Arena

Run only when the user explicitly invokes arena, such as $arena in Codex or /arena in Claude Code. A generic comparison request or a reference from another skill does not activate this workflow.

Compare alternatives to one problem. Different slices of a task are ordinary delegation, not an arena.

Frame before delegating

Define the requested artifact, fixed input/revision, scope, real constraints, and observable acceptance criteria before seeing candidates. Include the current approach as a baseline when it is a viable option. Agree a bounded time or work budget appropriate to the task.

Use two candidates by default; increase only for distinct useful approaches. Use different available models when supported and permitted, without hardcoded model IDs or new services. Same-model independent runs are valid, but report the reduced diversity.

Isolate candidates

Give each worker the same problem and constraints, its permitted output location, and a request for an artifact plus rationale. Keep candidates from reading each other's output until their attempts finish. Carry user/repository instructions and approval limits into every brief. Do not delegate further fan-out by default.

Design candidates can return sketches without file edits. File-writing candidates need separate temporary directories or isolated worktrees from the same baseline; do not copy secrets or unrelated dirty work. If isolating a real checkout requires additional approval, obtain it before that operation. Never reset, stash, switch, or clean the user's checkout to make room for a candidate.

Isolate runtime state as well as files: ports, data directories, caches that cannot be shared safely, and external write targets. Use available runtime isolation guidance. If isolation cannot be established, use read-only sketches.

When delegation is unavailable or forbidden, compare sequential alternatives and label them as one-agent exploration. Do not claim independent review.

Judge and synthesize

Wait until writers have finished before reading or testing their outputs. Read every completed candidate and compare each acceptance criterion. When available and useful, use an independent read-only judge after candidate output is stable, with the same rubric and anonymous candidate labels.

Verify consequential claims instead of selecting by confidence or vote. Report failed candidates and missing checks; with fewer than two usable alternatives, the comparison is incomplete.

Choose a base and incorporate only improvements that preserve one coherent design. Apply changes to the target checkout only when implementation is authorized, after inspecting its current diff and preserving unrelated work. Do not merge branches or rewrite history as an implicit synthesis step.

Use prove-it-works on the synthesized result. Evidence from an individual candidate does not validate the combined result.

Deliver

Return the selected artifact or design, comparison criteria, reasons for the choice, useful ideas adopted or rejected, candidate failures, and verification limits. Keep task-specific comparison state temporary; promote only requested durable decisions. Stop owned processes; retain evidence and honor approval requirements before removing worktrees or branches.

Source

Adapted from Lauren Tan's pstack arena. Local changes: bounded fan-out, runtime capability detection, preserved approval boundaries, and explicit isolation/fallback behavior. See LICENSE.