fix(crew): send AgentReady before starting nodes (unbreaks greeting in start())

[abhishekmishragithub]

Problem

CrewSession.start() (session.py) called _start_nodes() and only afterwards sent SDKAgentReadyEvent. A node whose start() emits a speak — the natural way to greet — had that SDKAgentSpeakEvent flushed to the platform before Ready. Pipecat's AgentSDKProcessor._connect expects Ready as the first frame; getting a speak first leaves _connected=False, its receive loop never starts, and every subsequent event is a no-op — the whole call goes silent. This is why node-level self.speak() in start() silenced our test agents.

Fix

Send SDKAgentReadyEvent before _start_nodes(). Incoming events still buffer on the websocket until our own receive loop starts (after nodes are up), so nothing is lost by ordering Ready first. Any early speak from a node's start() is now legal.

Test

Adds tests/custom/test_crew_session_ready_ordering.py asserting Ready is sent before any node starts (fails if the ordering regresses).

Notes

• Surfaced during the crew-greeting investigation; pairs with pipecat #1028 (which makes firstMessage the primary greeting path). This fix makes an in-start() greeting also viable instead of silently breaking.
• No version bump — bundle into the next SDK release (5.4.0), verified end to end there. A second lifecycle footgun (OutputCrewNode.process_event requiring super()) is deferred to the DevX pass since its fix is an API change.

🤖 Generated with Claude Code