The analytics agent cannot run a Flow

Asking the chat agent "run the weekly revenue flow" does not work, and this is deliberate. There is no run_flow in agent/tools/. A Run is started from the Flows page or from the flow chat, whose agent is built one tier up.

The reason is a cycle. Running an agent Phase means calling run_turn, so the Flow runner sits above agent/. A tool in agent/tools/ that started a Run would have to import that runner, which is an upward import — the thing tests/test_layering.py exists to refuse. The tier table is not a convention that can be bent for one convenient tool; flows/ is tier 3 and api/ moved to tier 4 to make room for it.

The alternative considered was injecting the runner into the tool as a callback at build time. That works and would satisfy the import graph, because the edge would no longer be visible to a static parse. That is exactly the objection: it would be a cycle test_layering cannot see, and the next person to reason about the tiers would be reasoning about a graph that does not describe the program. if TYPE_CHECKING: is already called out in AGENTS.md as a cycle that has not failed yet; a callback is the same trick with better manners.

What is lost is smaller than it sounds. "Run it from the agent" is still true — it is the flow chat that does it, and that chat is where someone is already talking about the Flow. What is not possible is starting a Run from the middle of a conversation about last quarter's revenue, and that conversation has a Session pinned to main and a transcript in sqlite, neither of which a Run wants.