@stackfactor/agent-utils 1.3.4 → 1.4.1

package/dist/cjs/langChain.d.ts

@@ -49,6 +49,7 @@ declare const _default: {
     runAgent: (agent: any, prompt: string, config: any, onProgress?: Function | null, usageTracker?: UsageTracker | null) => Promise<any>;
     runPromptWithModel: (modelName: string, config: any, prompt: any, onProgressReport: any, minPercent?: number, maxPercent?: number, expectsJsonResponse?: boolean, schema?: any, agentName?: string, tools?: any[], usageTracker?: UsageTracker | null) => Promise<any>;
     runPromptWithModelForImageGeneration: (modelName: string, config: any, prompt: string, options?: any, usageTracker?: UsageTracker | null) => Promise<any>;
+    runPromptWithModelForVideoGeneration: (modelName: string, config: any, prompt: string, options?: any, usageTracker?: UsageTracker | null) => Promise<any>;
     throwErrorIfNotSuccessful: (response: any) => string;
     updateUsageTrackerForCharacters: (tracker: UsageTracker | null | undefined, modelName: string, characterCount: number, config: any) => void;
     validateModel: (selectedModel: string, supportedModels: string[]) => string;

package/dist/cjs/langChain.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"langChain.d.ts","sourceRoot":"","sources":["../../src/langChain.ts"],"names":[],"mappings":"AA4GA;;;;;GAKG;AACH,MAAM,MAAM,YAAY,GAAG;IACzB,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,EAAE;QAAE,CAAC,QAAQ,EAAE,MAAM,GAAG,MAAM,CAAA;KAAE,CAAC;IACvC;;;;;OAKG;IACH,gBAAgB,CAAC,EAAE;QAAE,GAAG,EAAE,MAAM,CAAC;QAAC,KAAK,CAAC,EAAE,MAAM,CAAA;KAAE,EAAE,CAAC;CACtD,CAAC;AA2lBF;;;;;;;;;;;;;;;;;;;;;GAqBG;AAGH,eAAO,MAAM,WAAW,GACtB,WAAW,MAAM,EACjB,QAAQ,GAAG,EACX,SAAQ,GAAU,KACjB,GAoHF,CAAC;;0CAzuB2C,GAAG,KAAG,IAAI;wBA4vB/C,MAAM,aACD,MAAM,gBACH,MAAM,SACb,GAAG,EAAE,kBACI,GAAG,UACX,GAAG,iBACG,YAAY,GAAG,IAAI,KAChC,GAAG;sBAoCG,GAAG,UACF,MAAM,UACN,GAAG,eACC,QAAQ,GAAG,IAAI,iBACb,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;oCAkYF,MAAM,UACT,GAAG,UACH,GAAG,oBACO,GAAG,eACT,MAAM,eACN,MAAM,wBACG,OAAO,WACpB,GAAG,cACA,MAAM,UACV,GAAG,EAAE,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;sDA2tBF,MAAM,UACT,GAAG,UACH,MAAM,YACL,GAAG,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;0CArgC8B,GAAG,KAAG,MAAM;+CAzuB9C,YAAY,GAAG,IAAI,GAAG,SAAS,aAC7B,MAAM,kBACD,MAAM,UACd,GAAG,KACV,IAAI;mCAyvBU,MAAM,mBACJ,MAAM,EAAE,KACxB,MAAM;;AA4iCT,wBASE"}
+{"version":3,"file":"langChain.d.ts","sourceRoot":"","sources":["../../src/langChain.ts"],"names":[],"mappings":"AA4GA;;;;;GAKG;AACH,MAAM,MAAM,YAAY,GAAG;IACzB,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,EAAE;QAAE,CAAC,QAAQ,EAAE,MAAM,GAAG,MAAM,CAAA;KAAE,CAAC;IACvC;;;;;OAKG;IACH,gBAAgB,CAAC,EAAE;QAAE,GAAG,EAAE,MAAM,CAAC;QAAC,KAAK,CAAC,EAAE,MAAM,CAAA;KAAE,EAAE,CAAC;CACtD,CAAC;AAmrBF;;;;;;;;;;;;;;;;;;;;;GAqBG;AAGH,eAAO,MAAM,WAAW,GACtB,WAAW,MAAM,EACjB,QAAQ,GAAG,EACX,SAAQ,GAAU,KACjB,GAoHF,CAAC;;0CAj0B2C,GAAG,KAAG,IAAI;wBAo1B/C,MAAM,aACD,MAAM,gBACH,MAAM,SACb,GAAG,EAAE,kBACI,GAAG,UACX,GAAG,iBACG,YAAY,GAAG,IAAI,KAChC,GAAG;sBAoCG,GAAG,UACF,MAAM,UACN,GAAG,eACC,QAAQ,GAAG,IAAI,iBACb,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;oCAkYF,MAAM,UACT,GAAG,UACH,GAAG,oBACO,GAAG,eACT,MAAM,eACN,MAAM,wBACG,OAAO,WACpB,GAAG,cACA,MAAM,UACV,GAAG,EAAE,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;sDA2tBF,MAAM,UACT,GAAG,UACH,MAAM,YACL,GAAG,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;sDAmTF,MAAM,UACT,GAAG,UACH,MAAM,YACL,GAAG,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;0CA7zC8B,GAAG,KAAG,MAAM;+CAh0B9C,YAAY,GAAG,IAAI,GAAG,SAAS,aAC7B,MAAM,kBACD,MAAM,UACd,GAAG,KACV,IAAI;mCAg1BU,MAAM,mBACJ,MAAM,EAAE,KACxB,MAAM;;AAg2CT,wBAUE"}

package/dist/cjs/langChain.js

@@ -251,6 +251,39 @@ const updateUsageTrackerForCharacters = (tracker, modelName, characterCount, con
     const key = `${modelName}_characters`;
     tracker.tokens[key] = (tracker.tokens[key] || 0) + characterCount;
 };
+/**
+ * Adds second-billed usage (video generation) to the caller-supplied tracker.
+ *
+ * Video is the first output here that providers bill by DURATION rather than by
+ * tokens or characters, so the rate constant `<model>-second-costs` is read as
+ * USD PER SECOND — not per million, as the `input-token` and `character` rates
+ * are. A Veo or Sora second costs cents, and a per-million reading would bill
+ * it as effectively free.
+ *
+ * The rate is per model, not per model-and-resolution, so a caller that lets a
+ * tenant change resolution on a model whose price varies by it (sora-2-pro is
+ * $0.30/$0.50/$0.70 at 720p/1024p/1080p) must pin the resolution its constant
+ * describes, or it will bill the wrong number.
+ *
+ * @param tracker - The accumulator to add to; no-ops when null
+ * @param modelName - The video model id, used for both the rate and the counter key
+ * @param seconds - Billed duration of the generated clip
+ * @param config - Integration config carrying `<model>-second-costs`
+ */
+const updateVideoUsageTracker = (tracker, modelName, seconds, config) => {
+    if (!tracker || !modelName || !seconds)
+        return;
+    if (typeof tracker.cost !== "number")
+        tracker.cost = 0;
+    if (!tracker.tokens || typeof tracker.tokens !== "object")
+        tracker.tokens = {};
+    const rate = getModelRate(modelName, config, "second");
+    const addedCost = seconds * rate;
+    if (Number.isFinite(addedCost) && addedCost > 0)
+        tracker.cost += addedCost;
+    const key = `${modelName}_seconds`;
+    tracker.tokens[key] = (tracker.tokens[key] || 0) + seconds;
+};
 /**
  * Extracts a normalized token-usage object from a single LangChain `invoke()`
  * response. Reads from `usage_metadata` first (standardized in LangChain v1),
@@ -610,6 +643,46 @@ const resolveTemperatureSetting = (modelName, config) => {
     }
     return { temperature: clamped };
 };
+/**
+ * Gemini's ceiling on `maxOutputTokens`. The API validates the field and
+ * rejects the request outright rather than clamping it:
+ * "Unable to submit request because it has a maxOutputTokens value of N but the
+ * supported range is from 1 (inclusive) to 65537 (exclusive)" (HTTP 400,
+ * INVALID_ARGUMENT).
+ *
+ * Set to 64000 rather than that 65536 validation ceiling: the Pro models report
+ * a generation capacity of 64000, so 64000 is the value every model across the
+ * Gemini 3 line (3.1 Pro, 3.5/3.7/3.8 Flash) can both accept and actually
+ * produce. One safe constant beats a per-model table that would need editing
+ * every time Google ships a model — and the 1536 tokens of headroom given up on
+ * Flash are far beyond what any caller asks for.
+ *
+ * Deliberately Gemini-only. Anthropic and OpenAI take the caller's `maxTokens`
+ * unchanged, since neither rejects the values configured today.
+ */
+const GEMINI_MAX_OUTPUT_TOKENS = 64000;
+/**
+ * Resolves `maxOutputTokens` for a Gemini model. Callers share one `maxTokens`
+ * setting across every provider, and the other providers tolerate values far
+ * above what Gemini accepts, so anything above Gemini's ceiling is clamped here
+ * (with a warning) instead of being sent through to a guaranteed HTTP 400. With
+ * no caller value the ceiling itself is used, which is what the previous
+ * 200000 default was reaching for.
+ * @param modelName - The Gemini model identifier being routed
+ * @param config - Configuration object; only `maxTokens` is consulted
+ * @returns A token count within Gemini's supported range
+ */
+const resolveGeminiMaxOutputTokens = (modelName, config) => {
+    const maxTokens = config?.maxTokens;
+    if (typeof maxTokens !== "number" || !Number.isFinite(maxTokens)) {
+        return GEMINI_MAX_OUTPUT_TOKENS;
+    }
+    if (maxTokens > GEMINI_MAX_OUTPUT_TOKENS) {
+        logger_js_1.default.log(null, logger_js_1.default.levels.warn, `maxTokens ${maxTokens} exceeds the ${GEMINI_MAX_OUTPUT_TOKENS}-token output ceiling of "${modelName}"; clamping to ${GEMINI_MAX_OUTPUT_TOKENS}.`);
+        return GEMINI_MAX_OUTPUT_TOKENS;
+    }
+    return maxTokens;
+};
 /**
  * Flattens LangChain message content into plain text. Content is a string for
  * ordinary completions, but providers switch to an array of blocks whenever the
@@ -701,7 +774,7 @@ const getLLMModel = (modelName, config, schema = null) => {
     else if (modelName.startsWith("gemini-")) {
         const model = new google_genai_1.ChatGoogleGenerativeAI({
             apiKey: config.googleAPIKey,
-            maxOutputTokens: config.maxTokens || 200000,
+            maxOutputTokens: resolveGeminiMaxOutputTokens(modelName, config),
             model: modelName,
             // `json: true` forces responseMimeType=application/json, guaranteeing the
             // model returns valid JSON text (never prose) for the parse pipeline.
@@ -1771,12 +1844,202 @@ const runPromptWithModelForImageGeneration = async (modelName, config, prompt, o
         throw errorHandling_js_1.default.create(const_js_1.default.HTTP_CODES.INTERNAL_SERVER_ERROR, `Image generation failed: ${error?.message || String(error)}`);
     }
 };
+/**
+ * Determines the video generation provider for a model name from its prefix.
+ * `veo-` maps to `"google"`, `sora-` to `"openai"`. Returns `null` for anything
+ * else so the caller can reject an unrecognised id rather than guess at it.
+ * @param modelName - The video model identifier to inspect
+ * @returns `"openai"`, `"google"`, or `null`
+ */
+const getVideoModelProvider = (modelName) => {
+    if (modelName.startsWith("sora-"))
+        return "openai";
+    if (modelName.startsWith("veo-"))
+        return "google";
+    return null;
+};
+/** How long to wait between polls of a video generation job. */
+const VIDEO_POLL_INTERVAL_MS = 10_000;
+/** How long to keep polling before giving up on a video generation job. */
+const VIDEO_POLL_TIMEOUT_MS = 600_000;
+/**
+ * Sleeps for the given number of milliseconds.
+ * @param ms - Milliseconds to wait
+ */
+const delay = (ms) => new Promise((resolve) => setTimeout(resolve, ms));
+/**
+ * Generates a video with Google Veo.
+ *
+ * Video generation is a long-running operation on both providers — a Veo clip
+ * takes on the order of a minute or two, far past any request timeout — so the
+ * SDK returns an operation handle and this polls it to completion.
+ *
+ * The finished clip comes back either as inline base64 (`videoBytes`) or as a
+ * `uri` on Google's file service, depending on model and size. The uri is NOT
+ * public: it needs the same API key as the generation call, sent as a header,
+ * which is why this fetches it here rather than handing the caller a link that
+ * would 403 the moment it left the process.
+ *
+ * @param modelName - A Veo model id, e.g. `"veo-3.1-fast-generate-preview"`
+ * @param config - Configuration object; must include `googleAPIKey`
+ * @param prompt - The text prompt describing the video
+ * @param options - `aspectRatio`, `resolution`, `durationSeconds`,
+ *   `negativePrompt`, `personGeneration`, `pollTimeoutMs`
+ * @param usageTracker - Accumulator billed per second of generated video
+ * @returns `{ b64_json, mimeType, seconds, model }`
+ */
+const generateVideoWithGoogle = async (modelName, config, prompt, options = {}, usageTracker = null) => {
+    const { aspectRatio = "16:9", resolution = "720p", durationSeconds = 8, negativePrompt, personGeneration, pollTimeoutMs = VIDEO_POLL_TIMEOUT_MS, } = options;
+    const ai = new genai_1.GoogleGenAI({ apiKey: config.googleAPIKey });
+    let operation = await ai.models.generateVideos({
+        model: modelName,
+        prompt,
+        config: {
+            aspectRatio,
+            resolution,
+            durationSeconds,
+            numberOfVideos: 1,
+            ...(negativePrompt ? { negativePrompt } : {}),
+            ...(personGeneration ? { personGeneration } : {}),
+        },
+    });
+    const deadline = Date.now() + pollTimeoutMs;
+    while (!operation.done) {
+        if (Date.now() > deadline) {
+            throw errorHandling_js_1.default.create(const_js_1.default.HTTP_CODES.INTERNAL_SERVER_ERROR, `Video generation with "${modelName}" did not finish within ${Math.round(pollTimeoutMs / 1000)} seconds.`);
+        }
+        await delay(VIDEO_POLL_INTERVAL_MS);
+        operation = await ai.operations.getVideosOperation({ operation });
+    }
+    if (operation.error) {
+        throw errorHandling_js_1.default.create(const_js_1.default.HTTP_CODES.INTERNAL_SERVER_ERROR, `Video generation with "${modelName}" failed: ${operation.error?.message || JSON.stringify(operation.error)}`);
+    }
+    const video = operation.response?.generatedVideos?.[0]?.video;
+    if (!video) {
+        throw errorHandling_js_1.default.create(const_js_1.default.HTTP_CODES.INTERNAL_SERVER_ERROR, `No video was generated by ${modelName}`);
+    }
+    let b64 = video.videoBytes;
+    if (!b64) {
+        if (!video.uri) {
+            throw errorHandling_js_1.default.create(const_js_1.default.HTTP_CODES.INTERNAL_SERVER_ERROR, `${modelName} returned a video with neither inline bytes nor a uri`);
+        }
+        const download = await fetch(video.uri, {
+            headers: { "x-goog-api-key": config.googleAPIKey },
+        });
+        if (!download.ok) {
+            throw errorHandling_js_1.default.create(const_js_1.default.HTTP_CODES.INTERNAL_SERVER_ERROR, `Downloading the generated video failed with ${download.status} ${download.statusText}`);
+        }
+        b64 = Buffer.from(await download.arrayBuffer()).toString("base64");
+    }
+    // Billed on the duration ASKED FOR: the provider bills the clip it rendered,
+    // and there is no returned duration to read back.
+    updateVideoUsageTracker(usageTracker, modelName, durationSeconds, config);
+    return {
+        b64_json: b64,
+        mimeType: video.mimeType || "video/mp4",
+        seconds: durationSeconds,
+        model: modelName,
+    };
+};
+/**
+ * Generates a video with OpenAI Sora.
+ *
+ * Like Veo this is a job rather than a request: `videos.create` returns
+ * immediately with a `queued` job, which is polled until it completes and then
+ * downloaded as bytes.
+ *
+ * @param modelName - A Sora model id, e.g. `"sora-2"`
+ * @param config - Configuration object; must include `openAIAPIKey`
+ * @param prompt - The text prompt describing the video
+ * @param options - `size` (`"1280x720"` etc.), `durationSeconds` (4, 8 or 12),
+ *   `pollTimeoutMs`
+ * @param usageTracker - Accumulator billed per second of generated video
+ * @returns `{ b64_json, mimeType, seconds, model }`
+ */
+const generateVideoWithOpenAI = async (modelName, config, prompt, options = {}, usageTracker = null) => {
+    const { size = "1280x720", durationSeconds = 8, pollTimeoutMs = VIDEO_POLL_TIMEOUT_MS, } = options;
+    const openai = new openai_2.OpenAI({ apiKey: config.openAIAPIKey });
+    // Sora takes the duration as one of a fixed set of STRINGS ("4", "8", "12").
+    // A number, or any other value, is rejected by the API.
+    const seconds = String(durationSeconds);
+    if (!["4", "8", "12"].includes(seconds)) {
+        throw errorHandling_js_1.default.create(const_js_1.default.HTTP_CODES.BAD_REQUEST, `Invalid durationSeconds "${durationSeconds}" for ${modelName}. Sora supports 4, 8 or 12.`);
+    }
+    let job = await openai.videos.create({
+        model: modelName,
+        prompt,
+        seconds: seconds,
+        size: size,
+    });
+    const deadline = Date.now() + pollTimeoutMs;
+    while (job.status === "queued" || job.status === "in_progress") {
+        if (Date.now() > deadline) {
+            throw errorHandling_js_1.default.create(const_js_1.default.HTTP_CODES.INTERNAL_SERVER_ERROR, `Video generation with "${modelName}" did not finish within ${Math.round(pollTimeoutMs / 1000)} seconds.`);
+        }
+        await delay(VIDEO_POLL_INTERVAL_MS);
+        job = await openai.videos.retrieve(job.id);
+    }
+    if (job.status !== "completed") {
+        throw errorHandling_js_1.default.create(const_js_1.default.HTTP_CODES.INTERNAL_SERVER_ERROR, `Video generation with "${modelName}" ${job.status}: ${job.error?.message || "no reason given"}`);
+    }
+    const content = await openai.videos.downloadContent(job.id);
+    const b64 = Buffer.from(await content.arrayBuffer()).toString("base64");
+    updateVideoUsageTracker(usageTracker, modelName, Number(job.seconds) || Number(seconds), config);
+    return {
+        b64_json: b64,
+        mimeType: "video/mp4",
+        seconds: Number(job.seconds) || Number(seconds),
+        model: modelName,
+    };
+};
+/**
+ * Top-level entry point for AI video generation, mirroring
+ * `runPromptWithModelForImageGeneration`. Routes to Veo or Sora on the model
+ * name prefix and logs the total generation time on success.
+ *
+ * Unlike image generation this always returns bytes — never a url. Both
+ * providers hand back a link that requires their API key, so a url would be
+ * useless to any caller that is not this process.
+ *
+ * @param modelName - The video model identifier; must start with `veo-` or `sora-`
+ * @param config - Configuration object with `googleAPIKey` or `openAIAPIKey`
+ * @param prompt - The text prompt describing the video to generate
+ * @param options - Provider-specific generation options
+ * @param usageTracker - Accumulator billed per second of generated video
+ * @returns `{ b64_json, mimeType, seconds, model }`
+ */
+const runPromptWithModelForVideoGeneration = async (modelName, config, prompt, options = {}, usageTracker = null) => {
+    const provider = getVideoModelProvider(modelName);
+    if (!provider) {
+        throw errorHandling_js_1.default.create(const_js_1.default.HTTP_CODES.BAD_REQUEST, `Unable to determine provider for model: ${modelName}. Model name should start with 'veo-' or 'sora-'.`);
+    }
+    const startTime = Date.now();
+    try {
+        const result = provider === "openai"
+            ? await generateVideoWithOpenAI(modelName, config, prompt, options, usageTracker)
+            : await generateVideoWithGoogle(modelName, config, prompt, options, usageTracker);
+        const duration = Date.now() - startTime;
+        logger_js_1.default.log(null, logger_js_1.default.levels.info, `Video generation with "${modelName}" completed in ${Math.round(duration / 1000)} seconds.`);
+        return result;
+    }
+    catch (error) {
+        if (error?.code && error?.message) {
+            throw error;
+        }
+        if (error?.status || error?.code) {
+            const statusCode = error.status || const_js_1.default.HTTP_CODES.INTERNAL_SERVER_ERROR;
+            throw errorHandling_js_1.default.create(statusCode, error.message || "Video generation failed");
+        }
+        throw errorHandling_js_1.default.create(const_js_1.default.HTTP_CODES.INTERNAL_SERVER_ERROR, `Video generation failed: ${error?.message || String(error)}`);
+    }
+};
 exports.default = {
     checkIfAIProviderConfigured,
     createAgent,
     runAgent,
     runPromptWithModel,
     runPromptWithModelForImageGeneration,
+    runPromptWithModelForVideoGeneration,
     throwErrorIfNotSuccessful,
     updateUsageTrackerForCharacters,
     validateModel,

package/dist/esm/langChain.d.ts

@@ -49,6 +49,7 @@ declare const _default: {
     runAgent: (agent: any, prompt: string, config: any, onProgress?: Function | null, usageTracker?: UsageTracker | null) => Promise<any>;
     runPromptWithModel: (modelName: string, config: any, prompt: any, onProgressReport: any, minPercent?: number, maxPercent?: number, expectsJsonResponse?: boolean, schema?: any, agentName?: string, tools?: any[], usageTracker?: UsageTracker | null) => Promise<any>;
     runPromptWithModelForImageGeneration: (modelName: string, config: any, prompt: string, options?: any, usageTracker?: UsageTracker | null) => Promise<any>;
+    runPromptWithModelForVideoGeneration: (modelName: string, config: any, prompt: string, options?: any, usageTracker?: UsageTracker | null) => Promise<any>;
     throwErrorIfNotSuccessful: (response: any) => string;
     updateUsageTrackerForCharacters: (tracker: UsageTracker | null | undefined, modelName: string, characterCount: number, config: any) => void;
     validateModel: (selectedModel: string, supportedModels: string[]) => string;

package/dist/esm/langChain.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"langChain.d.ts","sourceRoot":"","sources":["../../src/langChain.ts"],"names":[],"mappings":"AA4GA;;;;;GAKG;AACH,MAAM,MAAM,YAAY,GAAG;IACzB,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,EAAE;QAAE,CAAC,QAAQ,EAAE,MAAM,GAAG,MAAM,CAAA;KAAE,CAAC;IACvC;;;;;OAKG;IACH,gBAAgB,CAAC,EAAE;QAAE,GAAG,EAAE,MAAM,CAAC;QAAC,KAAK,CAAC,EAAE,MAAM,CAAA;KAAE,EAAE,CAAC;CACtD,CAAC;AA2lBF;;;;;;;;;;;;;;;;;;;;;GAqBG;AAGH,eAAO,MAAM,WAAW,GACtB,WAAW,MAAM,EACjB,QAAQ,GAAG,EACX,SAAQ,GAAU,KACjB,GAoHF,CAAC;;0CAzuB2C,GAAG,KAAG,IAAI;wBA4vB/C,MAAM,aACD,MAAM,gBACH,MAAM,SACb,GAAG,EAAE,kBACI,GAAG,UACX,GAAG,iBACG,YAAY,GAAG,IAAI,KAChC,GAAG;sBAoCG,GAAG,UACF,MAAM,UACN,GAAG,eACC,QAAQ,GAAG,IAAI,iBACb,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;oCAkYF,MAAM,UACT,GAAG,UACH,GAAG,oBACO,GAAG,eACT,MAAM,eACN,MAAM,wBACG,OAAO,WACpB,GAAG,cACA,MAAM,UACV,GAAG,EAAE,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;sDA2tBF,MAAM,UACT,GAAG,UACH,MAAM,YACL,GAAG,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;0CArgC8B,GAAG,KAAG,MAAM;+CAzuB9C,YAAY,GAAG,IAAI,GAAG,SAAS,aAC7B,MAAM,kBACD,MAAM,UACd,GAAG,KACV,IAAI;mCAyvBU,MAAM,mBACJ,MAAM,EAAE,KACxB,MAAM;;AA4iCT,wBASE"}
+{"version":3,"file":"langChain.d.ts","sourceRoot":"","sources":["../../src/langChain.ts"],"names":[],"mappings":"AA4GA;;;;;GAKG;AACH,MAAM,MAAM,YAAY,GAAG;IACzB,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,EAAE;QAAE,CAAC,QAAQ,EAAE,MAAM,GAAG,MAAM,CAAA;KAAE,CAAC;IACvC;;;;;OAKG;IACH,gBAAgB,CAAC,EAAE;QAAE,GAAG,EAAE,MAAM,CAAC;QAAC,KAAK,CAAC,EAAE,MAAM,CAAA;KAAE,EAAE,CAAC;CACtD,CAAC;AAmrBF;;;;;;;;;;;;;;;;;;;;;GAqBG;AAGH,eAAO,MAAM,WAAW,GACtB,WAAW,MAAM,EACjB,QAAQ,GAAG,EACX,SAAQ,GAAU,KACjB,GAoHF,CAAC;;0CAj0B2C,GAAG,KAAG,IAAI;wBAo1B/C,MAAM,aACD,MAAM,gBACH,MAAM,SACb,GAAG,EAAE,kBACI,GAAG,UACX,GAAG,iBACG,YAAY,GAAG,IAAI,KAChC,GAAG;sBAoCG,GAAG,UACF,MAAM,UACN,GAAG,eACC,QAAQ,GAAG,IAAI,iBACb,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;oCAkYF,MAAM,UACT,GAAG,UACH,GAAG,oBACO,GAAG,eACT,MAAM,eACN,MAAM,wBACG,OAAO,WACpB,GAAG,cACA,MAAM,UACV,GAAG,EAAE,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;sDA2tBF,MAAM,UACT,GAAG,UACH,MAAM,YACL,GAAG,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;sDAmTF,MAAM,UACT,GAAG,UACH,MAAM,YACL,GAAG,iBACE,YAAY,GAAG,IAAI,KAChC,OAAO,CAAC,GAAG,CAAC;0CA7zC8B,GAAG,KAAG,MAAM;+CAh0B9C,YAAY,GAAG,IAAI,GAAG,SAAS,aAC7B,MAAM,kBACD,MAAM,UACd,GAAG,KACV,IAAI;mCAg1BU,MAAM,mBACJ,MAAM,EAAE,KACxB,MAAM;;AAg2CT,wBAUE"}

package/dist/esm/langChain.js

@@ -245,6 +245,39 @@ const updateUsageTrackerForCharacters = (tracker, modelName, characterCount, con
     const key = `${modelName}_characters`;
     tracker.tokens[key] = (tracker.tokens[key] || 0) + characterCount;
 };
+/**
+ * Adds second-billed usage (video generation) to the caller-supplied tracker.
+ *
+ * Video is the first output here that providers bill by DURATION rather than by
+ * tokens or characters, so the rate constant `<model>-second-costs` is read as
+ * USD PER SECOND — not per million, as the `input-token` and `character` rates
+ * are. A Veo or Sora second costs cents, and a per-million reading would bill
+ * it as effectively free.
+ *
+ * The rate is per model, not per model-and-resolution, so a caller that lets a
+ * tenant change resolution on a model whose price varies by it (sora-2-pro is
+ * $0.30/$0.50/$0.70 at 720p/1024p/1080p) must pin the resolution its constant
+ * describes, or it will bill the wrong number.
+ *
+ * @param tracker - The accumulator to add to; no-ops when null
+ * @param modelName - The video model id, used for both the rate and the counter key
+ * @param seconds - Billed duration of the generated clip
+ * @param config - Integration config carrying `<model>-second-costs`
+ */
+const updateVideoUsageTracker = (tracker, modelName, seconds, config) => {
+    if (!tracker || !modelName || !seconds)
+        return;
+    if (typeof tracker.cost !== "number")
+        tracker.cost = 0;
+    if (!tracker.tokens || typeof tracker.tokens !== "object")
+        tracker.tokens = {};
+    const rate = getModelRate(modelName, config, "second");
+    const addedCost = seconds * rate;
+    if (Number.isFinite(addedCost) && addedCost > 0)
+        tracker.cost += addedCost;
+    const key = `${modelName}_seconds`;
+    tracker.tokens[key] = (tracker.tokens[key] || 0) + seconds;
+};
 /**
  * Extracts a normalized token-usage object from a single LangChain `invoke()`
  * response. Reads from `usage_metadata` first (standardized in LangChain v1),
@@ -604,6 +637,46 @@ const resolveTemperatureSetting = (modelName, config) => {
     }
     return { temperature: clamped };
 };
+/**
+ * Gemini's ceiling on `maxOutputTokens`. The API validates the field and
+ * rejects the request outright rather than clamping it:
+ * "Unable to submit request because it has a maxOutputTokens value of N but the
+ * supported range is from 1 (inclusive) to 65537 (exclusive)" (HTTP 400,
+ * INVALID_ARGUMENT).
+ *
+ * Set to 64000 rather than that 65536 validation ceiling: the Pro models report
+ * a generation capacity of 64000, so 64000 is the value every model across the
+ * Gemini 3 line (3.1 Pro, 3.5/3.7/3.8 Flash) can both accept and actually
+ * produce. One safe constant beats a per-model table that would need editing
+ * every time Google ships a model — and the 1536 tokens of headroom given up on
+ * Flash are far beyond what any caller asks for.
+ *
+ * Deliberately Gemini-only. Anthropic and OpenAI take the caller's `maxTokens`
+ * unchanged, since neither rejects the values configured today.
+ */
+const GEMINI_MAX_OUTPUT_TOKENS = 64000;
+/**
+ * Resolves `maxOutputTokens` for a Gemini model. Callers share one `maxTokens`
+ * setting across every provider, and the other providers tolerate values far
+ * above what Gemini accepts, so anything above Gemini's ceiling is clamped here
+ * (with a warning) instead of being sent through to a guaranteed HTTP 400. With
+ * no caller value the ceiling itself is used, which is what the previous
+ * 200000 default was reaching for.
+ * @param modelName - The Gemini model identifier being routed
+ * @param config - Configuration object; only `maxTokens` is consulted
+ * @returns A token count within Gemini's supported range
+ */
+const resolveGeminiMaxOutputTokens = (modelName, config) => {
+    const maxTokens = config?.maxTokens;
+    if (typeof maxTokens !== "number" || !Number.isFinite(maxTokens)) {
+        return GEMINI_MAX_OUTPUT_TOKENS;
+    }
+    if (maxTokens > GEMINI_MAX_OUTPUT_TOKENS) {
+        logger.log(null, logger.levels.warn, `maxTokens ${maxTokens} exceeds the ${GEMINI_MAX_OUTPUT_TOKENS}-token output ceiling of "${modelName}"; clamping to ${GEMINI_MAX_OUTPUT_TOKENS}.`);
+        return GEMINI_MAX_OUTPUT_TOKENS;
+    }
+    return maxTokens;
+};
 /**
  * Flattens LangChain message content into plain text. Content is a string for
  * ordinary completions, but providers switch to an array of blocks whenever the
@@ -695,7 +768,7 @@ export const getLLMModel = (modelName, config, schema = null) => {
     else if (modelName.startsWith("gemini-")) {
         const model = new ChatGoogleGenerativeAI({
             apiKey: config.googleAPIKey,
-            maxOutputTokens: config.maxTokens || 200000,
+            maxOutputTokens: resolveGeminiMaxOutputTokens(modelName, config),
             model: modelName,
             // `json: true` forces responseMimeType=application/json, guaranteeing the
             // model returns valid JSON text (never prose) for the parse pipeline.
@@ -1764,12 +1837,202 @@ const runPromptWithModelForImageGeneration = async (modelName, config, prompt, o
         throw errorHandlingHelper.create(constants.HTTP_CODES.INTERNAL_SERVER_ERROR, `Image generation failed: ${error?.message || String(error)}`);
     }
 };
+/**
+ * Determines the video generation provider for a model name from its prefix.
+ * `veo-` maps to `"google"`, `sora-` to `"openai"`. Returns `null` for anything
+ * else so the caller can reject an unrecognised id rather than guess at it.
+ * @param modelName - The video model identifier to inspect
+ * @returns `"openai"`, `"google"`, or `null`
+ */
+const getVideoModelProvider = (modelName) => {
+    if (modelName.startsWith("sora-"))
+        return "openai";
+    if (modelName.startsWith("veo-"))
+        return "google";
+    return null;
+};
+/** How long to wait between polls of a video generation job. */
+const VIDEO_POLL_INTERVAL_MS = 10_000;
+/** How long to keep polling before giving up on a video generation job. */
+const VIDEO_POLL_TIMEOUT_MS = 600_000;
+/**
+ * Sleeps for the given number of milliseconds.
+ * @param ms - Milliseconds to wait
+ */
+const delay = (ms) => new Promise((resolve) => setTimeout(resolve, ms));
+/**
+ * Generates a video with Google Veo.
+ *
+ * Video generation is a long-running operation on both providers — a Veo clip
+ * takes on the order of a minute or two, far past any request timeout — so the
+ * SDK returns an operation handle and this polls it to completion.
+ *
+ * The finished clip comes back either as inline base64 (`videoBytes`) or as a
+ * `uri` on Google's file service, depending on model and size. The uri is NOT
+ * public: it needs the same API key as the generation call, sent as a header,
+ * which is why this fetches it here rather than handing the caller a link that
+ * would 403 the moment it left the process.
+ *
+ * @param modelName - A Veo model id, e.g. `"veo-3.1-fast-generate-preview"`
+ * @param config - Configuration object; must include `googleAPIKey`
+ * @param prompt - The text prompt describing the video
+ * @param options - `aspectRatio`, `resolution`, `durationSeconds`,
+ *   `negativePrompt`, `personGeneration`, `pollTimeoutMs`
+ * @param usageTracker - Accumulator billed per second of generated video
+ * @returns `{ b64_json, mimeType, seconds, model }`
+ */
+const generateVideoWithGoogle = async (modelName, config, prompt, options = {}, usageTracker = null) => {
+    const { aspectRatio = "16:9", resolution = "720p", durationSeconds = 8, negativePrompt, personGeneration, pollTimeoutMs = VIDEO_POLL_TIMEOUT_MS, } = options;
+    const ai = new GoogleGenAI({ apiKey: config.googleAPIKey });
+    let operation = await ai.models.generateVideos({
+        model: modelName,
+        prompt,
+        config: {
+            aspectRatio,
+            resolution,
+            durationSeconds,
+            numberOfVideos: 1,
+            ...(negativePrompt ? { negativePrompt } : {}),
+            ...(personGeneration ? { personGeneration } : {}),
+        },
+    });
+    const deadline = Date.now() + pollTimeoutMs;
+    while (!operation.done) {
+        if (Date.now() > deadline) {
+            throw errorHandlingHelper.create(constants.HTTP_CODES.INTERNAL_SERVER_ERROR, `Video generation with "${modelName}" did not finish within ${Math.round(pollTimeoutMs / 1000)} seconds.`);
+        }
+        await delay(VIDEO_POLL_INTERVAL_MS);
+        operation = await ai.operations.getVideosOperation({ operation });
+    }
+    if (operation.error) {
+        throw errorHandlingHelper.create(constants.HTTP_CODES.INTERNAL_SERVER_ERROR, `Video generation with "${modelName}" failed: ${operation.error?.message || JSON.stringify(operation.error)}`);
+    }
+    const video = operation.response?.generatedVideos?.[0]?.video;
+    if (!video) {
+        throw errorHandlingHelper.create(constants.HTTP_CODES.INTERNAL_SERVER_ERROR, `No video was generated by ${modelName}`);
+    }
+    let b64 = video.videoBytes;
+    if (!b64) {
+        if (!video.uri) {
+            throw errorHandlingHelper.create(constants.HTTP_CODES.INTERNAL_SERVER_ERROR, `${modelName} returned a video with neither inline bytes nor a uri`);
+        }
+        const download = await fetch(video.uri, {
+            headers: { "x-goog-api-key": config.googleAPIKey },
+        });
+        if (!download.ok) {
+            throw errorHandlingHelper.create(constants.HTTP_CODES.INTERNAL_SERVER_ERROR, `Downloading the generated video failed with ${download.status} ${download.statusText}`);
+        }
+        b64 = Buffer.from(await download.arrayBuffer()).toString("base64");
+    }
+    // Billed on the duration ASKED FOR: the provider bills the clip it rendered,
+    // and there is no returned duration to read back.
+    updateVideoUsageTracker(usageTracker, modelName, durationSeconds, config);
+    return {
+        b64_json: b64,
+        mimeType: video.mimeType || "video/mp4",
+        seconds: durationSeconds,
+        model: modelName,
+    };
+};
+/**
+ * Generates a video with OpenAI Sora.
+ *
+ * Like Veo this is a job rather than a request: `videos.create` returns
+ * immediately with a `queued` job, which is polled until it completes and then
+ * downloaded as bytes.
+ *
+ * @param modelName - A Sora model id, e.g. `"sora-2"`
+ * @param config - Configuration object; must include `openAIAPIKey`
+ * @param prompt - The text prompt describing the video
+ * @param options - `size` (`"1280x720"` etc.), `durationSeconds` (4, 8 or 12),
+ *   `pollTimeoutMs`
+ * @param usageTracker - Accumulator billed per second of generated video
+ * @returns `{ b64_json, mimeType, seconds, model }`
+ */
+const generateVideoWithOpenAI = async (modelName, config, prompt, options = {}, usageTracker = null) => {
+    const { size = "1280x720", durationSeconds = 8, pollTimeoutMs = VIDEO_POLL_TIMEOUT_MS, } = options;
+    const openai = new OpenAI({ apiKey: config.openAIAPIKey });
+    // Sora takes the duration as one of a fixed set of STRINGS ("4", "8", "12").
+    // A number, or any other value, is rejected by the API.
+    const seconds = String(durationSeconds);
+    if (!["4", "8", "12"].includes(seconds)) {
+        throw errorHandlingHelper.create(constants.HTTP_CODES.BAD_REQUEST, `Invalid durationSeconds "${durationSeconds}" for ${modelName}. Sora supports 4, 8 or 12.`);
+    }
+    let job = await openai.videos.create({
+        model: modelName,
+        prompt,
+        seconds: seconds,
+        size: size,
+    });
+    const deadline = Date.now() + pollTimeoutMs;
+    while (job.status === "queued" || job.status === "in_progress") {
+        if (Date.now() > deadline) {
+            throw errorHandlingHelper.create(constants.HTTP_CODES.INTERNAL_SERVER_ERROR, `Video generation with "${modelName}" did not finish within ${Math.round(pollTimeoutMs / 1000)} seconds.`);
+        }
+        await delay(VIDEO_POLL_INTERVAL_MS);
+        job = await openai.videos.retrieve(job.id);
+    }
+    if (job.status !== "completed") {
+        throw errorHandlingHelper.create(constants.HTTP_CODES.INTERNAL_SERVER_ERROR, `Video generation with "${modelName}" ${job.status}: ${job.error?.message || "no reason given"}`);
+    }
+    const content = await openai.videos.downloadContent(job.id);
+    const b64 = Buffer.from(await content.arrayBuffer()).toString("base64");
+    updateVideoUsageTracker(usageTracker, modelName, Number(job.seconds) || Number(seconds), config);
+    return {
+        b64_json: b64,
+        mimeType: "video/mp4",
+        seconds: Number(job.seconds) || Number(seconds),
+        model: modelName,
+    };
+};
+/**
+ * Top-level entry point for AI video generation, mirroring
+ * `runPromptWithModelForImageGeneration`. Routes to Veo or Sora on the model
+ * name prefix and logs the total generation time on success.
+ *
+ * Unlike image generation this always returns bytes — never a url. Both
+ * providers hand back a link that requires their API key, so a url would be
+ * useless to any caller that is not this process.
+ *
+ * @param modelName - The video model identifier; must start with `veo-` or `sora-`
+ * @param config - Configuration object with `googleAPIKey` or `openAIAPIKey`
+ * @param prompt - The text prompt describing the video to generate
+ * @param options - Provider-specific generation options
+ * @param usageTracker - Accumulator billed per second of generated video
+ * @returns `{ b64_json, mimeType, seconds, model }`
+ */
+const runPromptWithModelForVideoGeneration = async (modelName, config, prompt, options = {}, usageTracker = null) => {
+    const provider = getVideoModelProvider(modelName);
+    if (!provider) {
+        throw errorHandlingHelper.create(constants.HTTP_CODES.BAD_REQUEST, `Unable to determine provider for model: ${modelName}. Model name should start with 'veo-' or 'sora-'.`);
+    }
+    const startTime = Date.now();
+    try {
+        const result = provider === "openai"
+            ? await generateVideoWithOpenAI(modelName, config, prompt, options, usageTracker)
+            : await generateVideoWithGoogle(modelName, config, prompt, options, usageTracker);
+        const duration = Date.now() - startTime;
+        logger.log(null, logger.levels.info, `Video generation with "${modelName}" completed in ${Math.round(duration / 1000)} seconds.`);
+        return result;
+    }
+    catch (error) {
+        if (error?.code && error?.message) {
+            throw error;
+        }
+        if (error?.status || error?.code) {
+            const statusCode = error.status || constants.HTTP_CODES.INTERNAL_SERVER_ERROR;
+            throw errorHandlingHelper.create(statusCode, error.message || "Video generation failed");
+        }
+        throw errorHandlingHelper.create(constants.HTTP_CODES.INTERNAL_SERVER_ERROR, `Video generation failed: ${error?.message || String(error)}`);
+    }
+};
 export default {
     checkIfAIProviderConfigured,
     createAgent,
     runAgent,
     runPromptWithModel,
     runPromptWithModelForImageGeneration,
+    runPromptWithModelForVideoGeneration,
     throwErrorIfNotSuccessful,
     updateUsageTrackerForCharacters,
     validateModel,

package/package.json

@@ -3,7 +3,7 @@
   "publishConfig": {
     "access": "public"
   },
-  "version": "1.3.4",
+  "version": "1.4.1",
   "description": "",
   "main": "dist/cjs/index.js",
   "module": "dist/esm/index.js",