fix: media timeouts + Qwen-Image-2.1 fixes + unified Team chat search

packages/core/src/llm/dashscope.ts

@@ -112,7 +112,7 @@ export class DashScopeProvider extends OpenAIProvider {
       method: 'POST',
       headers: { 'Content-Type': 'application/json', Authorization: authorization },
       body: JSON.stringify(body),
-      signal: AbortSignal.timeout(120_000),
+      signal: AbortSignal.timeout(this.imageGenerationTimeoutMs),
     });
 
     if (!res.ok) {
@@ -169,7 +169,7 @@ export class DashScopeProvider extends OpenAIProvider {
       method: 'POST',
       headers: { 'Content-Type': 'application/json', Authorization: authorization },
       body: JSON.stringify(body),
-      signal: AbortSignal.timeout(60_000),
+      signal: AbortSignal.timeout(this.ttsTimeoutMs),
     });
 
     if (!res.ok) {

packages/core/src/llm/fireworks.ts

@@ -64,7 +64,7 @@ export class FireworksProvider extends OpenAIProvider {
         Authorization: authorization,
       },
       body: JSON.stringify(body),
-      signal: AbortSignal.timeout(120_000),
+      signal: AbortSignal.timeout(this.imageGenerationTimeoutMs),
     });
 
     if (!res.ok) {

packages/core/src/llm/markus-provider.ts

@@ -303,6 +303,13 @@ export class MarkusProvider implements MultiModalProviderInterface {
   private maxTokens?: number;
   private chatTimeoutMs: number;
   private streamTimeoutMs: number;
+  /** Generative-media client timeouts; each independently configurable per
+   *  provider (defaults match the historical hardcoded ceilings). */
+  private imageGenerationTimeoutMs: number;
+  private ttsTimeoutMs: number;
+  private sttTimeoutMs: number;
+  private videoGenerationTimeoutMs: number;
+  private decisionTimeoutMs: number;
   private cuCache = new CUCache();
   /** Hub-served geo-aware model catalog URL. */
   private modelsUrl = '';
@@ -330,6 +337,14 @@ export class MarkusProvider implements MultiModalProviderInterface {
     // Stream idle is independent of chat timeoutMs — never inherit a lower chat
     // timeout (e.g. 90s) or long reasoning / sparse SSE gaps abort mid-reply.
     this.streamTimeoutMs = config?.streamTimeoutMs ?? STREAM_TIMEOUT_MS;
+    // Generative-media client timeouts; each independently configurable per
+    // provider. Default: generous 10min — media synthesis through OpenRouter
+    // and self-hosted servers can take minutes on first use.
+    this.imageGenerationTimeoutMs = config?.imageGenerationTimeoutMs ?? 600_000;
+    this.ttsTimeoutMs = config?.ttsTimeoutMs ?? 600_000;
+    this.sttTimeoutMs = config?.sttTimeoutMs ?? 600_000;
+    this.videoGenerationTimeoutMs = config?.videoGenerationTimeoutMs ?? 600_000;
+    this.decisionTimeoutMs = config?.decisionTimeoutMs ?? 60_000;
     this.applyRetryConfig(config);
     this.modelsUrl = config?.modelsUrl ?? process.env['MARKUS_MODELS_URL'] ?? '';
     this.hubUrl = config?.hubUrl ?? process.env['MARKUS_HUB_URL'] ?? '';
@@ -361,6 +376,11 @@ export class MarkusProvider implements MultiModalProviderInterface {
     if (config.hubToken !== undefined) this.hubToken = config.hubToken;
     if (config.timeoutMs) this.chatTimeoutMs = config.timeoutMs;
     if (config.streamTimeoutMs) this.streamTimeoutMs = config.streamTimeoutMs;
+    if (config.imageGenerationTimeoutMs !== undefined) this.imageGenerationTimeoutMs = config.imageGenerationTimeoutMs;
+    if (config.ttsTimeoutMs !== undefined) this.ttsTimeoutMs = config.ttsTimeoutMs;
+    if (config.sttTimeoutMs !== undefined) this.sttTimeoutMs = config.sttTimeoutMs;
+    if (config.videoGenerationTimeoutMs !== undefined) this.videoGenerationTimeoutMs = config.videoGenerationTimeoutMs;
+    if (config.decisionTimeoutMs !== undefined) this.decisionTimeoutMs = config.decisionTimeoutMs;
     // Retry knobs follow the same precedence on re-configure (env re-read too).
     this.applyRetryConfig(config);
   }
@@ -1549,7 +1569,7 @@ export class MarkusProvider implements MultiModalProviderInterface {
         'X-Title': 'Markus',
       },
       body: JSON.stringify(body),
-      signal: AbortSignal.timeout(60_000),
+      signal: AbortSignal.timeout(this.decisionTimeoutMs),
     });
 
     if (!res.ok) {
@@ -1608,7 +1628,7 @@ export class MarkusProvider implements MultiModalProviderInterface {
       method: 'POST',
       headers: { 'Content-Type': 'application/json', Authorization: this.bearerOpenRouter() },
       body: JSON.stringify(body),
-      signal: AbortSignal.timeout(180_000),
+      signal: AbortSignal.timeout(this.imageGenerationTimeoutMs),
     });
     if (!res.ok) {
       const errText = await res.text();
@@ -1705,7 +1725,7 @@ export class MarkusProvider implements MultiModalProviderInterface {
       headers: { 'Content-Type': 'application/json', Authorization: this.bearerOpenRouter() },
       body: JSON.stringify(body),
       // Long narration + MiniMax/Deepgram synthesis often exceeds 60s.
-      signal: AbortSignal.timeout(180_000),
+      signal: AbortSignal.timeout(this.ttsTimeoutMs),
     });
     if (!res.ok) {
       const errText = await res.text();
@@ -1758,7 +1778,7 @@ export class MarkusProvider implements MultiModalProviderInterface {
       method: 'POST',
       headers: { 'Content-Type': 'application/json', Authorization: this.bearerOpenRouter() },
       body: JSON.stringify(body),
-      signal: AbortSignal.timeout(120_000),
+      signal: AbortSignal.timeout(this.sttTimeoutMs),
     });
     if (!res.ok) {
       const errText = await res.text();
@@ -1823,7 +1843,7 @@ export class MarkusProvider implements MultiModalProviderInterface {
       method: 'POST',
       headers: { 'Content-Type': 'application/json', Authorization: this.bearerOpenRouter() },
       body: JSON.stringify(body),
-      signal: AbortSignal.timeout(120_000),
+      signal: AbortSignal.timeout(this.videoGenerationTimeoutMs),
     });
     if (!createRes.ok) {
       const errText = await createRes.text();

packages/core/src/llm/minimax.ts

@@ -126,7 +126,7 @@ export class MiniMaxProvider extends OpenAIProvider {
       method: 'POST',
       headers: { 'Content-Type': 'application/json', Authorization: authorization },
       body: JSON.stringify(body),
-      signal: AbortSignal.timeout(120_000),
+      signal: AbortSignal.timeout(this.imageGenerationTimeoutMs),
     });
 
     if (!res.ok) {
@@ -177,7 +177,7 @@ export class MiniMaxProvider extends OpenAIProvider {
       method: 'POST',
       headers: { 'Content-Type': 'application/json', Authorization: authorization },
       body: JSON.stringify(body),
-      signal: AbortSignal.timeout(180_000),
+      signal: AbortSignal.timeout(this.ttsTimeoutMs),
     });
 
     if (!res.ok) {
@@ -224,7 +224,7 @@ export class MiniMaxProvider extends OpenAIProvider {
       method: 'POST',
       headers: { 'Content-Type': 'application/json', Authorization: authorization },
       body: JSON.stringify(body),
-      signal: AbortSignal.timeout(120_000),
+      signal: AbortSignal.timeout(this.videoGenerationTimeoutMs),
     });
 
     if (!createRes.ok) {

packages/core/src/llm/openai.ts

@@ -102,6 +102,11 @@ export class OpenAIProvider implements MultiModalProviderInterface {
   protected maxTokens: number;
   protected chatTimeoutMs: number;
   protected streamTimeoutMs: number;
+  protected imageGenerationTimeoutMs: number;
+  protected ttsTimeoutMs: number;
+  protected sttTimeoutMs: number;
+  protected videoGenerationTimeoutMs: number;
+  protected decisionTimeoutMs: number;
   protected tokenResolver?: TokenResolver;
   /** Explicit decisions endpoint; when unset it is resolved from `baseUrl`. */
   protected decisionsUrl?: string;
@@ -122,6 +127,16 @@ export class OpenAIProvider implements MultiModalProviderInterface {
     this.chatTimeoutMs = config?.timeoutMs ?? 90_000;
     // Idle gap between chunks (reset on data). Independent of chat timeoutMs.
     this.streamTimeoutMs = config?.streamTimeoutMs ?? 180_000;
+    // Generative media (image/video/audio) is naturally slower than chat and
+    // each costs real compute — give every modality an independently configured
+    // client timeout (per provider), instead of one hardcoded ceiling for all.
+    // Default: generous 10min — local/self-hosted servers (diffusers, vLLM)
+    // must cold-load weights on the first request and can take minutes.
+    this.imageGenerationTimeoutMs = config?.imageGenerationTimeoutMs ?? 600_000;
+    this.ttsTimeoutMs = config?.ttsTimeoutMs ?? 600_000;
+    this.sttTimeoutMs = config?.sttTimeoutMs ?? 600_000;
+    this.videoGenerationTimeoutMs = config?.videoGenerationTimeoutMs ?? 600_000;
+    this.decisionTimeoutMs = config?.decisionTimeoutMs ?? 60_000;
     this.decisionsUrl = config?.decisionsUrl;
     this.tokenResolver = tokenResolver;
   }
@@ -133,6 +148,11 @@ export class OpenAIProvider implements MultiModalProviderInterface {
     if (config.maxTokens) this.maxTokens = config.maxTokens;
     if (config.timeoutMs) this.chatTimeoutMs = config.timeoutMs;
     if (config.streamTimeoutMs) this.streamTimeoutMs = config.streamTimeoutMs;
+    if (config.imageGenerationTimeoutMs !== undefined) this.imageGenerationTimeoutMs = config.imageGenerationTimeoutMs;
+    if (config.ttsTimeoutMs !== undefined) this.ttsTimeoutMs = config.ttsTimeoutMs;
+    if (config.sttTimeoutMs !== undefined) this.sttTimeoutMs = config.sttTimeoutMs;
+    if (config.videoGenerationTimeoutMs !== undefined) this.videoGenerationTimeoutMs = config.videoGenerationTimeoutMs;
+    if (config.decisionTimeoutMs !== undefined) this.decisionTimeoutMs = config.decisionTimeoutMs;
     if (config.decisionsUrl) this.decisionsUrl = config.decisionsUrl;
   }
 
@@ -588,7 +608,7 @@ export class OpenAIProvider implements MultiModalProviderInterface {
         'X-Title': 'Markus',
       },
       body: JSON.stringify(body),
-      signal: AbortSignal.timeout(60_000),
+      signal: AbortSignal.timeout(this.decisionTimeoutMs),
     });
 
     if (!res.ok) {
@@ -630,7 +650,7 @@ export class OpenAIProvider implements MultiModalProviderInterface {
       method: 'POST',
       headers: { 'Content-Type': 'application/json', Authorization: authorization },
       body: JSON.stringify(body),
-      signal: AbortSignal.timeout(120_000),
+      signal: AbortSignal.timeout(this.imageGenerationTimeoutMs),
     });
 
     if (!res.ok) {
@@ -675,7 +695,7 @@ export class OpenAIProvider implements MultiModalProviderInterface {
       method: 'POST',
       headers: { 'Content-Type': 'application/json', Authorization: authorization },
       body: JSON.stringify(body),
-      signal: AbortSignal.timeout(180_000),
+      signal: AbortSignal.timeout(this.ttsTimeoutMs),
     });
 
     if (!res.ok) {
@@ -707,7 +727,7 @@ export class OpenAIProvider implements MultiModalProviderInterface {
       method: 'POST',
       headers: { Authorization: authorization },
       body: formData,
-      signal: AbortSignal.timeout(120_000),
+      signal: AbortSignal.timeout(this.sttTimeoutMs),
     });
 
     if (!res.ok) {

packages/core/src/llm/router.ts

@@ -1333,8 +1333,16 @@ export class LLMRouter {
    * Resolve a provider instance for a non-text modality (image_generation, audio_tts, etc.).
    * Returns the provider and the assigned model name WITHOUT mutating provider state.
    * The caller is responsible for passing `model` into the API call options.
+   *
+   * For non-text capabilities this NEVER falls back to the global text routing
+   * default model id (deepseek/gpt-4o/…): posting a chat id to an image/tts/stt
+   * endpoint yields a confusing upstream 404 ("model not found" / "not served for
+   * this capability"). Unassigned capabilities resolve to the provider only; the
+   * provider's own modality-appropriate default (or an explicit per-call model)
+   * is used instead.
    */
   resolveModalityProvider(capabilityType: ModelCapabilityType): { provider: MultiModalProviderInterface; model?: string } | undefined {
+    const isText = capabilityType === 'text';
     const assignment = this._capabilityRouting.assignments[capabilityType];
     if (assignment) {
       const provider = this.providers.get(assignment.provider);
@@ -1349,17 +1357,51 @@ export class LLMRouter {
       }
     }
 
-    // Fallback: try routingDefaultModel, then defaultProvider
+    // Fallback: try routingDefaultModel, then defaultProvider. Only text routing
+    // carries the default text model id; non-text capabilities must NOT reuse it
+    // (see class doc above) — except when the routing default model provably
+    // serves the capability (declared in its catalog entry), so a sensible
+    // media default like "image-01" still resolves instead of posting a chat id
+    // like "deepseek/deepseek-v4-flash-0731" to an image endpoint (upstream 404).
     if (this._routingDefaultModel) {
       const p = this.providers.get(this._routingDefaultModel.provider);
       if (p && this.isAvailable(this._routingDefaultModel.provider)) {
-        return { provider: p as MultiModalProviderInterface, model: this._routingDefaultModel.model };
+        const model = isText || this.routingDefaultModelServesCapability(capabilityType)
+          ? this._routingDefaultModel.model
+          : undefined;
+        return { provider: p as MultiModalProviderInterface, ...(model ? { model } : {}) };
       }
     }
     const p = this.providers.get(this.defaultProvider);
     return p && this.isAvailable(this.defaultProvider) ? { provider: p as MultiModalProviderInterface } : undefined;
   }
 
+  /**
+   * True when the routing default model's catalog entry declares the requested
+   * capability. Used to decide whether a non-text fallback may carry the model
+   * id — a text/chat model must never be POSTed to a media endpoint.
+   */
+  private routingDefaultModelServesCapability(capabilityType: ModelCapabilityType): boolean {
+    const r = this._routingDefaultModel;
+    if (!r) return false;
+    try {
+      const entry = this.getProviderModels(r.provider).find(m => m.id === r.model);
+      if (!entry) return false;
+      switch (capabilityType) {
+        case 'image_generation': return !!entry.capabilities?.includes('imageGeneration');
+        case 'audio_tts': return !!entry.capabilities?.includes('tts');
+        case 'audio_stt': return !!entry.capabilities?.includes('stt');
+        case 'video_generation': return !!entry.capabilities?.includes('videoGeneration');
+        case 'decision': return !!entry.capabilities?.includes('decision');
+        case 'image_recognition':
+          return !!entry.inputTypes?.includes('image') || !!entry.capabilities?.includes('vision');
+        default: return false;
+      }
+    } catch {
+      return false;
+    }
+  }
+
   /**
    * Look up any registered provider by name for one-shot tool overrides
    * (`provider=` + `model=` on generate_image / text_to_speech / …).

packages/core/src/tools/multimodal.ts

@@ -238,6 +238,24 @@ function resolveEffectiveCandidates(
     };
   }
   if (agentModel) {
+    // Qualified "provider/model" ids (e.g. "openai/gpt-image-1") must NOT be
+    // blindly stamped onto every candidate model — if the caller passes a
+    // chat id like "deepseek/deepseek-v4-flash-0731" for image_generation, we
+    // would otherwise POST a text model to an image endpoint and surface a
+    // confusing upstream 404 ("model not found"). Split the prefix into an
+    // explicit provider so the modality-support check below applies and the
+    // error names the real problem (provider does not serve this capability).
+    const slash = agentModel.indexOf('/');
+    if (slash > 0) {
+      const prefix = agentModel.slice(0, slash);
+      const bare = agentModel.slice(slash + 1);
+      if (
+        bare
+        && (ctx.resolveProvider?.(prefix) || candidates.some(c => c.name === prefix))
+      ) {
+        return resolveEffectiveCandidates(candidates, prefix, bare, ctx, method);
+      }
+    }
     return { candidates: candidates.map(c => ({ ...c, model: agentModel })) };
   }
   return { candidates };
@@ -865,7 +883,13 @@ export function createMultiModalTools(ctx: MultiModalToolsContext): AgentToolHan
         log.error(`Image generation failed on all ${candidates.length} provider(s)`);
         return toolErr(
           `Image generation failed (tried: ${tried}): ${lastError instanceof Error ? lastError.message : String(lastError)}`,
-          { hint: ROUTING_HINT, tried_models: candidates.map(c => c.model).filter(Boolean) },
+          {
+            hint:
+              `${ROUTING_HINT} ` +
+              `If the error mentions "model not found" / "not served for this capability", the selected model is a ` +
+              `text/chat model, not an image model — pick from llm_get_capability_routing.usable_models.image_generation.`,
+            tried_models: candidates.map(c => c.model).filter(Boolean),
+          },
         );
       },
     },