feat: add vici dial controls from UI (#563)

* ViciDial Working

* feat(telephony): add FreeSWITCH provider to upstream-PBX seam

Generalize the upstream-PBX capture and control paths to dispatch by
provider. FreeSWITCH bridges in with X-PBX-* headers and is driven over
the Event Socket Library (uuid_kill / uuid_transfer) by the channel UUID,
alongside the existing VICIdial ra_call_control path.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* add vici specific configs

* feat(telephony): env-based VICIdial config + address3 post-call routing

Move VICIdial agent/non-agent API and FreeSWITCH ESL connection settings
out of hardcoded POC values into environment variables so the same image
works against a PBX on another server.

Add VICIdial update_lead forwarding: X-VICI-UPDATE-LEAD_* extracted
variables are mapped to lead columns and pushed via the non-agent API
before a transfer, with reserved API-control params dropped.

Add hardcoded address3 disposition routing (Y/N -> in-group transfer,
else hang up) and a synchronous final variable extraction so the
transfer path sees the freshest conversation state. Skip upstream_pbx
capture on non-PJSIP channels to avoid Asterisk 500s.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* chore: run formatter

* feat: make vici configurable from UI

* chore: clean up PR

* fix: incorporate review comments

* chore: incorporate review comments

* chore: generate client

* chore: incporporate review comments

* chore: incorporate review comments

---------

Co-authored-by: Dograh POC <payment@dograh.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

api/services/workflow/pipecat_engine.py

@@ -98,6 +98,10 @@ class PipecatEngine:
         self._gathered_context: dict = {}
         self._user_response_timeout_task: Optional[asyncio.Task] = None
         self._pending_extraction_tasks: set[asyncio.Task] = set()
+        # True once a final (synchronous) extraction has run, so the end-of-call
+        # and upstream-transfer paths don't redundantly re-extract the same
+        # terminal state.
+        self._final_extraction_done: bool = False
 
         # Will be set later in initialize() when we have
         # access to _context
@@ -507,6 +511,29 @@ class PipecatEngine:
                 f"Incomplete: {incomplete}"
             )
 
+    async def perform_final_variable_extraction(self) -> None:
+        """Flush in-flight + current-node variable extraction synchronously.
+
+        Awaits any background extractions still running from previous nodes,
+        then runs the current node's extraction inline so callers that need the
+        freshest extracted variables before acting can rely on them -- e.g.
+        end_call_with_reason before disposing the call, or an external-PBX
+        transfer that maps extracted variables into a provider lead update
+        call before handing the customer off.
+
+        Idempotent: only the first call does work. The external-PBX transfer
+        runs this just before forwarding update_lead, so the subsequent
+        end_call_with_reason would otherwise re-extract the same terminal state.
+        """
+        if self._final_extraction_done:
+            logger.debug("Final variable extraction already performed; skipping")
+            return
+        self._final_extraction_done = True
+        await self._await_pending_extractions()
+        await self._perform_variable_extraction_if_needed(
+            self._current_node, run_in_background=False
+        )
+
     async def _setup_llm_context(self, node: Node) -> None:
         """Common method to set up LLM context"""
         # Set OTel span name for tracing
@@ -742,13 +769,8 @@ class PipecatEngine:
             EndTaskReason.PIPELINE_ERROR.value,
             EndTaskReason.VOICEMAIL_DETECTED.value,
         ):
-            # Await any in-flight background extractions from previous nodes
-            await self._await_pending_extractions()
-
-            # Perform final variable extraction synchronously before ending
-            await self._perform_variable_extraction_if_needed(
-                self._current_node, run_in_background=False
-            )
+            # Flush in-flight + current-node extractions synchronously before ending
+            await self.perform_final_variable_extraction()
 
         frame_to_push = (
             CancelFrame(reason=reason) if abort_immediately else EndFrame(reason=reason)
@@ -766,6 +788,20 @@ class PipecatEngine:
                 call_tags.append(call_disposition)
             self._gathered_context["call_tags"] = call_tags
 
+        # Hangup strategies run while serializing the terminal frame. Persist
+        # the final extracted values first so external-PBX adapters can apply
+        # workflow lead-field mappings before terminating the customer leg.
+        try:
+            await db_client.update_workflow_run(
+                run_id=self._workflow_run_id,
+                gathered_context=self._gathered_context,
+            )
+        except Exception as exc:
+            # Call teardown must never be held hostage by an enrichment write.
+            logger.warning(
+                f"Could not persist final gathered context before hangup: {exc}"
+            )
+
         logger.debug(
             f"Finishing run with reason: {reason}, disposition: {call_disposition} "
             f"queueing frame {frame_to_push}"