code-arena
/code-arena
Uruchamia kilka równoległych prób tego samego zadania w osobnych worktree i wybiera wersję z najwyższą oceną.
Ocena i bezpieczeństwo
- Pliki w paczce
- 3
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 12.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Uruchamia kilka równoległych prób tego samego zadania w osobnych worktree i wybiera wersję z najwyższą oceną.
Argumenty
/code-arena [n] <task>
Oryginalny opis (EN)
Run N candidates at one coding task in parallel, each in its own git worktree, score them against a rubric built from h2xd:code-quality and the tests, pick a base, and graft the best parts of the others into it. Use when the user asks for an arena, or for several competing attempts at a non-trivial change whose shape one attempt would lock in.
Jak działa
- Steps
- Report
- Do
- Don't
Podgląd SKILL.md
--- name: code-arena description: Run N candidates at one coding task in parallel, each in its own git worktree, score them against a rubric built from h2xd:code-quality and the tests, pick a base, and graft the best parts of the others into it. Use when the user asks for an arena, or for several competing attempts at a non-trivial change whose shape one attempt would lock in. argument-hint: "[n] <task>" disable-model-invocation: true --- # Arena Give one task to three engineers and you see which choices were real, and which were just the first idea that worked. An arena runs N attempts, each in its own git worktree, and builds the result from the strongest one plus the best ideas of the rest. It costs N full builds, so only the user starts it. `$WORK` is a scratchpad directory, and every hand-off is a file in it. Run one command per shell call. Chain only stage and commit, as `h2xd:git-commit` does. ## Steps ### 1. Frame When the working tree is dirty, stop and ask the user to commit or stash. Set `$BASE` to the output of `git rev-parse HEAD`. Set `N` to the argument, or 3, and at most 5. Write `$WORK/task.md`: what the change must do and where, `$BASE`, the acceptance criteria and the check commands. Read the criteria from the linked issue, as `docs/agents/issue-tracker.md` says. The checks are the tests, lint and typecheck the agent instructions file names, or `mutation.testCommand` in `agents.config.json`. Write `$WORK/rubric.md` with 3 to 6 criteria a reader can grade on this task. Always include "the checks pass" and the parts of `../code-quality/SKILL.md` the task touches, such as data first, deep modules, or tests through the interface. Candidates never see the rubric. When the task has open design questions, run `h2xd:plan-grill` first, or the candidates diverge on guesses. Done when `task.md` and `rubric.md` exist and `N` is set. ### 2. Create the worktrees For each candidate `n` from 1 to `N`, run `git worktree add -b arena/<slug>-<n> ../<repo>-arena-<slug>-<n> $BASE`. In each, run every command that `jq -c '.worktreeSetup[]' agents.config.json` lists. Done when `git worktree list` shows `N` arena worktrees at `$BASE`. ### 3. Fan out Start `N` background subagents in one message, through the CLI's subagent tool, such as Agent, on the strongest model, such as Opus. When the CLI offers a model from another vendor, put some candidates on it, because one model repeats one shape. Each prompt holds only absolute paths: [`candidate.md`](candidate.md), `task.md`, `../code-quality/SKILL.md`, the worktree, and the rationale file `$WORK/candidate-<n>.md`. When a candidate fails or leaves no commit, note the dropout and go on. With fewer than two left, stop and report. Done when each surviving candidate has a commit and a rationale file. ### 4. Score Run the check commands in each surviving worktree. Record pass or fail per command in `$WORK/checks.md`. Start one read-only judge subagent with the paths to [`judge.md`](judge.md), `task.md`, `rubric.md`, `checks.md`, `../code-quality/SKILL.md`, `../code-quality/smells.md`, and the output `$WORK/verdict.md`. Label the worktrees by letter, A, B, C, never by model. When the CLI offers another model than the candidates', run the judge on it, so it does not favor its own shape. While the judge runs, read every diff end to end with `git -C <worktree> diff $BASE`, and its rationale. In `$WORK/scores.md`, score each criterion for each candidate, one criterion at a time, not by feel. Done when `scores.md` and `verdict.md` score every criterion for every candidate. ### 5. Pick the base - A candidate with failing checks is not the base while another passes. - Compare your scores with the verdict. Where they disagree, read both reasons. One of you misread, or a criterion is vague. - Pick the candidate a maintainer can extend without breaking an invariant. On a tie, take the smaller interface. - When all candidates share one shape, ship that shape with no grafts. - When they share almost nothing, the task was under-specified. Reframe in step 1 and run again. Do not average them. Done when `$WORK/synthesis.md` names the base and why. ### 6. Graft Read each other candidate once more for what beats the base: a test it lacks, a better name, a missed edge case, a simpler function. Expect one or two grafts per candidate. Port each graft by hand in the base worktree, in the base's types and names. Commit each graft with `h2xd:git-commit`, and name the source candidate in the body. Done when `synthesis.md` lists each graft with its source, and each rejected idea with why. ### 7. Verify In the base worktree, run the check commands. Run the changed code and read its real output. Run `h2xd:prove-tests` with `$BASE` as the base. A defect no candidate handled means the frame was wrong, so go back to step 1. A defect one candidate handled means a missed graft, so go back to step 6. Done when every check passes and no surviving mutant marks a behavior the task asked for. ### 8. Hand over Report in chat in the format below. Keep the base branch and worktree. Ask before you remove the others with `git worktree remove` and `git branch -D`. Done when the user has the report and chose what happens to the others. ## Report One line per item, worst first, then the checks run and their result. ``` 🔴 check: <command>: <what still fails, and the next action> 🟡 dropout: <label>: <why> 🟡 judge: <where the verdict disagreed, and how it was settled> 🟢 base: <label> on arena/<slug>-<n> at <path>: <why> 🟢 graft: <what> from <label> ``` ## Do - Give every candidate the same task file and nothing about the others. - Write your scores before you read the verdict. - Keep the base worktree until the user has merged or dropped it. ## Don't - Don't start more than 5 candidates. Split the task or sharpen the frame. - Don't pick a failing base while one passes. Port its idea as a graft. - Don't paste hunks or cherry-pick. Rewrite each graft in the base's shape. - Don't show candidates the rubric. Give them the task and its criteria. - Don't remove worktrees without asking. List them and let the user choose.
Pliki w paczce (3)
- SKILL.md 6 KB
- candidate.md 2 KB
- judge.md 1 KB
Wersje
W bazie nie ma innych skilli o nazwie code-arena.
Migawka bazy z 13.11.2026 · wersja paczki 80be0b260abd (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 85arena
Uruchamia równolegle N kandydatów do tego samego zadania, wybiera bazę i przeszczepia mocne części przegranych. - 90arena
Uruchamia równolegle kilka wariantów tego samego zadania, wybiera bazę i przenosi do niej najmocniejsze fragmenty pozostałych. - 90arena
Uruchamia równolegle kilka wariantów tego samego zadania, wybiera bazę i przenosi do niej najmocniejsze fragmenty pozostałych. - 90arena
Uruchamia równolegle kilka wariantów tego samego zadania, wybiera bazę i przenosi do niej najmocniejsze fragmenty pozostałych. - 90arena
Uruchamia równolegle kilka wariantów tego samego zadania, wybiera bazę i przenosi do niej najmocniejsze fragmenty pozostałych. - 90arena
Uruchamia równolegle kilka wariantów tego samego zadania, wybiera bazę i przenosi do niej najmocniejsze fragmenty pozostałych.