Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: let Deep Research finish a turn handed off from a chat generation (#11923)

Mohammad Hijjawid7eba2e2ec0e
diff --git a/studio/backend/storage/studio_db.py b/studio/backend/storage/studio_db.py
index 120298c69a1..474bfe5202b 100644
--- a/studio/backend/storage/studio_db.py
+++ b/studio/backend/storage/studio_db.py
@@ -3034,6 +3034,17 @@ def _research_message_ids(conn: sqlite3.Connection, thread_id: str) -> set[str]:
}
+def _research_assistant_message_ids(conn: sqlite3.Connection, thread_id: str) -> set[str]:
+ return {
+ str(row[0])
+ for row in conn.execute(
+ "SELECT assistant_message_id FROM research_runs "
+ "WHERE thread_id = ? AND assistant_message_id IS NOT NULL",
+ (thread_id,),
+ ).fetchall()
+ }
+
+
def _generation_message_ids(conn: sqlite3.Connection, thread_id: str) -> set[str]:
return {
str(message_id)
@@ -3275,9 +3286,17 @@ def _guard_server_managed_messages(
allow_research_update: bool = False,
) -> None:
generation = _generation_message_ids(conn, thread_id)
- protected = set(generation)
- if not allow_research_update:
- protected.update(_research_message_ids(conn, thread_id))
+ if allow_research_update:
+ # A deep research run is handed off from a chat generation and reports into that
+ # generation's assistant message. Once the generation has settled, the research run
+ # is the message's only writer, so its authorized updates must not be held to the
+ # generation's monotonic-update rules.
+ generation -= _research_assistant_message_ids(
+ conn, thread_id
+ ) & _terminal_generation_message_ids(conn, thread_id)
+ protected = set(generation)
+ else:
+ protected = generation | _research_message_ids(conn, thread_id)
if not protected:
return
for message in messages:
@@ -3293,6 +3312,33 @@ def _guard_server_managed_messages(
raise ChatMessageProtectedError("server-managed generation messages cannot be edited")
+def _settle_handed_off_generation(conn: sqlite3.Connection, message: dict) -> dict:
+ # The live tab hands off before settling, so an unsettled row would be replayed by generation
+ # recovery on the next load and its settle write would replace the research report.
+ metadata = message.get("metadata")
+ if not isinstance(metadata, dict):
+ return message
+ row = conn.execute(
+ """SELECT id, status, last_event_seq FROM chat_generation_runs
+ WHERE thread_id = ? AND assistant_message_id = ?
+ AND status IN ('cancelled', 'completed', 'failed')""",
+ (message["threadId"], str(message["id"])),
+ ).fetchone()
+ if row is None or metadata.get("generationRunId") != row["id"]:
+ return message
+ # The research status now reports the outcome, not the acknowledgement's length/interrupt mark.
+ metadata = {key: value for key, value in metadata.items() if key != "incomplete"}
+ return {
+ **message,
+ "metadata": {
+ **metadata,
+ "generationStatus": row["status"],
+ "generationSeq": int(row["last_event_seq"]),
+ "generationSettled": True,
+ },
+ }
+
+
def _detach_terminal_generation_for_edit(
conn: sqlite3.Connection, thread_id: str, message: dict
) -> bool:
@@ -3639,6 +3685,8 @@ def upsert_chat_message(
[message],
allow_research_update = allow_research_update,
)
+ if allow_research_update:
+ message = _settle_handed_off_generation(conn, message)
_raise_if_chat_message_thread_conflicts(
conn,
message["threadId"],
diff --git a/studio/backend/tests/test_research_runs_storage.py b/studio/backend/tests/test_research_runs_storage.py
index 09d7b0a26fa..614b6a556ba 100644
--- a/studio/backend/tests/test_research_runs_storage.py
+++ b/studio/backend/tests/test_research_runs_storage.py
@@ -3067,6 +3067,123 @@ def test_create_run_rejects_a_completed_answer_even_beside_the_handoff(research_
assert research_db.get_run("run-1") is None
+def _hand_off_from_generation(generation_status, settled = None):
+ # The chat generation that called the deep_research tool owns the assistant message first.
+ conn = studio_db.get_connection()
+ try:
+ conn.execute(
+ """INSERT INTO chat_generation_runs
+ (id, owner_subject, thread_id, user_message_id, assistant_message_id,
+ request_hash, request_json, worker_token, status, last_event_seq,
+ created_at, updated_at)
+ VALUES ('gen-1', 'alice', 'thread-1', 'user-1', 'assistant-1',
+ 'hash', '{}', 'token', ?, 3, 4, 4)""",
+ (generation_status,),
+ )
+ conn.execute(
+ "UPDATE chat_messages SET metadata_json = ? WHERE id = 'assistant-1'",
+ (
+ json.dumps(
+ {
+ "serverManaged": True,
+ "generationRunId": "gen-1",
+ "generationStatus": generation_status,
+ "generationSeq": 3 if settled is not False else 2,
+ "generationSettled": (
+ generation_status == "completed" if settled is None else settled
+ ),
+ }
+ ),
+ ),
+ )
+ conn.commit()
+ finally:
+ conn.close()
+ return _create()
+
+
+def test_update_assistant_writes_the_report_after_a_settled_generation_handoff(research_home):
+ from core.research_runs import _update_assistant
+
+ run = _hand_off_from_generation("completed")
+
+ _update_assistant(run, "final report", "completed")
+
+ message = studio_db.get_chat_message("thread-1", "assistant-1")
+ assert [part["text"] for part in message["content"] if part["type"] == "text"] == [
+ "final report"
+ ]
+ assert message["metadata"]["researchRunId"] == "run-1"
+ assert message["metadata"]["researchStatus"] == "completed"
+ # Only the research run is exempted: a plain client edit is still refused.
+ with pytest.raises(studio_db.ChatMessageProtectedError):
+ studio_db.upsert_chat_message({**message, "content": [{"type": "text", "text": "edit"}]})
+
+
+def test_research_report_survives_recovery_of_an_unsettled_handoff(research_home):
+ from core.research_runs import _update_assistant
+
+ # The live tab starts research straight after the stream, leaving its checkpoint unsettled.
+ run = _hand_off_from_generation("completed", settled = False)
+
+ _update_assistant(run, "final report", "completed")
+
+ message = studio_db.get_chat_message("thread-1", "assistant-1")
+ # A recovery follower replays the generation tail and settles at the run's last event.
+ recovered = {
+ **message,
+ "content": [{"type": "text", "text": "final report into it."}],
+ "metadata": {
+ **message["metadata"],
+ "generationSeq": 3,
+ "generationStatus": "completed",
+ "generationSettled": True,
+ },
+ }
+ with pytest.raises(studio_db.ChatMessageProtectedError):
+ studio_db.upsert_chat_message(recovered)
+ stored = studio_db.get_chat_message("thread-1", "assistant-1")
+ assert [part["text"] for part in stored["content"] if part["type"] == "text"] == [
+ "final report"
+ ]
+ assert stored["metadata"]["generationSettled"] is True
+ assert stored["metadata"]["researchStatus"] == "completed"
+
+
+def test_research_report_drops_the_acknowledgement_incomplete_mark(research_home):
+ from core.research_runs import _update_assistant
+
+ run = _hand_off_from_generation("completed", settled = False)
+ # The acknowledgement after the tool call hit Max Tokens, and finish_run stamped the message.
+ conn = studio_db.get_connection()
+ try:
+ row = conn.execute(
+ "SELECT metadata_json FROM chat_messages WHERE id = 'assistant-1'"
+ ).fetchone()
+ metadata = {**json.loads(row[0]), "incomplete": {"reason": "length"}}
+ conn.execute(
+ "UPDATE chat_messages SET metadata_json = ? WHERE id = 'assistant-1'",
+ (json.dumps(metadata),),
+ )
+ conn.commit()
+ finally:
+ conn.close()
+
+ _update_assistant(run, "final report", "completed")
+
+ message = studio_db.get_chat_message("thread-1", "assistant-1")
+ assert "incomplete" not in message["metadata"]
+ assert message["metadata"]["researchStatus"] == "completed"
+
+
+def test_update_assistant_still_waits_for_an_active_generation(research_home):
+ from core.research_runs import _update_assistant
+ run = _hand_off_from_generation("running")
+
+ with pytest.raises(studio_db.ChatMessageProtectedError):
+ _update_assistant(run, "final report", "completed")
+
+
def test_update_assistant_replaces_report_parts_without_duplication(research_home):
from core.research_runs import _update_assistant