@mcoda/mswarm 0.1.101 → 0.1.103

package/dist/runtime.js

@@ -6,7 +6,7 @@ import { createHash, createHmac, randomBytes, randomUUID } from "node:crypto";
 import { isIP } from "node:net";
 import { GlobalRepository } from "@mcoda/db";
 import { MswarmCodaliExecutor } from "./codali-executor.js";
-import { MSWARM_CAPABILITY_SCHEMA_VERSION, CryptoHelper, assertMswarmSafeRelativePath, validateMswarmArchiveEntry, buildMswarmCapabilityNames, buildMswarmPrivateCapabilityCatalogEntry, buildMswarmLocalArtifactUri, buildMswarmSandboxProfile, defaultMswarmArtifactAccessPolicy, defaultMswarmArtifactRetentionPolicy, projectMswarmPublicCapabilities, validateMswarmGenericJobRequest } from "@mcoda/shared";
+import { GENERATIVE_OPERATION_DEFAULT_PATHS, GENERATIVE_REQUEST_PARAMETER_ALLOWLISTS, MSWARM_CAPABILITY_SCHEMA_VERSION, CryptoHelper, assertMswarmSafeRelativePath, validateMswarmArchiveEntry, buildMswarmCapabilityNames, buildMswarmPrivateCapabilityCatalogEntry, buildMswarmLocalArtifactUri, buildMswarmSandboxProfile, defaultMswarmArtifactAccessPolicy, defaultMswarmArtifactRetentionPolicy, normalizeGenerativeModalities, normalizeSelfHostedGenerativeOperation, projectMswarmPublicCapabilities, validateMswarmGenericJobRequest } from "@mcoda/shared";
 const REGISTRATION_POLL_INTERVAL_MS = 5000;
 const REGISTRATION_POLL_TIMEOUT_MS = 30 * 60000;
 /**
@@ -197,6 +197,17 @@ const SECRET_LOCAL_RUNNER_HEADER_KEYS = new Set(["authorization", "proxy-authori
 const RESERVED_LOCAL_RUNNER_EXTRA_BODY_KEYS = new Set([
     "model",
     "messages",
+    "prompt",
+    "n",
+    "size",
+    "seed",
+    "steps",
+    "negative_prompt",
+    "duration_seconds",
+    "sample_rate",
+    "fps",
+    "video_frames",
+    "high_noise_steps",
     "stream",
     "tools",
     "tool_choice",
@@ -204,6 +215,20 @@ const RESERVED_LOCAL_RUNNER_EXTRA_BODY_KEYS = new Set([
     "max_tokens",
     "temperature"
 ]);
+const REQUIRED_GENERATIVE_OPERATION_REQUEST_PARAMETERS = {
+    "chat.completions": ["model", "messages"],
+    "images.generations": ["model", "prompt", "response_format"],
+    "audio.generations": ["model", "prompt", "response_format"],
+    "videos.generations": ["model", "prompt", "response_format"]
+};
+const AUDIO_GENERATION_RESPONSE_FORMATS = new Set(["wav", "flac", "mp3", "ogg"]);
+const VIDEO_GENERATION_RESPONSE_FORMATS = new Set(["webm", "webp", "avi"]);
+const STABLE_DIFFUSION_CPP_EXTRA_ARGS_TAG_PATTERN = /<sd_cpp_extra_args\b[^>]*>[\s\S]*?<\/sd_cpp_extra_args\s*>/gi;
+const STABLE_DIFFUSION_CPP_EXTRA_ARGS_TAG_FRAGMENT_PATTERN = /<\/?sd_cpp_extra_args\b[^>]*>/gi;
+const DEFAULT_IMAGE_GENERATIVE_OUTPUT_BYTES = 64 * 1024 * 1024;
+const DEFAULT_AUDIO_GENERATIVE_OUTPUT_BYTES = 64 * 1024 * 1024;
+const DEFAULT_VIDEO_GENERATIVE_OUTPUT_BYTES = 256 * 1024 * 1024;
+const GENERATIVE_RESPONSE_JSON_OVERHEAD_BYTES = 64 * 1024;
 const LOCAL_RUNNER_KIND_ALIASES = {
     vllm: "vllm",
     "llama-cpp": "llama-cpp",
@@ -226,6 +251,13 @@ const LOCAL_RUNNER_KIND_ALIASES = {
     text_generation_inference: "tgi",
     custom: "custom"
 };
+const GENERATIVE_RUNNER_KIND_ALIASES = {
+    "stable-diffusion-cpp": "stable-diffusion-cpp",
+    "stable-diffusion.cpp": "stable-diffusion-cpp",
+    stable_diffusion_cpp: "stable-diffusion-cpp",
+    "sd-cpp": "stable-diffusion-cpp",
+    "sd.cpp": "stable-diffusion-cpp"
+};
 const LOCAL_RUNNER_AUTH_MODE_ALIASES = {
     none: "none",
     bearer: "bearer",
@@ -254,6 +286,189 @@ function optionalText(value) {
 function isRecord(value) {
     return typeof value === "object" && value !== null && !Array.isArray(value);
 }
+export function normalizeSelfHostedGenerativeOperationName(value) {
+    const normalized = optionalText(value);
+    return normalized === "chat.completions" ||
+        normalized === "images.generations" ||
+        normalized === "audio.generations" ||
+        normalized === "videos.generations"
+        ? normalized
+        : null;
+}
+export function resolveSelfHostedInvocationOperation(value) {
+    const raw = optionalText(value);
+    if (!raw) {
+        return "chat.completions";
+    }
+    const operation = normalizeSelfHostedGenerativeOperationName(raw);
+    if (!operation) {
+        throw new Error(`unsupported self-hosted operation ${raw}`);
+    }
+    return operation;
+}
+function normalizeGenerativeOperationDescriptor(value) {
+    const record = isRecord(value)
+        ? {
+            ...value,
+            operation: value.operation ?? value.type,
+            requestParameterAllowlist: value.requestParameterAllowlist ??
+                value.request_parameter_allowlist ??
+                value.supported_parameters,
+            responseFormats: value.responseFormats ?? value.response_formats,
+            outputMimeTypes: value.outputMimeTypes ?? value.output_mime_types
+        }
+        : value;
+    const normalized = normalizeSelfHostedGenerativeOperation(record);
+    if (!normalized)
+        return null;
+    const operation = normalized.operation;
+    const requestParameterAllowlist = Array.from(new Set([
+        ...REQUIRED_GENERATIVE_OPERATION_REQUEST_PARAMETERS[operation],
+        ...(normalized.requestParameterAllowlist ??
+            GENERATIVE_REQUEST_PARAMETER_ALLOWLISTS[operation])
+    ]));
+    const responseFormats = normalized.responseFormats?.length
+        ? normalized.responseFormats
+        : operation === "chat.completions"
+            ? undefined
+            : operation === "images.generations"
+                ? ["b64_json"]
+                : operation === "audio.generations"
+                    ? ["wav", "flac", "mp3", "ogg"]
+                    : ["webm", "webp", "avi"];
+    return {
+        operation,
+        path: normalized.path,
+        method: "POST",
+        requestParameterAllowlist,
+        ...(responseFormats?.length ? { responseFormats } : {}),
+        ...(normalized.outputMimeTypes?.length ? { outputMimeTypes: normalized.outputMimeTypes } : {}),
+        ...(normalized.limits ? { limits: normalized.limits } : {})
+    };
+}
+function defaultGenerativeOperationDescriptor(operation, path = GENERATIVE_OPERATION_DEFAULT_PATHS[operation]) {
+    return {
+        operation,
+        path,
+        method: "POST",
+        requestParameterAllowlist: [...GENERATIVE_REQUEST_PARAMETER_ALLOWLISTS[operation]],
+        ...(operation === "images.generations"
+            ? { responseFormats: ["b64_json"] }
+            : operation === "audio.generations"
+                ? { responseFormats: ["wav", "flac", "mp3", "ogg"] }
+                : operation === "videos.generations"
+                    ? { responseFormats: ["webm", "webp", "avi"] }
+                    : {})
+    };
+}
+function normalizeGenerativeOperations(value) {
+    if (value === undefined)
+        return undefined;
+    if (!Array.isArray(value))
+        return [];
+    const operations = value
+        .map((entry) => normalizeGenerativeOperationDescriptor(entry))
+        .filter((entry) => Boolean(entry));
+    const deduplicated = new Map();
+    for (const operation of operations) {
+        if (!deduplicated.has(operation.operation)) {
+            deduplicated.set(operation.operation, operation);
+        }
+    }
+    return Array.from(deduplicated.values());
+}
+function inferGenerativeModalities(operations, direction) {
+    const modalities = new Set();
+    for (const descriptor of operations) {
+        if (direction === "input" || descriptor.operation === "chat.completions") {
+            modalities.add("text");
+        }
+        if (direction === "output" && descriptor.operation === "images.generations") {
+            modalities.add("image");
+        }
+        if (direction === "output" && descriptor.operation === "audio.generations") {
+            modalities.add("audio");
+        }
+        if (direction === "output" && descriptor.operation === "videos.generations") {
+            modalities.add("video");
+        }
+    }
+    return Array.from(modalities);
+}
+function generativeOperationCatalogInput(descriptor) {
+    const limits = descriptor.limits
+        ? {
+            ...(descriptor.limits.maxRequestBytes !== undefined
+                ? { max_request_bytes: descriptor.limits.maxRequestBytes }
+                : {}),
+            ...(descriptor.limits.maxOutputBytes !== undefined
+                ? { max_output_bytes: descriptor.limits.maxOutputBytes }
+                : {}),
+            ...(descriptor.limits.maxPromptChars !== undefined
+                ? { max_prompt_chars: descriptor.limits.maxPromptChars }
+                : {}),
+            ...(descriptor.limits.maxNegativePromptChars !== undefined
+                ? { max_negative_prompt_chars: descriptor.limits.maxNegativePromptChars }
+                : {}),
+            ...(descriptor.limits.maxN !== undefined ? { max_n: descriptor.limits.maxN } : {}),
+            ...(descriptor.limits.minWidth !== undefined
+                ? { min_width: descriptor.limits.minWidth }
+                : {}),
+            ...(descriptor.limits.maxWidth !== undefined
+                ? { max_width: descriptor.limits.maxWidth }
+                : {}),
+            ...(descriptor.limits.minHeight !== undefined
+                ? { min_height: descriptor.limits.minHeight }
+                : {}),
+            ...(descriptor.limits.maxHeight !== undefined
+                ? { max_height: descriptor.limits.maxHeight }
+                : {}),
+            ...(descriptor.limits.maxPixels !== undefined
+                ? { max_pixels: descriptor.limits.maxPixels }
+                : {}),
+            ...(descriptor.limits.minDurationSeconds !== undefined
+                ? { min_duration_seconds: descriptor.limits.minDurationSeconds }
+                : {}),
+            ...(descriptor.limits.maxDurationSeconds !== undefined
+                ? { max_duration_seconds: descriptor.limits.maxDurationSeconds }
+                : {}),
+            ...(descriptor.limits.maxSampleRate !== undefined
+                ? { max_sample_rate: descriptor.limits.maxSampleRate }
+                : {}),
+            ...(descriptor.limits.minFps !== undefined
+                ? { min_fps: descriptor.limits.minFps }
+                : {}),
+            ...(descriptor.limits.maxFps !== undefined
+                ? { max_fps: descriptor.limits.maxFps }
+                : {}),
+            ...(descriptor.limits.minVideoFrames !== undefined
+                ? { min_video_frames: descriptor.limits.minVideoFrames }
+                : {}),
+            ...(descriptor.limits.maxVideoFrames !== undefined
+                ? { max_video_frames: descriptor.limits.maxVideoFrames }
+                : {}),
+            ...(descriptor.limits.maxSteps !== undefined
+                ? { max_steps: descriptor.limits.maxSteps }
+                : {}),
+            ...(descriptor.limits.maxHighNoiseSteps !== undefined
+                ? { max_high_noise_steps: descriptor.limits.maxHighNoiseSteps }
+                : {})
+        }
+        : undefined;
+    return {
+        type: descriptor.operation,
+        path: descriptor.path,
+        method: descriptor.method,
+        supported_parameters: [...descriptor.requestParameterAllowlist],
+        ...(descriptor.responseFormats?.length
+            ? { response_formats: [...descriptor.responseFormats] }
+            : {}),
+        ...(descriptor.outputMimeTypes?.length
+            ? { output_mime_types: [...descriptor.outputMimeTypes] }
+            : {}),
+        ...(limits ? { limits } : {})
+    };
+}
 function normalizeLookupKey(value) {
     return value.trim().toLowerCase();
 }
@@ -316,6 +531,10 @@ function normalizeLocalRunnerKind(value) {
     const normalized = optionalText(value);
     return normalized ? LOCAL_RUNNER_KIND_ALIASES[normalizeLookupKey(normalized)] : undefined;
 }
+function normalizeGenerativeRunnerKind(value) {
+    const normalized = optionalText(value);
+    return normalized ? GENERATIVE_RUNNER_KIND_ALIASES[normalizeLookupKey(normalized)] : undefined;
+}
 function normalizeLocalRunnerAuthMode(value) {
     const normalized = optionalText(value);
     return normalized ? LOCAL_RUNNER_AUTH_MODE_ALIASES[normalizeLookupKey(normalized)] : undefined;
@@ -332,11 +551,14 @@ function normalizeMcodaLocalRunnerConfig(adapter, config) {
     const merged = { ...localRunnerRecord, ...configRecord };
     const authMode = normalizeLocalRunnerAuthMode(merged.authMode) ?? "none";
     const dummyBearerToken = readConfigString(merged, ["dummyBearerToken", "dummyApiKey"]);
+    const generativeRunnerKind = normalizeGenerativeRunnerKind(merged.runnerKind ?? merged.runner_kind);
     return {
         baseUrl: readConfigString(merged, ["baseUrl", "base_url", "endpoint", "apiBaseUrl", "api_base_url"]),
         endpoint: readConfigString(merged, ["endpoint"]),
         apiBaseUrl: readConfigString(merged, ["apiBaseUrl", "api_base_url"]),
-        runnerKind: normalizeLocalRunnerKind(merged.runnerKind) ?? defaultLocalRunnerKindForAdapter(adapter),
+        runnerKind: normalizeLocalRunnerKind(merged.runnerKind ?? merged.runner_kind) ??
+            (generativeRunnerKind ? "custom" : defaultLocalRunnerKindForAdapter(adapter)),
+        generativeRunnerKind,
         authMode,
         dummyBearerToken: authMode === "dummy-bearer" ? dummyBearerToken ?? "local" : dummyBearerToken,
         headers: normalizeStringMap(merged.headers),
@@ -348,7 +570,11 @@ function normalizeMcodaLocalRunnerConfig(adapter, config) {
         supportsStreaming: readConfigBoolean(merged, ["supportsStreaming", "supports_streaming"]),
         supportsTools: readConfigBoolean(merged, ["supportsTools", "supports_tools"]),
         supportsJsonSchema: readConfigBoolean(merged, ["supportsJsonSchema", "supports_json_schema"]),
-        supportsGbnf: readConfigBoolean(merged, ["supportsGbnf", "supports_gbnf"])
+        supportsGbnf: readConfigBoolean(merged, ["supportsGbnf", "supports_gbnf"]),
+        inputModalities: normalizeGenerativeModalities(merged.inputModalities ?? merged.input_modalities),
+        outputModalities: normalizeGenerativeModalities(merged.outputModalities ?? merged.output_modalities),
+        operations: normalizeGenerativeOperations(merged.operations),
+        publicModelId: readConfigString(merged, ["publicModelId", "public_model_id"])
     };
 }
 function parsePositiveInteger(value, fallback) {
@@ -681,8 +907,23 @@ function buildCatalogFingerprint(models) {
         adapter: optionalText(model.adapter) || null,
         source_agent_slug: optionalText(model.source_agent_slug) || null,
         model_id: optionalText(model.model_id) || optionalText(model.model) || null,
+        public_model_id: optionalText(model.public_model_id) || null,
+        upstream_model: optionalText(model.upstream_model) || optionalText(model.model) || null,
         exposed: model.exposed !== false,
         capabilities: normalizeCapabilities(model.capabilities).sort(),
+        input_modalities: [...(model.input_modalities ?? [])].sort(),
+        output_modalities: [...(model.output_modalities ?? [])].sort(),
+        operations: (model.operations ?? [])
+            .map((operation) => ({
+            type: operation.type,
+            path: operation.path,
+            method: operation.method,
+            supported_parameters: [...operation.supported_parameters].sort(),
+            response_formats: [...(operation.response_formats ?? [])].sort(),
+            output_mime_types: [...(operation.output_mime_types ?? [])].sort(),
+            limits: operation.limits ?? null
+        }))
+            .sort((left, right) => left.type.localeCompare(right.type)),
         health_status: normalizeHealthStatus(model.health_status)
     }))
         .sort((left, right) => `${left.provider || ""}:${left.name}`.localeCompare(`${right.provider || ""}:${right.name}`));
@@ -860,6 +1101,69 @@ async function fetchJson(fetchImpl, url, init, timeoutMs) {
         clearTimeout(timeout);
     }
 }
+function generativeOutputByteLimit(operation, configuredLimit) {
+    if (typeof configuredLimit === "number" &&
+        Number.isSafeInteger(configuredLimit) &&
+        configuredLimit > 0) {
+        return configuredLimit;
+    }
+    return operation === "images.generations"
+        ? DEFAULT_IMAGE_GENERATIVE_OUTPUT_BYTES
+        : operation === "audio.generations"
+            ? DEFAULT_AUDIO_GENERATIVE_OUTPUT_BYTES
+            : DEFAULT_VIDEO_GENERATIVE_OUTPUT_BYTES;
+}
+function generativeJsonResponseByteLimit(outputByteLimit) {
+    const encodedLimit = Math.ceil(outputByteLimit / 3) * 4;
+    return Math.min(Number.MAX_SAFE_INTEGER, encodedLimit + GENERATIVE_RESPONSE_JSON_OVERHEAD_BYTES);
+}
+async function readBoundedResponseText(response, maxBytes, overflowMessage) {
+    const declaredLength = Number(response.headers.get("content-length"));
+    if (Number.isFinite(declaredLength) &&
+        declaredLength >= 0 &&
+        declaredLength > maxBytes) {
+        await response.body?.cancel().catch(() => undefined);
+        throw new Error(overflowMessage);
+    }
+    if (!response.body) {
+        return "";
+    }
+    const reader = response.body.getReader();
+    const chunks = [];
+    let totalBytes = 0;
+    try {
+        while (true) {
+            const { done, value } = await reader.read();
+            if (done)
+                break;
+            totalBytes += value.byteLength;
+            if (totalBytes > maxBytes) {
+                await reader.cancel().catch(() => undefined);
+                throw new Error(overflowMessage);
+            }
+            chunks.push(Buffer.from(value));
+        }
+    }
+    finally {
+        reader.releaseLock();
+    }
+    return Buffer.concat(chunks, totalBytes).toString("utf8");
+}
+function base64DecodedByteLength(value) {
+    if (value.length === 0 ||
+        value.length % 4 !== 0 ||
+        !/^[A-Za-z0-9+/]*={0,2}$/.test(value)) {
+        return null;
+    }
+    const padding = value.endsWith("==") ? 2 : value.endsWith("=") ? 1 : 0;
+    return (value.length / 4) * 3 - padding;
+}
+function sanitizeStableDiffusionCppExtraArgsTags(value) {
+    return value
+        .replace(STABLE_DIFFUSION_CPP_EXTRA_ARGS_TAG_PATTERN, "")
+        .replace(STABLE_DIFFUSION_CPP_EXTRA_ARGS_TAG_FRAGMENT_PATTERN, "")
+        .trim();
+}
 export async function readSelfHostedNodeState(statePath) {
     try {
         const content = await readFile(statePath, "utf8");
@@ -1522,17 +1826,20 @@ export function mapOllamaModelToSelfHostedModel(model, config) {
     const parameterSize = optionalText(model.details?.parameter_size);
     const quantizationLevel = optionalText(model.details?.quantization_level);
     const embeddingOnly = isEmbeddingModel(name, family);
+    const operations = [defaultGenerativeOperationDescriptor("chat.completions", "/api/chat")];
+    const supportsVision = isVisionModel(name, family);
     return {
         name,
         provider: "ollama",
         adapter: "ollama",
         model_id: name,
+        upstream_model: name,
         digest: optionalText(model.digest),
         family,
         parameter_size: parameterSize,
         quantization_level: quantizationLevel,
         supports_tools: false,
-        supports_vision: isVisionModel(name, family),
+        supports_vision: supportsVision,
         openai_compatible: false,
         exposed: isModelExposed(name, family, config),
         best_usage: embeddingOnly
@@ -1544,7 +1851,10 @@ export function mapOllamaModelToSelfHostedModel(model, config) {
         reasoning_rating: 4,
         max_complexity: 3,
         health_status: "healthy",
-        metadata_quality: model.details ? "discovered" : "unknown"
+        metadata_quality: model.details ? "discovered" : "unknown",
+        input_modalities: supportsVision ? ["text", "image"] : ["text"],
+        output_modalities: ["text"],
+        operations: operations.map(generativeOperationCatalogInput)
     };
 }
 export function mapMcodaAgentToSelfHostedModel(agent, config) {
@@ -1568,6 +1878,17 @@ export function mapMcodaAgentToSelfHostedModel(agent, config) {
     const supportsTools = optionalBoolean(agent.supportsTools, agent.supports_tools, localRunner?.supportsTools) === true;
     const openaiCompatible = localRunner !== undefined || optionalBoolean(agent.openaiCompatible, agent.openai_compatible) === true;
     const model = defaultModel || slug;
+    const operations = localRunner?.operations ??
+        [defaultGenerativeOperationDescriptor("chat.completions", localRunner ? "/chat/completions" : undefined)];
+    const inferredInputModalities = inferGenerativeModalities(operations, "input");
+    const supportsVision = capabilities.some((capability) => capability.toLowerCase().includes("vision")) ||
+        capabilities.some((capability) => capability.toLowerCase().includes("visual"));
+    const inputModalities = localRunner?.inputModalities ??
+        (supportsVision && operations.some((entry) => entry.operation === "chat.completions")
+            ? Array.from(new Set([...inferredInputModalities, "image"]))
+            : inferredInputModalities);
+    const outputModalities = localRunner?.outputModalities ?? inferGenerativeModalities(operations, "output");
+    const publicModelId = localRunner?.publicModelId;
     return {
         name: slug,
         provider: "mcoda",
@@ -1575,9 +1896,11 @@ export function mapMcodaAgentToSelfHostedModel(agent, config) {
         source_agent_id: optionalText(agent.id),
         source_agent_slug: slug,
         model,
-        model_id: model,
+        model_id: publicModelId || model,
+        public_model_id: publicModelId,
+        upstream_model: model,
         base_url: localRunner?.baseUrl,
-        runner_kind: localRunner?.runnerKind,
+        runner_kind: localRunner?.generativeRunnerKind ?? localRunner?.runnerKind,
         auth_mode: localRunner?.authMode,
         response_format_strategy: localRunner?.responseFormatStrategy,
         health_path: localRunner?.healthPath,
@@ -1587,8 +1910,7 @@ export function mapMcodaAgentToSelfHostedModel(agent, config) {
         max_output_tokens: optionalNumber(agent.maxOutputTokens, agent.max_output_tokens),
         supports_tools: supportsTools,
         supports_streaming: localRunner?.supportsStreaming,
-        supports_vision: capabilities.some((capability) => capability.toLowerCase().includes("vision")) ||
-            capabilities.some((capability) => capability.toLowerCase().includes("visual")),
+        supports_vision: supportsVision,
         supports_json_schema: localRunner?.supportsJsonSchema,
         supports_gbnf: localRunner?.supportsGbnf,
         openai_compatible: openaiCompatible,
@@ -1601,7 +1923,10 @@ export function mapMcodaAgentToSelfHostedModel(agent, config) {
         reasoning_rating: optionalNumber(agent.reasoningRating, agent.reasoning_rating),
         max_complexity: optionalNumber(agent.maxComplexity, agent.max_complexity),
         health_status: healthStatus,
-        metadata_quality: "discovered"
+        metadata_quality: "discovered",
+        input_modalities: inputModalities,
+        output_modalities: outputModalities,
+        operations: operations.map(generativeOperationCatalogInput)
     };
 }
 function parseMcodaAgentListOutput(stdout) {
@@ -1905,6 +2230,8 @@ function mapMcodaAgentToCodaliAgent(agent, fallbackSlug) {
     const localRunner = normalizeMcodaLocalRunnerConfig(adapter, config);
     const configAuthMode = normalizeLocalRunnerAuthMode(config?.authMode);
     const supportsTools = optionalBoolean(agent.supportsTools, agent.supports_tools, localRunner?.supportsTools) === true;
+    const generativeOperations = localRunner?.operations ??
+        [defaultGenerativeOperationDescriptor("chat.completions", localRunner ? "/chat/completions" : undefined)];
     return {
         slug: optionalText(agent.slug) || fallbackSlug,
         adapter,
@@ -1918,6 +2245,7 @@ function mapMcodaAgentToCodaliAgent(agent, fallbackSlug) {
             : configText(config, "apiKey", "api_key"),
         localRunner,
         runnerKind: localRunner?.runnerKind,
+        generativeRunnerKind: localRunner?.generativeRunnerKind,
         authMode: localRunner?.authMode ?? configAuthMode,
         dummyBearerToken: localRunner?.dummyBearerToken,
         headers: localRunner?.headers,
@@ -1933,6 +2261,10 @@ function mapMcodaAgentToCodaliAgent(agent, fallbackSlug) {
         capabilities: normalizeCapabilities(agent.capabilities),
         contextWindow: optionalNumber(agent.contextWindow, agent.context_window) ?? undefined,
         maxOutputTokens: optionalNumber(agent.maxOutputTokens, agent.max_output_tokens) ?? undefined,
+        generativeOperations,
+        inputModalities: localRunner?.inputModalities ?? inferGenerativeModalities(generativeOperations, "input"),
+        outputModalities: localRunner?.outputModalities ?? inferGenerativeModalities(generativeOperations, "output"),
+        publicModelId: localRunner?.publicModelId
     };
 }
 function buildCodaliWorkspace(job) {
@@ -1958,6 +2290,7 @@ const PRE_START_JOB_ERROR_CODES = new Set([
     "selected_agent_unavailable",
     "selected_agent_auth_unavailable",
     "selected_agent_unhealthy",
+    "unsupported_operation",
     "validation_failed",
     "docdex_context_missing",
     "docdex_api_key_missing",
@@ -2148,7 +2481,7 @@ function buildCodaliJob(job) {
     return {
         id: optionalText(payload.id) || job.job_id,
         jobType,
-        input: payload.input ?? messagesToPrompt(job.openai_request.messages),
+        input: payload.input ?? messagesToPrompt(job.openai_request.messages ?? []),
         context: objectRecord(payload.context) || undefined,
         tenant: objectRecord(payload.tenant) || undefined,
         requester: objectRecord(payload.requester) || undefined,
@@ -2182,7 +2515,7 @@ function buildCodaliPolicy(job) {
 }
 function normalizeGatewayConversation(job, value) {
     const record = objectRecord(value);
-    const rawMessages = Array.isArray(record?.messages) ? record.messages : job.openai_request.messages;
+    const rawMessages = Array.isArray(record?.messages) ? record.messages : (job.openai_request.messages ?? []);
     const messages = rawMessages
         .map((entry) => {
         const message = objectRecord(entry);
@@ -2279,7 +2612,7 @@ function buildCodaliGateway(job) {
     const payload = job.codali_gateway;
     if (!payload)
         return undefined;
-    const query = optionalText(payload.query) || messagesToPrompt(job.openai_request.messages);
+    const query = optionalText(payload.query) || messagesToPrompt(job.openai_request.messages ?? []);
     return {
         id: optionalText(payload.id) || job.job_id,
         query,
@@ -3903,6 +4236,7 @@ export class SelfHostedNodeRuntime {
         this.recentFailures = [];
         this.lifecyclePollingDisabled = false;
         this.config = config;
+        this.fetchImpl = deps?.fetchImpl || fetch;
         this.gateway =
             deps?.gateway ||
                 new MswarmSelfHostedNodeClient({
@@ -3911,7 +4245,7 @@ export class SelfHostedNodeRuntime {
                     jobsStartPathTemplate: config.jobsStartPathTemplate,
                     jobsEventsPathTemplate: config.jobsEventsPathTemplate,
                     jobsResultPathTemplate: config.jobsResultPathTemplate,
-                    fetchImpl: deps?.fetchImpl,
+                    fetchImpl: this.fetchImpl,
                     timeoutMs: config.requestTimeoutMs
                 });
         this.mcoda =
@@ -3933,14 +4267,14 @@ export class SelfHostedNodeRuntime {
             deps?.ollama ||
                 new OllamaClient({
                     baseUrl: config.ollamaBaseUrl,
-                    fetchImpl: deps?.fetchImpl,
+                    fetchImpl: this.fetchImpl,
                     timeoutMs: config.requestTimeoutMs
                 });
         this.jobOllama =
             deps?.ollama ||
                 new OllamaClient({
                     baseUrl: config.ollamaBaseUrl,
-                    fetchImpl: deps?.fetchImpl,
+                    fetchImpl: this.fetchImpl,
                     timeoutMs: config.jobTimeoutMs
                 });
         this.capabilityRunner = deps?.capabilityRunner || defaultCommandRunner;
@@ -4025,11 +4359,16 @@ export class SelfHostedNodeRuntime {
         });
     }
     jobResultPayload(result) {
+        const operation = normalizeSelfHostedGenerativeOperationName(result.operation) ?? "chat.completions";
         return {
             ...result,
             node_id: this.config.nodeId,
             ...(result.status === "success"
-                ? { usage: result.usage || openAiUsage(result.openai_response) }
+                ? result.usage
+                    ? { usage: result.usage }
+                    : operation === "chat.completions"
+                        ? { usage: openAiUsage(result.openai_response) }
+                        : {}
                 : {})
         };
     }
@@ -4397,7 +4736,7 @@ export class SelfHostedNodeRuntime {
         }
         const body = {
             model: optionalText(agent.model) || job.openai_request.model,
-            messages: job.openai_request.messages,
+            messages: job.openai_request.messages ?? [],
             tools: job.openai_request.tools,
             stream: false
         };
@@ -4412,7 +4751,7 @@ export class SelfHostedNodeRuntime {
         if (job.openai_request.stop !== undefined)
             body.stop = job.openai_request.stop;
         const url = `${baseUrl.replace(/\/+$/, "")}/chat/completions`;
-        const response = await fetch(url, { method: "POST", headers, body: JSON.stringify(body) });
+        const response = await this.fetchImpl(url, { method: "POST", headers, body: JSON.stringify(body) });
         const text = await response.text();
         let payload = null;
         try {
@@ -4426,6 +4765,335 @@ export class SelfHostedNodeRuntime {
         }
         return payload;
     }
+    generativeOperationForAgent(agent, operation) {
+        const descriptor = agent.generativeOperations.find((entry) => entry.operation === operation);
+        if (!descriptor) {
+            throw new SelfHostedPreStartJobError("unsupported_operation", `selected local mcoda agent ${agent.slug} does not expose ${operation}`);
+        }
+        return descriptor;
+    }
+    prepareOpenAiGenerativePassthrough(job, agent, operation) {
+        const descriptor = this.generativeOperationForAgent(agent, operation);
+        if (job.openai_request.stream === true) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} does not support streaming through the self-hosted relay`);
+        }
+        const baseUrl = optionalText(agent.baseUrl);
+        if (!baseUrl) {
+            throw new SelfHostedPreStartJobError("selected_agent_unavailable", `agent has no local runner base URL for ${operation}`);
+        }
+        let parsedBaseUrl;
+        try {
+            parsedBaseUrl = new URL(baseUrl);
+        }
+        catch {
+            throw new SelfHostedPreStartJobError("selected_agent_unavailable", `agent has an invalid local runner base URL for ${operation}`);
+        }
+        if ((parsedBaseUrl.protocol !== "http:" && parsedBaseUrl.protocol !== "https:") ||
+            parsedBaseUrl.username ||
+            parsedBaseUrl.password ||
+            parsedBaseUrl.search ||
+            parsedBaseUrl.hash) {
+            throw new SelfHostedPreStartJobError("selected_agent_unavailable", `agent local runner base URL is not valid for ${operation}`);
+        }
+        const prompt = optionalText(job.openai_request.prompt);
+        if (!prompt) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} requires a non-empty prompt`);
+        }
+        const rawResponseFormat = job.openai_request.response_format;
+        if (rawResponseFormat !== undefined && rawResponseFormat !== null && typeof rawResponseFormat !== "string") {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} response_format must be a string`);
+        }
+        const responseFormats = descriptor.responseFormats?.map((entry) => entry.toLowerCase()) ??
+            (operation === "images.generations"
+                ? ["b64_json"]
+                : operation === "audio.generations"
+                    ? ["wav", "flac", "mp3", "ogg"]
+                    : ["webm", "webp", "avi"]);
+        const requestedResponseFormat = (optionalText(rawResponseFormat) ||
+            (operation === "images.generations"
+                ? "b64_json"
+                : operation === "audio.generations"
+                    ? responseFormats[0] || "wav"
+                    : responseFormats[0] || "webm")).toLowerCase();
+        if (!responseFormats.includes(requestedResponseFormat) ||
+            (operation === "images.generations"
+                ? requestedResponseFormat !== "b64_json"
+                : operation === "audio.generations"
+                    ? !AUDIO_GENERATION_RESPONSE_FORMATS.has(requestedResponseFormat)
+                    : !VIDEO_GENERATION_RESPONSE_FORMATS.has(requestedResponseFormat))) {
+            throw new SelfHostedPreStartJobError("validation_failed", operation === "images.generations"
+                ? `${operation} requires response_format b64_json through the self-hosted relay`
+                : operation === "audio.generations"
+                    ? `${operation} response_format must be one of wav, flac, mp3, or ogg`
+                    : `${operation} response_format must be one of webm, webp, or avi`);
+        }
+        const body = {};
+        for (const parameter of descriptor.requestParameterAllowlist) {
+            if (parameter === "model") {
+                body.model = optionalText(agent.model) || job.openai_request.model;
+            }
+            else if (parameter === "response_format") {
+                body.response_format = requestedResponseFormat;
+            }
+            else if (job.openai_request[parameter] !== undefined) {
+                body[parameter] = job.openai_request[parameter];
+            }
+        }
+        body.model = optionalText(agent.model) || job.openai_request.model;
+        body.prompt = prompt;
+        body.response_format = requestedResponseFormat;
+        const n = body.n;
+        if (n !== undefined && (!Number.isSafeInteger(n) || n <= 0)) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} n must be a positive integer`);
+        }
+        const durationSeconds = body.duration_seconds;
+        if (durationSeconds !== undefined &&
+            (typeof durationSeconds !== "number" || !Number.isFinite(durationSeconds) || durationSeconds <= 0)) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} duration_seconds must be a positive number`);
+        }
+        const sampleRate = body.sample_rate;
+        if (sampleRate !== undefined &&
+            (!Number.isSafeInteger(sampleRate) || sampleRate <= 0)) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} sample_rate must be a positive integer`);
+        }
+        const fps = body.fps;
+        if (fps !== undefined && (!Number.isSafeInteger(fps) || fps <= 0)) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} fps must be a positive integer`);
+        }
+        const videoFrames = body.video_frames;
+        if (videoFrames !== undefined &&
+            (!Number.isSafeInteger(videoFrames) || videoFrames <= 0)) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} video_frames must be a positive integer`);
+        }
+        if (operation === "videos.generations" &&
+            durationSeconds !== undefined &&
+            videoFrames !== undefined) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} accepts duration_seconds or video_frames, not both`);
+        }
+        const steps = body.steps;
+        if (steps !== undefined && (!Number.isSafeInteger(steps) || steps <= 0)) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} steps must be a positive integer`);
+        }
+        const highNoiseSteps = body.high_noise_steps;
+        if (highNoiseSteps !== undefined &&
+            (!Number.isSafeInteger(highNoiseSteps) || highNoiseSteps <= 0)) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} high_noise_steps must be a positive integer`);
+        }
+        const seed = body.seed;
+        if (seed !== undefined && !Number.isSafeInteger(seed)) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} seed must be an integer`);
+        }
+        if (body.size !== undefined && typeof body.size !== "string") {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} size must be a string`);
+        }
+        if (body.negative_prompt !== undefined && typeof body.negative_prompt !== "string") {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} negative_prompt must be a string`);
+        }
+        const limits = descriptor.limits;
+        if (limits?.maxPromptChars && prompt.length > limits.maxPromptChars) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} prompt exceeds the configured limit`);
+        }
+        if (limits?.maxNegativePromptChars &&
+            typeof body.negative_prompt === "string" &&
+            body.negative_prompt.length > limits.maxNegativePromptChars) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} negative_prompt exceeds the configured limit`);
+        }
+        if (limits?.maxN && typeof n === "number" && n > limits.maxN) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} n exceeds the configured limit`);
+        }
+        if (limits?.minDurationSeconds &&
+            typeof durationSeconds === "number" &&
+            durationSeconds < limits.minDurationSeconds) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} duration_seconds is below the configured minimum`);
+        }
+        if (limits?.maxDurationSeconds &&
+            typeof durationSeconds === "number" &&
+            durationSeconds > limits.maxDurationSeconds) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} duration_seconds exceeds the configured limit`);
+        }
+        if (limits?.maxSampleRate &&
+            typeof sampleRate === "number" &&
+            sampleRate > limits.maxSampleRate) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} sample_rate exceeds the configured limit`);
+        }
+        if (limits?.minFps && typeof fps === "number" && fps < limits.minFps) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} fps is below the configured minimum`);
+        }
+        if (limits?.maxFps && typeof fps === "number" && fps > limits.maxFps) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} fps exceeds the configured limit`);
+        }
+        if (limits?.minVideoFrames &&
+            typeof videoFrames === "number" &&
+            videoFrames < limits.minVideoFrames) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} video_frames is below the configured minimum`);
+        }
+        if (limits?.maxVideoFrames &&
+            typeof videoFrames === "number" &&
+            videoFrames > limits.maxVideoFrames) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} video_frames exceeds the configured limit`);
+        }
+        if (limits?.maxSteps && typeof steps === "number" && steps > limits.maxSteps) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} steps exceeds the configured limit`);
+        }
+        if (limits?.maxHighNoiseSteps &&
+            typeof highNoiseSteps === "number" &&
+            highNoiseSteps > limits.maxHighNoiseSteps) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} high_noise_steps exceeds the configured limit`);
+        }
+        if (typeof body.size === "string") {
+            const sizeMatch = /^([1-9][0-9]*)x([1-9][0-9]*)$/.exec(body.size);
+            if (!sizeMatch) {
+                throw new SelfHostedPreStartJobError("validation_failed", `${operation} size must use WIDTHxHEIGHT`);
+            }
+            const width = Number(sizeMatch[1]);
+            const height = Number(sizeMatch[2]);
+            if ((limits?.minWidth && width < limits.minWidth) ||
+                (limits?.minHeight && height < limits.minHeight)) {
+                throw new SelfHostedPreStartJobError("validation_failed", `${operation} size is below the configured minimum`);
+            }
+            if ((limits?.maxWidth && width > limits.maxWidth) ||
+                (limits?.maxHeight && height > limits.maxHeight) ||
+                (limits?.maxPixels && width * height > limits.maxPixels)) {
+                throw new SelfHostedPreStartJobError("validation_failed", `${operation} size exceeds the configured limit`);
+            }
+        }
+        if (operation === "images.generations" &&
+            agent.generativeRunnerKind === "stable-diffusion-cpp") {
+            const sanitizedPrompt = sanitizeStableDiffusionCppExtraArgsTags(prompt);
+            if (!sanitizedPrompt) {
+                throw new SelfHostedPreStartJobError("validation_failed", `${operation} prompt is empty after removing reserved stable-diffusion.cpp tags`);
+            }
+            const controlledArgs = {};
+            if (typeof body.negative_prompt === "string") {
+                controlledArgs.negative_prompt =
+                    sanitizeStableDiffusionCppExtraArgsTags(body.negative_prompt);
+            }
+            if (typeof seed === "number") {
+                controlledArgs.seed = seed;
+            }
+            if (typeof steps === "number") {
+                controlledArgs.sample_params = { sample_steps: steps };
+            }
+            body.prompt =
+                Object.keys(controlledArgs).length > 0
+                    ? `${sanitizedPrompt}\n<sd_cpp_extra_args>${JSON.stringify(controlledArgs)}</sd_cpp_extra_args>`
+                    : sanitizedPrompt;
+            delete body.negative_prompt;
+            delete body.seed;
+            delete body.steps;
+        }
+        const serializedBody = JSON.stringify(body);
+        if (limits?.maxRequestBytes && Buffer.byteLength(serializedBody, "utf8") > limits.maxRequestBytes) {
+            throw new SelfHostedPreStartJobError("validation_failed", `${operation} request exceeds the configured byte limit`);
+        }
+        const headers = {};
+        const bearer = agent.authMode === "bearer"
+            ? optionalText(agent.apiKey)
+            : agent.authMode === "dummy-bearer"
+                ? optionalText(agent.dummyBearerToken) || "local"
+                : undefined;
+        if (bearer) {
+            headers.authorization = `Bearer ${bearer}`;
+        }
+        for (const [key, value] of Object.entries(agent.headers ?? {})) {
+            if (typeof value === "string") {
+                headers[key.toLowerCase()] = value;
+            }
+        }
+        headers["content-type"] = "application/json";
+        return {
+            descriptor,
+            url: `${baseUrl.replace(/\/+$/, "")}${descriptor.path}`,
+            headers,
+            body
+        };
+    }
+    async executeOpenAiGenerativePassthrough(prepared, operation) {
+        const outputByteLimit = generativeOutputByteLimit(operation, prepared.descriptor.limits?.maxOutputBytes);
+        const jsonResponseByteLimit = generativeJsonResponseByteLimit(outputByteLimit);
+        const controller = new AbortController();
+        const timeout = setTimeout(() => controller.abort(), this.config.jobTimeoutMs);
+        let payload;
+        try {
+            const response = await this.fetchImpl(prepared.url, {
+                method: "POST",
+                headers: prepared.headers,
+                body: JSON.stringify(prepared.body),
+                signal: controller.signal
+            });
+            if (!response.ok) {
+                await response.body?.cancel().catch(() => undefined);
+                throw new Error(`local ${operation} runner request failed with status ${response.status}`);
+            }
+            const responseText = await readBoundedResponseText(response, jsonResponseByteLimit, `local ${operation} runner response exceeds the configured byte limit`);
+            try {
+                payload = JSON.parse(responseText);
+            }
+            catch {
+                throw new Error(`local ${operation} runner returned invalid JSON`);
+            }
+        }
+        catch (error) {
+            if (error instanceof Error && error.name === "AbortError") {
+                throw new Error(`local ${operation} runner timeout`);
+            }
+            throw error;
+        }
+        finally {
+            clearTimeout(timeout);
+        }
+        if (!isRecord(payload) || !Array.isArray(payload.data) || payload.data.length === 0) {
+            throw new Error(`local ${operation} runner returned an invalid base64 response`);
+        }
+        const limits = prepared.descriptor.limits;
+        if (limits?.maxN && payload.data.length > limits.maxN) {
+            throw new Error(`local ${operation} runner returned too many outputs`);
+        }
+        let decodedOutputBytes = 0;
+        const normalizedData = [];
+        for (const item of payload.data) {
+            const encodedPayload = isRecord(item) && operation === "audio.generations"
+                ? optionalText(item.b64_audio) || optionalText(item.b64_json)
+                : isRecord(item) && operation === "videos.generations"
+                    ? optionalText(item.b64_video) || optionalText(item.b64_json)
+                    : isRecord(item)
+                        ? optionalText(item.b64_json)
+                        : null;
+            if (!isRecord(item) || !encodedPayload) {
+                throw new Error(`local ${operation} runner returned an invalid base64 response`);
+            }
+            const decodedBytes = base64DecodedByteLength(encodedPayload);
+            if (decodedBytes === null) {
+                throw new Error(`local ${operation} runner returned an invalid base64 response`);
+            }
+            decodedOutputBytes += decodedBytes;
+            if (decodedOutputBytes > outputByteLimit) {
+                throw new Error(`local ${operation} runner response exceeds the configured byte limit`);
+            }
+            const mimeType = optionalText(item.mime_type);
+            if (mimeType &&
+                prepared.descriptor.outputMimeTypes?.length &&
+                !prepared.descriptor.outputMimeTypes.includes(mimeType.toLowerCase())) {
+                throw new Error(`local ${operation} runner returned unsupported mime_type ${mimeType}`);
+            }
+            if ((operation === "audio.generations" || operation === "videos.generations") &&
+                !optionalText(operation === "audio.generations" ? item.b64_audio : item.b64_video) &&
+                optionalText(item.b64_json)) {
+                const normalizedItem = {
+                    ...item,
+                    [operation === "audio.generations" ? "b64_audio" : "b64_video"]: encodedPayload
+                };
+                delete normalizedItem.b64_json;
+                normalizedData.push(normalizedItem);
+            }
+            else {
+                normalizedData.push(item);
+            }
+        }
+        return operation === "audio.generations" || operation === "videos.generations"
+            ? { ...payload, data: normalizedData }
+            : payload;
+    }
     async resolveMcodaAgentForJob(job) {
         const selectedSourceAgentSlug = optionalText(job.source_agent_slug);
         const selectedAgentSlug = optionalText(job.agent_slug);
@@ -4647,6 +5315,7 @@ export class SelfHostedNodeRuntime {
         const startedAt = Date.now();
         this.beginExecutionTelemetry("llm");
         let selectedAgent;
+        let operation = "chat.completions";
         let jobStarted = false;
         const progressEvents = [];
         const streamEvents = [];
@@ -4676,6 +5345,9 @@ export class SelfHostedNodeRuntime {
             const result = {
                 job_id: job.job_id,
                 request_id: job.request_id,
+                ...(normalizeSelfHostedGenerativeOperationName(job.operation)
+                    ? { operation: normalizeSelfHostedGenerativeOperationName(job.operation) }
+                    : {}),
                 status: "failed",
                 pre_start_failure: true,
                 error: { code: "validation_failed", message: "job node_id does not match this node" }
@@ -4689,7 +5361,20 @@ export class SelfHostedNodeRuntime {
             return result;
         }
         try {
+            try {
+                operation = resolveSelfHostedInvocationOperation(job.operation);
+            }
+            catch {
+                throw new SelfHostedPreStartJobError("unsupported_operation", `unsupported self-hosted operation ${optionalText(job.operation) || String(job.operation)}`);
+            }
             if (job.provider === "ollama") {
+                if (operation !== "chat.completions") {
+                    throw new SelfHostedPreStartJobError("unsupported_operation", `Ollama relay jobs do not support ${operation}`);
+                }
+                const chatMessages = job.openai_request.messages;
+                if (!Array.isArray(chatMessages)) {
+                    throw new SelfHostedPreStartJobError("validation_failed", "chat.completions requires messages");
+                }
                 const options = {};
                 if (job.openai_request.temperature !== undefined)
                     options.temperature = job.openai_request.temperature;
@@ -4702,7 +5387,7 @@ export class SelfHostedNodeRuntime {
                 await acknowledgeStarted();
                 const ollamaResult = await this.jobOllama.chat({
                     model: job.model || job.openai_request.model,
-                    messages: job.openai_request.messages,
+                    messages: chatMessages,
                     options,
                     format: resolveOllamaResponseFormat(job.openai_request.response_format)
                 });
@@ -4731,6 +5416,7 @@ export class SelfHostedNodeRuntime {
                 const invocationResult = {
                     job_id: job.job_id,
                     request_id: job.request_id,
+                    operation,
                     status: "success",
                     openai_response: buildOpenAIChatCompletion({
                         requestId: job.request_id,
@@ -4752,12 +5438,43 @@ export class SelfHostedNodeRuntime {
                 });
                 return invocationResult;
             }
-            const taskPreview = messagesToPrompt(job.openai_request.messages);
-            if (!taskPreview) {
-                throw new Error("mcoda invocation prompt is empty");
-            }
             const agent = await this.resolveMcodaAgentForJob(job);
             selectedAgent = agent;
+            this.generativeOperationForAgent(agent, operation);
+            if (operation === "images.generations" ||
+                operation === "audio.generations" ||
+                operation === "videos.generations") {
+                const prepared = this.prepareOpenAiGenerativePassthrough(job, agent, operation);
+                await acknowledgeStarted(agent);
+                await recordProgress({
+                    type: "agent_selected",
+                    operation,
+                    job_id: job.job_id,
+                    request_id: job.request_id,
+                    agent_slug: agent.slug,
+                    adapter: agent.adapter,
+                    local_model: optionalText(agent.model) || job.openai_request.model
+                });
+                const passthrough = await this.executeOpenAiGenerativePassthrough(prepared, operation);
+                this.finishExecutionTelemetry({ executionClass: "llm", startedAt, ok: true });
+                return {
+                    job_id: job.job_id,
+                    request_id: job.request_id,
+                    operation,
+                    status: "success",
+                    openai_response: passthrough,
+                    ...(progressEvents.length ? { progress_events: progressEvents } : {}),
+                    timing: { local_latency_ms: Date.now() - startedAt }
+                };
+            }
+            const chatMessages = job.openai_request.messages;
+            if (!Array.isArray(chatMessages)) {
+                throw new SelfHostedPreStartJobError("validation_failed", "chat.completions requires messages");
+            }
+            const taskPreview = messagesToPrompt(chatMessages);
+            if (!taskPreview) {
+                throw new SelfHostedPreStartJobError("validation_failed", "mcoda invocation prompt is empty");
+            }
             validateRequiredDocdexContext(job, options.attachedMswarmApiKey);
             const attachedMswarmApiKey = attachedMswarmApiKeyForDocdex(job, options.attachedMswarmApiKey);
             await acknowledgeStarted(agent);
@@ -4779,6 +5496,7 @@ export class SelfHostedNodeRuntime {
                 const passthroughResult = {
                     job_id: job.job_id,
                     request_id: job.request_id,
+                    operation,
                     status: "success",
                     openai_response: passthrough,
                     ...(progressEvents.length ? { progress_events: progressEvents } : {}),
@@ -4793,14 +5511,16 @@ export class SelfHostedNodeRuntime {
                 jobId: job.job_id,
                 requestId: job.request_id,
                 model: job.openai_request.model,
-                messages: job.openai_request.messages,
+                messages: chatMessages,
                 agent,
                 workspace: buildCodaliWorkspace(job),
                 docdex: buildCodaliDocdex(job),
                 attachedMswarmApiKey,
                 policy: buildCodaliPolicy(job),
                 temperature: job.openai_request.temperature,
-                responseFormat: job.openai_request.response_format ?? null,
+                responseFormat: isRecord(job.openai_request.response_format)
+                    ? job.openai_request.response_format
+                    : null,
                 stream: job.openai_request.stream === true,
                 codaliGateway,
                 codaliJob,
@@ -4860,6 +5580,7 @@ export class SelfHostedNodeRuntime {
             const result = {
                 job_id: job.job_id,
                 request_id: job.request_id,
+                operation,
                 status: "success",
                 openai_response: buildOpenAIChatCompletion({
                     requestId: job.request_id,
@@ -4932,6 +5653,7 @@ export class SelfHostedNodeRuntime {
             const result = {
                 job_id: job.job_id,
                 request_id: job.request_id,
+                operation,
                 status: "failed",
                 ...(!jobStarted ? { pre_start_failure: true } : {}),
                 error: {
@@ -5147,6 +5869,9 @@ export class SelfHostedNodeRuntime {
         if (!job) {
             return { executed: false };
         }
+        const chatStreaming = (normalizeSelfHostedGenerativeOperationName(job.operation) ?? "chat.completions") ===
+            "chat.completions" &&
+            job.openai_request.stream === true;
         const pendingStreamEvents = [];
         let streamEventForwardingFailed = false;
         let forwardedStreamEventCount = 0;
@@ -5193,7 +5918,7 @@ export class SelfHostedNodeRuntime {
                 }
             },
             onOpenAIChunk: async (chunk) => {
-                if (job.openai_request.stream !== true || streamEventForwardingFailed) {
+                if (!chatStreaming || streamEventForwardingFailed) {
                     return;
                 }
                 pendingStreamEvents.push(chunk);
@@ -5212,6 +5937,7 @@ export class SelfHostedNodeRuntime {
             result = {
                 job_id: job.job_id,
                 request_id: job.request_id,
+                operation: result.operation,
                 status: "failed",
                 error: { code: SELF_HOSTED_PROTOCOL_MISMATCH_CODE, message: error.message },
                 ...(result.stream_events?.length ? { stream_events: result.stream_events } : {}),
@@ -5219,7 +5945,7 @@ export class SelfHostedNodeRuntime {
                 timing: result.timing
             };
         }
-        const postedResult = job.openai_request.stream === true
+        const postedResult = chatStreaming
             ? streamEventForwardingFailed
                 ? {
                     ...result,