dt) + pause (skip updates,
  show "⏸ 已暂停" tag) — power-user and accessibility controls.
• Personal best + share-on-end (the "surpass Red Alert" differentiator):
  on game over the screen shows 个人最佳：最快通关 / 最高食材 / 累计胜场,
  persisted to localStorage (kw_best); a 分享战绩 button (calls the
  screenshot export) lets the player brag a win straight to socials — RA never
  had persistent personal records or one-click share.
• 每日挑战 (Daily Challenge) — the 2026 browser-game retention hook. The
  📅 今日挑战 button seeds the map with the day's date (todaySeed()), so
  everyone who plays on the same day gets the identical, reproducible map +
  resource scatter (deterministic via mulberry32 + scatterResources).
  Natural daily competition without any server.
• 挑战分享码 (Share Code) — every match mints a short KW-<seed36>-<map><diff>
  code (makeCode/parseCode); the game-over screen and the screenshot watermark
  both carry it. A friend pastes the code → parseCode → startGame reproduces the
  exact same challenge (audit_challenge.js proves round-trip + replay). Local
  localStorage keeps a per-code best time, so players chase faster clears.
• 结算屏引流 CTA — on victory the screen shows the challenge code, the per-code
  best time, and a "去 SkillHub 搜 kitchen-war-rts 给个五星⭐" share prompt, turning
  every win into organic spread (the promo plan's "让用户替你传播" lever).

Quality Gate (MANDATORY after any change)

A change is NOT done until scripts/verify_game.js runs green with 0 runtime
errors. This test drives the real game loop (not internal stubs):

• starts the game via the real start button,
• builds a structure through the real side-panel button + ghost placement,
• produces and box-selects units, issues a group attack,
• forces and confirms the enemy AI launches an attack wave,
• destroys the enemy yard through real combat and confirms "战役胜利",
• fires the superweapon and confirms area damage,
• asserts zero pageerror / console.error throughout.

The full real-browser e2e suite (≈83 assertions across 6 scripts) must all
pass with 0 console errors / warnings / pageerrors before declaring done. The
only known non-green line is _probe.js's "拆除微波塔" check, which is a
test-harness artifact (Playwright's synthetic click at that viewport doesn't
land a valid turret tile); the underlying game logic is proven correct by
_power_logic.js (place turret → powerUse 40→90, sell → 40, no crash) and by
_audit2.js (places 4 turrets → powerUse 280). Treat every other failure as a
real bug.

Run command (managed Node + Playwright):

cd <skill>/scripts
export PLAYWRIGHT_BROWSERS_PATH="$LOCALAPPDATA/ms-playwright"
export NODE_PATH="C:/Users/www74/.workbuddy/binaries/node/workspace/node_modules"
"C:/Users/www74/.workbuddy/binaries/node/versions/22.22.2/node.exe" verify_game.js

The script auto-resolves the game at ../assets/index.html.

Why a real-browser gate (not a logic-only check)

Instant Playwright clicks that land inside one frame pass even when the real
game is broken: the sidebar used to rebuild its buttons every frame, so a human's
cross-frame click never triggered onclick; and unit separation once exactly
cancelled per-frame movement, freezing units after a few steps. Only a test that
samples per-frame state across many frames catches these. See references/testing.md.

Editing the Game

• The game is one file; edit assets/index.html directly.
• Key invariants to preserve (regression traps):
• Movement reads the top-level u.speed / u.slowTimer (set by spawnUnit),
  NOT u.def.speed. Units created anywhere must include these fields.
• applySeparation push is capped well below per-frame movement so units
  never deadlock behind allies.
• updateCombatUnit must keep making progress (repath cooldown + seekDirect
  fallback) so a findPath failure never leaves a unit permanently stopped.
• Victory triggers in damage() when a yard is destroyed.
• NaN coordinate propagation is a silent killer. If any unit x/y,
  destX/destY, or mouse-derived coordinate becomes NaN (e.g. via a
  clamp call, since clamp(NaN,a,b) returns NaN with the naive
  v<a?v:a:v>b?b:v form), it spreads: tileBlocked(NaN,NaN) then does
  G.map[ty][tx] → reads undefined → throws every frame, crashing the
  whole game loop (symptom: "worked for a few moves then broke"). Guard rails
  in place (do NOT remove): clamp coerces v!==v → a; clientToGame
  guards r.width/r.height===0; tileBlocked/commandUnits/moveDirect/
  moveAlongPath/seekDirect/ensureMove all early-return on non-finite
  inputs. Keep them.
• clearBaseZone must preserve ALL T_RES tiles (not just those outside
  the base's bbox). Each map deliberately places the starting resource patch
  adjacent to the base (so the harvester mines immediately), and the patch
  overlaps the base bounding box. The early form
  if(T_RES && !insideFootprint) continue wiped the near-base resource, so
  the harvester had to cross the entire map → first delivery ~30s instead of
  ~5s, devastating the early economy. Fixed: if(T_RES) continue;.
• updateUnits must also purge dead units from G.selUnits, not only
  from sd.units. Otherwise a dead unit stays in the selection set, its
  ghosts render, and stale references leak into the order/move pipeline.
• updateCombatUnit must drop a target whose side matches the attacker's
  side (e.g. an engineer-captured building flips to player and the
  remaining attackers must not keep "attacking" their own building). The
  re-acquisition condition is
  !u.target || u.target.hp<=0 || (u.target.side && u.target.side===u.side).
• Re-run the gate after every edit.

Publishing / Marketplace Listing

两套独立的系统，不要混淆（2026-09-08 实测结论）：

1. clawhub.ai 注册表（用户 clawhub install 拉取的来源）：用 clawhub publish 更新。
2. SkillHub 商店前台（skillhub.cn / skillhub.cloud.tencent.com，用户浏览的目录）：
   clawhub publish --site https://skillhub.cn 只写 clawhub 兼容注册表镜像，并不会在
   SkillHub 商店前台建/更新条目。 商店前台是独立 catalog，用独立的 skh_... token
   （存于 ~/.skillhub/credentials.json，不是 clawhub 的 clh_... token）。

本技能在 SkillHub 商店的真实条目 slug 是 war-skill（owner user_98676efe），
不是 kitchen-war-rts。用同一 slug + 更高 version 提交即为"更新到新版"。

发布到 clawhub.ai（注册表）：

clawhub publish "C:/Users/www74/.workbuddy/skills/kitchen-war-rts" ^
  --slug kitchen-war-rts ^
  --name "厨房战争：红警风RTS" ^
  --tags "rts,game,html5-game,browser-game,strategy,red-alert,canvas,entertainment,mini-game,红警,战略游戏,厨房战争,每日挑战,挑战分享" ^
  --changelog "兵种外观全面重做——10个兵种各有独特剪影+专属兵种色+阵营环(红蓝)，彻底解决战车/士兵外观雷同难分辨问题；修复自行火炮(高压锅炮车)贴脸卡死BUG(太近自动后撤拉开射程)；修复生产取消/拆除建筑两处经济退款精度漏洞(按实际折扣价退还)；62项真机门禁全绿。"

发布/更新到 SkillHub 商店前台（必须走 store API，不能用 clawhub publish --site）：

# 用托管 Node 运行本技能自带的脚本（自动读 ~/.skillhub/credentials.json 的 skh_ token）
"C:/Users/www74/.workbuddy/binaries/node/versions/22.22.2-2/node.exe" ^
  "C:/Users/www74/.workbuddy/skills/kitchen-war-rts/scripts/publish_skillhub_store.js" 1.0.4

真实接口（从 skillhub 网页前端 bundle 反查确认）：
POST /api/v1/community/skills/publish，multipart：payload(JSON {slug,displayName,version,summaryZh?}) + files[](每个文件 filename=相对路径)，Header Authorization: Bearer <skh_token>。
提交后 reviewStatus/contentAuditStatus/securityScanStatus 进入 pending，审核通过后自动翻成最新版（与 clawhub 的 pending-publication 机制一致，属正常上架流程）。

> 注：clawhub 拒绝二进制，发布前把 listing-assets/ 临时移出技能目录；封面/图标/预览图在
> SkillHub 控制台单独上传（非技能包内容）。

Then open the SkillHub web console (skillhub.cloud.tencent.com / skillhub.cn) and attach the
listing visuals from listing-assets/. These are NOT part of the skill package —
clawhub rejects binary files, so they must be uploaded through the console UI, not the
publish command:
- 封面：listing-assets/cover.jpg（1240×808 红警风厨房战场）
- 图标：listing-assets/icon.png（256×256 厨师帽红星坦克徽章）
- 预览图：listing-assets/gameplay_preview.jpg（1280×769 真实游戏画面）

Listing 文案要点：
- 一句话生成可玩的红警风 RTS，单文件 HTML，零依赖，即开即玩。
- 触发词示例：做一个红色警戒游戏、生成 RTS 战略游戏、厨房战争、能玩的网页游戏。
- 差异化：10 项成就、个人最佳战绩、截图分享、1x/2x/3x 速度、暂停。

Resources

• assets/index.html — the game (deliverable).
• listing-assets/cover.jpg / icon.png / gameplay_preview.jpg — marketplace
  listing visuals (upload via the SkillHub console; NOT included in the agent package).
• scripts/verify_game.js — mandatory real-browser e2e quality gate (29 asserts).
• scripts/_audit2.js — expanded regression suite (22 asserts).
• scripts/_smoke_features.js — engagement-feature smoke (11 asserts).
• scripts/_probe.js — adversarial probe (pause/3x/super-corners/resize/double-start/friendly-fire…).
• scripts/_power_logic.js — turret place/sell powerUse logic proof.
• scripts/_best_test.js — personal-best + share-on-end proof.
• references/testing.md — the verification methodology and known bug classes.