Fix orchestration follow-ups and selected model routing

application/single_app/config.py

@@ -97,7 +97,7 @@
 EXECUTOR_TYPE = 'thread'
 EXECUTOR_MAX_WORKERS = 30
 SESSION_TYPE = 'filesystem'
-VERSION = "0.261.102"
+VERSION = "0.261.103"
 IS_DEVELOPMENT = is_development_env_enabled()
 
 SESSION_COOKIE_SAMESITE = os.getenv('SESSION_COOKIE_SAMESITE', 'Lax')

application/single_app/functions_orchestration_adapters.py

@@ -44,7 +44,7 @@
 lives in ``route_backend_chats``, importing which at module load would be a circular import --
 so the same lazy pattern is used uniformly rather than only where it is strictly forced.
 
-Version: 0.261.099
+Version: 0.261.102
 """
 
 import json
@@ -1091,17 +1091,14 @@ def _finalize_source_review(
     )
 
 
-def _resolve_source_review_planner(settings):
+def _resolve_source_review_planner(settings, context=None):
     """The optional client for research query and link-selection planning.
 
-    perform_source_review takes a planner client/model so it can decide which discovered links
-    are worth reading. The context's ``invoke_prompt`` closure has already resolved a client,
-    but it is a ``call(prompt) -> text`` seam by design and does not expose the client object,
-    so we resolve one the same way the planner does. ``resolve_planner_client`` handles APIM,
-    managed identity and key auth and returns the planner deployment -- the right model for an
-    internal planning call rather than for writing the final answer. Expected configuration
-    failures leave the existing backup query/link planning available.
+    A running orchestration supplies its already-authorized, protocol-aware client.
+    Standalone callers retain the legacy optional planner and backup planning behavior.
     """
+    if getattr(context, 'planner_client', None) is not None:
+        return context.planner_client, context.planner_deployment
     from functions_orchestration_planner import PlannerError, resolve_planner_client
 
     try:
@@ -1229,7 +1226,7 @@ def run_deep_research(step, context, *, settings, user_id, emit, cancel_requeste
             'Deep research is not enabled or permitted.',
         )
 
-    planner_client, planner_model = _resolve_source_review_planner(settings)
+    planner_client, planner_model = _resolve_source_review_planner(settings, context)
     if _is_cancelled(cancel_requested):
         return _cancelled_result('Cancelled before deep research.')
 

application/single_app/functions_orchestration_context.py

@@ -27,7 +27,7 @@
 A user who picked a document and then watched the planner search their whole workspace
 would rightly conclude the control did nothing.
 
-Version: 0.261.099
+Version: 0.261.102
 """
 
 import hashlib
@@ -180,6 +180,7 @@ def resolve_seeds(request_data):
         'document_filter_mode': filter_mode,
         'agent': agent,
         'model': model or None,
+        'reasoning_effort': _text(request_data.get('reasoning_effort')),
         'prompt': prompt,
         # A user who switched web search on has said something about intent even in
         # orchestration mode, so it is carried through as a constraint rather than dropped.

application/single_app/functions_orchestration_events.py

@@ -30,7 +30,7 @@
     plan card ticks specific steps by id, and reverse-engineering that from prose would be
     guesswork.
 
-Version: 0.261.085
+Version: 0.261.102
 """
 
 import json
@@ -283,6 +283,10 @@ def build_run_done_event(
     plan_summary=None,
     status='completed',
     agent_citations=None,
+    model_deployment_name=None,
+    model_provider=None,
+    model_endpoint_id=None,
+    model_id=None,
 ):
     """Terminal frame of the run endpoint.
 
@@ -307,6 +311,14 @@ def build_run_done_event(
         'generated_artifacts': list(artifacts or ()),
         'orchestration': plan_summary or {},
         'status': status,
+        **{
+            key: value for key, value in {
+                'model_deployment_name': model_deployment_name,
+                'model_provider': model_provider,
+                'model_endpoint_id': model_endpoint_id,
+                'model_id': model_id,
+            }.items() if value is not None
+        },
     })
 
 

application/single_app/functions_orchestration_executor.py

@@ -13,10 +13,10 @@
 Two properties are worth stating because they are the reason this is an engine and not a
 loop:
 
-**A plan always produces an answer.** A gather step can fail, be skipped because its
+**A plan always attempts an answer.** A gather step can fail, be skipped because its
 dependency failed, or be cut off by a budget, and the run still reaches ``respond`` and
 answers with whatever evidence survived. The terminal step is therefore exempt from every
-skip rule; the only thing that stops it is an explicit cancellation.
+skip rule except an explicit cancellation. A failed answer completion still fails the run.
 
 **Access is re-checked at answer time, not trusted from plan time.** Between the planner
 naming a document and the executor answering from it, the user's access to that document can
@@ -31,7 +31,7 @@
 itself. The route owns that loop, because only the route can decide to spend another planner
 round trip.
 
-Version: 0.261.099
+Version: 0.261.102
 """
 
 import logging
@@ -169,6 +169,8 @@ def __init__(
         user_id=None,
         turn_index=0,
         invoke_prompt=None,
+        planner_client=None,
+        planner_deployment=None,
         user_message='',
         user_message_id=None,
         answered_questions=None,
@@ -209,6 +211,8 @@ def __init__(
         self.turn_index = turn_index
 
         self.invoke_prompt = invoke_prompt
+        self.planner_client = planner_client
+        self.planner_deployment = planner_deployment
         self.user_message = user_message
         self.answered_questions = deepcopy(answered_questions or [])
         self.resolved_message = resolved_message if resolved_message is not None else user_message