Skip to content

fix: empty response recovery for reasoning models (mimo, qwen, GLM) #8609

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
teknium1 merged 1 commit into from
Apr 12, 2026
+37 −17
Merged

fix: empty response recovery for reasoning models (mimo, qwen, GLM) #8609

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
40 changes: 30 additions & 10 deletions run_agent.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9736,12 +9736,25 @@ def _stop_spinner():

# Pop thinking-only prefill message(s) before appending
# (tool-call path — same rationale as the final-response path).
_had_prefill = False
while (
messages
and isinstance(messages[-1], dict)
and messages[-1].get("_thinking_prefill")
):
messages.pop()
_had_prefill = True

# Reset prefill counter when tool calls follow a prefill
# recovery. Without this, the counter accumulates across
# the whole conversation — a model that intermittently
# empties (empty → prefill → tools → empty → prefill →
# tools) burns both prefill attempts and the third empty
# gets zero recovery. Resetting here treats each tool-
# call success as a fresh start.
if _had_prefill:
self._thinking_prefill_retries = 0
self._empty_content_retries = 0

messages.append(assistant_msg)
self._emit_interim_assistant_message(assistant_msg)
Expand Down Expand Up @@ -9917,16 +9930,23 @@ def _stop_spinner():
self._save_session_log(messages)
continue

# ── Empty response retry (no reasoning) ──────
# Model returned nothing — no content, no
# structured reasoning, no tool calls. Common
# with open models (transient provider issues,
# rate limits, sampling flukes). Retry up to 3
# times before attempting fallback. Skip when
# content has inline <think> tags (model chose
# to reason, just no visible text).
_truly_empty = not final_response.strip()
if _truly_empty and not _has_structured and self._empty_content_retries < 3:
# ── Empty response retry ──────────────────────
# Model returned nothing usable. Retry up to 3
# times before attempting fallback. This covers
# both truly empty responses (no content, no
# reasoning) AND reasoning-only responses after
# prefill exhaustion — models like mimo-v2-pro
# always populate reasoning fields via OpenRouter,
# so the old `not _has_structured` guard blocked
# retries for every reasoning model after prefill.
_truly_empty = not self._strip_think_blocks(
final_response
).strip()
_prefill_exhausted = (
_has_structured
and self._thinking_prefill_retries >= 2
)
if _truly_empty and (not _has_structured or _prefill_exhausted) and self._empty_content_retries < 3:
self._empty_content_retries += 1
logger.warning(
"Empty response (no content or reasoning) — "
Expand Down
14 changes: 7 additions & 7 deletions tests/run_agent/test_run_agent.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1741,9 +1741,9 @@ def test_reasoning_only_local_resumed_no_compression_triggered(self, agent):
{"role": "assistant", "content": "old answer"},
]

# 3 responses: original + 2 prefill continuations (structured reasoning triggers prefill)
# 6 responses: original + 2 prefill + 3 retries after prefill exhaustion
with (
patch.object(agent, "_interruptible_api_call", side_effect=[empty_resp, empty_resp, empty_resp]),
patch.object(agent, "_interruptible_api_call", side_effect=[empty_resp] * 6),
patch.object(agent, "_compress_context") as mock_compress,
patch.object(agent, "_persist_session"),
patch.object(agent, "_save_trajectory"),
Expand All @@ -1754,18 +1754,18 @@ def test_reasoning_only_local_resumed_no_compression_triggered(self, agent):
mock_compress.assert_not_called() # no compression triggered
assert result["completed"] is True
assert result["final_response"] == "(empty)"
assert result["api_calls"] == 3 # 1 original + 2 prefill continuations
assert result["api_calls"] == 6 # 1 original + 2 prefill + 3 retries

def test_reasoning_only_response_prefill_then_empty(self, agent):
"""Structured reasoning-only triggers prefill continuation (up to 2), then falls through to (empty)."""
"""Structured reasoning-only triggers prefill (2), then retries (3), then (empty)."""
self._setup_agent(agent)
empty_resp = _mock_response(
content=None,
finish_reason="stop",
reasoning_content="structured reasoning answer",
)
# 3 responses: original + 2 prefill continuations, all reasoning-only
agent.client.chat.completions.create.side_effect = [empty_resp, empty_resp, empty_resp]
# 6 responses: 1 original + 2 prefill + 3 retries after prefill exhaustion
agent.client.chat.completions.create.side_effect = [empty_resp] * 6
with (
patch.object(agent, "_persist_session"),
patch.object(agent, "_save_trajectory"),
Expand All @@ -1774,7 +1774,7 @@ def test_reasoning_only_response_prefill_then_empty(self, agent):
result = agent.run_conversation("answer me")
assert result["completed"] is True
assert result["final_response"] == "(empty)"
assert result["api_calls"] == 3 # 1 original + 2 prefill continuations
assert result["api_calls"] == 6 # 1 original + 2 prefill + 3 retries

def test_reasoning_only_prefill_succeeds_on_continuation(self, agent):
"""When prefill continuation produces content, it becomes the final response."""
Expand Down
Loading