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.

fix(studio): support preserve thinking for llama.cpp connections (#11706)

Lee Jackson4fabb25e07dc
diff --git a/studio/backend/core/inference/external_provider.py b/studio/backend/core/inference/external_provider.py
index 7a078ba3688..9a5c3cdd706 100644
--- a/studio/backend/core/inference/external_provider.py
+++ b/studio/backend/core/inference/external_provider.py
@@ -1465,6 +1465,7 @@ class ExternalProviderClient:
continue_final_message: Optional[bool] = None,
response_format: Optional[dict[str, Any]] = None,
stream: bool = True,
+ preserve_thinking: Optional[bool] = None,
) -> AsyncGenerator[str, None]:
"""Yield OpenAI-format SSE lines from the external provider. OpenAI-compatible providers
forward lines verbatim; for Anthropic the native Messages API SSE is translated.
@@ -1634,6 +1635,9 @@ class ExternalProviderClient:
for field in provider_info.get("body_omit", ()):
body.pop(field, None)
+ if self.provider_type == "llama_cpp" and preserve_thinking is not None:
+ body["chat_template_kwargs"] = {"preserve_thinking": preserve_thinking}
+
# Kimi thinking is a top-level body field. kimi-k2-thinking is always on (ignore the toggle); kimi-k2.6
# defaults on, can be disabled. `keep: all` preserves every chunk for the UI panel.
if self.provider_type == "kimi" and enable_thinking is not None:
diff --git a/studio/backend/core/inference/external_tool_transport.py b/studio/backend/core/inference/external_tool_transport.py
index 5a815178577..6817e52bdf6 100644
--- a/studio/backend/core/inference/external_tool_transport.py
+++ b/studio/backend/core/inference/external_tool_transport.py
@@ -62,6 +62,9 @@ class OAICompatTransport:
self._model = model
self._continue_final_message = continue_final_message
self._request_kwargs = request_kwargs
+ self.preserves_reasoning = (
+ client.provider_type == "llama_cpp" and request_kwargs.get("preserve_thinking") is True
+ )
def stream(
self,
diff --git a/studio/backend/core/inference/studio_tool_loop.py b/studio/backend/core/inference/studio_tool_loop.py
index fd6faf43304..e6e6df11dde 100644
--- a/studio/backend/core/inference/studio_tool_loop.py
+++ b/studio/backend/core/inference/studio_tool_loop.py
@@ -548,6 +548,7 @@ class _Turn:
round: int = 0
healed: list[dict[str, Any]] = field(default_factory = list)
text: list[str] = field(default_factory = list)
+ reasoning: list[str] = field(default_factory = list)
reasoning_extra: dict[str, Any] | None = None
finish_reason: str | None = None
# Results from tools the PROVIDER ran this turn, keyed by call id so a repeated end event cannot record the same
@@ -1363,6 +1364,9 @@ async def stream_with_studio_tools(
delta = choice.get("delta")
delta = delta if isinstance(delta, dict) else {}
+ reasoning = delta.get("reasoning_content")
+ if getattr(transport, "preserves_reasoning", False) and isinstance(reasoning, str):
+ turn.reasoning.append(reasoning)
content = delta.get("content")
raw_calls = delta.get("tool_calls")
extra = delta.get("extra_content")
@@ -1871,6 +1875,8 @@ async def stream_with_studio_tools(
if assistant_message["content"]
else hosted_text
)
+ if turn.reasoning:
+ assistant_message["reasoning_content"] = "".join(turn.reasoning)
if turn.reasoning_extra:
assistant_message["extra_content"] = turn.reasoning_extra
if assistant_tool_calls:
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index dee15877cfb..2f24fec742e 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -22733,6 +22733,11 @@ def _build_external_messages(
result = []
for msg in messages:
+ reasoning = (
+ {"reasoning_content": msg.reasoning_content}
+ if provider_type == "llama_cpp" and msg.role == "assistant" and msg.reasoning_content
+ else {}
+ )
# Drop role=tool messages whose matching server-builtin tool_call was
# filtered above. An orphan tool_result with no matching tool_call is
# rejected by OpenAI Responses and Anthropic.
@@ -22742,19 +22747,24 @@ def _build_external_messages(
and msg.tool_call_id in dropped_server_builtin_tool_call_ids
):
continue
- if isinstance(msg.content, str):
+ if isinstance(msg.content, str) or (msg.content is None and reasoning):
# Drop bare assistant messages with no content AND no tool_calls
# (some providers reject empty assistant turns). Preserve assistant
# turns whose only payload is tool_calls so multi-turn
# function-call loops round-trip.
- if msg.role == "assistant" and not msg.content.strip() and not msg.tool_calls:
+ if (
+ msg.role == "assistant"
+ and not (msg.content or "").strip()
+ and not msg.tool_calls
+ and not reasoning
+ ):
continue
- out: dict[str, Any] = {"role": msg.role, "content": msg.content}
+ out: dict[str, Any] = {"role": msg.role, "content": msg.content or "", **reasoning}
if msg.role == "assistant" and msg.tool_calls:
_tcs = _filter_tool_calls(msg.tool_calls)
if _tcs:
out["tool_calls"] = _tcs
- elif not msg.content.strip():
+ elif not (msg.content or "").strip() and not reasoning:
# Every tool_call was a dropped synthetic provider card;
# the turn would be an empty
# `{"role":"assistant","content":""}` that some providers
@@ -22829,17 +22839,17 @@ def _build_external_messages(
# `compaction` block; every other provider would 400 on
# the unknown part, so gate by provider_type.
parts.append({"type": "compaction", "content": part.content})
- entry: dict[str, Any] = {"role": msg.role, "content": parts}
+ entry: dict[str, Any] = {"role": msg.role, "content": parts, **reasoning}
if msg.role == "assistant" and msg.tool_calls:
_tcs = _filter_tool_calls(msg.tool_calls)
if _tcs:
entry["tool_calls"] = _tcs
- elif not parts:
+ elif not parts and not reasoning:
# All tool_calls were synthetic and dropped, and no
# content parts survived. Skip rather than forward an
# empty assistant turn that downstream providers reject.
continue
- elif msg.role == "assistant" and not parts:
+ elif msg.role == "assistant" and not parts and not reasoning:
continue
if msg.role == "tool":
if msg.tool_call_id:
@@ -22866,14 +22876,14 @@ def _build_external_messages(
preserved.append(_rp)
elif p.type == "compaction" and anthropic:
preserved.append({"type": "compaction", "content": p.content})
- if msg.role == "assistant" and not preserved:
+ if msg.role == "assistant" and not preserved and not reasoning:
continue
if len(preserved) == 1 and preserved[0]["type"] == "text":
# Single text part collapses to a string for providers that
# don't accept content arrays.
- entry = {"role": msg.role, "content": preserved[0]["text"]}
+ entry = {"role": msg.role, "content": preserved[0]["text"], **reasoning}
else:
- entry = {"role": msg.role, "content": preserved}
+ entry = {"role": msg.role, "content": preserved, **reasoning}
if msg.role == "assistant" and msg.tool_calls:
_tcs = _filter_tool_calls(msg.tool_calls)
if _tcs:
@@ -22885,7 +22895,7 @@ def _build_external_messages(
_has_text = (
isinstance(_entry_content, str) and _entry_content.strip()
) or (isinstance(_entry_content, list) and len(_entry_content) > 0)
- if not _has_text:
+ if not _has_text and not reasoning:
continue
if msg.role == "tool":
if msg.tool_call_id:
@@ -23753,6 +23763,7 @@ async def _proxy_to_external_provider(
repetition_penalty = _repetition_penalty_explicit,
enable_thinking = payload.enable_thinking,
reasoning_effort = payload.reasoning_effort,
+ preserve_thinking = payload.preserve_thinking,
enable_prompt_caching = payload.enable_prompt_caching,
openai_code_exec_container_id = payload.openai_code_exec_container_id,
anthropic_code_exec_container_id = payload.anthropic_code_exec_container_id,
diff --git a/studio/backend/tests/test_external_confirm_gate_and_saved_keys.py b/studio/backend/tests/test_external_confirm_gate_and_saved_keys.py
index afeeab3c622..ea369029c0a 100644
--- a/studio/backend/tests/test_external_confirm_gate_and_saved_keys.py
+++ b/studio/backend/tests/test_external_confirm_gate_and_saved_keys.py
@@ -34,6 +34,7 @@ class FakeExternalClient:
def __init__(self, **kwargs):
FakeExternalClient.last = {"ctor": kwargs, "passthrough": None}
+ self.provider_type = kwargs.get("provider_type")
def stream_chat_completion(self, **kwargs):
FakeExternalClient.last["passthrough"] = kwargs
@@ -404,6 +405,8 @@ def test_transport_cancellation_is_wired_through_the_loop():
from core.inference.external_tool_transport import OAICompatTransport
class _Stalling:
+ provider_type = "custom"
+
def __init__(self):
self.torn_down = False
self.released = asyncio.Event()
diff --git a/studio/backend/tests/test_external_hosted_tool_passthrough.py b/studio/backend/tests/test_external_hosted_tool_passthrough.py
index 36d8f9ae1c7..3a105a096d0 100644
--- a/studio/backend/tests/test_external_hosted_tool_passthrough.py
+++ b/studio/backend/tests/test_external_hosted_tool_passthrough.py
@@ -49,6 +49,7 @@ class FakeExternalClient:
def __init__(self, **kwargs):
FakeExternalClient.last = {"ctor": kwargs, "passthrough": None}
+ self.provider_type = kwargs.get("provider_type")
def stream_chat_completion(self, **kwargs):
FakeExternalClient.last["passthrough"] = kwargs
diff --git a/studio/backend/tests/test_external_hosted_tool_selection.py b/studio/backend/tests/test_external_hosted_tool_selection.py
index 3c483c178fd..c262d9b479f 100644
--- a/studio/backend/tests/test_external_hosted_tool_selection.py
+++ b/studio/backend/tests/test_external_hosted_tool_selection.py
@@ -34,6 +34,7 @@ class _FakeExternalClient:
def __init__(self, **kwargs):
_FakeExternalClient.last = {"ctor": kwargs}
+ self.provider_type = kwargs.get("provider_type")
def stream_chat_completion(self, **kwargs):
async def gen():
diff --git a/studio/backend/tests/test_external_preserve_thinking.py b/studio/backend/tests/test_external_preserve_thinking.py
new file mode 100644
index 00000000000..978b8268874
--- /dev/null
+++ b/studio/backend/tests/test_external_preserve_thinking.py
@@ -0,0 +1,65 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import pytest
+
+from models.inference import ChatMessage
+from routes.inference import _build_external_messages
+from core.inference.external_provider import ExternalProviderClient
+from core.inference.external_tool_transport import OAICompatTransport
+
+
[email protected]("content", ["answer", "", None, [], [{"type": "text", "text": "answer"}]])
[email protected]("vision", [True, False])
[email protected]("with_tool", [True, False])
+def test_llama_history_keeps_reasoning_and_tool_calls(content, vision, with_tool):
+ calls = [
+ {"id": "call_1", "type": "function", "function": {"name": "lookup", "arguments": "{}"}}
+ ]
+ message = ChatMessage(
+ role = "assistant",
+ content = content,
+ reasoning_content = "prior thought",
+ tool_calls = calls if with_tool else None,
+ )
+ result = _build_external_messages([message], vision, provider_type = "llama_cpp")
+ assert len(result) == 1
+ assert result[0]["reasoning_content"] == "prior thought"
+ if with_tool:
+ assert result[0]["tool_calls"][0]["function"]["name"] == "lookup"
+
+
[email protected]("content", ["", None, []])
[email protected]("vision", [True, False])
+def test_llama_reasoning_survives_dropped_server_tool_cards(content, vision):
+ card = {
+ "id": "srv_1",
+ "type": "function",
+ "function": {"name": "web_search", "arguments": '{"_server_tool": true}'},
+ }
+ message = ChatMessage(
+ role = "assistant", content = content, reasoning_content = "prior thought", tool_calls = [card]
+ )
+ result = _build_external_messages([message], vision, provider_type = "llama_cpp")
+ assert len(result) == 1
+ assert result[0]["reasoning_content"] == "prior thought"
+ assert "tool_calls" not in result[0]
+
+
+ "provider", ["custom", "openai", "vllm", "ollama", "anthropic", "openrouter"]
+)
+def test_other_providers_keep_existing_reasoning_policy(provider):
+ message = ChatMessage(role = "assistant", content = "answer", reasoning_content = "prior thought")
+ result = _build_external_messages([message], False, provider_type = provider)
+ assert "reasoning_content" not in result[0]
+
+
[email protected]("provider", ["llama_cpp", "custom", "vllm"])
[email protected]("value", [True, False, None])
+def test_tool_transport_reasoning_policy(provider, value):
+ client = ExternalProviderClient(
+ provider_type = provider, base_url = "http://localhost:8080/v1", api_key = ""
+ )
+ transport = OAICompatTransport(client, model = "test", preserve_thinking = value)
+ assert transport.preserves_reasoning is (provider == "llama_cpp" and value is True)
diff --git a/studio/backend/tests/test_external_provider_sampling_forwarding.py b/studio/backend/tests/test_external_provider_sampling_forwarding.py
index 31ea9af8f2b..b7753b9c884 100644
--- a/studio/backend/tests/test_external_provider_sampling_forwarding.py
+++ b/studio/backend/tests/test_external_provider_sampling_forwarding.py
@@ -164,3 +164,23 @@ def test_the_registry_strips_the_three_for_providers_that_cannot_take_them(provi
assert "repetition_penalty" not in body
assert body["temperature"] == 0.31
assert body["presence_penalty"] == 0.11
+
+
[email protected]("value", [True, False, None])
+def test_llama_preserve_thinking_template_option(value):
+ body = _capture_body("llama_cpp", preserve_thinking = value)
+ if value is None:
+ assert "chat_template_kwargs" not in body
+ else:
+ assert body["chat_template_kwargs"] == {"preserve_thinking": value}
+
+
+def test_llama_preserve_thinking_keeps_enable_thinking():
+ body = _capture_body("llama_cpp", preserve_thinking = True, enable_thinking = False)
+ assert body["chat_template_kwargs"] == {"preserve_thinking": True, "enable_thinking": False}
+
+
[email protected]("provider", ["custom", "vllm", "ollama", "openrouter"])
+def test_preserve_thinking_does_not_leak_to_other_providers(provider):
+ body = _capture_body(provider, preserve_thinking = True)
+ assert "preserve_thinking" not in body.get("chat_template_kwargs", {})
diff --git a/studio/backend/tests/test_external_provider_sampling_over_the_wire.py b/studio/backend/tests/test_external_provider_sampling_over_the_wire.py
index 211ecc78b62..d4572ec1f31 100644
--- a/studio/backend/tests/test_external_provider_sampling_over_the_wire.py
+++ b/studio/backend/tests/test_external_provider_sampling_over_the_wire.py
@@ -152,7 +152,13 @@ def _client_capture(provider_type: str, **kwargs) -> dict:
return server.sampling()
-def _route_capture(**payload_fields) -> dict:
+def _route_capture(
+ *,
+ provider_type = "vllm",
+ messages = None,
+ full_body = False,
+ **payload_fields,
+) -> dict:
"""POST through `_proxy_to_external_provider`, not around it."""
import routes.inference as ri
from starlette.requests import Request
@@ -162,9 +168,9 @@ def _route_capture(**payload_fields) -> dict:
with _Server() as server:
payload = ChatCompletionRequest(
- provider_type = "vllm",
+ provider_type = provider_type,
provider_base_url = server.base_url,
- messages = [{"role": "user", "content": "hi"}],
+ messages = messages or [{"role": "user", "content": "hi"}],
model = "a-model",
stream = True,
**payload_fields,
@@ -193,7 +199,7 @@ def _route_capture(**payload_fields) -> dict:
_run(go)
assert server.bodies, "the route never reached the provider"
- return server.sampling()
+ return server.bodies[-1] if full_body else server.sampling()
def test_a_request_that_never_mentioned_them_forwards_nothing():
@@ -366,3 +372,20 @@ def test_a_stale_frontend_bundle_does_not_start_400ing_a_custom_gateway():
_run(go)
assert server.sampling() == {}, "custom leaked an extension the gateway rejects"
assert "Unrecognized request argument" not in "".join(lines)
+
+
[email protected]("value", [True, False, None])
+def test_preserve_thinking_reaches_llama_server_through_the_route(value):
+ messages = [
+ {"role": "user", "content": "one"},
+ {"role": "assistant", "content": "answer", "reasoning_content": "prior thought"},
+ {"role": "user", "content": "two"},
+ ]
+ body = _route_capture(
+ provider_type = "llama_cpp", messages = messages, full_body = True, preserve_thinking = value
+ )
+ assert body["messages"][-2]["reasoning_content"] == "prior thought"
+ if value is None:
+ assert "chat_template_kwargs" not in body
+ else:
+ assert body["chat_template_kwargs"] == {"preserve_thinking": value}
diff --git a/studio/backend/tests/test_external_tool_transport_cancel.py b/studio/backend/tests/test_external_tool_transport_cancel.py
index 7776e28e357..967d0e8f9da 100644
--- a/studio/backend/tests/test_external_tool_transport_cancel.py
+++ b/studio/backend/tests/test_external_tool_transport_cancel.py
@@ -18,6 +18,8 @@ from core.inference.external_tool_transport import OAICompatTransport
class _StallingClient:
"""One chunk, then silence: an upstream mid-answer or still in prefill."""
+ provider_type = "custom"
+
def __init__(self) -> None:
self.torn_down = False
self.released = asyncio.Event()
diff --git a/studio/backend/tests/test_studio_tool_loop.py b/studio/backend/tests/test_studio_tool_loop.py
index 4fa391ea649..014d664d1d7 100644
--- a/studio/backend/tests/test_studio_tool_loop.py
+++ b/studio/backend/tests/test_studio_tool_loop.py
@@ -1337,3 +1337,39 @@ def test_mcp_images_are_not_sent_to_a_text_only_provider(mcp_image_result):
follow_up = transport.requests[1]["messages"]
assert follow_up[-1]["role"] == "tool"
assert "__MCP_IMAGES__" not in follow_up[-1]["content"]
+
+
[email protected]("preserve", [True, False])
+def test_reasoning_replay_during_tool_rounds_is_opt_in(executed, preserve):
+ transport = FakeTransport(
+ [
+ [
+ _sse({"reasoning_content": "First "}),
+ _sse({"reasoning_content": "thought."}),
+ _sse(
+ {
+ "tool_calls": [
+ {
+ "index": 0,
+ "id": "call_1",
+ "type": "function",
+ "function": {
+ "name": "web_search",
+ "arguments": '{"query":"unsloth"}',
+ },
+ }
+ ]
+ }
+ ),
+ _sse(finish = "tool_calls"),
+ _DONE,
+ ],
+ [_sse({"content": "Done."}), _sse(finish = "stop"), _DONE],
+ ]
+ )
+ transport.preserves_reasoning = preserve
+ _run(transport)
+ assistant = transport.requests[1]["messages"][-2]
+ assert assistant.get("reasoning_content") == ("First thought." if preserve else None)
+ assert assistant["tool_calls"][0]["function"]["name"] == "web_search"
+ assert transport.requests[1]["messages"][-1]["role"] == "tool"
diff --git a/studio/frontend/src/components/assistant-ui/thread.tsx b/studio/frontend/src/components/assistant-ui/thread.tsx
index 86c069a553b..27d1b337479 100644
--- a/studio/frontend/src/components/assistant-ui/thread.tsx
+++ b/studio/frontend/src/components/assistant-ui/thread.tsx
@@ -5805,7 +5805,8 @@ const ReasoningToggle: FC<{ side?: "top" | "bottom" }> = ({
: reasoningEffort;
const effectiveReasoningVisualEnabled =
effectiveReasoningEnabled && displayedEffort !== "none";
- const disabled = !(modelLoaded && effectiveSupportsReasoning);
+ const disabled =
+ !modelLoaded || !(effectiveSupportsReasoning || supportsPreserveThinking);
const formatEffortLabel = (level: typeof reasoningEffort): string => {
if (level !== "xhigh")
return level.charAt(0).toUpperCase() + level.slice(1);
@@ -5820,8 +5821,8 @@ const ReasoningToggle: FC<{ side?: "top" | "bottom" }> = ({
};
const effortLabel = formatEffortLabel(displayedEffort);
- // Only rendered for models that can reason.
- if (!effectiveSupportsReasoning) {
+ // A connection may support history preservation without a generation toggle.
+ if (!effectiveSupportsReasoning && !supportsPreserveThinking) {
return null;
}
@@ -5832,9 +5833,11 @@ const ReasoningToggle: FC<{ side?: "top" | "bottom" }> = ({
effectiveReasoningStyle === "enable_thinking_effort";
// Dropdown when there are effort levels or preserve-thinking; else a toggle.
const useDropdown = isEffort || supportsPreserveThinking;
- const activeLook = isEffort
- ? reasoningLockedOn || (effectiveReasoningVisualEnabled && !disabled)
- : reasoningLockedOn || (effectiveReasoningEnabled && !disabled);
+ const activeLook = !effectiveSupportsReasoning
+ ? preserveThinking && !disabled
+ : isEffort
+ ? reasoningLockedOn || (effectiveReasoningVisualEnabled && !disabled)
+ : reasoningLockedOn || (effectiveReasoningEnabled && !disabled);
if (useDropdown) {
return (
@@ -5927,6 +5930,7 @@ const ReasoningToggle: FC<{ side?: "top" | "bottom" }> = ({
))}
</>
) : (
+ effectiveSupportsReasoning &&
effectiveSupportsReasoningOff &&
!reasoningLockedOn && (
<DropdownMenuItem
@@ -5960,8 +5964,8 @@ const ReasoningToggle: FC<{ side?: "top" | "bottom" }> = ({
e.preventDefault();
const next = !preserveThinking;
setPreserveThinking(next);
- // Preserve thinking requires thinking on.
- if (next) {
+ // Only local models couple this setting to generation controls.
+ if (next && externalSelection === null) {
setReasoningEnabled(true);
applyQwenThinkingParams(true);
}
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index 43b988d7c46..43efec3ebd7 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -186,6 +186,7 @@ import {
getGroundedExternalMaxOutputTokens,
getExternalMinOutputTokens,
getExternalReasoningCapabilities,
+ providerSupportsPreserveThinking,
getProviderCapabilities,
isGeminiCustomOpenAICompatBase,
providerHostsCodeExecution,
@@ -1502,8 +1503,8 @@ function isAbandonedAssistantTurn(
): boolean {
if (message.role !== "assistant") return false;
if (assistantTurnCarriesPayload(message)) return false;
- // A turn that finished on reasoning alone is a reply, even though external requests strip
- // reasoning and serialise it empty.
+ // A turn that finished on reasoning alone is a reply, even when the selected provider
+ // omits reasoning and serialises it empty.
if (
!assistantTurnEndedEarly(message) &&
(message.content ?? []).some((part) => part.type === "reasoning")
@@ -4952,14 +4953,15 @@ export function createOpenAIStreamAdapter(
readsImages: targetReadsImages,
localMarkers: mcpImagesLocalMarkers,
});
- const survivingMessages = pruneOutboundHistory(
- messages,
- !isExternalRequest,
+ const replayReasoning = !isExternalRequest || (
+ providerSupportsPreserveThinking(externalProvider?.providerType) &&
+ runtime.preserveThinking
);
+ const survivingMessages = pruneOutboundHistory(messages, replayReasoning);
// toOpenAIMessages emits assistant tool_calls plus role="tool" follow-ups; the backend Gemini
// translator rebuilds the functionCall/functionResponse parts.
let outboundMessages = survivingMessages
- .flatMap((message) => toOpenAIMessages(message, !isExternalRequest))
+ .flatMap((message) => toOpenAIMessages(message, replayReasoning))
.filter((message): message is NonNullable<typeof message> =>
Boolean(message),
);
@@ -6226,6 +6228,9 @@ export function createOpenAIStreamAdapter(
return {
model: externalSelection.modelId,
messages: outboundMessages,
+ ...(providerSupportsPreserveThinking(externalProvider?.providerType)
+ ? { preserve_thinking: runtime.preserveThinking }
+ : {}),
stream: true,
stream_options: { include_usage: true },
// Never forwarded upstream (the proxy sends an explicit field list); the trailing assistant
diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index 445b13a6f2d..5ffe17fcb6b 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -179,6 +179,7 @@ import {
import {
externalReasoningTakesEffort,
getExternalReasoningCapabilities,
+ providerSupportsPreserveThinking,
getProviderCapabilities,
modelCatalogVersion,
providerHostsCodeExecution,
@@ -222,6 +223,7 @@ import {
reconcilePinnedReasoningEffort,
takeEffortDisplacedByPin,
threadScopedOverride,
+ resolvePreserveThinkingOnLoad,
useChatRuntimeStore,
} from "./stores/chat-runtime-store";
import { wantsDownloadManagerStaging } from "./utils/model-download-staging";
@@ -2630,7 +2632,10 @@ export function ChatPage({
: state.reasoningEnabled
: true
: state.reasoningEnabled,
- supportsPreserveThinking: false,
+ supportsPreserveThinking: providerSupportsPreserveThinking(provider?.providerType),
+ preserveThinking: resolvePreserveThinkingOnLoad({
+ supports_preserve_thinking: providerSupportsPreserveThinking(provider?.providerType),
+ }),
supportsTools: supportsStudioToolsHere,
supportsBuiltinWebSearch,
supportsBuiltinCodeExecution,
@@ -3327,7 +3332,10 @@ export function ChatPage({
: store.reasoningEnabled
: true
: store.reasoningEnabled,
- supportsPreserveThinking: false,
+ supportsPreserveThinking: providerSupportsPreserveThinking(selectedProvider?.providerType),
+ preserveThinking: resolvePreserveThinkingOnLoad({
+ supports_preserve_thinking: providerSupportsPreserveThinking(selectedProvider?.providerType),
+ }),
supportsTools: supportsStudioToolsHere,
supportsBuiltinWebSearch,
supportsBuiltinCodeExecution,
diff --git a/studio/frontend/src/features/chat/provider-capabilities.ts b/studio/frontend/src/features/chat/provider-capabilities.ts
index a5f4f055426..79171e81608 100644
--- a/studio/frontend/src/features/chat/provider-capabilities.ts
+++ b/studio/frontend/src/features/chat/provider-capabilities.ts
@@ -1387,3 +1387,10 @@ export function reasoningFieldsAfterCatalogRefresh(
caps.supportsReasoning && !caps.supportsReasoningOff ? true : current.reasoningEnabled,
};
}
+
+/** Sent as a llama-server template kwarg, so the model's template must still read it. */
+export function providerSupportsPreserveThinking(
+ providerType: string | null | undefined,
+): boolean {
+ return providerType === "llama_cpp";
+}
diff --git a/studio/frontend/src/features/chat/shared-composer.tsx b/studio/frontend/src/features/chat/shared-composer.tsx
index d914dcceaa2..708fd3e9eb0 100644
--- a/studio/frontend/src/features/chat/shared-composer.tsx
+++ b/studio/frontend/src/features/chat/shared-composer.tsx
@@ -797,9 +797,12 @@ export function SharedComposer({
: reasoningEffort;
const effectiveReasoningVisualEnabled =
effectiveReasoningEnabled && displayedEffort !== "none";
- const reasoningDisabled = !modelLoaded || !effectiveSupportsReasoning;
+ const reasoningDisabled =
+ !modelLoaded || !(effectiveSupportsReasoning || supportsPreserveThinking);
const showReasoningControl =
- effectiveSupportsReasoning || effectiveReasoningAlwaysOn;
+ effectiveSupportsReasoning ||
+ effectiveReasoningAlwaysOn ||
+ supportsPreserveThinking;
// enable_thinking_effort (GLM-5.2: high|max + disable) reuses the effort dropdown; it just also
// carries an Off row via supportsReasoningOff.
const isEffort =
@@ -810,9 +813,11 @@ export function SharedComposer({
const narrowEffortMenu =
effectiveReasoningStyle === "enable_thinking_effort" &&
!supportsPreserveThinking;
- const thinkingActiveLook = isEffort
- ? reasoningLockedOn || (effectiveReasoningVisualEnabled && !reasoningDisabled)
- : reasoningLockedOn || (effectiveReasoningEnabled && !reasoningDisabled);
+ const thinkingActiveLook = !effectiveSupportsReasoning
+ ? preserveThinking && !reasoningDisabled
+ : isEffort
+ ? reasoningLockedOn || (effectiveReasoningVisualEnabled && !reasoningDisabled)
+ : reasoningLockedOn || (effectiveReasoningEnabled && !reasoningDisabled);
// Search can use Unsloth tools or provider web search independently of Code.
// Code follows the provider's sandbox placement and never falls back to local
// execution when a hosted model lacks code support.
@@ -2779,6 +2784,7 @@ export function SharedComposer({
))}
</>
) : (
+ effectiveSupportsReasoning &&
effectiveSupportsReasoningOff &&
!reasoningLockedOn && (
<DropdownMenuItem
@@ -2812,8 +2818,8 @@ export function SharedComposer({
e.preventDefault();
const next = !preserveThinking;
setPreserveThinking(next);
- // Preserve thinking requires thinking on.
- if (next) {
+ // Only local models couple this setting to generation controls.
+ if (next && !isExternalModel) {
setReasoningEnabled(true);
applyQwenThinkingParams(true);
}
diff --git a/studio/frontend/tests/cancelled-turn-history-prune.test.ts b/studio/frontend/tests/cancelled-turn-history-prune.test.ts
index 7efca12844d..41b9a1c5217 100644
--- a/studio/frontend/tests/cancelled-turn-history-prune.test.ts
+++ b/studio/frontend/tests/cancelled-turn-history-prune.test.ts
@@ -142,6 +142,6 @@ test("the send and token-count paths prune through the same helper", () => {
);
assert.match(
adapter,
- /const survivingMessages = pruneOutboundHistory\(\s*messages,\s*!isExternalRequest,\s*\);/,
+ /const survivingMessages = pruneOutboundHistory\(\s*messages,\s*replayReasoning\);/,
);
});
diff --git a/studio/frontend/tests/external-preserve-thinking.test.ts b/studio/frontend/tests/external-preserve-thinking.test.ts
new file mode 100644
index 00000000000..9bfcc988ada
--- /dev/null
+++ b/studio/frontend/tests/external-preserve-thinking.test.ts
@@ -0,0 +1,48 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import assert from "node:assert/strict";
+import { readFileSync } from "node:fs";
+import test from "node:test";
+import ts from "typescript";
+import { registerBundlerResolver } from "./helpers/kit.ts";
+
+registerBundlerResolver();
+const { providerSupportsPreserveThinking } = await import("../src/features/chat/provider-capabilities.ts");
+const source = readFileSync(new URL("../src/features/chat/api/chat-adapter.ts", import.meta.url), "utf8");
+const tree = ts.createSourceFile("adapter.ts", source, ts.ScriptTarget.Latest, true);
+let replayExpression = "";
+let payloadExpression = "";
+function visit(node: ts.Node): void {
+ if (ts.isVariableDeclaration(node) && node.name.getText(tree) === "replayReasoning") {
+ replayExpression = node.initializer!.getText(tree);
+ }
+ if (ts.isSpreadAssignment(node) && node.getText(tree).includes("providerSupportsPreserveThinking")) {
+ payloadExpression = node.expression.getText(tree);
+ }
+ ts.forEachChild(node, visit);
+}
+visit(tree);
+assert.ok(replayExpression && payloadExpression);
+// Evaluates the real send-path expressions rather than a copied policy.
+const evaluate = new Function("providerSupportsPreserveThinking", "externalProvider", "runtime", "isExternalRequest",
+ `return { replay: ${replayExpression}, fields: ${payloadExpression} };`) as
+ (supports: typeof providerSupportsPreserveThinking, provider: {providerType: string}, runtime: {preserveThinking: boolean}, external: boolean) =>
+ {replay: boolean; fields: {preserve_thinking?: boolean}};
+
+test("llama.cpp sends the selected value and replays reasoning only when enabled", () => {
+ for (const enabled of [true, false]) {
+ const result = evaluate(providerSupportsPreserveThinking, {providerType: "llama_cpp"}, {preserveThinking: enabled}, true);
+ assert.deepEqual(result, {replay: enabled, fields: {preserve_thinking: enabled}});
+ }
+ assert.match(source, /pruneOutboundHistory\(messages, replayReasoning\)/);
+ assert.match(source, /toOpenAIMessages\(message, replayReasoning\)/);
+});
+
+test("switching providers never leaks a retained enabled preference", () => {
+ for (const providerType of ["custom", "openai", "vllm", "ollama", "anthropic"]) {
+ assert.deepEqual(evaluate(providerSupportsPreserveThinking, {providerType}, {preserveThinking: true}, true),
+ {replay: false, fields: {}});
+ }
+ assert.equal(evaluate(providerSupportsPreserveThinking, {providerType: ""}, {preserveThinking: false}, false).replay, true);
+});
diff --git a/studio/frontend/tests/mcp-images.test.ts b/studio/frontend/tests/mcp-images.test.ts
index 3d3b8216de1..6dc6d7ad8d2 100644
--- a/studio/frontend/tests/mcp-images.test.ts
+++ b/studio/frontend/tests/mcp-images.test.ts
@@ -424,7 +424,7 @@ test("the send path bounds the run's own results before serializing them", () =>
);
assert.match(
adapter,
- /const messages = boundMcpImageResults\(rawMessages, \{\n\s*readsImages: targetReadsImages,\n\s*localMarkers: mcpImagesLocalMarkers,\n\s*\}\);\n\s*const survivingMessages = pruneOutboundHistory\(\n\s*messages,/,
+ /const messages = boundMcpImageResults\(rawMessages, \{\n\s*readsImages: targetReadsImages,\n\s*localMarkers: mcpImagesLocalMarkers,\n\s*\}\);\n\s*const replayReasoning = [\s\S]*?;\n\s*const survivingMessages = pruneOutboundHistory\(messages, replayReasoning\);/,
);
assert.doesNotMatch(adapter, /boundMcpImageEnvelopes\(outboundMessages/);
assert.doesNotMatch(adapter, /stripMcpImageEnvelopes\(outboundMessages/);
diff --git a/studio/frontend/tests/provider-capabilities-current-models.test.ts b/studio/frontend/tests/provider-capabilities-current-models.test.ts
index 5e18bed1bcc..78992561b23 100644
--- a/studio/frontend/tests/provider-capabilities-current-models.test.ts
+++ b/studio/frontend/tests/provider-capabilities-current-models.test.ts
@@ -21,6 +21,7 @@ const {
providerSupportsBuiltinWebSearch,
providerSupportsFastMode,
+ providerSupportsPreserveThinking,
} = await import("../src/features/chat/provider-capabilities.ts");
const {
@@ -508,3 +509,11 @@ test("hosted Code is only claimed where the backend registry hosts code_executio
}
}
});
+
+
+test("preserve thinking is available only for explicit llama.cpp connections", () => {
+ assert.equal(providerSupportsPreserveThinking("llama_cpp"), true);
+ for (const provider of [undefined, null, "custom", "vllm", "ollama", "openai", "anthropic", "openrouter"]) {
+ assert.equal(providerSupportsPreserveThinking(provider), false);
+ }
+});
diff --git a/tests/studio/test_cancelled_turn_history_prune.py b/tests/studio/test_cancelled_turn_history_prune.py
index 74999645803..c2ce5e46d48 100644
--- a/tests/studio/test_cancelled_turn_history_prune.py
+++ b/tests/studio/test_cancelled_turn_history_prune.py
@@ -92,13 +92,13 @@ def _send_path_slice() -> str:
"""
body = slice_between(
read(ADAPTER),
- " const survivingMessages = pruneOutboundHistory(\n"
- " messages,\n"
- " !isExternalRequest,\n",
+ " const survivingMessages = pruneOutboundHistory(messages, replayReasoning);",
"if (selectedImageEditReference) {",
)
return (
"export function buildSendPathOutbound(messages: any, isExternalRequest: boolean) {\n"
+ # The provider-dependent flag is covered by external-preserve-thinking.test.ts.
+ + " const replayReasoning = !isExternalRequest;\n"
+ body
+ " return outboundMessages;\n}\n"
)