mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-04-25 00:51:20 +00:00
fix: activate fallback provider on repeated empty responses + user-visible status (#7505)
When models return empty responses (no content, no tool calls, no reasoning), Hermes previously retried 3 times silently then fell through to '(empty)' — without ever trying the fallback provider chain. Users on GLM-4.5-Air and similar models experienced what appeared to be a complete hang, especially in gateway (Telegram/Discord) contexts where the silent retries produced zero feedback. Changes: - After exhausting 3 empty retries, attempt _try_activate_fallback() before giving up with '(empty)'. If fallback succeeds, reset retry counter and continue the conversation loop with the new provider. - Replace all _vprint() calls in recovery paths with _emit_status(), which surfaces messages through both CLI (_vprint with force=True) and gateway (status_callback -> adapter.send). Users now see: * '⚠️ Empty response from model — retrying (N/3)' during retries * '⚠️ Model returning empty responses — switching to fallback...' * '↻ Switched to fallback: <model> (<provider>)' on success * '❌ Model returned no content after all retries [and fallback]' - Add logger.warning() throughout empty response paths for log file visibility (model name, provider, retry counts). - Upgrade _last_content_with_tools fallback from logger.debug to logger.info + _emit_status so recovery is visible. - Upgrade thinking-only prefill continuation to use _emit_status. Tests: - test_empty_response_triggers_fallback_provider: verifies fallback activation after 3 empty retries produces content from fallback model - test_empty_response_fallback_also_empty_returns_empty: verifies graceful degradation when fallback also returns empty - test_empty_response_emits_status_for_gateway: verifies _emit_status is called during retries so gateway users see feedback Addresses #7180.
This commit is contained in:
Teknium
GitHub
parent
992422910c
commit
842e669a13
2 changed files with 178 additions and 15 deletions
88
run_agent.py
88
run_agent.py
|
|
@ -9459,7 +9459,8 @@ class AIAgent:
|
|||
fallback = getattr(self, '_last_content_with_tools', None)
|
||||
if fallback:
|
||||
_turn_exit_reason = "fallback_prior_turn_content"
|
||||
logger.debug("Empty follow-up after tool calls — using prior turn content as final response")
|
||||
logger.info("Empty follow-up after tool calls — using prior turn content as final response")
|
||||
self._emit_status("↻ Empty response after tool calls — using earlier content as final answer")
|
||||
self._last_content_with_tools = None
|
||||
self._empty_content_retries = 0
|
||||
for i in range(len(messages) - 1, -1, -1):
|
||||
|
|
@ -9490,9 +9491,13 @@ class AIAgent:
|
|||
)
|
||||
if _has_structured and self._thinking_prefill_retries < 2:
|
||||
self._thinking_prefill_retries += 1
|
||||
self._vprint(
|
||||
f"{self.log_prefix}↻ Thinking-only response — "
|
||||
f"prefilling to continue "
|
||||
logger.info(
|
||||
"Thinking-only response (no visible content) — "
|
||||
"prefilling to continue (%d/2)",
|
||||
self._thinking_prefill_retries,
|
||||
)
|
||||
self._emit_status(
|
||||
f"↻ Thinking-only response — prefilling to continue "
|
||||
f"({self._thinking_prefill_retries}/2)"
|
||||
)
|
||||
interim_msg = self._build_assistant_message(
|
||||
|
|
@ -9508,23 +9513,57 @@ class AIAgent:
|
|||
# Model returned nothing — no content, no
|
||||
# structured reasoning, no tool calls. Common
|
||||
# with open models (transient provider issues,
|
||||
# rate limits, sampling flukes). Silently retry
|
||||
# up to 3 times before giving up. Skip when
|
||||
# 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:
|
||||
self._empty_content_retries += 1
|
||||
self._vprint(
|
||||
f"{self.log_prefix}↻ Empty response (no content or reasoning) "
|
||||
f"— retrying ({self._empty_content_retries}/3)",
|
||||
force=True,
|
||||
logger.warning(
|
||||
"Empty response (no content or reasoning) — "
|
||||
"retry %d/3 (model=%s)",
|
||||
self._empty_content_retries, self.model,
|
||||
)
|
||||
self._emit_status(
|
||||
f"⚠️ Empty response from model — retrying "
|
||||
f"({self._empty_content_retries}/3)"
|
||||
)
|
||||
continue
|
||||
|
||||
# Exhausted prefill attempts, empty retries, or
|
||||
# structured reasoning with no content —
|
||||
# fall through to "(empty)" terminal.
|
||||
# ── Exhausted retries — try fallback provider ──
|
||||
# Before giving up with "(empty)", attempt to
|
||||
# switch to the next provider in the fallback
|
||||
# chain. This covers the case where a model
|
||||
# (e.g. GLM-4.5-Air) consistently returns empty
|
||||
# due to context degradation or provider issues.
|
||||
if _truly_empty and self._fallback_chain:
|
||||
logger.warning(
|
||||
"Empty response after %d retries — "
|
||||
"attempting fallback (model=%s, provider=%s)",
|
||||
self._empty_content_retries, self.model,
|
||||
self.provider,
|
||||
)
|
||||
self._emit_status(
|
||||
"⚠️ Model returning empty responses — "
|
||||
"switching to fallback provider..."
|
||||
)
|
||||
if self._try_activate_fallback():
|
||||
self._empty_content_retries = 0
|
||||
self._emit_status(
|
||||
f"↻ Switched to fallback: {self.model} "
|
||||
f"({self.provider})"
|
||||
)
|
||||
logger.info(
|
||||
"Fallback activated after empty responses: "
|
||||
"now using %s on %s",
|
||||
self.model, self.provider,
|
||||
)
|
||||
continue
|
||||
|
||||
# Exhausted retries and fallback chain (or no
|
||||
# fallback configured). Fall through to the
|
||||
# "(empty)" terminal.
|
||||
_turn_exit_reason = "empty_response_exhausted"
|
||||
reasoning_text = self._extract_reasoning(assistant_message)
|
||||
assistant_msg = self._build_assistant_message(assistant_message, finish_reason)
|
||||
|
|
@ -9533,9 +9572,28 @@ class AIAgent:
|
|||
|
||||
if reasoning_text:
|
||||
reasoning_preview = reasoning_text[:500] + "..." if len(reasoning_text) > 500 else reasoning_text
|
||||
self._vprint(f"{self.log_prefix}ℹ️ Reasoning-only response (no visible content). Reasoning: {reasoning_preview}")
|
||||
logger.warning(
|
||||
"Reasoning-only response (no visible content) "
|
||||
"after exhausting retries and fallback. "
|
||||
"Reasoning: %s", reasoning_preview,
|
||||
)
|
||||
self._emit_status(
|
||||
"⚠️ Model produced reasoning but no visible "
|
||||
"response after all retries. Returning empty."
|
||||
)
|
||||
else:
|
||||
self._vprint(f"{self.log_prefix}ℹ️ Empty response (no content or reasoning) after 3 retries.")
|
||||
logger.warning(
|
||||
"Empty response (no content or reasoning) "
|
||||
"after %d retries. No fallback available. "
|
||||
"model=%s provider=%s",
|
||||
self._empty_content_retries, self.model,
|
||||
self.provider,
|
||||
)
|
||||
self._emit_status(
|
||||
"❌ Model returned no content after all retries"
|
||||
+ (" and fallback attempts." if self._fallback_chain else
|
||||
". No fallback providers configured.")
|
||||
)
|
||||
|
||||
final_response = "(empty)"
|
||||
break
|
||||
|
|
|
|||
|
|
@ -1823,6 +1823,111 @@ class TestRunConversation:
|
|||
assert result["final_response"] == "Here is the actual answer."
|
||||
assert result["api_calls"] == 2 # 1 original + 1 nudge retry
|
||||
|
||||
def test_empty_response_triggers_fallback_provider(self, agent):
|
||||
"""After 3 empty retries, fallback provider is activated and produces content."""
|
||||
self._setup_agent(agent)
|
||||
agent.base_url = "http://127.0.0.1:1234/v1"
|
||||
# Configure a fallback chain
|
||||
agent._fallback_chain = [{"provider": "openrouter", "model": "anthropic/claude-sonnet-4"}]
|
||||
agent._fallback_index = 0
|
||||
agent._fallback_activated = False
|
||||
|
||||
empty_resp = _mock_response(content=None, finish_reason="stop")
|
||||
content_resp = _mock_response(content="Fallback answer.", finish_reason="stop")
|
||||
# 4 empty (1 orig + 3 retries), then fallback model answers
|
||||
agent.client.chat.completions.create.side_effect = [
|
||||
empty_resp, empty_resp, empty_resp, empty_resp, content_resp,
|
||||
]
|
||||
|
||||
fallback_called = {"called": False}
|
||||
|
||||
def _mock_fallback():
|
||||
fallback_called["called"] = True
|
||||
# Simulate what _try_activate_fallback does: just advance the
|
||||
# index and set the flag (the client is already mocked).
|
||||
agent._fallback_index = 1
|
||||
agent._fallback_activated = True
|
||||
agent.model = "anthropic/claude-sonnet-4"
|
||||
agent.provider = "openrouter"
|
||||
return True
|
||||
|
||||
with (
|
||||
patch.object(agent, "_persist_session"),
|
||||
patch.object(agent, "_save_trajectory"),
|
||||
patch.object(agent, "_cleanup_task_resources"),
|
||||
patch.object(agent, "_try_activate_fallback", side_effect=_mock_fallback),
|
||||
):
|
||||
result = agent.run_conversation("answer me")
|
||||
assert fallback_called["called"], "Fallback should have been triggered"
|
||||
assert result["completed"] is True
|
||||
assert result["final_response"] == "Fallback answer."
|
||||
|
||||
def test_empty_response_fallback_also_empty_returns_empty(self, agent):
|
||||
"""If fallback also returns empty, final response is (empty)."""
|
||||
self._setup_agent(agent)
|
||||
agent.base_url = "http://127.0.0.1:1234/v1"
|
||||
agent._fallback_chain = [{"provider": "openrouter", "model": "anthropic/claude-sonnet-4"}]
|
||||
agent._fallback_index = 0
|
||||
agent._fallback_activated = False
|
||||
|
||||
empty_resp = _mock_response(content=None, finish_reason="stop")
|
||||
# 4 empty from primary (1 + 3 retries), fallback activated,
|
||||
# then 4 more empty from fallback (1 + 3 retries), no more fallbacks
|
||||
agent.client.chat.completions.create.side_effect = [
|
||||
empty_resp, empty_resp, empty_resp, empty_resp, # primary exhausted
|
||||
empty_resp, empty_resp, empty_resp, empty_resp, # fallback exhausted
|
||||
]
|
||||
|
||||
def _mock_fallback():
|
||||
if agent._fallback_index >= len(agent._fallback_chain):
|
||||
return False
|
||||
agent._fallback_index += 1
|
||||
agent._fallback_activated = True
|
||||
agent.model = "anthropic/claude-sonnet-4"
|
||||
agent.provider = "openrouter"
|
||||
return True
|
||||
|
||||
with (
|
||||
patch.object(agent, "_persist_session"),
|
||||
patch.object(agent, "_save_trajectory"),
|
||||
patch.object(agent, "_cleanup_task_resources"),
|
||||
patch.object(agent, "_try_activate_fallback", side_effect=_mock_fallback),
|
||||
):
|
||||
result = agent.run_conversation("answer me")
|
||||
assert result["completed"] is True
|
||||
assert result["final_response"] == "(empty)"
|
||||
|
||||
def test_empty_response_emits_status_for_gateway(self, agent):
|
||||
"""_emit_status is called during empty retries so gateway users see feedback."""
|
||||
self._setup_agent(agent)
|
||||
agent.base_url = "http://127.0.0.1:1234/v1"
|
||||
|
||||
empty_resp = _mock_response(content=None, finish_reason="stop")
|
||||
# 4 empty: 1 original + 3 retries, all empty, no fallback
|
||||
agent.client.chat.completions.create.side_effect = [
|
||||
empty_resp, empty_resp, empty_resp, empty_resp,
|
||||
]
|
||||
|
||||
status_messages = []
|
||||
|
||||
def _capture_status(msg):
|
||||
status_messages.append(msg)
|
||||
|
||||
with (
|
||||
patch.object(agent, "_persist_session"),
|
||||
patch.object(agent, "_save_trajectory"),
|
||||
patch.object(agent, "_cleanup_task_resources"),
|
||||
patch.object(agent, "_emit_status", side_effect=_capture_status),
|
||||
):
|
||||
result = agent.run_conversation("answer me")
|
||||
|
||||
assert result["final_response"] == "(empty)"
|
||||
# Should have emitted retry statuses (3 retries) + final failure
|
||||
retry_msgs = [m for m in status_messages if "retrying" in m.lower()]
|
||||
assert len(retry_msgs) == 3, f"Expected 3 retry status messages, got {len(retry_msgs)}: {status_messages}"
|
||||
failure_msgs = [m for m in status_messages if "no content" in m.lower() or "no fallback" in m.lower()]
|
||||
assert len(failure_msgs) >= 1, f"Expected at least 1 failure status, got: {status_messages}"
|
||||
|
||||
def test_nous_401_refreshes_after_remint_and_retries(self, agent):
|
||||
self._setup_agent(agent)
|
||||
agent.provider = "nous"
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue