Przejdź do treści
Biblioteka skilli
AI i agenci · Automatyzacja

arena

/arena

Uruchamia N równoległych kandydatów do tego samego zadania, wybiera bazę i przenosi do niej najmocniejsze części przegranych, gdy jedna próba nietrywialnego artefaktu utrwaliłaby zły kształt.

Repozytorium
pstack ↗
Licencja
MIT
Ostatnia zmiana repozytorium
26.10.2026

Ocena i bezpieczeństwo

100/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść35 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
1
Skrypty
0
Komentarze HTML
0
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 26.10.2026, bez odwołań do wycofanych modeli i API.

Licencja

MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.

Co robi i kiedy

Uruchamia N równoległych kandydatów do tego samego zadania, wybiera bazę i przenosi do niej najmocniejsze części przegranych, gdy jedna próba nietrywialnego artefaktu utrwaliłaby zły kształt.

Wyzwalacze z opisu

arena thisthrow it in the arena
Oryginalny opis (EN)

Spawn N parallel candidates at the same task, pick a base, graft the strongest parts of the losers into it. Use for /arena, 'arena this', 'throw it in the arena', or when one attempt at a non-trivial artifact would lock in the wrong shape.

Jak działa

  1. Phase A: Frame
  2. Phase B: Fan out
  3. Phase C: Cross-judge
  4. Phase D: Pick a base
  5. Phase E: Graft
  6. Phase F: Verify
  7. Outputs
Sekcje z nagłówków SKILL.md, w oryginale · 51 linii, 1 136 słów
Podgląd SKILL.md
---
name: arena
description: "Spawn N parallel candidates at the same task, pick a base, graft the strongest parts of the losers into it. Use for /arena, 'arena this', 'throw it in the arena', or when one attempt at a non-trivial artifact would lock in the wrong shape."
disable-model-invocation: true
---

# Arena

Open a todolist with one entry per phase (Frame, Fan out, Cross-judge, Pick, Graft, Verify) before launching anything. The arena runs autonomously and the list keeps phases from silently disappearing.

## Phase A: Frame

Every candidate gets the same prompt, so the prompt is the contract. Get it right before spawning anything.

1. State the artifact each candidate is producing.
2. Derive the rubric, 3-6 concrete gradeable criteria for what success looks like on *this* task. Concrete: `Adds a --dry-run flag that skips writes`. Vague: `code is correct`. The rubric is the picker's tool in Phase D; candidates only see the task.
3. Pick the runners from canonical role key `arena-runners` in the host's setup projection per [`../setup-pstack/SKILL.md`](../setup-pstack/SKILL.md) §5. If the catalog or key is absent, spawn **one** runner using the host default model resolved from the active profile (omit if rejected). Spawn more only when the projection lists more detected slugs, or when the arena covers multiple design directions on the **same** named-or-omitted slug. Same model N times when the work is generation-bound rather than judgment-sensitive. Every candidate seat maps to a real tracked child agent and discrete worktree; never claim N independent models from one repeated alias or without verified spawn. When multiple seats use the same model family, record generation diversity only; never claim cross-family independence without verified diverse model families. Pick each subagent's reasoning level and fast or standard speed for the job, per the Subagents section of the **poteto-mode** skill.
4. Give each candidate its own output location: a project worktree under `.worktrees/` for repo-backed work (per using-git-worktrees, `fix/<slug>` leaves, removed after the run), otherwise `/tmp/arena-<slug>/candidate-<n>/` only for scratch with no repo checkout. N candidates writing to the same path is shared mutable state and fails the the **separate-before-serializing-shared-state** principle skill test.
5. Freeze the task and the rubric before any candidate spawns: a rubric rewritten after a candidate returned invalidates the comparison, so the run re-runs instead of scoring the old candidates against the new rubric.

## Phase B: Fan out

Spawn all N subagents in one message using the host's `agent.spawn` primitive (per active harness profile and host references), background/async execution, role/type `pstack:arena-runners` ([`../setup-pstack/SKILL.md`](../setup-pstack/SKILL.md)), each with the brief, the path to the shared grounding, its own output path, and instructions to produce the artifact and a short rationale naming the alternatives it considered and what it rejected. Do not send `reasoning_effort` unless supported. Architect passes role key `architect-runners` and `pstack:architect-runners` instead.

The rationale is mandatory. Without it, the parent cannot tell whether a candidate's structure is principled or accidental, which makes Phase E grafting unreliable.

If a candidate fails to produce output, proceed with N-1 and note the dropout in the synthesis record. Join under a bounded join deadline named before spawn (e.g. `join deadline: 10m`); a candidate that has not returned by the deadline is a reported dropout, not silence. A candidate you cancel, or that the host kills in flight, is the same reported dropout, never a pass.

Dispatch the candidates through the fan-out route the active harness profile binds (its host reference names the concrete mechanism). Each candidate runs in a fresh context: a warm or keep-alive worker reused across candidates carries the previous candidate's context, so a run that reuses one reports that reuse as an adaptation or a gap rather than an independent attempt. A dropped or cancelled candidate keeps its gap in the synthesis record, and synthesis never grades or backfills it.

## Phase C: Cross-judge

After all Phase B candidates complete, choose one model from canonical role key `arena-cross-judge-pool` in the host's setup projection per [`../setup-pstack/SKILL.md`](../setup-pstack/SKILL.md) §5. If the catalog or key is absent, spawn **one** judge using the host default model resolved from the active profile (omit if rejected). Prefer a pool entry whose evidence-backed family differs from the candidate family when the projection names more than one detected slug. When no differing family is available, report cross-judge family diversity as unproven and mark the evaluation `DEGRADED` (or `BLOCKED` if distinct family independence was required). Spawn one judge using the host's `agent.spawn` primitive with role/type `pstack:arena-cross-judge-pool` ([`../setup-pstack/SKILL.md`](../setup-pstack/SKILL.md)), operating in read-only mode (`workspace.readonly`). Do not send `reasoning_effort` unless supported. It sees the rubric and the candidates by path label, scores each criterion, and recommends a base with rationale. It runs in parallel with the parent's read.

## Phase D: Pick a base

Read every candidate end to end before picking. Skimming N candidates surfaces only the candidate whose surface looks most familiar. Score each against the rubric criterion by criterion, not on holistic feel, and compare with the cross-judge. Agreement on the base confirms the pick. Disagreement means one of you is biased or the rubric was ambiguous. Read both rationales before deciding.

Pick the base a future maintainer can extend most easily without breaking invariants. When two feel tied, prefer the cleaner boundary or smaller surface area, per the Laziness Protocol.

## Phase E: Graft

Walk each losing candidate once more for what is worth porting into the base, usually one or two things per candidate, not most of it. Fold each graft in by hand, per the **redesign-from-first-principles** principle skill, so the result stays coherent under one mental model. Don't paste mechanically.

The rejection notes are the highest-signal part of the record. Future readers learn from what you considered and dropped, not just what you kept.

If the candidates converge on one shape, that is strong agreement. Note it and ship the consensus shape. No graft is needed. If they wildly diverge, Phase A was under-specified. Reframe and re-run rather than averaging the divergence.

## Phase F: Verify

Verify the synthesized artifact like any other output, per the **prove-it-works** principle skill. The arena does not earn you a pass. If verification finds a problem the arena missed, either Phase A was wrong (re-frame and re-run) or a candidate caught it and you missed the graft (go back to Phase E). Don't paper over it.

## Outputs

One synthesized artifact, with a short synthesis note alongside naming the base and why, the cross-judge's verdict, each graft with its source candidate, what was rejected and why, any convergence or dropouts, and the verification result. Name each candidate and judge model with its evidence-backed family. When the harness cannot identify a family from live discovery, record it as `unknown` and mark the run degraded. Never derive a family from a slug or an inferred provider. A single-family run stays degraded and never counts as multi-model proof. The note binds to evidence: the receipt scenario id, file hash, or PR head SHA the artifact was verified against. A prose-only verdict fails. When the arena claims host parity, name the declared census read from `pstack.package.json#host_targets` via `get_declared_hosts` in `scripts/verify_portable.py`, with each host's evidence tier; a missing host is a stated gap, not silence.

Wersje

RepozytoriumWynikLicencja
Ta strona · tommy-ca/pstack100MIT
nathanvale/side-quest-marketplace100MIT
hamzzaqureshi/gravidex100MIT
requie/endostack100MIT
Sup-Film/pstack-omp-lean100MIT
jakeschincariol/arena-skill100MIT
samhvw8/dotfiles90MIT
Snestra/pi-t3-bridge90MIT
chmonitor/chmonitor90GPL-3.0

Migawka bazy z 13.11.2026 · wersja paczki aa59f9ba86f0 (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

  • 100code-arena
    Uruchamia kilka równoległych prób tego samego zadania w osobnych worktree i wybiera wersję z najwyższą oceną.
  • 100agni-kai
    Prowadzi pojedynek implementacji (bakeoff): uruchamia N równoległych kandydatów do tego samego zadania, wybiera bazę sędzią krzyżowym i przenosi mocne fragmenty przegranych do zwycięzcy.
  • 90st-best-of-n
    Uruchamia równolegle N niezależnych prób tego samego zadania, wybiera najmocniejszą i przenosi do niej mocne fragmenty pozostałych.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.