arena
Put several agents on one task. The default arena mode runs parallel candidates and grafts the best parts into one result, and benchmark ranks models or skills with a blind judge. Runs only when the user types /arena.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-13
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- arena
- metadata
- { "opencode/autoinvoke": false }
- description
- Put several agents on one task. The default arena mode runs parallel candidates and grafts the best parts into one result, and benchmark ranks models or skills with a blind judge. Runs only when the user types /arena.
- disable-model-invocation
- true
02 资源文件 8
- 参考文档agents/openai.yaml96 B
- 其他includes.toml724 B
- 参考文档modes/benchmark.md7.0 KB
- 参考文档references/pstack-arena.md6.8 KB
- 参考文档references/pstack-principle-laziness-protocol.md543 B
- 参考文档references/pstack-principle-prove-it-works.md2.4 KB
- 参考文档references/pstack-principle-redesign-from-first-principles.md749 B
- 参考文档references/pstack-principle-separate-before-serializing-shared-state.md1.5 KB
03 SKILL.md 正文
arena
Use several agents where one attempt or one opinion is not enough.
Pick the mode
The first word of the argument picks the mode. The rest of the argument is the task.
- arena (the default): several candidates make the same artifact in parallel. You pick a base and graft the best parts of the others into it. Use it to make one thing when a single attempt could commit to the wrong design.
- benchmark: contenders do the same task, a blind judge scores them, and an HTML report ranks them. Use it to compare models or skills. It does not make a deliverable.
With no mode word, choose from the request. A comparison of models or skills runs benchmark. Anything else runs arena. Name the mode in your first line.
A review or a second opinion on existing work is not an arena task. Tell the user to type /council.
Models and harness
These rules apply in every mode.
- Read role lines from
agent-models.mdin${CLAUDE_CONFIG_DIR:-$HOME/.claude}/when that file exists. Without it, use the defaults in the mode's file. - Route every seat with the Routing rules in the Models section of
../build/SKILL.md: the|fallback choices, the route for each harness, and the effort field. In T3 Code, loaddelegate_taskandorchestrator_capabilitiesfirst if they are deferred tools. - Set each agent's reasoning effort for its job. Go higher for hard design or review work. Go lower for work that is mostly generation.
arena mode
Follow references/pstack-arena.md. It names pstack files and skills. h00man is standalone, so use these in their place:
- Its platform-mapping links and the poteto-mode Subagents section: the rules under Models and harness above.
- separate-before-serializing-shared-state:
references/pstack-principle-separate-before-serializing-shared-state.md. - redesign-from-first-principles:
references/pstack-principle-redesign-from-first-principles.md. - prove-it-works:
references/pstack-principle-prove-it-works.md. The show-me-your-work skill it names:../build/references/show-me-your-work/SKILL.ref.mdwhenbuildis installed. Either way, keep the proof as a local file next to the artifact. - The Laziness Protocol:
references/pstack-principle-laziness-protocol.md. /setup-pstackin its Models section: not in the stack. Edit the role lines inagent-models.mdby hand.
Read each principle file at the phase that names it.
benchmark mode
Follow modes/benchmark.md. It is the user's own method. The words after benchmark are the argument its setup step parses, for example /arena benchmark skill deep.
Sources
- h00man, by h00man23 (https://github.com/h00mankind/workflow, Apache-2.0): the earlier
h00man:benchmarkskill, nowmodes/benchmark.md. It gives the benchmark mode: calibrated tasks, hermetic runs, the blind judge, and the report. - pstack, by Lauren Tan (https://github.com/cursor/plugins/tree/main/pstack, ported at https://github.com/h00mankind/pstack, MIT): the
arenaskill, copied toreferences/pstack-arena.md, for the arena mode. The principle skills it names, copied toreferences/pstack-principle-*.md.
04 同仓库更多 Skills
teach
Teach the user a subject, in explain mode (one plain account of code, a change, or a concept, in chat) or course mode (a learning workspace with a mission, lessons, and records over many sessions). Runs only when the user types /teach.
council
Get a review or a second opinion on existing work from several strong models of different providers. Runs only when the user types /council.
variant
Builds multiple variants of a component you're working on and helps you iterate and pick one.
teardown
Creates technical illustrations of physical objects and of how things work, such as display pixels or interface layers, with coherent views, fine linework and part fills.
state-machine
Renders every state of a component you choose on a throwaway page, with mock data and a switcher, so you can work on each state.
interface-review
Reviews a branch, pull request or uncommitted change for the interface problems it introduced or regressed, across accessibility, layout, writing, typography, color and UI.