ouyangbo/hermes-agent
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Merge pull request #85581 from bbednarski9/codex/fix-openai-sparse-response-objects

Browse Source 
fix(openai): tolerate sparse response objects
This commit is contained in:
Jeffrey Quesnelle
2026-08-18 11:42:24 -04:00
committed by GitHub
parent aa09c8fe8c a51337ecee
commit 9664e386f6
4 changed files with 197 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files
agent/chat_completion_helpers.py
+27 -19
View File
@@ -4140,8 +4140,9 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
agent._fire_reasoning_delta(reasoning_text)
# Accumulate text content — fire callback only when no tool calls
if delta and delta.content:
content_parts.append(delta.content)
delta_content = getattr(delta, "content", None)
if delta_content:
content_parts.append(delta_content)
if not tool_calls_acc:
if pending_text_parts or _provider_stream_text_may_be_sse(delta.content):
pending_text_parts.append(delta.content)
@@ -4151,7 +4152,7 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
_flush_pending_stream_text()
continue
_fire_first_delta()
agent._fire_stream_delta(delta.content)
agent._fire_stream_delta(delta_content)
deltas_were_sent["yes"] = True
# Tool calls suppress regular content streaming (avoids
# displaying chatty "I'll use the tool..." text alongside
@@ -4166,17 +4167,19 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
# box is already closed (tool boundary flush).
elif agent.stream_delta_callback:
try:
agent.stream_delta_callback(delta.content)
agent._record_streamed_assistant_text(delta.content)
agent.stream_delta_callback(delta_content)
agent._record_streamed_assistant_text(delta_content)
except Exception:
pass
# Accumulate tool call deltas — notify display on first name
if delta and delta.tool_calls:
delta_tool_calls = getattr(delta, "tool_calls", None)
if delta_tool_calls:
_flush_pending_stream_text()
for tc_delta in delta.tool_calls:
raw_idx = tc_delta.index if tc_delta.index is not None else 0
delta_id = tc_delta.id or ""
for tc_delta in delta_tool_calls:
raw_index = getattr(tc_delta, "index", None)
raw_idx = raw_index if raw_index is not None else 0
delta_id = getattr(tc_delta, "id", None) or ""
# Ollama fix: detect a new tool call reusing the same
# raw index (different id) and redirect to a fresh slot.
@@ -4195,7 +4198,7 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
if idx not in tool_calls_acc:
# Poolside may send integer id instead of string
_tc_id = tc_delta.id
_tc_id = getattr(tc_delta, "id", None)
if isinstance(_tc_id, int):
_tc_id = str(_tc_id)
tool_calls_acc[idx] = {
@@ -4205,14 +4208,17 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
"extra_content": None,
}
entry = tool_calls_acc[idx]
if tc_delta.id is not None:
_new_id = tc_delta.id
tc_id = getattr(tc_delta, "id", None)
if tc_id is not None:
_new_id = tc_id
if isinstance(_new_id, int):
_new_id = str(_new_id)
if _new_id:
entry["id"] = _new_id
if tc_delta.function:
if tc_delta.function.name:
tc_function = getattr(tc_delta, "function", None)
if tc_function:
function_name = getattr(tc_function, "name", None)
if function_name:
# Use assignment, not +=. Function names are
# atomic identifiers delivered complete in the
# first chunk (OpenAI spec). Some providers
@@ -4221,9 +4227,10 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
# produce "read_fileread_file". Assignment
# (matching the OpenAI Node SDK / LiteLLM /
# Vercel AI patterns) is immune to this.
entry["function"]["name"] = tc_delta.function.name
if tc_delta.function.arguments:
entry["function"]["arguments"] += tc_delta.function.arguments
entry["function"]["name"] = function_name
function_arguments = getattr(tc_function, "arguments", None)
if function_arguments:
entry["function"]["arguments"] += function_arguments
extra = getattr(tc_delta, "extra_content", None)
if extra is None and hasattr(tc_delta, "model_extra"):
extra = (tc_delta.model_extra if isinstance(tc_delta.model_extra, dict) else {}).get("extra_content")
@@ -4249,8 +4256,9 @@ def interruptible_streaming_api_call(agent, api_kwargs: dict, *, on_first_delta=
# discarding the attempted action.
result["partial_tool_names"].append(name)
if chunk.choices[0].finish_reason:
finish_reason = chunk.choices[0].finish_reason
chunk_finish_reason = getattr(chunk.choices[0], "finish_reason", None)
if chunk_finish_reason:
finish_reason = chunk_finish_reason
# Usage in the final chunk
if hasattr(chunk, "usage") and chunk.usage:
agent/transports/chat_completions.py
+20 -8
View File
@@ -854,17 +854,25 @@ class ChatCompletionsTransport(ProviderTransport):
preserved for downstream replay.
"""
choice = response.choices[0]
msg = choice.message
msg = getattr(choice, "message", None)
# Poolside returns integer finish_reason (e.g. 24) instead of string
_fr = choice.finish_reason
_fr = getattr(choice, "finish_reason", None)
if isinstance(_fr, int):
_fr = str(_fr)
finish_reason = _fr or "stop"
tool_calls = None
if msg.tool_calls:
message_tool_calls = getattr(msg, "tool_calls", None)
if message_tool_calls:
tool_calls = []
for tc in msg.tool_calls:
for tc in message_tool_calls:
tc_function = getattr(tc, "function", None)
function_name = getattr(tc_function, "name", None)
# Match Relay's codec: skip absent function/name fields, but
# preserve an explicit blank name for Hermes's recovery path.
if tc_function is None or function_name is None:
continue
function_arguments = getattr(tc_function, "arguments", None)
# Preserve provider-specific extras on the tool call.
# Gemini 3 thinking models attach extra_content with
# thought_signature — without replay on the next turn the API
@@ -887,9 +895,13 @@ class ChatCompletionsTransport(ProviderTransport):
tc_provider_data["extra_content"] = extra
tool_calls.append(
ToolCall(
id=tc.id,
name=tc.function.name,
arguments=tc.function.arguments,
id=getattr(tc, "id", None),
name=function_name,
arguments=(
function_arguments
if function_arguments is not None
else "{}"
),
provider_data=tc_provider_data or None,
)
)
@@ -932,7 +944,7 @@ class ChatCompletionsTransport(ProviderTransport):
# Promote it to content + a ``content_filter`` finish reason so the
# loop's refusal handler surfaces it clearly and stops. ``refusal`` is
# ``None`` for normal responses, so this is a no-op in the common case.
content = msg.content
content = getattr(msg, "content", None)
refusal = getattr(msg, "refusal", None)
if refusal is None and hasattr(msg, "model_extra"):
_msg_extra = getattr(msg, "model_extra", None) or {}
tests/agent/transports/test_chat_completions.py
+62
View File
@@ -18,8 +18,70 @@ def transport():
class TestChatCompletionsBasic:
@pytest.mark.parametrize(
"choice",
[SimpleNamespace(message=SimpleNamespace()), SimpleNamespace()],
)
def test_normalize_response_allows_missing_optional_message_fields(
self, transport, choice
):
response = SimpleNamespace(choices=[choice], usage=None)
normalized = transport.normalize_response(response)
assert normalized.content is None
assert normalized.tool_calls is None
assert normalized.finish_reason == "stop"
def test_normalize_response_allows_sparse_tool_call_fields(self, transport):
response = SimpleNamespace(
choices=[
SimpleNamespace(
message=SimpleNamespace(
tool_calls=[
SimpleNamespace(
function=SimpleNamespace(arguments='{"city":"Paris"}')
),
SimpleNamespace(),
SimpleNamespace(
id="call-3",
function=SimpleNamespace(name="lookup"),
),
SimpleNamespace(
id="call-4",
function=SimpleNamespace(name="", arguments="{}"),
),
SimpleNamespace(
function=SimpleNamespace(
name="weather", arguments="{}"
)
),
]
)
)
],
usage=None,
)
normalized = transport.normalize_response(response)
assert normalized.finish_reason == "stop"
assert normalized.tool_calls is not None
assert [tool.id for tool in normalized.tool_calls] == [
"call-3",
"call-4",
None,
]
assert [tool.name for tool in normalized.tool_calls] == [
"lookup",
"",
"weather",
]
assert [tool.arguments for tool in normalized.tool_calls] == [
"{}",
"{}",
"{}",
]
@pytest.mark.parametrize("provider", ["nous", "openrouter"])
def test_gpt56_ultra_uses_max_wire_effort(self, transport, provider):
tests/run_agent/test_streaming.py
+88 -1
View File
@@ -96,6 +96,94 @@ class TestStreamingAccumulator:
assert response.usage is not None
assert response.usage.completion_tokens == 3
@patch("run_agent.AIAgent._create_request_openai_client")
@patch("run_agent.AIAgent._close_request_openai_client")
def test_sparse_delta_allows_missing_optional_fields(self, mock_close, mock_create):
"""Managed stream deltas may omit both content and tool_calls."""
from run_agent import AIAgent
sparse_delta = SimpleNamespace(reasoning_content=None, reasoning=None)
chunks = [
SimpleNamespace(
choices=[
SimpleNamespace(
index=0,
delta=sparse_delta,
finish_reason=None,
)
],
model="test-model",
usage=None,
),
_make_stream_chunk(
content="done", finish_reason="stop", model="test-model"
),
]
mock_client = MagicMock()
mock_client.chat.completions.create.return_value = iter(chunks)
mock_create.return_value = mock_client
agent = AIAgent(
api_key="test-key",
base_url="https://openrouter.ai/api/v1",
model="test/model",
quiet_mode=True,
skip_context_files=True,
skip_memory=True,
)
agent.api_mode = "chat_completions"
agent._interrupt_requested = False
response = agent._interruptible_streaming_api_call({})
assert response.choices[0].message.content == "done"
assert response.choices[0].message.tool_calls is None
@patch("run_agent.AIAgent._create_request_openai_client")
@patch("run_agent.AIAgent._close_request_openai_client")
def test_sparse_tool_delta_allows_missing_nested_fields(
self, mock_close, mock_create
):
"""A partial tool delta may contain arguments before its other fields."""
from run_agent import AIAgent
sparse_tool_delta = SimpleNamespace(
index=0,
function=SimpleNamespace(arguments='{"city":"Paris"}'),
)
chunks = [
SimpleNamespace(
choices=[
SimpleNamespace(
index=0,
delta=SimpleNamespace(tool_calls=[sparse_tool_delta]),
)
],
model="test-model",
usage=None,
),
_make_stream_chunk(finish_reason="tool_calls", model="test-model"),
]
mock_client = MagicMock()
mock_client.chat.completions.create.return_value = iter(chunks)
mock_create.return_value = mock_client
agent = AIAgent(
api_key="test-key",
base_url="https://openrouter.ai/api/v1",
model="test/model",
quiet_mode=True,
skip_context_files=True,
skip_memory=True,
)
agent.api_mode = "chat_completions"
agent._interrupt_requested = False
response = agent._interruptible_streaming_api_call({})
tool_call = response.choices[0].message.tool_calls[0]
assert tool_call.function.name == ""
assert tool_call.function.arguments == '{"city":"Paris"}'
assert response.choices[0].finish_reason == "tool_calls"
@patch("run_agent.AIAgent._create_request_openai_client")
@patch("run_agent.AIAgent._close_request_openai_client")
def test_chat_stream_closes_original_provider_resource(
@@ -1634,4 +1722,3 @@ class TestBedrockReasoningStaleFloor:
from agent.chat_completion_helpers import _bedrock_reasoning_stale_floor
assert _bedrock_reasoning_stale_floor(model_id) == expected