Skip to content

Use Claude Code, Codex, VSCode, Pi, and OpenCode (and 6 other harnesses) for free (1.3B+ free tokens) from your terminal, app, IDE, or phone, and now from the browser with native browser sessions (multi-harness + multi-model) like OpenClaw (voice supported + ToS friendly)

Python

Fix null usage in SSE for OpenAI-compatible streams (#209, #123)

Alishahryar1abae61d85bc9
diff --git a/core/anthropic/sse.py b/core/anthropic/sse.py
index 3923f0f4..1a99f6c4 100644
--- a/core/anthropic/sse.py
+++ b/core/anthropic/sse.py
@@ -37,6 +37,11 @@ def map_stop_reason(openai_reason: str | None) -> str:
)
+def _safe_usage_int(value: object) -> int:
+ """Coerce streamed usage counters to int; non-integers become 0."""
+ return value if isinstance(value, int) else 0
+
+
def format_sse_event(event_type: str, data: dict) -> str:
"""Format one Anthropic-style SSE event (no logging)."""
return f"event: {event_type}\ndata: {json.dumps(data)}\n\n"
@@ -191,7 +196,8 @@ class SSEBuilder:
return event_str
def message_start(self) -> str:
- usage = {"input_tokens": self.input_tokens, "output_tokens": 1}
+ safe_input = _safe_usage_int(self.input_tokens)
+ usage = {"input_tokens": safe_input, "output_tokens": 1}
return self._format_event(
"message_start",
{
@@ -209,15 +215,17 @@ class SSEBuilder:
},
)
- def message_delta(self, stop_reason: str, output_tokens: int) -> str:
+ def message_delta(self, stop_reason: str, output_tokens: int | None) -> str:
+ safe_in = _safe_usage_int(self.input_tokens)
+ safe_out = output_tokens if isinstance(output_tokens, int) else 0
return self._format_event(
"message_delta",
{
"type": "message_delta",
"delta": {"stop_reason": stop_reason, "stop_sequence": None},
"usage": {
- "input_tokens": self.input_tokens,
- "output_tokens": output_tokens,
+ "input_tokens": safe_in,
+ "output_tokens": safe_out,
},
},
)
diff --git a/providers/openai_compat.py b/providers/openai_compat.py
index b01f5dad..e6aa0c5b 100644
--- a/providers/openai_compat.py
+++ b/providers/openai_compat.py
@@ -409,11 +409,15 @@ class OpenAIChatTransport(BaseProvider):
for event in sse.close_all_blocks():
yield event
- output_tokens = (
- usage_info.completion_tokens
- if usage_info and hasattr(usage_info, "completion_tokens")
- else sse.estimate_output_tokens()
+ completion = (
+ getattr(usage_info, "completion_tokens", None)
+ if usage_info is not None
+ else None
)
+ if isinstance(completion, int):
+ output_tokens = completion
+ else:
+ output_tokens = sse.estimate_output_tokens()
if usage_info and hasattr(usage_info, "prompt_tokens"):
provider_input = usage_info.prompt_tokens
if isinstance(provider_input, int):
diff --git a/tests/providers/test_sse_builder.py b/tests/providers/test_sse_builder.py
index 01426e5b..c68f0526 100644
--- a/tests/providers/test_sse_builder.py
+++ b/tests/providers/test_sse_builder.py
@@ -1,5 +1,6 @@
"""Tests for core.anthropic.sse."""
+from typing import Any, cast
from unittest.mock import patch
import pytest
@@ -109,6 +110,27 @@ class TestSSEBuilderMessageLifecycle:
assert data["delta"]["stop_reason"] == "end_turn"
assert data["usage"]["output_tokens"] == 42
+ def test_message_start_coerces_non_int_input_tokens(self):
+ builder = SSEBuilder("msg_1", "model", input_tokens=0)
+ builder.input_tokens = cast(Any, "not_an_int")
+ sse = builder.message_start()
+ data = _parse_sse(sse)
+ assert data["message"]["usage"]["input_tokens"] == 0
+ assert data["message"]["usage"]["output_tokens"] == 1
+
+ def test_message_delta_coerces_none_output_tokens(self):
+ builder = SSEBuilder("msg_1", "model", input_tokens=3)
+ sse = builder.message_delta("end_turn", None)
+ data = _parse_sse(sse)
+ assert data["usage"]["input_tokens"] == 3
+ assert data["usage"]["output_tokens"] == 0
+
+ def test_message_delta_preserves_zero_output_tokens(self):
+ builder = SSEBuilder("msg_1", "model")
+ sse = builder.message_delta("end_turn", 0)
+ data = _parse_sse(sse)
+ assert data["usage"]["output_tokens"] == 0
+
def test_message_stop(self):
builder = SSEBuilder("msg_1", "model")
sse = builder.message_stop()
diff --git a/tests/providers/test_streaming_errors.py b/tests/providers/test_streaming_errors.py
index 4bc5ace7..09113a74 100644
--- a/tests/providers/test_streaming_errors.py
+++ b/tests/providers/test_streaming_errors.py
@@ -302,6 +302,46 @@ class TestStreamingExceptionHandling:
assert '"text_delta"' in event_text
assert "message_stop" in event_text
+ @pytest.mark.asyncio
+ async def test_upstream_completion_tokens_null_emits_int_usage(self):
+ """NIM/GLM may send usage.completion_tokens=null; final SSE must not use JSON null."""
+ provider = _make_provider()
+ request = _make_request()
+
+ delta = SimpleNamespace(
+ content="hello",
+ tool_calls=None,
+ reasoning_content=None,
+ )
+ choice = SimpleNamespace(delta=delta, finish_reason="stop")
+ usage = SimpleNamespace(completion_tokens=None, prompt_tokens=None)
+ chunk = SimpleNamespace(choices=[choice], usage=usage)
+ stream_mock = AsyncStreamMock([chunk])
+
+ with (
+ patch.object(
+ provider._client.chat.completions,
+ "create",
+ new_callable=AsyncMock,
+ return_value=stream_mock,
+ ),
+ patch.object(
+ provider._global_rate_limiter,
+ "wait_if_blocked",
+ new_callable=AsyncMock,
+ return_value=False,
+ ),
+ ):
+ events = await _collect_stream(provider, request)
+
+ parsed = parse_sse_text("".join(events))
+ delta_events = [e for e in parsed if e.event == "message_delta"]
+ assert len(delta_events) == 1
+ usage_out = delta_events[0].data.get("usage", {})
+ assert isinstance(usage_out.get("output_tokens"), int)
+ assert usage_out["output_tokens"] is not None
+ assert '"output_tokens": null' not in "".join(events)
+
@pytest.mark.asyncio
async def test_reasoning_only_stream_emits_placeholder_text(self):
"""When the model streams only ``reasoning_content`` (no ``content``), add text block.