Bug

After the run_agent.py refactor (splitting ~12k LOC into ~15 modules under agent/), the fallback rate-limit handling in agent/conversation_loop.py calls _pool_may_recover_from_rate_limit() as a bare name — but the function is defined in run_agent.py and was never imported.

When the primary model returns a 429 (rate limit or billing quota), the agent crashes with:

NameError: name '_pool_may_recover_from_rate_limit' is not defined

This breaks the entire fallback chain: the user sees the error and /reset does not help because the import is missing at module level.

Root Cause

The call site at agent/conversation_loop.py:2254 was moved from run_agent.py during the refactor but the import was not added. A top-level from run_agent import _pool_may_recover_from_rate_limit would risk circular imports since run_agent.py imports from agent/ modules.

Fix

Use the existing _ra() lazy-import helper (already used by other run_agent symbols in conversation_loop.py, e.g. _ra()._set_interrupt(...), _ra().AIAgent._get_tool_call_name_static(...)) to resolve the function at call time:

# Before (crashes):
pool_may_recover = _pool_may_recover_from_rate_limit(...)

# After (works):
pool_may_recover = _ra()._pool_may_recover_from_rate_limit(...)

Testing

Added 3 regression tests in tests/run_agent/test_provider_fallback.py:

• test_ra_resolves_pool_may_recover — verifies _ra() resolves the function
• test_call_with_none_returns_false — end-to-end: None pool → False
• test_call_with_pool_passthrough — end-to-end: multi-credential pool → True

All 25 tests in the file pass (including 22 pre-existing + 3 new).

Files Changed