@truefoundry/ai-gateway-sdk 0.0.0-stage → 0.0.2

package/dist/globals.js

@@ -0,0 +1,782 @@
+/**
+ * ================================
+ * Inherited upstream file
+ * But need few custom changes for TFY
+ * 1. Chaning name of providers
+ * 2. Add response format tool name
+ * 3. Add gateway cached credentials prefix
+ * 4. Add model unsupported params
+ * ================================
+ */
+/**
+ * ================================
+ * NOTE: Keep this aligned with the tfy header prefix
+ * ================================
+ */
+export const POWERED_BY = 'tfy';
+export const API_TFY_AI = 'api.tfy.ai';
+export const MAX_RETRY_LIMIT_MS = 60 * 1000; // 60 seconds
+export const POSSIBLE_RETRY_STATUS_HEADERS = ['retry-after-ms', 'x-ms-retry-after-ms', 'retry-after'];
+// Jitter constant to stagger operations across multiple pods during rollout
+// This prevents thundering herd problems when all pods start simultaneously
+export const POD_STARTUP_JITTER_MS = 15000; // 15 seconds
+/**
+ * ================================
+ * NOTE: Add response format tool name for anthropic and gateway cached credentials prefix
+ * ================================
+ */
+export const RESPONSE_FORMAT_TOOL_NAME = 'json_extractor';
+export const GATEWAY_V2_CACHE_PREFIX = 'v2-cache.';
+export const GATEWAY_V2_COMMON_PREFIX = 'v2-common.';
+export const GATEWAY_V2_ENFORCEMENT_PREFIX = 'v2-enforcement.';
+// Same blocked rules, plus the limits each block was computed against. Deliberately not a
+// sub-prefix of the above, so older gateways never receive these keys.
+export const GATEWAY_V2_ENFORCEMENT_LIMITS_PREFIX = 'v2-enforcement-with-limits.';
+export const GATEWAY_V2_METRICS_LOADBALANCER_PREFIX = 'v2-metrics-loadbalancer.';
+export const GATEWAY_V2_STATISTICS_PREFIX = 'v2-statistics.';
+export const PROMPTS_CACHE_NATS_KEY_PREFIX = 'prompts-cache.';
+export const GIT_SKILL_CONTENT_CACHE_NATS_KEY_PREFIX = 'git-skill-content-cache.';
+export const GATEWAY_SECRET_CACHE_NATS_KEY_PREFIX = 'gateway-secret-cache.';
+export const GATEWAY_APPROVAL_CACHE_NATS_KEY_PREFIX = 'gateway-approval-cache.';
+// V2 Cache Entity Types
+export const V2_CACHE_ENTITIES = {
+    PROVIDER_CREDENTIALS: 'providerCredentials',
+    PROMPTS_MANIFESTS: 'promptsManifests',
+    STICKY_FALLBACK: 'sticky_fallback',
+};
+// V2 Enforcement Entity Types (for blocked state)
+// `BUDGET` is the v1 engine map; `BUDGET_V2` is the v2 engine map at
+// `v2-enforcement.{tenant}.BUDGET_V2`. The two engines run side-by-side.
+export const V2_ENFORCEMENT_ENTITIES = {
+    BUDGET: 'BUDGET',
+    BUDGET_V2: 'BUDGET_V2',
+    RATELIMIT: 'RATELIMIT',
+};
+// V2 Common Entity Types (shared across tenants)
+export const V2_COMMON_ENTITIES = {
+    MODEL_CONFIG: 'model-config',
+    PROVIDER_CONFIG: 'provider-config',
+};
+// V2 Statistics Entity Types (old flow - bucket data)
+export const V2_STATISTICS_ENTITIES = {
+    RATE_LIMIT_BUCKET: 'gateway-rate-limit-bucket',
+    BUDGET_BUCKET: 'gateway-budget-bucket',
+    LOADBALANCE_BUCKET: 'gateway-loadbalance-bucket',
+};
+export const HEADER_KEYS = {
+    API_KEY: `x-${POWERED_BY}-api-key`,
+    MODE: `x-${POWERED_BY}-mode`,
+    RETRIES: `x-${POWERED_BY}-retry-count`,
+    PROVIDER: `x-${POWERED_BY}-provider`,
+    CONFIG: `x-${POWERED_BY}-config`,
+    TRACE_ID: `x-${POWERED_BY}-trace-id`,
+    CACHE: `x-${POWERED_BY}-cache`,
+    METADATA: `x-${POWERED_BY}-metadata`,
+    FORWARD_HEADERS: `x-${POWERED_BY}-forward-headers`,
+    CUSTOM_HOST: `x-${POWERED_BY}-custom-host`,
+    SSE_COMMENTS: `x-${POWERED_BY}-sse-comments`,
+    REQUEST_TIMEOUT: `x-${POWERED_BY}-request-timeout`,
+    STRICT_OPEN_AI_COMPLIANCE: `x-${POWERED_BY}-strict-open-ai-compliance`,
+    CONTENT_TYPE: `Content-Type`,
+    VIRTUAL_KEY: `x-${POWERED_BY}-virtual-key`,
+    /**
+     * ================================
+     * NOTE: Add guardrails header
+     * ================================
+     */
+    GUARDRAILS: `x-${POWERED_BY}-guardrails`,
+    /**
+     * Guardrail selectors for the coding-assistant hook endpoint (`/hooks/claude-code`).
+     * A plain JSON array of selectors (e.g. `["group/name"]`) — the hook event decides which
+     * pipeline they run through, so this is intentionally a distinct header from the object-shaped
+     * GUARDRAILS header used on the model request path.
+     */
+    HOOK_GUARDRAILS: `x-${POWERED_BY}-hook-guardrails`,
+    TTFT_TIMEOUT: `x-${POWERED_BY}-ttft-timeout-ms`,
+};
+export const RESPONSE_HEADER_KEYS = {
+    RETRY_ATTEMPT_COUNT: `x-${POWERED_BY}-retry-attempt-count`,
+    LAST_USED_OPTION_INDEX: `x-${POWERED_BY}-last-used-option-index`,
+    LAST_USED_OPTION_PARAMS: `x-${POWERED_BY}-last-used-option-params`,
+    CACHE_STATUS: `x-${POWERED_BY}-cache-status`,
+    TRACE_ID: `x-${POWERED_BY}-trace-id`,
+    /**
+     * ================================
+     * NOTE: Add model latency header
+     * ================================
+     */
+    MODEL_LATENCY: `x-${POWERED_BY}-model-latency`,
+    APPLIED_RULES: `x-${POWERED_BY}-applied-rules`,
+    FEEDBACK_TARGET_ID: `x-${POWERED_BY}-feedback-target-id`,
+};
+export const RETRY_STATUS_CODES = [429, 500, 502, 503, 504];
+export const MAX_RETRIES = 5;
+export const REQUEST_TIMEOUT_STATUS_CODE = 408;
+export const PRECONDITION_CHECK_FAILED_STATUS_CODE = 412;
+export const PROVIDER_MALFORMED_RESPONSE_STATUS_CODE = 500;
+/**
+ * ================================
+ * NOTE: Change name of vertex-ai to vertex, bedrock to aws-bedrock, google to gcp, azure-ai to azure-foundry, and add google-vertex
+ * ================================
+ */
+export const DATABRICKS = 'databricks';
+export const OPEN_AI = 'openai';
+export const COHERE = 'cohere';
+export const AZURE_OPEN_AI = 'azure-openai';
+export const AZURE_AI_INFERENCE = 'azure-foundry';
+export const MICROSOFT_FOUNDRY = 'microsoft-foundry';
+export const ANTHROPIC = 'anthropic';
+/** Provider names whose deployments are priced by foundation_model / deployment_type-derived region, Azure-style. */
+export const AZURE_OPENAI_COMPATIBLE_COST_PROVIDERS = new Set([AZURE_OPEN_AI, MICROSOFT_FOUNDRY]);
+/** Prefix used by Anthropic to identify server-side tool call IDs (web search, code execution, etc.). */
+export const ANTHROPIC_SERVER_TOOL_CALL_ID_PREFIX = 'srvtoolu_';
+/**
+ * Normalizes a Claude model ID by stripping provider-specific prefixes, version
+ * suffixes, and separators so version extraction works uniformly.
+ *
+ * Bedrock prefixes:   "anthropic.", "us.anthropic.", "global.anthropic-"
+ * Vertex prefixes:    "anthropic/"
+ * Bedrock suffixes:   "-v1:0", "-v2:0", "-v1-0", "-v1"
+ * Vertex separators:  "@" between model name and date (e.g. claude-opus-4@20250514)
+ *
+ * E.g. "global.anthropic-claude-sonnet-4-20250514-v1-0" → "claude-sonnet-4-20250514"
+ *      "anthropic/claude-opus-4@20250514"               → "claude-opus-4-20250514"
+ */
+function normalizeClaudeModelId(model) {
+    const claudeIdx = model.indexOf('claude');
+    const stripped = claudeIdx > 0 ? model.slice(claudeIdx) : model;
+    return stripped.replace(/-v\d+(?:[:-]\d+)?$/, '').replace(/@/g, '-');
+}
+export function isClaudeOpus45(model) {
+    if (!model.includes('claude'))
+        return false;
+    const normalized = normalizeClaudeModelId(model);
+    return /(?:opus-4-5|4-5-opus)\b/.test(normalized);
+}
+/**
+ * Extracts the major.minor version from a normalized Claude model ID.
+ *
+ * Handles both naming conventions:
+ *   New format (variant before version): claude-opus-4, claude-sonnet-4-5-20250929
+ *   Old format (version before variant): claude-3-opus-20240229, claude-3-5-sonnet-20241022
+ *
+ * Minor version is distinguished from a date suffix by being a single digit
+ * followed by either '-' or end-of-string (dates are 8 digits).
+ */
+function extractClaudeVersion(normalized) {
+    // New format: claude-{variant}-{major}[-{minor}][-{date}]
+    // e.g. claude-opus-4, claude-opus-4-5, claude-opus-4-6-20260101
+    const newFmt = normalized.match(/^claude-[a-z]+-(\d+)(?:-(\d)(?:-|$))?/);
+    if (newFmt) {
+        return { major: parseInt(newFmt[1]), minor: newFmt[2] ? parseInt(newFmt[2]) : 0 };
+    }
+    // Old format: claude-{major}[-{minor}]-{variant}[-{date}]
+    // e.g. claude-3-opus-20240229, claude-3-5-sonnet-20241022, claude-4-opus-20250514
+    const oldFmt = normalized.match(/^claude-(\d+)(?:-(\d))?-[a-z]/);
+    if (oldFmt) {
+        return { major: parseInt(oldFmt[1]), minor: oldFmt[2] ? parseInt(oldFmt[2]) : 0 };
+    }
+    return null;
+}
+const ADAPTIVE_THINKING_MIN_VERSION = { major: 4, minor: 6 };
+/**
+ * Returns true when the model is a Claude model that should use adaptive thinking
+ * ({type: 'adaptive'}, no budget_tokens).
+ *
+ * Versions >= 4.6 use adaptive thinking. Older versions use {type: 'enabled', budget_tokens}.
+ * If the version can't be parsed, defaults to legacy (enabled) as a safe fallback.
+ *
+ * Handles Bedrock-style model IDs (e.g. "global.anthropic-claude-sonnet-4-20250514-v1-0")
+ * and Vertex AI model IDs (e.g. "anthropic/claude-opus-4@20250514").
+ */
+export function isClaudeAdaptiveThinkingModel(model) {
+    if (!model.includes('claude'))
+        return false;
+    const normalized = normalizeClaudeModelId(model);
+    const version = extractClaudeVersion(normalized);
+    if (!version)
+        return false;
+    return (version.major > ADAPTIVE_THINKING_MIN_VERSION.major ||
+        (version.major === ADAPTIVE_THINKING_MIN_VERSION.major && version.minor >= ADAPTIVE_THINKING_MIN_VERSION.minor));
+}
+const BEDROCK_EAGER_STREAMING_MIN_VERSION = { major: 4, minor: 6 };
+/**
+ * Returns true when the Bedrock model natively supports eager_input_streaming on tools.
+ *
+ * Older models (< 4.6, e.g. claude-opus-4-5) reject eager_input_streaming with
+ * "Extra inputs are not permitted". Newer models (>= 4.6) accept it directly.
+ */
+export function isBedrockEagerInputStreamingSupported(model) {
+    if (!model.includes('claude'))
+        return false;
+    const normalized = normalizeClaudeModelId(model);
+    const version = extractClaudeVersion(normalized);
+    if (!version)
+        return false;
+    return (version.major > BEDROCK_EAGER_STREAMING_MIN_VERSION.major ||
+        (version.major === BEDROCK_EAGER_STREAMING_MIN_VERSION.major &&
+            version.minor >= BEDROCK_EAGER_STREAMING_MIN_VERSION.minor));
+}
+export const ANYSCALE = 'anyscale';
+export const PALM = 'palm';
+export const TOGETHER_AI = 'together-ai';
+export const GOOGLE = 'gcp';
+export const GOOGLE_VERTEX_AI = 'vertex';
+export const HUGGING_FACE = 'huggingface';
+export const GOOGLE_VERTEX = 'google-vertex';
+export const GOOGLE_GEMINI = 'google-gemini';
+export const PERPLEXITY_AI = 'perplexity-ai';
+export const REKA_AI = 'reka-ai';
+export const MISTRAL_AI = 'mistral-ai';
+export const DEEPINFRA = 'deepinfra';
+export const NCOMPASS = 'ncompass';
+export const STABILITY_AI = 'stability-ai';
+export const NOMIC = 'nomic';
+export const OLLAMA = 'ollama';
+export const AI21 = 'ai21';
+export const BEDROCK = 'aws-bedrock';
+export const AWS_BEDROCK_MANTLE = 'aws-bedrock-mantle';
+export const AWS_CLAUDE_PLATFORM = 'aws-claude-platform';
+/**
+ * AWS Partner Network (APN) user-agent tag for TrueFoundry. AWS reads this from
+ * the User-Agent header to attribute API usage to our APN partner code. It must
+ * be appended to every outbound AWS API request (Bedrock model invocations,
+ * Bedrock guardrails, Bedrock Mantle, Claude on AWS, SageMaker, STS, ...).
+ * Ref: https://docs.aws.amazon.com/PRM/latest/aws-prm-onboarding-guide/user-agent-samples.html
+ */
+export const AWS_APN_USER_AGENT = 'APN_1.1/pc_bjgqgh3eh39h62mdfdwks5bwp$';
+/**
+ * Appends {@link AWS_APN_USER_AGENT} to a headers object, preserving any existing
+ * User-Agent value. `user-agent` is an unsignable SigV4 header, so this is safe
+ * to call before signing without invalidating the signature. Idempotent and
+ * case-insensitive on the existing header key.
+ */
+export const appendApnUserAgent = (headers) => {
+    const existingKey = Object.keys(headers).find(k => k.toLowerCase() === 'user-agent');
+    const existing = existingKey ? headers[existingKey] : undefined;
+    if (existing?.includes(AWS_APN_USER_AGENT)) {
+        return headers;
+    }
+    if (existingKey && existingKey !== 'User-Agent') {
+        delete headers[existingKey];
+    }
+    headers['User-Agent'] = existing ? `${existing} ${AWS_APN_USER_AGENT}` : AWS_APN_USER_AGENT;
+    return headers;
+};
+export const GROQ = 'groq';
+export const SEGMIND = 'segmind';
+export const JINA = 'jina';
+export const FIREWORKS_AI = 'fireworks-ai';
+export const WORKERS_AI = 'workers-ai';
+export const MOONSHOT = 'moonshot';
+export const OPENROUTER = 'openrouter';
+export const LINGYI = 'lingyi';
+export const ZHIPU = 'zhipu';
+export const NOVITA_AI = 'novita-ai';
+export const MONSTERAPI = 'monsterapi';
+export const DEEPSEEK = 'deepseek';
+export const PREDIBASE = 'predibase';
+export const TRITON = 'triton';
+export const VOYAGE = 'voyage';
+export const GITHUB = 'github';
+export const DEEPBRICKS = 'deepbricks';
+export const SILICONFLOW = 'siliconflow';
+export const CEREBRAS = 'cerebras';
+export const WAFER = 'wafer';
+export const TYPESAFE = 'typesafe';
+export const INFERENCENET = 'inference-net';
+export const SAMBANOVA = 'sambanova';
+export const LEMONFOX_AI = 'lemonfox-ai';
+export const UPSTAGE = 'upstage';
+export const LAMBDA = 'lambda';
+export const DASHSCOPE = 'dashscope';
+export const X_AI = 'xai';
+export const BASETEN = 'baseten';
+export const TRUEFOUNDRY_MODELS = 'truefoundry-models';
+export const CORTEX = 'cortex';
+export const SAGEMAKER = 'sagemaker';
+export const NEBIUS = 'nebius';
+export const RECRAFTAI = 'recraft-ai';
+export const MILVUS = 'milvus';
+export const REPLICATE = 'replicate';
+export const LEPTON = 'lepton';
+export const KLUSTER_AI = 'kluster-ai';
+export const NSCALE = 'nscale';
+export const HYPERBOLIC = 'hyperbolic';
+export const BYTEZ = 'bytez';
+export const FEATHERLESS_AI = 'featherless-ai';
+export const KRUTRIM = 'krutrim';
+export const QDRANT = 'qdrant';
+export const THREE_ZERO_TWO_AI = '302ai';
+export const CLOUDERA = 'cloudera';
+export const CARTESIA = 'cartesia';
+export const DEEPGRAM = 'deepgram';
+/**
+ * ================================
+ * NOTE: Add self-hosted provider
+ * ================================
+ */
+export const SELF_HOSTED = 'self-hosted';
+/**
+ * ================================
+ * OpenAI-Compatible Provider Base URLs
+ * Base URLs for OpenAI-compatible providers
+ * ================================
+ */
+export const OPENAI_COMPATIBLE_BASE_URLS = {
+    xai: 'https://api.x.ai/v1',
+};
+/**
+ * ================================
+ * Provider Name to Constant Mapping
+ * Maps provider names from providerOptions to constants from globals.ts
+ * ================================
+ */
+export const PROVIDER_CONSTANTS = {
+    xai: X_AI,
+};
+export const VALID_PROVIDERS = [
+    ANTHROPIC,
+    ANYSCALE,
+    AZURE_OPEN_AI,
+    COHERE,
+    GOOGLE,
+    GOOGLE_VERTEX_AI,
+    MISTRAL_AI,
+    OPEN_AI,
+    PALM,
+    PERPLEXITY_AI,
+    REKA_AI,
+    TOGETHER_AI,
+    DEEPINFRA,
+    NCOMPASS,
+    STABILITY_AI,
+    NOMIC,
+    OLLAMA,
+    AI21,
+    BEDROCK,
+    GROQ,
+    SEGMIND,
+    JINA,
+    FIREWORKS_AI,
+    WORKERS_AI,
+    MOONSHOT,
+    OPENROUTER,
+    LINGYI,
+    ZHIPU,
+    NOVITA_AI,
+    MONSTERAPI,
+    DEEPSEEK,
+    PREDIBASE,
+    TRITON,
+    VOYAGE,
+    AZURE_AI_INFERENCE,
+    MICROSOFT_FOUNDRY,
+    GITHUB,
+    DEEPBRICKS,
+    SILICONFLOW,
+    HUGGING_FACE,
+    CEREBRAS,
+    WAFER,
+    TYPESAFE,
+    INFERENCENET,
+    SAMBANOVA,
+    LEMONFOX_AI,
+    UPSTAGE,
+    LAMBDA,
+    DASHSCOPE,
+    X_AI,
+    BASETEN,
+    TRUEFOUNDRY_MODELS,
+    CORTEX,
+    SAGEMAKER,
+    NEBIUS,
+    RECRAFTAI,
+    MILVUS,
+    REPLICATE,
+    POWERED_BY,
+    LEPTON,
+    KLUSTER_AI,
+    NSCALE,
+    HYPERBOLIC,
+    BYTEZ,
+    FEATHERLESS_AI,
+    KRUTRIM,
+    QDRANT,
+    THREE_ZERO_TWO_AI,
+];
+/**
+ * ================================
+ * Query parameter constants
+ * ================================
+ */
+export const QUERY_PARAMS = {
+    REFRESH: 'refresh',
+    INCLUDE_ALL_TOOLS: 'includeAllTools',
+    BYPASS_METADATA_OVERRIDES: 'bypassMetadataOverrides',
+};
+export const CONTENT_TYPES = {
+    APPLICATION_JSON: 'application/json',
+    MULTIPART_FORM_DATA: 'multipart/form-data',
+    EVENT_STREAM: 'text/event-stream',
+    AUDIO_MPEG: 'audio/mpeg',
+    APPLICATION_OCTET_STREAM: 'application/octet-stream',
+    BINARY_OCTET_STREAM: 'binary/octet-stream',
+    GENERIC_AUDIO_PATTERN: 'audio',
+    PLAIN_TEXT: 'text/plain',
+    HTML: 'text/html',
+    GENERIC_IMAGE_PATTERN: 'image/',
+};
+export const MULTIPART_FORM_DATA_ENDPOINTS = ['createTranscription', 'createTranslation', 'uploadFile'];
+export const fileExtensionMimeTypeMap = {
+    mp4: 'video/mp4',
+    jpeg: 'image/jpeg',
+    jpg: 'image/jpeg',
+    png: 'image/png',
+    bmp: 'image/bmp',
+    tiff: 'image/tiff',
+    webp: 'image/webp',
+    gif: 'image/gif',
+    pdf: 'application/pdf',
+    csv: 'text/csv',
+    doc: 'application/msword',
+    docx: 'application/vnd.openxmlformats-officedocument.wordprocessingml.document',
+    xls: 'application/vnd.ms-excel',
+    xlsx: 'application/vnd.openxmlformats-officedocument.spreadsheetml.sheet',
+    html: 'text/html',
+    md: 'text/markdown',
+    mp3: 'audio/mp3',
+    wav: 'audio/wav',
+    txt: 'text/plain',
+    mov: 'video/mov',
+    mpeg: 'video/mpeg',
+    mpg: 'video/mpg',
+    avi: 'video/avi',
+    wmv: 'video/wmv',
+    mpegps: 'video/mpegps',
+    flv: 'video/flv',
+    webm: 'video/webm',
+    mkv: 'video/mkv',
+    threegpp: 'video/three_gpp',
+};
+/**
+ * Reverse mapping from MIME type to file extension.
+ * Example: "text/plain" → "txt", "application/pdf" → "pdf"
+ *
+ * When multiple extensions map to the same MIME type (e.g. both `jpeg` and `jpg`
+ * → `image/jpeg`), the FIRST extension declared in `fileExtensionMimeTypeMap`
+ * wins. This guarantees the canonical IANA-style extension is returned, which
+ * is what downstream providers (e.g. AWS Bedrock Converse, which only accepts
+ * `[gif, jpeg, png, webp]` as image formats) expect.
+ */
+export const mimeTypeToExtensionMap = (() => {
+    const map = {};
+    for (const [extension, mime] of Object.entries(fileExtensionMimeTypeMap)) {
+        if (!(mime in map)) {
+            map[mime] = extension;
+        }
+    }
+    return map;
+})();
+export const imagesMimeTypes = [
+    fileExtensionMimeTypeMap.jpeg,
+    fileExtensionMimeTypeMap.jpg,
+    fileExtensionMimeTypeMap.png,
+    fileExtensionMimeTypeMap.bmp,
+    fileExtensionMimeTypeMap.tiff,
+    fileExtensionMimeTypeMap.webp,
+];
+export const documentMimeTypes = [
+    fileExtensionMimeTypeMap.pdf,
+    fileExtensionMimeTypeMap.csv,
+    fileExtensionMimeTypeMap.doc,
+    fileExtensionMimeTypeMap.docx,
+    fileExtensionMimeTypeMap.xls,
+    fileExtensionMimeTypeMap.xlsx,
+    fileExtensionMimeTypeMap.html,
+    fileExtensionMimeTypeMap.md,
+    fileExtensionMimeTypeMap.txt,
+];
+export const videoMimeTypes = [
+    fileExtensionMimeTypeMap.mkv,
+    fileExtensionMimeTypeMap.mov,
+    fileExtensionMimeTypeMap.mp4,
+    fileExtensionMimeTypeMap.webm,
+    fileExtensionMimeTypeMap.flv,
+    fileExtensionMimeTypeMap.mpeg,
+    fileExtensionMimeTypeMap.mpg,
+    fileExtensionMimeTypeMap.wmv,
+    fileExtensionMimeTypeMap.threegpp,
+    fileExtensionMimeTypeMap.avi,
+];
+export var BatchEndpoints;
+(function (BatchEndpoints) {
+    BatchEndpoints["CHAT_COMPLETIONS"] = "/v1/chat/completions";
+    BatchEndpoints["COMPLETIONS"] = "/v1/completions";
+    BatchEndpoints["EMBEDDINGS"] = "/v1/embeddings";
+})(BatchEndpoints || (BatchEndpoints = {}));
+/**
+ * ================================
+ * Context Keys
+ * ================================
+ */
+// Timing keys store performance.now() values (monotonic ms since performance.timeOrigin), not epoch timestamps.
+// To convert stored monotonic value t to epoch: performance.timeOrigin + t
+// To convert current monotonic time to epoch: performance.timeOrigin + performance.now()
+// Do not mix Date.now() with performance.now() when computing deltas.
+export const CONTEXT_KEYS = {
+    TRANSFORMED_BODY: 'transformedBody',
+    SKIP_GUARDRAIL_AUTHORIZATION: 'skipGuardrailAuthorization',
+    SKIP_MODEL_AUTHORIZATION: 'skipModelAuthorization',
+    SKIP_METADATA_LOGGING: 'skipMetadataLogging',
+    SKIP_GLOBAL_STREAM_CONVERSION: 'skipGlobalStreamConversion',
+    MONOTONIC_MODEL_LATENCY_IN_MS: 'monotonicModelLatencyInMs',
+    IS_STREAMING_REQUEST: 'isStreamingRequest',
+    // Timing keys
+    MONOTONIC_MODEL_REQUEST_START_TIME: 'monotonicModelRequestStartTime',
+    MONOTONIC_REQUEST_START_TIME: 'monotonicRequestStartTime',
+    MONOTONIC_REQUEST_FIRST_TOKEN_TIME: 'monotonicRequestFirstTokenTime',
+    /** TTFT timeout: AbortController set by routing layer to abort provider fetch when budget expires */
+    TTFT_ABORT_CONTROLLER: 'ttftAbortController',
+    /** TTFT timeout: timeout in ms, set when TTFT is enabled so span can build message */
+    TTFT_TIMEOUT_MS: 'ttftTimeoutMs',
+    /** TTFT timeout: timer set when TTFT is enabled so span can build message */
+    TTFT_TIMER: 'ttftTimer',
+    // Tracing keys
+    TRACING_RESPONSE_ERROR_MESSAGE: 'tracingResponseErrorMessage',
+    TRACING_RESPONSE_STATUS_CODE: 'tracingResponseStatusCode',
+    TRACING_STREAMING_MODEL_PROVIDER_TERMINATED: 'tracingStreamingModelProviderTerminated', // Flag set when model provider terminates the stream
+    TRACING_PROVIDER_STREAM_STATUS_CODE: 'tracingProviderStreamStatusCode', // Set when the provider reports a failure inside a 200 stream; the status that failure maps to
+    TRACING_MODEL_PROVIDER_STATUS_CODE: 'tracingModelProviderStatusCode', // Only set for actual provider HTTP responses
+    TRACING_MODEL_SPAN_EVER_CREATED: 'tracingModelSpanEverCreated', // Set when any model span is created so we don't create a duplicate failure span
+    TRACING_PROVIDER_REQUEST_FORMAT: 'tracingProviderRequestFormat', // Upstream request format, when it differs from the provider default
+    TRACING_ANTHROPIC_BETAS_SENT: 'tracingAnthropicBetasSent', // Anthropic betas sent upstream, post-allowlist
+    TRACING_ANTHROPIC_BETAS_REQUESTED: 'tracingAnthropicBetasRequested', // Anthropic betas the caller asked for
+    TRACING_ROOT_SPAN_MODEL: 'tracingRootSpanModel', // Model from request body, set when root span is created (for failure model span)
+    TRACING_MODEL_REQUEST_URL: 'tracingModelRequestUrl',
+    DATA_ROUTING_DESTINATION: 'dataRoutingDestination',
+    TRACING_PROVIDER_INPUT: 'tracingProviderInput',
+    // Guardrails execution tracking
+    GUARDRAILS_INPUT_EXECUTED: 'guardrailsInputExecuted',
+    GUARDRAILS_RESOLVED_MODEL: 'guardrailsResolvedModel',
+    RESOLVED_GUARDRAIL_CHECKS: 'guardrailsConfigCache',
+    RESOLVED_GUARDRAIL_CUSTOM_ERROR_MESSAGE_MAP: 'resolvedGuardrailCustomErrorMessageMap',
+    INPUT_GUARDRAIL_PROMISE: 'inputGuardrailPromise',
+    HOOK_SPAN_ID: 'hookSpanId',
+    TFY_PROVIDER_NAME: 'tfyProviderName',
+    // Authentication context
+    AUTH_TYPE: 'authType',
+    // Router context
+    IS_GEMINI_ROUTER: 'isGeminiRouter',
+    IS_VERTEX_AI_ROUTER: 'isVertexAIRouter',
+    IS_BEDROCK_ROUTER: 'isBedrockRouter',
+    NEED_ORIGINAL_BINARY_CHUNKS: 'needOriginalBinaryChunks',
+    /** TTS router: /tts/{providerName}/* - provider name from path, model from body */
+    IS_TTS_ROUTER: 'isTtsRouter',
+    /** STT router: /stt/{providerName}/* - provider name from path, model from body (multipart/form-data) */
+    IS_STT_ROUTER: 'isSttRouter',
+    /** Video blind-proxy router: /videos-proxy/{providerName}/* - provider name from path, model from native URL/header */
+    IS_VIDEO_PROXY_ROUTER: 'isVideoProxyRouter',
+    /** Custom proxy router: /proxy-api/{providerAccountName}/{endpointName}/* - transparent HTTP proxy */
+    IS_CUSTOM_PROXY_ROUTER: 'isCustomProxyRouter',
+    IS_CURSOR_REQUEST: 'isCursorRequest',
+    /** Model id extracted from Gemini/Vertex proxy path (e.g. /v1beta/models/{model}:action) for auth and cost */
+    TFY_PROXY_PATH_MODEL_ID: 'tfyProxyPathModelId',
+    /** When proxying via tryTargetsRecursively, the actual endpoint fn (tts-proxy | stt-proxy | proxy) for cost model type mapping */
+    PROXY_ENDPOINT_FN: 'proxyEndpointFn',
+    /** Parsed request body for proxy cost attribution (e.g. Vertex TTS JSON before streaming audio response) */
+    PROXY_COST_REQUEST_BODY: 'proxyCostRequestBody',
+    /** Unicode code-point count billed for Cloud TTS (Vertex tts-proxy), excluding SSML mark tags */
+    AUDIO_BILLED_CHARS: 'audioBilledChars',
+    /** Billable audio seconds from Cloud STT response (Vertex stt-proxy), after parsing totalBilledTime / totalBilledDuration */
+    AUDIO_BILLED_SECONDS: 'audioBilledSeconds',
+    /** Billable video seconds (durationSeconds * sampleCount), set at submit time since Veo's cost is fully known from the request. */
+    VIDEO_BILLED_SECONDS: 'videoBilledSeconds',
+    /** Resolved video resolution (e.g. "720p", "1080p", "4k") used for resolution-tiered video pricing. */
+    VIDEO_BILLED_RESOLUTION: 'videoBilledResolution',
+    /** Whether the requested video includes audio; selects the audio-on vs audio-off rate tier. */
+    VIDEO_BILLED_AUDIO: 'videoBilledAudio',
+    TFY_INTEGRATION_TO_GUARDRAIL_MANIFEST_MAP: 'tfyIntegrationToGuardrailManifestMap',
+    PROMPT_LOGGING_ENABLED: 'promptLoggingEnabled',
+    LOGGING_REDACTION_PATTERNS: 'loggingRedactionPatterns',
+    PROMPT_SUBSTITUTED: 'promptSubstituted',
+    // CrowdStrike AIDR FPE context map (Map<tfy_integration, fpeContext>)
+    // Keyed by the guardrail's tfy_integration selector so multiple CrowdStrike
+    // AIDR guardrails in the same request never overwrite each other - even when
+    // they share the same regional base_url.
+    CROWDSTRIKE_FPE_CONTEXT_MAP: 'crowdstrikeAidrFpeContextMap',
+    // A2A streaming wrapper
+    CUSTOM_STREAM_WRAPPER: 'customStreamWrapper',
+    // Rate Limit context
+    NOMINATED_RATE_LIMIT_RULES: 'nominatedRateLimitRules',
+    APPLIED_RATE_LIMIT_RULE_IDS: 'appliedRateLimitRuleIds',
+    AGGREGATED_APPLIED_RATE_LIMIT_RULE_IDS: 'aggregatedAppliedRateLimitRuleIds',
+    // Budget context
+    NOMINATED_BUDGET_RULES: 'nominatedBudgetRules',
+    APPLIED_BUDGET_RULE_IDS: 'appliedBudgetRuleIds',
+    AGGREGATED_APPLIED_BUDGET_RULE_IDS: 'aggregatedAppliedBudgetRuleIds',
+    BUDGET_AUDIT_MODE: 'budgetAuditMode',
+    // Which budget engine (v1 vs v2) produced the applied budget for this request.
+    APPLIED_BUDGET_SOURCE: 'appliedBudgetSource',
+    // Memoised v2 matches for this request (hot path + request logger). Cleared/replaced per target.
+    NOMINATED_BUDGETS_V2: 'nominatedBudgetsV2',
+    // Virtual-model–scoped budgets matched once before load-balancer routing. Survives target
+    // attempts (unlike NOMINATED_*) so each backend check can union them with concrete-target matches.
+    // Legacy v1 budgets do not use this path — they remain concrete-model only.
+    PREMATCHED_VM_BUDGETS: 'prematchedVmBudgets',
+    // Virtual model id (`account/model`) set by prematch so phase-2 target checks can dual-match
+    // when.models without re-deriving the VM id from IS_VIRTUAL_MODEL / slug flags.
+    PREMATCHED_VIRTUAL_MODEL_ID: 'prematchedVirtualModelId',
+    REQUESTED_MODEL: 'requestedModel',
+    MODEL_COST_IN_USD: 'modelCostInUsd',
+    // Cache context
+    CACHE_CONFIG: 'cacheConfig',
+    CACHE_EMBEDDINGS: 'cacheEmbeddings',
+    CACHE_STATUS: 'cacheStatus',
+    // Subscription passthrough (Claude Max, Codex / ChatGPT): override provider account API key with the request-supplied bearer token
+    OVERRIDE_PROVIDER_BEARER_TOKEN: 'overrideProviderBearerToken',
+    // Virtual model context
+    IS_VIRTUAL_MODEL: 'isVirtualModel',
+    VIRTUAL_MODEL_RESOLVED_FROM_SLUG: 'virtualModelResolvedFromSlug',
+    // Responses origin-pinning (GATEWAY-2691), only set for /responses on virtual models.
+    RESPONSES_STATEFUL: 'responsesStateful', // request has server-side state → returned id is encoded
+    RESPONSES_PINNED_TARGET: 'responsesPinnedTarget', // follow-up must go to this target; routing pins here, no fallback
+    RESPONSES_SERVED_TARGET: 'responsesServedTarget', // target that served this turn; read when encoding the reply id
+    RESPONSES_PROVIDER_RESPONSE_ID: 'responsesProviderResponseId', // provider's real id, for the GET/DELETE URL
+    // Resolved metadata context (header + config + default merged)
+    RESOLVED_METADATA: 'resolvedMetadata',
+    TRUNCATED_METADATA: 'truncatedMetadata',
+    REQUEST_SIGNAL: 'requestSignal',
+    // Streaming lifecycle: set to true in processChunk / handleProxyPassthroughWithTTFT
+    // when the [DONE] SSE marker is written to the client. Set synchronously before the
+    // generator resumes reader.read(), so the abort handler sees it immediately.
+    STREAMING_COMPLETED: 'streamingCompleted',
+    /** Resolved ClientCancelBilling policy for this request (see utils/clientCancelPolicy). */
+    CLIENT_CANCEL_BILLING_TYPE: 'clientCancelBillingType',
+    /** Case B: detach the client but keep draining upstream so the provider's own usage block is captured. */
+    COMPLETE_REQUEST_ON_CLIENT_CANCEL: 'completeRequestOnClientCancel',
+    /** Case C: the pump will estimate tokens from the partial stream, so spans must stay open for it. */
+    KEEP_SPANS_OPEN_ON_CLIENT_CANCEL: 'keepSpansOpenOnClientCancel',
+    // # Memory Optimization Block
+    PARSED_REQUEST_BODY: 'parsedRequestBody',
+    // Adapter context
+    ADAPTER: 'adapter',
+    ADAPTER_ORIGINAL_ENDPOINT: 'adapterOriginalEndpoint',
+    ADAPTER_ORIGINAL_REQUEST: 'adapterOriginalRequest',
+    /** IMAGE-modality billable tokens from image gen/edit responses (set before model span cost) */
+    IMAGE_BILLED_INPUT_TOKENS: 'imageBilledInputTokens',
+    IMAGE_BILLED_OUTPUT_TOKENS: 'imageBilledOutputTokens',
+    /** Set to true by the response loggers when token counts are produced via the estimation fallback (provider returned no usage). */
+    TOKENS_ESTIMATED: 'tokensEstimated',
+};
+/**
+ * ================================
+ * OpenAI model patterns for native response_format support
+ * ================================
+ */
+export const OPENAI_MODELS_WITHOUT_NATIVE_RESPONSE_FORMAT = ['gpt-5-chat', 'gpt-4o-2024-05-13'];
+export const OPENAI_MODELS_WITH_NATIVE_RESPONSE_FORMAT = ['gpt-4o', 'gpt-5', 'gpt-4.1', 'o3', 'o4'];
+const NATIVE_STRUCTURED_OUTPUT_MIN_VERSION = { major: 4, minor: 5 };
+/**
+ * Returns true when the Claude model natively supports structured output via
+ * output_config.format (e.g. claude-sonnet-4-5, claude-opus-4-6); others fall
+ * back to tool conversion.
+ *
+ * Versions >= 4.5 support native structured output. If the version can't be
+ * parsed, defaults to false (tool conversion) as a safe fallback.
+ */
+export function isClaudeNativeStructuredOutputModel(model) {
+    if (!model.toLowerCase().includes('claude'))
+        return false;
+    const normalized = normalizeClaudeModelId(model);
+    const version = extractClaudeVersion(normalized);
+    if (!version)
+        return false;
+    return (version.major > NATIVE_STRUCTURED_OUTPUT_MIN_VERSION.major ||
+        (version.major === NATIVE_STRUCTURED_OUTPUT_MIN_VERSION.major &&
+            version.minor >= NATIVE_STRUCTURED_OUTPUT_MIN_VERSION.minor));
+}
+/**
+ * ================================
+ * Model-specific parameter exclusions
+ * ================================
+ *
+ *
+ * Sampling params the GPT-5 family rejects. Sending any of them to a model that does not
+ * accept them is a 400 from OpenAI, so they are dropped rather than forwarded.
+ */
+export const GPT5_UNSUPPORTED_SAMPLING_PARAMS = ['temperature', 'top_p', 'presence_penalty', 'stop'];
+/**
+ * True for the original GPT-5 reasoning models — `gpt-5`, `gpt-5-mini`, `gpt-5-nano`,
+ * `gpt-5-pro`, `gpt-5-codex` and their dated snapshots — which reject the params above
+ * unconditionally. https://developers.openai.com/api/docs/guides/latest-model
+ *
+ * Scoped to that family on purpose. From gpt-5.1 onwards support depends on `reasoning_effort`,
+ * and per-model quirks (`-pro` cannot disable reasoning, some `-chat` aliases still reason)
+ * can't be read off the model name reliably — an earlier attempt to encode them here got two
+ * families wrong. So those are forwarded and the provider validates them, which fails loudly
+ * instead of silently discarding what the caller asked for. The durable fix is the model
+ * catalogue; see the TODO above.
+ *
+ * The `includes('gpt-5')` gate is deliberately loose: Azure deployment names are arbitrary and
+ * callers rely on a deployment named e.g. `prod-gpt-5-east` matching when no foundation model
+ * is set.
+ */
+export function isBaseGpt5ReasoningModel(model) {
+    if (!model.includes('gpt-5'))
+        return false;
+    // gpt-5.1, gpt-5.2, gpt-5.6, ... — conditional on reasoning_effort, so not our call.
+    if (/gpt-5\.\d/.test(model))
+        return false;
+    // gpt-5-chat / gpt-5-chat-latest are non-reasoning and take sampling params normally.
+    if (model.includes('gpt-5-chat'))
+        return false;
+    return true;
+}
+/**
+ * ================================
+ * TLS Certificate Error Codes and Messages
+ * Reference: https://nodejs.org/api/tls.html
+ * These are official Node.js TLS error codes from OpenSSL
+ * Keys are the error codes, values are user-friendly error messages
+ * ================================
+ */
+export const TLS_ERROR_CONFIG = {
+    // Certificate validation errors
+    UNABLE_TO_VERIFY_LEAF_SIGNATURE: 'Request rejected due to certificate verification failure',
+    UNABLE_TO_GET_ISSUER_CERT: 'Request rejected: certificate issuer not found',
+    UNABLE_TO_GET_ISSUER_CERT_LOCALLY: 'Request rejected: local certificate issuer not found',
+    DEPTH_ZERO_SELF_SIGNED_CERT: 'Request rejected: self-signed certificate not trusted',
+    SELF_SIGNED_CERT_IN_CHAIN: 'Request rejected: self-signed certificate in chain not trusted',
+    // Certificate expiration errors
+    CERT_HAS_EXPIRED: 'Request rejected: certificate has expired',
+    CERT_NOT_YET_VALID: 'Request rejected: certificate not yet valid',
+    CRL_HAS_EXPIRED: 'Request rejected: certificate revocation list has expired',
+    CRL_NOT_YET_VALID: 'Request rejected: certificate revocation list not yet valid',
+    // Certificate signature errors
+    CERT_SIGNATURE_FAILURE: 'Request rejected: certificate signature invalid',
+    UNABLE_TO_DECRYPT_CERT_SIGNATURE: 'Request rejected: unable to decrypt certificate signature',
+    UNABLE_TO_DECODE_ISSUER_PUBLIC_KEY: 'Request rejected: unable to decode issuer public key',
+    // Certificate revocation errors
+    CERT_REVOKED: 'Request rejected: certificate has been revoked',
+    UNABLE_TO_GET_CRL: 'Request rejected: certificate revocation check failed',
+    UNABLE_TO_DECRYPT_CRL_SIGNATURE: 'Request rejected: unable to decrypt CRL signature',
+    CRL_SIGNATURE_FAILURE: 'Request rejected: CRL signature invalid',
+    // Certificate chain errors
+    CERT_CHAIN_TOO_LONG: 'Request rejected: certificate chain too long',
+    PATH_LENGTH_EXCEEDED: 'Request rejected: certificate path length exceeded',
+    // Certificate trust errors
+    CERT_UNTRUSTED: 'Request rejected: certificate not trusted',
+    CERT_REJECTED: 'Request rejected: certificate rejected',
+    INVALID_CA: 'Request rejected: invalid CA certificate',
+    INVALID_PURPOSE: 'Request rejected: unsupported certificate purpose',
+    // Hostname errors
+    HOSTNAME_MISMATCH: 'Request rejected: certificate hostname mismatch',
+    ERR_TLS_CERT_ALTNAME_INVALID: 'Request rejected: certificate hostname mismatch',
+    // Format errors
+    ERROR_IN_CERT_NOT_BEFORE_FIELD: 'Request rejected: certificate format error in notBefore field',
+    ERROR_IN_CERT_NOT_AFTER_FIELD: 'Request rejected: certificate format error in notAfter field',
+    ERROR_IN_CRL_LAST_UPDATE_FIELD: 'Request rejected: CRL format error in lastUpdate field',
+    ERROR_IN_CRL_NEXT_UPDATE_FIELD: 'Request rejected: CRL format error in nextUpdate field',
+};
+//# sourceMappingURL=globals.js.map