fix(studio): ignore reasoning in tool reprompts (#7134)

studio/backend/core/inference/inference.py

@@ -888,6 +888,7 @@ class InferenceBackend:
         thread_id: Optional[str] = None,
         rag_scope: Optional[dict] = None,
         presence_penalty: float = 0.0,
+        reasoning_prefilled: bool = False,
     ):
         """Run an agentic tool loop on top of ``generate_chat_response``.
 
@@ -941,6 +942,7 @@ class InferenceBackend:
             session_id = session_id,
             thread_id = thread_id,
             rag_scope = rag_scope,
+            reasoning_prefilled = reasoning_prefilled,
         )
 
     def generate_chat_response(

studio/backend/core/inference/orchestrator.py

@@ -1407,6 +1407,7 @@ class InferenceOrchestrator:
         use_adapter: Optional[Union[bool, str]] = None,
         stats_holder: Optional[dict] = None,
         presence_penalty: float = 0.0,
+        reasoning_prefilled: bool = False,
         **_unused,
     ):
         """Run the safetensors agentic tool loop in the parent process,
@@ -1487,6 +1488,7 @@ class InferenceOrchestrator:
             confirm_tool_calls = confirm_tool_calls,
             bypass_permissions = bypass_permissions,
             permission_mode = permission_mode,
+            reasoning_prefilled = reasoning_prefilled,
         )
 
     def generate_with_adapter_control(

studio/backend/core/inference/safetensors_agentic.py

@@ -50,6 +50,7 @@ from core.inference.tool_call_parser import (
 # pattern lists, so the safetensors streaming strip stays aligned with the parser.
 from core.tool_healing import (
     _REHEARSAL_TAIL_STRIP_RE,
+    _THINK_CLOSE_RE,
     _strip_bracket_tag_calls,
     _think_spans_outside_tool_markup,
     apply_tool_strip_patterns,
@@ -304,6 +305,45 @@ def _status_for_tool(tool_name: str, arguments: dict) -> str:
     return status_for_tool(tool_name, arguments)
 
 
+def _reprompt_intent_text(text: str, *, reasoning_prefilled: bool = False) -> str:
+    """Return visible answer text for the plan-without-action classifier.
+
+    Safetensors reasoning shares the cumulative text channel with the answer.
+    Forward-looking phrases inside ``<think>`` / ``[THINK]`` are private
+    planning, not a user-visible promise to call a tool. Match GGUF's behavior:
+    classify visible content when present and fall back to reasoning only for a
+    reasoning-only stall.
+    """
+    prefilled_reasoning = ""
+    if reasoning_prefilled:
+        close = _THINK_CLOSE_RE.search(text)
+        if close is None:
+            return text.strip()
+        prefilled_reasoning = text[: close.end()].strip()
+        text = text[close.end() :].strip()
+        if not text:
+            return prefilled_reasoning
+
+    spans = _think_spans_outside_tool_markup(text)
+    if not spans:
+        return text.strip()
+
+    visible: list[str] = []
+    reasoning: list[str] = []
+    cursor = 0
+    for start, end in spans:
+        visible.append(text[cursor:start])
+        reasoning.append(text[start:end])
+        cursor = end
+    visible.append(text[cursor:])
+
+    visible_text = "".join(visible).strip()
+    reasoning_text = "".join(reasoning).strip()
+    if visible_text:
+        return visible_text
+    return "\n".join(part for part in (prefilled_reasoning, reasoning_text) if part).strip()
+
+
 def _looks_like_enabled_bare_json(text: str, enabled_tool_names: Optional[set]) -> bool:
     """True when ``text`` opens with an ENABLED markerless bare-JSON call; an ordinary JSON answer returns False."""
     probe = strip_llama3_leading_sentinels(text.lstrip())
@@ -448,6 +488,7 @@ def run_safetensors_tool_loop(
     confirm_tool_calls: bool = False,
     bypass_permissions: bool = False,
     permission_mode: Optional[str] = None,
+    reasoning_prefilled: bool = False,
 ) -> Generator[dict, None, None]:
     """Drive an agentic tool loop on top of a cumulative-text generator.
 
@@ -956,7 +997,10 @@ def run_safetensors_tool_loop(
                 # (GGUF loop parity). The retry is gated on nudge_tool_calls so
                 # Studio callers (which send True) always nudge, while API callers
                 # who omit the flag keep today's no-reprompt behavior (opt-in).
-                stripped_answer = content_accum.strip()
+                intent_text = _reprompt_intent_text(
+                    content_accum,
+                    reasoning_prefilled = reasoning_prefilled,
+                )
                 if (
                     auto_heal_tool_calls
                     and nudge_tool_calls
@@ -965,7 +1009,7 @@ def run_safetensors_tool_loop(
                     and not rag_autoinjected
                     and not tool_denied
                     and not any(record.executed for record in tool_controller.history)
-                    and is_short_intent_without_action(stripped_answer)
+                    and is_short_intent_without_action(intent_text)
                 ):
                     reprompt_count += 1
                     logger.info(
@@ -973,9 +1017,9 @@ def run_safetensors_tool_loop(
                         "calling tools (%d chars)",
                         reprompt_count,
                         MAX_ACT_REPROMPTS,
-                        len(stripped_answer),
+                        len(intent_text),
                     )
-                    conversation.append({"role": "assistant", "content": stripped_answer})
+                    conversation.append({"role": "assistant", "content": intent_text})
                     tool_hint = " or ".join(_active_tool_names(active_tools)) or "an available tool"
                     conversation.append(
                         {

studio/backend/routes/inference.py

@@ -8886,6 +8886,7 @@ async def openai_chat_completions(
                 permission_mode = payload.permission_mode,
                 use_adapter = payload.use_adapter,
                 stats_holder = _sf_stats_holder,
+                reasoning_prefilled = _sf_reasoning_prefilled,
             )
 
         _sf_tool_sentinel = object()

studio/backend/tests/test_safetensors_tool_loop.py

@@ -3205,6 +3205,196 @@ class TestLoopBehaviour:
 class TestLoopRePrompt:
     """Plan-without-action re-prompt parity with GGUF: nudge instead of terminating, up to ``MAX_ACT_REPROMPTS`` extra slots. Studio always nudges, so these drive the loop with ``nudge_tool_calls=True``."""
 
+    def test_reasoning_intent_does_not_reprompt_a_visible_answer(self):
+        generations = 0
+
+        def _gen(_messages, active_tools = None):
+            nonlocal generations
+            generations += 1
+            yield (
+                "<think>Let me prepare the requested summary carefully.</think>"
+                "This is the final visible answer."
+            )
+
+        exec_fn = FakeExecuteTool([])
+        events = _collect_events(
+            run_safetensors_tool_loop(
+                single_turn = _gen,
+                messages = [{"role": "user", "content": "summarize this"}],
+                tools = [{"type": "function", "function": {"name": "web_search"}}],
+                execute_tool = exec_fn,
+                nudge_tool_calls = True,
+            )
+        )
+
+        assert generations == 1
+        assert exec_fn.calls == []
+        contents = [e["text"] for e in events if e["type"] == "content"]
+        assert contents[-1].endswith("This is the final visible answer.")
+
+    def test_prefilled_reasoning_intent_does_not_reprompt_a_visible_answer(self):
+        generations = 0
+
+        def _gen(_messages, active_tools = None):
+            nonlocal generations
+            generations += 1
+            yield "Let me prepare the requested summary carefully.</think>This is the final visible answer."
+
+        exec_fn = FakeExecuteTool([])
+        events = _collect_events(
+            run_safetensors_tool_loop(
+                single_turn = _gen,
+                messages = [{"role": "user", "content": "summarize this"}],
+                tools = [{"type": "function", "function": {"name": "web_search"}}],
+                execute_tool = exec_fn,
+                nudge_tool_calls = True,
+                reasoning_prefilled = True,
+            )
+        )
+
+        assert generations == 1
+        assert exec_fn.calls == []
+        contents = [e["text"] for e in events if e["type"] == "content"]
+        assert contents[-1].endswith("This is the final visible answer.")
+
+    def test_prefilled_reasoning_with_reemitted_think_does_not_reprompt(self):
+        generations = 0
+
+        def _gen(_messages, active_tools = None):
+            nonlocal generations
+            generations += 1
+            yield (
+                "Let me prepare the requested summary carefully."
+                "<think>more private planning</think>This is the final visible answer."
+            )
+
+        exec_fn = FakeExecuteTool([])
+        events = _collect_events(
+            run_safetensors_tool_loop(
+                single_turn = _gen,
+                messages = [{"role": "user", "content": "summarize this"}],
+                tools = [{"type": "function", "function": {"name": "web_search"}}],
+                execute_tool = exec_fn,
+                nudge_tool_calls = True,
+                reasoning_prefilled = True,
+            )
+        )
+
+        assert generations == 1
+        assert exec_fn.calls == []
+        contents = [e["text"] for e in events if e["type"] == "content"]
+        assert contents[-1].endswith("This is the final visible answer.")
+
+    def test_prefilled_reasoning_with_later_think_does_not_reprompt(self):
+        generations = 0
+
+        def _gen(_messages, active_tools = None):
+            nonlocal generations
+            generations += 1
+            yield (
+                "private prefilled planning</think>"
+                "<think>Let me prepare the requested summary carefully.</think>"
+                "This is the final visible answer."
+            )
+
+        exec_fn = FakeExecuteTool([])
+        events = _collect_events(
+            run_safetensors_tool_loop(
+                single_turn = _gen,
+                messages = [{"role": "user", "content": "summarize this"}],
+                tools = [{"type": "function", "function": {"name": "web_search"}}],
+                execute_tool = exec_fn,
+                nudge_tool_calls = True,
+                reasoning_prefilled = True,
+            )
+        )
+
+        assert generations == 1
+        assert exec_fn.calls == []
+        contents = [e["text"] for e in events if e["type"] == "content"]
+        assert contents[-1].endswith("This is the final visible answer.")
+
+    def test_reasoning_only_intent_still_reprompts_and_uses_a_tool(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["<think>Let me search for that.</think>"],
+                ['<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>'],
+                ["Here is the answer."],
+            ],
+            exec_results = ["result"],
+            nudge_tool_calls = True,
+        )
+
+        events = _collect_events(loop)
+
+        assert exec_fn.calls == [("web_search", {"query": "cats"})]
+        contents = [e["text"] for e in events if e["type"] == "content"]
+        assert contents[-1] == "Here is the answer."
+
+    def test_prefilled_no_close_reasoning_intent_still_reprompts(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["I need more context.<think>Let me search for that."],
+                ['<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>'],
+                ["Here is the answer."],
+            ],
+            exec_results = ["result"],
+            nudge_tool_calls = True,
+            reasoning_prefilled = True,
+        )
+
+        events = _collect_events(loop)
+
+        assert exec_fn.calls == [("web_search", {"query": "cats"})]
+        contents = [e["text"] for e in events if e["type"] == "content"]
+        assert contents[-1] == "Here is the answer."
+
+    def test_prefilled_reasoning_prefix_is_kept_for_reasoning_only_reprompt(self):
+        loop, exec_fn = _make_loop(
+            turns = [
+                ["Let me search for that.</think><think>checking details</think>"],
+                ['<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>'],
+                ["Here is the answer."],
+            ],
+            exec_results = ["result"],
+            nudge_tool_calls = True,
+            reasoning_prefilled = True,
+        )
+
+        events = _collect_events(loop)
+
+        assert exec_fn.calls == [("web_search", {"query": "cats"})]
+        contents = [e["text"] for e in events if e["type"] == "content"]
+        assert contents[-1] == "Here is the answer."
+
+    def test_reprompt_history_uses_visible_intent_text(self):
+        captured: list[list[dict]] = []
+
+        def _gen(messages, active_tools = None):
+            captured.append([dict(message) for message in messages])
+            if len(captured) == 1:
+                yield "<think>private planning details</think>Let me search for that."
+            elif len(captured) == 2:
+                yield '<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>'
+            else:
+                yield "Here is the answer."
+
+        exec_fn = FakeExecuteTool(["result"])
+        events = _collect_events(
+            run_safetensors_tool_loop(
+                single_turn = _gen,
+                messages = [{"role": "user", "content": "find cats"}],
+                tools = [{"type": "function", "function": {"name": "web_search"}}],
+                execute_tool = exec_fn,
+                nudge_tool_calls = True,
+            )
+        )
+
+        assert exec_fn.calls == [("web_search", {"query": "cats"})]
+        assert captured[1][1] == {"role": "assistant", "content": "Let me search for that."}
+        contents = [e["text"] for e in events if e["type"] == "content"]
+        assert contents[-1] == "Here is the answer."
+
     def test_intent_signal_triggers_reprompt(self):
         # Turn 1: intent signal, no tool call.
         # Turn 2 (re-prompt): proper tool call -> executes.