wx44wx/ai-video-fullstack
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: add deterministic message interaction stages

Browse Source
This commit is contained in:
Xin Wang
2026-08-03 07:10:41 +08:00
parent 479a516546
commit 9daa46ed4d
26 changed files with 1356 additions and 414 deletions
Download Patch File Download Diff File Expand all files Collapse all files

100
backend/services/action_stage.py Normal file
View File

@@ -0,0 +1,100 @@
"""Shared deterministic stage for one or more tool Actions."""
from __future__ import annotations
from collections.abc import Awaitable, Callable
from dataclasses import dataclass, field
from typing import Any, Literal
from models import RuntimeTool
from services.action_runtime import ActionOutcome, ActionRunner, ActionStatus
InputPolicy = Literal["queue", "block"]
OutcomeHook = Callable[["StageAction", ActionOutcome], Awaitable[None]]
StartedHook = Callable[[], Awaitable[None]]
@dataclass(frozen=True)
class StageAction:
"""One deterministic tool invocation inside an Action stage."""
id: str
tool: RuntimeTool | None
arguments: dict[str, Any] = field(default_factory=dict)
result_assignments: dict[str, str] | None = None
required: bool = True
invocation_id: str | None = None
@dataclass(frozen=True)
class ActionStageSpec:
"""Mode-independent description produced by Prompt or Workflow config."""
actions: tuple[StageAction, ...] = ()
input_policy: InputPolicy = "queue"
@dataclass(frozen=True)
class ActionStageResult:
"""Ordered Action outcomes; optional failures do not fail the stage."""
succeeded: bool
outcomes: tuple[ActionOutcome, ...]
class ActionStageRunner:
"""Run deterministic tool Actions under one optional user-input gate."""
def __init__(self, actions: ActionRunner) -> None:
self._actions = actions
async def run(
self,
spec: ActionStageSpec,
*,
set_input_enabled: Callable[[bool], None] | None = None,
input_already_blocked: bool = False,
release_input_on_failure: bool = True,
on_started: StartedHook | None = None,
on_outcome: OutcomeHook | None = None,
) -> ActionStageResult:
input_setter = set_input_enabled
block_input = spec.input_policy == "block" and input_setter is not None
if block_input and not input_already_blocked:
input_setter(False)
result: ActionStageResult | None = None
try:
if on_started is not None:
await on_started()
outcomes: list[ActionOutcome] = []
succeeded = True
for action in spec.actions:
outcome = await self._actions.execute(
action.tool,
action.arguments,
result_assignments=action.result_assignments,
invocation_id=action.invocation_id,
)
outcomes.append(outcome)
if on_outcome is not None:
await on_outcome(action, outcome)
if outcome.status == ActionStatus.SUCCESS:
continue
if outcome.status == ActionStatus.CANCELLED or action.required:
succeeded = False
break
result = ActionStageResult(
succeeded=succeeded,
outcomes=tuple(outcomes),
)
return result
finally:
if block_input and (
release_input_on_failure
or (result is not None and result.succeeded)
):
input_setter(True)

2
backend/services/brains/base.py
View File

@@ -64,7 +64,7 @@ class CallEndPort(Protocol):
def arm_after_speech(self) -> None: ...
def track_speech(self) -> None: ...
def track_speech(self) -> Awaitable[None] | None: ...
async def arm_after_tracked_speech(self) -> None: ...

200
backend/services/brains/prompt_brain.py
View File

@@ -3,6 +3,7 @@
from __future__ import annotations
import asyncio
from collections.abc import Awaitable
from typing import Any
from uuid import uuid4
@@ -25,10 +26,18 @@ from services.brains.base import (
SessionVariableUpdate,
)
from services.action_runtime import (
ActionOutcome,
ActionInvocationCancelled,
ActionRunner,
ActionStatus,
)
from services.action_stage import ActionStageRunner, ActionStageSpec, StageAction
from services.fixed_speech import FixedSpeechOutput
from services.message_stage import (
MessageDisplaySpec,
MessageStageRunner,
MessageStageSpec,
)
from services.runtime_variables import DynamicVariableStore
from services.tool_executor import ToolExecutionError, ToolExecutor
from services.tool_policy import policy_for_tool
@@ -50,17 +59,31 @@ class PromptBrain(BaseBrain):
self._store = DynamicVariableStore.from_config(cfg)
self._tools = ToolExecutor(self._store)
self._actions = ActionRunner(self._tools)
self._action_stages = ActionStageRunner(self._actions)
self._message_stages = MessageStageRunner()
self._tool_by_id = {tool.id: tool for tool in cfg.tools}
self._runtime: BrainRuntime | None = None
self._output: FixedSpeechOutput | None = None
self._waiting_for_generated_end_speech = False
self._greeting_finished = True
self._preflight_finished = False
self._opening_started = False
self._opening_finished = False
self._opening_input_blocked = False
self._startup_failed = False
async def greeting(self, cfg: AssistantConfig) -> str:
return self._store.render(cfg.greeting) if self._dynamic_enabled else cfg.greeting
# The built-in opening Message owns the greeting so speech and the
# client dialog can start as one atomic stage.
if self._opening_message() is not None:
return ""
return self._render_greeting(cfg)
def _render_greeting(self, cfg: AssistantConfig) -> str:
return (
self._store.render(cfg.greeting)
if self._dynamic_enabled
else cfg.greeting
)
def system_prompt(self, cfg: AssistantConfig) -> str:
return self._store.render(cfg.prompt) if self._dynamic_enabled else cfg.prompt
@@ -77,12 +100,15 @@ class PromptBrain(BaseBrain):
self._tools,
is_session_ending=lambda: runtime.call_end.ending,
)
self._action_stages = ActionStageRunner(self._actions)
self._message_stages = MessageStageRunner(runtime.client_tools)
self._output = FixedSpeechOutput(self._store, runtime)
self._tool_by_id = {tool.id: tool for tool in cfg.tools}
self._waiting_for_generated_end_speech = False
self._greeting_finished = True
self._preflight_finished = False
self._opening_started = False
self._opening_finished = not bool(self._startup_actions("opening"))
self._opening_finished = not self._has_opening_stage()
self._opening_input_blocked = False
self._startup_failed = False
llm_tool_ids = (
set(cfg.llm_tool_ids) if cfg.llm_tool_ids is not None else None
@@ -119,32 +145,83 @@ class PromptBrain(BaseBrain):
self._preflight_finished = True
async def on_connected(self, *, greeting_pending: bool = False) -> None:
self._greeting_finished = not greeting_pending
if (
self._startup_actions("opening")
self._has_opening_stage()
and self._runtime is not None
and self._runtime.set_input_enabled is not None
):
self._runtime.set_input_enabled(False)
self._opening_input_blocked = True
async def on_client_ready(self) -> None:
if self._output is not None:
await self._output.mark_client_ready()
if self._opening_started or self._opening_finished or self._startup_failed:
return
self._opening_started = True
runtime = self._runtime
if runtime is None:
raise RuntimeError("PromptBrain 尚未初始化")
opening_message = self._opening_message()
opening_actions = self._startup_actions("opening")
speech = (
self._render_greeting(self._cfg).strip()
if opening_message is not None
else ""
)
if speech:
self.prepare_greeting_context(speech, runtime.context)
try:
succeeded = await self._run_startup_actions("opening")
if opening_message is not None:
message_result = await self._message_stages.run(
self._opening_message_stage_spec(speech, opening_message),
speak=self._speak_opening,
set_input_enabled=runtime.set_input_enabled,
input_already_blocked=self._opening_input_blocked,
release_input_on_success=not bool(opening_actions),
release_input_on_failure=False,
)
if not message_result.succeeded:
await self._fail_opening(
message_result.error or "开场消息显示失败"
)
return
if opening_actions:
result = await self._action_stages.run(
self._opening_actions_stage_spec(),
set_input_enabled=runtime.set_input_enabled,
input_already_blocked=self._opening_input_blocked,
release_input_on_failure=False,
on_outcome=self._publish_opening_outcome,
)
if not result.succeeded:
await self._fail_opening("必需的开场 Action 执行失败")
return
except ActionInvocationCancelled:
self._startup_failed = True
raise
if not succeeded:
await self._fail_opening("必需的开场 Action 执行失败")
return
self._opening_finished = True
self._release_startup_gate_if_ready()
self._opening_input_blocked = False
async def on_greeting_finished(self) -> None:
self._greeting_finished = True
self._release_startup_gate_if_ready()
async def _speak_opening(self, content: str) -> Awaitable[None] | None:
if self._output is None:
raise RuntimeError("Prompt 固定播报输出尚未初始化")
return await self._output.speak(
content,
source="prompt-opening-speech",
record_history=False,
)
def _opening_message(self) -> dict[str, Any] | None:
startup = self._cfg.startup if isinstance(self._cfg.startup, dict) else {}
value = startup.get("opening_message", startup.get("openingMessage"))
return value if isinstance(value, dict) else None
def _has_opening_stage(self) -> bool:
return self._opening_message() is not None or bool(
self._startup_actions("opening")
)
def _startup_actions(self, phase: str) -> list[dict[str, Any]]:
startup = self._cfg.startup if isinstance(self._cfg.startup, dict) else {}
@@ -154,6 +231,78 @@ class PromptBrain(BaseBrain):
if isinstance(action, dict) and action.get("phase", "opening") == phase
]
def _opening_actions_stage_spec(self) -> ActionStageSpec:
actions = tuple(
StageAction(
id=str(action.get("id") or "startup_action"),
tool=self._tool_by_id.get(
str(action.get("tool_id") or action.get("toolId") or "")
),
arguments=action.get("arguments") or {},
required=bool(action.get("required", True)),
invocation_id=self._actions.new_invocation_id(),
)
for action in self._startup_actions("opening")
)
return ActionStageSpec(
actions=actions,
input_policy="block",
)
def _opening_message_stage_spec(
self,
speech: str,
config: dict[str, Any],
) -> MessageStageSpec:
return MessageStageSpec(
speech=speech,
display=MessageDisplaySpec(
title=self._store.render(
str(config.get("title") or "重要提示")
).strip(),
message=self._store.render(
str(config.get("message") or "")
).strip(),
confirm_label=self._store.render(
str(
config.get("confirm_label")
or config.get("confirmLabel")
or "确认"
)
).strip(),
),
require_confirmation=True,
)
async def _publish_opening_outcome(
self,
action: StageAction,
outcome: ActionOutcome,
) -> None:
if outcome.updated_variables:
self._refresh_prompt()
if self._runtime is not None:
await self._runtime.queue_frame(
OutputTransportMessageUrgentFrame(
message={
"type": "startup-action-result",
"actionId": action.id,
"phase": "opening",
"outcome": outcome.trace_payload(),
}
)
)
if outcome.status == ActionStatus.FAILURE and action.required:
logger.warning(
f"必需的 Prompt opening Action 失败: "
f"action={action.id} error={outcome.error}"
)
elif outcome.status == ActionStatus.FAILURE:
logger.warning(
f"忽略可选 Prompt opening Action 失败: "
f"action={action.id} error={outcome.error}"
)
async def _run_startup_actions(self, phase: str) -> bool:
for action in self._startup_actions(phase):
action_id = str(action.get("id") or "startup_action")
@@ -170,17 +319,6 @@ class PromptBrain(BaseBrain):
)
if outcome.updated_variables:
self._refresh_prompt()
if phase == "opening" and self._runtime is not None:
await self._runtime.queue_frame(
OutputTransportMessageUrgentFrame(
message={
"type": "startup-action-result",
"actionId": action_id,
"phase": phase,
"outcome": outcome.trace_payload(),
}
)
)
if outcome.status == ActionStatus.SUCCESS:
continue
if outcome.status == ActionStatus.CANCELLED:
@@ -197,18 +335,6 @@ class PromptBrain(BaseBrain):
)
return True
def _release_startup_gate_if_ready(self) -> None:
runtime = self._runtime
if (
runtime is not None
and runtime.set_input_enabled is not None
and self._greeting_finished
and self._opening_finished
and not self._startup_failed
and not runtime.call_end.ending
):
runtime.set_input_enabled(True)
async def _fail_opening(self, message: str) -> None:
self._startup_failed = True
runtime = self._runtime

183
backend/services/brains/workflow_brain.py
View File

@@ -3,6 +3,7 @@
from __future__ import annotations
import asyncio
from collections.abc import Awaitable
from copy import deepcopy
from dataclasses import replace
from typing import Any
@@ -40,7 +41,14 @@ from services.action_runtime import (
ActionRunner,
ActionStatus,
)
from services.action_stage import ActionStageRunner, ActionStageSpec, StageAction
from services.knowledge import search as search_knowledge
from services.message_stage import (
MessageDisplaySpec,
MessageStageResult,
MessageStageRunner,
MessageStageSpec,
)
from services.runtime_variables import DynamicVariableStore
from services.tool_executor import ToolExecutionError, ToolExecutor
from services.tool_policy import policy_for_tool
@@ -110,6 +118,8 @@ class WorkflowBrain(BaseBrain):
self._store = DynamicVariableStore.from_config(cfg or AssistantConfig(type="workflow"))
self._tools = ToolExecutor(self._store)
self._actions = ActionRunner(self._tools)
self._action_stages = ActionStageRunner(self._actions)
self._message_stages = MessageStageRunner()
self._tool_by_id: dict[str, RuntimeTool] = {
tool.id: tool for tool in (cfg.tools if cfg else [])
}
@@ -126,11 +136,10 @@ class WorkflowBrain(BaseBrain):
self._output: WorkflowOutput | None = None
self._agent_stage: WorkflowAgentStage | None = None
self._ended = False
self._greeting_context_message: dict[str, str] | None = None
self._startup_waiting_for_greeting = False
async def greeting(self, cfg: AssistantConfig) -> str:
return self._engine.greeting(self._store) or cfg.greeting
async def greeting(self, _cfg: AssistantConfig) -> str:
"""Workflow opening speech belongs to an explicit Message or Agent."""
return ""
def system_prompt(self, cfg: AssistantConfig) -> str:
return self._store.render(self._engine.global_prompt())
@@ -151,6 +160,8 @@ class WorkflowBrain(BaseBrain):
self._tools,
is_session_ending=lambda: runtime.call_end.ending,
)
self._action_stages = ActionStageRunner(self._actions)
self._message_stages = MessageStageRunner(runtime.client_tools)
self._tool_by_id = {tool.id: tool for tool in cfg.tools}
self._router = WorkflowLLMRouter(cfg)
self._edge_evaluator = WorkflowEdgeEvaluator(
@@ -168,8 +179,6 @@ class WorkflowBrain(BaseBrain):
runtime=runtime,
)
self._ended = False
self._greeting_context_message = None
self._startup_waiting_for_greeting = False
self._manager = ConfiguredFlowManager(
worker=runtime.worker,
llm=runtime.llm,
@@ -179,15 +188,6 @@ class WorkflowBrain(BaseBrain):
)
self._manager.state["variables"] = self._store.values
def prepare_greeting_context(
self,
greeting: str,
context: LLMContext,
) -> dict[str, str] | None:
message = super().prepare_greeting_context(greeting, context)
self._greeting_context_message = deepcopy(message) if message else None
return message
async def on_connected(self, *, greeting_pending: bool = False) -> None:
self._state.enter(self._engine.start_id, WorkflowStatus.STARTING)
await self._emit_node_active(self._engine.start_id)
@@ -198,39 +198,11 @@ class WorkflowBrain(BaseBrain):
if self._manager is None:
raise RuntimeError("Workflow FlowManager 尚未初始化")
self._startup_waiting_for_greeting = greeting_pending
if greeting_pending:
# Keep the Workflow on Start until the transport confirms that the
# shared greeting has finished. This prevents an initial Agent's
# fixed speech (or generated reply) from racing the greeting.
await self._manager.initialize(
self._passive_node_config(self._engine.start_id)
)
logger.info("工作流等待 Start 开场白播放完毕")
return
node_config = await self._initial_node_config()
await self._manager.initialize(node_config)
await self._after_node_activated(node_config)
logger.info(f"工作流模式启用: 当前节点={self._manager.current_node}")
async def on_greeting_finished(self) -> None:
"""Enter the first node only after Start's greeting reaches playback end."""
if not self._startup_waiting_for_greeting or self._ended:
return
self._startup_waiting_for_greeting = False
manager = self._require_manager()
if manager.current_node != self._engine.start_id:
return
node_config = await self._initial_node_config()
if node_config.get("name") == self._engine.start_id:
self._state.enter(self._engine.start_id, WorkflowStatus.WAITING_USER)
return
await manager.set_node_from_config(node_config)
await self._after_node_activated(node_config)
logger.info(f"Start 开场白结束,进入节点: {manager.current_node}")
async def _initial_node_config(self) -> NodeConfig:
"""Only a default-only Start advances before the first user turn."""
outgoing = self._engine.outgoing(self._engine.start_id)
@@ -409,7 +381,6 @@ class WorkflowBrain(BaseBrain):
return self._require_agent_stage().node_config(
node_id,
functions=functions,
greeting_context_message=self._greeting_context_message,
leading_messages=leading_messages,
)
@@ -455,8 +426,8 @@ class WorkflowBrain(BaseBrain):
*,
source: str = "workflow-speech",
node_id: str | None = None,
) -> None:
await self._require_output().speak(
) -> Awaitable[None] | None:
return await self._require_output().speak(
text,
source=source,
node_id=node_id,
@@ -700,6 +671,14 @@ class WorkflowBrain(BaseBrain):
outcome = await self._enter_action(node_id)
if not outcome.should_route:
return self._passive_node_config(node_id, context_messages)
elif node_type == "message":
message_result = await self._enter_message(node_id)
if not message_result.succeeded:
return self._passive_node_config(node_id, context_messages)
if message_result.speech:
context_messages.append(
{"role": "assistant", "content": message_result.speech}
)
elif node_type == "handoff":
await self._enter_handoff(node_id)
elif node_type == "start":
@@ -735,29 +714,36 @@ class WorkflowBrain(BaseBrain):
data = self._engine.data(node_id)
runtime = self._require_runtime()
invocation_id = self._actions.new_invocation_id()
block_user_input = data.get("userInputPolicy") == "block"
if block_user_input and runtime.set_input_enabled:
# Blocking only suppresses new audio/text input while the Action
# runs. It deliberately does not cancel the tool. The default
# queue policy leaves input enabled; the turn lock serializes any
# completed user turn until this automatic path has finished.
runtime.set_input_enabled(False)
tool_id = str(data.get("toolId") or "")
tool = self._tool_by_id.get(tool_id)
try:
await self._emit_trace(
"action_started",
nodeId=node_id,
invocationId=invocation_id,
toolId=tool_id,
toolType=tool.type if tool else None,
)
outcome = await self._actions.execute(
tool,
data.get("arguments") or {},
result_assignments=self._action_result_assignments(data),
invocation_id=invocation_id,
stage_result = await self._action_stages.run(
ActionStageSpec(
actions=(
StageAction(
id=node_id,
tool=tool,
arguments=data.get("arguments") or {},
result_assignments=self._action_result_assignments(data),
invocation_id=invocation_id,
),
),
input_policy=(
"block"
if data.get("userInputPolicy") == "block"
else "queue"
),
),
set_input_enabled=runtime.set_input_enabled,
on_started=lambda: self._emit_trace(
"action_started",
nodeId=node_id,
invocationId=invocation_id,
toolId=tool_id,
toolType=tool.type if tool else None,
),
)
outcome = stage_result.outcomes[0]
updated_variables = list(outcome.updated_variables)
if updated_variables:
await self._emit_variables(
@@ -770,13 +756,74 @@ class WorkflowBrain(BaseBrain):
self._set_last_action(outcome)
await self._emit_action_outcome(node_id, outcome)
raise
finally:
if block_user_input and runtime.set_input_enabled:
runtime.set_input_enabled(True)
self._set_last_action(outcome)
await self._emit_action_outcome(node_id, outcome)
return outcome
async def _enter_message(self, node_id: str) -> MessageStageResult:
self._state.enter(node_id, WorkflowStatus.RUNNING_MESSAGE)
await self._emit_node_active(node_id)
data = self._engine.data(node_id)
runtime = self._require_runtime()
speech = self._store.render(str(data.get("speech") or "")).strip()
show_message = bool(data.get("showMessage", False))
require_confirmation = bool(data.get("requireConfirmation", False))
display = (
MessageDisplaySpec(
title=self._store.render(
str(data.get("title") or "重要提示")
).strip(),
message=self._store.render(
str(data.get("message") or "")
).strip(),
confirm_label=self._store.render(
str(data.get("confirmLabel") or "确认")
).strip(),
)
if show_message
else None
)
result = await self._message_stages.run(
MessageStageSpec(
speech=speech,
display=display,
require_confirmation=require_confirmation,
),
speak=lambda content: self._queue_visible_speech(
content,
source="workflow-message-speech",
node_id=node_id,
),
set_input_enabled=runtime.set_input_enabled,
on_started=lambda: self._emit_trace(
"message_started",
nodeId=node_id,
hasSpeech=bool(speech),
showsMessage=show_message,
requiresConfirmation=require_confirmation,
),
)
if result.succeeded:
await self._emit_trace(
"message_completed",
nodeId=node_id,
action=result.action,
)
return result
self._state.enter(node_id, WorkflowStatus.WAITING_USER)
await self._emit_trace(
"message_failed",
nodeId=node_id,
error=result.error or "Message 节点执行失败",
)
await self._require_output().emit_error(
result.error or "Message 节点执行失败",
node_id=node_id,
code="workflow_message_error",
)
return result
def _set_last_action(self, outcome: ActionOutcome) -> None:
legacy_status = {
ActionStatus.SUCCESS: "ok",

73
backend/services/fixed_speech.py Normal file
View File

@@ -0,0 +1,73 @@
"""Shared client-visible output for deterministic fixed speech."""
from __future__ import annotations
from collections.abc import Awaitable
from typing import Any
from pipecat.frames.frames import OutputTransportMessageUrgentFrame, TTSSpeakFrame
from pipecat.utils.time import time_now_iso8601
from services.brains.base import BrainRuntime
from services.runtime_variables import DynamicVariableStore
class FixedSpeechOutput:
"""Display and synthesize fixed speech without waiting for playback."""
def __init__(
self,
store: DynamicVariableStore,
runtime: BrainRuntime,
) -> None:
self._store = store
self._runtime = runtime
self._client_ready = False
self._pending_transcripts: list[dict[str, Any]] = []
async def mark_client_ready(self) -> None:
self._client_ready = True
pending = self._pending_transcripts
self._pending_transcripts = []
for message in pending:
await self.emit(message)
async def speak(
self,
text: str,
*,
source: str,
node_id: str | None = None,
record_history: bool = True,
) -> Awaitable[None] | None:
content = text.strip()
if not content:
return None
if record_history:
self._store.record("agent", content)
transcript = {
"type": "transcript",
"role": "assistant",
"content": content,
"timestamp": time_now_iso8601(),
"source": source,
**({"nodeId": node_id} if node_id else {}),
}
if self._client_ready:
await self.emit(transcript)
else:
self._pending_transcripts.append(transcript)
track_speech = getattr(self._runtime.call_end, "track_speech", None)
playback_completion: Awaitable[None] | None = None
if callable(track_speech):
playback_completion = track_speech()
await self._runtime.queue_frame(
TTSSpeakFrame(content, append_to_context=False)
)
return playback_completion
async def emit(self, message: dict[str, Any]) -> None:
await self._runtime.queue_frame(
OutputTransportMessageUrgentFrame(message=message)
)

197
backend/services/message_stage.py Normal file
View File

@@ -0,0 +1,197 @@
"""Deterministic speech and client-message interaction shared by all brains."""
from __future__ import annotations
import asyncio
from collections.abc import Awaitable, Callable
from dataclasses import dataclass
from services.client_tools import ClientToolError, ClientToolPort
BUILTIN_SHOW_MESSAGE = "show_message"
SpeechCompletion = Awaitable[None] | None
Speak = Callable[[str], Awaitable[SpeechCompletion]]
StartedHook = Callable[[], Awaitable[None]]
@dataclass(frozen=True)
class MessageDisplaySpec:
"""Content rendered by the platform-provided client message dialog."""
title: str
message: str
confirm_label: str = "确认"
@dataclass(frozen=True)
class MessageStageSpec:
"""Mode-independent fixed speech and optional client interaction."""
speech: str = ""
display: MessageDisplaySpec | None = None
require_confirmation: bool = False
@dataclass(frozen=True)
class MessageStageResult:
"""Result used by Workflow routing and Prompt opening failure handling."""
succeeded: bool
speech: str = ""
action: str | None = None
error: str | None = None
class MessageStageRunner:
"""Run one atomic user-visible message stage.
Speech is queued before the client message is dispatched. A confirmation
stage completes when the user confirms, even if audio is still playing. A
speech-only stage completes at the real transport playback boundary.
"""
def __init__(self, client_tools: ClientToolPort | None = None) -> None:
self._client_tools = client_tools
def set_client_tools(self, client_tools: ClientToolPort | None) -> None:
self._client_tools = client_tools
async def run(
self,
spec: MessageStageSpec,
*,
speak: Speak | None = None,
set_input_enabled: Callable[[bool], None] | None = None,
input_already_blocked: bool = False,
release_input_on_success: bool = True,
release_input_on_failure: bool = True,
on_started: StartedHook | None = None,
) -> MessageStageResult:
input_setter = set_input_enabled
if input_setter is not None and not input_already_blocked:
input_setter(False)
result: MessageStageResult | None = None
try:
if on_started is not None:
await on_started()
speech = spec.speech.strip()
if spec.require_confirmation and spec.display is None:
result = MessageStageResult(
succeeded=False,
speech=speech,
error="等待用户确认时必须显示客户端消息",
)
return result
if not speech and spec.display is None:
result = MessageStageResult(
succeeded=False,
error="Message 阶段至少需要播报或客户端消息",
)
return result
playback_completion: SpeechCompletion = None
if speech and speak is not None:
playback_completion = await speak(speech)
action: str | None = None
if spec.display is not None:
result = await self._show_message(spec, speech=speech)
if not result.succeeded:
return result
action = result.action
# Confirmation is the gate. It deliberately does not wait for the
# audio completion future, so the user can continue immediately.
if (
playback_completion is not None
and not spec.require_confirmation
):
await playback_completion
result = MessageStageResult(
succeeded=True,
speech=speech,
action=action,
)
return result
except asyncio.CancelledError:
raise
except Exception as exc: # noqa: BLE001 - surface deterministic stage failure
result = MessageStageResult(
succeeded=False,
speech=spec.speech.strip(),
error=str(exc),
)
return result
finally:
should_release = (
result is not None
and (
(result.succeeded and release_input_on_success)
or (not result.succeeded and release_input_on_failure)
)
)
if input_setter is not None and should_release:
input_setter(True)
async def _show_message(
self,
spec: MessageStageSpec,
*,
speech: str,
) -> MessageStageResult:
display = spec.display
if display is None:
return MessageStageResult(succeeded=True, speech=speech)
if self._client_tools is None:
return MessageStageResult(
succeeded=False,
speech=speech,
error="当前运行模式不支持客户端消息",
)
try:
response = await self._client_tools.call(
BUILTIN_SHOW_MESSAGE,
{
"title": display.title,
"message": display.message,
"actions": [
{
"id": "confirmed",
"label": display.confirm_label,
"style": "primary",
}
],
"dismissible": not spec.require_confirmation,
},
timeout_seconds=3,
wait_for_response=spec.require_confirmation,
response_wait_mode=(
"session" if spec.require_confirmation else "timeout"
),
)
except ClientToolError as exc:
return MessageStageResult(
succeeded=False,
speech=speech,
error=str(exc),
)
if response.get("status") != "ok":
return MessageStageResult(
succeeded=False,
speech=speech,
error=str(response.get("message") or "客户端消息显示失败"),
)
data = response.get("data")
action = (
str(data.get("action") or "") or None
if isinstance(data, dict)
else None
)
return MessageStageResult(
succeeded=True,
speech=speech,
action=action,
)

86
backend/services/node_specs.py
View File

@@ -8,12 +8,12 @@ from typing import Any
SPEC_VERSION = "3"
NODE_TYPES = {"start", "agent", "action", "handoff", "end"}
NODE_TYPES = {"start", "agent", "message", "action", "handoff", "end"}
EDGE_MODES = {"llm", "expression", "always"}
AGENT_ENTRY_MODES = {"wait_user", "generate", "fixed_speech"}
ACTION_RESULT_ASSIGNMENT_MODES = {"inherit", "override", "none"}
ACTION_USER_INPUT_POLICIES = {"queue", "block"}
AUTOMATIC_NODE_TYPES = {"start", "action", "handoff"}
AUTOMATIC_NODE_TYPES = {"start", "message", "action", "handoff"}
EXPRESSION_OPERATORS = {
"eq",
"neq",
@@ -31,7 +31,7 @@ NODE_SPECS: list[dict[str, Any]] = [
"name": "start",
"displayName": "Start",
"category": "control_node",
"description": "初始化会话、动态变量和全局观察器，可播放固定开场白。",
"description": "初始化会话、动态变量和全局观察器。",
"icon": "Play",
"accent": "mint",
"addable": False,
@@ -44,7 +44,6 @@ NODE_SPECS: list[dict[str, Any]] = [
},
"fields": [
{"key": "name", "label": "节点名称", "type": "text", "default": "Start"},
{"key": "greeting", "label": "固定开场白", "type": "textarea", "default": ""},
],
},
{
@@ -67,6 +66,19 @@ NODE_SPECS: list[dict[str, Any]] = [
},
],
},
{
"name": "message",
"displayName": "Message",
"category": "interaction_node",
"description": "固定播报，并可同时显示内置客户端消息、等待用户确认。",
"icon": "MessageSquareText",
"accent": "lavender",
"addable": True,
"constraints": {"minIncoming": 1, "minOutgoing": 0},
"fields": [
{"key": "name", "label": "节点名称", "type": "text", "default": "Message"},
],
},
{
"name": "action",
"displayName": "Action",
@@ -172,6 +184,17 @@ def _normalize_action_data(data: dict[str, Any]) -> None:
)
data.setdefault("resultAssignments", {})
data.setdefault("userInputPolicy", "queue")
data.pop("speech", None)
def _normalize_message_data(data: dict[str, Any]) -> None:
"""Fill the small built-in Message contract used by runtime and editor."""
data.setdefault("speech", "")
data.setdefault("showMessage", False)
data.setdefault("title", "重要提示")
data.setdefault("message", "")
data.setdefault("confirmLabel", "确认")
data.setdefault("requireConfirmation", False)
def _normalize_settings(settings: dict[str, Any], *, global_prompt: str = "") -> None:
@@ -200,8 +223,12 @@ def normalize_graph(graph: dict[str, Any] | None) -> dict[str, Any]:
source.setdefault("edges", [])
for node in source["nodes"]:
data = node.setdefault("data", {})
if node.get("type") == "agent":
if node.get("type") == "start":
data.pop("greeting", None)
elif node.get("type") == "agent":
_normalize_agent_data(data)
elif node.get("type") == "message":
_normalize_message_data(data)
elif node.get("type") == "action":
_normalize_action_data(data)
return source
@@ -218,6 +245,7 @@ def normalize_graph(graph: dict[str, Any] | None) -> dict[str, Any]:
"endCall": "end",
"start": "start",
"agent": "agent",
"message": "message",
"action": "action",
"handoff": "handoff",
"end": "end",
@@ -236,10 +264,13 @@ def normalize_graph(graph: dict[str, Any] | None) -> dict[str, Any]:
data.setdefault("scope", "session")
elif new_type == "agent":
_normalize_agent_data(data)
elif new_type == "message":
_normalize_message_data(data)
elif new_type == "action":
_normalize_action_data(data)
elif new_type == "start":
prompt = str(data.pop("prompt", "") or "").strip()
data.pop("greeting", None)
if prompt:
start_prompt_nodes[str(node.get("id"))] = prompt
for key in ("allowInterrupt", "addGlobalPrompt"):
@@ -349,6 +380,49 @@ def validate_graph(graph: dict[str, Any]) -> list[str]:
data.get("entrySpeech") or ""
).strip():
errors.append(f"Agent 节点 {node_id} 的固定进入语不能为空")
elif node_type == "message":
data = node.get("data") or {}
speech = data.get("speech")
show_message = data.get("showMessage")
require_confirmation = data.get("requireConfirmation")
if not isinstance(speech, str):
errors.append(f"Message 节点 {node_id} 的播报内容必须是文本")
if not isinstance(show_message, bool):
errors.append(f"Message 节点 {node_id} 的弹窗开关必须是布尔值")
if not isinstance(require_confirmation, bool):
errors.append(f"Message 节点 {node_id} 的确认开关必须是布尔值")
if require_confirmation and show_message is not True:
errors.append(f"Message 节点 {node_id} 等待确认时必须显示弹窗")
if not str(speech or "").strip() and show_message is not True:
errors.append(f"Message 节点 {node_id} 至少需要播报或显示弹窗")
if show_message is True:
title = data.get("title")
message = data.get("message")
confirm_label = data.get("confirmLabel")
if (
not isinstance(title, str)
or not title.strip()
or len(title) > 120
):
errors.append(
f"Message 节点 {node_id} 的弹窗标题必须为 1-120 个字符"
)
if (
not isinstance(message, str)
or not message.strip()
or len(message) > 2000
):
errors.append(
f"Message 节点 {node_id} 的弹窗消息必须为 1-2000 个字符"
)
if (
not isinstance(confirm_label, str)
or not confirm_label.strip()
or len(confirm_label) > 40
):
errors.append(
f"Message 节点 {node_id} 的按钮文字必须为 1-40 个字符"
)
elif node_type == "action":
data = node.get("data") or {}
assignment_mode = data.get("resultAssignmentMode")
@@ -477,7 +551,7 @@ def validate_graph(graph: dict[str, Any]) -> list[str]:
if node.get("type") != "agent"
)
if any(visit(node_id) for node_id in automatic_node_ids):
errors.append("Start/Action/Handoff/End 之间不能形成无等待循环")
errors.append("自动节点之间不能形成无等待循环")
return list(dict.fromkeys(errors))

13
backend/services/pipecat/call_lifecycle.py
View File

@@ -2,6 +2,8 @@
from __future__ import annotations
import asyncio
from collections import deque
from collections.abc import Awaitable, Callable
from loguru import logger
@@ -19,6 +21,7 @@ class CallEndCoordinator:
self._speaking = False
self._response_speech_started = False
self._tracked_speeches = 0
self._tracked_speech_completions: deque[asyncio.Future[None]] = deque()
self._finish_after_tracked_speech = False
self._finished = False
self._reason = "completed"
@@ -39,9 +42,12 @@ class CallEndCoordinator:
"""Wait for the next observed bot speech to finish."""
self._armed = True
def track_speech(self) -> None:
"""Register one fixed utterance before its TTSSpeakFrame is queued."""
def track_speech(self) -> Awaitable[None]:
"""Register fixed speech and return its transport completion signal."""
completion = asyncio.get_running_loop().create_future()
self._tracked_speech_completions.append(completion)
self._tracked_speeches += 1
return completion
async def arm_after_tracked_speech(self) -> None:
"""Finish after every already queued fixed utterance has played."""
@@ -73,6 +79,9 @@ class CallEndCoordinator:
self._speaking = False
if self._tracked_speeches > 0:
self._tracked_speeches -= 1
completion = self._tracked_speech_completions.popleft()
if not completion.done():
completion.set_result(None)
if (
self._finish_after_tracked_speech
and self._tracked_speeches == 0

9
backend/services/workflow/agent.py
View File

@@ -2,8 +2,6 @@
from __future__ import annotations
from copy import deepcopy
from models import AssistantConfig
from pipecat.flows import ContextStrategy, ContextStrategyConfig, NodeConfig
from pipecat.frames.frames import LLMUpdateSettingsFrame
@@ -108,7 +106,6 @@ class WorkflowAgentStage:
node_id: str,
*,
functions: list,
greeting_context_message: dict[str, str] | None,
leading_messages: list[dict[str, str]] | None = None,
) -> NodeConfig:
data = self._engine.data(node_id)
@@ -119,11 +116,6 @@ class WorkflowAgentStage:
if data.get("contextPolicy") == "fresh"
else ContextStrategy.APPEND
)
greeting_messages = (
[deepcopy(greeting_context_message)]
if strategy == ContextStrategy.RESET and greeting_context_message
else []
)
fixed_reply_messages = (
[{"role": "assistant", "content": entry_speech}]
if entry_mode == "fixed_speech" and entry_speech
@@ -133,7 +125,6 @@ class WorkflowAgentStage:
"name": node_id,
"role_message": self.role_message(node_id),
"task_messages": [
*greeting_messages,
*(leading_messages or []),
*fixed_reply_messages,
],

1
backend/services/workflow/models.py
View File

@@ -15,6 +15,7 @@ class WorkflowStatus(StrEnum):
ROUTING = "routing"
RUNNING_AGENT = "running_agent"
RUNNING_ACTION = "running_action"
RUNNING_MESSAGE = "running_message"
HANDOFF = "handoff"
ENDED = "ended"

56
backend/services/workflow/output.py
View File

@@ -5,65 +5,15 @@ from __future__ import annotations
from typing import Any
from uuid import uuid4
from pipecat.frames.frames import OutputTransportMessageUrgentFrame, TTSSpeakFrame
from pipecat.frames.frames import OutputTransportMessageUrgentFrame
from pipecat.utils.time import time_now_iso8601
from services.brains.base import BrainRuntime
from services.runtime_variables import DynamicVariableStore
from services.fixed_speech import FixedSpeechOutput
class WorkflowOutput:
class WorkflowOutput(FixedSpeechOutput):
"""Publish debug events and fixed speech without duplicating persistence."""
def __init__(
self,
store: DynamicVariableStore,
runtime: BrainRuntime,
) -> None:
self._store = store
self._runtime = runtime
self._client_ready = False
self._pending_transcripts: list[dict[str, Any]] = []
async def mark_client_ready(self) -> None:
self._client_ready = True
pending = self._pending_transcripts
self._pending_transcripts = []
for message in pending:
await self.emit(message)
async def speak(
self,
text: str,
*,
source: str,
node_id: str | None = None,
) -> None:
"""Record, display and synthesize one Workflow-owned utterance."""
content = text.strip()
if not content:
return
self._store.record("agent", content)
transcript = {
"type": "transcript",
"role": "assistant",
"content": content,
"timestamp": time_now_iso8601(),
"source": source,
**({"nodeId": node_id} if node_id else {}),
}
if self._client_ready:
await self.emit(transcript)
else:
self._pending_transcripts.append(transcript)
track_speech = getattr(self._runtime.call_end, "track_speech", None)
if callable(track_speech):
track_speech()
await self._runtime.queue_frame(
TTSSpeakFrame(content, append_to_context=False)
)
async def emit_node_active(self, node_id: str | None) -> None:
if node_id:
await self.emit({"type": "node-active", "nodeId": node_id})

10
backend/services/workflow_engine.py
View File

@@ -190,20 +190,12 @@ class WorkflowEngine:
sections.append(f"[当前阶段任务]\n{prompt}")
return "\n\n".join(sections)
def greeting(self, store: DynamicVariableStore) -> str:
return store.render(str(self.data(self.start_id).get("greeting") or ""))
def routing_prompt(self, node_id: str, store: DynamicVariableStore) -> str:
"""Describe the current node to the small LLM edge router."""
if self.node_type(node_id) == "agent":
return self.prompt_for(node_id, store)
data = self.data(node_id)
details = (
data.get("greeting")
or data.get("message")
or data.get("target")
or ""
)
details = data.get("message") or data.get("target") or ""
rendered = store.render(str(details)).strip()
return f"{self.node_type(node_id) or 'workflow'} 节点：{rendered or self.name(node_id)}"