fix: cap consecutive empty LLM response retries to prevent infinite loop

[Kailigithub]

Closes #201

Summary

Issue #201 reports that the agent enters an infinite retry loop when the LLM returns empty responses, continuously printing [Warn] LLM returned an empty response. Retrying... without any upper bound.

Root Cause

do_no_tool() in ga.py returns a StepOutcome with a retry prompt whenever the LLM response is empty. The agent loop (agent_runner_loop) then re-invokes the LLM, which can produce another empty response, creating an unbounded cycle until max_turns (40) is exhausted.

Fix

• Added a consecutive empty response counter (_consecutive_empty_retries) to GenericAgentHandler
• After 3 consecutive empty responses, the loop breaks by returning StepOutcome({}, next_prompt=None) instead of a retry prompt
• The counter resets on any non-empty response
• Retry messages now show progress: Retrying (1/3)..., Retrying (2/3)..., etc.

Changes

• ga.py — __init__: added _consecutive_empty_retries = 0
• ga.py — do_no_tool(): capped retries at 3, added counter reset on valid response

Verification

✓ python3 -m py_compile ga.py — no syntax errors
✓ ruff check ga.py — no new lint issues (existing project-level warnings unchanged)
✓ 8 lines added, 1 removed — surgical change, no collateral modification

[lsdefine]

Thanks for identifying this issue. We've implemented a simpler fix on our end. Closing this PR.