AIX: Transmit token stop errors, if provided

AIX: OpenAI: mark the need for roundtrip of hosted tool pairs
Beam: Fix ghost columns. Fixes #1073
2026-05-10 21:50:14 -07:00 · 2026-04-24 17:08:40 -07:00 · 2026-04-24 17:08:40 -07:00 · 2026-04-24 16:04:29 -07:00 · 2026-04-24 16:04:29 -07:00 · 2026-04-24 15:56:13 -07:00
48 changed files with 521 additions and 207 deletions
@@ -583,9 +583,11 @@ export function AppChat() {
  }, []);

  useGlobalShortcuts('AppChat', React.useMemo(() => [
-    // focused conversation
-    { 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' },
+    // focused conversation (excluded when Beam is open so the keystroke passes through to the browser)
+    ...(beamOpenStoreInFocusedPane ? [] : [
+      { 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 <>
@@ -33,7 +33,10 @@ const _styles = {
    } as const,
    '& nav > ol > li:first-of-type': {
      overflow: 'hidden',
-      maxWidth: { xs: '110px', md: '140px' },
+      // allow the chat title to use available space, shrinking gracefully when the bar is narrow
+      // NOTE: already performed by virtue of the breadcrumb having agi-ellipsize on the crumbs
+      // flexShrink: 1,
+      // minWidth: '60px',
    } as const,

  } as const,
@@ -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 isExpired = expiresAt != null && Date.now() > expiresAt;
+  const { expiresAt /*, runId = ''*/ } = props.upstreamHandle;
+  // 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>
        )}
@@ -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) */}
-          {!messagePendingIncomplete && props.isBottom && fromAssistant && messageGenerator?.upstreamHandle && (!!onMessageUpstreamResume || !!onMessageUpstreamDelete) && (
+          {/* Upstream Resume - shows whenever there's a stored handle (incl. post-reload, and while streaming so Stop can cancel the upstream run) */}
+          {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}
            />
          )}
@@ -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>Inline</b>: download and embed text/images.</li>
-          <li><b>Inline + Free</b>: embed, then delete from Anthropic to free storage.</li>
+          <li><b>Show</b>: keep as references.</li>
+          <li><b>Embed</b>: download and embed text/images (default).</li>
+          <li><b>Embed + Free</b>: embed, then delete from Anthropic to free storage.</li>
        </ul>
        Only affects Anthropic models.
      </>}
@@ -23,7 +23,7 @@ export const Release = {

  // this is here to trigger revalidation of data, e.g. models refresh
  Monotonics: {
-    Aix: 67,
+    Aix: 69,
    NewsVersion: 204,
  },

@@ -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);
@@ -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)
  }),

@@ -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')}
@@ -905,9 +905,12 @@ export class ContentReassembler {
  /**
   * Stores raw termination data from the wire - classification deferred to finalizeReassembly()
   */
-  private onCGEnd({ terminationReason, tokenStopReason }: Extract<AixWire_Particles.ChatGenerateOp, { cg: 'end' }>): void {
+  private onCGEnd({ terminationReason, tokenStopReason, tokenStopError }: Extract<AixWire_Particles.ChatGenerateOp, { cg: 'end' }>): void {
    this.S.terminationReason = terminationReason;
    this.S.dialectStopReason = tokenStopReason;
+    // Vendor-composed stop error, surfaced as a complementary error fragment alongside the generic classification message
+    if (tokenStopError)
+      this._appendErrorFragment(tokenStopError);
  }

  /**
@@ -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 };

 }

@@ -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);
 }

 /**
@@ -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
-  let aixStreaming = !!onTextStreamUpdate;
+  // Caller streaming preference - implicit: stream if a callback is provided
+  const callerStreaming = !!onTextStreamUpdate;


  // Client-side late stage model HotFixes
-  const { shallDisableStreaming } = await clientHotFixGenerateRequest_ApplyAll(llm.interfaces, aixChatGenerate, llmParameters.llmRef || llm.id);
-  if (shallDisableStreaming || aixModel.forceNoStream)
-    aixStreaming = false;
+  const { hotfixNoStream } = await clientHotFixGenerateRequest_ApplyAll(llm.interfaces, aixChatGenerate, llmParameters.llmRef || llm.id);
+  const wireStreaming = !hotfixNoStream && !aixModel.forceNoStream ? callerStreaming : 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
-  const { shallDisableStreaming } = await clientHotFixGenerateRequest_ApplyAll(llm.interfaces, aixChatGenerate, llmParameters.llmRef || llm.id);
-  if (shallDisableStreaming || aixModel.forceNoStream)
-    aixStreaming = false;
+  // 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 { hotfixNoStream } = await clientHotFixGenerateRequest_ApplyAll(llm.interfaces, aixChatGenerate, llmParameters.llmRef || llm.id);
+  const wireStreaming = !hotfixNoStream && !aixModel.forceNoStream ? 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)
-    ({ clientSideChatGenerate, clientSideReattachUpstream } = await _loadCsfModuleOrThrow());
+  if (aixAccess.clientSideFetch) {
+    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 })

@@ -7,6 +7,7 @@ import { Box, Card, Chip, Divider, Sheet, Typography } from '@mui/joy';
 import { RenderCodeMemo } from '~/modules/blocks/code/RenderCode';

 import { ExpanderControlledBox } from '~/common/components/ExpanderControlledBox';
+import { objectDeepCloneWithStringLimit } from '~/common/util/objectUtils';
 import TimelapseIcon from '@mui/icons-material/Timelapse';

 import type { AixClientDebugger } from './memstore-aix-client-debugger';
@@ -184,12 +185,10 @@ export function AixDebuggerFrame(props: {
          {/* List of particles */}
          {frame.particles.map((particle, idx) => {

-            // truncated preview of particle content
+            // preview of particle content: preserve structure, trim long string fields
            let jsonPreview = '';
            try {
-              const content = particle.content;
-              jsonPreview = JSON.stringify(content).substring(0, 1024);
-              if (jsonPreview.length >= 1024) jsonPreview += '...';
+              jsonPreview = JSON.stringify(objectDeepCloneWithStringLimit(particle.content, 'aix-debugger-particle', 64));
            } catch (e) {
              jsonPreview = 'Error parsing content';
            }
@@ -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);
    }),

  /**
@@ -689,7 +689,7 @@ export namespace AixWire_Particles {

  export type ChatControlOp =
  // | { cg: 'start' } // not really used for now
-    | { cg: 'end', terminationReason: CGEndReason /* we know why we're sending 'end' */, tokenStopReason?: GCTokenStopReason /* we may or not have gotten a logical token stop reason from the dispatch */ }
+    | { cg: 'end', terminationReason: CGEndReason /* we know why we're sending 'end' */, tokenStopReason?: GCTokenStopReason /* we may or not have gotten a logical token stop reason from the dispatch */, tokenStopError?: string /* optional vendor-composed human-readable detail paired with tokenStopReason */ }
    | { cg: 'issue', issueId: CGIssueId, issueText: string }
    | { cg: 'aix-info', ait: 'flow-cont' /* important: establishes a checkpoint */, text: string }
    | { cg: 'aix-retry-reset', rScope: 'srv-dispatch' | 'srv-op' | 'cli-ll', rClearStrategy: 'none' | 'since-checkpoint' | 'all', reason: string, attempt: number, maxAttempts: number, delayMs: number, causeHttp?: number, causeConn?: string }
@@ -56,6 +56,7 @@ export class ChatGenerateTransmitter implements IParticleTransmitter {

  // Token stop reason
  private tokenStopReason: AixWire_Particles.GCTokenStopReason | undefined = undefined;
+  private tokenStopError: string | undefined = undefined;

  // Metrics
  private accMetrics: AixWire_Particles.CGSelectMetrics | undefined = undefined;
@@ -105,6 +106,7 @@ export class ChatGenerateTransmitter implements IParticleTransmitter {
        cg: 'end',
        terminationReason: this.terminationReason,
        tokenStopReason: this.tokenStopReason, // See NOTE above - || (dispatchOrDialectIssue ? 'cg-issue' : 'ok'),
+        ...(this.tokenStopError && { tokenStopError: this.tokenStopError }),
      });
      // Keep this in a terminated state, so that every subsequent call will yield errors (not implemented)
      // this.terminationReason = null;
@@ -201,12 +203,13 @@ export class ChatGenerateTransmitter implements IParticleTransmitter {
    this.setDialectEnded('issue-dialect');
  }

-  setTokenStopReason(reason: AixWire_Particles.GCTokenStopReason) {
+  setTokenStopReason(reason: AixWire_Particles.GCTokenStopReason, errorText?: string) {
    if (SERVER_DEBUG_WIRE)
-      console.log('|token-stop|', reason);
+      console.log('|token-stop|', reason, errorText ?? '');
    if (this.tokenStopReason && this.tokenStopReason !== reason)
      console.warn(`[Aix.${this.prettyDialect}] setTokenStopReason('${reason}'): already has token stop reason '${this.tokenStopReason}' (overriding)`);
    this.tokenStopReason = reason;
+    if (errorText) this.tokenStopError = errorText;
  }


@@ -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
-      throw new Error(`OpenAI ChatCompletions API does not support '${reasoningEffort}' reasoning effort`);
+    // for: 'alibaba' | 'azure' | 'groq' | 'lmstudio' | 'localai' | 'mistral' | 'openai' | 'openpipe' | 'togetherai' | 'xai'
    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
-      throw new Error(`${openAIDialect} only supports reasoning effort 'none' or 'high', got '${reasoningEffort}'`);
+    const allowedEffort = openAIDialect === 'deepseek' ? ['none', 'high', 'max'] : ['none', 'high'];
+    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;
    }
  }
@@ -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) {
@@ -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;
@@ -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);
@@ -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,
      });

@@ -15,8 +15,8 @@ export interface IParticleTransmitter {
  /** End the current part and flush it, which also calls `setDialectEnded('issue-dialect')` */
  setDialectTerminatingIssue(dialectText: string, symbol: string | null, serverLog: ParticleServerLogLevel): void;

-  /** Communicates the finish reason to the client - Data only, this does not do Control, like the above */
-  setTokenStopReason(reason: AixWire_Particles.GCTokenStopReason): void;
+  /** Communicates the finish reason to the client - Data only. Optional `errorText` is a vendor-composed string rendered as a complementary error fragment alongside the generic classification message. */
+  setTokenStopReason(reason: AixWire_Particles.GCTokenStopReason, errorText?: string): void;


  // Parts data //
@@ -404,7 +404,7 @@ export function createAnthropicMessageParser(): ChatGenerateParseFunction {
        // -> Token Stop Reason
        const tokenStopReason = _fromAnthropicStopReason(delta.stop_reason, 'message_delta');
        if (tokenStopReason !== null)
-          pt.setTokenStopReason(tokenStopReason);
+          pt.setTokenStopReason(tokenStopReason, _formatAnthropicStopError(delta.stop_details));

        // NOTE: we have more fields we're not parsing yet - https://platform.claude.com/docs/en/api/typescript/messages#message_delta_usage
        if (usage?.output_tokens && messageStartTime) {
@@ -511,6 +511,7 @@ export function createAnthropicMessageParserNS(): ChatGenerateParseFunction {
      content,
      container,
      stop_reason,
+      stop_details,
      usage,
    } = AnthropicWire_API_Message_Create.Response_schema.parse(JSON.parse(fullData));

@@ -653,7 +654,7 @@ export function createAnthropicMessageParserNS(): ChatGenerateParseFunction {
    // -> Token Stop Reason (pause_turn already thrown above)
    const tokenStopReason = _fromAnthropicStopReason(stop_reason, 'parser_NS');
    if (tokenStopReason !== null)
-      pt.setTokenStopReason(tokenStopReason);
+      pt.setTokenStopReason(tokenStopReason, _formatAnthropicStopError(stop_details));
  };
 }

@@ -681,6 +682,19 @@ function _emitContainerState(pt: IParticleTransmitter, container: { id: string;
  });
 }

+/** Compose a human-readable error string from Anthropic's stop_details. Returns undefined when nothing useful to surface. */
+function _formatAnthropicStopError(stopDetails: { type: string; category?: string | null; explanation?: string | null } | null | undefined): string | undefined {
+  if (!stopDetails) return undefined;
+  if (stopDetails.type !== 'refusal') {
+    aixResilientUnknownValue('Anthropic', 'stopDetailsType', stopDetails.type);
+    return undefined;
+  }
+  const parts: string[] = [];
+  if (stopDetails.category) parts.push(`[${stopDetails.category}]`);
+  if (stopDetails.explanation) parts.push(stopDetails.explanation);
+  return parts.length ? `Refusal: ${parts.join(' ')}` : undefined;
+}
+

 // --- Shared server tool result handlers (used by both S and NS parsers) ---

@@ -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
-        // continues with further events and eventually an interaction.complete. Silent-skip empty
-        // payloads (Beta noise); warn only when actual error info is present.
-        if (event.error?.message || event.error?.code)
-          console.warn('[GeminiInteractions] SSE error event:', event.error);
+        // Two observed shapes:
+        //  1) Empty payload mid-stream (Beta noise): the stream continues with further events and
+        //     eventually an interaction.complete - silent-skip.
+        //  2) Populated payload with message/code: terminal upstream error (also how Gemini reports
+        //     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: {
@@ -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) {
@@ -18,6 +18,21 @@ const OPENAI_RESPONSES_SAME_PART_SPACER = '\n\n';
 const INLINE_IMAGE_SKIP_RESIZE_MAX_B64_BYTES = 250_000; // skip resize for small images (e.g. code interpreter charts)


+/**
+ * Wishlist marker: hosted tool calls (web_search_call, image_generation_call, code_interpreter_call, ...)
+ * are rendered via ephemeral OperationState/inline-asset paths and are NOT round-tripped as structured
+ * fragments. This breaks stateless multi-turn with reasoning models. See PRD.FUTURE-atol.md "Wishlist:
+ * Hosted tool invocations as first-class fragments".
+ */
+// const _hostedToolWishlistSeen = new Set<string>();
+function _hostedToolWishlistHint(family: 'web_search' | 'image_generation' | 'code_interpreter' | 'custom_tool'): void {
+  // if (_hostedToolWishlistSeen.has(family)) return;
+  // _hostedToolWishlistSeen.add(family);
+  // NOTE: disable the log because it's logging all the time evenrwyehre; just implement this
+  // console.log(`[DEV] AIX: ATOL wishlist - hosted '${family}' call observed; not round-tripped as a structured fragment yet (see kb/product/PRD.FUTURE-atol.md)`);
+}
+
+
 /**
 * Safely sanitizes a URL for display in placeholders by removing query parameters and paths
 * to prevent leaking sensitive information while keeping the domain recognizable.
@@ -46,6 +61,11 @@ type TEventType = OpenAIWire_API_Responses.StreamingEvent['type'];
 // cached config for the image_generation hosted tool, captured at response.created
 type TImageGenToolCfg = Extract<OpenAIWire_Responses_Tools.Tool, { type: 'image_generation' }>;

+/** Extract the image_generation tool config from the echoed tools array (API does not echo `model` per-item). Shared by streaming and non-streaming paths. */
+function _findImageGenToolCfg(tools: TResponse['tools']): TImageGenToolCfg | undefined {
+  return tools?.find((t): t is TImageGenToolCfg => t.type === 'image_generation');
+}
+

 /**
 * We need this just to ensure events are not out of order, as out streaming is progressive
@@ -248,8 +268,7 @@ class ResponseParserStateMachine {
  // Hosted tool config capture

  captureHostedToolConfigs(tools: TResponse['tools']) {
-    if (!tools?.length) return;
-    this.#imageGenToolCfg = tools.find((t): t is TImageGenToolCfg => t.type === 'image_generation');
+    this.#imageGenToolCfg = _findImageGenToolCfg(tools);
  }

  get imageGenToolCfg() {
@@ -448,6 +467,7 @@ export function createOpenAIResponsesEventParser(): ChatGenerateParseFunction {
            break;

          case 'image_generation_call':
+            _hostedToolWishlistHint('image_generation');
            // -> IGC: process completed image generation using 'ii' particle for inline images
            const { id: igId, result: igResult, revised_prompt: igRevisedPrompt } = doneItem;
            const igDoneText = !igRevisedPrompt?.length ? 'Image generated'
@@ -765,6 +785,9 @@ export function createOpenAIResponseParserNS(): ChatGenerateParseFunction {
    if (response.model)
      pt.setModelName(response.model);

+    // -> Hosted tool config capture (needed for enriching done-item particles with tool params the API does not echo per-item, e.g. image_generation.model)
+    const imageGenToolCfg = _findImageGenToolCfg(response.tools);
+
    // -> Upstream Handle (for remote control: resume, cancel, delete)
    // NOTE: we don't do it for full responses, because they're supposed to be 'complete' - i.e. no 'background' execution

@@ -875,11 +898,16 @@ export function createOpenAIResponseParserNS(): ChatGenerateParseFunction {
            pt.appendReasoningText(item.text);
          }

+          // [DEV] surface cases that diverge from our continuity round-trip expectations
+          if (!reasoningId && !reasoningEC)
+            console.warn('[DEV] AIX: OpenAI-Response-NS: reasoning item has neither id nor encrypted_content - no continuity handle captured for this turn', { oItem });
+          else if (!reasoningEC)
+            console.log('[DEV] AIX: OpenAI-Response-NS: reasoning item has id but no encrypted_content - stateless round-trip requires include:[\'reasoning.encrypted_content\'] on the request');
+
          // Capture the continuity handle (encrypted_content + id) for stateless multi-turn round-tripping.
-          // Attached to the ma fragment produced by the summary above; if no summary was emitted, this may
-          // attach to an unrelated preceding fragment - tolerable as the worst case is a misfiled blob.
-          // FIXME: make sure we are attaching to an 'ma' (i.e. reasoning text or somehting was emitted)
-          if (reasoningEC || reasoningId)
+          if (reasoningEC || reasoningId) {
+            // Defensive: ensure an ma fragment exists as the attach target for the svs particle below (parity with the streaming path).
+            pt.appendReasoningText('');
            pt.sendSetVendorState({
              p: 'svs',
              vendor: 'openai',
@@ -890,10 +918,7 @@ export function createOpenAIResponseParserNS(): ChatGenerateParseFunction {
                },
              },
            });
-          else if (!reasoningId && !reasoningEC)
-            console.warn('[DEV] AIX: OpenAI-Response-NS: reasoning item has neither id nor encrypted_content - no continuity handle captured for this turn', { oItem });
-          else if (!reasoningEC)
-            console.log('[DEV] AIX: OpenAI-Response-NS: reasoning item has id but no encrypted_content - stateless round-trip requires include:[\'reasoning.encrypted_content\'] on the request');
+          }
          break;

        // Message contains the main 'assistant' response
@@ -957,6 +982,7 @@ export function createOpenAIResponseParserNS(): ChatGenerateParseFunction {
          break;

        case 'image_generation_call':
+          _hostedToolWishlistHint('image_generation');
          // -> IGC: process completed image generation using 'ii' particle for inline images
          const { result: igResult, revised_prompt: igRevisedPrompt } = oItem;
          // Create inline image with base64 data
@@ -965,7 +991,7 @@ export function createOpenAIResponseParserNS(): ChatGenerateParseFunction {
              _imageGenerationMimeType(oItem), // infer from output_format echoed in the item
              igResult,
              igRevisedPrompt || 'Generated image',
-              AIX_OAI_DEFAULT_IMAGE_GEN_MODEL, // generator: non-streaming path has no captured tool config, use current default
+              imageGenToolCfg?.model || AIX_OAI_DEFAULT_IMAGE_GEN_MODEL, // generator: read from echoed tools (API does not echo model per-item), fallback to current default
              igRevisedPrompt || '', // prompt used
            );
          else
@@ -1150,6 +1176,7 @@ function _imageGenerationMimeType(item: { output_format?: string }): string {
 * - citations: High-quality links (2-3) via annotations in message content
 */
 function _forwardDoneWebSearchCallItem(pt: IParticleTransmitter, webSearchCall: Extract<OpenAIWire_API_Responses.Response['output'][number], { type: 'web_search_call' }>, opId: string): void {
+  _hostedToolWishlistHint('web_search');
  const { action, status } = webSearchCall;

  const doneOpts = { opId, state: 'done' } as const;
@@ -1203,6 +1230,7 @@ function _forwardDoneWebSearchCallItem(pt: IParticleTransmitter, webSearchCall:
 * - addCodeExecutionResponse for each output result
 */
 function _forwardDoneCodeInterpreterCallItem(pt: IParticleTransmitter, codeInterpreterCall: Extract<OpenAIWire_API_Responses.Response['output'][number], { type: 'code_interpreter_call' }>): void {
+  _hostedToolWishlistHint('code_interpreter');
  const { id, code, outputs, status /*,container_id*/ } = codeInterpreterCall;

  // <- Emit code (like Gemini's executableCode)
@@ -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:
-        return _nullStreamDemuxerWarn;
+      default:
+        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: () => [],
-};
@@ -1,7 +1,7 @@
 <!--
  Upstream snapshot - DO NOT EDIT - run _upstream/sync.sh to refresh
  Source: https://platform.claude.com/docs/en/api/messages/create.md
-  Synced: 2026-04-23
+  Synced: 2026-04-24
  Consumed by: anthropic.wiretypes.ts, anthropic.parser.ts, anthropic.messageCreate.ts, anthropic.transform-fileInline.ts
 -->

@@ -2429,7 +2429,7 @@ Learn more about the Messages API in our [user guide](https://docs.claude.com/en

  Configuration options for the model's output, such as the output format.

-  - `effort: optional "low" or "medium" or "high" or 2 more`
+  - `effort: optional "low" or "medium" or "high" or "max"`

    All possible effort levels.

@@ -2439,8 +2439,6 @@ Learn more about the Messages API in our [user guide](https://docs.claude.com/en

    - `"high"`

-    - `"xhigh"`
-
    - `"max"`

  - `format: optional JSONOutputFormat`
@@ -3822,15 +3820,15 @@ Learn more about the Messages API in our [user guide](https://docs.claude.com/en

  Used to remove "long tail" low probability responses. [Learn more technical details here](https://towardsdatascience.com/how-to-sample-from-language-models-682bceb97277).

-  Recommended for advanced use cases only. You usually only need to use `temperature`.
+  Recommended for advanced use cases only.

 - `top_p: optional number`

  Use nucleus sampling.

-  In nucleus sampling, we compute the cumulative distribution over all the options for each subsequent token in decreasing probability order and cut it off once it reaches a particular probability specified by `top_p`. You should either alter `temperature` or `top_p`, but not both.
+  In nucleus sampling, we compute the cumulative distribution over all the options for each subsequent token in decreasing probability order and cut it off once it reaches a particular probability specified by `top_p`.

-  Recommended for advanced use cases only. You usually only need to use `temperature`.
+  Recommended for advanced use cases only.

 ### Returns

@@ -1,7 +1,7 @@
 <!--
  Upstream snapshot - DO NOT EDIT - run _upstream/sync.sh to refresh
  Source: https://ai.google.dev/gemini-api/docs/deep-research.md.txt
-  Synced: 2026-04-23
+  Synced: 2026-04-24
  Consumed by: gemini.interactions.wiretypes.ts, gemini.interactions.parser.ts, gemini.interactionsCreate.ts, gemini.interactionsPoller.ts
  Companion: ./gemini.interactions.guide.md (the Interactions API guide)
 -->
@@ -1,7 +1,7 @@
 <!--
  Upstream snapshot - DO NOT EDIT - run _upstream/sync.sh to refresh
  Source: https://ai.google.dev/gemini-api/docs/interactions.md.txt
-  Synced: 2026-04-23
+  Synced: 2026-04-24
  Consumed by: gemini.interactions.wiretypes.ts, gemini.interactions.parser.ts, gemini.interactionsCreate.ts, gemini.interactionsPoller.ts
  Companion: ./gemini.interactions.spec.md (the Interactions API reference spec), ./gemini.deep-research.guide.md (the Deep Research agent guide)
 -->
@@ -1,7 +1,7 @@
 <!--
  Upstream snapshot - DO NOT EDIT - run _upstream/sync.sh to refresh
  Source: https://ai.google.dev/api/interactions-api.md.txt
-  Synced: 2026-04-23
+  Synced: 2026-04-24
  Consumed by: gemini.interactions.wiretypes.ts, gemini.interactions.parser.ts, gemini.interactionsCreate.ts, gemini.interactionsPoller.ts
  Companion: ./gemini.interactions.guide.md (the Interactions API guide)
 -->
@@ -1,7 +1,7 @@
 <!--
  Upstream snapshot - DO NOT EDIT - run _upstream/sync.sh to refresh
  Source: https://developers.openai.com/api/reference/resources/responses/methods/create/index.md
-  Synced: 2026-04-23
+  Synced: 2026-04-24
  Consumed by: openai.wiretypes.ts, openai.responses.parser.ts, openai.responsesCreate.ts
 -->

@@ -13,6 +13,10 @@ const hotFixAntShipNoEmptyTextBlocks = true; // Replace empty text blocks with a
 *
 * ## Updates
 *
+ * ### 2026-04-24 - API Sync: stop_details for structured refusals
+ * - Response: added `stop_details` ({ type: 'refusal', category: 'cyber'|'bio'|null, explanation: string|null })
+ * - event_MessageDelta.delta: added `stop_details` (arrives alongside stop_reason in streaming)
+ *
 * ### 2026-03-21 - API Sync: GA tool versions, thinking display, caller updates, cache_control
 * - Tools: Added web_search_20260209 (GA), web_fetch_20260209/20260309 (GA), code_execution_20260120 (GA REPL)
 * - Request: Added top-level `cache_control` for automatic caching (Feb 2026)
@@ -825,6 +829,16 @@ export namespace AnthropicWire_API_Message_Create {
    'model_context_window_exceeded',
  ]);

+  /**
+   * Structured stop details, paired with stop_reason. Currently only populated when stop_reason === 'refusal'.
+   * Both `type` and `category` are loosely typed for forward-compat - parser warns on unknown `type`.
+   */
+  const StopDetails_schema = z.object({
+    type: z.enum(['refusal']).or(z.string()),
+    category: z.enum(['cyber', 'bio']).or(z.string()).nullish(),
+    explanation: z.string().nullish(),
+  });
+
  /// Request

  export type Request = z.infer<typeof Request_schema>;
@@ -1030,6 +1044,12 @@ export namespace AnthropicWire_API_Message_Create {
    // Which custom stop sequence was generated, if any.
    stop_sequence: z.string().nullable(),

+    /**
+     * Structured stop details. Present when stop_reason === 'refusal' (carries category + explanation).
+     * In streaming, stop_details is null at message_start and appears on message_delta alongside stop_reason.
+     */
+    stop_details: StopDetails_schema.nullish(),
+
    /**
     * Billing and rate-limit usage.
     * Token counts represent the underlying cost to Anthropic's systems.
@@ -1088,6 +1108,10 @@ export namespace AnthropicWire_API_Message_Create {
    delta: z.object({
      stop_reason: StopReason_schema.nullable(),
      stop_sequence: z.string().nullable(),
+      /**
+       * Structured stop details - present alongside stop_reason === 'refusal' (category + explanation).
+       */
+      stop_details: StopDetails_schema.nullish(),
      /**
       * Container state updates - present when Skills/code_execution tools are used.
       * Provides container id/expiry that may differ from message_start if the container was created mid-stream.
@@ -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) --

@@ -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(),
@@ -1174,9 +1183,11 @@ export namespace OpenAIWire_Responses_Items {
        // [OpenAI 2026-03-xx] DEPRECATED query might not always be present in done event
        query: z.string().optional(),
        // the output websites, if any [{"type":"url","url":"https://www.enricoros.com/"}, {"type":"url","url": "https://linkedin.com/in/enricoros/"}, ...]
+        // [OpenAI 2026-04-23, GPT-5.5] new source types: { type: 'api', name: 'oai-calculator' } for hosted-tool invocations (no url)
        sources: z.array(z.object({
-          type: z.literal('url').optional(), // source type
-          url: z.string(),
+          type: z.enum(['url', 'api']).or(z.string()).optional(), // 'url' (default) | 'api' (GPT-5.5 hosted tools) | future types
+          url: z.string().nullish(), // optional: 'api' sources have no url, only name
+          name: z.string().nullish(), // for 'api' sources (e.g., 'oai-calculator')
          // [OpenAI 2026-03-xx] not present anymore
          // title: z.string().optional(),
          // snippet: z.string().optional(),
@@ -1437,6 +1448,7 @@ export namespace OpenAIWire_Responses_Tools {
  const WebSearchTool_schema = z.object({
    type: z.enum(['web_search', 'web_search_preview', 'web_search_preview_2025_03_11']),
    search_context_size: z.enum(['low', 'medium', 'high']).optional(),
+    // [OpenAI 2026-04-23, GPT-5.5] API echoes user_location as `null` (not undefined) when unset - so .nullish()
    user_location: z.object({
      type: z.literal('approximate'),
      // API echoes these as `null` when unset, not omitted - so .nullish()
@@ -1444,7 +1456,7 @@ export namespace OpenAIWire_Responses_Tools {
      country: z.string().nullish(),
      region: z.string().nullish(),
      timezone: z.string().nullish(),
-    }).optional(),
+    }).nullish(),
    external_web_access: z.boolean().optional(),
  });

@@ -1641,7 +1653,7 @@ 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),
@@ -1,7 +1,9 @@
 import * as React from 'react';
 import { useShallow } from 'zustand/react/shallow';

-import { Alert, Box, CircularProgress } from '@mui/joy';
+import { Alert, Box, Button, CircularProgress } from '@mui/joy';
+import ContentCopyIcon from '@mui/icons-material/ContentCopy';
+import TelegramIcon from '@mui/icons-material/Telegram';

 import { ConfirmationModal } from '~/common/components/modals/ConfirmationModal';
 import { ShortcutKey, useGlobalShortcuts } from '~/common/components/shortcuts/useGlobalShortcuts';
@@ -204,13 +206,30 @@ export function BeamView(props: {
        isMobile={props.isMobile}
        rayIds={rayIds}
        showRayAdd={cardAdd}
-        showRaysOps={(isScattering || raysReady < 2) ? undefined : raysReady}
        hadImportedRays={hadImportedRays}
        onIncreaseRayCount={handleRayIncreaseCount}
-        onRaysOperation={handleRaysOperation}
        // linkedLlmId={currentGatherLlmId}
      />

+      {/* Rays Action Bar (2+ ready beams) - sibling of the grid (NOT a grid child); an in-grid spanning element with gridColumn:'1/-1' pins all auto-fit tracks open and leaves dead whitespace when raysCount < tracksCount. Fixes #1073. */}
+      {(!isScattering && raysReady >= 2) && (
+        <Box sx={{ display: 'flex', justifyContent: 'center', gap: 2, mx: 'var(--Pad)' }}>
+          <Button size='sm' variant='outlined' color='neutral' onClick={() => handleRaysOperation('copy')} endDecorator={<ContentCopyIcon sx={{ fontSize: 'md' }} />} sx={{
+            backgroundColor: 'background.surface',
+            '&:hover': { backgroundColor: 'background.popup' },
+          }}>
+            Copy {raysReady}
+          </Button>
+          <Button size='sm' variant='outlined' color='success' onClick={() => handleRaysOperation('use')} endDecorator={<TelegramIcon sx={{ fontSize: 'xl' }} />} sx={{
+            justifyContent: 'space-between',
+            backgroundColor: 'background.surface',
+            '&:hover': { backgroundColor: 'background.popup' },
+          }}>
+            Use {raysReady === 2 ? 'both' : 'all ' + raysReady} messages
+          </Button>
+        </Box>
+      )}
+

      {/* Gapper between Rays and Merge, without compromising the auto margin of the Ray Grid */}
      <Box />
@@ -246,9 +265,9 @@ export function BeamView(props: {
        onPositive={handleStartMergeConfirmation}
        // lowStakes
        noTitleBar
-        confirmationText='Some responses are still being generated. Do you want to stop and proceed with merging the available responses now?'
-        positiveActionText='Proceed with Merge'
-        negativeActionText='Wait for All Responses'
+        confirmationText={'Some replies are still generating. Merge what\'s ready?'}
+        positiveActionText='Merge now'
+        negativeActionText='Wait for all'
        negativeActionStartDecorator={
          <CircularProgress color='neutral' sx={{ '--CircularProgress-size': '24px', '--CircularProgress-trackThickness': '1px' }} />
        }
@@ -149,7 +149,8 @@ export function BeamFusionGrid(props: {
          </Box> : (
            <Typography level='body-sm' sx={{ opacity: 0.8 }}>
              {/*You need two or more replies for a {currentFactory?.shortLabel?.toLocaleLowerCase() ?? ''} merge.*/}
-              Waiting for multiple responses.
+              {/*Waiting for multiple responses.*/}
+              Merge needs 2+ replies. Beam some first.
            </Typography>
          )}
        </BeamCard>
@@ -49,7 +49,7 @@ export async function executeGatherInstruction(_i: GatherInstruction, inputs: Ex
  if (!inputs.chatMessages.length)
    throw new Error('No conversation history available');
  if (!inputs.rayMessages.length)
-    throw new Error('No responses available');
+    throw new Error('Needs two Beams at least');
  for (let rayMessage of inputs.rayMessages)
    if (rayMessage.role !== 'assistant')
      throw new Error('Invalid response role');
@@ -58,7 +58,7 @@ export function gatherStartFusion(
  if (chatMessages.length < 1)
    return onError('No conversation history available');
  if (rayMessages.length <= 1)
-    return onError('No responses available');
+    return onError('Needs two Beams at least');
  if (!initialFusion.llmId)
    return onError('No Merge model selected');

@@ -122,7 +122,7 @@ The final output should reflect a deep understanding of the user's preferences a
    addLabel: 'Add Breakdown',
    cardTitle: 'Evaluation Table',
    Icon: TableViewRoundedIcon as typeof SvgIcon,
-    description: 'Analyzes and compares AI responses, offering a structured framework to support your response choice.',
+    description: 'Analyzes and compares replies, with a structured framework to support your choice.',
    createInstructions: () => [
      {
        type: 'gather',
@@ -3,8 +3,6 @@ import * as React from 'react';
 import type { SxProps } from '@mui/joy/styles/types';
 import { Box, Button } from '@mui/joy';
 import AddCircleOutlineRoundedIcon from '@mui/icons-material/AddCircleOutlineRounded';
-import ContentCopyIcon from '@mui/icons-material/ContentCopy';
-import TelegramIcon from '@mui/icons-material/Telegram';

 import type { BeamStoreApi } from '../store-beam.hooks';
 import { BeamCard } from '../BeamCard';
@@ -32,10 +30,8 @@ export function BeamRayGrid(props: {
  hadImportedRays: boolean,
  isMobile: boolean,
  onIncreaseRayCount: () => void,
-  onRaysOperation: (operation: 'copy' | 'use') => void,
  rayIds: string[],
  showRayAdd: boolean,
-  showRaysOps: undefined | number,
 }) {

  const raysCount = props.rayIds.length;
@@ -71,25 +67,6 @@ export function BeamRayGrid(props: {
        </BeamCard>
      )}

-      {/* Multi-Use and Copy Buttons */}
-      {!!props.showRaysOps && (
-        <Box sx={{ gridColumn: '1 / -1', display: 'flex', justifyContent: 'center', gap: 2, mt: 2 }}>
-          <Button size='sm' variant='outlined' color='neutral' onClick={() => props.onRaysOperation('copy')} endDecorator={<ContentCopyIcon sx={{ fontSize: 'md' }} />} sx={{
-            backgroundColor: 'background.surface',
-            '&:hover': { backgroundColor: 'background.popup' },
-          }}>
-            Copy {props.showRaysOps}
-          </Button>
-          <Button size='sm' variant='outlined' color='success' onClick={() => props.onRaysOperation('use')} endDecorator={<TelegramIcon sx={{ fontSize: 'xl' }} />} sx={{
-            justifyContent: 'space-between',
-            backgroundColor: 'background.surface',
-            '&:hover': { backgroundColor: 'background.popup' },
-          }}>
-            Use {props.showRaysOps == 2 ? 'both' : 'all ' + props.showRaysOps} messages
-          </Button>
-        </Box>
-      )}
-
      {/*/!* Takes a full row *!/*/}
      {/*<Divider sx={{*/}
      {/*  gridColumn: '1 / -1',*/}
@@ -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();

@@ -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,
@@ -72,7 +72,7 @@ const geminiExpFree: ModelDescriptionSchema['chatPrice'] = {
 };


-// Pricing based on https://ai.google.dev/pricing (Apr 22, 2026)
+// Pricing based on https://ai.google.dev/pricing (Apr 24, 2026)

 const gemini31FlashLitePricing: ModelDescriptionSchema['chatPrice'] = {
  input: 0.25, // text/image/video; audio is $0.50 but we don't differentiate yet
@@ -262,8 +262,10 @@ const _knownGeminiModels: ({

  /// Generation 3.0

-  // 3.0 Pro (Preview) - Released November 18, 2025; DEPRECATED: shutdown March 9, 2026 (still served by API as of Apr 17, 2026)
+  // 3.0 Pro (Preview) - Released November 18, 2025; SHUT DOWN March 9, 2026 - now silently routed to gemini-3.1-pro-preview
+  // Kept hidden (still returned by API) to avoid confusing users with a silently-redirected model.
  {
+    hidden: true, // March 9, 2026: API silently routes 'gemini-3-pro-preview' to 'gemini-3.1-pro-preview' - hide to prevent user confusion
    id: 'models/gemini-3-pro-preview',
    labelOverride: 'Gemini 3 Pro Preview',
    isPreview: true,
@@ -335,6 +337,7 @@ const _knownGeminiModels: ({

  // 2.5 Pro (Stable) - Released June 17, 2025; DEPRECATED: shutdown June 17, 2026
  {
+    hidden: true, // outperformed by 3.1 Pro (1493) and even 3 Flash (1474) - deprecated in 2 months
    id: 'models/gemini-2.5-pro',
    labelOverride: 'Gemini 2.5 Pro',
    deprecated: '2026-06-17',
@@ -378,7 +381,7 @@ 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 +393,7 @@ 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 +405,7 @@ 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
@@ -412,6 +415,7 @@ const _knownGeminiModels: ({

  // 2.5 Flash
  {
+    hidden: true, // outperformed by 3 Flash Preview (1474 vs 1411) - deprecated in 2 months
    id: 'models/gemini-2.5-flash',
    labelOverride: 'Gemini 2.5 Flash',
    deprecated: '2026-06-17',
@@ -467,6 +471,7 @@ const _knownGeminiModels: ({

  // 2.5 Flash-Based: Gemini Robotics-ER 1.5 Preview - Released September 25, 2025; DEPRECATED: shutdown April 30, 2026
  {
+    hidden: true, // superseded by Robotics-ER 1.6 - shutdown April 30, 2026
    id: 'models/gemini-robotics-er-1.5-preview',
    labelOverride: 'Gemini Robotics-ER 1.5 Preview',
    isPreview: true,
@@ -573,6 +578,7 @@ const _knownGeminiModels: ({

  // 2.0 Flash - DEPRECATED: shutdown June 1, 2026 (announced Feb 18, 2026)
  {
+    hidden: true, // outclassed by all Flash models in 2.5/3.x series - shutdown in ~5 weeks
    id: 'models/gemini-2.0-flash-001',
    deprecated: '2026-06-01',
    chatPrice: gemini20FlashPricing,
@@ -580,6 +586,7 @@ const _knownGeminiModels: ({
    benchmark: { cbaElo: 1360 }, // gemini-2.0-flash-001
  },
  {
+    hidden: true, // outclassed by all Flash models in 2.5/3.x series - shutdown in ~5 weeks
    id: 'models/gemini-2.0-flash',
    symLink: 'models/gemini-2.0-flash-001',
    deprecated: '2026-06-01',
@@ -591,6 +598,7 @@ const _knownGeminiModels: ({

  // 2.0 Flash Lite - DEPRECATED: shutdown June 1, 2026 (announced Feb 18, 2026)
  {
+    hidden: true, // outclassed by 2.5/3.1 Flash-Lite - shutdown in ~5 weeks
    id: 'models/gemini-2.0-flash-lite',
    chatPrice: gemini20FlashLitePricing,
    symLink: 'models/gemini-2.0-flash-lite-001',
@@ -599,6 +607,7 @@ const _knownGeminiModels: ({
    benchmark: { cbaElo: 1310 },
  },
  {
+    hidden: true, // outclassed by 2.5/3.1 Flash-Lite - shutdown in ~5 weeks
    id: 'models/gemini-2.0-flash-lite-001',
    chatPrice: gemini20FlashLitePricing,
    deprecated: '2026-06-01',
@@ -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)
-  // [Release Notes - V3.2](https://api-docs.deepseek.com/news/news251201) - Released 2025-12-01
+  {
+    idPrefix: 'deepseek-v4-pro',
+    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)',
-    description: 'Reasoning model with Chain-of-Thought capabilities, 128K context length. Supports JSON output and function calling.',
-    contextWindow: 131072, // 128K
-    interfaces: [...IF_3, LLM_IF_OAI_Reasoning],
-    // parameterSpecs: [
-    //   { paramId: 'llmVndMiscEffort', enumValues: ['none', 'high'] }, // not supported: this model is reasoning only
-    // ],
-    maxCompletionTokens: 32768, // default, max: 65536
-    chatPrice: { input: 0.28, output: 0.42, cache: { cType: 'oai-ac', read: 0.028 } },
-    benchmark: { cbaElo: 1425 }, // deepseek-v3.2-exp-thinking
+    label: 'DeepSeek Reasoner (legacy)',
+    description: 'Legacy alias: routes to DeepSeek V4 Flash with thinking enabled. Retires 2026-07-24.',
+    contextWindow: 1_048_576,
+    interfaces: [...IF_4, LLM_IF_OAI_Reasoning],
+    maxCompletionTokens: 65536,
+    chatPrice: { input: 0.14, output: 0.28, cache: { cType: 'oai-ac', read: 0.028 } },
+    benchmark: { cbaElo: 1439 }, // lmarena: deepseek-v4-flash-thinking
+    isLegacy: true,
  },
  {
    idPrefix: 'deepseek-chat',
-    label: 'DeepSeek V3.2',
-    description: 'General-purpose model with 128K context length. Supports JSON output and function calling.',
-    contextWindow: 131072, // 128K
-    interfaces: IF_3,
-    maxCompletionTokens: 8192, // default is 4096, max is 8192
-    chatPrice: { input: 0.28, output: 0.42, cache: { cType: 'oai-ac', read: 0.028 } },
-    benchmark: { cbaElo: 1424 }, // deepseek-v3.2
+    label: 'DeepSeek Chat (legacy)',
+    description: 'Legacy alias: routes to DeepSeek V4 Flash with thinking disabled. Retires 2026-07-24.',
+    contextWindow: 1_048_576,
+    interfaces: IF_4,
+    maxCompletionTokens: 65536,
+    chatPrice: { input: 0.14, output: 0.28, cache: { cType: 'oai-ac', read: 0.028 } },
+    benchmark: { cbaElo: 1433 }, // lmarena: deepseek-v4-flash (non-thinking)
+    isLegacy: true,
  },
 ];

@@ -12,6 +12,23 @@ import { fromManualMapping, KnownModel, llmDevCheckModels_DEV, ManualMappings }
 // OpenAI Model Variants
 export const hardcodedOpenAIVariants: ModelVariantMap = {

+  // GPT-5.5 with reasoning disabled (non-thinking) - supports temperature control
+  'gpt-5.5-2026-04-23': {
+    idVariant: '::thinking-none',
+    label: 'GPT-5.5 (No-thinking)',
+    hidden: true, // hidden by default as redundant, user can unhide in settings
+    description: 'Supports temperature control for creative applications. GPT-5.5 with reasoning disabled (reasoning_effort=none).',
+    interfaces: [LLM_IF_OAI_Responses, LLM_IF_OAI_Chat, LLM_IF_OAI_Vision, LLM_IF_OAI_Fn, LLM_IF_OAI_PromptCaching], // NO LLM_IF_OAI_Reasoning, NO LLM_IF_HOTFIX_NoTemperature
+    parameterSpecs: [
+      { paramId: 'llmVndOaiEffort', enumValues: ['none', 'low', 'medium', 'high', 'xhigh'], initialValue: 'none', hidden: true }, // factory 'none', not changeable
+      { paramId: 'llmVndOaiWebSearchContext' },
+      { paramId: 'llmVndOaiVerbosity' },
+      { paramId: 'llmVndOaiImageGeneration' },
+      { paramId: 'llmVndOaiCodeInterpreter' },
+      { paramId: 'llmForceNoStream' },
+    ],
+  },
+
  // GPT-5.4 with reasoning disabled (non-thinking) - supports temperature control
  'gpt-5.4-2026-03-05': {
    idVariant: '::thinking-none',
@@ -88,6 +105,58 @@ const PS_DEEP_RESEARCH = [{ paramId: 'llmVndOaiWebSearchContext' as const, initi
 // https://platform.openai.com/docs/pricing
 export const _knownOpenAIChatModels: ManualMappings = [

+  /// GPT-5.5 series - Released April 23, 2026
+
+  // GPT-5.5
+  {
+    idPrefix: 'gpt-5.5-2026-04-23',
+    label: 'GPT-5.5 (2026-04-23)',
+    description: 'New baseline for complex production workflows. Stronger task execution, more precise tool use, more efficient reasoning with fewer tokens. 1M token context.',
+    contextWindow: 1050000,
+    maxCompletionTokens: 128000,
+    interfaces: [LLM_IF_OAI_Responses, ...IFS_CHAT_CACHE_REASON, LLM_IF_HOTFIX_NoTemperature],
+    parameterSpecs: [
+      { paramId: 'llmVndOaiEffort', enumValues: ['none', 'low', 'medium', 'high', 'xhigh'], initialValue: 'medium' }, // medium is the new default for 5.5
+      { paramId: 'llmVndOaiWebSearchContext' },
+      { paramId: 'llmVndOaiVerbosity' },
+      { paramId: 'llmVndOaiImageGeneration' },
+      { paramId: 'llmVndOaiCodeInterpreter' },
+      { paramId: 'llmForceNoStream' },
+    ],
+    chatPrice: { input: 5, cache: { cType: 'oai-ac', read: 0.5 }, output: 30 },
+    // benchmark: TBD - no CBA ELO yet
+  },
+  {
+    idPrefix: 'gpt-5.5',
+    label: 'GPT-5.5',
+    symLink: 'gpt-5.5-2026-04-23',
+  },
+
+  // GPT-5.5 Pro
+  {
+    idPrefix: 'gpt-5.5-pro-2026-04-23',
+    label: 'GPT-5.5 Pro (2026-04-23)',
+    description: 'Most capable model for complex tasks. Uses more compute for smarter, more precise responses on the hardest problems.',
+    contextWindow: 1050000,
+    maxCompletionTokens: 272000,
+    interfaces: [LLM_IF_OAI_Responses, ...IFS_CHAT_MIN, LLM_IF_OAI_Reasoning, LLM_IF_HOTFIX_NoTemperature],
+    parameterSpecs: [
+      { paramId: 'llmVndOaiEffort', enumValues: ['medium', 'high', 'xhigh'] }, // Pro: no low/none
+      { paramId: 'llmVndOaiWebSearchContext' },
+      { paramId: 'llmVndOaiVerbosity' },
+      { paramId: 'llmVndOaiImageGeneration' },
+      { paramId: 'llmForceNoStream' },
+    ],
+    chatPrice: { input: 30, output: 180 },
+    // benchmark: TBD
+  },
+  {
+    idPrefix: 'gpt-5.5-pro',
+    label: 'GPT-5.5 Pro',
+    symLink: 'gpt-5.5-pro-2026-04-23',
+  },
+
+
  /// GPT-5.4 series - Released March 5, 2026

  // GPT-5.4
@@ -250,6 +319,7 @@ export const _knownOpenAIChatModels: ManualMappings = [

  // GPT-5.2
  {
+    hidden: true, // superseded by GPT-5.4/5.5
    idPrefix: 'gpt-5.2-2025-12-11',
    label: 'GPT-5.2 (2025-12-11)',
    description: 'Most capable model for professional work and long-running agents. Improvements in general intelligence, long-context, agentic tool-calling, and vision.',
@@ -268,6 +338,7 @@ export const _knownOpenAIChatModels: ManualMappings = [
    benchmark: { cbaElo: 1441 }, // gpt-5.2-high
  },
  {
+    hidden: true, // superseded by GPT-5.4/5.5
    idPrefix: 'gpt-5.2',
    label: 'GPT-5.2',
    symLink: 'gpt-5.2-2025-12-11',
@@ -275,6 +346,7 @@ export const _knownOpenAIChatModels: ManualMappings = [

  // GPT-5.2 Codex
  {
+    hidden: true, // superseded by GPT-5.3 Codex
    idPrefix: 'gpt-5.2-codex',
    label: 'GPT-5.2 Codex',
    description: 'GPT-5.2 optimized for long-horizon, agentic coding tasks in Codex or similar environments. Supports low, medium, high, and xhigh reasoning effort settings.',
@@ -293,6 +365,7 @@ export const _knownOpenAIChatModels: ManualMappings = [

  // GPT-5.2 Chat Latest
  {
+    hidden: true, // superseded by GPT-5.3 Instant
    idPrefix: 'gpt-5.2-chat-latest',
    label: 'GPT-5.2 Instant',
    description: 'GPT-5.2 model powering ChatGPT. Fast, capable for everyday work with clear improvements in info-seeking, how-tos, technical writing.',
@@ -311,6 +384,7 @@ export const _knownOpenAIChatModels: ManualMappings = [

  // GPT-5.2 Pro
  {
+    hidden: true, // superseded by GPT-5.4/5.5 Pro
    idPrefix: 'gpt-5.2-pro-2025-12-11',
    label: 'GPT-5.2 Pro (2025-12-11)',
    description: 'Smartest and most trustworthy option for difficult questions. Uses more compute for harder thinking on complex domains like programming.',
@@ -328,6 +402,7 @@ export const _knownOpenAIChatModels: ManualMappings = [
    // benchmark: TBD
  },
  {
+    hidden: true, // superseded by GPT-5.4/5.5 Pro
    idPrefix: 'gpt-5.2-pro',
    label: 'GPT-5.2 Pro',
    symLink: 'gpt-5.2-pro-2025-12-11',
@@ -338,6 +413,7 @@ export const _knownOpenAIChatModels: ManualMappings = [

  // GPT-5.1
  {
+    hidden: true, // superseded by GPT-5.4/5.5
    idPrefix: 'gpt-5.1-2025-11-13',
    label: 'GPT-5.1 (2025-11-13)',
    description: 'The best model for coding and agentic tasks with configurable reasoning effort.',
@@ -355,6 +431,7 @@ export const _knownOpenAIChatModels: ManualMappings = [
    benchmark: { cbaElo: 1455 }, // gpt-5.1-high
  },
  {
+    hidden: true, // superseded by GPT-5.4/5.5
    idPrefix: 'gpt-5.1',
    label: 'GPT-5.1',
    symLink: 'gpt-5.1-2025-11-13',
@@ -362,6 +439,7 @@ export const _knownOpenAIChatModels: ManualMappings = [

  // GPT-5.1 Chat Latest
  {
+    hidden: true, // superseded by GPT-5.3 Instant
    idPrefix: 'gpt-5.1-chat-latest',
    label: 'GPT-5.1 Instant',
    description: 'GPT-5.1 Instant with adaptive reasoning. More conversational with improved instruction following.',
@@ -381,6 +459,7 @@ export const _knownOpenAIChatModels: ManualMappings = [

  // GPT-5.1 Codex Max
  {
+    hidden: true, // superseded by GPT-5.3 Codex
    idPrefix: 'gpt-5.1-codex-max',
    label: 'GPT-5.1 Codex Max',
    description: 'Our most intelligent coding model optimized for long-horizon, agentic coding tasks.',
@@ -398,6 +477,7 @@ export const _knownOpenAIChatModels: ManualMappings = [
  },
  // GPT-5.1 Codex
  {
+    hidden: true, // superseded by GPT-5.3 Codex
    idPrefix: 'gpt-5.1-codex',
    label: 'GPT-5.1 Codex',
    description: 'A version of GPT-5.1 optimized for agentic coding tasks in Codex or similar environments.',
@@ -415,6 +495,7 @@ export const _knownOpenAIChatModels: ManualMappings = [
  },
  // GPT-5.1 Codex Mini
  {
+    hidden: true, // superseded by GPT-5.3 Codex
    idPrefix: 'gpt-5.1-codex-mini',
    label: 'GPT-5.1 Codex Mini',
    description: 'Smaller, faster version of GPT-5.1 Codex for efficient coding tasks.',
@@ -436,6 +517,7 @@ export const _knownOpenAIChatModels: ManualMappings = [

  // GPT-5
  {
+    hidden: true, // superseded by GPT-5.4/5.5
    idPrefix: 'gpt-5-2025-08-07',
    label: 'GPT-5 (2025-08-07)',
    description: 'The best model for coding and agentic tasks across domains.',
@@ -453,6 +535,7 @@ export const _knownOpenAIChatModels: ManualMappings = [
    benchmark: { cbaElo: 1433 }, // gpt-5-high
  },
  {
+    hidden: true, // superseded by GPT-5.4/5.5
    idPrefix: 'gpt-5',
    label: 'GPT-5',
    symLink: 'gpt-5-2025-08-07',
@@ -460,6 +543,7 @@ export const _knownOpenAIChatModels: ManualMappings = [

  // GPT-5 Pro
  {
+    hidden: true, // superseded by GPT-5.4/5.5 Pro
    idPrefix: 'gpt-5-pro-2025-10-06',
    label: 'GPT-5 Pro (2025-10-06)',
    description: 'Version of GPT-5 that uses more compute to produce smarter and more precise responses. Designed for tough problems.',
@@ -471,6 +555,7 @@ export const _knownOpenAIChatModels: ManualMappings = [
    // benchmark: has not been measured yet
  },
  {
+    hidden: true, // superseded by GPT-5.4/5.5 Pro
    idPrefix: 'gpt-5-pro',
    label: 'GPT-5 Pro',
    symLink: 'gpt-5-pro-2025-10-06',
@@ -511,6 +596,7 @@ export const _knownOpenAIChatModels: ManualMappings = [

  // GPT-5 Search API
  {
+    hidden: true, // poor quality - use llmVndOaiWebSearchContext on regular models instead
    idPrefix: 'gpt-5-search-api-2025-10-14',
    label: 'GPT-5 Search API (2025-10-14)',
    description: 'Updated web search model in Chat Completions API. 60% cheaper with domain filtering support.',
@@ -522,6 +608,7 @@ export const _knownOpenAIChatModels: ManualMappings = [
    // benchmark: TBD
  },
  {
+    hidden: true, // poor quality - use llmVndOaiWebSearchContext on regular models instead
    idPrefix: 'gpt-5-search-api',
    label: 'GPT-5 Search API',
    symLink: 'gpt-5-search-api-2025-10-14',
@@ -529,6 +616,7 @@ export const _knownOpenAIChatModels: ManualMappings = [

  // GPT-5 mini
  {
+    hidden: true, // superseded by GPT-5.4 Mini
    idPrefix: 'gpt-5-mini-2025-08-07',
    label: 'GPT-5 Mini (2025-08-07)',
    description: 'A faster, more cost-efficient version of GPT-5 for well-defined tasks.',
@@ -540,6 +628,7 @@ export const _knownOpenAIChatModels: ManualMappings = [
    benchmark: { cbaElo: 1390 }, // gpt-5-mini-high
  },
  {
+    hidden: true, // superseded by GPT-5.4 Mini
    idPrefix: 'gpt-5-mini',
    label: 'GPT-5 Mini',
    symLink: 'gpt-5-mini-2025-08-07',
@@ -547,6 +636,7 @@ export const _knownOpenAIChatModels: ManualMappings = [

  // GPT-5 nano
  {
+    hidden: true, // superseded by GPT-5.4 Nano
    idPrefix: 'gpt-5-nano-2025-08-07',
    label: 'GPT-5 Nano (2025-08-07)',
    description: 'Fastest, most cost-efficient version of GPT-5 for summarization and classification tasks.',
@@ -558,6 +648,7 @@ export const _knownOpenAIChatModels: ManualMappings = [
    benchmark: { cbaElo: 1337 }, // gpt-5-nano-high
  },
  {
+    hidden: true, // superseded by GPT-5.4 Nano
    idPrefix: 'gpt-5-nano',
    label: 'GPT-5 Nano',
    symLink: 'gpt-5-nano-2025-08-07',
@@ -608,8 +699,9 @@ export const _knownOpenAIChatModels: ManualMappings = [
  // o4-mini-deep-research - (v1/responses API)
  {
    idPrefix: 'o4-mini-deep-research-2025-06-26',
-    label: 'o4 Mini Deep Research (2025-06-26)',
-    description: 'Faster, more affordable deep research model for complex, multi-step research tasks.',
+    label: 'o4 Mini Deep Research [Deprecated]',
+    isLegacy: true,
+    description: 'Faster, more affordable deep research model for complex, multi-step research tasks. [Shutdown: 2026-07-23 - migrate to GPT-5.5 with web search.]',
    contextWindow: 200000,
    maxCompletionTokens: 100000,
    interfaces: [LLM_IF_OAI_Responses, ...IFS_CHAT_CACHE_REASON, LLM_IF_HOTFIX_NoTemperature],
@@ -625,8 +717,9 @@ export const _knownOpenAIChatModels: ManualMappings = [
  /// o4-mini
  {
    idPrefix: 'o4-mini-2025-04-16',
-    label: 'o4 Mini (2025-04-16)',
-    description: 'Latest o4-mini model. Optimized for fast, effective reasoning with exceptionally efficient performance in coding and visual tasks.',
+    label: 'o4 Mini [Deprecated]',
+    isLegacy: true,
+    description: 'Latest o4-mini model. Optimized for fast, effective reasoning with exceptionally efficient performance in coding and visual tasks. [Shutdown: 2026-10-23 - migrate to GPT-5.4 Mini.]',
    contextWindow: 200000,
    maxCompletionTokens: 100000,
    interfaces: IFS_CHAT_CACHE_REASON,
@@ -643,8 +736,9 @@ export const _knownOpenAIChatModels: ManualMappings = [
  // o3-deep-research - (v1/responses API)
  {
    idPrefix: 'o3-deep-research-2025-06-26',
-    label: 'o3 Deep Research (2025-06-26)',
-    description: 'Our most powerful deep research model for complex, multi-step research tasks.',
+    label: 'o3 Deep Research [Deprecated]',
+    isLegacy: true,
+    description: 'Our most powerful deep research model for complex, multi-step research tasks. [Shutdown: 2026-07-23 - migrate to GPT-5.5 Pro with web search.]',
    contextWindow: 200000,
    maxCompletionTokens: 100000,
    interfaces: [LLM_IF_OAI_Responses, ...IFS_CHAT_CACHE_REASON, LLM_IF_HOTFIX_NoTemperature],
@@ -696,8 +790,9 @@ export const _knownOpenAIChatModels: ManualMappings = [
  // o3-mini
  {
    idPrefix: 'o3-mini-2025-01-31',
-    label: 'o3 Mini (2025-01-31)',
-    description: 'Latest o3-mini model snapshot. High intelligence at the same cost and latency targets of o1-mini. Excels at science, math, and coding tasks.',
+    label: 'o3 Mini [Deprecated]',
+    isLegacy: true,
+    description: 'Latest o3-mini model snapshot. High intelligence at the same cost and latency targets of o1-mini. Excels at science, math, and coding tasks. [Shutdown: 2026-10-23 - migrate to GPT-5.4 Mini.]',
    contextWindow: 200000,
    maxCompletionTokens: 100000,
    interfaces: [LLM_IF_OAI_Chat, LLM_IF_OAI_Fn, LLM_IF_OAI_PromptCaching, LLM_IF_OAI_Reasoning, LLM_IF_HOTFIX_StripImages],
@@ -733,8 +828,9 @@ export const _knownOpenAIChatModels: ManualMappings = [
  // o1
  {
    idPrefix: 'o1-2024-12-17',
-    label: 'o1 (2024-12-17)',
-    description: 'Previous full o-series reasoning model.',
+    label: 'o1 [Deprecated]',
+    isLegacy: true,
+    description: 'Previous full o-series reasoning model. [Shutdown: 2026-10-23 - migrate to GPT-5.5 or o3.]',
    contextWindow: 200000,
    maxCompletionTokens: 100000,
    interfaces: IFS_CHAT_CACHE_REASON,
@@ -788,8 +884,9 @@ export const _knownOpenAIChatModels: ManualMappings = [
  // GPT-4.1 nano
  {
    idPrefix: 'gpt-4.1-nano-2025-04-14',
-    label: 'GPT-4.1 Nano (2025-04-14)',
-    description: 'Fastest, most cost-effective GPT 4.1 model. Delivers exceptional performance with low latency, ideal for tasks like classification or autocompletion.',
+    label: 'GPT-4.1 Nano [Deprecated]',
+    isLegacy: true,
+    description: 'Fastest, most cost-effective GPT 4.1 model. Delivers exceptional performance with low latency, ideal for tasks like classification or autocompletion. [Shutdown: 2026-10-23 - migrate to GPT-5.4 Nano.]',
    contextWindow: 1047576,
    maxCompletionTokens: 32768,
    interfaces: IFS_CHAT_CACHE,
@@ -819,6 +916,7 @@ export const _knownOpenAIChatModels: ManualMappings = [

  // gpt-audio
  {
+    hidden: true, // superseded by GPT Audio 1.5
    idPrefix: 'gpt-audio-2025-08-28',
    label: 'GPT Audio (2025-08-28)',
    description: 'First generally available audio model. Accepts audio inputs and outputs, and can be used in the Chat Completions REST API.',
@@ -829,6 +927,7 @@ export const _knownOpenAIChatModels: ManualMappings = [
    // benchmark: TBD
  },
  {
+    hidden: true, // superseded by GPT Audio 1.5
    idPrefix: 'gpt-audio',
    label: 'GPT Audio',
    symLink: 'gpt-audio-2025-08-28',
@@ -1220,6 +1319,12 @@ export function openAIInjectVariants(acc: ModelDescriptionSchema[], model: Model


 const _manualOrderingIdPrefixes = [
+  // GPT-5.5
+  'gpt-5.5-20',
+  'gpt-5.5-pro-20',
+  'gpt-5.5-pro',
+  'gpt-5.5-chat-latest',
+  'gpt-5.5',
  // GPT-5.4
  'gpt-5.4-20',
  'gpt-5.4-pro-20',
@@ -1419,6 +1524,7 @@ export function llmOrtOaiLookup(orModelName: string): OrtVendorLookupResult | un
  // typemap to known models
  const ortOaiRefMap: Record<string, string | null> = {
    // renames
+    'gpt-5.5-chat': 'gpt-5.5-2026-04-23', // no chat-latest yet, map to snapshot
    'gpt-5.4-chat': 'gpt-5.4-2026-03-05', // no chat-latest yet, map to snapshot
    'gpt-5.3-chat': 'gpt-5.3-chat-latest',
    'gpt-5.2-chat': 'gpt-5.2-chat-latest',
@@ -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;
Author	SHA1	Message	Date
Enrico Ros	9fc0b39730	AIX: Transmit token stop errors, if provided	2026-04-24 17:08:40 -07:00
Enrico Ros	194bfe23a1	AIX: OpenAI: mark the need for roundtrip of hosted tool pairs	2026-04-24 17:08:40 -07:00
Enrico Ros	35110480ef	Beam: Fix ghost columns. Fixes #1073	2026-04-24 16:04:29 -07:00
Enrico Ros	959595e33a	Merge: smaller copy update	2026-04-24 16:04:29 -07:00
Enrico Ros	a960424dfb	Merge: copy update. Fixes #1083	2026-04-24 15:56:13 -07:00
Enrico Ros	0df6c7d08b	Merge: copy. Fixes #1083	2026-04-24 15:48:56 -07:00
Enrico Ros	65c841e7a7	Roll AIX	2026-04-24 15:23:30 -07:00
Enrico Ros	b21b8cc982	AIX: Anthropic: show refusal details, if present, as inline text	2026-04-24 15:20:10 -07:00
Enrico Ros	aa2c4f06b7	AI Inspector: compress intermediate large string fields	2026-04-24 15:19:35 -07:00
Enrico Ros	b8d7b4ec10	AIX: OpenAI: fix svs on !ma for for NS	2026-04-24 15:19:35 -07:00
Enrico Ros	c48520255a	AIX: OpenAI: fix tool reparsing for NS	2026-04-24 15:19:34 -07:00
Enrico Ros	0790da989d	Don't truncate the Beam Title on Edit. Fix #1085 part 1.	2026-04-24 15:19:34 -07:00
Enrico Ros	506d24d2fd	AIX: OpenAI Response: fix reparse of tools	2026-04-24 15:19:34 -07:00
Enrico Ros	1348dbf493	AIX: update _upstreams	2026-04-24 15:19:33 -07:00
Enrico Ros	ce677f3cd9	LLMs: OpenAI: GPT 5.5	2026-04-24 15:19:33 -07:00
Enrico Ros	39203d78e3	LLMs: OpenAI: hide lots of older models, so by default the lastest are shown	2026-04-24 15:19:33 -07:00
Enrico Ros	2ef7daf369	LLMs: Gemini: hide 3.0 Pro (silently remapped to 3.1 by Gemini). Fixes #1082	2026-04-24 15:19:33 -07:00
Enrico Ros	cff3d90613	AIX: DeepSeek V4: fix function calling	2026-04-24 05:45:53 -07:00
Enrico Ros	9f89243d7f	AIX: DeepSeek V4: fix swalling of tool parts	2026-04-24 05:45:53 -07:00
Enrico Ros	784ee9a4da	AIX: DeepSeek V4: wires and parser NS	2026-04-24 05:45:53 -07:00
Enrico Ros	678e6b8ba1	AIX: Gemini Interactions: terminate on error	2026-04-24 05:45:53 -07:00
Enrico Ros	30e301c496	BlockOpUpstreamResume: Stop/Cancel	2026-04-24 03:59:50 -07:00
Enrico Ros	b22904f6bb	AIX: Gemini Interactions: Cancel + Delete Also see: googleapis/python-genai#1971	2026-04-24 03:40:34 -07:00
Enrico Ros	3f0de7ddca	CH: Auto-Title beam chats when done. Fixes #1078	2026-04-24 03:32:04 -07:00
Enrico Ros	9a6f0f9202	AppChat: never re-open an opened beam. Fixes #1079	2026-04-24 03:24:56 -07:00
Enrico Ros	4f0bae5657	AppChat: do not re-beam or regenerate while beam is open. Fixes #1079	2026-04-24 03:19:17 -07:00
Enrico Ros	2101f06195	Roll AIX	2026-04-24 03:04:09 -07:00
Enrico Ros	6d54b5594c	Autotitle: Use natural capitalization. Fixes #1077	2026-04-24 02:48:28 -07:00
Enrico Ros	36b8e5b1df	Chat: show Stop/Cancel on streaming upstream runs	2026-04-24 02:47:17 -07:00
Enrico Ros	8252d671c7	LLMs: Gemini: Deep Research models support images	2026-04-24 02:47:13 -07:00
Enrico Ros	30d97c94aa	LLMs: DeepSeek: bits (note: vision is still not available)	2026-04-24 02:47:13 -07:00
Enrico Ros	82654a00d4	AIX: Streaming (hinting) review and Gemini Interactions API fix	2026-04-24 02:47:09 -07:00
Enrico Ros	9595f14ddc	LLM: DeepSeek V4 (flash, pro) + thinking/reasoning_effort fix	2026-04-23 23:59:09 -07:00
Enrico Ros	8c496074b2	LLMs: DeepSeek: add V4 models	2026-04-23 23:30:41 -07:00
Enrico Ros	4d097d7136	LLMs: DeepSeek: add V4 support infra	2026-04-23 23:30:34 -07:00
Enrico Ros	178619d275	AI Settings: match the defaults description. Fixes #1076	2026-04-23 23:29:20 -07:00
Enrico Ros	59c8b2538d	Merge pull request #1074 from tredondo/patch-1 chore: fix Zod 4 type-strictness issue (#1072)	2026-04-23 22:57:01 -07:00
Enrico Ros	443b72c52a	AIX: OpenAI Responses: fix Zod 4 build error in tools .catch() Bare `return;` produced `void`, which Zod 4 rejects for a `.catch()` on `z.array(...).optional()` expecting `Tool[] \| undefined`. Return `undefined` explicitly, matching the existing pattern at line 1204. Fixes #1072	2026-04-23 22:56:19 -07:00
Enrico Ros	ae13abef45	Nobody can tell @fredliubojin what to resume	2026-04-23 22:22:16 -07:00
Ted Robertson	83ae02ef9b	chore: fix Zod 4 type-strictness issue (#1072 )	2026-04-23 19:51:49 -07:00