Turn competing plans from several agents (“such as Codex and Claude Code”) into one executable direction. It aims to “produce a clear handoff instead of a blended mush.”
Five steps: collect the source plans (resolving the original artifacts where possible, “so you can see prompt changes and assumptions that may be missing from a final summary”); normalize each into comparable claims (objective, assumptions, touched surfaces, sequence, validation, rollback, cost, executor fit); cross-review them against each other and the real codebase; decide; and hand off. “Do not reward verbosity.”
There are three outcomes: Adopt, Hybrid, and Revise first. The tie-break order is correctness and fit, then grounding in real files, then a simpler first cut, then a better validation and rollback story, then lower execution cost. It deliberately separates plan quality from executor choice: “A cheaper/faster executor can be the right choice for implementation even when another model produced the best critique.” The decision memo has sections for Decision, Why, Execution Plan, Borrowed From Other Plans, Rejected, Verification, and Executor Recommendation.
The Borrowed From Other Plans section makes this a planning-stage instance of pattern-candidate-grafting: keep the pieces of the losing plans worth keeping instead of discarding them. It differs from pstack-arena, which generates its own candidates. The arbiter judges candidates that independent agents produced. Read-only unless execution was asked for.
See Also
- pstack-arena — generate N candidates and graft the losers; the arbiter judges candidates it did not generate.
- builder-agent-watchdog — the same stance, applied to finished work.
- stage-plan — the canonical stage this implements.