dv-pipecat-flows 0.0.26.dev142__tar.gz → 0.0.26.dev156__tar.gz

{dv_pipecat_flows-0.0.26.dev142 → dv_pipecat_flows-0.0.26.dev156}/.gitignore

@@ -187,3 +187,8 @@ MANIFEST
 
 # Asterisk configs
 remote-asterisk-code/pjsip.conf
+
+
+# local files
+scratch/
+tmp/

{dv_pipecat_flows-0.0.26.dev142/src/dv_pipecat_flows.egg-info → dv_pipecat_flows-0.0.26.dev156}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dv-pipecat-flows
-Version: 0.0.26.dev142
+Version: 0.0.26.dev156
 Summary: Conversation Flow management for Pipecat AI applications
 License: BSD 2-Clause License
 Project-URL: Source, https://github.com/pipecat-ai/pipecat-flows

{dv_pipecat_flows-0.0.26.dev142 → dv_pipecat_flows-0.0.26.dev156/src/dv_pipecat_flows.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dv-pipecat-flows
-Version: 0.0.26.dev142
+Version: 0.0.26.dev156
 Summary: Conversation Flow management for Pipecat AI applications
 License: BSD 2-Clause License
 Project-URL: Source, https://github.com/pipecat-ai/pipecat-flows

{dv_pipecat_flows-0.0.26.dev142 → dv_pipecat_flows-0.0.26.dev156}/src/dv_pipecat_flows.egg-info/scm_file_list.json

@@ -385,6 +385,7 @@
     "ringg-bot/tests/conftest.py",
     "ringg-bot/tests/test_abandoned_call_report.py",
     "ringg-bot/tests/test_action_call_variables.py",
+    "ringg-bot/tests/test_anthropic_llm.py",
     "ringg-bot/tests/test_app_metrics.py",
     "ringg-bot/tests/test_arabic_amount.py",
     "ringg-bot/tests/test_artifact_store.py",
@@ -506,6 +507,7 @@
     "ringg-bot/tests/test_regional_vendor_endpoints.py",
     "ringg-bot/tests/test_regression_monitor_sync.py",
     "ringg-bot/tests/test_replay_kb_flag.py",
+    "ringg-bot/tests/test_replay_mock_websocket.py",
     "ringg-bot/tests/test_ringg_stt_mapping.py",
     "ringg-bot/tests/test_router_model_client.py",
     "ringg-bot/tests/test_router_shadow.py",
@@ -529,6 +531,7 @@
     "ringg-bot/tests/test_simulate_agent_config.py",
     "ringg-bot/tests/test_simulate_batch.py",
     "ringg-bot/tests/test_simulate_progress.py",
+    "ringg-bot/tests/test_simulation_action_widget_guard.py",
     "ringg-bot/tests/test_single_node_kb_id_resolution.py",
     "ringg-bot/tests/test_stall_watchdog.py",
     "ringg-bot/tests/test_status_router.py",
@@ -571,6 +574,7 @@
     "ringg-bot/transports/factory.py",
     "ringg-bot/utils/action_call_handler.py",
     "ringg-bot/utils/analyzer_pool.py",
+    "ringg-bot/utils/anthropic_llm.py",
     "ringg-bot/utils/api_calls.py",
     "ringg-bot/utils/app_metrics.py",
     "ringg-bot/utils/arabic_amount.py",

dv_pipecat_flows-0.0.26.dev156/src/dv_pipecat_flows.egg-info/scm_version.json

@@ -0,0 +1,8 @@
+{
+  "tag": "0.0.25",
+  "distance": 156,
+  "node": "g7ace808e",
+  "dirty": false,
+  "branch": "dv-stage",
+  "node_date": "2026-10-09"
+}

{dv_pipecat_flows-0.0.26.dev142 → dv_pipecat_flows-0.0.26.dev156}/src/pipecat_flows/manager.py

@@ -53,6 +53,9 @@ from utils.collect_prompts import collect_prompt
 from pipecat.frames.frames import (
     DTMFUpdateSettingsFrame,
     FunctionCallResultProperties,
+    InterruptionFrame,
+    LLMFullResponseEndFrame,
+    LLMFullResponseStartFrame,
     LLMMessagesAppendFrame,
     LLMMessagesUpdateFrame,
     LLMRunFrame,
@@ -177,6 +180,50 @@ def _paired_transition_call_ids(
     return call_ids & result_ids
 
 
+# Appended after the carried conversation when the node prompt is folded into the
+# system prompt, so the context still ends on a customer-side turn. Without it Claude
+# sees the bot's own last message as the end of the conversation and replies with
+# nothing after a node change (measured on the PB chatbot replay, 9 Oct 2026).
+CONTINUE_NUDGE = (
+    "Continue the conversation: act on the customer's latest message using the "
+    "instructions in the system prompt above. Do not reply to this note or mention it."
+)
+
+
+def _system_text(content: Any) -> str:
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        return "\n".join(
+            part.get("text", "") for part in content if isinstance(part, dict) and part.get("text")
+        )
+    return ""
+
+
+def fold_system_messages(messages: List[dict]) -> List[dict]:
+    """Put every system message into one leading system message for Claude.
+
+    The conversation keeps its order after it, and a system-role nudge closes the
+    context: the adapters turn that into a user message (merged into the customer's
+    turn when it is last), so the request ends on the customer's side. The nudge is
+    a system message so that the next RESET_WITH_CONVERSATION drops it instead of
+    carrying it as something the customer said. A context with no conversation is
+    left as one system message, as before.
+    """
+    system_parts = [
+        text
+        for m in messages
+        if m.get("role") == "system" and (text := _system_text(m.get("content")))
+    ]
+    if not system_parts:
+        return messages
+    rest = [m for m in messages if m.get("role") != "system"]
+    folded = [{"role": "system", "content": "\n\n".join(system_parts)}, *rest]
+    if rest:
+        folded.append({"role": "system", "content": CONTINUE_NUDGE})
+    return folded
+
+
 def extract_conversation_messages(
     messages: List[dict], transition_names: Optional[Set[str]] = None
 ) -> List[dict]:
@@ -265,6 +312,11 @@ def extract_conversation_messages(
 # drained the terminal frame. Coarse on purpose — nothing waits on it.
 _CALL_END_POLL_INTERVAL_S = 5.0
 
+# How long a RESET_WITH_CONVERSATION carry waits for the reply in flight to be
+# saved (see FlowManager._wait_for_reply_saved). In a chat session the end frame
+# lands within milliseconds of the transition.
+_REPLY_SAVE_TIMEOUT_S = 1.5
+
 # How long to let an utterance that straddled the wait ceiling finish
 # committing. The caller started speaking just before the ceiling fired, so STT
 # has a final in flight; deciding anything before it lands means classifying
@@ -453,6 +505,29 @@ class FlowManager:
         # Signature:  is_injected(message) -> bool
         self.injected_message_filter: Optional[Callable[[Any], bool]] = None
 
+        # Chat sessions: before RESET_WITH_CONVERSATION carries the conversation into
+        # the next node, wait for the assistant aggregator to save the reply in flight,
+        # so text the LLM wrote alongside the transition is carried too (see
+        # `_wait_for_reply_saved`). Off by default: in a voice call the reply is saved
+        # only once it has been spoken, so waiting would hold the hop for the speech.
+        self.carry_waits_for_reply: bool = False
+        # Claude (Anthropic API / Bedrock): both take one system prompt at the top, and
+        # their adapters turn every later system message into a user message merged
+        # with the customer's last turn. A node's task prompt placed after the carried
+        # conversation then reads to Claude as something the customer said, and Claude
+        # answers the instructions instead of the customer. When set, a context replace
+        # folds every system message into one leading system message and ends on a
+        # short nudge (see `fold_system_messages`). Off by default: GPT reads a trailing
+        # system message as an instruction, which is why the task prompt sits there.
+        self.fold_task_into_system: bool = False
+        # Replies the assistant aggregator has started but not yet saved, counted from
+        # its own frame events (`_track_assistant_reply`); `_reply_saved` is set while
+        # there are none.
+        self._replies_in_flight = 0
+        self._reply_saved = asyncio.Event()
+        self._reply_saved.set()
+        self._watch_assistant_replies()
+
         # Wired in `initialize()` once the pipeline is up. Tracks the timestamp
         # of the latest UserStoppedSpeakingFrame so router-mode nodes can gate
         # their LLM classification on "user has actually replied" instead of
@@ -5078,6 +5153,54 @@ In all of these cases, you can provide a `name` in your new node's config for de
             resolved.append(message)
         return resolved
 
+    def _watch_assistant_replies(self) -> None:
+        """Follow the assistant aggregator's replies through its on_after_process_frame event."""
+        assistant = self._context_aggregator.assistant() if self._context_aggregator else None
+        add_event_handler = getattr(assistant, "add_event_handler", None)
+        if add_event_handler is None:
+            return
+        try:
+            add_event_handler("on_after_process_frame", self._track_assistant_reply)
+        except Exception as exc:  # never let reply tracking break flow setup
+            logger.debug(f"Not tracking assistant replies: {exc}")
+
+    async def _track_assistant_reply(self, _aggregator: Any, frame: Any) -> None:
+        """Count replies in flight. The aggregator saves a reply's text to the context
+        while processing its LLMFullResponseEndFrame, so after that frame it is saved."""
+        if isinstance(frame, LLMFullResponseStartFrame):
+            self._replies_in_flight += 1
+            self._reply_saved.clear()
+        elif isinstance(frame, (LLMFullResponseEndFrame, InterruptionFrame)):
+            if isinstance(frame, InterruptionFrame):
+                self._replies_in_flight = 0
+            else:
+                self._replies_in_flight = max(0, self._replies_in_flight - 1)
+            if not self._replies_in_flight:
+                self._reply_saved.set()
+
+    async def _wait_for_reply_saved(self) -> None:
+        """Wait, briefly, until the assistant aggregator has saved the reply in flight.
+
+        A transition fires on its function-call result, a SystemFrame that overtakes
+        the reply's queued text and its LLMFullResponseEndFrame. The aggregator adds
+        that text to the context only at the end frame, so a carry taken before it
+        leaves out what the LLM wrote alongside the transition ("Plans that work well
+        for NRIs -") and the LLMMessagesUpdateFrame then erases it from the live
+        context: the customer saw it, the LLM never does, and the next node re-routes
+        on a turn that was already answered (Policybazaar chat, Oct 2026).
+
+        Capped, so a lost end frame only delays the hop.
+        """
+        if self._reply_saved.is_set():
+            return
+        try:
+            await asyncio.wait_for(self._reply_saved.wait(), timeout=_REPLY_SAVE_TIMEOUT_S)
+        except asyncio.TimeoutError:
+            logger.warning(
+                f"Reply still in flight after {_REPLY_SAVE_TIMEOUT_S}s; carrying the "
+                "conversation without it"
+            )
+
     async def _update_llm_context(
         self,
         role_messages: Optional[List[dict]],
@@ -5156,6 +5279,8 @@ In all of these cases, you can provide a `name` in your new node's config for de
                 # prompt and tool-call/tool-result logs are dropped. Placed after
                 # role_messages (the new system instruction) and before task
                 # messages, mirroring the RESET_WITH_SUMMARY layout.
+                if self.carry_waits_for_reply:
+                    await self._wait_for_reply_saved()
                 carried = extract_conversation_messages(
                     self._without_injected(
                         self._context_aggregator.user()._context.get_messages()
@@ -5191,6 +5316,9 @@ In all of these cases, you can provide a `name` in your new node's config for de
                 except Exception as exc:  # never let a provider break a transition
                     logger.warning(f"sticky_context_provider failed: {exc}")
 
+            if is_replace and self.fold_task_into_system:
+                messages = fold_system_messages(messages)
+
             frame_type = LLMMessagesUpdateFrame if is_replace else LLMMessagesAppendFrame
 
             await self._task.queue_frames(

dv_pipecat_flows-0.0.26.dev142/src/dv_pipecat_flows.egg-info/scm_version.json

@@ -1,8 +0,0 @@
-{
-  "tag": "0.0.25",
-  "distance": 142,
-  "node": "g5494275d",
-  "dirty": false,
-  "branch": "dv-stage",
-  "node_date": "2026-10-07"
-}