evolcore 0.0.5 → 0.0.6

package/ecagent/src/runtime/utils/diagnostics.ts

@@ -0,0 +1,45 @@
+export interface DiagnosticErrorInfo {
+	name?: string;
+	message: string;
+	stack?: string;
+	code?: string | number;
+}
+
+export interface AssistantMessageDiagnostic {
+	type: string;
+	timestamp: number;
+	error?: DiagnosticErrorInfo;
+	details?: Record<string, unknown>;
+}
+
+export function formatThrownValue(value: unknown): string {
+	if (value instanceof Error) return value.message || value.name;
+	if (typeof value === "string") return value;
+	return String(value);
+}
+
+export function extractDiagnosticError(error: unknown): DiagnosticErrorInfo {
+	if (!(error instanceof Error)) return { name: "ThrownValue", message: formatThrownValue(error) };
+	const code = (error as Error & { code?: unknown }).code;
+	return {
+		name: error.name || undefined,
+		message: error.message || error.name,
+		stack: error.stack,
+		code: typeof code === "string" || typeof code === "number" ? code : undefined,
+	};
+}
+
+export function createAssistantMessageDiagnostic(
+	type: string,
+	error: unknown,
+	details?: Record<string, unknown>,
+): AssistantMessageDiagnostic {
+	return { type, timestamp: Date.now(), error: extractDiagnosticError(error), details };
+}
+
+export function appendAssistantMessageDiagnostic<T extends { diagnostics?: AssistantMessageDiagnostic[] }>(
+	message: T,
+	diagnostic: AssistantMessageDiagnostic,
+): void {
+	message.diagnostics = [...(message.diagnostics ?? []), diagnostic];
+}

package/ecagent/src/runtime/utils/event-stream.ts

@@ -0,0 +1,88 @@
+import type { AssistantMessage, AssistantMessageEvent } from "../types.ts";
+
+// Generic event stream class for async iteration
+export class EventStream<T, R = T> implements AsyncIterable<T> {
+	private queue: T[] = [];
+	private waiting: ((value: IteratorResult<T>) => void)[] = [];
+	private done = false;
+	private finalResultPromise: Promise<R>;
+	private resolveFinalResult!: (result: R) => void;
+	private isComplete: (event: T) => boolean;
+	private extractResult: (event: T) => R;
+
+	constructor(isComplete: (event: T) => boolean, extractResult: (event: T) => R) {
+		this.isComplete = isComplete;
+		this.extractResult = extractResult;
+		this.finalResultPromise = new Promise((resolve) => {
+			this.resolveFinalResult = resolve;
+		});
+	}
+
+	push(event: T): void {
+		if (this.done) return;
+
+		if (this.isComplete(event)) {
+			this.done = true;
+			this.resolveFinalResult(this.extractResult(event));
+		}
+
+		// Deliver to waiting consumer or queue it
+		const waiter = this.waiting.shift();
+		if (waiter) {
+			waiter({ value: event, done: false });
+		} else {
+			this.queue.push(event);
+		}
+	}
+
+	end(result?: R): void {
+		this.done = true;
+		if (result !== undefined) {
+			this.resolveFinalResult(result);
+		}
+		// Notify all waiting consumers that we're done
+		while (this.waiting.length > 0) {
+			const waiter = this.waiting.shift()!;
+			waiter({ value: undefined as any, done: true });
+		}
+	}
+
+	async *[Symbol.asyncIterator](): AsyncIterator<T> {
+		while (true) {
+			if (this.queue.length > 0) {
+				yield this.queue.shift()!;
+			} else if (this.done) {
+				return;
+			} else {
+				const result = await new Promise<IteratorResult<T>>((resolve) => this.waiting.push(resolve));
+				if (result.done) return;
+				yield result.value;
+			}
+		}
+	}
+
+	result(): Promise<R> {
+		return this.finalResultPromise;
+	}
+}
+
+export class AssistantMessageEventStream extends EventStream<AssistantMessageEvent, AssistantMessage> {
+	constructor() {
+		super(
+			(event) => event.type === "done" || event.type === "error",
+			(event) => {
+				if (event.type === "done") {
+					return event.message;
+				} else if (event.type === "error") {
+					return event.error;
+				}
+				throw new Error("Unexpected event type for final result");
+			},
+		);
+	}
+}
+
+/** Factory function for AssistantMessageEventStream (for use in extensions) */
+export function createAssistantMessageEventStream(): AssistantMessageEventStream {
+	return new AssistantMessageEventStream();
+}

package/ecagent/src/runtime/utils/headers.ts

@@ -0,0 +1,18 @@
+import type { ProviderHeaders } from "../types.ts";
+
+export function headersToRecord(headers: Headers): Record<string, string> {
+	const result: Record<string, string> = {};
+	for (const [key, value] of headers.entries()) {
+		result[key] = value;
+	}
+	return result;
+}
+
+export function providerHeadersToRecord(headers: ProviderHeaders | undefined): Record<string, string> | undefined {
+	if (!headers) return undefined;
+	const result: Record<string, string> = {};
+	for (const [key, value] of Object.entries(headers)) {
+		if (value !== null) result[key] = value;
+	}
+	return Object.keys(result).length > 0 ? result : undefined;
+}

package/ecagent/src/runtime/utils/json-parse.ts

@@ -0,0 +1,124 @@
+import { parse as partialParse } from "partial-json";
+
+const VALID_JSON_ESCAPES = new Set(['"', "\\", "/", "b", "f", "n", "r", "t", "u"]);
+
+function isControlCharacter(char: string): boolean {
+	const codePoint = char.codePointAt(0);
+	return codePoint !== undefined && codePoint >= 0x00 && codePoint <= 0x1f;
+}
+
+function escapeControlCharacter(char: string): string {
+	switch (char) {
+		case "\b":
+			return "\\b";
+		case "\f":
+			return "\\f";
+		case "\n":
+			return "\\n";
+		case "\r":
+			return "\\r";
+		case "\t":
+			return "\\t";
+		default:
+			return `\\u${char.codePointAt(0)?.toString(16).padStart(4, "0") ?? "0000"}`;
+	}
+}
+
+/**
+ * Repairs malformed JSON string literals by:
+ * - escaping raw control characters inside strings
+ * - doubling backslashes before invalid escape characters
+ */
+export function repairJson(json: string): string {
+	let repaired = "";
+	let inString = false;
+
+	for (let index = 0; index < json.length; index++) {
+		const char = json[index];
+
+		if (!inString) {
+			repaired += char;
+			if (char === '"') {
+				inString = true;
+			}
+			continue;
+		}
+
+		if (char === '"') {
+			repaired += char;
+			inString = false;
+			continue;
+		}
+
+		if (char === "\\") {
+			const nextChar = json[index + 1];
+			if (nextChar === undefined) {
+				repaired += "\\\\";
+				continue;
+			}
+
+			if (nextChar === "u") {
+				const unicodeDigits = json.slice(index + 2, index + 6);
+				if (/^[0-9a-fA-F]{4}$/.test(unicodeDigits)) {
+					repaired += `\\u${unicodeDigits}`;
+					index += 5;
+					continue;
+				}
+			}
+
+			if (VALID_JSON_ESCAPES.has(nextChar)) {
+				repaired += `\\${nextChar}`;
+				index += 1;
+				continue;
+			}
+
+			repaired += "\\\\";
+			continue;
+		}
+
+		repaired += isControlCharacter(char) ? escapeControlCharacter(char) : char;
+	}
+
+	return repaired;
+}
+
+export function parseJsonWithRepair<T>(json: string): T {
+	try {
+		return JSON.parse(json) as T;
+	} catch (error) {
+		const repairedJson = repairJson(json);
+		if (repairedJson !== json) {
+			return JSON.parse(repairedJson) as T;
+		}
+		throw error;
+	}
+}
+
+/**
+ * Attempts to parse potentially incomplete JSON during streaming.
+ * Always returns a valid object, even if the JSON is incomplete.
+ *
+ * @param partialJson The partial JSON string from streaming
+ * @returns Parsed object or empty object if parsing fails
+ */
+export function parseStreamingJson<T = Record<string, unknown>>(partialJson: string | undefined): T {
+	if (!partialJson || partialJson.trim() === "") {
+		return {} as T;
+	}
+
+	try {
+		return parseJsonWithRepair<T>(partialJson);
+	} catch {
+		try {
+			const result = partialParse(partialJson);
+			return (result ?? {}) as T;
+		} catch {
+			try {
+				const result = partialParse(repairJson(partialJson));
+				return (result ?? {}) as T;
+			} catch {
+				return {} as T;
+			}
+		}
+	}
+}

package/ecagent/src/runtime/utils/oauth/types.ts

@@ -0,0 +1,79 @@
+import type { Api, Model } from "../../types.ts";
+
+export type OAuthCredentials = {
+	refresh: string;
+	access: string;
+	expires: number;
+	[key: string]: unknown;
+};
+
+export type OAuthProviderId = string;
+
+/** @deprecated Use OAuthProviderId instead */
+export type OAuthProvider = OAuthProviderId;
+
+export type OAuthPrompt = {
+	message: string;
+	placeholder?: string;
+	allowEmpty?: boolean;
+};
+
+export type OAuthAuthInfo = {
+	url: string;
+	instructions?: string;
+};
+
+export type OAuthDeviceCodeInfo = {
+	userCode: string;
+	verificationUri: string;
+	intervalSeconds?: number;
+	expiresInSeconds?: number;
+};
+
+export type OAuthSelectOption = {
+	id: string;
+	label: string;
+};
+
+export type OAuthSelectPrompt = {
+	message: string;
+	options: OAuthSelectOption[];
+};
+
+export interface OAuthLoginCallbacks {
+	onAuth: (info: OAuthAuthInfo) => void;
+	onDeviceCode: (info: OAuthDeviceCodeInfo) => void;
+	onPrompt: (prompt: OAuthPrompt) => Promise<string>;
+	onProgress?: (message: string) => void;
+	onManualCodeInput?: () => Promise<string>;
+	/** Show an interactive selector and return the selected option id, or undefined on cancel. */
+	onSelect: (prompt: OAuthSelectPrompt) => Promise<string | undefined>;
+	signal?: AbortSignal;
+}
+
+export interface OAuthProviderInterface {
+	readonly id: OAuthProviderId;
+	readonly name: string;
+
+	/** Run the login flow, return credentials to persist */
+	login(callbacks: OAuthLoginCallbacks): Promise<OAuthCredentials>;
+
+	/** Whether login uses a local callback server and supports manual code input. */
+	usesCallbackServer?: boolean;
+
+	/** Refresh expired credentials, return updated credentials to persist */
+	refreshToken(credentials: OAuthCredentials): Promise<OAuthCredentials>;
+
+	/** Convert credentials to API key string for the provider */
+	getApiKey(credentials: OAuthCredentials): string;
+
+	/** Optional: modify models for this provider (e.g., update baseUrl) */
+	modifyModels?(models: Model<Api>[], credentials: OAuthCredentials): Model<Api>[];
+}
+
+/** @deprecated Use OAuthProviderInterface instead */
+export interface OAuthProviderInfo {
+	id: OAuthProviderId;
+	name: string;
+	available: boolean;
+}

package/ecagent/src/runtime/utils/overflow.ts

@@ -0,0 +1,165 @@
+import type { AssistantMessage } from "../types.ts";
+
+/**
+ * Regex patterns to detect context overflow errors from different providers.
+ *
+ * These patterns match error messages returned when the input exceeds
+ * the model's context window.
+ *
+ * Provider-specific patterns (with example error messages):
+ *
+ * - Anthropic: "prompt is too long: 213462 tokens > 200000 maximum"
+ * - Anthropic: "413 {\"error\":{\"type\":\"request_too_large\",\"message\":\"Request exceeds the maximum size\"}}"
+ * - OpenAI: "Your input exceeds the context window of this model"
+ * - OpenAI/LiteLLM: "Requested token count exceeds the model's maximum context length of 131072 tokens"
+ * - OpenAI-compatible: "Input length (265330) exceeds model's maximum context length (262144)."
+ * - Google: "The input token count (1196265) exceeds the maximum number of tokens allowed (1048575)"
+ * - xAI: "This model's maximum prompt length is 131072 but the request contains 537812 tokens"
+ * - Groq: "Please reduce the length of the messages or completion"
+ * - OpenRouter: "This endpoint's maximum context length is X tokens. However, you requested about Y tokens"
+ * - OpenRouter/Poolside: "Input length X exceeds the maximum allowed input length of Y tokens."
+ * - Together AI: "The input (X tokens) is longer than the model's context length (Y tokens)."
+ * - llama.cpp: "the request exceeds the available context size, try increasing it"
+ * - LM Studio: "tokens to keep from the initial prompt is greater than the context length"
+ * - GitHub Copilot: "prompt token count of X exceeds the limit of Y"
+ * - MiniMax: "invalid params, context window exceeds limit"
+ * - Kimi For Coding: "Your request exceeded model token limit: X (requested: Y)"
+ * - DS4: "Prompt has X tokens, but the configured context size is Y tokens"
+ * - Cerebras: "400/413 status code (no body)"
+ * - Mistral: "Prompt contains X tokens ... too large for model with Y maximum context length"
+ * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow
+ * - Xiaomi MiMo: Truncates input to fill contextWindow exactly, then returns finish_reason "length"
+ *   with output=0 (no room left to generate). Detected via stopReason "length" + zero output +
+ *   input filling the context window.
+ * - Ollama: Some deployments truncate silently, others return errors like "prompt too long; exceeded max context length by X tokens"
+ */
+const OVERFLOW_PATTERNS = [
+	/prompt is too long/i, // Anthropic token overflow
+	/request_too_large/i, // Anthropic request byte-size overflow (HTTP 413)
+	/input is too long for requested model/i, // Amazon Bedrock
+	/exceeds the context window/i, // OpenAI (Completions & Responses API)
+	/exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\d,]+ tokens?|\s*\([\d,]+\))/i, // OpenAI-compatible proxies (LiteLLM)
+	/input token count.*exceeds the maximum/i, // Google (Gemini)
+	/maximum prompt length is \d+/i, // xAI (Grok)
+	/reduce the length of the messages/i, // Groq
+	/maximum context length is \d+ tokens/i, // OpenRouter (most backends)
+	/exceeds (?:the )?maximum allowed input length of [\d,]+ tokens?/i, // OpenRouter/Poolside
+	/input \(\d+ tokens\) is longer than the model'?s context length \(\d+ tokens\)/i, // Together AI
+	/exceeds the limit of \d+/i, // GitHub Copilot
+	/exceeds the available context size/i, // llama.cpp server
+	/greater than the context length/i, // LM Studio
+	/context window exceeds limit/i, // MiniMax
+	/exceeded model token limit/i, // Kimi For Coding
+	/too large for model with \d+ maximum context length/i, // Mistral
+	/prompt has [\d,]+ tokens?, but the configured context size is [\d,]+ tokens?/i, // DS4 server
+	/model_context_window_exceeded/i, // z.ai non-standard finish_reason surfaced as error text
+	/prompt too long; exceeded (?:max )?context length/i, // Ollama explicit overflow error
+	/context[_ ]length[_ ]exceeded/i, // Generic fallback
+	/too many tokens/i, // Generic fallback
+	/token limit exceeded/i, // Generic fallback
+	/^4(?:00|13)\s*(?:status code)?\s*\(no body\)/i, // Cerebras: 400/413 with no body
+];
+
+/**
+ * Patterns that indicate non-overflow errors (e.g. rate limiting, server errors).
+ * Error messages matching any of these are excluded from overflow detection
+ * even if they also match an OVERFLOW_PATTERN.
+ *
+ * Example: Bedrock formats throttling errors as "ThrottlingException: Too many tokens,
+ * please wait before trying again." which would match the /too many tokens/i overflow
+ * pattern without this exclusion.
+ */
+const NON_OVERFLOW_PATTERNS = [
+	/^(Throttling error|Service unavailable):/i, // AWS Bedrock non-overflow errors (human-readable prefixes from formatBedrockError)
+	/rate limit/i, // Generic rate limiting
+	/too many requests/i, // Generic HTTP 429 style
+];
+
+/**
+ * Check if an assistant message represents a context overflow error.
+ *
+ * This handles two cases:
+ * 1. Error-based overflow: Most providers return stopReason "error" with a
+ *    specific error message pattern.
+ * 2. Silent overflow: Some providers accept overflow requests and return
+ *    successfully. For these, we check if usage.input exceeds the context window.
+ *
+ * ## Reliability by Provider
+ *
+ * **Reliable detection (returns error with detectable message):**
+ * - Anthropic: "prompt is too long: X tokens > Y maximum" or "request_too_large"
+ * - OpenAI (Completions & Responses): "exceeds the context window", "exceeds the model's maximum context length of X tokens", or "exceeds model's maximum context length (X)"
+ * - Google Gemini: "input token count exceeds the maximum"
+ * - xAI (Grok): "maximum prompt length is X but request contains Y"
+ * - Groq: "reduce the length of the messages"
+ * - Cerebras: 400/413 status code (no body)
+ * - Mistral: "Prompt contains X tokens ... too large for model with Y maximum context length"
+ * - OpenRouter (most backends): "maximum context length is X tokens"
+ * - OpenRouter/Poolside: "Input length X exceeds the maximum allowed input length of Y tokens."
+ * - Together AI: "The input (X tokens) is longer than the model's context length (Y tokens)."
+ * - llama.cpp: "exceeds the available context size"
+ * - LM Studio: "greater than the context length"
+ * - Kimi For Coding: "exceeded model token limit: X (requested: Y)"
+ * - DS4: "Prompt has X tokens, but the configured context size is Y tokens"
+ *
+ * **Unreliable detection:**
+ * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),
+ *   sometimes returns rate limit errors. Pass contextWindow param to detect silent overflow.
+ * - Xiaomi MiMo: Truncates input to fit contextWindow then returns stopReason "length" with
+ *   output=0. Pass contextWindow param to detect via the "filled context + zero output" signal.
+ * - Ollama: May truncate input silently for some setups, but may also return explicit
+ *   overflow errors that match the patterns above. Silent truncation still cannot be
+ *   detected here because we do not know the expected token count.
+ *
+ * ## Custom Providers
+ *
+ * If you've added custom models via settings.json, this function may not detect
+ * overflow errors from those providers. To add support:
+ *
+ * 1. Send a request that exceeds the model's context window
+ * 2. Check the errorMessage in the response
+ * 3. Create a regex pattern that matches the error
+ * 4. The pattern should be added to OVERFLOW_PATTERNS in this file, or
+ *    check the errorMessage yourself before calling this function
+ *
+ * @param message - The assistant message to check
+ * @param contextWindow - Optional context window size for detecting silent overflow (z.ai)
+ * @returns true if the message indicates a context overflow
+ */
+export function isContextOverflow(message: AssistantMessage, contextWindow?: number): boolean {
+	// Case 1: Check error message patterns
+	if (message.stopReason === "error" && message.errorMessage) {
+		// Skip messages matching known non-overflow patterns (e.g. throttling / rate-limit)
+		const isNonOverflow = NON_OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!));
+		if (!isNonOverflow && OVERFLOW_PATTERNS.some((p) => p.test(message.errorMessage!))) {
+			return true;
+		}
+	}
+
+	// Case 2: Silent overflow (z.ai style) - successful but usage exceeds context
+	if (contextWindow && message.stopReason === "stop") {
+		const inputTokens = message.usage.input + message.usage.cacheRead;
+		if (inputTokens > contextWindow) {
+			return true;
+		}
+	}
+
+	// Case 3: Length-stop overflow (Xiaomi MiMo style) - server truncates oversized input
+	// to fit the context window, leaving no room for output. Returns stopReason "length"
+	// with output=0 and input+cacheRead filling the context window.
+	if (contextWindow && message.stopReason === "length" && message.usage.output === 0) {
+		const inputTokens = message.usage.input + message.usage.cacheRead;
+		if (inputTokens >= contextWindow * 0.99) {
+			return true;
+		}
+	}
+
+	return false;
+}
+
+/**
+ * Get the overflow patterns for testing purposes.
+ */
+export function getOverflowPatterns(): RegExp[] {
+	return [...OVERFLOW_PATTERNS];
+}

package/ecagent/src/runtime/utils/provider-env.ts

@@ -0,0 +1,52 @@
+import type { ProviderEnv } from "../types.ts";
+
+let procEnvCache: Map<string, string> | null = null;
+
+/**
+ * Fallback for https://github.com/oven-sh/bun/issues/27802.
+ * Bun compiled binaries can expose an empty process.env inside Linux sandboxes
+ * even though /proc/self/environ contains the environment.
+ *
+ * This intentionally duplicates restoreSandboxEnv() in
+ * packages/coding-agent/src/bun/restore-sandbox-env.ts. The ai package can be
+ * used directly, without going through that entrypoint, so provider env lookup
+ * must not depend on process.env having been patched.
+ */
+function getBunSandboxEnvValue(name: string): string | undefined {
+	if (typeof process === "undefined" || !process.versions?.bun || Object.keys(process.env).length > 0) {
+		return undefined;
+	}
+
+	if (procEnvCache === null) {
+		procEnvCache = new Map();
+		try {
+			const { readFileSync } = require("node:fs") as {
+				readFileSync(path: string, encoding: BufferEncoding): string;
+			};
+			const data = readFileSync("/proc/self/environ", "utf-8");
+			for (const entry of data.split("\0")) {
+				const idx = entry.indexOf("=");
+				if (idx > 0) {
+					procEnvCache.set(entry.slice(0, idx), entry.slice(idx + 1));
+				}
+			}
+		} catch {
+			// /proc/self/environ may not exist or may not be readable.
+		}
+	}
+
+	return procEnvCache.get(name);
+}
+
+/**
+ * Resolve a provider env value from scoped overrides, normal process.env, then
+ * the duplicated Bun sandbox fallback for direct pi-ai consumers.
+ */
+export function getProviderEnvValue(name: string, env?: ProviderEnv): string | undefined {
+	return (
+		env?.[name] ||
+		(typeof process !== "undefined" ? process.env[name] : undefined) ||
+		getBunSandboxEnvValue(name) ||
+		undefined
+	);
+}

package/ecagent/src/runtime/utils/retry.ts

@@ -0,0 +1,101 @@
+import type { AssistantMessage } from "../types.ts";
+
+function buildProviderErrorPattern(patterns: readonly string[]): RegExp {
+	return new RegExp(patterns.join("|"), "i");
+}
+
+const NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN = buildProviderErrorPattern([
+	// OpenCode Go/free-tier limits returned as 429 JSON error types by OpenCode's
+	// Zen API. These are subscription/account limits, not transient throttles.
+	"GoUsageLimitError",
+	"FreeUsageLimitError",
+
+	// OpenCode Go subscription-limit text asks users to enable available-balance
+	// usage after rolling/weekly/monthly limits are reached.
+	"Monthly usage limit reached",
+	"available balance",
+
+	// Generic quota/budget/billing exhaustion. `insufficient_quota` is OpenAI's
+	// quota/billing error code; the other strings cover common gateway wording.
+	"insufficient_quota",
+	"out of budget",
+	"quota exceeded",
+	"billing",
+]);
+
+const RETRYABLE_PROVIDER_ERROR_PATTERN = buildProviderErrorPattern([
+	// Generic provider load, HTTP status, and server-side transient failures.
+	"overloaded",
+	"rate.?limit",
+	"too many requests",
+	"429",
+	"500",
+	"502",
+	"503",
+	"504",
+	"524",
+	"service.?unavailable",
+	"server.?error",
+	"internal.?error",
+
+	// Wrapper/provider text for transient upstream failures, including OpenRouter
+	// "Provider returned error" responses (#2264).
+	"provider.?returned.?error",
+
+	// Network, proxy, and fetch transport failures. This includes OpenAI Codex
+	// raw-fetch failures such as "upstream connect", "connection refused", and
+	// "reset before headers" (#733), plus OpenRouter connection drops (#3317).
+	"network.?error",
+	"connection.?error",
+	"connection.?refused",
+	"connection.?lost",
+	"other side closed",
+	"fetch failed",
+	"upstream.?connect",
+	"reset before headers",
+	"socket hang up",
+	"socket connection was closed",
+	"timed? out",
+	"timeout",
+	"terminated",
+
+	// WebSocket transports can report close/error text instead of HTTP/fetch text.
+	"websocket.?closed",
+	"websocket.?error",
+
+	// Premature stream endings from SDKs and transports. Anthropic can throw
+	// "stream ended without ..." and "Anthropic stream ended before message_stop"
+	// (#4433); Bedrock/Smithy can throw an HTTP/2 no-response error (#3594).
+	"ended without",
+	"stream ended before message_stop",
+	"http2 request did not get a response",
+
+	// Provider-requested retry delay cap failures should flow through the outer
+	// retry policy so callers can surface/abort the backoff (#1123).
+	"retry delay",
+
+	// Explicit retry guidance emitted mid-stream by OpenAI Responses and Bedrock
+	// stream exceptions (#6019).
+	"you can retry your request",
+	"try your request again",
+	"please retry your request",
+
+	// gRPC based providers (e.g. NVIDIA NIM)
+	"ResourceExhausted",
+]);
+
+/**
+ * Classifies whether a failed assistant message looks like a transient provider
+ * or transport error, so callers can decide if the last assistant turn should be
+ * restarted.
+ *
+ * This does not implement retry policy. Callers should first handle context
+ * overflow separately, then apply their own retry budget, backoff, and reporting
+ * before restarting the assistant turn.
+ */
+export function isRetryableAssistantError(message: AssistantMessage): boolean {
+	if (message.stopReason !== "error" || !message.errorMessage) return false;
+	const errorMessage = message.errorMessage;
+	if (NON_RETRYABLE_PROVIDER_LIMIT_ERROR_PATTERN.test(errorMessage)) return false;
+	return RETRYABLE_PROVIDER_ERROR_PATTERN.test(errorMessage);
+}