Spawns N parallel attempts at the same task, picks the strongest as a base, and grafts the best parts of the losers into it. The framework’s design-and-code bakeoff primitive, and the engine pstack-architect Phase B runs on. Its division of labour against pstack-swarm is explicit in pstack-poteto-mode: swarm is for coverage matrices, races, gauntlets, and exploration partitions; arena is for bakeoffs with base selection and grafting.
Six phases.
- A. Frame. The N candidates receive the same prompt, so the prompt is the contract. State the artifact, then derive a rubric of three to six concrete gradeable criteria — held back from the candidates, who see only the task, because the rubric is the picker’s tool in Phase D. Runners default to one candidate each on four different models; the same model N times is correct when the work is generation-bound rather than judgment-sensitive. Each candidate gets its own output location, a git worktree where possible, per pstack-principle-separate-before-serializing-shared-state.
- B. Fan out. All N spawned in one message with
run_in_background: true. Each returns the artifact and a short rationale naming the alternatives it considered and what it rejected. A dropout proceeds with N-1 and is noted in the synthesis record. - C. Cross-judge. One readonly judge on a model from a configured pool, preferring a different model family from the parent’s. It sees the rubric and the candidates by path label, scores each criterion, and recommends a base. It runs in parallel with the parent’s own reading in Phase D, never while candidates are still writing.
- D. Pick a base. Read every candidate end to end first. Score criterion by criterion, not on holistic feel, then compare against the cross-judge. Agreement confirms the pick. Disagreement means one of you is biased or the rubric was ambiguous. The tie-break is maintainability — which candidate a future maintainer can extend most easily without breaking invariants — with the cleaner boundary or smaller API preferred per pstack-principle-laziness-protocol.
- E. Graft. Walk each losing candidate once more for what is worth porting into the base. The signal is usually one or two things per candidate, not most of it.
- F. Verify the synthesized result.
Phase E is the contribution. pattern-throwaway-prototype as practiced elsewhere (including gstack-design-shotgun’s mockup variants) discards the losers; arena treats them as a parts bin, which is what pattern-candidate-grafting names. The parent-versus-judge disagreement rule in Phase D is also unusual: a split verdict is read as evidence that the rubric was ambiguous, not merely that one reader was wrong.
See Also
- pstack-swarm — the sibling fan-out primitive, partitioned by coverage rather than competition.
- pstack-architect — the main consumer; Phase B is an arena.
- pstack-blast-radius — escalates to an arena for a big or wide change.
- gstack-design-shotgun — the visual-variant counterpart, which discards rather than grafts.
- pattern-candidate-grafting — the technique this capability defines.