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: add a run setting to turn off tool-call deduplication (#11685)

frostyautumnleaff5aff5827bd5
diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index 52b7e9a1d13..cdc17ebcf0b 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -1221,6 +1221,7 @@ class InferenceBackend:
rag_scope: Optional[dict] = None,
presence_penalty: float = 0.0,
reasoning_prefilled: bool = False,
+ deduplicate_tool_calls: bool = True,
):
"""Run an agentic tool loop on top of ``generate_chat_response``.
@@ -1308,6 +1309,7 @@ class InferenceBackend:
execute_tool = execute_tool,
cancel_event = cancel_event,
auto_heal_tool_calls = auto_heal_tool_calls,
+ deduplicate_tool_calls = deduplicate_tool_calls,
nudge_tool_calls = nudge_tool_calls,
max_tool_iterations = max_tool_iterations,
tool_call_timeout = tool_call_timeout,
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index d64d2dea36c..779463eea31 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -36724,6 +36724,7 @@ class LlamaCppBackend:
on_decode_slot: Optional[Callable[[str, int], None]] = None,
thinking_budget_tokens: Optional[int] = None,
mcp_image = None,
+ deduplicate_tool_calls: bool = True,
instruction_anchor_ids = None,
sandbox_level: Optional[str] = None,
) -> Generator[dict, None, None]:
@@ -37045,6 +37046,7 @@ class LlamaCppBackend:
tool_controller = ToolLoopController(
tools = controller_tools,
auto_heal_tool_calls = auto_heal_tool_calls,
+ deduplicate_tool_calls = deduplicate_tool_calls,
session_id = session_id,
thread_id = thread_id,
)
@@ -38901,7 +38903,10 @@ class LlamaCppBackend:
_novel_kept = 0
_novel_at_last_keep: dict = {}
_deduped: list = []
- for _tc in tool_calls:
+ if not deduplicate_tool_calls:
+ _deduped = tool_calls[:_MAX_TOOL_CALLS_PER_TURN]
+ _over_cap = tool_calls[_MAX_TOOL_CALLS_PER_TURN:]
+ for _tc in tool_calls if deduplicate_tool_calls else ():
_fn = _tc.get("function", {}) or {}
_key = (_fn.get("name", ""), str(_fn.get("arguments", "")))
if _fn.get("name") in _WORKSPACE_TOOLS:
diff --git a/studio/backend/core/inference/openai_codex_tool_loop.py b/studio/backend/core/inference/openai_codex_tool_loop.py
index e586ae8ce51..d76560f44d1 100644
--- a/studio/backend/core/inference/openai_codex_tool_loop.py
+++ b/studio/backend/core/inference/openai_codex_tool_loop.py
@@ -50,6 +50,7 @@ class CodexToolPolicy:
bypass_permissions: bool
rag_scope: dict[str, Any] | None
nudge_tool_calls: bool | None = None
+ deduplicate_tool_calls: bool | None = None
sandbox_level: str = "high"
@@ -123,6 +124,7 @@ def stream_codex_with_studio_tools(
rag_scope = policy.rag_scope,
auto_heal = False,
nudge_tool_calls = policy.nudge_tool_calls,
+ deduplicate_tool_calls = policy.deduplicate_tool_calls,
sandbox_level = policy.sandbox_level,
),
cancel_event = cancel_event,
diff --git a/studio/backend/core/inference/orchestrator.py b/studio/backend/core/inference/orchestrator.py
index 894ec5f3777..769d1d65064 100644
--- a/studio/backend/core/inference/orchestrator.py
+++ b/studio/backend/core/inference/orchestrator.py
@@ -3098,6 +3098,7 @@ class InferenceOrchestrator:
seed: Optional[int] = None,
caller_image_indexes: "tuple[int, ...]" = (),
mcp_image = None,
+ deduplicate_tool_calls: bool = True,
context_overflow: Optional[str] = None,
context_policy: Optional[str] = None,
compaction_headroom_ratio: Optional[float] = None,
@@ -3280,6 +3281,7 @@ class InferenceOrchestrator:
execute_tool = execute_tool,
cancel_event = cancel_event,
auto_heal_tool_calls = auto_heal_tool_calls,
+ deduplicate_tool_calls = deduplicate_tool_calls,
nudge_tool_calls = nudge_tool_calls,
max_tool_iterations = max_tool_iterations,
tool_call_timeout = tool_call_timeout,
diff --git a/studio/backend/core/inference/safetensors_agentic.py b/studio/backend/core/inference/safetensors_agentic.py
index e665f76b96e..481d1463e9b 100644
--- a/studio/backend/core/inference/safetensors_agentic.py
+++ b/studio/backend/core/inference/safetensors_agentic.py
@@ -621,6 +621,7 @@ def run_safetensors_tool_loop(
execute_tool: Callable[..., str],
cancel_event: Optional[threading.Event] = None,
auto_heal_tool_calls: bool = True,
+ deduplicate_tool_calls: bool = True,
nudge_tool_calls: Optional[bool] = None,
max_tool_iterations: int = 25,
tool_call_timeout: int = 300,
@@ -771,6 +772,7 @@ def run_safetensors_tool_loop(
tool_controller = ToolLoopController(
tools = (None if unrestricted_tools else _authorized),
auto_heal_tool_calls = auto_heal_tool_calls,
+ deduplicate_tool_calls = deduplicate_tool_calls,
session_id = session_id,
thread_id = thread_id,
)
@@ -1421,7 +1423,10 @@ def run_safetensors_tool_loop(
novel_kept = 0
novel_at_last_keep: dict = {}
deduped: list = []
- for _tc in tool_calls:
+ if not deduplicate_tool_calls:
+ deduped = tool_calls[:_MAX_TOOL_CALLS_PER_TURN]
+ over_cap = tool_calls[_MAX_TOOL_CALLS_PER_TURN:]
+ for _tc in tool_calls if deduplicate_tool_calls else ():
_fn = _tc.get("function", {}) or {}
_key = (_fn.get("name", ""), str(_fn.get("arguments", "")))
if _fn.get("name") in _WORKSPACE_TOOLS:
diff --git a/studio/backend/core/inference/studio_tool_loop.py b/studio/backend/core/inference/studio_tool_loop.py
index 67ce6fa4dc8..0613c095f66 100644
--- a/studio/backend/core/inference/studio_tool_loop.py
+++ b/studio/backend/core/inference/studio_tool_loop.py
@@ -407,6 +407,8 @@ class ToolLoopPolicy:
# Called when a provider turn ends, however it ended. Headerless only: clears the stripper's withheld-call flag,
# which the wire cannot always close because a turn may end on [DONE] alone.
on_provider_turn_end: Callable[[], None] | None = None
+ # None keeps the request default (on); explicit booleans win.
+ deduplicate_tool_calls: bool | None = None
sandbox_level: str = "high"
@@ -1486,6 +1488,7 @@ async def stream_with_studio_tools(
controller = ToolLoopController(
tools = tools,
auto_heal_tool_calls = policy.auto_heal is not False,
+ deduplicate_tool_calls = policy.deduplicate_tool_calls is not False,
session_id = session_id,
thread_id = thread_id,
)
diff --git a/studio/backend/core/inference/tool_loop_controller.py b/studio/backend/core/inference/tool_loop_controller.py
index 07a95186b61..5210e8675c6 100644
--- a/studio/backend/core/inference/tool_loop_controller.py
+++ b/studio/backend/core/inference/tool_loop_controller.py
@@ -1159,6 +1159,7 @@ class ToolLoopController:
*,
tools: Sequence[Mapping[str, Any]] | None,
auto_heal_tool_calls: bool = True,
+ deduplicate_tool_calls: bool = True,
one_shot_tools: frozenset[str] = _ONE_SHOT_TOOLS,
duplicate_noop_limit: int = 2,
session_id: str | None = None,
@@ -1172,6 +1173,7 @@ class ToolLoopController:
name for name in (_tool_name_from_schema(tool) for tool in self._tools) if name
}
self._auto_heal_tool_calls = auto_heal_tool_calls
+ self._deduplicate_tool_calls = deduplicate_tool_calls
self._one_shot_tools = one_shot_tools
self._completed_one_shot_tools: set[str] = set()
self._successful_keys: set[str] = set()
@@ -1247,7 +1249,7 @@ class ToolLoopController:
elif self._restrict_to_allowed and tool_name not in self._allowed_tool_names:
action = "disabled"
noop = _noop_result("disabled", tool_name)
- elif key in self._successful_keys:
+ elif self._deduplicate_tool_calls and key in self._successful_keys:
action = "duplicate"
noop = _noop_result("duplicate", tool_name)
diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index 57b130cab47..7c3ec4c945c 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -2858,6 +2858,14 @@ class ChatCompletionRequest(BaseModel):
"UNSLOTH_TOOL_CALL_NUDGE=1 flips the process default."
),
)
+ deduplicate_tool_calls: Optional[bool] = Field(
+ True,
+ description = (
+ "[x-unsloth] When false, a tool call identical to one that already "
+ "succeeded in this response runs again instead of being answered "
+ "with a duplicate notice. Default on."
+ ),
+ )
context_overflow: Optional[Literal["error", "truncate_middle", "truncate_oldest"]] = Field(
None,
description = (
diff --git a/studio/backend/routes/chat_history.py b/studio/backend/routes/chat_history.py
index 3c17c3fe167..885256008e7 100644
--- a/studio/backend/routes/chat_history.py
+++ b/studio/backend/routes/chat_history.py
@@ -497,6 +497,7 @@ class ChatSettingsPayload(BaseModel):
searchImages: Optional[bool] = None
autoHealToolCalls: Optional[bool] = None
nudgeToolCalls: Optional[bool] = None
+ deduplicateToolCalls: Optional[bool] = None
maxToolCallsPerMessage: Optional[int] = Field(default = None, ge = 0)
toolCallTimeout: Optional[int] = Field(default = None, ge = 1)
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 73447d78f26..9b59515f6c7 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -28678,6 +28678,7 @@ async def _proxy_to_external_provider(
bypass_permissions = bool(payload.bypass_permissions),
rag_scope = payload.rag_scope,
nudge_tool_calls = payload.nudge_tool_calls,
+ deduplicate_tool_calls = payload.deduplicate_tool_calls,
)
if studio_tool_payloads
else None
@@ -29172,6 +29173,7 @@ async def _proxy_to_external_provider(
rag_scope = payload.rag_scope,
auto_heal = payload.auto_heal_tool_calls,
nudge_tool_calls = payload.nudge_tool_calls,
+ deduplicate_tool_calls = payload.deduplicate_tool_calls,
# Matches the strip below: only a headerless caller has its calls
# withheld, so only it needs a healed one the wire never carried flagged.
on_withheld_tool_call = (None if _ui_events else _tool_call_stripper.arm),
@@ -31664,6 +31666,7 @@ async def produce_openai_chat_completions(
preserve_thinking = payload.preserve_thinking,
continue_final_message = _continue_final_message(payload, thought = True),
auto_heal_tool_calls = _gguf_auto_heal_tool_calls,
+ deduplicate_tool_calls = payload.deduplicate_tool_calls is not False,
nudge_tool_calls = payload.nudge_tool_calls,
tool_choice = payload.tool_choice,
max_tool_iterations = payload.max_tool_calls_per_message
@@ -33795,6 +33798,7 @@ async def produce_openai_chat_completions(
preserve_thinking = payload.preserve_thinking,
continue_final_message = _sf_continue,
auto_heal_tool_calls = _sf_auto_heal_tool_calls,
+ deduplicate_tool_calls = payload.deduplicate_tool_calls is not False,
nudge_tool_calls = payload.nudge_tool_calls,
max_tool_iterations = _sf_tool_budget,
tool_call_timeout = payload.tool_call_timeout
diff --git a/studio/backend/routes/managed_engine_chat.py b/studio/backend/routes/managed_engine_chat.py
index a6de6230056..82dd680c2c0 100644
--- a/studio/backend/routes/managed_engine_chat.py
+++ b/studio/backend/routes/managed_engine_chat.py
@@ -234,6 +234,7 @@ async def managed_tool_chat(
rag_scope = payload.rag_scope,
auto_heal = payload.auto_heal_tool_calls,
nudge_tool_calls = payload.nudge_tool_calls,
+ deduplicate_tool_calls = payload.deduplicate_tool_calls,
on_withheld_tool_call = None if ui_events else stripper.arm,
on_provider_turn_end = None if ui_events else stripper.end_turn,
),
diff --git a/studio/backend/tests/test_llama_cpp_tool_loop.py b/studio/backend/tests/test_llama_cpp_tool_loop.py
index 567a76ede62..44c251c6091 100644
--- a/studio/backend/tests/test_llama_cpp_tool_loop.py
+++ b/studio/backend/tests/test_llama_cpp_tool_loop.py
@@ -1704,6 +1704,31 @@ def test_duplicate_web_search_noop_allows_distinct_followup_tool(monkeypatch):
assert len(duplicate_nudges) == 1
[email protected]("deduplicate", [True, False])
+def test_rebuild_without_an_edit_runs_again_only_when_deduplication_is_off(
+ monkeypatch, deduplicate
+):
+ # #10379: a second identical build command after no file edit is a duplicate by default.
+ build = {"command": "./gradlew build"}
+ streams = [
+ [_tool_call_sse("terminal", build, "call_build_1"), _done()],
+ [_tool_call_sse("terminal", build, "call_build_2"), _done()],
+ [_sse({"content": "Built."}), _done()],
+ ]
+ backend = _make_backend(monkeypatch, streams, [])
+ calls = _record_tool_calls(monkeypatch, lambda name: "BUILD SUCCESSFUL")
+
+ _run_tool_loop(
+ backend,
+ [{"role": "user", "content": "build it twice"}],
+ [{"type": "function", "function": {"name": "terminal"}}],
+ max_tool_iterations = 3,
+ deduplicate_tool_calls = deduplicate,
+ )
+
+ assert calls == [("terminal", build)] * (1 if deduplicate else 2)
+
+
def test_repeated_duplicate_noop_transitions_to_final_pass(monkeypatch):
first_search = [
_tool_call_sse("web_search", {"query": "gpu prices 2026"}, "call_search_1"),
@@ -5226,6 +5251,33 @@ def test_gguf_textual_fallback_collapses_duplicate_tool_calls(monkeypatch):
assert len(calls) == 1, [c[0] for c in calls]
[email protected]("deduplicate", [True, False])
+def test_gguf_textual_fallback_keeps_identical_calls_when_deduplication_is_off(
+ monkeypatch, deduplicate
+):
+ blocks = '<tool_call>{"name":"web_search","arguments":{"query":"cats"}}</tool_call>' * 3
+ first_stream = [_sse({"content": blocks}), _done()]
+ final_stream = [_sse({"content": "done"}), _done()]
+ backend, _payloads = _backend_and_payloads(monkeypatch, [first_stream, final_stream])
+
+ calls: list[tuple[str, dict]] = []
+ monkeypatch.setattr(
+ "core.inference.tools.execute_tool",
+ lambda name, arguments, **_k: calls.append((name, arguments)) or "OK",
+ )
+
+ list(
+ backend.generate_chat_completion_with_tools(
+ messages = [{"role": "user", "content": "cats"}],
+ tools = [{"type": "function", "function": {"name": "web_search"}}],
+ max_tool_iterations = 1,
+ deduplicate_tool_calls = deduplicate,
+ )
+ )
+
+ assert len(calls) == (1 if deduplicate else 3), [c[0] for c in calls]
+
+
def test_gguf_drain_truncated_enabled_name_json_preserved_when_auto_heal_disabled(monkeypatch):
"""Auto-Heal OFF keeps a truncated enabled-name fragment visible; ON suppresses it (strip gated on auto_heal_tool_calls)."""
diff --git a/studio/backend/tests/test_managed_engine_tools.py b/studio/backend/tests/test_managed_engine_tools.py
index 7f0addc1fb5..b287eeb1532 100644
--- a/studio/backend/tests/test_managed_engine_tools.py
+++ b/studio/backend/tests/test_managed_engine_tools.py
@@ -289,6 +289,33 @@ def test_run_automatically_streams_arm_the_no_sandbox_gate(native, monkeypatch):
assert seen == [True]
[email protected]("dedup", [None, False, True])
+def test_deduplicate_tool_calls_reaches_the_managed_loop(native, monkeypatch, dedup):
+ from core.inference import studio_tool_loop
+ from routes import managed_engine_chat
+
+ seen = []
+ original = managed_engine_chat.stream_with_studio_tools
+
+ def capture(*args, **kwargs):
+ seen.append(kwargs["policy"].deduplicate_tool_calls is not False)
+ return original(*args, **kwargs)
+
+ monkeypatch.setattr(managed_engine_chat, "stream_with_studio_tools", capture)
+ monkeypatch.setattr(studio_tool_loop, "execute_tool", lambda *a, **k: "3973")
+ extra = {} if dedup is None else {"deduplicate_tool_calls": dedup}
+ run(
+ route_test._request(
+ enable_tools = True,
+ enabled_tools = ["python"],
+ permission_mode = "off",
+ max_tool_calls_per_message = 1,
+ **extra,
+ )
+ )
+ assert seen == [dedup is not False]
+
+
@pytest.mark.parametrize(
"choice", ["auto", "required", "none", {"type": "function", "function": {"name": "lookup"}}]
)
diff --git a/studio/backend/tests/test_safetensors_tool_loop.py b/studio/backend/tests/test_safetensors_tool_loop.py
index 9dba6d99f1c..38c929328e7 100644
--- a/studio/backend/tests/test_safetensors_tool_loop.py
+++ b/studio/backend/tests/test_safetensors_tool_loop.py
@@ -2795,6 +2795,31 @@ class TestLoopBehaviour:
)
assert "previous tool request" not in after[1]["content"].lower()
+ @pytest.mark.parametrize("deduplicate", [True, False])
+ def test_identical_calls_in_one_turn_run_when_deduplication_is_off(self, deduplicate):
+ call = '<tool_call>{"name":"python","arguments":{"code":"print(1)"}}</tool_call>'
+ turns = iter([[call * 2], ["final"]])
+
+ def fake_single_turn(messages, active_tools = None):
+ acc = ""
+ for chunk in next(turns):
+ acc += chunk
+ yield acc
+
+ exec_fn = FakeExecuteTool(["1", "1"])
+ _collect_events(
+ run_safetensors_tool_loop(
+ single_turn = fake_single_turn,
+ messages = [{"role": "user", "content": "run it twice"}],
+ tools = [{"type": "function", "function": {"name": "python"}}],
+ execute_tool = exec_fn,
+ max_tool_iterations = 4,
+ deduplicate_tool_calls = deduplicate,
+ )
+ )
+
+ assert exec_fn.calls == [("python", {"code": "print(1)"})] * (1 if deduplicate else 2)
+
def test_duplicate_tool_call_internal_noop_allows_distinct_followup_tool(self):
captured_messages: list[list[dict]] = []
captured_tool_names: list[list[str]] = []
diff --git a/studio/backend/tests/test_tool_loop_controller.py b/studio/backend/tests/test_tool_loop_controller.py
index 6f61834470b..792dbb1a1eb 100644
--- a/studio/backend/tests/test_tool_loop_controller.py
+++ b/studio/backend/tests/test_tool_loop_controller.py
@@ -264,6 +264,22 @@ def test_repeated_successful_duplicate_becomes_terminal_after_one_recovery_nudge
assert controller.active_tools() == []
+def test_deduplication_can_be_disabled():
+ controller = ToolLoopController(
+ tools = [_tool("web_search"), _tool("python")],
+ deduplicate_tool_calls = False,
+ )
+ first = controller.prepare_call(_call("web_search", {"query": "gpu prices"}, "call_a"))
+ controller.record_result(first, "ok")
+
+ repeat = controller.prepare_call(_call("web_search", {"query": "gpu prices"}, "call_b"))
+ assert repeat.action == "execute"
+ assert repeat.should_execute
+ controller.record_result(repeat, "ok")
+ assert not controller.force_final_answer
+ _shared_setup_1(controller)
+
+
def test_command_can_run_again_after_a_file_edit():
controller = ToolLoopController(
tools = [_tool("terminal"), _tool("edit_file"), _tool("web_search")]
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index aacb4b30ae5..f0a417c9ef9 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -6778,6 +6778,7 @@ export function createOpenAIStreamAdapter(
// false, not omitted: omission follows UNSLOTH_TOOL_CALL_NUDGE, which the
// launchers set to 1 when unset, so this loop would keep nudging (#9686, #9125).
nudge_tool_calls: false,
+ deduplicate_tool_calls: runtime.deduplicateToolCalls,
// This branch runs the tools here, so say so by name: enabled_tools ["web_search"] is
// byte-identical to an older bundle's hosted search.
run_tools_locally: true,
@@ -7019,6 +7020,7 @@ export function createOpenAIStreamAdapter(
: {}),
auto_heal_tool_calls: runtime.autoHealToolCalls,
nudge_tool_calls: runtime.nudgeToolCalls,
+ deduplicate_tool_calls: runtime.deduplicateToolCalls,
max_tool_calls_per_message: runtime.maxToolCallsPerMessage,
tool_call_timeout: (() => {
const mins = runtime.toolCallTimeout;
diff --git a/studio/frontend/src/features/chat/api/chat-settings-api.ts b/studio/frontend/src/features/chat/api/chat-settings-api.ts
index f5723b9d946..d547722c7c0 100644
--- a/studio/frontend/src/features/chat/api/chat-settings-api.ts
+++ b/studio/frontend/src/features/chat/api/chat-settings-api.ts
@@ -49,6 +49,7 @@ export interface PersistedChatSettings {
searchImages?: boolean;
autoHealToolCalls?: boolean;
nudgeToolCalls?: boolean;
+ deduplicateToolCalls?: boolean;
maxToolCallsPerMessage?: number;
toolCallTimeout?: number;
reasoningEnabled?: boolean;
diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
index 45b50113f17..48840e7e355 100644
--- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx
+++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
@@ -1702,6 +1702,7 @@ export function ChatSettingsPanel({
<div className="flex flex-col gap-5">
<AutoHealToolCallsToggle />
<NudgeToolCallsToggle />
+ <DeduplicateToolCallsToggle />
<ConfirmToolCallsToggle />
<BypassPermissionsToggle />
<MaxToolCallsSlider />
@@ -2040,6 +2041,35 @@ function NudgeToolCallsToggle() {
);
}
+function DeduplicateToolCallsToggle() {
+ const deduplicateToolCalls = useChatRuntimeStore(
+ (s) => s.deduplicateToolCalls,
+ );
+ const setDeduplicateToolCalls = useChatRuntimeStore(
+ (s) => s.setDeduplicateToolCalls,
+ );
+
+ return (
+ <div className="flex min-h-8 items-center justify-between gap-3">
+ <div className="flex min-w-0 items-center gap-1.5">
+ <span className="min-w-0 text-ui-13 font-medium leading-[1.25] tracking-nav text-nav-fg">
+ Deduplicate Tool Calls
+ </span>
+ <InfoHint>
+ Skips a tool call identical to one that already succeeded in this
+ response and tells the model it was a duplicate. Turn off to let a
+ repeated call run again.
+ </InfoHint>
+ </div>
+ <Switch
+ className="panel-switch shrink-0"
+ checked={deduplicateToolCalls}
+ onCheckedChange={setDeduplicateToolCalls}
+ />
+ </div>
+ );
+}
+
function ConfirmToolCallsToggle() {
const setConfirmToolCalls = useChatRuntimeStore((s) => s.setConfirmToolCalls);
const setPermissionMode = useChatRuntimeStore((s) => s.setPermissionMode);
diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
index 06a3d99e575..395d5affecf 100644
--- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
@@ -2444,6 +2444,7 @@ type ChatRuntimeStore = {
generatingStatus: string | null;
autoHealToolCalls: boolean;
nudgeToolCalls: boolean;
+ deduplicateToolCalls: boolean;
autoCompactEnabled: boolean;
maxToolCallsPerMessage: number;
toolCallTimeout: number;
@@ -2741,6 +2742,7 @@ type ChatRuntimeStore = {
clearActiveDiffusionCanvasForThread: (threadId: string | null) => void;
setAutoHealToolCalls: (enabled: boolean) => void;
setNudgeToolCalls: (enabled: boolean) => void;
+ setDeduplicateToolCalls: (enabled: boolean) => void;
setAutoCompactEnabled: (enabled: boolean) => void;
setMaxToolCallsPerMessage: (value: number) => void;
setToolCallTimeout: (value: number) => void;
@@ -2786,6 +2788,7 @@ type ScalarSettingKey =
| "searchImages"
| "autoHealToolCalls"
| "nudgeToolCalls"
+ | "deduplicateToolCalls"
| "autoCompactEnabled"
| "maxToolCallsPerMessage"
| "toolCallTimeout"
@@ -2837,6 +2840,7 @@ const SCALAR_SETTING_KEYS = [
"searchImages",
"autoHealToolCalls",
"nudgeToolCalls",
+ "deduplicateToolCalls",
"autoCompactEnabled",
"maxToolCallsPerMessage",
"toolCallTimeout",
@@ -4235,6 +4239,7 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
activeDiffusionCanvasByThreadId: {},
autoHealToolCalls: true,
nudgeToolCalls: true,
+ deduplicateToolCalls: true,
autoCompactEnabled: DEFAULT_AUTO_COMPACT_ENABLED,
maxToolCallsPerMessage: 25,
toolCallTimeout: 5,
@@ -5876,6 +5881,18 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
queuedSettingsEpoch: state.queuedSettingsEpoch + 1,
};
}),
+ setDeduplicateToolCalls: (deduplicateToolCalls) =>
+ set((state) => {
+ setScalarSettingVersion(
+ "deduplicateToolCalls",
+ deduplicateToolCalls,
+ state.deduplicateToolCalls,
+ );
+ return {
+ deduplicateToolCalls,
+ queuedSettingsEpoch: state.queuedSettingsEpoch + 1,
+ };
+ }),
setAutoCompactEnabled: (autoCompactEnabled) =>
set((state) => {
setScalarSettingVersion(
diff --git a/studio/frontend/src/features/chat/types/api.ts b/studio/frontend/src/features/chat/types/api.ts
index 0e30f98736a..7e2d05800eb 100644
--- a/studio/frontend/src/features/chat/types/api.ts
+++ b/studio/frontend/src/features/chat/types/api.ts
@@ -735,6 +735,7 @@ export interface OpenAIChatCompletionsRequest {
/** Run the selected tools here rather than as the provider's hosted builtins. */
run_tools_locally?: boolean;
nudge_tool_calls?: boolean;
+ deduplicate_tool_calls?: boolean;
/** Local GGUF overflow policy. Rolling mode preserves the transcript but omits oldest turns. */
context_overflow?: "error" | "truncate_middle" | "truncate_oldest";
/** Override UNSLOTH_CONTEXT_POLICY for this local GGUF request. */
diff --git a/studio/frontend/src/features/chat/utils/chat-settings-storage.ts b/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
index 713717412af..a7ab63cf135 100644
--- a/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
+++ b/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
@@ -298,6 +298,7 @@ export function sanitizeChatSettings(value: unknown): PersistedChatSettings {
);
const autoHealToolCalls = sanitizeBool(value.autoHealToolCalls);
const nudgeToolCalls = sanitizeBool(value.nudgeToolCalls);
+ const deduplicateToolCalls = sanitizeBool(value.deduplicateToolCalls);
const autoCompactEnabled = sanitizeBool(value.autoCompactEnabled);
const maxToolCallsPerMessage = sanitizeInt(value.maxToolCallsPerMessage, 0);
const toolCallTimeout = sanitizeInt(value.toolCallTimeout, 1);
@@ -330,6 +331,9 @@ export function sanitizeChatSettings(value: unknown): PersistedChatSettings {
if (nudgeToolCalls !== undefined) {
settings.nudgeToolCalls = nudgeToolCalls;
}
+ if (deduplicateToolCalls !== undefined) {
+ settings.deduplicateToolCalls = deduplicateToolCalls;
+ }
if (autoCompactEnabled !== undefined) {
settings.autoCompactEnabled = autoCompactEnabled;
}
@@ -397,6 +401,7 @@ export function isEmptyChatSettings(settings: PersistedChatSettings): boolean {
settings.allowArtifactNetworkAccess === undefined &&
settings.autoHealToolCalls === undefined &&
settings.nudgeToolCalls === undefined &&
+ settings.deduplicateToolCalls === undefined &&
settings.autoCompactEnabled === undefined &&
settings.maxToolCallsPerMessage === undefined &&
settings.toolCallTimeout === undefined &&
diff --git a/studio/frontend/src/features/chat/utils/queued-chat-run-settings.ts b/studio/frontend/src/features/chat/utils/queued-chat-run-settings.ts
index 7f33397c174..68a7ff4793c 100644
--- a/studio/frontend/src/features/chat/utils/queued-chat-run-settings.ts
+++ b/studio/frontend/src/features/chat/utils/queued-chat-run-settings.ts
@@ -44,6 +44,7 @@ const QUEUED_SETTING_KEYS = [
"loadedIsMlx",
"autoHealToolCalls",
"nudgeToolCalls",
+ "deduplicateToolCalls",
"maxToolCallsPerMessage",
"toolCallTimeout",
"autoCompactEnabled",