@juspay/neurolink 11.13.0 → 11.13.2

package/dist/lib/providers/anthropic/client.js

@@ -1,1862 +0,0 @@
-import Anthropic from "@anthropic-ai/sdk";
-import { SpanKind, trace } from "@opentelemetry/api";
-import { existsSync, mkdirSync, readFileSync, renameSync, writeFileSync, } from "fs";
-import { homedir } from "os";
-import { join } from "path";
-import { ANTHROPIC_TOKEN_URL, CLAUDE_CLI_USER_AGENT, CLAUDE_CODE_CLIENT_ID, CLAUDE_CODE_OAUTH_BETAS, } from "../../auth/anthropicOAuth.js";
-import { AnthropicModels, TOKEN_EXPIRY_BUFFER_MS, } from "../../constants/enums.js";
-import { BaseProvider } from "../../core/baseProvider.js";
-import { DEFAULT_MAX_STEPS } from "../../core/constants.js";
-import { streamAnalyticsCollector } from "../../core/streamAnalytics.js";
-import { getModelCapabilities, getRecommendedModelForTier, isModelAvailableForTier, } from "../../models/anthropicModels.js";
-import { createOAuthFetch } from "../../proxy/oauthFetch.js";
-import { createProxyFetch } from "../../proxy/proxyFetch.js";
-import { getCapturedLimitSnapshot, getCapturedResponseHeaders, logClaudeLimitSnapshot, runInLimitCaptureScope, setLimitSpanAttributes, withLimitCapture, wrapFetchWithLimitCapture, } from "./rateLimitCapture.js";
-import { AuthenticationError, NetworkError, ProviderError, RateLimitError, } from "../../types/index.js";
-import { classifyProviderError } from "../../utils/errorClassifier.js";
-import { logger } from "../../utils/logger.js";
-import { ANTHROPIC_ELISION_NOTE, planAnthropicLoopReclaim, previewAnthropicToolResultText, } from "../../context/anthropicLoopGuard.js";
-import { getAvailableInputTokens } from "../../constants/contextWindows.js";
-import { estimateTokens } from "../../utils/tokenEstimation.js";
-import { redactUrlCredentials } from "../../utils/logSanitize.js";
-import { ANTHROPIC_MAX_CACHE_BREAKPOINTS, applyAnthropicHistoryCacheBreakpoints, countAnthropicCacheMarkers, } from "../../utils/anthropicCacheBreakpoints.js";
-import { calculateCost } from "../../utils/pricing.js";
-import { stringifyAnthropicToolOutput } from "./toolOutput.js";
-import { createAnthropicLoopAdapter } from "./loopAdapter.js";
-import { runAgenticLoop } from "../../core/loopEngine.js";
-import { createAnthropicConfig, getProviderModel, validateApiKey, } from "../../utils/providerConfig.js";
-import { composeAbortSignals, createTimeoutController, mergeAbortSignals, } from "../../utils/timeout.js";
-import { resolveToolChoice } from "../../utils/toolChoice.js";
-import { emitToolEndFromStepFinish } from "../../utils/toolEndEmitter.js";
-import { NoOutputGeneratedError } from "../../utils/generationErrors.js";
-import { buildNoOutputSentinel, stampNoOutputSpan, } from "../../utils/noOutputSentinel.js";
-import { convertZodToJsonSchema } from "../../utils/schemaConversion.js";
-import { resolveClaudeMaxTokens } from "../../utils/tokenLimits.js";
-import { toAnthropicImageBlock, fileToAnthropicBlock, } from "../anthropicImageBlocks.js";
-import { resolveSamplingParams } from "../../models/modelRegistry.js";
-import { createDeferredAnalytics, stringifyToolInput, } from "../openaiChatCompletionsClient.js";
-import { createStreamChannel } from "../../core/streamChannel.js";
-import { toNativeToolDeclarations } from "../../core/nativeToolFormat.js";
-import { ANTHROPIC_BETA_HEADERS } from "./constants.js";
-import { cacheControlOf } from "./cacheControl.js";
-import { appendFinalResultInstruction, appendFinalResultTool, FINAL_RESULT_TOOL_NAME, } from "./structuredOutput.js";
-// AnthropicProviderConfig is imported from types/providers.ts
-// Re-export for backward compatibility
-// Configuration helpers - now using consolidated utility
-const getAnthropicApiKey = () => {
-    return validateApiKey(createAnthropicConfig());
-};
-const getDefaultAnthropicModel = () => {
-    return getProviderModel("ANTHROPIC_MODEL", AnthropicModels.CLAUDE_SONNET_4_6);
-};
-const streamTracer = trace.getTracer("neurolink.provider.anthropic");
-/**
- * Get OAuth token from stored credentials file or environment.
- * Priority:
- * 1. Stored credentials file (~/.neurolink/anthropic-credentials.json)
- * 2. Environment variables (ANTHROPIC_OAUTH_TOKEN or CLAUDE_OAUTH_TOKEN)
- */
-const getOAuthToken = () => {
-    // First, check stored credentials file (highest priority)
-    try {
-        const credentialsPath = join(homedir(), ".neurolink", "anthropic-credentials.json");
-        if (existsSync(credentialsPath)) {
-            const credentialsContent = readFileSync(credentialsPath, "utf-8");
-            const credentials = JSON.parse(credentialsContent);
-            if (credentials.type === "oauth" && credentials.oauth?.accessToken) {
-                logger.debug("[AnthropicProvider] Using OAuth token from stored credentials file");
-                return credentials.oauth;
-            }
-        }
-    }
-    catch (error) {
-        logger.debug("[AnthropicProvider] Failed to read stored credentials:", error);
-    }
-    // Fallback to environment variables
-    const tokenString = process.env.ANTHROPIC_OAUTH_TOKEN || process.env.CLAUDE_OAUTH_TOKEN;
-    if (!tokenString) {
-        return null;
-    }
-    // Try to parse as JSON (for full token object with refresh token and expiry)
-    try {
-        const parsed = JSON.parse(tokenString);
-        if (typeof parsed === "object" && parsed.accessToken) {
-            return parsed;
-        }
-        // If it's a simple string in JSON, use it as access token
-        if (typeof parsed === "string") {
-            return { accessToken: parsed };
-        }
-    }
-    catch {
-        // Not JSON, treat as plain access token string
-    }
-    // Treat as plain access token string
-    return { accessToken: tokenString };
-};
-/**
- * Detect subscription tier from environment or token.
- * Environment variable ANTHROPIC_SUBSCRIPTION_TIER takes precedence.
- */
-const detectSubscriptionTier = (oauthToken) => {
-    // Check explicit environment variable first
-    const envTier = process.env.ANTHROPIC_SUBSCRIPTION_TIER?.toLowerCase();
-    if (envTier) {
-        const validTiers = [
-            "free",
-            "pro",
-            "max",
-            "max_5",
-            "max_20",
-            "api",
-        ];
-        if (validTiers.includes(envTier)) {
-            logger.debug("[detectSubscriptionTier] Using environment override", {
-                tier: envTier,
-            });
-            return envTier;
-        }
-        logger.warn("[detectSubscriptionTier] Invalid ANTHROPIC_SUBSCRIPTION_TIER", {
-            value: envTier,
-            validTiers,
-        });
-    }
-    // If using OAuth, default to 'pro' (most common subscription tier)
-    if (oauthToken) {
-        // Check if token scopes indicate tier (future-proofing)
-        const scopes = oauthToken.scopes ?? [];
-        let detectedTier = "pro";
-        if (scopes.includes("max_20")) {
-            detectedTier = "max_20";
-        }
-        else if (scopes.includes("max_5")) {
-            detectedTier = "max_5";
-        }
-        else if (scopes.includes("max")) {
-            detectedTier = "max";
-        }
-        logger.debug("[detectSubscriptionTier] Detected from OAuth token", {
-            tier: detectedTier,
-            scopes,
-        });
-        return detectedTier;
-    }
-    // Default to 'api' for API key authentication
-    logger.debug("[detectSubscriptionTier] No OAuth token, defaulting to API tier");
-    return "api";
-};
-/**
- * Determine authentication method based on available credentials.
- * OAuth takes precedence over API key if both are available.
- */
-const detectAuthMethod = (oauthToken) => {
-    // Explicit env var takes highest precedence — allows forcing api_key mode
-    // even when OAuth credentials exist (e.g., when using a proxy that handles auth)
-    const explicit = process.env.ANTHROPIC_AUTH_METHOD?.toLowerCase();
-    if (explicit === "api_key" || explicit === "apikey") {
-        logger.debug("[detectAuthMethod] Forced to api_key by ANTHROPIC_AUTH_METHOD env var");
-        return "api_key";
-    }
-    if (explicit === "oauth") {
-        if (oauthToken) {
-            logger.debug("[detectAuthMethod] Forced to oauth by ANTHROPIC_AUTH_METHOD env var");
-            return "oauth";
-        }
-        logger.warn("[detectAuthMethod] ANTHROPIC_AUTH_METHOD=oauth but no OAuth token found; falling through to auto-detection");
-    }
-    else if (explicit) {
-        logger.warn("[detectAuthMethod] Unrecognized ANTHROPIC_AUTH_METHOD value; falling through to auto-detection", {
-            value: explicit,
-        });
-    }
-    // Auto-detect: OAuth takes precedence if available
-    const method = oauthToken ? "oauth" : "api_key";
-    logger.debug("[detectAuthMethod] Auth method resolved", {
-        method,
-        hasOAuthToken: !!oauthToken,
-    });
-    return method;
-};
-// Rate-limit header parsing lives in `rateLimitCapture.ts`, which sees the raw
-// fetch Response and understands both header families (unified subscription
-// windows and the legacy per-tier counters). The module-private copy that used
-// to sit here parsed only the legacy family and had no callers.
-// ───────────────────────────────────────────────────────────────────────────
-// Native Messages-API conversion helpers (NeuroLink/V3 shapes → Anthropic)
-// ───────────────────────────────────────────────────────────────────────────
-/**
- * Convert NeuroLink/V3-shaped messages (the shape produced by
- * buildMessagesForStream and by the AI-SDK prompt on the V3 doGenerate path)
- * into the Anthropic Messages payload: a top-level `system` string plus
- * alternating user/assistant messages with typed content blocks.
- */
-const messagesToAnthropic = (msgs) => {
-    const systemBlocks = [];
-    const messages = [];
-    const partsOf = (content) => Array.isArray(content) ? content : [content];
-    // Message-level cache breakpoints apply to the message's LAST content
-    // block (the AI-SDK convention MessageBuilder relies on).
-    const applyMessageCacheControl = (blocks, msg) => {
-        const cc = cacheControlOf(msg);
-        if (cc && blocks.length > 0) {
-            const last = blocks[blocks.length - 1];
-            last.cache_control = cc;
-        }
-    };
-    for (const msg of msgs) {
-        switch (msg.role) {
-            case "system": {
-                const text = typeof msg.content === "string"
-                    ? msg.content
-                    : partsOf(msg.content)
-                        .map((p) => typeof p === "string"
-                        ? p
-                        : String(p?.text ?? ""))
-                        .join("\n");
-                const cc = cacheControlOf(msg);
-                systemBlocks.push({
-                    type: "text",
-                    text,
-                    ...(cc ? { cache_control: cc } : {}),
-                });
-                break;
-            }
-            case "user": {
-                const blocks = [];
-                for (const part of partsOf(msg.content)) {
-                    if (typeof part === "string") {
-                        if (part.length > 0) {
-                            blocks.push({ type: "text", text: part });
-                        }
-                        continue;
-                    }
-                    const p = part;
-                    if (p?.type === "text" && typeof p.text === "string") {
-                        const cc = cacheControlOf(p);
-                        blocks.push({
-                            type: "text",
-                            text: p.text,
-                            ...(cc ? { cache_control: cc } : {}),
-                        });
-                    }
-                    else if (p?.type === "image" || p?.type === "image_url") {
-                        const img = toAnthropicImageBlock(p.image ?? p.data ?? p.url);
-                        if (img) {
-                            const cc = cacheControlOf(p);
-                            blocks.push(cc ? { ...img, cache_control: cc } : img);
-                        }
-                    }
-                    else if (p?.type === "file") {
-                        // AI-SDK v6 encodes images AND PDFs as `type:"file"` parts in the
-                        // LanguageModel prompt that `doGenerate` receives. Without this
-                        // branch the image is dropped on the tool-using generate path and
-                        // the model never sees it ("no image detected").
-                        //
-                        // Runtime guard: p comes from message parsing and may not match the
-                        // expected shape. Verify p is an object and that mediaType, if
-                        // present, is a string (not an object/array from a malformed part).
-                        // Skip gracefully rather than passing a bad shape to fileToAnthropicBlock.
-                        const isValidFilePart = typeof p === "object" &&
-                            p !== null &&
-                            ("mediaType" in p
-                                ? typeof p.mediaType === "string"
-                                : true);
-                        const block = isValidFilePart
-                            ? fileToAnthropicBlock(p)
-                            : undefined;
-                        if (block) {
-                            const cc = cacheControlOf(p);
-                            if (cc) {
-                                // block is a fresh object from fileToAnthropicBlock; mutate in
-                                // place to keep the discriminated-union type (a spread widens it
-                                // past ContentBlockParam).
-                                block.cache_control = cc;
-                            }
-                            blocks.push(block);
-                        }
-                    }
-                }
-                if (blocks.length > 0) {
-                    applyMessageCacheControl(blocks, msg);
-                    messages.push({ role: "user", content: blocks });
-                }
-                break;
-            }
-            case "assistant": {
-                const blocks = [];
-                for (const part of partsOf(msg.content)) {
-                    if (typeof part === "string") {
-                        if (part.length > 0) {
-                            blocks.push({ type: "text", text: part });
-                        }
-                        continue;
-                    }
-                    const p = part;
-                    if (p?.type === "text" && typeof p.text === "string") {
-                        if (p.text.length > 0) {
-                            const cc = cacheControlOf(p);
-                            blocks.push({
-                                type: "text",
-                                text: p.text,
-                                ...(cc ? { cache_control: cc } : {}),
-                            });
-                        }
-                    }
-                    else if (p?.type === "tool-call") {
-                        let input = p.input;
-                        if (typeof input === "string") {
-                            try {
-                                input = JSON.parse(input);
-                            }
-                            catch {
-                                input = {};
-                            }
-                        }
-                        blocks.push({
-                            type: "tool_use",
-                            id: p.toolCallId ?? "",
-                            name: p.toolName ?? "",
-                            input: input ?? {},
-                        });
-                    }
-                }
-                if (blocks.length > 0) {
-                    applyMessageCacheControl(blocks, msg);
-                    messages.push({ role: "assistant", content: blocks });
-                }
-                break;
-            }
-            case "tool": {
-                // Tool results are user-role messages with tool_result blocks.
-                const blocks = [];
-                if (Array.isArray(msg.content)) {
-                    for (const part of msg.content) {
-                        const p = part;
-                        if (p?.type === "tool-result") {
-                            blocks.push({
-                                type: "tool_result",
-                                tool_use_id: p.toolCallId ?? "",
-                                content: stringifyAnthropicToolOutput(p.output),
-                            });
-                        }
-                    }
-                }
-                else if (typeof msg.content === "string") {
-                    blocks.push({
-                        type: "tool_result",
-                        tool_use_id: msg.toolCallId ?? "",
-                        content: msg.content,
-                    });
-                }
-                if (blocks.length > 0) {
-                    messages.push({ role: "user", content: blocks });
-                }
-                break;
-            }
-        }
-    }
-    // Plain-string system when no cache breakpoints are present (matches the
-    // previous wire shape); block form only when cache_control must ride along.
-    const system = systemBlocks.length === 0
-        ? undefined
-        : systemBlocks.some((b) => b.cache_control)
-            ? systemBlocks
-            : systemBlocks.map((b) => b.text).join("\n\n");
-    return {
-        ...(system !== undefined ? { system } : {}),
-        messages,
-    };
-};
-/** Map a NeuroLink tool choice onto Anthropic's tool_choice shape. */
-const toolChoiceToAnthropic = (choice) => {
-    if (!choice || choice === "auto") {
-        return undefined; // Anthropic defaults to auto when tools are present
-    }
-    if (choice === "none") {
-        return { type: "none" };
-    }
-    if (choice === "required") {
-        return { type: "any" };
-    }
-    if (typeof choice === "object") {
-        const c = choice;
-        if (c.type === "tool" && c.toolName) {
-            return { type: "tool", name: c.toolName };
-        }
-    }
-    return undefined;
-};
-/** Map Anthropic stop_reason onto the V3 unified finish reason. */
-const mapAnthropicStopReason = (raw) => {
-    switch (raw) {
-        case "max_tokens":
-            return "length";
-        case "tool_use":
-            return "tool-calls";
-        case "refusal":
-            return "content-filter";
-        default:
-            return "stop";
-    }
-};
-// Anthropic's Messages API requires max_tokens on every request. When the
-// caller omits it, default to the model's real output ceiling via
-// resolveClaudeMaxTokens (e.g. 64K for Sonnet 4.x) instead of the legacy 4096,
-// which silently truncated large structured responses mid-JSON.
-//
-// Client-level request timeout. The Anthropic SDK throws "Streaming is required
-// for long requests" from a NON-streaming `messages.create` when `max_tokens`
-// is large AND no client-level timeout is configured (it can't estimate a safe
-// timeout). Setting an explicit client timeout — equal to the SDK's own default
-// for the non-throwing path — suppresses that pre-flight throw so large
-// max_tokens (our model-ceiling default) works. Per-request duration is still
-// bounded by the abort signal NeuroLink composes for each call.
-const ANTHROPIC_CLIENT_TIMEOUT_MS = 600_000;
-/**
- * Anthropic Provider v2 - BaseProvider Implementation
- * Enhanced with OAuth support, subscription tiers, and beta headers for Claude Code integration.
- */
-export class AnthropicProvider extends BaseProvider {
-    client;
-    authMethod;
-    subscriptionTier;
-    enableBetaFeatures;
-    oauthToken;
-    lastResponseMetadata = null;
-    usageInfo = null;
-    refreshPromise;
-    /**
-     * Create a new Anthropic provider instance.
-     *
-     * @param modelName - Optional model name to use (defaults to CLAUDE_3_5_SONNET)
-     * @param sdk - Optional NeuroLink SDK instance
-     * @param config - Optional configuration options for auth, subscription tier, and beta features
-     */
-    constructor(modelName, sdk, config, credentials) {
-        // Pre-compute effective model with tier validation before calling super.
-        //
-        // When per-request credentials supply an apiKey (without oauthToken),
-        // force api_key auth — skip OAuth detection entirely so the caller's
-        // key is used rather than a stale OAuth token from ~/.neurolink/.
-        const forceApiKey = !!(credentials?.apiKey && !credentials?.oauthToken);
-        const oauthToken = forceApiKey
-            ? null
-            : ((credentials?.oauthToken
-                ? { accessToken: credentials.oauthToken }
-                : null) ??
-                config?.oauthToken ??
-                getOAuthToken());
-        // Resolve auth method FIRST so that tier detection uses the chosen method.
-        // If ANTHROPIC_AUTH_METHOD=api_key wins over an existing OAuth token, the
-        // tier must reflect api_key mode (full model access) rather than the OAuth
-        // token's subscription level.
-        const authMethod = forceApiKey
-            ? "api_key"
-            : (config?.authMethod ?? detectAuthMethod(oauthToken));
-        const subscriptionTier = config?.subscriptionTier ??
-            (authMethod === "oauth" ? detectSubscriptionTier(oauthToken) : "api");
-        const targetModel = modelName || getDefaultAnthropicModel();
-        // Determine effective model based on tier access.
-        // Skip tier validation when a proxy is in use (ANTHROPIC_BASE_URL is set)
-        // — the proxy handles model access and auth, so the SDK should pass
-        // the requested model through without downgrading.
-        let effectiveModel = targetModel;
-        const usingProxy = !!process.env.ANTHROPIC_BASE_URL;
-        if (!usingProxy &&
-            subscriptionTier !== "api" &&
-            !isModelAvailableForTier(targetModel, subscriptionTier)) {
-            effectiveModel = getRecommendedModelForTier(subscriptionTier);
-            logger.warn("Model not available for subscription tier, using recommended model", {
-                requestedModel: targetModel,
-                subscriptionTier,
-                recommendedModel: effectiveModel,
-            });
-        }
-        super(effectiveModel, "anthropic", sdk);
-        // Apply configuration with defaults
-        this.enableBetaFeatures = config?.enableBetaFeatures ?? true;
-        // Store computed values
-        this.oauthToken = oauthToken;
-        this.subscriptionTier = subscriptionTier;
-        // Use the auth method already resolved above (before tier computation)
-        this.authMethod = authMethod;
-        // Build headers based on auth method and subscription tier
-        const headers = this.getAuthHeaders();
-        // Create the official Anthropic SDK client based on auth method
-        let client;
-        logger.debug("[AnthropicProvider] Constructor - checking OAuth:", {
-            authMethod: this.authMethod,
-            hasOAuthToken: !!this.oauthToken,
-            hasAccessToken: !!this.oauthToken?.accessToken,
-        });
-        if (this.authMethod === "oauth" && this.oauthToken) {
-            // OAuth authentication - use custom fetch wrapper that handles:
-            // - Bearer token authorization
-            // - OAuth beta headers (oauth-2025-04-20, NOT claude-code-20250219)
-            // - User-Agent spoofing
-            // - ?beta=true query param
-            // - Tool name prefixing/stripping
-            logger.debug("[AnthropicProvider] Creating OAuth fetch wrapper...");
-            // Pass a getter so the fetch wrapper always uses the current token,
-            // even after an automatic token refresh.
-            // oauthToken is guaranteed non-null here (checked by the enclosing if-guard).
-            const tokenRef = this.oauthToken;
-            // skipBodyTransform=true: For the SDK client path, body transforms ARE
-            // intentionally skipped because the official Anthropic SDK builds its
-            // own request format (system prompts, metadata, tool definitions). The
-            // billing header, agent block, user_id injection, and mcp_ tool-name
-            // prefixing are only needed for proxy passthrough of raw Claude API
-            // requests where we must make the request look like it came from
-            // Claude Code / CLIProxyAPI.
-            const oauthFetch = createOAuthFetch(() => tokenRef.accessToken, this.enableBetaFeatures, false, // No mcp_ prefix — tool names pass through as-is (matches CLIProxyAPI)
-            true);
-            // For OAuth, we use a dummy API key since our fetch wrapper handles auth
-            // IMPORTANT: Do NOT pass beta headers here - our fetch wrapper handles them
-            // The claude-code-20250219 beta header triggers "credential only for Claude Code" error
-            client = new Anthropic({
-                apiKey: "oauth-authenticated", // Placeholder, actual auth is in fetch wrapper
-                // Note: No headers passed - fetch wrapper sets oauth-2025-04-20 beta header
-                // Limit capture wraps the OAuth fetch so subscription quota headers
-                // (anthropic-ratelimit-unified-*) are recorded on every request —
-                // streaming and non-streaming alike.
-                fetch: wrapFetchWithLimitCapture(oauthFetch),
-                timeout: ANTHROPIC_CLIENT_TIMEOUT_MS,
-                // The SDK's built-in retry honors Retry-After hints without any
-                // upper bound (a 429 with retry-after: 8549 sleeps 2.4h per retry,
-                // invisible to fallback orchestration). Retries are the
-                // orchestrator's job; transient network blips are still retried by
-                // the fetch wrapper.
-                maxRetries: 0,
-            });
-            logger.debug("[AnthropicProvider] Anthropic SDK client created with OAuth fetch wrapper");
-            logger.debug("Anthropic Provider initialized with OAuth", {
-                modelName: this.modelName,
-                provider: this.providerName,
-                authMethod: this.authMethod,
-                subscriptionTier: this.subscriptionTier,
-                enableBetaFeatures: this.enableBetaFeatures,
-                hasRefreshToken: !!this.oauthToken.refreshToken,
-                tokenExpiry: this.oauthToken.expiresAt
-                    ? new Date(this.oauthToken.expiresAt).toISOString()
-                    : "none",
-            });
-        }
-        else {
-            // Traditional API key authentication
-            const apiKeyToUse = credentials?.apiKey ?? config?.apiKey ?? getAnthropicApiKey();
-            // The official Anthropic SDK builds `${baseURL}/v1/messages` itself, so
-            // a version-suffixed base URL — the form the previous @ai-sdk/anthropic
-            // implementation REQUIRED (`https://api.anthropic.com/v1`) — would
-            // double up as `/v1/v1/messages`. Normalize the inverse way now: strip
-            // a trailing `/vN` segment when present so both historical forms of
-            // ANTHROPIC_BASE_URL keep working.
-            const normalizedBaseURL = (() => {
-                const raw = process.env.ANTHROPIC_BASE_URL;
-                if (!raw) {
-                    return undefined;
-                }
-                const trimmed = raw.replace(/\/+$/, "");
-                const stripped = trimmed.replace(/\/v\d+$/, "");
-                if (stripped !== trimmed) {
-                    logger.debug("[AnthropicProvider] Stripping the version suffix from " +
-                        "ANTHROPIC_BASE_URL — the official Anthropic SDK appends /v1 " +
-                        "to the base URL itself.", {
-                        baseURL: redactUrlCredentials(raw),
-                        rewrittenTo: redactUrlCredentials(stripped),
-                    });
-                }
-                return stripped;
-            })();
-            client = new Anthropic({
-                apiKey: apiKeyToUse,
-                defaultHeaders: headers,
-                ...(normalizedBaseURL && { baseURL: normalizedBaseURL }),
-                // Same capture as the OAuth branch: works for direct API-key traffic
-                // (legacy requests/tokens counters) and for the NeuroLink Claude proxy
-                // (verbatim unified quota plus x-neurolink-* account/pool state).
-                fetch: wrapFetchWithLimitCapture(createProxyFetch()),
-                timeout: ANTHROPIC_CLIENT_TIMEOUT_MS,
-                // See the OAuth-path client above: unbounded Retry-After sleeps in
-                // the SDK's retry loop must never stall fallback orchestration.
-                maxRetries: 0,
-            });
-            logger.debug("Anthropic Provider initialized with API key", {
-                modelName: this.modelName,
-                provider: this.providerName,
-                authMethod: this.authMethod,
-                subscriptionTier: this.subscriptionTier,
-                enableBetaFeatures: this.enableBetaFeatures,
-            });
-        }
-        this.client = client;
-        // Initialize usage tracking
-        this.usageInfo = {
-            messagesUsed: 0,
-            messagesRemaining: -1, // Unknown until we get rate limit headers
-            tokensUsed: 0,
-            tokensRemaining: -1,
-            inputTokensUsed: 0,
-            outputTokensUsed: 0,
-            lastRequestTimestamp: 0,
-            isRateLimited: false,
-            requestCount: 0,
-            messageQuotaPercent: 0,
-            tokenQuotaPercent: 0,
-        };
-        logger.debug("Anthropic Provider v2 initialized", {
-            modelName: this.modelName,
-            provider: this.providerName,
-            authMethod: this.authMethod,
-            subscriptionTier: this.subscriptionTier,
-            enableBetaFeatures: this.enableBetaFeatures,
-            betaFeatures: this.enableBetaFeatures
-                ? ANTHROPIC_BETA_HEADERS["anthropic-beta"]
-                : "disabled",
-        });
-    }
-    /**
-     * Get authentication headers based on current auth method and configuration.
-     *
-     * @returns Headers object containing auth and beta feature headers
-     */
-    getAuthHeaders() {
-        const headers = {};
-        // When routing through proxy (ANTHROPIC_BASE_URL set), use the full
-        // OAuth beta set so the proxy forwards them upstream. Without these,
-        // Anthropic treats the request with tighter non-subscription rate limits.
-        const usingProxy = !!process.env.ANTHROPIC_BASE_URL;
-        if (this.enableBetaFeatures) {
-            if (usingProxy) {
-                // The 1M-context beta requires a plan upgrade on most accounts;
-                // surfacing it by default forces a "The long context beta is not
-                // yet available for this subscription." failure for everyone else.
-                // Gate behind ANTHROPIC_ENABLE_LONG_CONTEXT_BETA=1 so default-tier
-                // accounts (and CI) can use the proxy without the gated feature.
-                const longContextOptIn = process.env.ANTHROPIC_ENABLE_LONG_CONTEXT_BETA === "1" ||
-                    process.env.ANTHROPIC_ENABLE_LONG_CONTEXT_BETA === "true";
-                const betas = [
-                    ...CLAUDE_CODE_OAUTH_BETAS,
-                    "fine-grained-tool-streaming-2025-05-14",
-                    "interleaved-thinking-2025-05-14",
-                    "redact-thinking-2026-02-12",
-                ];
-                if (longContextOptIn) {
-                    betas.push("context-1m-2025-08-07");
-                }
-                headers["anthropic-beta"] = betas.join(",");
-            }
-            else {
-                headers["anthropic-beta"] = ANTHROPIC_BETA_HEADERS["anthropic-beta"];
-            }
-        }
-        if (usingProxy) {
-            // WAFs in front of ANTHROPIC_BASE_URL proxies commonly block the bare
-            // SDK UA ("Anthropic/JS x.y.z"); send the claude-cli UA the OAuth path
-            // already uses. Direct-to-Anthropic traffic keeps the honest SDK UA.
-            headers["User-Agent"] = CLAUDE_CLI_USER_AGENT;
-        }
-        // Add subscription-specific headers if applicable
-        if (this.subscriptionTier !== "api") {
-            headers["x-subscription-tier"] = this.subscriptionTier;
-        }
-        return headers;
-    }
-    /**
-     * Validate if a model is accessible with the current subscription tier.
-     *
-     * @param model - The model ID to validate
-     * @returns true if the model is accessible, false otherwise
-     *
-     * @example
-     * ```typescript
-     * const provider = new AnthropicProvider();
-     * if (provider.validateModelAccess("claude-opus-4-5-20251101")) {
-     *   // Use the model
-     * } else {
-     *   // Fall back to a different model or show upgrade prompt
-     * }
-     * ```
-     */
-    validateModelAccess(model) {
-        // Proxy mode: bypass tier validation entirely — the proxy handles model
-        // access. Log at debug level so users can tell why an unknown model name
-        // "validated" when their proxy may not actually expose it.
-        if (process.env.ANTHROPIC_BASE_URL) {
-            logger.debug("[validateModelAccess] Bypassing tier check (ANTHROPIC_BASE_URL set — proxy enforces access)", { model });
-            return true;
-        }
-        // API tier has access to all models
-        if (this.subscriptionTier === "api") {
-            return true;
-        }
-        const hasAccess = isModelAvailableForTier(model, this.subscriptionTier);
-        if (!hasAccess) {
-            logger.debug("[validateModelAccess] Model not available for tier", {
-                model,
-                tier: this.subscriptionTier,
-            });
-        }
-        return hasAccess;
-    }
-    /**
-     * Get current usage information.
-     *
-     * Returns usage tracking data including messages sent, tokens consumed,
-     * and remaining quotas. This information is updated after each API request.
-     *
-     * @returns Current usage info or null if no requests have been made
-     *
-     * @example
-     * ```typescript
-     * const usage = provider.getUsageInfo();
-     * if (usage && usage.tokenQuotaPercent > 80) {
-     *   console.warn("Approaching token quota limit");
-     * }
-     * ```
-     */
-    getUsageInfo() {
-        return this.usageInfo;
-    }
-    /**
-     * Check if beta features are enabled for this provider instance.
-     *
-     * @returns true if beta features are enabled
-     */
-    areBetaFeaturesEnabled() {
-        return this.enableBetaFeatures;
-    }
-    /**
-     * Get model capabilities for the current model.
-     *
-     * @returns The model capabilities or undefined if not found
-     */
-    getModelCapabilities() {
-        return getModelCapabilities(this.modelName || this.getDefaultModel());
-    }
-    /**
-     * Get the current subscription tier.
-     * @returns The detected or configured subscription tier
-     */
-    getSubscriptionTier() {
-        return this.subscriptionTier;
-    }
-    /**
-     * Get the authentication method being used.
-     * @returns The current authentication method
-     */
-    getAuthMethod() {
-        return this.authMethod;
-    }
-    /**
-     * Refresh OAuth token if needed and possible.
-     * This method checks if the token is expired or about to expire,
-     * and attempts to refresh it using the refresh token if available.
-     *
-     * @returns Promise that resolves when refresh is complete (or not needed)
-     * @throws Error if refresh is needed but fails
-     */
-    async refreshAuthIfNeeded() {
-        // Only applicable for OAuth authentication
-        if (this.authMethod !== "oauth" || !this.oauthToken) {
-            logger.debug("Token refresh not applicable for API key authentication");
-            return;
-        }
-        // Check if token has expiry information
-        if (!this.oauthToken.expiresAt) {
-            logger.debug("Token has no expiry information, assuming valid");
-            return;
-        }
-        // expiresAt is stored as Unix milliseconds (matching how auth status/refresh stores it).
-        // Compare against Date.now() so both sides are in milliseconds.
-        const now = Date.now();
-        const isExpired = this.oauthToken.expiresAt <= now;
-        const isExpiringSoon = this.oauthToken.expiresAt <= now + TOKEN_EXPIRY_BUFFER_MS;
-        if (!isExpired && !isExpiringSoon) {
-            logger.debug("OAuth token is still valid", {
-                expiresInMs: this.oauthToken.expiresAt - now,
-            });
-            return;
-        }
-        // Check if we have a refresh token
-        if (!this.oauthToken.refreshToken) {
-            if (isExpired) {
-                throw new AuthenticationError("OAuth token expired and no refresh token available. Please re-authenticate.", this.providerName);
-            }
-            logger.warn("OAuth token expiring soon but no refresh token available", {
-                expiresInMs: this.oauthToken.expiresAt - now,
-            });
-            return;
-        }
-        // Serialize concurrent refresh attempts — if a refresh is already in flight,
-        // wait for it rather than issuing a duplicate request.
-        if (this.refreshPromise) {
-            await this.refreshPromise;
-            return;
-        }
-        // Attempt to refresh the token using the correct Anthropic token endpoint.
-        logger.info("Refreshing OAuth token", {
-            isExpired,
-            expiresInMs: this.oauthToken.expiresAt - now,
-        });
-        // Capture the token reference before entering the async IIFE;
-        // the enclosing guards already verified both fields are non-null.
-        const tokenRef = this.oauthToken;
-        const refreshToken = tokenRef.refreshToken;
-        this.refreshPromise = (async () => {
-            const REFRESH_TIMEOUT_MS = 30_000;
-            const controller = new AbortController();
-            const timeoutId = setTimeout(() => controller.abort(), REFRESH_TIMEOUT_MS);
-            // User-Agent is set to CLAUDE_CLI_USER_AGENT so the refresh request
-            // matches what the official Claude CLI / CLIProxyAPI sends. Anthropic
-            // gates parts of the OAuth flow on this UA (the same `client_id` is
-            // rejected by `ANTHROPIC_TOKEN_URL` if the UA looks like a generic
-            // SDK), so this is required for OAuth refresh to succeed — not a
-            // cosmetic choice. If Anthropic ever publishes a separate UA for
-            // third-party OAuth clients, switch to that. See `auth/anthropicOAuth.ts`
-            // for the source of `CLAUDE_CLI_USER_AGENT` / `CLAUDE_CODE_CLIENT_ID`.
-            const response = await fetch(ANTHROPIC_TOKEN_URL, {
-                method: "POST",
-                headers: {
-                    "Content-Type": "application/x-www-form-urlencoded",
-                    "User-Agent": CLAUDE_CLI_USER_AGENT,
-                },
-                body: new URLSearchParams({
-                    grant_type: "refresh_token",
-                    refresh_token: refreshToken,
-                    client_id: CLAUDE_CODE_CLIENT_ID,
-                }),
-                signal: controller.signal,
-            });
-            clearTimeout(timeoutId);
-            if (!response.ok) {
-                const errorText = await response.text();
-                throw new AuthenticationError(`Failed to refresh OAuth token: ${response.status} ${errorText}`, this.providerName);
-            }
-            const newToken = (await response.json());
-            // Mutate the existing oauthToken object in-place so that the fetch wrapper
-            // closure (which captured the object reference, not a copy) picks up the
-            // new accessToken automatically on the next request.
-            // Store expiresAt as milliseconds to match the format used by auth status/refresh.
-            tokenRef.accessToken = newToken.access_token;
-            tokenRef.refreshToken = newToken.refresh_token || tokenRef.refreshToken;
-            tokenRef.expiresAt = newToken.expires_in
-                ? Date.now() + newToken.expires_in * 1000
-                : undefined;
-            tokenRef.tokenType = newToken.token_type || "Bearer";
-            const updatedToken = tokenRef;
-            // Persist the refreshed token to disk atomically (tmp + rename) so
-            // subsequent provider instances and the CLI pick up the new credentials.
-            try {
-                const credentialsDir = join(homedir(), ".neurolink");
-                if (!existsSync(credentialsDir)) {
-                    mkdirSync(credentialsDir, { recursive: true });
-                }
-                const credentialsPath = join(credentialsDir, "anthropic-credentials.json");
-                const tmpPath = `${credentialsPath}.tmp`;
-                const existingRaw = existsSync(credentialsPath)
-                    ? JSON.parse(readFileSync(credentialsPath, "utf-8"))
-                    : {};
-                const updated = {
-                    ...existingRaw,
-                    type: "oauth",
-                    oauth: updatedToken,
-                    updatedAt: Date.now(),
-                };
-                writeFileSync(tmpPath, JSON.stringify(updated, null, 2), {
-                    mode: 0o600,
-                });
-                renameSync(tmpPath, credentialsPath);
-                logger.debug("Refreshed OAuth credentials persisted to disk");
-            }
-            catch (persistError) {
-                // Non-fatal: in-memory token is already updated; next CLI start will
-                // need a manual refresh but the current session will work.
-                logger.warn("Failed to persist refreshed OAuth token to disk", {
-                    error: persistError instanceof Error
-                        ? persistError.message
-                        : String(persistError),
-                });
-            }
-            logger.info("OAuth token refreshed successfully", {
-                hasNewRefreshToken: !!newToken.refresh_token,
-                expiresIn: newToken.expires_in,
-            });
-        })();
-        try {
-            await this.refreshPromise;
-        }
-        catch (error) {
-            if (error instanceof AuthenticationError) {
-                throw error;
-            }
-            throw new AuthenticationError(`Failed to refresh OAuth token: ${error instanceof Error ? error.message : String(error)}`, this.providerName);
-        }
-        finally {
-            this.refreshPromise = undefined;
-        }
-    }
-    /**
-     * Get the last response metadata including rate limit information.
-     * @returns The last response metadata or null if no request has been made
-     */
-    getLastResponseMetadata() {
-        return this.lastResponseMetadata;
-    }
-    /**
-     * Update response metadata from a captured limit snapshot.
-     *
-     * Takes already-parsed rate-limit info rather than raw headers: parsing now
-     * lives in `rateLimitCapture`, which is the only layer that sees the raw
-     * response and understands both header families (unified subscription
-     * windows and legacy per-tier counters).
-     *
-     * @param rateLimit - Parsed rate-limit figures
-     * @param requestId - Optional Anthropic request ID
-     * @param usageUpdate - Optional token counts to fold into usage tracking
-     */
-    updateResponseMetadata(rateLimit, requestId, usageUpdate) {
-        this.lastResponseMetadata = {
-            rateLimit,
-            ...(requestId ? { requestId } : {}),
-        };
-        if (this.usageInfo) {
-            this.usageInfo.requestCount++;
-            this.usageInfo.messagesUsed++;
-            this.usageInfo.lastRequestTimestamp = Date.now();
-            // Update token usage if provided
-            if (usageUpdate) {
-                if (usageUpdate.inputTokens !== undefined) {
-                    this.usageInfo.inputTokensUsed += usageUpdate.inputTokens;
-                    this.usageInfo.tokensUsed += usageUpdate.inputTokens;
-                }
-                if (usageUpdate.outputTokens !== undefined) {
-                    this.usageInfo.outputTokensUsed += usageUpdate.outputTokens;
-                    this.usageInfo.tokensUsed += usageUpdate.outputTokens;
-                }
-            }
-            // Update remaining quotas from rate limit headers
-            if (rateLimit?.requestsRemaining !== undefined) {
-                this.usageInfo.messagesRemaining = rateLimit.requestsRemaining;
-            }
-            if (rateLimit?.tokensRemaining !== undefined) {
-                this.usageInfo.tokensRemaining = rateLimit.tokensRemaining;
-            }
-            // Calculate quota percentages
-            if (rateLimit?.requestsLimit && rateLimit.requestsLimit > 0) {
-                this.usageInfo.messageQuotaPercent = Math.round(((rateLimit.requestsLimit - (rateLimit.requestsRemaining ?? 0)) /
-                    rateLimit.requestsLimit) *
-                    100);
-            }
-            if (rateLimit?.tokensLimit && rateLimit.tokensLimit > 0) {
-                this.usageInfo.tokenQuotaPercent = Math.round(((rateLimit.tokensLimit - (rateLimit.tokensRemaining ?? 0)) /
-                    rateLimit.tokensLimit) *
-                    100);
-            }
-            // Check for rate limiting
-            if (rateLimit?.retryAfter !== undefined) {
-                this.usageInfo.isRateLimited = true;
-                this.usageInfo.rateLimitExpiresAt =
-                    Date.now() + rateLimit.retryAfter * 1000;
-            }
-            else {
-                this.usageInfo.isRateLimited = false;
-                this.usageInfo.rateLimitExpiresAt = undefined;
-            }
-        }
-        // Log rate limit warnings if approaching limits
-        if (rateLimit?.requestsRemaining !== undefined) {
-            if (rateLimit.requestsRemaining <= 5) {
-                logger.warn("Approaching Anthropic request rate limit", {
-                    remaining: rateLimit.requestsRemaining,
-                    limit: rateLimit.requestsLimit,
-                    reset: rateLimit.requestsReset,
-                });
-            }
-        }
-        if (rateLimit?.tokensRemaining !== undefined) {
-            if (rateLimit.tokensLimit &&
-                rateLimit.tokensRemaining < rateLimit.tokensLimit * 0.1) {
-                logger.warn("Approaching Anthropic token rate limit", {
-                    remaining: rateLimit.tokensRemaining,
-                    limit: rateLimit.tokensLimit,
-                    reset: rateLimit.tokensReset,
-                });
-            }
-        }
-    }
-    getProviderName() {
-        return "anthropic";
-    }
-    getDefaultModel() {
-        return getDefaultAnthropicModel();
-    }
-    /**
-     * Returns a V3-shaped delegating model whose `doGenerate` drives the
-     * official Anthropic Messages API directly. BaseProvider's `generate()`
-     * path (and its middleware wrapping) keeps working unchanged; the
-     * streaming path bypasses this entirely via `executeStream`.
-     */
-    getAISDKModel() {
-        const client = this.client;
-        const providerName = this.providerName;
-        const modelId = this.modelName || getDefaultAnthropicModel();
-        const getTimeoutForOptions = (opts) => this.getTimeout((opts ?? {}));
-        const refreshAuth = () => this.refreshAuthIfNeeded();
-        // Structural AI-SDK model shape (`SageMakerAsLanguageModel`, the
-        // sanctioned intermediate from src/lib/types/providers.ts): assigning the
-        // literal to it keeps the compiler's member checks, and the intermediate
-        // is single-assertable to the ai-package `LanguageModel` handle.
-        const delegatingModel = {
-            specificationVersion: "v3",
-            provider: providerName,
-            modelId,
-            supportedUrls: {},
-            doGenerate: async (options) => {
-                await refreshAuth();
-                const built = messagesToAnthropic(options.prompt);
-                const messages = built.messages;
-                // `let`: the additive structured-output path below appends the
-                // final_result instruction to the system prompt.
-                let system = built.system;
-                let tools = (options.tools ?? [])
-                    .filter((t) => t.type === "function")
-                    .map((t) => {
-                    // GenerationHandler marks the last tool definition with a cache
-                    // breakpoint when prompt caching is active — keep honoring it.
-                    const cc = cacheControlOf(t);
-                    return {
-                        name: t.name,
-                        ...(t.description ? { description: t.description } : {}),
-                        input_schema: (t.inputSchema ?? {
-                            type: "object",
-                            properties: {},
-                        }),
-                        ...(cc ? { cache_control: cc } : {}),
-                    };
-                });
-                if (tools && tools.length === 0) {
-                    tools = undefined;
-                }
-                let toolChoice = options.toolChoice
-                    ? toolChoiceToAnthropic(options.toolChoice.type === "tool"
-                        ? { type: "tool", toolName: options.toolChoice.toolName }
-                        : options.toolChoice.type)
-                    : undefined;
-                // JSON/structured output: Anthropic has no response_format — emulate
-                // with a forced synthetic tool whose input IS the JSON payload (the
-                // same object-tool strategy @ai-sdk/anthropic used).
-                let jsonTool;
-                if (options.responseFormat?.type === "json") {
-                    jsonTool = options.responseFormat.name ?? "json";
-                    const schema = (options.responseFormat.schema ?? {
-                        type: "object",
-                    });
-                    tools = [
-                        {
-                            name: jsonTool,
-                            description: options.responseFormat.description ??
-                                "Respond by calling this tool with the answer as its input.",
-                            input_schema: schema,
-                        },
-                    ];
-                    toolChoice = { type: "tool", name: jsonTool };
-                }
-                // Additive structured output: when the caller wants a schema AND real
-                // tools, the forced-json path above cannot be used (it replaces the
-                // tools array), and the AI-SDK experimental_output path is excluded
-                // for this surface by structuredOutputPolicy. GenerationHandler hands
-                // the JSON Schema down here instead, and we APPEND a `final_result`
-                // tool — tool_choice stays auto, so every real tool keeps working and
-                // the model self-selects final_result when it is ready to answer.
-                const finalResultSchema = options.providerOptions?.anthropic
-                    ?.finalResultSchema;
-                let finalResultActive = false;
-                if (!jsonTool && finalResultSchema) {
-                    const appended = appendFinalResultTool(tools, finalResultSchema);
-                    tools = appended.tools;
-                    finalResultActive = appended.applied;
-                    if (appended.applied) {
-                        system = appendFinalResultInstruction(system);
-                    }
-                }
-                // Extended thinking passthrough (providerOptions.anthropic.thinking).
-                const thinking = options.providerOptions?.anthropic?.thinking;
-                // Prompt-cache parity with the native Vertex+Claude path: upstream
-                // layers mark only the stable prefix (system via MessageBuilder,
-                // last tool via GenerationHandler) — the growing conversation
-                // history has no breakpoint, so on every turn it falls after the
-                // last marker and is re-billed as fresh input. Add rolling history
-                // breakpoints in whatever budget remains under Anthropic's
-                // four-marker ceiling; pre-existing markers are counted so the
-                // request can never exceed the cap.
-                const cacheMarkersUsed = countAnthropicCacheMarkers({
-                    system,
-                    tools,
-                    messages: messages,
-                });
-                const cachedMessages = applyAnthropicHistoryCacheBreakpoints(messages, ANTHROPIC_MAX_CACHE_BREAKPOINTS - cacheMarkersUsed);
-                // Registry-driven strip: Sonnet 5 / Opus 4.7+ / Fable 5 families
-                // reject sampling params (also covers the retry paths — this params
-                // object is what any rebuild spreads).
-                const samplingParams = resolveSamplingParams("anthropic", modelId, {
-                    ...(options.temperature !== undefined &&
-                        options.temperature !== null
-                        ? { temperature: options.temperature }
-                        : {}),
-                    ...(options.topP !== undefined && options.topP !== null
-                        ? { topP: options.topP }
-                        : {}),
-                }, "anthropic.doGenerate");
-                const params = {
-                    model: modelId,
-                    messages: cachedMessages,
-                    max_tokens: resolveClaudeMaxTokens(modelId, options.maxOutputTokens),
-                    ...(system ? { system } : {}),
-                    ...(samplingParams.temperature !== undefined
-                        ? { temperature: samplingParams.temperature }
-                        : {}),
-                    ...(samplingParams.topP !== undefined
-                        ? { top_p: samplingParams.topP }
-                        : {}),
-                    ...(options.stopSequences && options.stopSequences.length > 0
-                        ? { stop_sequences: options.stopSequences }
-                        : {}),
-                    ...(tools ? { tools } : {}),
-                    ...(toolChoice ? { tool_choice: toolChoice } : {}),
-                    ...(thinking ? { thinking } : {}),
-                };
-                // The 60s anthropic generate default was tuned for the old ~4096
-                // max_tokens. Now that the default ceiling is the model's real max,
-                // a large structured response needs more wall-clock to be produced —
-                // otherwise the inner controller aborts mid-generation (the AI-SDK
-                // doGenerate layer doesn't see the caller's `timeout`). Raise the
-                // floor to 5 min when a large output budget is in play — but only
-                // when the caller did NOT set an explicit timeout: an explicit value
-                // is a contract and must never be silently extended. The abort
-                // signal stays the real bound.
-                const callerTimeout = options
-                    .timeout;
-                const callerSpecifiedTimeout = callerTimeout !== undefined && callerTimeout !== null;
-                const generateTimeoutMs = params.max_tokens > 8192 && !callerSpecifiedTimeout
-                    ? Math.max(getTimeoutForOptions(options), 300_000)
-                    : getTimeoutForOptions(options);
-                const timeoutController = createTimeoutController(generateTimeoutMs, providerName, "generate");
-                let response;
-                try {
-                    response = await client.messages.create(params, {
-                        signal: composeAbortSignals(options.abortSignal, timeoutController?.controller.signal),
-                    });
-                }
-                finally {
-                    timeoutController?.cleanup();
-                }
-                const content = [];
-                let finalResultText;
-                // Text emitted in forced-json mode, kept only as a fallback (see below).
-                const jsonModeText = [];
-                let jsonToolAnswered = false;
-                for (const block of response.content) {
-                    if (block.type === "thinking") {
-                        content.push({ type: "reasoning", text: block.thinking });
-                    }
-                    else if (block.type === "text") {
-                        // In forced-json mode the payload arrives via the tool input, not
-                        // text — pass text through only in normal mode.
-                        if (jsonTool) {
-                            jsonModeText.push(block.text);
-                        }
-                        else {
-                            content.push({ type: "text", text: block.text });
-                        }
-                    }
-                    else if (block.type === "tool_use") {
-                        if (jsonTool && block.name === jsonTool) {
-                            // Unwrap the synthetic tool call back into text JSON.
-                            jsonToolAnswered = true;
-                            content.push({
-                                type: "text",
-                                text: stringifyToolInput(block.input),
-                            });
-                        }
-                        else if (finalResultActive &&
-                            block.name === FINAL_RESULT_TOOL_NAME) {
-                            // Internal pattern: never surfaced as a tool call. Its arguments
-                            // ARE the structured answer.
-                            finalResultText = stringifyToolInput(block.input);
-                        }
-                        else {
-                            content.push({
-                                type: "tool-call",
-                                toolCallId: block.id,
-                                toolName: block.name,
-                                input: stringifyToolInput(block.input),
-                            });
-                        }
-                    }
-                }
-                // Forced-json mode normally drops text blocks because the payload rides
-                // in the synthetic tool's input. But when the response is cut short
-                // (stop_reason "max_tokens") the tool call can be missing entirely, and
-                // dropping the text would leave an EMPTY completion with nothing for
-                // coerceJsonToSchema to recover. Fall back to the text so a partial
-                // object can still be salvaged and flagged truncated.
-                if (jsonTool && !jsonToolAnswered && jsonModeText.length > 0) {
-                    content.push({ type: "text", text: jsonModeText.join("") });
-                }
-                // final_result is terminal — parity with the native Claude-on-Vertex
-                // and Gemini loops, which break out of the tool loop the moment it
-                // arrives. Reasoning blocks are kept; any prose preamble and any tool
-                // calls issued alongside it are dropped so `text` is exactly the
-                // structured payload and the AI-SDK loop stops here.
-                if (finalResultText !== undefined) {
-                    const reasoning = content.filter((part) => part.type === "reasoning");
-                    content.length = 0;
-                    content.push(...reasoning, { type: "text", text: finalResultText });
-                    logger.debug("[Anthropic] Extracted structured output from final_result tool (generate)", { chars: finalResultText.length });
-                }
-                const cacheRead = response.usage.cache_read_input_tokens ?? 0;
-                const cacheWrite = response.usage.cache_creation_input_tokens ?? 0;
-                return {
-                    content,
-                    finishReason: {
-                        // A final_result call ends the turn: the provider reports
-                        // stop_reason "tool_use", but no tool call is surfaced, so
-                        // reporting "tool-calls" would misread as a step-capped turn.
-                        // `raw` still carries the provider's verbatim stop_reason.
-                        unified: finalResultText !== undefined
-                            ? "stop"
-                            : mapAnthropicStopReason(response.stop_reason),
-                        raw: response.stop_reason ?? "stop",
-                    },
-                    usage: {
-                        inputTokens: {
-                            total: response.usage.input_tokens + cacheRead + cacheWrite,
-                            noCache: response.usage.input_tokens,
-                            cacheRead,
-                            cacheWrite,
-                        },
-                        outputTokens: {
-                            total: response.usage.output_tokens,
-                            text: response.usage.output_tokens,
-                            reasoning: undefined,
-                        },
-                    },
-                    warnings: [],
-                    request: { body: params },
-                    response: {
-                        id: response.id,
-                        modelId: response.model,
-                        // Real response headers, captured by the fetch wrapper. This used
-                        // to be a hardcoded `{}`, which silently discarded every
-                        // rate-limit and quota header Anthropic returns.
-                        headers: getCapturedResponseHeaders() ?? {},
-                        body: response,
-                    },
-                };
-            },
-            doStream: () => {
-                throw new Error(`${providerName}: doStream is not implemented on the delegating model — the streaming path uses executeStream directly.`);
-            },
-        };
-        return delegatingModel;
-    }
-    formatProviderError(error) {
-        const rules = [
-            {
-                // Plan 02's mocked-contract work documented a pre-existing gap: the
-                // Anthropic SDK formats auth failures as a bare "401 <msg>" string,
-                // which the message-text regex alone does not catch — mirrors
-                // Vertex's statusCode === 401 fallback in this same task.
-                match: (ctx) => ctx.statusCode === 401 ||
-                    /API_KEY_INVALID|Invalid API key/i.test(ctx.message),
-                errorClass: AuthenticationError,
-                message: "Invalid Anthropic API key. Please check your ANTHROPIC_API_KEY environment variable.",
-            },
-            {
-                match: (ctx) => ctx.statusCode === 429 ||
-                    /rate limit|too_many_requests|429/i.test(ctx.message),
-                errorClass: RateLimitError,
-                message: "Anthropic rate limit exceeded. Please try again later.",
-            },
-            {
-                match: (ctx) => /ECONNRESET|ENOTFOUND|ECONNREFUSED|network|connection/i.test(ctx.message),
-                errorClass: NetworkError,
-                message: (ctx) => `Connection error: ${ctx.message}`,
-            },
-            {
-                match: (ctx) => /500|502|503|504|server error/i.test(ctx.message),
-                errorClass: ProviderError,
-                message: (ctx) => `Server error: ${ctx.message}`,
-            },
-            {
-                match: () => true,
-                errorClass: ProviderError,
-                message: (ctx) => `Anthropic error: ${ctx.message}`,
-            },
-        ];
-        return classifyProviderError(error, rules, this.providerName, this.modelName);
-    }
-    // executeGenerate removed - BaseProvider handles all generation with tools
-    /**
-     * Override generate to refresh the OAuth token before delegating to
-     * BaseProvider so that expired tokens are renewed automatically.
-     */
-    async generate(optionsOrPrompt, analysisSchema) {
-        await this.refreshAuthIfNeeded();
-        // Open a per-request capture scope around the whole turn. Scoping here
-        // rather than on the instance is what makes it concurrency-safe: several
-        // generate() calls can be in flight on one provider instance, and an
-        // instance field would attribute one call's limits to another.
-        const { result, snapshot } = await withLimitCapture(() => super.generate(optionsOrPrompt, analysisSchema));
-        if (result && snapshot) {
-            this.recordLimitSnapshot(snapshot);
-            result.limits = snapshot;
-            if (result.analytics) {
-                result.analytics.limits = snapshot;
-            }
-        }
-        return result;
-    }
-    /**
-     * Fold a captured snapshot into the provider's usage bookkeeping and log it.
-     *
-     * `updateResponseMetadata` had no callers before this — the metadata it
-     * maintains, and the public `getLastResponseMetadata()` / `getUsageInfo()`
-     * that read it, were never populated by anything.
-     */
-    recordLimitSnapshot(snapshot) {
-        this.updateResponseMetadata(snapshot.rateLimit, snapshot.requestId);
-        setLimitSpanAttributes(snapshot);
-        logClaudeLimitSnapshot(snapshot, this.modelName);
-    }
-    async executeStream(options, analysisSchema) {
-        // The capture scope must outlive this call: the SSE loop keeps running in
-        // the background after executeStream returns, and its per-step HTTP
-        // requests are what carry the limit headers. AsyncLocalStorage.run
-        // propagates into every continuation started inside, so the whole stream
-        // lifetime shares one slot.
-        return runInLimitCaptureScope(() => this.executeStreamInCaptureScope(options, analysisSchema));
-    }
-    async executeStreamInCaptureScope(options, _analysisSchema) {
-        // Refresh OAuth token if needed before making any API request.
-        await this.refreshAuthIfNeeded();
-        this.validateStreamOptions(options);
-        const timeout = this.getTimeout(options);
-        const timeoutController = createTimeoutController(timeout, this.providerName, "stream");
-        // Consumer-driven abort: fires when the async iterator is closed early
-        // (caller breaks out of `for await`) so the background loop stops
-        // reading SSE and running tools.
-        const consumerAbortController = new AbortController();
-        const abortSignal = mergeAbortSignals([
-            options.abortSignal,
-            timeoutController?.controller.signal,
-            consumerAbortController.signal,
-        ]).signal;
-        let toolsRecord;
-        let anthropicTools;
-        let payload;
-        let shouldUseTools;
-        // True once the additive `final_result` tool is in the request — the
-        // streaming twin of the doGenerate path above.
-        let finalResultActive = false;
-        try {
-            // options.tools is pre-merged by BaseProvider.stream() with base tools
-            // (MCP/built-in) + user-provided tools (RAG, etc.)
-            shouldUseTools = !options.disableTools && this.supportsTools();
-            toolsRecord = shouldUseTools
-                ? options.tools || (await this.getAllTools())
-                : {};
-            anthropicTools = shouldUseTools
-                ? toNativeToolDeclarations(toolsRecord, "input_schema")
-                : undefined;
-            // Build message array from options with multimodal support, then
-            // convert to the Anthropic Messages payload (system + content blocks).
-            const built = await this.buildMessagesForStream(options);
-            payload = messagesToAnthropic(built);
-            // Schema + tools: append final_result rather than pinning tool_choice to
-            // a json tool, so the real tools stay callable for the whole turn.
-            // Unlike generate, no plumbing is needed — this is a native loop, so the
-            // caller's Zod/JSON schema is right here on the options.
-            if (options.schema && anthropicTools && anthropicTools.length > 0) {
-                const appended = appendFinalResultTool(anthropicTools, convertZodToJsonSchema(options.schema));
-                anthropicTools = appended.tools;
-                finalResultActive = appended.applied;
-                if (appended.applied) {
-                    payload.system = appendFinalResultInstruction(payload.system);
-                }
-            }
-        }
-        catch (setupErr) {
-            timeoutController?.cleanup();
-            throw this.handleProviderError(setupErr);
-        }
-        const modelId = this.modelName || getDefaultAnthropicModel();
-        const anthropicToolChoice = shouldUseTools && anthropicTools && anthropicTools.length > 0
-            ? toolChoiceToAnthropic(resolveToolChoice(options, toolsRecord, shouldUseTools))
-            : undefined;
-        // Extended thinking: enabled when the caller supplies an explicit token
-        // budget (mirrors the previous experimental_thinking gating). Thinking
-        // deltas stream out on the `reasoning` chunk channel.
-        const thinking = options.thinkingConfig?.enabled && options.thinkingConfig.budgetTokens
-            ? {
-                type: "enabled",
-                budget_tokens: options.thinkingConfig.budgetTokens,
-            }
-            : undefined;
-        // Wrap the native stream in an OTel span to capture provider-level
-        // latency and token usage (same span name as the pre-migration path so
-        // dashboards stay continuous).
-        const streamSpan = streamTracer.startSpan("neurolink.provider.streamText", {
-            kind: SpanKind.CLIENT,
-            attributes: {
-                "gen_ai.system": "anthropic",
-                "gen_ai.request.model": modelId,
-            },
-        });
-        const maxSteps = options.maxSteps || DEFAULT_MAX_STEPS;
-        const emitter = this.neurolink?.getEventEmitter();
-        const channel = createStreamChannel();
-        const { push: pushChunk } = channel;
-        const { usagePromise, finishPromise, resolveUsage, resolveFinish } = createDeferredAnalytics();
-        usagePromise
-            .then((usage) => {
-            streamSpan.setAttribute("gen_ai.usage.input_tokens", usage.promptTokens || 0);
-            streamSpan.setAttribute("gen_ai.usage.output_tokens", usage.completionTokens || 0);
-            if (usage.cacheReadTokens) {
-                streamSpan.setAttribute("gen_ai.usage.cached_input_tokens", usage.cacheReadTokens);
-            }
-            const cost = calculateCost(this.providerName, this.modelName, {
-                input: usage.promptTokens || 0,
-                output: usage.completionTokens || 0,
-                total: usage.totalTokens || 0,
-                ...(usage.cacheReadTokens
-                    ? { cacheReadTokens: usage.cacheReadTokens }
-                    : {}),
-                ...(usage.cacheCreationTokens
-                    ? { cacheCreationTokens: usage.cacheCreationTokens }
-                    : {}),
-            });
-            if (cost && cost > 0) {
-                streamSpan.setAttribute("neurolink.cost", cost);
-            }
-        })
-            .catch(() => {
-            // usage may never resolve if the stream is aborted before completion
-        });
-        finishPromise
-            .then((reason) => {
-            streamSpan.setAttribute("gen_ai.response.finish_reason", reason || "unknown");
-            streamSpan.end();
-        })
-            .catch(() => {
-            streamSpan.end();
-        });
-        let capturedProviderError;
-        const client = this.client;
-        const toolsUsed = [];
-        const streamStartTime = Date.now();
-        // Hoisted out of runLoop so the error path can resolve the usage
-        // accumulated by steps that completed BEFORE the failure — those steps
-        // were billed and must not be reported as zero.
-        let totalInput = 0;
-        let totalOutput = 0;
-        let totalCacheRead = 0;
-        let totalCacheWrite = 0;
-        let lastStop = null;
-        const buildDeferredUsage = () => ({
-            promptTokens: totalInput,
-            completionTokens: totalOutput,
-            totalTokens: totalInput + totalCacheRead + totalCacheWrite + totalOutput,
-            ...(totalCacheRead > 0 ? { cacheReadTokens: totalCacheRead } : {}),
-            ...(totalCacheWrite > 0 ? { cacheCreationTokens: totalCacheWrite } : {}),
-        });
-        // Structured-output turns are delivered as ONE chunk, not incrementally:
-        // a caller that passed a schema needs parseable JSON, and text deltas
-        // emitted before the model calls final_result would prefix the payload
-        // with prose and break every JSON.parse on the consumer side. Same
-        // contract as the native Vertex loops. Non-schema streams are untouched
-        // and stay fully incremental.
-        let bufferedText = "";
-        let finalResultText;
-        /** System prompt + tool definitions: they ride outside `messages`. */
-        const estimateAnthropicFixedOverhead = (system, tools) => {
-            const text = (value) => {
-                if (typeof value === "string") {
-                    return value;
-                }
-                try {
-                    return JSON.stringify(value) ?? "";
-                }
-                catch {
-                    return "";
-                }
-            };
-            return (estimateTokens(text(system), "anthropic") +
-                estimateTokens(text(tools), "anthropic"));
-        };
-        const runLoop = async () => {
-            // The provider's REAL prompt-token count for the previous step,
-            // calibrating the guard's char-based estimate for free, paired with the
-            // guard's own estimate for that same request — a ratio between counts of
-            // two different payloads would be meaningless.
-            let lastObservedPromptTokens;
-            let lastSentEstimate;
-            /**
-             * Reclaim, as a PURE function of the conversation.
-             *
-             * The hand-rolled loop this replaces rebuilt in place
-             * (`conversation.length = 0; conversation.push(...rebuilt)`) because it
-             * owned the array. The engine owns it now and assigns what this
-             * returns, so mutating here would corrupt a retried or reclaimed step.
-             */
-            const planReclaim = (conversation) => {
-                const reclaim = planAnthropicLoopReclaim({
-                    conversation,
-                    availableInputTokens: getAvailableInputTokens("anthropic", modelId, options.maxTokens ?? undefined),
-                    fixedOverheadTokens: estimateAnthropicFixedOverhead(payload.system, anthropicTools),
-                    provider: "anthropic",
-                    observedPromptTokens: lastObservedPromptTokens,
-                    // Both halves of the calibration ratio must describe the same
-                    // request: the tokens the provider reported, and this guard's own
-                    // estimate for what was sent to earn them.
-                    previousSentEstimate: lastSentEstimate,
-                    onSentEstimate: (tokens) => {
-                        lastSentEstimate = tokens;
-                    },
-                });
-                if (!reclaim) {
-                    return undefined;
-                }
-                // Dropping an assistant tool_use message together with its user
-                // tool_result message is what keeps blocks paired.
-                const dropSet = new Set(reclaim.drop);
-                const truncateSet = new Set(reclaim.truncate);
-                const rebuilt = [];
-                for (let i = 0; i < conversation.length; i++) {
-                    if (dropSet.has(i)) {
-                        continue;
-                    }
-                    const message = conversation[i];
-                    if (truncateSet.has(i) && Array.isArray(message.content)) {
-                        rebuilt.push({
-                            ...message,
-                            content: message.content.map((block) => block.type === "tool_result"
-                                ? {
-                                    ...block,
-                                    content: previewAnthropicToolResultText(typeof block.content === "string"
-                                        ? block.content
-                                        : (JSON.stringify(block.content) ?? "")),
-                                }
-                                : block),
-                        });
-                        continue;
-                    }
-                    rebuilt.push(message);
-                }
-                if (dropSet.size > 0) {
-                    // Anthropic requires user/assistant alternation around tool blocks;
-                    // the note is a user turn placed immediately before the first
-                    // surviving assistant tool_use turn, which preserves it.
-                    let noteIndex = rebuilt.findIndex((m) => Array.isArray(m.content) &&
-                        m.content.some((b) => b.type === "tool_use" || b.type === "tool_result"));
-                    if (noteIndex < 0) {
-                        noteIndex = Math.min(1, rebuilt.length);
-                    }
-                    rebuilt.splice(noteIndex, 0, {
-                        role: "user",
-                        content: [{ type: "text", text: ANTHROPIC_ELISION_NOTE }],
-                    });
-                }
-                return rebuilt;
-            };
-            const buildParams = (conversation) => {
-                // Mid-turn discovery sync: search_tools (tools.discovery) hydrates
-                // new tools into toolsRecord between steps; Claude only calls tools
-                // declared in the request, so advertise them now.
-                if (anthropicTools) {
-                    const declared = new Set(anthropicTools.map((t) => t.name));
-                    const hydrated = Object.fromEntries(Object.entries(toolsRecord).filter(([name]) => !declared.has(name)));
-                    if (Object.keys(hydrated).length > 0) {
-                        const extra = toNativeToolDeclarations(hydrated, "input_schema");
-                        if (extra && extra.length > 0) {
-                            anthropicTools = [...anthropicTools, ...extra];
-                        }
-                    }
-                }
-                // Prompt-cache parity with the native Vertex+Claude path — rolling
-                // history breakpoints, re-applied per step so the stable prefix stays
-                // byte-identical while the breakpoint follows the growing tail.
-                const cacheMarkersUsed = countAnthropicCacheMarkers({
-                    system: payload.system,
-                    tools: anthropicTools,
-                    messages: conversation,
-                });
-                const cachedConversation = applyAnthropicHistoryCacheBreakpoints(conversation, ANTHROPIC_MAX_CACHE_BREAKPOINTS - cacheMarkersUsed);
-                const streamSamplingParams = resolveSamplingParams("anthropic", modelId, options.temperature !== undefined && options.temperature !== null
-                    ? { temperature: options.temperature }
-                    : {}, "anthropic.executeStream");
-                return {
-                    model: modelId,
-                    messages: cachedConversation,
-                    max_tokens: resolveClaudeMaxTokens(modelId, options.maxTokens),
-                    stream: true,
-                    ...(payload.system ? { system: payload.system } : {}),
-                    ...(streamSamplingParams.temperature !== undefined
-                        ? { temperature: streamSamplingParams.temperature }
-                        : {}),
-                    ...(anthropicTools && anthropicTools.length > 0
-                        ? { tools: anthropicTools }
-                        : {}),
-                    ...(anthropicToolChoice ? { tool_choice: anthropicToolChoice } : {}),
-                    ...(thinking ? { thinking } : {}),
-                };
-            };
-            const baseAdapter = createAnthropicLoopAdapter({
-                client,
-                maxSteps,
-                toolsRecord,
-                buildParams,
-                planReclaim,
-                noteObservedPromptTokens: (tokens) => {
-                    lastObservedPromptTokens = tokens;
-                },
-                ...(finalResultActive
-                    ? {
-                        finalResultToolName: FINAL_RESULT_TOOL_NAME,
-                        onTerminalResult: (text) => {
-                            finalResultText = text;
-                            logger.debug("[Anthropic] Extracted structured output from final_result tool (stream)", { chars: text.length });
-                        },
-                    }
-                    : {}),
-            });
-            // Wrapped rather than folded into the adapter: analytics emission and
-            // tool-execution storage fire ONCE PER STEP in the loop this replaces,
-            // and buildToolResultMessages is the only per-step hook — it receives
-            // exactly that step's results. Doing this from the turn's final result
-            // instead would batch every step's tools into one late write.
-            const adapter = {
-                ...baseAdapter,
-                buildToolResultMessages: (conversation, stepResult, toolResults) => {
-                    for (const result of toolResults) {
-                        toolsUsed.push(result.name);
-                    }
-                    const toolCallsForStorage = toolResults.map((result) => ({
-                        type: "tool-call",
-                        toolCallId: result.id,
-                        toolName: result.name,
-                        args: result.args,
-                    }));
-                    const toolResultsForStorage = toolResults.map((result) => result.error
-                        ? {
-                            type: "tool-result",
-                            toolCallId: result.id,
-                            toolName: result.name,
-                            error: result.error,
-                        }
-                        : {
-                            type: "tool-result",
-                            toolCallId: result.id,
-                            toolName: result.name,
-                            result: result.output,
-                        });
-                    emitToolEndFromStepFinish(emitter, toolResultsForStorage.map((tr) => ({
-                        toolName: tr.toolName,
-                        result: "result" in tr ? tr.result : undefined,
-                        error: "error" in tr ? tr.error : undefined,
-                    })));
-                    this.handleToolExecutionStorage(toolCallsForStorage, toolResultsForStorage, options, new Date()).catch((storageErr) => {
-                        logger.warn("[AnthropicProvider] Failed to store tool executions", {
-                            provider: this.providerName,
-                            error: storageErr instanceof Error
-                                ? storageErr.message
-                                : String(storageErr),
-                        });
-                    });
-                    return baseAdapter.buildToolResultMessages(conversation, stepResult, toolResults);
-                },
-            };
-            // Presented in the shape the engine dispatches through. The engine
-            // supplies `{ toolCallId, abortSignal }`; the loop this replaces also
-            // supplied `messages: []`, so it is kept — a tool that reads it would
-            // otherwise see undefined where it used to see an empty array.
-            const engineTools = {};
-            for (const [name, tool] of Object.entries(toolsRecord)) {
-                const execute = tool.execute;
-                if (!execute) {
-                    continue;
-                }
-                engineTools[name] = {
-                    execute: async (args, opts) => {
-                        const ctx = opts;
-                        return execute(args, {
-                            toolCallId: ctx.toolCallId ?? "",
-                            ...(ctx.abortSignal ? { abortSignal: ctx.abortSignal } : {}),
-                            messages: [],
-                        });
-                    },
-                };
-            }
-            const { stream, resultPromise } = runAgenticLoop(adapter, payload.messages.slice(), {
-                tools: engineTools,
-                ...(abortSignal ? { abortSignal } : {}),
-            });
-            // Structured turns buffer their text rather than streaming it: a caller
-            // that passed a schema needs parseable JSON, and deltas emitted before
-            // the model calls final_result would prefix the payload with prose.
-            // Reasoning is forwarded either way — it is not part of the payload.
-            const pump = (async () => {
-                for await (const chunk of stream) {
-                    if (chunk.reasoning) {
-                        pushChunk({ content: "", reasoning: chunk.reasoning });
-                    }
-                    if (chunk.content) {
-                        if (finalResultActive) {
-                            bufferedText += chunk.content;
-                        }
-                        else {
-                            pushChunk({ content: chunk.content });
-                        }
-                    }
-                }
-            })();
-            const result = await resultPromise;
-            await pump;
-            totalInput += result.usage.inputTokens;
-            totalOutput += result.usage.outputTokens;
-            totalCacheRead += result.usage.cacheReadTokens ?? 0;
-            totalCacheWrite += result.usage.cacheWriteTokens ?? 0;
-            lastStop = result.rawStopReason ?? lastStop;
-            resolveUsage(buildDeferredUsage());
-            resolveFinish(lastStop ?? "stop");
-        };
-        const loopPromise = runLoop()
-            // Parameter named `error` so the compiled `capturedProviderError = error`
-            // assignment matches the regression-grep in test:context 6.14.
-            .catch((error) => {
-            capturedProviderError = error;
-            logger.error("Anthropic: Stream error", {
-                error: error instanceof Error ? error.message : String(error),
-            });
-            // Report whatever the completed steps accumulated — they were billed
-            // — and unblock any consumer awaiting the usage promise.
-            resolveUsage(buildDeferredUsage());
-            resolveFinish("error");
-            throw this.formatProviderError(error);
-        })
-            .finally(() => {
-            // Deliver the buffered structured-output turn: `finalResultText` when
-            // the model called final_result, otherwise the prose it produced
-            // instead — never nothing, so a model that ignores the instruction
-            // degrades to today's plain-text behaviour rather than an empty
-            // stream. In `finally` so a turn that dies mid-loop still surfaces
-            // the text it had already buffered, exactly as the unbuffered path
-            // surfaces its partial deltas.
-            if (finalResultActive) {
-                const output = finalResultText ?? bufferedText;
-                if (output.length > 0) {
-                    pushChunk({ content: output });
-                }
-            }
-            timeoutController?.cleanup();
-            channel.close();
-        });
-        loopPromise.catch(() => {
-            // Swallowed by design: the generator below surfaces loop errors after
-            // draining the channel; this guard only prevents an unhandled-rejection
-            // crash when the consumer abandons the stream early.
-        });
-        const providerName = this.providerName;
-        const transformedStream = async function* () {
-            let contentYielded = 0;
-            try {
-                for await (const chunk of channel.iterable) {
-                    if ("content" in chunk &&
-                        typeof chunk.content === "string" &&
-                        chunk.content.length > 0) {
-                        contentYielded++;
-                    }
-                    yield chunk;
-                }
-                // Surface any error the loop threw after draining the channel.
-                await loopPromise;
-                // No-output path: stream completed normally but yielded zero text.
-                if (contentYielded === 0 && toolsUsed.length === 0) {
-                    logger.warn(`${providerName}: Stream produced no output — emitting enriched sentinel`);
-                    const fauxNoOutput = new NoOutputGeneratedError({
-                        message: "Stream produced no output",
-                    });
-                    const sentinel = await buildNoOutputSentinel(fauxNoOutput, undefined, capturedProviderError);
-                    stampNoOutputSpan(sentinel);
-                    yield sentinel;
-                }
-            }
-            catch (streamError) {
-                const sentinel = await buildNoOutputSentinel(streamError, undefined, capturedProviderError);
-                stampNoOutputSpan(sentinel);
-                yield sentinel;
-                if (!NoOutputGeneratedError.isInstance(streamError)) {
-                    throw streamError;
-                }
-            }
-            finally {
-                if (!consumerAbortController.signal.aborted) {
-                    consumerAbortController.abort();
-                }
-            }
-        };
-        return {
-            stream: transformedStream(),
-            provider: this.providerName,
-            model: this.modelName,
-            toolCalls: [],
-            toolResults: [],
-            // Wire the deferred usage/finish promises into the analytics collector
-            // (mirrors openaiChatCompletionsBase). Without this the loop computed a
-            // fully correct aggregate that was consumed only by the OTel span —
-            // stream consumers and session cost tracking saw no usage at all.
-            // Chained off finishPromise so requestDuration reflects the DRAINED
-            // stream, not the milliseconds it took to construct this result object.
-            analytics: finishPromise.then(async () => {
-                const analytics = await streamAnalyticsCollector.createAnalytics(this.providerName, modelId, {
-                    textStream: (async function* () { })(),
-                    usage: usagePromise,
-                    finishReason: finishPromise,
-                }, Date.now() - streamStartTime, {
-                    requestId: options.requestId ??
-                        `${this.providerName}-stream-${Date.now()}`,
-                    streamingMode: true,
-                });
-                // Still inside the capture scope opened by executeStream, so this sees
-                // the limits reported by the last upstream step of the stream.
-                const snapshot = getCapturedLimitSnapshot();
-                if (snapshot && analytics) {
-                    this.recordLimitSnapshot(snapshot);
-                    analytics.limits = snapshot;
-                }
-                return analytics;
-            }),
-        };
-    }
-    async isAvailable() {
-        try {
-            // Check OAuth token first
-            const oauthToken = getOAuthToken();
-            if (oauthToken) {
-                return true;
-            }
-            // Fall back to API key check
-            getAnthropicApiKey();
-            return true;
-        }
-        catch {
-            return false;
-        }
-    }
-    getModel() {
-        return this.getAISDKModel();
-    }
-}
-// Re-export types and utilities for convenience
-export { getModelCapabilities, getRecommendedModelForTier, isModelAvailableForTier, ModelAccessError, } from "../../models/anthropicModels.js";
-//# sourceMappingURL=client.js.map