Empty collaboration waits can keep a provider turn running indefinitely

Problem

A live Codex provider turn can enter a permanent loop that invokes the collaboration wait tool every 30 seconds even though each completed tool item reports both:

• agentsStates: {}
• receiverThreadIds: []

The provider process remains alive, so ordinary provider-process liveness checks correctly consider the session live. The projection therefore remains running indefinitely even though there is no agent left to wait for and no useful work occurs.

This is distinct from a provider process dying during restart.

Expected behavior

When collaboration wait completes with no live/receivable agents, the turn must not silently re-enter the same empty wait forever. The provider integration should surface a terminal/no-agents result that lets the root continue or finish, or apply a narrowly scoped guard against repeated empty waits.

Evidence

In the reproduced turn, consecutive collab_agent_tool_call items completed at 30-second intervals with empty agent state and empty receiver IDs while the orchestration turn remained running. Stopping through the supported thread.session.stop command immediately produced a clean interrupted turn, stopped session/runtime, and cleared activeTurnId.

Any fix should target this exact empty-wait condition and must not impose a generic timeout on legitimate long-running tools or active subagents.

The issue names no files or tests; start by tracing the provider integration's collaboration wait completion path and how agentsStates and receiverThreadIds affect turn progression. Done means an empty wait cannot repeat indefinitely, while legitimate long-running tools and active subagents remain unaffected, and the root turn can finish or continue appropriately.