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 search_conversation find this turn's results on API providers (#16666)

Daniel Hand80856cf9af8
diff --git a/studio/backend/core/inference/studio_tool_loop.py b/studio/backend/core/inference/studio_tool_loop.py
index 7ecc09e2df0..07c0e180bda 100644
--- a/studio/backend/core/inference/studio_tool_loop.py
+++ b/studio/backend/core/inference/studio_tool_loop.py
@@ -1465,8 +1465,18 @@ async def _stream_with_studio_tools(
resumes_partial = run.continue_final_message
# cap run-owned image parts across the full loop without counting caller attachments.
loop_mcp_image_parts: list = list(run.promoted_image_parts)
- # preserve the request branch for tools that search conversation history.
- request_branch = list(restored)
+ # search_conversation filters on this: the request plus every message the loop added, so a
+ # result of this turn that a refit evicted stays findable (llama_cpp's _extend_live_branch).
+ live_branch = list(restored)
+ live_branch_ids = {id(message) for message in conversation}
+
+ def extend_live_branch() -> list[dict]:
+ for message in conversation:
+ if id(message) not in live_branch_ids:
+ live_branch_ids.add(id(message))
+ live_branch.append(message)
+ return list(live_branch)
+
remaining = policy.max_calls
unlimited = remaining >= 9999
session_id = run.session_id
@@ -2104,7 +2114,7 @@ async def _stream_with_studio_tools(
# once a thread has an archive and needs the same branch: the stored rows are the whole DAG, and Retry
# leaves the replaced response in them.
if accepts_kwarg(execute_tool, "conversation_branch"):
- kwargs["conversation_branch"] = request_branch
+ kwargs["conversation_branch"] = extend_live_branch()
if approved and accepts_kwarg(execute_tool, "host_access_approved"):
kwargs["host_access_approved"] = True
# External window is unknowable: 0 keeps the default page cap; unmeasurable budget means one recall, since 8 chunks is ~4K tokens replayed each later call.
diff --git a/studio/backend/tests/test_external_tool_edge_cases.py b/studio/backend/tests/test_external_tool_edge_cases.py
index 9113bc2a3a1..81f93d9be90 100644
--- a/studio/backend/tests/test_external_tool_edge_cases.py
+++ b/studio/backend/tests/test_external_tool_edge_cases.py
@@ -285,6 +285,23 @@ def test_intermediate_done_sentinel_is_not_relayed(executed):
assert "the real answer" in _visible_text(lines)
+def test_search_conversation_branch_holds_this_turns_tool_results(executed):
+ """A refit can evict this turn's results; the branch must still hold them or the archive filters them out (#16602)."""
+ transport = FakeTransport(
+ [
+ _call_turn("c1", "web_search", '{"query":"valve"}'),
+ _call_turn("c2", "search_conversation", '{"query":"valve"}'),
+ _answer_turn(),
+ ]
+ )
+ _run(transport, tools = [WEB, _tool("search_conversation")])
+
+ [search] = [call for call in executed if call["name"] == "search_conversation"]
+ branch = search["conversation_branch"]
+ assert branch[0] == {"role": "user", "content": "hi"}
+ assert any(m.get("role") == "tool" and m.get("content") == "RESULT<web_search>" for m in branch)
+
+
def test_non_json_data_line_is_relayed_untouched(executed):
"""Garbage on the wire is passed through, not parsed and not fatal."""
transport = FakeTransport(