@mixio-pro/kalaasetu-mcp 2.0.11-beta → 2.1.1-beta

package/src/tools/image-to-video.ts

@@ -179,357 +179,366 @@ export const imageToVideo = {
       };
     },
   ) {
-    return safeToolExecute(async () => {
-      const projectId = "mixio-pro";
-      const location = "us-central1";
-      const modelId = args.model_id || "veo-3.1-fast-generate-preview";
-
-      // Validate and parse duration_seconds - snap to nearest 4, 6, or 8
-      let durationSeconds = parseInt(args.duration_seconds || "6");
-      if (isNaN(durationSeconds)) durationSeconds = 6;
-
-      const validDurations = [4, 6, 8];
-      // Find nearest valid duration
-      durationSeconds = validDurations.reduce((prev, curr) => {
-        return Math.abs(curr - durationSeconds) <
-          Math.abs(prev - durationSeconds)
-          ? curr
-          : prev;
-      });
-
-      // Tie-breaking: if equidistant (e.g. 5), the reduce above keeps the first one (4) because < is strict.
-      // However, user requested "nearest duration with the ceil", effectively meaning round up if equidistant.
-      // Let's explicitly handle the equidistant cases or just use a custom finder.
-      // 5 -> equidistant to 4 and 6. "With ceil" implies 6.
-      // 7 -> equidistant to 6 and 8. "With ceil" implies 8.
-
-      // Simpler logic for these specific values:
-      if (
-        durationSeconds === 4 &&
-        parseInt(args.duration_seconds || "6") === 5
-      ) {
-        durationSeconds = 6;
-      }
-      if (
-        durationSeconds === 6 &&
-        parseInt(args.duration_seconds || "6") === 7
-      ) {
-        durationSeconds = 8;
-      }
-      // Stream diagnostic info about auth
-      let token: string;
-      try {
-        if (context?.streamContent) {
-          await context.streamContent({
-            type: "text" as const,
-            text: `[Vertex] Authenticating with Google Cloud (project: ${projectId}, location: ${location})...`,
-          });
+    return safeToolExecute(
+      async () => {
+        const projectId = "mixio-pro";
+        const location = "us-central1";
+        const modelId = args.model_id || "veo-3.1-fast-generate-preview";
+
+        // Validate and parse duration_seconds - snap to nearest 4, 6, or 8
+        let durationSeconds = parseInt(args.duration_seconds || "6");
+        if (isNaN(durationSeconds)) durationSeconds = 6;
+
+        const validDurations = [4, 6, 8];
+        // Find nearest valid duration
+        durationSeconds = validDurations.reduce((prev, curr) => {
+          return Math.abs(curr - durationSeconds) <
+            Math.abs(prev - durationSeconds)
+            ? curr
+            : prev;
+        });
+
+        // Tie-breaking: if equidistant (e.g. 5), the reduce above keeps the first one (4) because < is strict.
+        // However, user requested "nearest duration with the ceil", effectively meaning round up if equidistant.
+        // Let's explicitly handle the equidistant cases or just use a custom finder.
+        // 5 -> equidistant to 4 and 6. "With ceil" implies 6.
+        // 7 -> equidistant to 6 and 8. "With ceil" implies 8.
+
+        // Simpler logic for these specific values:
+        if (
+          durationSeconds === 4 &&
+          parseInt(args.duration_seconds || "6") === 5
+        ) {
+          durationSeconds = 6;
         }
-        token = await getGoogleAccessToken();
-        if (context?.streamContent) {
-          await context.streamContent({
-            type: "text" as const,
-            text: `[Vertex] ✓ Authentication successful. Token acquired.`,
-          });
+        if (
+          durationSeconds === 6 &&
+          parseInt(args.duration_seconds || "6") === 7
+        ) {
+          durationSeconds = 8;
         }
-      } catch (authError: any) {
-        const errorMsg = authError?.message || String(authError);
-        if (context?.streamContent) {
-          await context.streamContent({
-            type: "text" as const,
-            text: `[Vertex] ✗ Authentication FAILED: ${errorMsg}. Check GOOGLE_APPLICATION_CREDENTIALS or run 'gcloud auth application-default login'.`,
-          });
-        }
-        throw new Error(`Google Cloud authentication failed: ${errorMsg}`);
-      }
-
-      const fetchUrl = `https://${location}-aiplatform.googleapis.com/v1/projects/${projectId}/locations/${location}/publishers/google/models/${modelId}:fetchPredictOperation`;
-
-      // If resuming, reconstruct the full operation path from the UUID
-      let operationName: string | undefined;
-      if (args.resume_endpoint) {
-        // Support both UUID-only and full path formats
-        if (args.resume_endpoint.includes("/")) {
-          operationName = args.resume_endpoint; // Already a full path
-        } else {
-          // Reconstruct full path from UUID
-          operationName = `projects/${projectId}/locations/${location}/publishers/google/models/${modelId}/operations/${args.resume_endpoint}`;
+        // Stream diagnostic info about auth
+        let token: string;
+        try {
+          if (context?.streamContent) {
+            await context.streamContent({
+              type: "text" as const,
+              text: `[Vertex] Authenticating with Google Cloud (project: ${projectId}, location: ${location})...`,
+            });
+          }
+          token = await getGoogleAccessToken();
+          if (context?.streamContent) {
+            await context.streamContent({
+              type: "text" as const,
+              text: `[Vertex] ✓ Authentication successful. Token acquired.`,
+            });
+          }
+        } catch (authError: any) {
+          const errorMsg = authError?.message || String(authError);
+          if (context?.streamContent) {
+            await context.streamContent({
+              type: "text" as const,
+              text: `[Vertex] ✗ Authentication FAILED: ${errorMsg}. Check GOOGLE_APPLICATION_CREDENTIALS or run 'gcloud auth application-default login'.`,
+            });
+          }
+          throw new Error(`Google Cloud authentication failed: ${errorMsg}`);
         }
-      }
-      let current: any;
 
-      if (!operationName) {
-        if (!args.prompt) {
-          throw new Error("prompt is required when starting a new generation.");
-        }
+        const fetchUrl = `https://${location}-aiplatform.googleapis.com/v1/projects/${projectId}/locations/${location}/publishers/google/models/${modelId}:fetchPredictOperation`;
 
-        if (context?.streamContent) {
-          await context.streamContent({
-            type: "text" as const,
-            text: `[Vertex] Submitting video generation request to Veo model: ${modelId}...`,
-          });
+        // If resuming, reconstruct the full operation path from the UUID
+        let operationName: string | undefined;
+        if (args.resume_endpoint) {
+          // Support both UUID-only and full path formats
+          if (args.resume_endpoint.includes("/")) {
+            operationName = args.resume_endpoint; // Already a full path
+          } else {
+            // Reconstruct full path from UUID
+            operationName = `projects/${projectId}/locations/${location}/publishers/google/models/${modelId}/operations/${args.resume_endpoint}`;
+          }
         }
+        let current: any;
 
-        const url = `https://${location}-aiplatform.googleapis.com/v1/projects/${projectId}/locations/${location}/publishers/google/models/${modelId}:predictLongRunning`;
+        if (!operationName) {
+          if (!args.prompt) {
+            throw new Error(
+              "prompt is required when starting a new generation.",
+            );
+          }
 
-        let imagePart: any = undefined;
-        if (args.image_path) {
-          const { data, mimeType } = await fileToBase64(args.image_path);
-          imagePart = {
-            image: {
-              bytesBase64Encoded: data,
-              mimeType,
-            },
-          };
-        }
+          if (context?.streamContent) {
+            await context.streamContent({
+              type: "text" as const,
+              text: `[Vertex] Submitting video generation request to Veo model: ${modelId}...`,
+            });
+          }
 
-        let lastFramePart: any = undefined;
-        if (args.last_frame_path) {
-          const { data, mimeType } = await fileToBase64(args.last_frame_path);
-          lastFramePart = {
-            lastFrame: {
-              bytesBase64Encoded: data,
-              mimeType,
-            },
-          };
-        }
+          const url = `https://${location}-aiplatform.googleapis.com/v1/projects/${projectId}/locations/${location}/publishers/google/models/${modelId}:predictLongRunning`;
+
+          let imagePart: any = undefined;
+          if (args.image_path) {
+            const { data, mimeType } = await fileToBase64(args.image_path);
+            imagePart = {
+              image: {
+                bytesBase64Encoded: data,
+                mimeType,
+              },
+            };
+          }
+
+          let lastFramePart: any = undefined;
+          if (args.last_frame_path) {
+            const { data, mimeType } = await fileToBase64(args.last_frame_path);
+            lastFramePart = {
+              lastFrame: {
+                bytesBase64Encoded: data,
+                mimeType,
+              },
+            };
+          }
 
-        let referenceImages: any[] | undefined = undefined;
-        if (args.reference_images) {
-          let refImages: string[];
-          if (typeof args.reference_images === "string") {
-            if (
-              args.reference_images.startsWith("[") &&
-              args.reference_images.endsWith("]")
-            ) {
-              try {
-                refImages = JSON.parse(args.reference_images);
-              } catch {
-                throw new Error("Invalid reference_images format");
+          let referenceImages: any[] | undefined = undefined;
+          if (args.reference_images) {
+            let refImages: string[];
+            if (typeof args.reference_images === "string") {
+              if (
+                args.reference_images.startsWith("[") &&
+                args.reference_images.endsWith("]")
+              ) {
+                try {
+                  refImages = JSON.parse(args.reference_images);
+                } catch {
+                  throw new Error("Invalid reference_images format");
+                }
+              } else {
+                refImages = [args.reference_images];
               }
+            } else if (Array.isArray(args.reference_images)) {
+              refImages = args.reference_images;
             } else {
-              refImages = [args.reference_images];
+              throw new Error(
+                "Invalid reference_images: must be array or string",
+              );
             }
-          } else if (Array.isArray(args.reference_images)) {
-            refImages = args.reference_images;
-          } else {
-            throw new Error(
-              "Invalid reference_images: must be array or string",
-            );
-          }
 
-          if (refImages.length > 0) {
-            referenceImages = await Promise.all(
-              refImages.slice(0, 3).map(async (p) => {
-                const { data, mimeType } = await fileToBase64(p);
-                return {
-                  image: {
-                    bytesBase64Encoded: data,
-                    mimeType,
-                  },
-                  referenceType: "asset",
-                };
-              }),
-            );
+            if (refImages.length > 0) {
+              referenceImages = await Promise.all(
+                refImages.slice(0, 3).map(async (p) => {
+                  const { data, mimeType } = await fileToBase64(p);
+                  return {
+                    image: {
+                      bytesBase64Encoded: data,
+                      mimeType,
+                    },
+                    referenceType: "asset",
+                  };
+                }),
+              );
+            }
           }
-        }
-
-        const personGeneration =
-          args.person_generation ||
-          (args.image_path ? "allow_adult" : "allow_all");
 
-        // Apply prompt enhancement logic
-        let enhancedPrompt = args.prompt;
-        let enhancedNegativePrompt = args.negative_prompt;
+          const personGeneration =
+            args.person_generation ||
+            (args.image_path ? "allow_adult" : "allow_all");
 
-        // Determine which preset to use
-        let presetToUse = args.enhancer_preset;
+          // Apply prompt enhancement logic
+          let enhancedPrompt = args.prompt;
+          let enhancedNegativePrompt = args.negative_prompt;
 
-        // If auto_enhance is true and no preset specified, default to 'veo'
-        if (args.auto_enhance === true && !presetToUse) {
-          presetToUse = "veo";
-        }
+          // Determine which preset to use
+          let presetToUse = args.enhancer_preset;
 
-        // Disable enhancement if auto_enhance is explicitly false
-        if (args.auto_enhance === false) {
-          presetToUse = undefined;
-        }
+          // If auto_enhance is true and no preset specified, default to 'veo'
+          if (args.auto_enhance === true && !presetToUse) {
+            presetToUse = "veo";
+          }
 
-        if (presetToUse && args.prompt) {
-          // Use LLM-based enhancement for 'veo' preset
-          if (presetToUse === "veo") {
-            const { enhancePromptWithLLM, isLLMEnhancerAvailable } =
-              await import("../utils/llm-prompt-enhancer");
-
-            if (isLLMEnhancerAvailable()) {
-              if (context?.streamContent) {
-                await context.streamContent({
-                  type: "text" as const,
-                  text: `[VEO] Enhancing prompt with Gemini for optimal Veo 3.1 generation...`,
-                });
-              }
+          // Disable enhancement if auto_enhance is explicitly false
+          if (args.auto_enhance === false) {
+            presetToUse = undefined;
+          }
 
-              try {
-                enhancedPrompt = await enhancePromptWithLLM(args.prompt, "veo");
-                context?.log?.info(
-                  `LLM-enhanced prompt for Veo: "${args.prompt}" → "${enhancedPrompt}"`,
-                );
+          if (presetToUse && args.prompt) {
+            // Use LLM-based enhancement for 'veo' preset
+            if (presetToUse === "veo") {
+              const { enhancePromptWithLLM, isLLMEnhancerAvailable } =
+                await import("../utils/llm-prompt-enhancer");
 
+              if (isLLMEnhancerAvailable()) {
                 if (context?.streamContent) {
                   await context.streamContent({
                     type: "text" as const,
-                    text: `[VEO] ✓ Prompt enhanced. Length: ${args.prompt.length} → ${enhancedPrompt.length} chars`,
+                    text: `[VEO] Enhancing prompt with Gemini for optimal Veo 3.1 generation...`,
                   });
                 }
-              } catch (err: any) {
+
+                try {
+                  enhancedPrompt = await enhancePromptWithLLM(
+                    args.prompt,
+                    "veo",
+                  );
+                  context?.log?.info(
+                    `LLM-enhanced prompt for Veo: "${args.prompt}" → "${enhancedPrompt}"`,
+                  );
+
+                  if (context?.streamContent) {
+                    await context.streamContent({
+                      type: "text" as const,
+                      text: `[VEO] ✓ Prompt enhanced. Length: ${args.prompt.length} → ${enhancedPrompt.length} chars`,
+                    });
+                  }
+                } catch (err: any) {
+                  context?.log?.info(
+                    `LLM enhancement failed, using original: ${err.message}`,
+                  );
+                }
+              } else {
                 context?.log?.info(
-                  `LLM enhancement failed, using original: ${err.message}`,
+                  "GEMINI_API_KEY not set, skipping Veo LLM enhancement",
                 );
               }
             } else {
-              context?.log?.info(
-                "GEMINI_API_KEY not set, skipping Veo LLM enhancement",
-              );
-            }
-          } else {
-            // Fall back to static string-based enhancement for other presets
-            const enhancer = resolveEnhancer(presetToUse);
-            if (enhancer.hasTransformations()) {
-              enhancedPrompt = enhancer.enhance(args.prompt);
-              // Apply negative elements if not already set
-              const negatives = enhancer.getNegativeElements();
-              if (negatives && !enhancedNegativePrompt) {
-                enhancedNegativePrompt = negatives;
+              // Fall back to static string-based enhancement for other presets
+              const enhancer = resolveEnhancer(presetToUse);
+              if (enhancer.hasTransformations()) {
+                enhancedPrompt = enhancer.enhance(args.prompt);
+                // Apply negative elements if not already set
+                const negatives = enhancer.getNegativeElements();
+                if (negatives && !enhancedNegativePrompt) {
+                  enhancedNegativePrompt = negatives;
+                }
               }
             }
           }
+
+          const instances: any[] = [
+            {
+              prompt: enhancedPrompt,
+              ...(imagePart || {}),
+              ...(lastFramePart || {}),
+              ...(referenceImages ? { referenceImages } : {}),
+            },
+          ];
+
+          const parameters: any = {
+            aspectRatio: args.aspect_ratio || "9:16",
+            durationSeconds: durationSeconds,
+            resolution: args.resolution || "720p",
+            negativePrompt: enhancedNegativePrompt,
+            generateAudio: args.generate_audio || false,
+            personGeneration,
+          };
+
+          const res = await fetch(url, {
+            method: "POST",
+            headers: {
+              Authorization: `Bearer ${token}`,
+              "Content-Type": "application/json",
+            },
+            body: JSON.stringify({ instances, parameters }),
+          });
+
+          if (!res.ok) {
+            const text = await res.text();
+            throw new Error(`Vertex request failed: ${res.status} ${text}`);
+          }
+
+          const op = (await res.json()) as any;
+          operationName = op.name || op.operation || "";
+          current = op;
         }
 
-        const instances: any[] = [
-          {
-            prompt: enhancedPrompt,
-            ...(imagePart || {}),
-            ...(lastFramePart || {}),
-            ...(referenceImages ? { referenceImages } : {}),
-          },
-        ];
-
-        const parameters: any = {
-          aspectRatio: args.aspect_ratio || "9:16",
-          durationSeconds: durationSeconds,
-          resolution: args.resolution || "720p",
-          negativePrompt: enhancedNegativePrompt,
-          generateAudio: args.generate_audio || false,
-          personGeneration,
-        };
-
-        const res = await fetch(url, {
-          method: "POST",
-          headers: {
-            Authorization: `Bearer ${token}`,
-            "Content-Type": "application/json",
-          },
-          body: JSON.stringify({ instances, parameters }),
-        });
+        if (!operationName) {
+          throw new Error(
+            "Vertex did not return an operation name for long-running request",
+          );
+        }
+
+        // Construct the composite resume_endpoint: fetchUrl||operationName||outputPath
+        // This allows get_generation_status to use the URL directly and preserve output_path
+        const outputPathPart = args.output_path || "";
+        const compositeResumeEndpoint = `${fetchUrl}||${operationName}||${outputPathPart}`;
 
-        if (!res.ok) {
-          const text = await res.text();
-          throw new Error(`Vertex request failed: ${res.status} ${text}`);
+        // Stream the resume_endpoint to the LLM immediately (before polling starts)
+        // This way the LLM has it even if MCP client times out during polling
+        if (context?.streamContent) {
+          const isResume = !!args.resume_endpoint;
+          await context.streamContent({
+            type: "text" as const,
+            text: isResume
+              ? `[Vertex] Resuming status check for job`
+              : `[Vertex] Video generation started. resume_endpoint: ${compositeResumeEndpoint} (use this to check status if needed)`,
+          });
         }
 
-        const op = (await res.json()) as any;
-        operationName = op.name || op.operation || "";
-        current = op;
-      }
-
-      if (!operationName) {
-        throw new Error(
-          "Vertex did not return an operation name for long-running request",
-        );
-      }
-
-      // Construct the composite resume_endpoint: fetchUrl||operationName||outputPath
-      // This allows get_generation_status to use the URL directly and preserve output_path
-      const outputPathPart = args.output_path || "";
-      const compositeResumeEndpoint = `${fetchUrl}||${operationName}||${outputPathPart}`;
-
-      // Stream the resume_endpoint to the LLM immediately (before polling starts)
-      // This way the LLM has it even if MCP client times out during polling
-      if (context?.streamContent) {
-        const isResume = !!args.resume_endpoint;
-        await context.streamContent({
-          type: "text" as const,
-          text: isResume
-            ? `[Vertex] Resuming status check for job`
-            : `[Vertex] Video generation started. resume_endpoint: ${compositeResumeEndpoint} (use this to check status if needed)`,
-        });
-      }
+        // Poll for status - keep polling until done
+        // Resume_endpoint was already streamed, so if MCP client times out the LLM still has it
+        let done = current ? !!current.done || !!current.response : false;
+        const startTime = Date.now();
+        const MAX_POLL_TIME = 60000; // 60 seconds internal timeout - then return resume_endpoint
 
-      // Poll for status - keep polling until done
-      // Resume_endpoint was already streamed, so if MCP client times out the LLM still has it
-      let done = current ? !!current.done || !!current.response : false;
-      const startTime = Date.now();
-      const MAX_POLL_TIME = 60000; // 60 seconds internal timeout - then return resume_endpoint
+        while (!done && Date.now() - startTime < MAX_POLL_TIME) {
+          await wait(10000); // 10 second intervals
 
-      while (!done && Date.now() - startTime < MAX_POLL_TIME) {
-        await wait(10000); // 10 second intervals
+          current = await checkVertexStatus(compositeResumeEndpoint);
+          done = !!current.done || !!current.response;
 
-        current = await checkVertexStatus(compositeResumeEndpoint);
-        done = !!current.done || !!current.response;
+          if (context?.reportProgress) {
+            const elapsed = Date.now() - startTime;
+            const progressPercent = Math.min(
+              Math.round((elapsed / MAX_POLL_TIME) * 100),
+              99,
+            );
+            await context.reportProgress({
+              progress: progressPercent,
+              total: 100,
+            });
+          }
 
-        if (context?.reportProgress) {
-          const elapsed = Date.now() - startTime;
-          const progressPercent = Math.min(
-            Math.round((elapsed / MAX_POLL_TIME) * 100),
-            99,
-          );
-          await context.reportProgress({
-            progress: progressPercent,
-            total: 100,
+          if (context?.streamContent && !done) {
+            await context.streamContent({
+              type: "text" as const,
+              text: `[Vertex] Still processing... (${Math.round(
+                (Date.now() - startTime) / 1000,
+              )}s elapsed)`,
+            });
+          }
+        }
+
+        if (!done) {
+          return JSON.stringify({
+            status: "IN_PROGRESS",
+            request_id: operationName,
+            resume_endpoint: compositeResumeEndpoint,
+            message:
+              "Still in progress. Call this tool again with resume_endpoint to continue checking.",
           });
         }
 
-        if (context?.streamContent && !done) {
-          await context.streamContent({
-            type: "text" as const,
-            text: `[Vertex] Still processing... (${Math.round(
-              (Date.now() - startTime) / 1000,
-            )}s elapsed)`,
+        const resp = current.response || current;
+
+        // checkVertexStatus already handles saving videos and sanitizing base64
+        if (Array.isArray(resp.saved_videos) && resp.saved_videos.length > 0) {
+          return JSON.stringify({
+            videos: resp.saved_videos,
+            message: "Video(s) generated successfully",
           });
         }
-      }
 
-      if (!done) {
+        // If nothing saved, return a clean error without any raw JSON that could contain base64
+        // CRITICAL: Never return raw response data to prevent context window poisoning
+        const respKeys = resp ? Object.keys(resp) : [];
         return JSON.stringify({
-          status: "IN_PROGRESS",
-          request_id: operationName,
-          resume_endpoint: compositeResumeEndpoint,
+          status: "ERROR",
           message:
-            "Still in progress. Call this tool again with resume_endpoint to continue checking.",
-        });
-      }
-
-      const resp = current.response || current;
-      
-      // checkVertexStatus already handles saving videos and sanitizing base64
-      if (Array.isArray(resp.saved_videos) && resp.saved_videos.length > 0) {
-        return JSON.stringify({
-          videos: resp.saved_videos,
-          message: "Video(s) generated successfully",
+            "Vertex operation completed but no videos were found in the response.",
+          operationName,
+          responseKeys: respKeys,
+          hint: "The response structure may have changed. Check the Vertex AI documentation or search for the expected response format.",
         });
-      }
-
-      // If nothing saved, return a clean error without any raw JSON that could contain base64
-      // CRITICAL: Never return raw response data to prevent context window poisoning
-      const respKeys = resp ? Object.keys(resp) : [];
-      return JSON.stringify({
-        status: "ERROR",
-        message:
-          "Vertex operation completed but no videos were found in the response.",
-        operationName,
-        responseKeys: respKeys,
-        hint: "The response structure may have changed. Check the Vertex AI documentation or search for the expected response format.",
-      });
-    }, "imageToVideo");
+      },
+      "imageToVideo",
+      { toolName: "generateVideoi2v" },
+    );
   },
 };