Documented experiment
Maestro (Codex/Cursor) delegates over MCP; FIFO daemon and OpenCode workers execute with state in SQLite. Orchestration worked, but measurement did not confirm cost or time reduction.
Maestro delega; daemon FIFO coordena um worker OpenCode por vez; SQLite guarda estado. A hipótese de eficiência não se confirmou.
A maestro (Codex or Cursor) needs to delegate work to OpenCode workers with an explicit lifecycle: plan, start, follow, respond, cancel, and recover, without infinite agent recursion.
Everything is local. The daemon authenticates on loopback. There are global and per-workspace limits. Interrupted tasks must be reconciled. State corruption must not become blind writes. On the happy path: one OpenCode worker at a time, without automatic decomposition that leaves the maestro without control.
Go implementation: MCP gateways over stdio, single daemon, state machine, and FIFO executor. SQLite persistence with WAL; JSONL artifacts. OpenCode adapter with server as the primary path and CLI as fallback. Isolation against recursive delegation and read-only diagnostics if state corrupts.
Repository with tests, ADRs, and benchmarks. Orchestration worked. The measurements published in the project itself did not confirm the hypothesis of reducing time and cost.
It is an experiment. It does not claim agent-engineering productivity gains in production. The negative result of the hypothesis is part of the artifact.