Skip to content

vllm.entrypoints.openai.responses.streaming_events

Streaming SSE event builders for the Responses API.

Pure functions that translate streaming state + delta data into OpenAI Response API SSE events. Used by the streaming event processors in serving.py.

The file is organized as
  1. StreamingState dataclass + utility helpers
  2. Shared leaf helpers — delta events (take plain strings, no context)
  3. Shared leaf helpers — done events (take plain strings, no context)
  4. Harmony-specific dispatchers (route ctx/previous_item → leaf helpers)
  5. Harmony-specific tool lifecycle helpers

SimpleStreamingEventProcessor

State-machine processor for the simple (non-Harmony) streaming path.

Core flow
  1. Resolve the target state from the delta_message (CONTENT / REASONING / TOOL_CALL).
  2. If the target state differs from the current one, close_current() then open() the new state.
  3. emit_delta() produces the incremental events for the state.
State lifecycle

open() -> repeated emit_delta() -> close_current()

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
class SimpleStreamingEventProcessor:
    """
    State-machine processor for the simple (non-Harmony) streaming path.

    Core flow:
      1. Resolve the target state from the delta_message
         (CONTENT / REASONING / TOOL_CALL).
      2. If the target state differs from the current one,
         close_current() then open() the new state.
      3. emit_delta() produces the incremental events for the state.

    State lifecycle:
      open()  ->  repeated emit_delta()  ->  close_current()
    """

    _STATE_HANDLERS: ClassVar[dict[_StateType, _StateHandlers]] = {
        _StateType.CONTENT: _StateHandlers(
            emit_simple_content_open,
            emit_simple_content_delta,
            emit_simple_content_done,
        ),
        _StateType.REASONING: _StateHandlers(
            emit_simple_reasoning_open,
            emit_simple_reasoning_delta,
            emit_simple_reasoning_done,
        ),
        _StateType.TOOL_CALL: _StateHandlers(
            emit_simple_tool_call_open,
            emit_simple_tool_call_delta,
            emit_simple_tool_call_done,
        ),
    }

    def __init__(self, state: SimpleStreamingState | None = None) -> None:
        self.state = state or SimpleStreamingState()

    def resolve_target_state(
        self, delta_message: DeltaMessage
    ) -> tuple[_StateType, Any]:
        """
        Decide which state the next delta belongs to.

        Priority: TOOL_CALL > REASONING > CONTENT, fallback to NONE.
        For TOOL_CALL the first tool_call object is also returned so
        callers can detect a switch between consecutive tools.
        """
        if (
            delta_message.tool_calls
            and delta_message.tool_calls[0].function is not None
        ):
            return _StateType.TOOL_CALL, delta_message.tool_calls[0]
        if delta_message.reasoning is not None:
            return _StateType.REASONING, None
        if delta_message.content:
            return _StateType.CONTENT, None
        return _StateType.NONE, None

    def needs_transition(self, target_state: _StateType, tool_call: Any) -> bool:
        """
        Return True when we must close the current state and open a new one.

        Two cases trigger a transition:
          1. The target state differs from the current state
             (e.g. CONTENT -> TOOL_CALL).
          2. We are already in TOOL_CALL but the next tool_call has a
             different index (multiple consecutive tool calls).
        """
        if self.state.current_state != target_state:
            return True
        return (
            target_state == _StateType.TOOL_CALL
            and tool_call is not None
            and self.state.tool_call_index is not None
            and tool_call.index is not None
            and self.state.tool_call_index != tool_call.index
        )

    def close_current(self) -> list[StreamingResponsesResponse]:
        """Close the current state and emit its 'done' event sequence."""
        handlers = self._STATE_HANDLERS.get(self.state.current_state)
        if handlers is None:
            return []
        return handlers.done_fn(self.state)

    def open(
        self, target_state: _StateType, tool_call: Any = None
    ) -> list[StreamingResponsesResponse]:
        """Open a new state and emit its 'added' / 'open' event sequence."""
        handlers = self._STATE_HANDLERS[target_state]
        if target_state == _StateType.TOOL_CALL:
            assert tool_call is not None
            return handlers.open_fn(
                self.state, tool_call.function.name, tool_call.index
            )
        return handlers.open_fn(self.state)

    def emit_delta(
        self,
        delta_message: DeltaMessage,
        output: CompletionOutput,
        get_logprobs: Callable[
            [CompletionOutput], list[response_text_delta_event.Logprob]
        ]
        | None = None,
    ) -> list[StreamingResponsesResponse]:
        """
        Emit incremental events for the current state from the delta.

        Special case: when already in REASONING and the same delta also
        carries content, we emit the reasoning delta, close reasoning,
        open content, and then emit the content delta.
        """
        handlers = self._STATE_HANDLERS[self.state.current_state]
        events: list[StreamingResponsesResponse] = []

        # Special case: reasoning -> content inside a single delta.
        if (
            self.state.current_state == _StateType.REASONING
            and delta_message.reasoning is not None
            and delta_message.content is not None
        ):
            events.extend(handlers.delta_fn(self.state, delta_message.reasoning))
            events.extend(self.close_current())
            events.extend(self.open(_StateType.CONTENT))
            content_handlers = self._STATE_HANDLERS[_StateType.CONTENT]
            logprobs = get_logprobs(output) if get_logprobs else []
            events.extend(
                content_handlers.delta_fn(self.state, delta_message.content, logprobs)
            )
            return events

        if self.state.current_state == _StateType.TOOL_CALL:
            assert delta_message.tool_calls is not None
            tool_call_function = delta_message.tool_calls[0].function
            assert tool_call_function is not None
            if tool_call_function.arguments:
                return handlers.delta_fn(self.state, tool_call_function.arguments)
            return []
        elif self.state.current_state == _StateType.REASONING:
            assert delta_message.reasoning is not None
            return handlers.delta_fn(self.state, delta_message.reasoning)
        elif self.state.current_state == _StateType.CONTENT:
            assert delta_message.content is not None
            logprobs = get_logprobs(output) if get_logprobs else []
            return handlers.delta_fn(self.state, delta_message.content, logprobs)
        return []

close_current 

close_current() -> list[StreamingResponsesResponse]

Close the current state and emit its 'done' event sequence.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
1190
1191
1192
1193
1194
1195
def close_current(self) -> list[StreamingResponsesResponse]:
    """Close the current state and emit its 'done' event sequence."""
    handlers = self._STATE_HANDLERS.get(self.state.current_state)
    if handlers is None:
        return []
    return handlers.done_fn(self.state)

emit_delta 

emit_delta(
    delta_message: DeltaMessage,
    output: CompletionOutput,
    get_logprobs: Callable[
        [CompletionOutput], list[Logprob]
    ]
    | None = None,
) -> list[StreamingResponsesResponse]

Emit incremental events for the current state from the delta.

Special case: when already in REASONING and the same delta also carries content, we emit the reasoning delta, close reasoning, open content, and then emit the content delta.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
def emit_delta(
    self,
    delta_message: DeltaMessage,
    output: CompletionOutput,
    get_logprobs: Callable[
        [CompletionOutput], list[response_text_delta_event.Logprob]
    ]
    | None = None,
) -> list[StreamingResponsesResponse]:
    """
    Emit incremental events for the current state from the delta.

    Special case: when already in REASONING and the same delta also
    carries content, we emit the reasoning delta, close reasoning,
    open content, and then emit the content delta.
    """
    handlers = self._STATE_HANDLERS[self.state.current_state]
    events: list[StreamingResponsesResponse] = []

    # Special case: reasoning -> content inside a single delta.
    if (
        self.state.current_state == _StateType.REASONING
        and delta_message.reasoning is not None
        and delta_message.content is not None
    ):
        events.extend(handlers.delta_fn(self.state, delta_message.reasoning))
        events.extend(self.close_current())
        events.extend(self.open(_StateType.CONTENT))
        content_handlers = self._STATE_HANDLERS[_StateType.CONTENT]
        logprobs = get_logprobs(output) if get_logprobs else []
        events.extend(
            content_handlers.delta_fn(self.state, delta_message.content, logprobs)
        )
        return events

    if self.state.current_state == _StateType.TOOL_CALL:
        assert delta_message.tool_calls is not None
        tool_call_function = delta_message.tool_calls[0].function
        assert tool_call_function is not None
        if tool_call_function.arguments:
            return handlers.delta_fn(self.state, tool_call_function.arguments)
        return []
    elif self.state.current_state == _StateType.REASONING:
        assert delta_message.reasoning is not None
        return handlers.delta_fn(self.state, delta_message.reasoning)
    elif self.state.current_state == _StateType.CONTENT:
        assert delta_message.content is not None
        logprobs = get_logprobs(output) if get_logprobs else []
        return handlers.delta_fn(self.state, delta_message.content, logprobs)
    return []

needs_transition 

needs_transition(
    target_state: _StateType, tool_call: Any
) -> bool

Return True when we must close the current state and open a new one.

Two cases trigger a transition
  1. The target state differs from the current state (e.g. CONTENT -> TOOL_CALL).
  2. We are already in TOOL_CALL but the next tool_call has a different index (multiple consecutive tool calls).
Source code in vllm/entrypoints/openai/responses/streaming_events.py 
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
def needs_transition(self, target_state: _StateType, tool_call: Any) -> bool:
    """
    Return True when we must close the current state and open a new one.

    Two cases trigger a transition:
      1. The target state differs from the current state
         (e.g. CONTENT -> TOOL_CALL).
      2. We are already in TOOL_CALL but the next tool_call has a
         different index (multiple consecutive tool calls).
    """
    if self.state.current_state != target_state:
        return True
    return (
        target_state == _StateType.TOOL_CALL
        and tool_call is not None
        and self.state.tool_call_index is not None
        and tool_call.index is not None
        and self.state.tool_call_index != tool_call.index
    )

open 

open(
    target_state: _StateType, tool_call: Any = None
) -> list[StreamingResponsesResponse]

Open a new state and emit its 'added' / 'open' event sequence.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
def open(
    self, target_state: _StateType, tool_call: Any = None
) -> list[StreamingResponsesResponse]:
    """Open a new state and emit its 'added' / 'open' event sequence."""
    handlers = self._STATE_HANDLERS[target_state]
    if target_state == _StateType.TOOL_CALL:
        assert tool_call is not None
        return handlers.open_fn(
            self.state, tool_call.function.name, tool_call.index
        )
    return handlers.open_fn(self.state)

resolve_target_state 

resolve_target_state(
    delta_message: DeltaMessage,
) -> tuple[_StateType, Any]

Decide which state the next delta belongs to.

Priority: TOOL_CALL > REASONING > CONTENT, fallback to NONE. For TOOL_CALL the first tool_call object is also returned so callers can detect a switch between consecutive tools.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
def resolve_target_state(
    self, delta_message: DeltaMessage
) -> tuple[_StateType, Any]:
    """
    Decide which state the next delta belongs to.

    Priority: TOOL_CALL > REASONING > CONTENT, fallback to NONE.
    For TOOL_CALL the first tool_call object is also returned so
    callers can detect a switch between consecutive tools.
    """
    if (
        delta_message.tool_calls
        and delta_message.tool_calls[0].function is not None
    ):
        return _StateType.TOOL_CALL, delta_message.tool_calls[0]
    if delta_message.reasoning is not None:
        return _StateType.REASONING, None
    if delta_message.content:
        return _StateType.CONTENT, None
    return _StateType.NONE, None

StreamingState dataclass

Mutable state for streaming event processing.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
@dataclass
class StreamingState:
    """Mutable state for streaming event processing."""

    current_content_index: int = -1
    current_output_index: int = 0
    current_item_id: str = ""
    current_call_id: str = ""
    sent_output_item_added: bool = False
    is_first_function_call_delta: bool = False

    def reset_for_new_item(self) -> None:
        """Reset state when expecting a new output item."""
        self.current_output_index += 1
        self.sent_output_item_added = False
        self.is_first_function_call_delta = False
        self.current_call_id = ""

reset_for_new_item 

reset_for_new_item() -> None

Reset state when expecting a new output item.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
106
107
108
109
110
111
def reset_for_new_item(self) -> None:
    """Reset state when expecting a new output item."""
    self.current_output_index += 1
    self.sent_output_item_added = False
    self.is_first_function_call_delta = False
    self.current_call_id = ""

_StateHandlers

Bases: NamedTuple

Tuple for each state: open(start), delta(chunk), done(finish).

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
1105
1106
1107
1108
1109
1110
class _StateHandlers(NamedTuple):
    """Tuple for each state: open(start), delta(chunk), done(finish)."""

    open_fn: Callable[..., list[StreamingResponsesResponse]]
    delta_fn: Callable[..., list[StreamingResponsesResponse]]
    done_fn: Callable[..., list[StreamingResponsesResponse]]

_resolve_mcp_name_label 

_resolve_mcp_name_label(recipient: str) -> tuple[str, str]

Resolve MCP tool name and server label from a recipient string.

  • mcp.* recipients: strip prefix, use the bare name as both name and server_label.
  • Everything else: use the recipient as the name and look up the server_label in TOOL_NAME_TO_MCP_SERVER_LABEL.
Source code in vllm/entrypoints/openai/responses/streaming_events.py 
81
82
83
84
85
86
87
88
89
90
91
92
def _resolve_mcp_name_label(recipient: str) -> tuple[str, str]:
    """Resolve MCP tool name and server label from a recipient string.

    - ``mcp.*`` recipients: strip prefix, use the bare name as both
      name and server_label.
    - Everything else: use the recipient as the name and look up the
      server_label in TOOL_NAME_TO_MCP_SERVER_LABEL.
    """
    if recipient.startswith("mcp."):
        name = recipient[len("mcp.") :]
        return name, name
    return recipient, TOOL_NAME_TO_MCP_SERVER_LABEL.get(recipient, recipient)

emit_browser_tool_events 

emit_browser_tool_events(
    previous_item: Message, state: StreamingState
) -> list[StreamingResponsesResponse]

Emit events for browser tool calls (web search).

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
def emit_browser_tool_events(
    previous_item: HarmonyMessage,
    state: StreamingState,
) -> list[StreamingResponsesResponse]:
    """Emit events for browser tool calls (web search)."""
    function_name = previous_item.recipient[len("browser.") :]
    parsed_args = json.loads(previous_item.content[0].text)
    action = None

    if function_name == "search":
        action = response_function_web_search.ActionSearch(
            type="search",
            query=parsed_args["query"],
        )
    elif function_name == "open":
        action = response_function_web_search.ActionOpenPage(
            type="open_page",
            # TODO: translate to url
            url=f"cursor:{parsed_args.get('cursor', '')}",
        )
    elif function_name == "find":
        action = response_function_web_search.ActionFind(
            type="find",
            pattern=parsed_args["pattern"],
            # TODO: translate to url
            url=f"cursor:{parsed_args.get('cursor', '')}",
        )
    else:
        raise ValueError(f"Unknown function name: {function_name}")

    state.current_item_id = f"tool_{random_uuid()}"
    events: list[StreamingResponsesResponse] = []
    events.append(
        ResponseOutputItemAddedEvent(
            type="response.output_item.added",
            sequence_number=-1,
            output_index=state.current_output_index,
            item=response_function_web_search.ResponseFunctionWebSearch(
                # TODO: generate a unique id for web search call
                type="web_search_call",
                id=state.current_item_id,
                action=action,
                status="in_progress",
            ),
        )
    )
    events.append(
        ResponseWebSearchCallInProgressEvent(
            type="response.web_search_call.in_progress",
            sequence_number=-1,
            output_index=state.current_output_index,
            item_id=state.current_item_id,
        )
    )
    events.append(
        ResponseWebSearchCallSearchingEvent(
            type="response.web_search_call.searching",
            sequence_number=-1,
            output_index=state.current_output_index,
            item_id=state.current_item_id,
        )
    )
    # enqueue
    events.append(
        ResponseWebSearchCallCompletedEvent(
            type="response.web_search_call.completed",
            sequence_number=-1,
            output_index=state.current_output_index,
            item_id=state.current_item_id,
        )
    )
    events.append(
        ResponseOutputItemDoneEvent(
            type="response.output_item.done",
            sequence_number=-1,
            output_index=state.current_output_index,
            item=ResponseFunctionWebSearch(
                type="web_search_call",
                id=state.current_item_id,
                action=action,
                status="completed",
            ),
        )
    )
    return events

emit_code_interpreter_completion_events 

emit_code_interpreter_completion_events(
    previous_item: Message, state: StreamingState
) -> list[StreamingResponsesResponse]

Emit events when code interpreter completes.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
def emit_code_interpreter_completion_events(
    previous_item: HarmonyMessage,
    state: StreamingState,
) -> list[StreamingResponsesResponse]:
    """Emit events when code interpreter completes."""
    events: list[StreamingResponsesResponse] = []
    events.append(
        ResponseCodeInterpreterCallCodeDoneEvent(
            type="response.code_interpreter_call_code.done",
            sequence_number=-1,
            output_index=state.current_output_index,
            item_id=state.current_item_id,
            code=previous_item.content[0].text,
        )
    )
    events.append(
        ResponseCodeInterpreterCallInterpretingEvent(
            type="response.code_interpreter_call.interpreting",
            sequence_number=-1,
            output_index=state.current_output_index,
            item_id=state.current_item_id,
        )
    )
    events.append(
        ResponseCodeInterpreterCallCompletedEvent(
            type="response.code_interpreter_call.completed",
            sequence_number=-1,
            output_index=state.current_output_index,
            item_id=state.current_item_id,
        )
    )
    events.append(
        ResponseOutputItemDoneEvent(
            type="response.output_item.done",
            sequence_number=-1,
            output_index=state.current_output_index,
            item=ResponseCodeInterpreterToolCallParam(
                type="code_interpreter_call",
                id=state.current_item_id,
                code=previous_item.content[0].text,
                container_id="auto",
                outputs=[],
                status="completed",
            ),
        )
    )
    return events

emit_code_interpreter_delta_events 

emit_code_interpreter_delta_events(
    delta: str, state: StreamingState
) -> list[StreamingResponsesResponse]

Emit events for code interpreter delta streaming.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
def emit_code_interpreter_delta_events(
    delta: str,
    state: StreamingState,
) -> list[StreamingResponsesResponse]:
    """Emit events for code interpreter delta streaming."""
    events: list[StreamingResponsesResponse] = []
    if not state.sent_output_item_added:
        state.sent_output_item_added = True
        state.current_item_id = f"tool_{random_uuid()}"
        events.append(
            ResponseOutputItemAddedEvent(
                type="response.output_item.added",
                sequence_number=-1,
                output_index=state.current_output_index,
                item=ResponseCodeInterpreterToolCallParam(
                    type="code_interpreter_call",
                    id=state.current_item_id,
                    code=None,
                    container_id="auto",
                    outputs=None,
                    status="in_progress",
                ),
            )
        )
        events.append(
            ResponseCodeInterpreterCallInProgressEvent(
                type="response.code_interpreter_call.in_progress",
                sequence_number=-1,
                output_index=state.current_output_index,
                item_id=state.current_item_id,
            )
        )
    events.append(
        ResponseCodeInterpreterCallCodeDeltaEvent(
            type="response.code_interpreter_call_code.delta",
            sequence_number=-1,
            output_index=state.current_output_index,
            item_id=state.current_item_id,
            delta=delta,
        )
    )
    return events

emit_content_delta_events 

emit_content_delta_events(
    ctx: StreamingHarmonyContext, state: StreamingState
) -> list[StreamingResponsesResponse]

Emit events for content delta streaming based on channel type.

This is a Harmony-specific dispatcher that extracts values from the Harmony context and delegates to shared leaf helpers.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
def emit_content_delta_events(
    ctx: StreamingHarmonyContext,
    state: StreamingState,
) -> list[StreamingResponsesResponse]:
    """Emit events for content delta streaming based on channel type.

    This is a Harmony-specific dispatcher that extracts values from the
    Harmony context and delegates to shared leaf helpers.
    """
    delta = ctx.last_content_delta
    if not delta:
        return []

    channel = ctx.parser.current_channel
    recipient = ctx.parser.current_recipient

    if channel in ("final", "commentary") and recipient is None:
        # Preambles (commentary with no recipient) and final messages
        # are both user-visible text.
        return emit_text_delta_events(delta, state)
    elif channel == "analysis" and recipient is None:
        return emit_reasoning_delta_events(delta, state)
    # built-in tools will be triggered on the analysis channel
    # However, occasionally built-in tools will
    # still be output to commentary.
    elif channel in ("commentary", "analysis") and recipient is not None:
        if recipient.startswith("functions."):
            function_name = recipient[len("functions.") :]
            return emit_function_call_delta_events(delta, function_name, state)
        elif recipient == "python":
            return emit_code_interpreter_delta_events(delta, state)
        elif recipient.startswith("mcp.") or is_mcp_tool_by_namespace(recipient):
            return emit_mcp_delta_events(delta, state, recipient)

    return []

emit_function_call_delta_events 

emit_function_call_delta_events(
    delta: str, function_name: str, state: StreamingState
) -> list[StreamingResponsesResponse]

Emit events for function call argument deltas.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
def emit_function_call_delta_events(
    delta: str,
    function_name: str,
    state: StreamingState,
) -> list[StreamingResponsesResponse]:
    """Emit events for function call argument deltas."""
    events: list[StreamingResponsesResponse] = []
    if state.is_first_function_call_delta is False:
        state.is_first_function_call_delta = True
        state.current_item_id = f"fc_{random_uuid()}"
        state.current_call_id = f"call_{random_uuid()}"
        tool_call_item = ResponseFunctionToolCall(
            name=function_name,
            type="function_call",
            id=state.current_item_id,
            call_id=state.current_call_id,
            arguments="",
            status="in_progress",
        )
        events.append(
            ResponseOutputItemAddedEvent(
                type="response.output_item.added",
                sequence_number=-1,
                output_index=state.current_output_index,
                item=tool_call_item,
            )
        )
    # Always emit the delta (including on first call)
    events.append(
        ResponseFunctionCallArgumentsDeltaEvent(
            item_id=state.current_item_id,
            delta=delta,
            output_index=state.current_output_index,
            sequence_number=-1,
            type="response.function_call_arguments.delta",
        )
    )
    return events

emit_function_call_done_events 

emit_function_call_done_events(
    function_name: str,
    arguments: str,
    state: StreamingState,
) -> list[StreamingResponsesResponse]

Emit events when a function call completes.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
def emit_function_call_done_events(
    function_name: str,
    arguments: str,
    state: StreamingState,
) -> list[StreamingResponsesResponse]:
    """Emit events when a function call completes."""
    events: list[StreamingResponsesResponse] = []
    events.append(
        ResponseFunctionCallArgumentsDoneEvent(
            type="response.function_call_arguments.done",
            arguments=arguments,
            name=function_name,
            item_id=state.current_item_id,
            output_index=state.current_output_index,
            sequence_number=-1,
        )
    )
    function_call_item = ResponseFunctionToolCall(
        type="function_call",
        arguments=arguments,
        name=function_name,
        item_id=state.current_item_id,
        output_index=state.current_output_index,
        sequence_number=-1,
        call_id=state.current_call_id,
        status="completed",
    )
    events.append(
        ResponseOutputItemDoneEvent(
            type="response.output_item.done",
            sequence_number=-1,
            output_index=state.current_output_index,
            item=function_call_item,
        )
    )
    return events

emit_mcp_completion_events 

emit_mcp_completion_events(
    recipient: str, arguments: str, state: StreamingState
) -> list[StreamingResponsesResponse]

Emit events when an MCP tool call completes.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
def emit_mcp_completion_events(
    recipient: str,
    arguments: str,
    state: StreamingState,
) -> list[StreamingResponsesResponse]:
    """Emit events when an MCP tool call completes."""
    name, server_label = _resolve_mcp_name_label(recipient)
    events: list[StreamingResponsesResponse] = []
    events.append(
        ResponseMcpCallArgumentsDoneEvent(
            type="response.mcp_call_arguments.done",
            arguments=arguments,
            name=name,
            item_id=state.current_item_id,
            output_index=state.current_output_index,
            sequence_number=-1,
        )
    )
    events.append(
        ResponseMcpCallCompletedEvent(
            type="response.mcp_call.completed",
            sequence_number=-1,
            output_index=state.current_output_index,
            item_id=state.current_item_id,
        )
    )
    events.append(
        ResponseOutputItemDoneEvent(
            type="response.output_item.done",
            sequence_number=-1,
            output_index=state.current_output_index,
            item=McpCall(
                type="mcp_call",
                arguments=arguments,
                name=name,
                id=state.current_item_id,
                server_label=server_label,
                status="completed",
            ),
        )
    )
    return events

emit_mcp_delta_events 

emit_mcp_delta_events(
    delta: str, state: StreamingState, recipient: str
) -> list[StreamingResponsesResponse]

Emit events for MCP tool delta streaming.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
def emit_mcp_delta_events(
    delta: str,
    state: StreamingState,
    recipient: str,
) -> list[StreamingResponsesResponse]:
    """Emit events for MCP tool delta streaming."""
    name, server_label = _resolve_mcp_name_label(recipient)
    events: list[StreamingResponsesResponse] = []
    if not state.sent_output_item_added:
        state.sent_output_item_added = True
        state.current_item_id = f"mcp_{random_uuid()}"
        events.append(
            ResponseOutputItemAddedEvent(
                type="response.output_item.added",
                sequence_number=-1,
                output_index=state.current_output_index,
                item=McpCall(
                    type="mcp_call",
                    id=state.current_item_id,
                    name=name,
                    arguments="",
                    server_label=server_label,
                    status="in_progress",
                ),
            )
        )
        events.append(
            ResponseMcpCallInProgressEvent(
                type="response.mcp_call.in_progress",
                sequence_number=-1,
                output_index=state.current_output_index,
                item_id=state.current_item_id,
            )
        )
    events.append(
        ResponseMcpCallArgumentsDeltaEvent(
            type="response.mcp_call_arguments.delta",
            sequence_number=-1,
            output_index=state.current_output_index,
            item_id=state.current_item_id,
            delta=delta,
        )
    )
    return events

emit_previous_item_done_events 

emit_previous_item_done_events(
    previous_item: Message, state: StreamingState
) -> list[StreamingResponsesResponse]

Emit done events for the previous item when expecting a new start.

This is a Harmony-specific dispatcher that extracts values from the Harmony parser's message object and delegates to shared leaf helpers.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
def emit_previous_item_done_events(
    previous_item: HarmonyMessage,
    state: StreamingState,
) -> list[StreamingResponsesResponse]:
    """Emit done events for the previous item when expecting a new start.

    This is a Harmony-specific dispatcher that extracts values from the
    Harmony parser's message object and delegates to shared leaf helpers.
    """
    text = previous_item.content[0].text
    if previous_item.recipient is not None:
        # Deal with tool call
        if previous_item.recipient.startswith("functions."):
            function_name = previous_item.recipient[len("functions.") :]
            return emit_function_call_done_events(function_name, text, state)
        elif previous_item.recipient == "python":
            return emit_code_interpreter_completion_events(previous_item, state)
        elif (
            is_mcp_tool_by_namespace(previous_item.recipient)
            and state.current_item_id is not None
            and state.current_item_id.startswith("mcp_")
        ):
            return emit_mcp_completion_events(previous_item.recipient, text, state)
    elif previous_item.channel == "analysis":
        return emit_reasoning_done_events(text, state)
    elif previous_item.channel in ("commentary", "final"):
        # Preambles (commentary with no recipient) and final messages
        # are both user-visible text.
        return emit_text_output_done_events(text, state)
    return []

emit_reasoning_delta_events 

emit_reasoning_delta_events(
    delta: str, state: StreamingState
) -> list[StreamingResponsesResponse]

Emit events for reasoning text delta streaming.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
def emit_reasoning_delta_events(
    delta: str,
    state: StreamingState,
) -> list[StreamingResponsesResponse]:
    """Emit events for reasoning text delta streaming."""
    events: list[StreamingResponsesResponse] = []
    if not state.sent_output_item_added:
        state.sent_output_item_added = True
        state.current_item_id = f"msg_{random_uuid()}"
        events.append(
            ResponseOutputItemAddedEvent(
                type="response.output_item.added",
                sequence_number=-1,
                output_index=state.current_output_index,
                item=ResponseReasoningItem(
                    type="reasoning",
                    id=state.current_item_id,
                    summary=[],
                    status="in_progress",
                ),
            )
        )
        state.current_content_index += 1
        events.append(
            ResponseReasoningPartAddedEvent(
                type="response.reasoning_part.added",
                sequence_number=-1,
                output_index=state.current_output_index,
                item_id=state.current_item_id,
                content_index=state.current_content_index,
                part=ResponseReasoningTextContent(
                    text="",
                    type="reasoning_text",
                ),
            )
        )
    events.append(
        ResponseReasoningTextDeltaEvent(
            type="response.reasoning_text.delta",
            item_id=state.current_item_id,
            output_index=state.current_output_index,
            content_index=state.current_content_index,
            delta=delta,
            sequence_number=-1,
        )
    )
    return events

emit_reasoning_done_events 

emit_reasoning_done_events(
    text: str, state: StreamingState
) -> list[StreamingResponsesResponse]

Emit events when a reasoning (analysis) item completes.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
def emit_reasoning_done_events(
    text: str,
    state: StreamingState,
) -> list[StreamingResponsesResponse]:
    """Emit events when a reasoning (analysis) item completes."""
    content = ResponseReasoningTextContent(
        text=text,
        type="reasoning_text",
    )
    reasoning_item = ResponseReasoningItem(
        type="reasoning",
        content=[content],
        status="completed",
        id=state.current_item_id,
        summary=[],
    )
    events: list[StreamingResponsesResponse] = []
    events.append(
        ResponseReasoningTextDoneEvent(
            type="response.reasoning_text.done",
            item_id=state.current_item_id,
            sequence_number=-1,
            output_index=state.current_output_index,
            content_index=state.current_content_index,
            text=text,
        )
    )
    events.append(
        ResponseReasoningPartDoneEvent(
            type="response.reasoning_part.done",
            sequence_number=-1,
            item_id=state.current_item_id,
            output_index=state.current_output_index,
            content_index=state.current_content_index,
            part=content,
        )
    )
    events.append(
        ResponseOutputItemDoneEvent(
            type="response.output_item.done",
            sequence_number=-1,
            output_index=state.current_output_index,
            item=reasoning_item,
        )
    )
    return events

emit_text_delta_events 

emit_text_delta_events(
    delta: str, state: StreamingState
) -> list[StreamingResponsesResponse]

Emit events for text content delta streaming.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
def emit_text_delta_events(
    delta: str,
    state: StreamingState,
) -> list[StreamingResponsesResponse]:
    """Emit events for text content delta streaming."""
    events: list[StreamingResponsesResponse] = []
    if not state.sent_output_item_added:
        state.sent_output_item_added = True
        state.current_item_id = f"msg_{random_uuid()}"
        events.append(
            ResponseOutputItemAddedEvent(
                type="response.output_item.added",
                sequence_number=-1,
                output_index=state.current_output_index,
                item=ResponseOutputMessage(
                    id=state.current_item_id,
                    type="message",
                    role="assistant",
                    content=[],
                    status="in_progress",
                ),
            )
        )
        state.current_content_index += 1
        events.append(
            ResponseContentPartAddedEvent(
                type="response.content_part.added",
                sequence_number=-1,
                output_index=state.current_output_index,
                item_id=state.current_item_id,
                content_index=state.current_content_index,
                part=ResponseOutputText(
                    type="output_text",
                    text="",
                    annotations=[],
                    logprobs=[],
                ),
            )
        )
    events.append(
        ResponseTextDeltaEvent(
            type="response.output_text.delta",
            sequence_number=-1,
            content_index=state.current_content_index,
            output_index=state.current_output_index,
            item_id=state.current_item_id,
            delta=delta,
            # TODO, use logprobs from ctx.last_request_output
            logprobs=[],
        )
    )
    return events

emit_text_output_done_events 

emit_text_output_done_events(
    text: str, state: StreamingState
) -> list[StreamingResponsesResponse]

Emit events when a final text output item completes.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
def emit_text_output_done_events(
    text: str,
    state: StreamingState,
) -> list[StreamingResponsesResponse]:
    """Emit events when a final text output item completes."""
    text_content = ResponseOutputText(
        type="output_text",
        text=text,
        annotations=[],
    )
    events: list[StreamingResponsesResponse] = []
    events.append(
        ResponseTextDoneEvent(
            type="response.output_text.done",
            sequence_number=-1,
            output_index=state.current_output_index,
            content_index=state.current_content_index,
            text=text,
            logprobs=[],
            item_id=state.current_item_id,
        )
    )
    events.append(
        ResponseContentPartDoneEvent(
            type="response.content_part.done",
            sequence_number=-1,
            item_id=state.current_item_id,
            output_index=state.current_output_index,
            content_index=state.current_content_index,
            part=text_content,
        )
    )
    events.append(
        ResponseOutputItemDoneEvent(
            type="response.output_item.done",
            sequence_number=-1,
            output_index=state.current_output_index,
            item=ResponseOutputMessage(
                id=state.current_item_id,
                type="message",
                role="assistant",
                content=[text_content],
                status="completed",
            ),
        )
    )
    return events

emit_tool_action_events 

emit_tool_action_events(
    ctx: StreamingHarmonyContext,
    state: StreamingState,
    tool_server: ToolServer | None,
) -> list[StreamingResponsesResponse]

Emit events for tool action turn.

Source code in vllm/entrypoints/openai/responses/streaming_events.py 
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
def emit_tool_action_events(
    ctx: StreamingHarmonyContext,
    state: StreamingState,
    tool_server: ToolServer | None,
) -> list[StreamingResponsesResponse]:
    """Emit events for tool action turn."""
    if not ctx.is_assistant_action_turn() or len(ctx.parser.messages) == 0:
        return []

    events: list[StreamingResponsesResponse] = []
    previous_item = ctx.parser.messages[-1]

    # Handle browser tool
    if (
        tool_server is not None
        and tool_server.has_tool("browser")
        and previous_item.recipient is not None
        and previous_item.recipient.startswith("browser.")
    ):
        events.extend(emit_browser_tool_events(previous_item, state))

    # Handle tool completion
    if (
        tool_server is not None
        and previous_item.recipient is not None
        and state.current_item_id is not None
        and state.sent_output_item_added
    ):
        recipient = previous_item.recipient
        if recipient == "python":
            events.extend(emit_code_interpreter_completion_events(previous_item, state))
        elif recipient.startswith("mcp.") or is_mcp_tool_by_namespace(recipient):
            events.extend(
                emit_mcp_completion_events(
                    recipient, previous_item.content[0].text, state
                )
            )

    return events

is_mcp_tool_by_namespace 

is_mcp_tool_by_namespace(recipient: str | None) -> bool

Determine if a tool call is an MCP tool based on recipient prefix.

  • Tools starting with "functions." are function calls
  • Everything else is an MCP tool
Source code in vllm/entrypoints/openai/responses/streaming_events.py 
114
115
116
117
118
119
120
121
122
123
124
125
126
def is_mcp_tool_by_namespace(recipient: str | None) -> bool:
    """
    Determine if a tool call is an MCP tool based on recipient prefix.

    - Tools starting with "functions." are function calls
    - Everything else is an MCP tool
    """
    if recipient is None:
        return False

    # Function calls have "functions." prefix
    # Everything else is an MCP tool
    return not recipient.startswith("functions.")