AIX: DeepSeek V4: fix function calling

Also see: googleapis/python-genai#1971

src/apps/chat/AppChat.tsx

@@ -583,9 +583,11 @@ export function AppChat() {
   }, []);
 
   useGlobalShortcuts('AppChat', React.useMemo(() => [
-    // focused conversation
+    // focused conversation (excluded when Beam is open so the keystroke passes through to the browser)
-    { key: 'z', ctrl: true, shift: true, disabled: isFocusedChatEmpty, action: handleMessageRegenerateLastInFocusedPane, description: 'Retry' },
+    ...(beamOpenStoreInFocusedPane ? [] : [
-    { key: 'b', ctrl: true, shift: true, disabled: isFocusedChatEmpty, action: handleMessageBeamLastInFocusedPane, description: 'Beam Edit' },
+      { key: 'z', ctrl: true, shift: true, disabled: isFocusedChatEmpty, action: handleMessageRegenerateLastInFocusedPane, description: 'Retry' },
+      { key: 'b', ctrl: true, shift: true, disabled: isFocusedChatEmpty, action: handleMessageBeamLastInFocusedPane, description: 'Beam Edit' },
+    ]),
     { key: 'o', ctrl: true, action: handleConversationsImportFormFilePicker },
     { key: 's', ctrl: true, action: () => handleFileSaveConversation(focusedPaneConversationId) },
     { key: 'n', ctrl: true, shift: true, action: () => handleConversationNewInFocusedPane(false, false) },
@@ -603,7 +605,7 @@ export function AppChat() {
     { key: 'p', ctrl: true, action: () => personaDropdownRef.current?.openListbox() /*, description: 'Open Persona Dropdown'*/ },
     // focused conversation llm
     { key: 'o', ctrl: true, shift: true, action: handleOpenChatLlmOptions },
-  ], [focusedPaneConversationId, handleConversationNewInFocusedPane, handleConversationReset, handleConversationsImportFormFilePicker, handleDeleteConversations, handleFileSaveConversation, handleMessageBeamLastInFocusedPane, handleMessageRegenerateLastInFocusedPane, handleMoveFocus, handleNavigateHistoryInFocusedPane, handleOpenChatLlmOptions, isFocusedChatEmpty]));
+  ], [beamOpenStoreInFocusedPane, focusedPaneConversationId, handleConversationNewInFocusedPane, handleConversationReset, handleConversationsImportFormFilePicker, handleDeleteConversations, handleFileSaveConversation, handleMessageBeamLastInFocusedPane, handleMessageRegenerateLastInFocusedPane, handleMoveFocus, handleNavigateHistoryInFocusedPane, handleOpenChatLlmOptions, isFocusedChatEmpty]));
 
 
   return <>

src/apps/chat/components/message/BlockOpUpstreamResume.tsx

@@ -16,6 +16,7 @@ const ARM_TIMEOUT_MS = 4000;
  */
 export function BlockOpUpstreamResume(props: {
   upstreamHandle: Exclude<DMessageGenerator['upstreamHandle'], undefined>,
+  pending?: boolean; // true while the message is actively streaming; labels the Delete button as "Stop"
   onResume?: () => void | Promise<void>;
   onCancel?: () => void | Promise<void>;
   onDelete?: () => void | Promise<void>;
@@ -30,8 +31,8 @@ export function BlockOpUpstreamResume(props: {
 
   // expiration: boolean is evaluated at render (may lag briefly if nothing re-renders past expiry).
   // TimeAgo handles its own tick for the label; the button's disabled state is the only consumer of this flag.
-  const { expiresAt, runId = '' } = props.upstreamHandle;
+  const { expiresAt /*, runId = ''*/ } = props.upstreamHandle;
-  const isExpired = expiresAt != null && Date.now() > expiresAt;
+  // const isExpired = expiresAt != null && Date.now() > expiresAt;
 
   // handlers
 
@@ -102,7 +103,7 @@ export function BlockOpUpstreamResume(props: {
         {props.onResume && (
           <Tooltip title='Resume generation from last checkpoint'>
             <Button
-              disabled={isResuming || isCancelling || isDeleting || isExpired}
+              disabled={isResuming || isCancelling || isDeleting}
               loading={isResuming}
               startDecorator={<PlayArrowRoundedIcon color='success' />}
               onClick={handleResume}
@@ -126,7 +127,7 @@ export function BlockOpUpstreamResume(props: {
         )}
 
         {props.onDelete && (
-          <Tooltip title={deleteArmed ? 'Click again to confirm - cancels the run upstream (no resume after)' : 'Cancel the upstream run'}>
+          <Tooltip title={deleteArmed ? 'Click again to confirm - cancels the run upstream (no resume after)' : (props.pending ? 'Stop this response and cancel the upstream run' : 'Cancel the upstream run')}>
             <Button
               loading={isDeleting}
               color={deleteArmed ? 'danger' : 'neutral'}
@@ -135,7 +136,7 @@ export function BlockOpUpstreamResume(props: {
               onClick={handleDelete}
               disabled={isResuming || isCancelling || isDeleting}
             >
-              {deleteArmed ? 'Confirm?' : 'Cancel'}
+              {deleteArmed ? 'Confirm?' : (props.pending ? 'Stop' : 'Cancel')}
             </Button>
           </Tooltip>
         )}

src/apps/chat/components/message/ChatMessage.tsx

@@ -898,11 +898,12 @@ export function ChatMessage(props: {
             />
           )}
 
-          {/* Upstream Resume - shows whenever there's a stored handle (incl. post-reload, where no error fragment is present) */}
+          {/* Upstream Resume - shows whenever there's a stored handle (incl. post-reload, and while streaming so Stop can cancel the upstream run) */}
-          {!messagePendingIncomplete && props.isBottom && fromAssistant && messageGenerator?.upstreamHandle && (!!onMessageUpstreamResume || !!onMessageUpstreamDelete) && (
+          {props.isBottom && fromAssistant && messageGenerator?.upstreamHandle && (!!onMessageUpstreamResume || !!onMessageUpstreamDelete) && (
             <BlockOpUpstreamResume
               upstreamHandle={messageGenerator.upstreamHandle}
-              onResume={onMessageUpstreamResume ? handleUpstreamResume : undefined}
+              pending={messagePendingIncomplete}
+              onResume={(!messagePendingIncomplete && onMessageUpstreamResume) ? handleUpstreamResume : undefined}
               onDelete={onMessageUpstreamDelete ? handleUpstreamDelete : undefined}
             />
           )}

src/apps/settings-modal/AppChatSettingsAI.tsx

@@ -166,9 +166,9 @@ export function AppChatSettingsAI() {
       tooltip={<>
         When Claude uses tools like code execution, it may produce text and image files stored in Anthropic&apos;s File API. This setting controls whether Big-AGI should automatically download and embed them in the chat.
         <ul>
-          <li><b>Off</b>: keep as references (default).</li>
+          <li><b>Show</b>: keep as references.</li>
-          <li><b>Inline</b>: download and embed text/images.</li>
+          <li><b>Embed</b>: download and embed text/images (default).</li>
-          <li><b>Inline + Free</b>: embed, then delete from Anthropic to free storage.</li>
+          <li><b>Embed + Free</b>: embed, then delete from Anthropic to free storage.</li>
         </ul>
         Only affects Anthropic models.
       </>}

src/common/app.release.ts

@@ -23,7 +23,7 @@ export const Release = {
 
   // this is here to trigger revalidation of data, e.g. models refresh
   Monotonics: {
-    Aix: 67,
+    Aix: 68,
     NewsVersion: 204,
   },
 

src/common/chat-overlay/ConversationHandler.ts

@@ -5,6 +5,7 @@ import { bareBonesPromptMixer } from '~/modules/persona/pmix/pmix';
 import { SystemPurposes } from '../../data';
 
 import { BeamStore, createBeamVanillaStore } from '~/modules/beam/store-beam_vanilla';
+import { autoConversationTitle } from '~/modules/aifn/autotitle/autoTitle';
 import { useModuleBeamStore } from '~/modules/beam/store-module-beam';
 
 import type { DConversationId } from '~/common/stores/chat/chat.conversation';
@@ -275,6 +276,10 @@ export class ConversationHandler {
 
       // close beam
       terminateKeepingSettings();
+
+      // auto-title the conversation if enabled (parity with chat-persona flow — fixes #1078)
+      if (getChatAutoAI().autoTitleChat)
+        void autoConversationTitle(this.conversationId, false);
     };
 
     beamOpen(viewHistory, getChatLLMId(), !!destReplaceMessageId, onBeamSuccess);

src/common/stores/llms/llms.parameters.ts

@@ -175,7 +175,8 @@ export const DModelParameterRegistry = {
     label: 'Thinking',
     type: 'enum',
     description: 'Enable or disable extended thinking mode.',
-    values: ['none', 'high'],
+    values: ['none', 'high', 'max'],
+    // 'max' is for now DeepSeek V4-specific (reasoning_effort=max); other vendors restrict via enumValues
     // undefined means vendor default (usually 'high', i.e. thinking enabled)
   }),
 

src/modules/aifn/autotitle/autoTitle.ts

@@ -49,7 +49,7 @@ export async function autoConversationTitle(conversationId: string, forceReplace
       autoTitleLlmId,
       'You are an AI conversation titles assistant who specializes in creating expressive yet few-words chat titles.',
       `Analyze the given short conversation (every line is truncated) and extract a concise chat title that summarizes the conversation in as little as a couple of words.
-Only respond with the lowercase short title and nothing else.
+Only respond with the short title and nothing else.
 
 \`\`\`
 ${historyLines.join('\n')}

src/modules/aix/client/aix.client.chatGenerateRequest.ts

@@ -409,11 +409,15 @@ export async function aixCGR_ChatSequence_FromDMessagesOrThrow(
             break;
 
           case 'ma':
-            // Preserve reasoning continuity across turns. Two channels, any one is sufficient:
+            // Preserve reasoning continuity across turns. Three channels, any one is sufficient:
             // - Anthropic: part.textSignature / part.redactedData (bespoke fields, see Anthropic extended thinking docs)
-            // - OpenAI/Gemini: _vnd sidecar (reasoningItem.* / thoughtSignature, generic vendor-state mechanism)
+            // - OpenAI Responses / Gemini: _vnd sidecar (reasoningItem.* / thoughtSignature, opaque continuity handle)
+            // - DeepSeek V4 (OpenAI chat-completions): plain reasoning text in aText is the payload itself
             const oaiReasoning = _vnd?.openai?.reasoningItem;
-            const hasReasoningHandle = aPart.textSignature || aPart.redactedData?.length || oaiReasoning?.encryptedContent || oaiReasoning?.id;
+            const hasReasoningHandle =
+              (aPart.textSignature || aPart.redactedData?.length)
+              || (oaiReasoning?.encryptedContent || oaiReasoning?.id)
+              || (aPart.aText && aPart.aType === 'reasoning'); // DeepSeek V4 reasoning in plain text - NOTE: will send LOTS of 'ma' parts (e.g. to Gemini, which doesn't even need them)
             if (hasReasoningHandle) {
               const aModelAuxPart = aPart as AixParts_ModelAuxPart; // NOTE: this is a forced cast from readonly string[] to string[], but not a big deal here
               modelMessage.parts.push(_vnd ? { ...aModelAuxPart, _vnd } : aModelAuxPart);
@@ -653,7 +657,7 @@ function _clientCreateAixMetaInReferenceToPart(items: DMetaReferenceItem[]): Aix
 
 
 export async function clientHotFixGenerateRequest_ApplyAll(llmInterfaces: DLLM['interfaces'], aixChatGenerate: AixAPIChatGenerate_Request, modelName: string): Promise<{
-  shallDisableStreaming: boolean;
+  hotfixNoStream: boolean;
   workaroundsCount: number;
 }> {
 
@@ -676,12 +680,12 @@ export async function clientHotFixGenerateRequest_ApplyAll(llmInterfaces: DLLM['
     workaroundsCount += await clientHotFixGenerateRequest_ConvertWebP(aixChatGenerate, 'image/jpeg');
 
   // Disable streaming for select chat models that don't support it (e.g. o1-preview (old) and o1-2024-12-17)
-  const shallDisableStreaming = llmInterfaces.includes(LLM_IF_HOTFIX_NoStream);
+  const hotfixNoStream = llmInterfaces.includes(LLM_IF_HOTFIX_NoStream);
 
   if (workaroundsCount > 0)
     console.warn(`[DEV] Working around '${modelName}' model limitations: client-side applied ${workaroundsCount} workarounds`);
 
-  return { shallDisableStreaming, workaroundsCount };
+  return { hotfixNoStream, workaroundsCount };
 
 }
 

src/modules/aix/client/aix.client.direct-chatGenerate.ts

@@ -37,7 +37,7 @@ export async function* clientSideChatGenerate(
       return dispatch;
     });
 
-  yield* executeChatGenerateWithContinuation(dispatchCreator, streaming, abortSignal, _d);
+  yield* executeChatGenerateWithContinuation(dispatchCreator, abortSignal, _d);
 }
 
 /**
@@ -48,7 +48,7 @@ export async function* clientSideReattachUpstream(
   access: AixAPI_Access,
   resumeHandle: AixAPI_ResumeHandle,
   context: AixAPI_Context_ChatGenerate,
-  streaming: true,
+  streaming: boolean,
   connectionOptions: AixAPI_ConnectionOptions_ChatGenerate,
   abortSignal: AbortSignal,
 ): AsyncGenerator<AixWire_Particles.ChatGenerateOp, void> {
@@ -56,7 +56,7 @@ export async function* clientSideReattachUpstream(
   const _d: AixDebugObject = _createClientDebugConfig(access, connectionOptions, context.name);
   const dispatchCreator = () => createChatGenerateResumeDispatch(access, resumeHandle, streaming);
 
-  yield * executeChatGenerateWithContinuation(dispatchCreator, streaming, abortSignal, _d);
+  yield * executeChatGenerateWithContinuation(dispatchCreator, abortSignal, _d);
 }
 
 /**

src/modules/aix/client/aix.client.ts

@@ -342,7 +342,7 @@ export async function aixChatGenerateText_Simple(
   aixContextRef: AixAPI_Context_ChatGenerate['ref'],
   // optional options
   clientOptions?: Partial<AixClientOptions>, // this makes the abortController optional
-  // optional callback for streaming
+  // optional callback - if provided, streaming is activated
   onTextStreamUpdate?: (text: string, isDone: boolean, generator: DMessageGenerator) => MaybePromise<void>,
 ): Promise<string> {
 
@@ -363,14 +363,13 @@ export async function aixChatGenerateText_Simple(
   // Aix Context
   const aixContext = aixCreateChatGenerateContext(aixContextName, aixContextRef);
 
-  // Aix Streaming - implicit if the callback is provided
+  // Caller streaming preference - implicit: stream if a callback is provided
-  let aixStreaming = !!onTextStreamUpdate;
+  const callerStreaming = !!onTextStreamUpdate;
 
 
   // Client-side late stage model HotFixes
-  const { shallDisableStreaming } = await clientHotFixGenerateRequest_ApplyAll(llm.interfaces, aixChatGenerate, llmParameters.llmRef || llm.id);
+  const { hotfixNoStream } = await clientHotFixGenerateRequest_ApplyAll(llm.interfaces, aixChatGenerate, llmParameters.llmRef || llm.id);
-  if (shallDisableStreaming || aixModel.forceNoStream)
+  const wireStreaming = !hotfixNoStream && !aixModel.forceNoStream ? callerStreaming : false;
-    aixStreaming = false;
 
 
   // Variable to store the final text
@@ -398,11 +397,11 @@ export async function aixChatGenerateText_Simple(
     aixModel,
     aixChatGenerate,
     aixContext,
-    aixStreaming,
+    wireStreaming,
     state.generator,
     abortSignal,
     clientOptions?.throttleParallelThreads ?? 0,
-    !aixStreaming ? undefined : async (ll: AixChatGenerateContent_LL, _isDone: boolean /* we want to issue this, in case the next action is an exception */) => {
+    !onTextStreamUpdate ? undefined : async (ll: AixChatGenerateContent_LL, _isDone: boolean /* we want to issue this, in case the next action is an exception */) => {
       _llToL2Simple(ll, state);
       if (onTextStreamUpdate && state.text !== null)
         await onTextStreamUpdate(state.text, false, state.generator);
@@ -521,7 +520,7 @@ type _AixChatGenerateContent_DMessageGuts_WithOutcome = AixChatGenerateContent_D
  * @param llmId - ID of the Language Model to use
  * @param aixChatGenerate - Multi-modal chat generation request specifics, including Tools and high-level metadata
  * @param aixContext - Information about how this chat generation is being used
- * @param aixStreaming - Whether to use streaming for generation
+ * @param aixStreaming - Caller's wire-streaming preference. Subject to override by model/hotfix constraints, or dispatch constraints
  * @param clientOptions - Client options for the operation
  * @param onStreamingUpdate - Optional callback for streaming updates
  *
@@ -551,10 +550,9 @@ export async function aixChatGenerateContent_DMessage_orThrow<TServiceSettings e
     vndAntTransformInlineFiles: aixAccess.dialect === 'anthropic' ? getVndAntInlineFiles() : undefined,
   });
 
-  // Client-side late stage model HotFixes
+  // Client-side late stage model HotFixes - collapse the caller's requested streaming preference into the effective wire-streaming decision after constraints (hotfix gate, model.forceNoStream)
-  const { shallDisableStreaming } = await clientHotFixGenerateRequest_ApplyAll(llm.interfaces, aixChatGenerate, llmParameters.llmRef || llm.id);
+  const { hotfixNoStream } = await clientHotFixGenerateRequest_ApplyAll(llm.interfaces, aixChatGenerate, llmParameters.llmRef || llm.id);
-  if (shallDisableStreaming || aixModel.forceNoStream)
+  const wireStreaming = !hotfixNoStream && !aixModel.forceNoStream ? aixStreaming : false;
-    aixStreaming = false;
 
   // Legacy Note: awaited OpenAI moderation check was removed (was only on this codepath)
 
@@ -584,7 +582,7 @@ export async function aixChatGenerateContent_DMessage_orThrow<TServiceSettings e
     aixModel,
     aixChatGenerate,
     aixContext,
-    aixStreaming,
+    wireStreaming,
     dMessage.generator,
     clientOptions.abortSignal,
     clientOptions.throttleParallelThreads ?? 0,
@@ -753,7 +751,7 @@ export type AixChatGenerateTerminal_LL = 'completed' | 'aborted' | 'failed';
  *
  * Contract:
  * - empty fragments means no content yet, and no error
- * - aixStreaming hints the source, but can be respected or not
+ * - wireStreaming hints the wire transport (SSE vs single response), but can be respected or not by the dispatch (e.g. SSE-only APIs ignore a `false` value)
  *   - onReassemblyUpdate is optional, you can ignore the updates and await the final result
  * - errors become Error fragments, and they can be dialect-sent, dispatch-excepts, client-read issues or even user aborts
  *   - DOES NOT THROW, but the final accumulator may contain error fragments
@@ -772,7 +770,7 @@ export type AixChatGenerateTerminal_LL = 'completed' | 'aborted' | 'failed';
  *    - special parts include 'In Reference To' (a decorator of messages)
  *    - other special parts include the Anthropic Caching hints, on select message
  * @param aixContext specifies the scope of the caller, such as what's the high level objective of this call
- * @param aixStreaming requests the source to provide incremental updates
+ * @param wireStreaming the effective wire-level streaming decision (already collapsed from caller preference + model/hotfix constraints); drives tRPC `streaming` field and downstream dispatch body shape
  * @param initialGenerator generator initial value, which will be updated for every new piece of information received
  * @param abortSignal allows the caller to stop the operation
  * @param throttleParallelThreads allows the caller to limit the number of parallel threads
@@ -790,7 +788,7 @@ async function _aixChatGenerateContent_LL(
   aixModel: AixAPI_Model,
   aixChatGenerate: AixAPIChatGenerate_Request,
   aixContext: AixAPI_Context_ChatGenerate,
-  aixStreaming: boolean,
+  wireStreaming: boolean,
   // others
   initialGenerator: DMessageGenerator,
   abortSignal: AbortSignal,
@@ -804,10 +802,13 @@ async function _aixChatGenerateContent_LL(
   const inspectorTransport = !inspectorEnabled ? undefined : aixAccess.clientSideFetch ? 'csf' : 'trpc';
   const inspectorContext = !inspectorEnabled ? undefined : { contextName: aixContext.name, contextRef: aixContext.ref };
 
-  // [DEV] Inspector - request body override
+  // Inspector - override request body
   const requestBodyOverrideJson = inspectorEnabled && aixClientDebuggerGetRBO();
   const debugRequestBodyOverride = !requestBodyOverrideJson ? false : JSON.parse(requestBodyOverrideJson);
 
+  // Inspector - force disable streaming (note: dispatches may still override this)
+  if (getAixDebuggerNoStreaming()) wireStreaming = false;
+
   /**
    * FIXME: implement client selection of resumability - aixAccess option?
    * NOTE: for Gemini Deep Research, it's on by default, so both auto-reattach on network breaks (currently disabled)
@@ -827,8 +828,11 @@ async function _aixChatGenerateContent_LL(
   // [CSF] Pre-load client-side executor if needed - type inference works here, no need to type
   let clientSideChatGenerate;
   let clientSideReattachUpstream;
-  if (aixAccess.clientSideFetch)
+  if (aixAccess.clientSideFetch) {
-    ({ clientSideChatGenerate, clientSideReattachUpstream } = await _loadCsfModuleOrThrow());
+    const csf = await _loadCsfModuleOrThrow();
+    clientSideChatGenerate = csf.clientSideChatGenerate;
+    clientSideReattachUpstream = csf.clientSideReattachUpstream;
+  }
 
 
   // Client-side particle transforms:
@@ -891,7 +895,7 @@ async function _aixChatGenerateContent_LL(
           aixModel,
           aixChatGenerate,
           aixContext,
-          getAixDebuggerNoStreaming() ? false : aixStreaming,
+          wireStreaming,
           aixConnectionOptions,
           abortSignal,
         ) :
@@ -901,7 +905,7 @@ async function _aixChatGenerateContent_LL(
           model: aixModel,
           chatGenerate: aixChatGenerate,
           context: aixContext,
-          streaming: getAixDebuggerNoStreaming() ? false : aixStreaming, // [DEV] disable streaming if set in the UX (testing)
+          streaming: wireStreaming,
           connectionOptions: aixConnectionOptions,
         }, { signal: abortSignal })
 
@@ -912,7 +916,7 @@ async function _aixChatGenerateContent_LL(
           aixAccess,
           accumulator_LL.generator.upstreamHandle,
           aixContext,
-          true, // streaming - reattach is only validated for streaming for now
+          wireStreaming,
           aixConnectionOptions,
           abortSignal,
         ) :
@@ -921,7 +925,7 @@ async function _aixChatGenerateContent_LL(
           access: aixAccess,
           upstreamHandle: accumulator_LL.generator.upstreamHandle,
           context: aixContext,
-          streaming: true,
+          streaming: wireStreaming,
           connectionOptions: aixConnectionOptions,
         }, { signal: abortSignal })
 

src/modules/aix/server/api/aix.router.ts

@@ -30,7 +30,7 @@ export const aixRouter = createTRPCRouter({
       const _d = _createDebugConfig(input.access, input.connectionOptions, input.context.name);
       const dispatchCreator = () => createChatGenerateDispatch(input.access, input.model, input.chatGenerate, input.streaming, !!input.connectionOptions?.enableResumability);
 
-      yield* executeChatGenerateWithContinuation(dispatchCreator, input.streaming, ctx.reqSignal, _d);
+      yield* executeChatGenerateWithContinuation(dispatchCreator, ctx.reqSignal, _d);
     }),
 
   /**
@@ -42,14 +42,14 @@ export const aixRouter = createTRPCRouter({
       access: AixWire_API.Access_schema,
       upstreamHandle: AixWire_API.UpstreamHandle_schema, // reattach uses a handle instead of 'model + chatGenerate'
       context: AixWire_API.ContextChatGenerate_schema,
-      streaming: z.literal(true), // reattach is always streaming
+      streaming: z.boolean(),
       connectionOptions: AixWire_API.ConnectionOptionsChatGenerate_schema.pick({ debugDispatchRequest: true }).optional(), // debugDispatchRequest
     }))
     .mutation(async function* ({ input, ctx }) {
       const _d = _createDebugConfig(input.access, input.connectionOptions, input.context.name);
       const dispatchCreator = () => createChatGenerateResumeDispatch(input.access, input.upstreamHandle, input.streaming);
 
-      yield* executeChatGenerateWithContinuation(dispatchCreator, input.streaming, ctx.reqSignal, _d);
+      yield* executeChatGenerateWithContinuation(dispatchCreator, ctx.reqSignal, _d);
     }),
 
   /**

src/modules/aix/server/dispatch/chatGenerate/adapters/openai.chatCompletions.ts

@@ -37,6 +37,7 @@ export function aixToOpenAIChatCompletions(openAIDialect: OpenAIDialects, model:
   const chatGenerate = aixSpillSystemToUser(_chatGenerate);
 
   // Dialect incompatibilities -> Hotfixes
+  // [DeepSeek, 2026-04-24] V4 doesn't require strict alternation but we keep coalescing for cleanliness; the reducer only merges assistant/user, tool messages stay separate (parallel tool_calls).
   const hotFixAlternateUserAssistantRoles = openAIDialect === 'deepseek' || openAIDialect === 'perplexity';
   const hotFixRemoveEmptyMessages = openAIDialect === 'moonshot' || openAIDialect === 'perplexity'; // [Moonshot, 2026-02-10] consecutive assistant messages (empty + content) break Moonshot - coalesce to fix
   const hotFixRemoveStreamOptions = openAIDialect === 'azure' || openAIDialect === 'mistral';
@@ -59,7 +60,7 @@ export function aixToOpenAIChatCompletions(openAIDialect: OpenAIDialects, model:
     throw new Error('This service does not support function calls');
 
   // Convert the chat messages to the OpenAI 4-Messages format
-  let chatMessages = _toOpenAIMessages(chatGenerate.systemMessage, chatGenerate.chatSequence, hotFixOpenAIOFamily);
+  let chatMessages = _toOpenAIMessages(openAIDialect, chatGenerate.systemMessage, chatGenerate.chatSequence, hotFixOpenAIOFamily);
 
   // Apply hotfixes
 
@@ -69,6 +70,13 @@ export function aixToOpenAIChatCompletions(openAIDialect: OpenAIDialects, model:
   if (hotFixAlternateUserAssistantRoles)
     chatMessages = _fixAlternateUserAssistantRoles(chatMessages);
 
+  // [DeepSeek, 2026-04-24] When tools are present and thinking isn't disabled, V4 demands reasoning_content on EVERY assistant message in history
+  // Inject '' placeholder where missing; real reasoning is attached by _toOpenAIMessages
+  if (openAIDialect === 'deepseek' && chatGenerate.tools?.length)
+    for (const m of chatMessages)
+      if (m.role === 'assistant' && m.reasoning_content === undefined)
+        m.reasoning_content = '';
+
 
   // constrained output modes - both JSON and tool invocations
   // const strictJsonOutput = !!model.strictJsonOutput;
@@ -145,18 +153,23 @@ export function aixToOpenAIChatCompletions(openAIDialect: OpenAIDialects, model:
     && openAIDialect !== 'deepseek' && openAIDialect !== 'moonshot' && openAIDialect !== 'zai' // MoonShot maps to none->disabled / high->enabled
     && openAIDialect !== 'perplexity' // Perplexity has its own block below with stricter validation
   ) {
-    if (reasoningEffort === 'max') // domain validation
+    // for: 'alibaba' | 'azure' | 'groq' | 'lmstudio' | 'localai' | 'mistral' | 'openai' | 'openpipe' | 'togetherai' | 'xai'
-      throw new Error(`OpenAI ChatCompletions API does not support '${reasoningEffort}' reasoning effort`);
     payload.reasoning_effort = reasoningEffort;
   }
 
   // [Moonshot] Kimi K2.5 reasoning effort -> thinking mode (only 'none' and 'high' supported for now)
   // [Z.ai] GLM thinking mode: binary enabled/disabled (supports GLM-4.5 series and higher) - https://docs.z.ai/guides/capabilities/thinking-mode
+  // [DeepSeek, 2026-04-23] V4 thinking control https://api-docs.deepseek.com/guides/thinking_mode
   if (reasoningEffort && (openAIDialect === 'deepseek' || openAIDialect === 'moonshot' || openAIDialect === 'zai')) {
-    if (reasoningEffort !== 'none' && reasoningEffort !== 'high') // domain validation
+    const allowedEffort = openAIDialect === 'deepseek' ? ['none', 'high', 'max'] : ['none', 'high'];
-      throw new Error(`${openAIDialect} only supports reasoning effort 'none' or 'high', got '${reasoningEffort}'`);
+    if (!allowedEffort.includes(reasoningEffort)) // domain validation
+      throw new Error(`${openAIDialect} only supports reasoning effort ${allowedEffort.join(', ')}, got '${reasoningEffort}'`);
 
-    payload.thinking = { type: reasoningEffort === 'none' ? 'disabled' : 'enabled' };
+    payload.thinking = { type: reasoningEffort !== 'none' ? 'enabled' : 'disabled' };
+
+    // [DeepSeek, 2026-04-23] DeepSeek also supports effort control for reasoning-enabled requests - set it here as it was carved from the reasoningEffort setter before
+    if (openAIDialect === 'deepseek' && reasoningEffort !== 'none')
+      payload.reasoning_effort = reasoningEffort;
   }
 
 
@@ -348,19 +361,23 @@ function _fixAlternateUserAssistantRoles(chatMessages: TRequestMessages): TReque
       };
     }
 
-    // if the current item has the same role as the last item, concatenate their content
+    // If current item has the same role as the last, coalesce ONLY assistant/user.
+    // Tool/system/developer must stay separate - tool messages each pair with a tool_call_id; merging corrupts the protocol.
     if (acc.length > 0) {
       const lastItem = acc[acc.length - 1];
       if (lastItem.role === historyItem.role) {
         if (lastItem.role === 'assistant') {
           lastItem.content += hotFixSquashTextSeparator + historyItem.content;
-        } else if (lastItem.role === 'user') {
+          return acc;
+        }
+        if (lastItem.role === 'user') {
           lastItem.content = [
             ...(Array.isArray(lastItem.content) ? lastItem.content : [OpenAIWire_ContentParts.TextContentPart(lastItem.content)]),
             ...(Array.isArray(historyItem.content) ? historyItem.content : historyItem.content ? [OpenAIWire_ContentParts.TextContentPart(historyItem.content)] : []),
           ];
+          return acc;
         }
-        return acc;
+        // fall through to push for tool/system/developer - each stays its own message
       }
     }
 
@@ -442,7 +459,10 @@ function _fixVndOaiRestoreMarkdown_Inline(payload: TRequest) {
 }*/
 
 
-function _toOpenAIMessages(systemMessage: AixMessages_SystemMessage | null, chatSequence: AixMessages_ChatMessage[], hotFixOpenAIo1Family: boolean): TRequestMessages {
+function _toOpenAIMessages(openAIDialect: OpenAIDialects, systemMessage: AixMessages_SystemMessage | null, chatSequence: AixMessages_ChatMessage[], hotFixOpenAIo1Family: boolean): TRequestMessages {
+
+  // [DeepSeek, 2026-04-24] V4 thinking-by-default - reasoning_content must round-trip on tool-call turns; payload is the 'ma' part's aText (unlike Gemini/OpenAI-Responses which carry opaque handles).
+  const echoDeepseekReasoning = openAIDialect === 'deepseek';
 
   // Transform the chat messages into OpenAI's format (an array of 'system', 'user', 'assistant', and 'tool' messages)
   const chatMessages: TRequestMessages = [];
@@ -555,6 +575,8 @@ function _toOpenAIMessages(systemMessage: AixMessages_SystemMessage | null, chat
         break;
 
       case 'model':
+        // Accumulate 'ma' reasoning text across this turn; echoed below onto the assistant message if it carries tool_calls (DeepSeek only).
+        let pendingReasoningText = '';
         for (const part of parts) {
           const currentMessage = chatMessages[chatMessages.length - 1];
           switch (part.pt) {
@@ -630,7 +652,9 @@ function _toOpenAIMessages(systemMessage: AixMessages_SystemMessage | null, chat
               break;
 
             case 'ma':
-              // ignore this thinking block - Anthropic only
+              // [DeepSeek only] accumulate reasoning text for the echo-back below. Other dialects ignore 'ma' (reasoning continuity flows via _vnd opaque handles, not via this adapter).
+              if (echoDeepseekReasoning && part.aType === 'reasoning' && part.aText)
+                pendingReasoningText += part.aText;
               break;
 
             case 'tool_response':
@@ -651,6 +675,18 @@ function _toOpenAIMessages(systemMessage: AixMessages_SystemMessage | null, chat
           }
 
         }
+
+        // [DeepSeek] attach accumulated reasoning to this turn's assistant message only if it carries tool_calls; plain-text turns don't need the echo per docs.
+        if (echoDeepseekReasoning && pendingReasoningText) {
+          for (let i = chatMessages.length - 1; i >= 0; i--) {
+            const m = chatMessages[i];
+            if (m.role !== 'assistant') continue;
+            if (m.tool_calls?.length)
+              m.reasoning_content = pendingReasoningText;
+            break; // stop at the most recent assistant message from this turn
+          }
+        }
+
         break;
     }
   }

src/modules/aix/server/dispatch/chatGenerate/chatGenerate.continuation.ts

@@ -55,7 +55,6 @@ export class DispatchContinuationSignal extends Error {
  */
 export async function* executeChatGenerateWithContinuation(
   dispatchCreatorFn: () => Promise<ChatGenerateDispatch>,
-  streaming: boolean,
   abortSignal: AbortSignal,
   _d: AixDebugObject,
 ): AsyncGenerator<AixWire_Particles.ChatGenerateOp, void> {
@@ -65,7 +64,7 @@ export async function* executeChatGenerateWithContinuation(
   for (let turn = 0; turn <= MAX_CONTINUATION_TURNS; turn++) {
     try {
 
-      yield* executeChatGenerateWithOperationRetry(currentCreator, streaming, abortSignal, _d);
+      yield* executeChatGenerateWithOperationRetry(currentCreator, abortSignal, _d);
       return; // normal completion
 
     } catch (error) {

src/modules/aix/server/dispatch/chatGenerate/chatGenerate.dispatch.ts

@@ -25,7 +25,7 @@ import { createAnthropicFileInlineTransform } from './parsers/anthropic.transfor
 import { createAnthropicMessageParser, createAnthropicMessageParserNS } from './parsers/anthropic.parser';
 import { createBedrockConverseParserNS, createBedrockConverseStreamParser } from './parsers/bedrock-converse.parser';
 import { createGeminiGenerateContentResponseParser } from './parsers/gemini.parser';
-import { createGeminiInteractionsParser } from './parsers/gemini.interactions.parser';
+import { createGeminiInteractionsParserSSE } from './parsers/gemini.interactions.parser';
 import { createOpenAIChatCompletionsChunkParser, createOpenAIChatCompletionsParserNS } from './parsers/openai.parser';
 import { createOpenAIResponseParserNS, createOpenAIResponsesEventParser } from './parsers/openai.responses.parser';
 
@@ -37,7 +37,8 @@ export type ChatGenerateDispatch = {
   /** Used by dialects that need multi-step I/O. The returned response is consumed normally via demuxerFormat/chatGenerateParse */
   customConnect?: (signal: AbortSignal) => Promise<Response>;
   bodyTransform?: AixDemuxers.StreamBodyTransform;
-  demuxerFormat: AixDemuxers.StreamDemuxerFormat;
+  /** Source of truth for the consumer mode: null = NS */
+  demuxerFormat: null | AixDemuxers.StreamDemuxerFormat;
   chatGenerateParse: ChatGenerateParseFunction;
   particleTransform?: ChatGenerateParticleTransformFunction;
 };
@@ -173,6 +174,7 @@ export async function createChatGenerateDispatch(access: AixAPI_Access, model: A
 
       // [Gemini Interactions API - ALPHA TEST] SSE-native: POST with stream=true, upstream returns event-stream we pipe through the fast-sse demuxer.
       if (model.vndGeminiAPI === 'interactions-agent') {
+        if (!streaming) console.warn(`[DEV] Gemini Interactions API - only supported in SSE mode, ignoring streaming=false for model ${model.id}`);
         const request: ChatGenerateDispatchRequest = {
           ...geminiAccess(access, null, GeminiInteractionsWire_API_Interactions.postPath, false),
           method: 'POST',
@@ -186,8 +188,9 @@ export async function createChatGenerateDispatch(access: AixAPI_Access, model: A
               if (signal.aborted) throw error; // preserve abort identity for the executor's abort classifier
               throw new Error(`Gemini Interactions POST: ${error?.message || 'upstream error'}`); // rewrapping TRPCFetcherError as plain Error makes the retrier treat it as non-retryable
             }),
+          /** Upstream hardcodes stream=true + background=true (required by deep-research agents) and has no non-streaming alternative. */
           demuxerFormat: 'fast-sse',
-          chatGenerateParse: createGeminiInteractionsParser(requestedModelName),
+          chatGenerateParse: createGeminiInteractionsParserSSE(requestedModelName),
         };
       }
 
@@ -323,11 +326,13 @@ export async function createChatGenerateResumeDispatch(access: AixAPI_Access, re
       // [Gemini Interactions] Reattach via SSE stream - GET /interactions/{id}?stream=true replays all events from the start (intentional - client's ContentReassembler replaces message content on reattach; partial resume via last_event_id is deliberately NOT used).
       if (resumeHandle.uht !== 'vnd.gem.interactions')
         throw new Error(`Resume handle mismatch for gemini: expected 'vnd.gem.interactions', got '${resumeHandle.uht}'`);
+      if (!streaming) console.warn(`[DEV] Gemini Interactions API - Resume only supported in SSE mode, ignoring streaming=false for ${resumeHandle.runId}`);
       const { url: _baseUrl, headers: _headers } = geminiAccess(access, null, GeminiInteractionsWire_API_Interactions.getPath(resumeHandle.runId /* Gemini interaction.id */), false);
       return {
         request: { url: `${_baseUrl}${_baseUrl.includes('?') ? '&' : '?'}stream=true`, method: 'GET', headers: _headers },
+        /** Again, only support SSE here, for now (see comment in `createChatGenerateDispatch`) */
         demuxerFormat: 'fast-sse',
-        chatGenerateParse: createGeminiInteractionsParser(null /* model name unknown at resume time - caller's DMessage already has it */),
+        chatGenerateParse: createGeminiInteractionsParserSSE(null /* model name unknown at resume time - caller's DMessage already has it */),
       };
     }
 
@@ -393,6 +398,21 @@ export async function executeChatGenerateDelete(access: AixAPI_Access, handle: A
     case 'gemini':
       if (handle.uht !== 'vnd.gem.interactions')
         throw new Error(`Delete handle mismatch for gemini: expected 'vnd.gem.interactions', got '${handle.uht}'`);
+
+      // Gemini: cancel the background run first (stops token generation), then DELETE the stored record.
+      // The DELETE endpoint only removes the resource; it does NOT cancel an in-flight run.
+      // Cancel may 404 "Method not found" on the Developer API (API-key mode, googleapis/python-genai#1971) -
+      // we log the outcome and proceed to DELETE so local cleanup still happens.
+      const { url: cancelUrl, headers: cancelHeaders } = geminiAccess(access, null, GeminiInteractionsWire_API_Interactions.cancelPath(handle.runId), false);
+      try {
+        const cancelResp = await fetchResponseOrTRPCThrow({ url: cancelUrl, method: 'POST', body: {}, headers: cancelHeaders, signal: abortSignal, name: 'Aix.Gemini.Interactions.cancel', throwWithoutName: true });
+        console.log(`[AIX] Gemini.Interactions.cancel: ok=${cancelResp.ok} status=${cancelResp.status}`);
+      } catch (error: any) {
+        if (abortSignal.aborted) throw error;
+        const status = error instanceof TRPCFetcherError ? error.httpStatus : undefined;
+        console.log(`[AIX] Gemini.Interactions.cancel: failed status=${status ?? '?'} msg=${error?.message ?? 'unknown'}`);
+      }
+
       ({ url, headers } = geminiAccess(access, null, GeminiInteractionsWire_API_Interactions.deletePath(handle.runId), false));
       name = 'Aix.Gemini.Interactions.delete';
       break;

src/modules/aix/server/dispatch/chatGenerate/chatGenerate.executor.ts

@@ -26,7 +26,6 @@ import { heartbeatsWhileAwaiting } from '../heartbeatsWhileAwaiting';
  */
 export async function* executeChatGenerateDispatch(
   dispatchCreatorFn: () => Promise<ChatGenerateDispatch>,
-  streaming: boolean,
   intakeAbortSignal: AbortSignal,
   _d: AixDebugObject,
   parseContext?: { retriesAvailable: boolean },
@@ -59,7 +58,7 @@ export async function* executeChatGenerateDispatch(
   const innerStream = (async function* () {
 
     // Consume dispatch response
-    if (!streaming)
+    if (dispatch.demuxerFormat === null /* NS */)
       yield* _consumeDispatchUnified(dispatchResponse, dispatch.chatGenerateParse, chatGenerateTx, _d, parseContext);
     else
       yield* _consumeDispatchStream(dispatchResponse, dispatch.bodyTransform ?? null, dispatch.demuxerFormat, dispatch.chatGenerateParse, chatGenerateTx, _d, parseContext);

src/modules/aix/server/dispatch/chatGenerate/chatGenerate.operation-retry.ts

@@ -44,7 +44,6 @@ export class OperationRetrySignal extends Error {
  */
 export async function* executeChatGenerateWithOperationRetry(
   dispatchCreatorFn: () => Promise<ChatGenerateDispatch>,
-  streaming: boolean,
   abortSignal: AbortSignal,
   _d: AixDebugObject,
 ): AsyncGenerator<AixWire_Particles.ChatGenerateOp, void> {
@@ -55,7 +54,7 @@ export async function* executeChatGenerateWithOperationRetry(
   while (true) {
     try {
 
-      yield* executeChatGenerateDispatch(dispatchCreatorFn, streaming, abortSignal, _d, {
+      yield* executeChatGenerateDispatch(dispatchCreatorFn, abortSignal, _d, {
         retriesAvailable: attemptNumber < maxAttempts,
       });
 

src/modules/aix/server/dispatch/chatGenerate/parsers/gemini.interactions.parser.ts

@@ -5,6 +5,7 @@ import type { ChatGenerateParseFunction } from '../chatGenerate.dispatch';
 import type { IParticleTransmitter } from './IParticleTransmitter';
 
 import { GeminiInteractionsWire_API_Interactions } from '../../wiretypes/gemini.interactions.wiretypes';
+import { IssueSymbols } from '../ChatGenerateTransmitter';
 import { geminiConvertPCM2WAV } from './gemini.audioutils';
 
 
@@ -44,7 +45,7 @@ type BlockState = {
  * the cursor (or from start if omitted). Our parser is position-idempotent within a single run
  * because the transmitter's state carries across events.
  */
-export function createGeminiInteractionsParser(requestedModelName: string | null): ChatGenerateParseFunction {
+export function createGeminiInteractionsParserSSE(requestedModelName: string | null): ChatGenerateParseFunction {
 
   const parserCreationTimestamp = Date.now();
   let timeToFirstContent: number | undefined;
@@ -218,11 +219,16 @@ export function createGeminiInteractionsParser(requestedModelName: string | null
       }
 
       case 'error':
-        // Observed mid-stream with an empty payload between content blocks - non-fatal, the stream
+        // Two observed shapes:
-        // continues with further events and eventually an interaction.complete. Silent-skip empty
+        //  1) Empty payload mid-stream (Beta noise): the stream continues with further events and
-        // payloads (Beta noise); warn only when actual error info is present.
+        //     eventually an interaction.complete - silent-skip.
-        if (event.error?.message || event.error?.code)
+        //  2) Populated payload with message/code: terminal upstream error (also how Gemini reports
-          console.warn('[GeminiInteractions] SSE error event:', event.error);
+        //     cancelled interactions: HTTP 500 to the cancel call + an error SSE on the stream).
+        //     Surface as a dialect-terminating issue so the UI renders it and the stream ends cleanly.
+        if (event.error?.message || event.error?.code) {
+          const errorText = `${event.error.code ? `${event.error.code}: ` : ''}${event.error.message || 'Upstream error.'}`;
+          pt.setDialectTerminatingIssue(errorText, IssueSymbols.Generic, 'srv-warn');
+        }
         break;
 
       default: {

src/modules/aix/server/dispatch/chatGenerate/parsers/openai.parser.ts

@@ -494,6 +494,10 @@ export function createOpenAIChatCompletionsParserNS(): ChatGenerateParseFunction
       } else if (message.content !== undefined && message.content !== null)
         throw new Error(`unexpected message content type: ${typeof message.content}`);
 
+      // [DeepSeek, 2026-04-24] Non-streaming reasoning_content -> 'ma' reasoning part (mirror of streaming path above)
+      if (typeof message.reasoning_content === 'string' && message.reasoning_content)
+        pt.appendReasoningText(message.reasoning_content);
+
       // [OpenRouter, 2025-01-20] Handle structured reasoning_details
       if (Array.isArray(message.reasoning_details)) {
         for (const reasoningDetail of message.reasoning_details) {

src/modules/aix/server/dispatch/stream.demuxers.ts

@@ -21,7 +21,7 @@ export namespace AixDemuxers {
    * - 'fast-sse' is our own parser, optimized for performance. to be preferred when possible over 'sse' (check for full compatibility with the upstream)
    * - 'json-nl' is used by Ollama
    */
-  export type StreamDemuxerFormat = 'fast-sse' | 'json-nl' | null;
+  export type StreamDemuxerFormat = 'fast-sse' | 'json-nl';
 
 
   /**
@@ -34,8 +34,8 @@ export namespace AixDemuxers {
         return createFastEventSourceDemuxer();
       case 'json-nl':
         return _createJsonNlDemuxer();
-      case null:
+      default:
-        return _nullStreamDemuxerWarn;
+        throw new Error(`Unsupported stream demuxer format: ${format}`);
     }
   }
 
@@ -115,12 +115,3 @@ function _createJsonNlDemuxer(): AixDemuxers.StreamDemuxer {
     },
   };
 }
-
-
-const _nullStreamDemuxerWarn: AixDemuxers.StreamDemuxer = {
-  demux: () => {
-    console.warn('Null demuxer called - shall not happen, as it is only created in non-streaming');
-    return [];
-  },
-  flushRemaining: () => [],
-};

src/modules/aix/server/dispatch/wiretypes/gemini.interactions.wiretypes.ts

@@ -23,8 +23,12 @@ export namespace GeminiInteractionsWire_API_Interactions {
 
   export const getPath = (id: string) => `/v1beta/interactions/${encodeURIComponent(id)}`;
 
+  // DELETE. Removes the stored record. Orthogonal to cancel; when removed the original connection may still be running and streaming
   export const deletePath = (id: string) => `/v1beta/interactions/${encodeURIComponent(id)}`;
 
+  // POST. Only cancels background interactions that are still running
+  export const cancelPath = (id: string) => `/v1beta/interactions/${encodeURIComponent(id)}/cancel`;
+
 
   // -- Request Body (POST /v1beta/interactions) --
 

src/modules/aix/server/dispatch/wiretypes/openai.wiretypes.ts

@@ -189,6 +189,13 @@ export namespace OpenAIWire_Messages {
     /** [OpenRouter, 2025-01-20] Reasoning traces with multiple blocks (summary, text, encrypted). */
     reasoning_details: z.array(OpenAIWire_ContentParts.OpenRouter_ReasoningDetail_schema).optional(),
 
+    /**
+     * [DeepSeek, 2026-04-24] Chain-of-thought reasoning text.
+     * - Response: emitted by V4 thinking-by-default; parsed into a 'ma' reasoning part.
+     * - (this) Request: MUST be echoed back on assistant turns that carry tool_calls (otherwise HTTP 400: "The reasoning_content in the thinking mode must be passed back to the API.").
+     */
+    reasoning_content: z.string().nullable().optional(),
+
     // function_call: // ignored, as it's deprecated
     // name: _optionalParticipantName, // omitted by choice: generally unsupported
   });
@@ -331,7 +338,7 @@ export namespace OpenAIWire_API_Chat_Completions {
     stream_options: z.object({
       include_usage: z.boolean().optional(), // If set, an additional chunk will be streamed with a 'usage' field on the entire request.
     }).optional(),
-    reasoning_effort: z.enum(['none', 'minimal', 'low', 'medium', 'high', 'xhigh']).optional(), // [OpenAI, 2024-12-17] [Perplexity, 2025-06-23] reasoning effort
+    reasoning_effort: z.enum(['none', 'minimal', 'low', 'medium', 'high', 'xhigh', 'max']).optional(), // [OpenAI, 2024-12-17] [Perplexity, 2025-06-23] reasoning effort; [DeepSeek, 2026-04-23] 'max' added for V4
     // OpenAI and [OpenRouter, 2025-01-20] Verbosity parameter - maps to output_config.effort for Anthropic models
     // https://openrouter.ai/docs/api/reference/parameters#verbosity
     verbosity: z.enum([
@@ -342,7 +349,7 @@ export namespace OpenAIWire_API_Chat_Completions {
     // [OpenRouter, 2025-11-11] Unified reasoning parameter for all models
     reasoning: z.object({
       max_tokens: z.int().optional(), // Token-based control (Anthropic, Gemini): 1024-32000
-      effort: z.enum(['none', 'minimal', 'low', 'medium', 'high', 'xhigh']).optional(), // Effort-based control (OpenAI o1/o3/GPT-5, xAI, DeepSeek): allocates % of max_tokens
+      effort: z.enum(['none', 'minimal', 'low', 'medium', 'high', 'xhigh', 'max']).optional(), // Effort-based control (OpenAI o1/o3/GPT-5, xAI, DeepSeek): allocates % of max_tokens
       enabled: z.boolean().optional(), // Simple enable with medium effort defaults
       exclude: z.boolean().optional(), // Use reasoning internally without returning it in response
     }).optional(),
@@ -447,6 +454,8 @@ export namespace OpenAIWire_API_Chat_Completions {
     search_after_date_filter: z.string().optional(), // Date filter in MM/DD/YYYY format
 
     // [Moonshot, 2026-01-26] Kimi K2.5 thinking mode control
+    // [Z.ai, 2025-xx] GLM thinking mode: type 'enabled' | 'disabled'
+    // [DeepSeek, 2026-04-23] V4 thinking mode: same binary shape; depth is controlled via top-level `reasoning_effort`
     thinking: z.object({
       type: z.enum(['enabled', 'disabled']),
     }).optional(),
@@ -1641,7 +1650,7 @@ export namespace OpenAIWire_API_Responses {
     // NOTE: .catch() gracefully degrades to undefined since this is a non-critical enrichment path
     tools: z.array(OpenAIWire_Responses_Tools.Tool_schema).optional().catch((ctx) => {
       console.warn('[DEV] AIX: OpenAI Responses: unable to parse echoed tools, ignoring:', { tools: ctx.value });
-      return;
+      return undefined;
     }),
 
     output: z.array(OpenAIWire_Responses_Items.OutputItem_schema),

src/modules/beam/store-beam_vanilla.ts

@@ -76,6 +76,12 @@ const createRootSlice: StateCreator<BeamStore, [], [], RootStoreSlice> = (_set,
   open: (chatHistory: Readonly<DMessage[]>, initialChatLlmId: DLLMId | null, isEditMode: boolean, callback: BeamSuccessCallback) => {
     const { isOpen: wasAlreadyOpen, terminateKeepingSettings, loadBeamConfig, hadImportedRays, setRayLlmIds, setCurrentGatherLlmId } = _get();
 
+    // if already open, preserve the live state (rays, fusions, callback) - re-invocation must never wipe an ongoing beam
+    if (wasAlreadyOpen) {
+      console.warn('[DEV] Beam is already open');
+      return;
+    }
+
     // reset pending operations
     terminateKeepingSettings();
 

src/modules/llms/models-modal/LLMParametersEditor.tsx

@@ -51,6 +51,7 @@ const _oaiEffortOptions = [
 ] as const;
 
 const _miscEffortOptions = [
+  { value: 'max', label: 'Max', description: 'Hardest thinking' } as const,
   { value: 'high', label: 'On', description: 'Multi-step reasoning' } as const,
   { value: 'none', label: 'Off', description: 'Disable thinking mode' } as const,
   { value: _UNSPECIFIED, label: 'Default', description: 'Model Default' } as const,

src/modules/llms/server/gemini/gemini.models.ts

@@ -378,7 +378,7 @@ const _knownGeminiModels: ({
     labelOverride: 'Deep Research Preview (2026-04)',
     isPreview: true,
     chatPrice: gemini25ProPricing, // pricing not explicitly listed; using 2.5 Pro as baseline
-    interfaces: [LLM_IF_OAI_Chat, LLM_IF_OAI_Reasoning, LLM_IF_GEM_Interactions],
+    interfaces: [LLM_IF_OAI_Chat, LLM_IF_OAI_Vision, LLM_IF_OAI_Reasoning, LLM_IF_GEM_Interactions],
     parameterSpecs: [],
     benchmark: undefined, // Deep research model, not benchmarkable on standard tests
     // 128K input, 64K output
@@ -390,7 +390,7 @@ const _knownGeminiModels: ({
     labelOverride: 'Deep Research Max Preview (2026-04)',
     isPreview: true,
     chatPrice: gemini25ProPricing, // baseline estimate (see note above)
-    interfaces: [LLM_IF_OAI_Chat, LLM_IF_OAI_Reasoning, LLM_IF_GEM_Interactions],
+    interfaces: [LLM_IF_OAI_Chat, LLM_IF_OAI_Vision, LLM_IF_OAI_Reasoning, LLM_IF_GEM_Interactions],
     parameterSpecs: [],
     benchmark: undefined, // Deep research model, not benchmarkable on standard tests
   },
@@ -402,7 +402,7 @@ const _knownGeminiModels: ({
     labelOverride: 'Deep Research Pro Preview',
     isPreview: true,
     chatPrice: gemini25ProPricing,
-    interfaces: [LLM_IF_OAI_Chat, LLM_IF_OAI_Reasoning, LLM_IF_GEM_Interactions],
+    interfaces: [LLM_IF_OAI_Chat, LLM_IF_OAI_Vision, LLM_IF_OAI_Reasoning, LLM_IF_GEM_Interactions],
     parameterSpecs: [{ paramId: 'llmVndGeminiThinkingBudget' }],
     benchmark: undefined, // Deep research model, not benchmarkable on standard tests
     // Note: 128K input context, 64K output context

src/modules/llms/server/openai/models/deepseek.models.ts

@@ -1,38 +1,70 @@
-import { LLM_IF_HOTFIX_StripImages, LLM_IF_OAI_Chat, LLM_IF_OAI_Fn, LLM_IF_OAI_Json, LLM_IF_OAI_Reasoning } from '~/common/stores/llms/llms.types';
+import { LLM_IF_HOTFIX_StripImages, LLM_IF_OAI_Chat, LLM_IF_OAI_Fn, LLM_IF_OAI_Reasoning } from '~/common/stores/llms/llms.types';
 
 import type { ModelDescriptionSchema } from '../../llm.server.types';
 
 import { fromManualMapping, ManualMappings } from '../../models.mappings';
 
 
-const IF_3 = [LLM_IF_HOTFIX_StripImages, LLM_IF_OAI_Chat, LLM_IF_OAI_Fn, LLM_IF_OAI_Json];
+const IF_4 = [LLM_IF_HOTFIX_StripImages, LLM_IF_OAI_Chat, LLM_IF_OAI_Fn];
 
+// [DeepSeek, 2026-04-24] V4 release - https://api-docs.deepseek.com/news/news260424
+// - V4-Pro: 1.6T total / 49B active params; V4-Flash: 284B total / 13B active params (Novel Attention: token-wise compression + DSA)
+// - Model IDs listed by /models: deepseek-v4-flash, deepseek-v4-pro
+// - 1M context is the default across services; text-only (no vision/multimodal)
+// - Legacy aliases still accepted until 2026-07-24: deepseek-chat -> v4-flash (thinking disabled), deepseek-reasoner -> v4-flash (thinking enabled)
+// - Reasoning control: object `thinking: { type: 'enabled'|'disabled', reasoning_effort?: 'high'|'max' }`
+//   (the live API also accepts type: 'adaptive', but it is undocumented and empirically behaves the same as 'enabled'
+//    on current builds -- deliberately not exposed here; add it once docs + semantics stabilize)
+// - V3.2 endpoints no longer accessible via direct model ID (API returns only v4-flash/v4-pro)
 const _knownDeepseekChatModels: ManualMappings = [
-  // [Models and Pricing](https://api-docs.deepseek.com/quick_start/pricing)
+  {
-  // [List Models](https://api-docs.deepseek.com/api/list-models)
+    idPrefix: 'deepseek-v4-pro',
-  // [Release Notes - V3.2](https://api-docs.deepseek.com/news/news251201) - Released 2025-12-01
+    label: 'DeepSeek V4 Pro',
+    description: 'Premium reasoning model with 1M context. Supports extended thinking modes, JSON output, and function calling.',
+    contextWindow: 1_048_576, // 1M
+    interfaces: [...IF_4, LLM_IF_OAI_Reasoning],
+    parameterSpecs: [
+      { paramId: 'llmVndMiscEffort', enumValues: ['none', 'high', 'max'] },
+    ],
+    maxCompletionTokens: 65536, // conservative default; docs advertise up to 384K
+    chatPrice: { input: 1.74, output: 3.48, cache: { cType: 'oai-ac', read: 0.145 } },
+    benchmark: { cbaElo: 1463 }, // lmarena: deepseek-v4-pro (thinking variant 1462, near-tied)
+  },
+  {
+    idPrefix: 'deepseek-v4-flash',
+    label: 'DeepSeek V4 Flash',
+    description: 'Fast general-purpose model with 1M context. Supports extended thinking modes, JSON output, and function calling.',
+    contextWindow: 1_048_576, // 1M
+    interfaces: [...IF_4, LLM_IF_OAI_Reasoning],
+    parameterSpecs: [
+      { paramId: 'llmVndMiscEffort', enumValues: ['none', 'high', 'max'] },
+    ],
+    maxCompletionTokens: 65536, // conservative default; docs advertise up to 384K
+    chatPrice: { input: 0.14, output: 0.28, cache: { cType: 'oai-ac', read: 0.028 } },
+    benchmark: { cbaElo: 1439 }, // lmarena: deepseek-v4-flash-thinking (non-thinking variant 1433)
+  },
+  // Legacy aliases - API routes both to deepseek-v4-flash with thinking pre-set
   {
     idPrefix: 'deepseek-reasoner',
-    label: 'DeepSeek V3.2 (Reasoner)',
+    label: 'DeepSeek Reasoner (legacy)',
-    description: 'Reasoning model with Chain-of-Thought capabilities, 128K context length. Supports JSON output and function calling.',
+    description: 'Legacy alias: routes to DeepSeek V4 Flash with thinking enabled. Retires 2026-07-24.',
-    contextWindow: 131072, // 128K
+    contextWindow: 1_048_576,
-    interfaces: [...IF_3, LLM_IF_OAI_Reasoning],
+    interfaces: [...IF_4, LLM_IF_OAI_Reasoning],
-    // parameterSpecs: [
+    maxCompletionTokens: 65536,
-    //   { paramId: 'llmVndMiscEffort', enumValues: ['none', 'high'] }, // not supported: this model is reasoning only
+    chatPrice: { input: 0.14, output: 0.28, cache: { cType: 'oai-ac', read: 0.028 } },
-    // ],
+    benchmark: { cbaElo: 1439 }, // lmarena: deepseek-v4-flash-thinking
-    maxCompletionTokens: 32768, // default, max: 65536
+    isLegacy: true,
-    chatPrice: { input: 0.28, output: 0.42, cache: { cType: 'oai-ac', read: 0.028 } },
-    benchmark: { cbaElo: 1425 }, // deepseek-v3.2-exp-thinking
   },
   {
     idPrefix: 'deepseek-chat',
-    label: 'DeepSeek V3.2',
+    label: 'DeepSeek Chat (legacy)',
-    description: 'General-purpose model with 128K context length. Supports JSON output and function calling.',
+    description: 'Legacy alias: routes to DeepSeek V4 Flash with thinking disabled. Retires 2026-07-24.',
-    contextWindow: 131072, // 128K
+    contextWindow: 1_048_576,
-    interfaces: IF_3,
+    interfaces: IF_4,
-    maxCompletionTokens: 8192, // default is 4096, max is 8192
+    maxCompletionTokens: 65536,
-    chatPrice: { input: 0.28, output: 0.42, cache: { cType: 'oai-ac', read: 0.028 } },
+    chatPrice: { input: 0.14, output: 0.28, cache: { cType: 'oai-ac', read: 0.028 } },
-    benchmark: { cbaElo: 1424 }, // deepseek-v3.2
+    benchmark: { cbaElo: 1433 }, // lmarena: deepseek-v4-flash (non-thinking)
+    isLegacy: true,
   },
 ];
 

src/modules/llms/server/openai/models/openrouter.models.ts

@@ -246,7 +246,10 @@ export function openRouterModelToModelDescription(wireModel: object): ModelDescr
       // 0-day: xAI/Grok/Moonshot/Z.ai/DeepSeek models get default reasoning effort if not inherited
       if (interfaces.includes(LLM_IF_OAI_Reasoning) && !parameterSpecs.some(p => p.paramId === 'llmVndMiscEffort')) {
         // console.log('[DEV] openRouterModelToModelDescription: unexpected xAI/Grok/DeepSeek reasoning model:', model.id);
-        parameterSpecs.push({ paramId: 'llmVndMiscEffort' }); // binary thinking for these vendors
+        // Binary thinking only: OpenRouter's unified reasoning API currently rejects 'max' (see openai.chatCompletions.ts).
+        // We pin enumValues here so the shared llmVndMiscEffort registry (which also includes 'max' for native DeepSeek V4)
+        // does not surface 'max' in the UI for OR-routed models that can't honor it.
+        parameterSpecs.push({ paramId: 'llmVndMiscEffort', enumValues: ['none', 'high'] });
       }
       break;