CodeQuest 2026 · Chaotic Crossing

🏆 First Prize — 2026 CodeCombat CodeQuest Hackathon

I won First Prize at the 2026 CodeCombat CodeQuest Hackathon.

I'm Jackson Tai. I'm sharing this collection of Chaotic Crossing arena bots so other contestants can study the strategies, run the code, and build their own approaches.

我在 2026 CodeCombat CodeQuest Hackathon 获得了 First Prize（一等奖）。 这里整理了国内与国际赛场使用和研究过的代码版本，供其他选手学习、运行与改进。

Also included: Algorithm Iteration Lab, an MIT-licensed agent skill for evidence-driven algorithm iteration with local simulation and a measured compute budget.

This is a collection of competition-era versions, not a claim that every file was the prize-winning submission. Historical ladder scores are not current performance guarantees.

Choose a version / 选择版本

File	Purpose
bots/cn/CN_final2.py	Later domestic snapshot: two-player role selection, goalkeeper feed cycles, passing and shooting. / 较后的国内版本。
bots/cn/CN_final.py	Alternative domestic snapshot: separate player roles and interception calculations. / 另一份国内版本。
bots/international/chaotic_crossing_bot_intl_v16.py	International v16 archive: side-aware positioning, shooting/pass evaluation, and throttled decisions. / 国际 v16 历史版。

The scripts are alternatives. Use one complete file at a time; do not concatenate them. Their original code is preserved, including historical quirks.

这些脚本互为备选。一次复制一份完整代码，不要拼接。 本次整理不改变算法。

Reusable workflow skill / 可复用工作流

The model proposes hypotheses and candidate code. The local machine runs the experiments. A trustworthy evaluator decides whether a candidate actually improves. More tokens or faster hardware cannot compensate for a wrong result parser or an inaccurate simulator.

algorithm-iteration-lab packages that workflow:

• Freeze a baseline and the evaluation inputs.
• Investigate a loss before changing tactics.
• Make one evidence-backed candidate and run the key regressions first.
• Expand to the acceptance set only after the gates pass.
• Measure CPU/memory needs before increasing match parallelism.
• Keep official-engine results, approximations, timeouts and unexecuted work distinct.

它不是“一键保证冠军”的提示词，也不自带官方引擎或自动抓取对手。模型负责找方向，本机负责做实验，评测负责判对错。

Install the skill

Copy skills/algorithm-iteration-lab into the skill directory used by your agent. For Codex, use $CODEX_HOME/skills when configured, otherwise ~/.codex/skills. Do not overwrite an existing customized copy without reviewing the changes.

PowerShell example, from this repository:

$skillRoot = if ($env:CODEX_HOME) { Join-Path $env:CODEX_HOME 'skills' } else { Join-Path $HOME '.codex/skills' }
$target = Join-Path $skillRoot 'algorithm-iteration-lab'
if (Test-Path -LiteralPath $target) { throw "Skill already exists: $target" }
New-Item -ItemType Directory -Path $skillRoot -Force | Out-Null
Copy-Item -LiteralPath './skills/algorithm-iteration-lab' -Destination $target -Recurse

Invoke it with a concrete algorithm and evaluator, for example:

Use $algorithm-iteration-lab to inspect my bot and local evaluator, reproduce the baseline loss, and validate a candidate within a 20-match budget. Do not submit anything online.

The bundled read-only probe needs Python 3 and no external packages:

python skills/algorithm-iteration-lab/scripts/resource_probe.py
python -B -m unittest discover -s skills/algorithm-iteration-lab/tests -p test_resource_probe.py -v

Supply measured per-match memory and runtime only after calibration. For example, if a measured match takes 20 seconds and a whole worker peaks at 512 MiB:

python skills/algorithm-iteration-lab/scripts/resource_probe.py --seconds-per-match 20 --peak-memory-mb 512 --matches 36 --workers 4

Those example numbers are not benchmark results. The probe does not launch a game, model or network request; its wall-time output is an estimate.

Quick start

1. Open the Chaotic Crossing arena in CodeCombat or the corresponding Chinese platform. Use the arena available to your account and competition.
2. Select Python in the game editor.
3. Open one of the files above and copy its complete contents into the editor.
4. Run a practice match. Check for errors and test both red and blue sides.
5. Adjust the strategy for the current arena rules before submitting it.

No pip or npm packages are needed to paste these bots into the game. The game supplies hero, params, player objects, and the spawn-mate event. These are game scripts, not standalone Python applications: python CN_final2.py will not start a match.

CN_final2.py registers its callback before the function declaration. That ordering relies on the game's Python execution environment; do not use ordinary CPython execution as a game-runtime test.

中文快速上手

1. 打开 CodeCombat／扣哒世界对应的 Chaotic Crossing 关卡，选择 Python。
2. 从上表选一份脚本，将全文复制到游戏编辑器。
3. 先运行练习对局，确认无报错，并分别检查红蓝两边。
4. 根据当前关卡规则调整，再提交比赛。

无需安装 Python 依赖。hero、params 和队员事件由游戏提供，不能在终端直接运行这些脚本来启动比赛。

Validation and known limits / 验证与限制

• All three published snapshots pass Python syntax compilation. This checks syntax only; it does not execute a bot or prove game compatibility.
• This publication does not include a fresh in-game match, live ladder evaluation, or tournament submission.
• Historical mechanics, runtime limits, and platform APIs may change. Test with the arena you are actually entering.
• The later domestic snapshot retains two known review findings: a whistle-distance condition repeats player 0, and one pass-safety condition mixes passer and receiver references. These have not been fixed or match-tested in this release.
• The international filename records v16; its opening comment still says v2. The source comment was left unchanged.

三份脚本均通过静态语法编译检查；本次没有重新运行官方对局或验证胜率。国内较后版本保留了两处已知索引问题，未擅自修复，也不将旧版本战绩套用到这些文件。

Optional local syntax check, with Python 3 installed:

python -B -c "from pathlib import Path; files=sorted(Path('bots').rglob('*.py')); assert len(files)==3; [(compile(p.read_text(encoding='utf-8-sig'), str(p), 'exec'), print('PASS:', p)) for p in files]"

This compiles source strings only. It does not import or execute the bots.

What's intentionally excluded / 未公开的内容

• Raw opponent-source collections and unrelated experimental variants.
• Captured sessions, account records, network payloads, ladder snapshots, and match logs.
• Downloaded game-engine bundles and the internal replay harness.
• Local agent settings, private notes, and the original internal handoff document.

The repository is an unofficial contestant project. It is not affiliated with or endorsed by CodeCombat. Third-party code retains its own attribution and applicable permissions; an award does not imply sole authorship of every technique or snapshot.

License and provenance / 许可与来源

The new workflow skill, helper, tests and repository documentation are MIT licensed; see LICENSE.

The archived bots are published with permission confirmed by the repository owner. They are not automatically covered by the MIT license: the domestic snapshots include third-party-derived code, and a named upstream license was not supplied for this release. See THIRD_PARTY_NOTICES.md for source context. Public availability alone does not establish unrestricted relicensing rights.

新编写的 Skill、工具、测试和文档按 MIT 开源。历史 bot 按仓库作者确认的许可公开，保留第三方归属，不擅自替第三方部分改签 MIT；需要进一步分发或改签时，应先明确相应授权条款。