@sayknow-cli/ai 0.3.5 → 0.3.7

package/dist/types/providers/anthropic.d.ts

@@ -134,6 +134,7 @@ export type AnthropicClientOptionsArgs = {
     onSseEvent?: AnthropicOptions["onSseEvent"];
     fetch?: FetchImpl;
     requestMaxRetries?: number;
+    maxRetryDelayMs?: number;
 };
 export type AnthropicClientOptionsResult = {
     isOAuthToken: boolean;

package/dist/types/utils/oauth/deepinfra.d.ts

@@ -0,0 +1 @@
+export declare const loginDeepInfra: (options: import("./types").OAuthController) => Promise<string>;

package/dist/types/utils.d.ts

@@ -2,6 +2,7 @@ import type { ResponseInput } from "openai/resources/responses/responses";
 import type { CacheRetention, OpenAIResponsesHistoryPayload, ProviderPayload } from "./types";
 export { isRecord } from "@sayknow-cli/utils";
 export declare function normalizeSystemPrompts(systemPrompt: readonly string[] | string | undefined | null): string[];
+export declare function sanitizeJsonStrings(value: unknown): unknown;
 export declare function toNumber(value: unknown): number | undefined;
 export declare function toPositiveNumber(value: unknown, fallback: number): number;
 export declare function toBoolean(value: unknown): boolean | undefined;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@sayknow-cli/ai",
-	"version": "0.3.5",
+	"version": "0.3.7",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"homepage": "https://github.com/jaybeyond/Sayknow_CLI",
 	"author": "jaybeyond",
@@ -43,7 +43,7 @@
 	"dependencies": {
 		"@anthropic-ai/sdk": "^0.94.0",
 		"@bufbuild/protobuf": "^2.12.0",
-		"@sayknow-cli/utils": "0.3.5",
+		"@sayknow-cli/utils": "0.3.7",
 		"openai": "^6.36.0",
 		"partial-json": "^0.1.7",
 		"zod": "4.4.3"

package/src/providers/amazon-bedrock.ts

@@ -28,7 +28,7 @@ import type {
 	ToolChoice,
 	ToolResultMessage,
 } from "../types";
-import { normalizeToolCallId, resolveCacheRetention } from "../utils";
+import { normalizeToolCallId, resolveCacheRetention, sanitizeJsonStrings } from "../utils";
 import { AssistantMessageEventStream } from "../utils/event-stream";
 import { appendRawHttpRequestDumpFor400, type RawHttpRequestDump, withHttpStatus } from "../utils/http-inspector";
 import { parseStreamingJson } from "../utils/json-parse";
@@ -701,7 +701,7 @@ function convertMessages(
 								toolUse: {
 									toolUseId: normalizeToolCallId(c.id),
 									name: c.name,
-									input: c.arguments,
+									input: sanitizeJsonStrings(c.arguments),
 								},
 							});
 							break;

package/src/providers/anthropic-messages-server.ts

@@ -13,6 +13,7 @@ import type {
 	ToolResultMessage,
 	UserMessage,
 } from "../types";
+import { sanitizeJsonStrings } from "../utils";
 import {
 	type AnthropicAssistantContentBlock,
 	type AnthropicMessage,
@@ -433,7 +434,7 @@ function encodeContentBlocks(message: AssistantMessage): Record<string, unknown>
 				blocks.push({ type: "redacted_thinking", data: c.data });
 				break;
 			case "toolCall":
-				blocks.push({ type: "tool_use", id: c.id, name: c.name, input: c.arguments ?? {} });
+				blocks.push({ type: "tool_use", id: c.id, name: c.name, input: sanitizeJsonStrings(c.arguments ?? {}) });
 				break;
 		}
 	}

package/src/providers/anthropic.ts

@@ -21,6 +21,7 @@ import {
 } from "@sayknow-cli/utils";
 import { hasOpus47ApiRestrictions, mapEffortToAnthropicAdaptiveEffort } from "../model-thinking";
 import { calculateCost } from "../models";
+import { isUsageLimitError } from "../rate-limit-utils";
 import { getEnvApiKey, OUTPUT_FALLBACK_BUFFER } from "../stream";
 import type {
 	Api,
@@ -52,6 +53,7 @@ import {
 	normalizeSystemPrompts,
 	normalizeToolCallId,
 	resolveCacheRetention,
+	sanitizeJsonStrings,
 } from "../utils";
 import { createAbortSourceTracker } from "../utils/abort";
 import { AssistantMessageEventStream } from "../utils/event-stream";
@@ -62,6 +64,7 @@ import { parseJsonWithRepair, parseStreamingJson } from "../utils/json-parse";
 import { parseGitHubCopilotApiKey } from "../utils/oauth/github-copilot";
 import { notifyProviderResponse } from "../utils/provider-response";
 import { isCopilotTransientModelError } from "../utils/retry";
+import { getRetryAfterMsFromHeaders } from "../utils/retry-after";
 import { resolveRetryBudget } from "../utils/retry-budget";
 import {
 	COMBINATOR_KEYS,
@@ -724,6 +727,7 @@ export type AnthropicClientOptionsArgs = {
 	onSseEvent?: AnthropicOptions["onSseEvent"];
 	fetch?: FetchImpl;
 	requestMaxRetries?: number;
+	maxRetryDelayMs?: number;
 };
 
 export type AnthropicClientOptionsResult = {
@@ -877,6 +881,101 @@ function mergeHeaders(...headerSources: (Record<string, string> | undefined)[]):
 	return merged;
 }
 
+const ANTHROPIC_RETRY_DELAY_CAP_MS = 60_000;
+const ANTHROPIC_RATE_LIMIT_HEADER_PREFIX = "anthropic-ratelimit-";
+
+function getSafeAnthropicHeaderEvidence(headers: Headers): string[] {
+	const evidence: string[] = [];
+	for (const [name, value] of headers) {
+		const lowerName = name.toLowerCase();
+		if (
+			lowerName === "retry-after" ||
+			lowerName === "retry-after-ms" ||
+			lowerName.startsWith(ANTHROPIC_RATE_LIMIT_HEADER_PREFIX)
+		) {
+			evidence.push(`${lowerName}=${value}`);
+		}
+	}
+	return evidence.sort();
+}
+
+function getStringProperty(source: unknown, key: string): string | undefined {
+	if (!isRecord(source)) return undefined;
+	const value = source[key];
+	return typeof value === "string" ? value : undefined;
+}
+
+function appendAnthropicRateLimitEvidence(bodyText: string, headers: Headers): string {
+	const evidence = getSafeAnthropicHeaderEvidence(headers);
+	if (evidence.length === 0) return bodyText;
+
+	const suffix = ` Anthropic rate-limit evidence: ${evidence.join(", ")}`;
+	try {
+		const parsed = JSON.parse(bodyText) as unknown;
+		if (isRecord(parsed)) {
+			const error = parsed.error;
+			if (isRecord(error) && typeof error.message === "string" && !error.message.includes(suffix)) {
+				return JSON.stringify({ ...parsed, error: { ...error, message: `${error.message}${suffix}` } });
+			}
+			if (typeof parsed.message === "string" && !parsed.message.includes(suffix)) {
+				return JSON.stringify({ ...parsed, message: `${parsed.message}${suffix}` });
+			}
+		}
+	} catch {}
+
+	return bodyText.includes(suffix) ? bodyText : `${bodyText}${suffix}`;
+}
+
+function isAnthropicUsageExhaustionResponse(
+	bodyText: string,
+	headers: Headers,
+	retryAfterMs: number | undefined,
+	retryDelayCapMs: number,
+): boolean {
+	const overageReason = headers.get("anthropic-ratelimit-unified-overage-disabled-reason")?.toLowerCase();
+	if (overageReason === "out_of_credits") return true;
+	if (retryAfterMs !== undefined && retryAfterMs > retryDelayCapMs) return true;
+
+	try {
+		const parsed = JSON.parse(bodyText) as unknown;
+		const error = isRecord(parsed) ? parsed.error : undefined;
+		const type = getStringProperty(error, "type") ?? getStringProperty(parsed, "type");
+		const message = getStringProperty(error, "message") ?? getStringProperty(parsed, "message") ?? bodyText;
+		return (
+			/rate_limit_error/i.test(type ?? "") &&
+			(/request would exceed your account.?s rate limit/i.test(message) || /out_of_credits/i.test(message))
+		);
+	} catch {
+		return /request would exceed your account.?s rate limit|out_of_credits/i.test(bodyText);
+	}
+}
+
+function wrapAnthropicFetchForBoundedRateLimits(baseFetch: FetchImpl, maxRetryDelayMs: number | undefined): FetchImpl {
+	const retryDelayCapMs = maxRetryDelayMs ?? ANTHROPIC_RETRY_DELAY_CAP_MS;
+	return Object.assign(
+		async (input: string | URL | Request, init?: RequestInit): Promise<Response> => {
+			const response = await baseFetch(input, init);
+			if (response.status !== 429 || retryDelayCapMs === 0) return response;
+
+			const headers = new Headers(response.headers);
+			const retryAfterMs = getRetryAfterMsFromHeaders(headers);
+			const bodyText = await response
+				.clone()
+				.text()
+				.catch(() => "");
+			if (!isAnthropicUsageExhaustionResponse(bodyText, headers, retryAfterMs, retryDelayCapMs)) return response;
+
+			headers.set("x-should-retry", "false");
+			return new Response(appendAnthropicRateLimitEvidence(bodyText, headers), {
+				status: response.status,
+				statusText: response.statusText,
+				headers,
+			});
+		},
+		baseFetch.preconnect ? { preconnect: baseFetch.preconnect } : {},
+	);
+}
+
 // The Anthropic SDK logs malformed SSE frames directly before rethrowing them.
 // We surface the resulting provider error ourselves, so keep the SDK quiet.
 const ANTHROPIC_SDK_LOG_LEVEL = "off" as const;
@@ -1042,6 +1141,7 @@ export function isProviderRetryableError(error: unknown, provider?: string): boo
 	if (!(error instanceof Error)) return false;
 	if (provider === "github-copilot" && isCopilotTransientModelError(error)) return true;
 	const msg = error.message.toLowerCase();
+	if (isUsageLimitError(error.message)) return false;
 	if (
 		isUnexpectedSocketCloseMessage(msg) ||
 		/rate.?limit|too many requests|overloaded|service.?unavailable|internal_error|stream error.*received from peer|1302|timed?\s*out while waiting for the first event|timeout waiting for first/i.test(
@@ -1166,6 +1266,7 @@ export const streamAnthropic: StreamFunction<"anthropic-messages"> = (
 					onSseEvent: options?.onSseEvent,
 					fetch: options?.fetch,
 					requestMaxRetries: options?.requestMaxRetries,
+					maxRetryDelayMs: options?.maxRetryDelayMs,
 				});
 				client = created.client;
 				isOAuthToken = created.isOAuthToken;
@@ -1723,11 +1824,8 @@ export function buildAnthropicClientOptions(args: AnthropicClientOptionsArgs): A
 	const foundryCustomHeaders = resolveAnthropicCustomHeaders(model);
 	const tlsFetchOptions = buildClaudeCodeTlsFetchOptions(model, baseUrl);
 	const baseFetch = args.fetch ?? fetch;
-	const debugFetch = onSseEvent
-		? wrapFetchForSseDebug(baseFetch, event => onSseEvent(event, model))
-		: args.fetch
-			? baseFetch
-			: undefined;
+	const boundedFetch = wrapAnthropicFetchForBoundedRateLimits(baseFetch, args.maxRetryDelayMs);
+	const debugFetch = onSseEvent ? wrapFetchForSseDebug(boundedFetch, event => onSseEvent(event, model)) : boundedFetch;
 	if (model.provider === "github-copilot") {
 		const copilotApiKey = parseGitHubCopilotApiKey(apiKey).accessToken;
 		const betaFeatures = [...extraBetas];
@@ -1755,7 +1853,7 @@ export function buildAnthropicClientOptions(args: AnthropicClientOptionsArgs): A
 			dangerouslyAllowBrowser: true,
 			defaultHeaders,
 			logLevel: ANTHROPIC_SDK_LOG_LEVEL,
-			...(debugFetch ? { fetch: debugFetch } : {}),
+			fetch: debugFetch,
 			...(tlsFetchOptions ? { fetchOptions: tlsFetchOptions } : {}),
 		};
 	}
@@ -1789,7 +1887,7 @@ export function buildAnthropicClientOptions(args: AnthropicClientOptionsArgs): A
 			dangerouslyAllowBrowser: true,
 			defaultHeaders,
 			logLevel: ANTHROPIC_SDK_LOG_LEVEL,
-			...(debugFetch ? { fetch: debugFetch } : {}),
+			fetch: debugFetch,
 		};
 	}
 
@@ -1802,7 +1900,7 @@ export function buildAnthropicClientOptions(args: AnthropicClientOptionsArgs): A
 		dangerouslyAllowBrowser: true,
 		defaultHeaders,
 		logLevel: ANTHROPIC_SDK_LOG_LEVEL,
-		...(debugFetch ? { fetch: debugFetch } : {}),
+		fetch: debugFetch,
 		...(tlsFetchOptions ? { fetchOptions: tlsFetchOptions } : {}),
 	};
 }
@@ -2394,7 +2492,7 @@ export function convertAnthropicMessages(
 						type: "tool_use",
 						id: block.id,
 						name: isOAuthToken ? applyClaudeToolPrefix(block.name) : block.name,
-						input: block.arguments ?? {},
+						input: sanitizeJsonStrings(block.arguments ?? {}),
 					});
 				}
 			}

package/src/providers/google-shared.ts

@@ -18,7 +18,7 @@ import type {
 	Tool,
 	ToolCall,
 } from "../types";
-import { normalizeSystemPrompts } from "../utils";
+import { normalizeSystemPrompts, sanitizeJsonStrings } from "../utils";
 import { AssistantMessageEventStream } from "../utils/event-stream";
 import { finalizeErrorMessage, type RawHttpRequestDump, withHttpStatus } from "../utils/http-inspector";
 import { normalizeSchemaForCCA, normalizeSchemaForGoogle, toolWireSchema } from "../utils/schema";
@@ -237,7 +237,7 @@ export function convertMessages<T extends GoogleApiType>(model: Model<T>, contex
 					const part: Part = {
 						functionCall: {
 							name: block.name,
-							args: block.arguments ?? {},
+							args: sanitizeJsonStrings(block.arguments ?? {}) as Record<string, unknown>,
 							...(requiresToolCallId(model.id) ? { id: block.id } : {}),
 						},
 					};
@@ -467,7 +467,7 @@ export function pushToolCallEvents(
 	stream.push({
 		type: "toolcall_delta",
 		contentIndex,
-		delta: JSON.stringify(toolCall.arguments),
+		delta: JSON.stringify(sanitizeJsonStrings(toolCall.arguments)),
 		partial: output,
 	});
 	stream.push({ type: "toolcall_end", contentIndex, toolCall, partial: output });

package/src/providers/openai-chat-server.ts

@@ -19,6 +19,7 @@ import type {
 	ToolResultMessage,
 	TSchema,
 } from "../types";
+import { sanitizeJsonStrings } from "../utils";
 import {
 	type OpenAIChatContentPart,
 	type OpenAIChatMessage,
@@ -453,9 +454,9 @@ function isOnlyRaw(args: Record<string, unknown>): boolean {
 
 function stringifyArgs(args: Record<string, unknown>): string {
 	// `__raw` is our fallback marker for un-parseable inbound args; preserve it verbatim on the way out.
-	if (typeof args.__raw === "string" && isOnlyRaw(args)) return args.__raw;
+	if (typeof args.__raw === "string" && isOnlyRaw(args)) return args.__raw.toWellFormed();
 	try {
-		return JSON.stringify(args);
+		return JSON.stringify(sanitizeJsonStrings(args));
 	} catch {
 		return "{}";
 	}

package/src/providers/openai-codex/request-transformer.ts

@@ -1,6 +1,7 @@
 import type { Effort } from "../../model-thinking";
 import { requireSupportedEffort } from "../../model-thinking";
 import type { Api, Model } from "../../types";
+import { sanitizeJsonStrings } from "../../utils";
 
 export interface ReasoningConfig {
 	effort: "none" | "minimal" | "low" | "medium" | "high" | "xhigh" | "max";
@@ -109,6 +110,10 @@ function normalizeInputTextPartFields(input: InputItem[] | undefined): InputItem
 		}
 		if (normalizedItem.type === "message") {
 			normalizedItem.content = normalizeTextPartFields(normalizedItem.content, `input[${itemIndex}].content`);
+		} else if (normalizedItem.type === "function_call" && "arguments" in itemRecord) {
+			itemRecord.arguments = sanitizeJsonStrings(itemRecord.arguments);
+		} else if (normalizedItem.type === "custom_tool_call" && typeof itemRecord.input === "string") {
+			itemRecord.input = itemRecord.input.toWellFormed();
 		}
 		return normalizedItem;
 	});

package/src/providers/openai-responses-server.ts

@@ -22,6 +22,7 @@ import type {
 	Tool,
 	ToolCall,
 } from "../types";
+import { sanitizeJsonStrings } from "../utils";
 import {
 	type OpenAIResponsesFunctionCallItem,
 	type OpenAIResponsesFunctionCallOutputItem,
@@ -612,7 +613,8 @@ function buildOutputItems(message: AssistantMessage): OutputItem[] {
 		} else if (part.type === "toolCall") {
 			flushMessage();
 			if (part.customWireName) {
-				const rawInput = typeof part.arguments?.input === "string" ? (part.arguments.input as string) : "";
+				const rawInput =
+					typeof part.arguments?.input === "string" ? (part.arguments.input as string).toWellFormed() : "";
 				out.push({
 					type: "custom_tool_call",
 					id: part.thoughtSignature ?? makeCustomCallId(),
@@ -627,7 +629,7 @@ function buildOutputItems(message: AssistantMessage): OutputItem[] {
 					id: part.thoughtSignature ?? makeFuncCallId(),
 					call_id: part.id,
 					name: part.name,
-					arguments: JSON.stringify(part.arguments ?? {}),
+					arguments: JSON.stringify(sanitizeJsonStrings(part.arguments ?? {})),
 					status: "completed",
 				});
 			}
@@ -1085,7 +1087,8 @@ export function encodeStream(
 							} else {
 								// Standard JSON tool: arguments object on the skc side, the
 								// wire wants the JSON string the model emitted (= streamed deltas).
-								const argsJson = cur.argsText || JSON.stringify(tc.arguments ?? {});
+								const argsJson =
+									cur.argsText.toWellFormed() || JSON.stringify(sanitizeJsonStrings(tc.arguments ?? {}));
 								cur.argsText = argsJson;
 								emit("response.function_call_arguments.done", {
 									item_id: cur.itemId,

package/src/providers/openai-responses-shared.ts

@@ -28,7 +28,7 @@ import {
 	type ToolCall,
 	type ToolResultMessage,
 } from "../types";
-import { normalizeResponsesToolCallId } from "../utils";
+import { normalizeResponsesToolCallId, sanitizeJsonStrings } from "../utils";
 import type { AssistantMessageEventStream } from "../utils/event-stream";
 import { isCompleteJson, parseStreamingJson } from "../utils/json-parse";
 import { joinTextWithImagePlaceholder, NON_VISION_IMAGE_PLACEHOLDER, partitionVisionContent } from "./vision-guard";
@@ -275,7 +275,7 @@ export function convertResponsesAssistantMessage<TApi extends Api>(
 		}
 		knownCallIds.add(normalized.callId);
 		if (block.customWireName) {
-			const rawInput = typeof block.arguments?.input === "string" ? block.arguments.input : "";
+			const rawInput = typeof block.arguments?.input === "string" ? block.arguments.input.toWellFormed() : "";
 			customCallIds?.add(normalized.callId);
 			outputItems.push({
 				type: "custom_tool_call",
@@ -291,7 +291,7 @@ export function convertResponsesAssistantMessage<TApi extends Api>(
 			id: itemId,
 			call_id: normalized.callId,
 			name: block.name,
-			arguments: JSON.stringify(block.arguments),
+			arguments: JSON.stringify(sanitizeJsonStrings(block.arguments ?? {})),
 		});
 	}
 

package/src/rate-limit-utils.ts

@@ -36,6 +36,14 @@ export function parseRateLimitReason(errorMessage: string): RateLimitReason {
 		return "MODEL_CAPACITY_EXHAUSTED";
 	}
 
+	if (
+		lower.includes("out_of_credits") ||
+		lower.includes("request would exceed your account's rate limit") ||
+		lower.includes("request would exceed your accounts rate limit")
+	) {
+		return "QUOTA_EXHAUSTED";
+	}
+
 	if (
 		lower.includes("per minute") ||
 		lower.includes("rate limit") ||
@@ -86,8 +94,7 @@ export function calculateRateLimitBackoffMs(reason: RateLimitReason): number {
 
 /** Detect usage/quota limit errors in error messages (persistent, requires credential switch). */
 const USAGE_LIMIT_PATTERN =
-	/usage.?limit|usage_limit_reached|usage_not_included|limit_reached|model.?limit|model_limit_reached|message.?limit|message_limit_reached|limit for this model|quota.?exceeded|resource has been exhausted[^\n]*(?:quota|limit)/i;
-
+	/usage.?limit|usage_limit_reached|usage_not_included|limit_reached|model.?limit|model_limit_reached|message.?limit|message_limit_reached|limit for this model|quota.?exceeded|out_of_credits|request would exceed your account.?s rate limit|resource has been exhausted[^\n]*(?:quota|limit)/i;
 export function isUsageLimitError(errorMessage: string): boolean {
 	return USAGE_LIMIT_PATTERN.test(errorMessage);
 }

package/src/utils/oauth/deepinfra.ts

@@ -0,0 +1,15 @@
+/** DeepInfra login flow (API key paste against https://api.deepinfra.com/v1/openai). */
+import { createApiKeyLogin } from "./api-key-login";
+
+export const loginDeepInfra = createApiKeyLogin({
+	providerLabel: "DeepInfra",
+	authUrl: "https://deepinfra.com/dash/api_keys",
+	instructions: "Create or copy your DeepInfra API key from the DeepInfra dashboard",
+	promptMessage: "Paste your DeepInfra API key",
+	placeholder: "sk-...",
+	validation: {
+		kind: "models-endpoint",
+		provider: "DeepInfra",
+		modelsUrl: "https://api.deepinfra.com/v1/openai/models",
+	},
+});

package/src/utils.ts

@@ -11,6 +11,34 @@ export function normalizeSystemPrompts(systemPrompt: readonly string[] | string
 	return prompts.map(prompt => prompt.toWellFormed()).filter(prompt => prompt.length > 0);
 }
 
+export function sanitizeJsonStrings(value: unknown): unknown {
+	return sanitizeJsonStringsInner(value, new WeakMap<object, unknown>());
+}
+
+function sanitizeJsonStringsInner(value: unknown, seen: WeakMap<object, unknown>): unknown {
+	if (typeof value === "string") return value.toWellFormed();
+	if (!value || typeof value !== "object") return value;
+
+	const cached = seen.get(value);
+	if (cached !== undefined) return cached;
+
+	if (Array.isArray(value)) {
+		const sanitized: unknown[] = [];
+		seen.set(value, sanitized);
+		for (const item of value) {
+			sanitized.push(sanitizeJsonStringsInner(item, seen));
+		}
+		return sanitized;
+	}
+
+	const sanitized: Record<string, unknown> = {};
+	seen.set(value, sanitized);
+	for (const [key, nestedValue] of Object.entries(value)) {
+		sanitized[key.toWellFormed()] = sanitizeJsonStringsInner(nestedValue, seen);
+	}
+	return sanitized;
+}
+
 export function toNumber(value: unknown): number | undefined {
 	if (typeof value === "number" && Number.isFinite(value)) return value;
 	if (typeof value === "string" && value.trim()) {