Daemons must run long tasks via Solo MCP spawned workers, NOT inline — flower-refine blocked its own poll/heartbeat loop 16+ min running the #95 design-loop in-process

Operating-model lesson (operator directive, 2026-07-03). flower-refine ran the Brief #95 design-loop (writer↔reviewer) INLINE via Agent-tool subagents in its own daemon session; the round-1 writer alone took ~12 min and blocked the daemon's poll/heartbeat loop for 16+ min. Directive: when a standing daemon (orchestrator/ops/refine) needs/wants to trigger LONG-RUNNING work (design loops, multi-round subagent passes, big audits/reviews), it must SPAWN a Solo MCP agent/worker (mcp__solo__spawn_agent) to do it and monitor via the normal poll cadence — NOT run it in-process. Keep the daemon session responsive to heartbeats + operator messages. This belongs codified in the daemon charters + shared conventions (Brief #97) and in the daemon-runtime review (Brief #106). Route accordingly.

{
    "incident": "flower-refine ran #95 design-loop inline",
    "codify_in": [
        "#97 charters",
        "#106 daemon-runtime review"
    ],
    "directive": "long-running work via Solo spawned workers, not inline",
    "block_minutes": "16+"
}

created

2d ago

triaged

—

resolved

—

resolved by

—