@prestyj/ai 4.6.4 → 4.11.4

package/dist/index.d.cts

@@ -255,8 +255,9 @@ declare class StreamResult implements AsyncIterable<StreamEvent> {
     private resolveResponse;
     private rejectResponse;
     private resolveWait;
-    constructor(generator: AsyncGenerator<StreamEvent, StreamResponse>);
+    constructor(generator: AsyncGenerator<StreamEvent, StreamResponse>, signal?: AbortSignal);
     private pump;
+    private _nextWithAbort;
     [Symbol.asyncIterator](): AsyncIterator<StreamEvent>;
     then<TResult1 = StreamResponse, TResult2 = never>(onfulfilled?: ((value: StreamResponse) => TResult1 | PromiseLike<TResult1>) | null, onrejected?: ((reason: unknown) => TResult2 | PromiseLike<TResult2>) | null): Promise<TResult1 | TResult2>;
 }
@@ -414,6 +415,19 @@ declare function formatError(err: unknown): FormattedError;
  */
 declare function formatErrorForDisplay(err: unknown): string;
 
+/**
+ * Inspect a raw provider error message and tag it with a clearer, actionable
+ * prefix so a worker orchestrator can route on intent instead of regexing JSON.
+ * Preserves the original message verbatim after the prefix — helpful for
+ * debugging.
+ *
+ * Order matters: context-overflow is checked first because some providers wrap
+ * overflow errors in HTTP 429 envelopes; we want the structural meaning, not
+ * the transport status. Billing comes before auth/rate-limit because "402
+ * Payment Required" must not be mis-routed as a rate-limit retry.
+ */
+declare function classifyProviderError(message: string): string;
+
 /**
  * Provider-level diagnostic hook. Mirrors the pattern used by gg-agent's
  * setStreamDiagnostic — the host app wires a callback (typically writing to
@@ -506,4 +520,4 @@ interface PalsuProviderConfig {
  */
 declare function registerPalsuProvider(config?: PalsuProviderConfig): PalsuProviderHandle;
 
-export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, EZCoderAIError, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, type ImageContent, type Message, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, type RawContent, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, formatError, formatErrorForDisplay, isHardBillingMessage, isUsageLimitError, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, providerRegistry, registerPalsuProvider, setProviderDiagnostic, stream, toAnthropicMessages, toOpenAIMessages };
+export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, EZCoderAIError, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, type ImageContent, type Message, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, type RawContent, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, classifyProviderError, formatError, formatErrorForDisplay, isHardBillingMessage, isUsageLimitError, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, providerRegistry, registerPalsuProvider, setProviderDiagnostic, stream, toAnthropicMessages, toOpenAIMessages };

package/dist/index.d.ts

@@ -255,8 +255,9 @@ declare class StreamResult implements AsyncIterable<StreamEvent> {
     private resolveResponse;
     private rejectResponse;
     private resolveWait;
-    constructor(generator: AsyncGenerator<StreamEvent, StreamResponse>);
+    constructor(generator: AsyncGenerator<StreamEvent, StreamResponse>, signal?: AbortSignal);
     private pump;
+    private _nextWithAbort;
     [Symbol.asyncIterator](): AsyncIterator<StreamEvent>;
     then<TResult1 = StreamResponse, TResult2 = never>(onfulfilled?: ((value: StreamResponse) => TResult1 | PromiseLike<TResult1>) | null, onrejected?: ((reason: unknown) => TResult2 | PromiseLike<TResult2>) | null): Promise<TResult1 | TResult2>;
 }
@@ -414,6 +415,19 @@ declare function formatError(err: unknown): FormattedError;
  */
 declare function formatErrorForDisplay(err: unknown): string;
 
+/**
+ * Inspect a raw provider error message and tag it with a clearer, actionable
+ * prefix so a worker orchestrator can route on intent instead of regexing JSON.
+ * Preserves the original message verbatim after the prefix — helpful for
+ * debugging.
+ *
+ * Order matters: context-overflow is checked first because some providers wrap
+ * overflow errors in HTTP 429 envelopes; we want the structural meaning, not
+ * the transport status. Billing comes before auth/rate-limit because "402
+ * Payment Required" must not be mis-routed as a rate-limit retry.
+ */
+declare function classifyProviderError(message: string): string;
+
 /**
  * Provider-level diagnostic hook. Mirrors the pattern used by gg-agent's
  * setStreamDiagnostic — the host app wires a callback (typically writing to
@@ -506,4 +520,4 @@ interface PalsuProviderConfig {
  */
 declare function registerPalsuProvider(config?: PalsuProviderConfig): PalsuProviderHandle;
 
-export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, EZCoderAIError, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, type ImageContent, type Message, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, type RawContent, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, formatError, formatErrorForDisplay, isHardBillingMessage, isUsageLimitError, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, providerRegistry, registerPalsuProvider, setProviderDiagnostic, stream, toAnthropicMessages, toOpenAIMessages };
+export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, EZCoderAIError, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, type ImageContent, type Message, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, type RawContent, type ServerToolCall, type ServerToolCallEvent, type ServerToolDefinition, type ServerToolResult, type ServerToolResultEvent, type StopReason, type StreamEvent, type StreamOptions, type StreamResponse, StreamResult, type SystemMessage, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, classifyProviderError, formatError, formatErrorForDisplay, isHardBillingMessage, isUsageLimitError, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, providerRegistry, registerPalsuProvider, setProviderDiagnostic, stream, toAnthropicMessages, toOpenAIMessages };

package/dist/index.js

@@ -84,10 +84,25 @@ function formatResetTime(resetsAt) {
     minute: "2-digit"
   });
 }
+function isMythosAccessError(message) {
+  const lower = message.toLowerCase();
+  return lower.includes("mythos") && (lower.includes("not_found") || lower.includes("not found") || lower.includes("no access"));
+}
 function formatError(err) {
   if (err instanceof ProviderError) {
     const name = providerDisplayName(err.provider);
     const cleanMessage = cleanProviderMessage(err.message);
+    if (isMythosAccessError(cleanMessage)) {
+      return {
+        headline: "Claude Mythos 5 is invitation-only.",
+        source: "provider",
+        message: "Your Anthropic account isn't approved for Project Glasswing, so the API reports the model as not found.",
+        provider: err.provider,
+        statusCode: err.statusCode,
+        ...err.requestId ? { requestId: err.requestId } : {},
+        guidance: "Request access via your Anthropic account team (see platform.claude.com/docs/en/about-claude/models/overview), or switch to claude-fable-5 with /model \u2014 same underlying model, generally available."
+      };
+    }
     if (isUsageLimitError(err)) {
       const resetClause = err.resetsAt ? ` It resets at ${formatResetTime(err.resetsAt)}.` : "";
       return {
@@ -274,21 +289,21 @@ var StreamResult = class {
   resolveResponse;
   rejectResponse;
   resolveWait = null;
-  constructor(generator) {
+  constructor(generator, signal) {
     this.response = new Promise((resolve, reject) => {
       this.resolveResponse = resolve;
       this.rejectResponse = reject;
     });
-    this.pump(generator);
+    this.pump(generator, signal);
   }
-  async pump(generator) {
+  async pump(generator, signal) {
     try {
-      let next = await generator.next();
+      let next = await this._nextWithAbort(generator, signal);
       while (!next.done) {
         this.buffer.push(next.value);
         this.resolveWait?.();
         this.resolveWait = null;
-        next = await generator.next();
+        next = await this._nextWithAbort(generator, signal);
       }
       this.done = true;
       this.resolveResponse(next.value);
@@ -303,6 +318,28 @@ var StreamResult = class {
       this.resolveWait = null;
     }
   }
+  async _nextWithAbort(generator, signal) {
+    if (!signal) {
+      return generator.next();
+    }
+    if (signal.aborted) {
+      return Promise.reject(new DOMException("Aborted", "AbortError"));
+    }
+    let onAbort;
+    const abortPromise = new Promise((_, reject) => {
+      onAbort = () => {
+        generator.return?.(void 0).catch(() => {
+        });
+        reject(new DOMException("Aborted", "AbortError"));
+      };
+      signal.addEventListener("abort", onAbort, { once: true });
+    });
+    try {
+      return await Promise.race([generator.next(), abortPromise]);
+    } finally {
+      if (onAbort) signal.removeEventListener("abort", onAbort);
+    }
+  }
   async *[Symbol.asyncIterator]() {
     let index = 0;
     while (true) {
@@ -399,6 +436,29 @@ function isRawThinking(part) {
   const t = part.data.type;
   return t === "thinking" || t === "redacted_thinking";
 }
+var ANTHROPIC_INPUT_BLOCK_TYPES = /* @__PURE__ */ new Set([
+  "bash_code_execution_tool_result",
+  "code_execution_tool_result",
+  "connector_text",
+  "container_upload",
+  "document",
+  "image",
+  "mid_conv_system",
+  "redacted_thinking",
+  "search_result",
+  "server_tool_use",
+  "text",
+  "text_editor_code_execution_tool_result",
+  "thinking",
+  "tool_result",
+  "tool_search_tool_result",
+  "tool_use",
+  "web_fetch_tool_result",
+  "web_search_tool_result"
+]);
+function isAnthropicCompatibleRaw(part) {
+  return ANTHROPIC_INPUT_BLOCK_TYPES.has(part.data.type);
+}
 function isPositionSensitiveThinking(part) {
   if (part.type === "thinking") return hasValidThinkingSignature(part);
   return isRawThinking(part);
@@ -425,7 +485,8 @@ function toAnthropicAssistantPart(part, idMap) {
     };
   if (part.type === "server_tool_result")
     return part.data;
-  if (part.type === "raw") return part.data;
+  if (part.type === "raw")
+    return isAnthropicCompatibleRaw(part) ? part.data : null;
   return null;
 }
 function toAnthropicAssistantContent(content, preserveThinking, idMap) {
@@ -684,12 +745,12 @@ function toAnthropicToolChoice(choice) {
   return { type: "tool", name: choice.name };
 }
 function isAdaptiveThinkingModel(model) {
-  return /fable-5|opus-4[-.]8|opus-4[-.]7|opus-4[-.]6|sonnet-4[-.]6/.test(model);
+  return /opus-4[-.]8|opus-4[-.]7|opus-4[-.]6|sonnet-4[-.]6|fable-5|mythos-5/.test(model);
 }
 function toAnthropicThinking(level, maxTokens, model) {
   if (isAdaptiveThinkingModel(model)) {
     let effort = level;
-    if (effort === "xhigh" && !/opus-4-8|opus-4-7|fable-5/.test(model)) {
+    if (effort === "xhigh" && !/opus-4-8|opus-4-7/.test(model)) {
       effort = "high";
     }
     return {
@@ -698,16 +759,17 @@ function toAnthropicThinking(level, maxTokens, model) {
       outputConfig: { effort }
     };
   }
+  const VISIBLE_FLOOR = 1024;
   const effectiveLevel = level === "xhigh" || level === "max" ? "high" : level;
   const budgetMap = {
-    low: Math.max(1024, Math.floor(maxTokens * 0.25)),
-    medium: Math.max(2048, Math.floor(maxTokens * 0.5)),
-    high: Math.max(4096, maxTokens)
+    low: Math.max(1024, Math.floor(maxTokens * 0.2)),
+    medium: Math.max(2048, Math.floor(maxTokens * 0.45)),
+    high: Math.max(4096, Math.floor(maxTokens * 0.8))
   };
-  const budget = budgetMap[effectiveLevel];
+  const budget = Math.max(0, Math.min(budgetMap[effectiveLevel], maxTokens - VISIBLE_FLOOR));
   return {
     thinking: { type: "enabled", budget_tokens: budget },
-    maxTokens: maxTokens + budget
+    maxTokens
   };
 }
 function remapToolCallId(id, idMap) {
@@ -919,10 +981,10 @@ function createClient(options) {
     ...isOAuth ? { apiKey: null, authToken: options.apiKey } : { apiKey: options.apiKey },
     ...options.baseUrl ? { baseURL: options.baseUrl } : {},
     ...options.fetch ? { fetch: options.fetch } : {},
-    // Allow SDK retries for connection-level failures (socket hang up, 500s,
-    // connection refused).  Our stall detection handles abort-initiated retries
-    // separately — SDK retries only fire on genuine transport errors.
-    maxRetries: 2,
+    // Disable SDK retries — the agent loop has its own stall/overload retry
+    // logic that surfaces errors properly. SDK retries on 429s can cause
+    // multi-minute hangs when the provider stops responding mid-retry.
+    maxRetries: 0,
     ...isOAuth ? {
       defaultHeaders: {
         // Anthropic's OAuth edge validates the claude-cli version. Callers
@@ -935,7 +997,7 @@ function createClient(options) {
   });
 }
 function streamAnthropic(options) {
-  return new StreamResult(runStream(options));
+  return new StreamResult(runStream(options), options.signal);
 }
 async function* runStream(options) {
   const client = createClient(options);
@@ -1030,7 +1092,6 @@ async function* runStream(options) {
       throw toError(err);
     }
   }
-  const stream2 = client.messages.stream(params, requestOptions);
   const contentParts = [];
   const blocks = /* @__PURE__ */ new Map();
   let inputTokens = 0;
@@ -1039,8 +1100,14 @@ async function* runStream(options) {
   let cacheWrite;
   let stopReason = null;
   const keepalive = { type: "keepalive" };
+  let receivedAnyEvent = false;
   try {
+    const stream2 = await client.messages.create(
+      params,
+      requestOptions
+    );
     for await (const event of stream2) {
+      receivedAnyEvent = true;
       switch (event.type) {
         case "message_start": {
           const usage = event.message.usage;
@@ -1077,7 +1144,7 @@ async function* runStream(options) {
             accum.toolId = block.id;
             accum.toolName = block.name;
             accum.input = block.input;
-          } else if (block.type === "redacted_thinking") {
+          } else if (block.type !== "text" && block.type !== "thinking") {
             accum.raw = block;
           }
           blocks.set(idx, accum);
@@ -1174,8 +1241,7 @@ async function* runStream(options) {
             contentParts.push({ type: "raw", data: accum.raw });
             yield keepalive;
           } else {
-            const msg = stream2.currentMessage;
-            const rawBlock = msg?.content[event.index];
+            const rawBlock = accum.raw;
             if (rawBlock) {
               const blockType = rawBlock.type;
               if (blockType === "web_search_tool_result") {
@@ -1221,6 +1287,11 @@ async function* runStream(options) {
   } catch (err) {
     throw toError(err);
   }
+  if (!receivedAnyEvent) {
+    throw new ProviderError("anthropic", "Stream ended without producing any events.", {
+      statusCode: 504
+    });
+  }
   const normalizedStop = normalizeAnthropicStopReason(stopReason);
   const response = {
     message: {
@@ -1512,7 +1583,7 @@ function createClient2(options) {
   });
 }
 function streamOpenAI(options) {
-  return new StreamResult(runStream2(options));
+  return new StreamResult(runStream2(options), options.signal);
 }
 async function* runStream2(options) {
   const providerName = options.provider ?? "openai";
@@ -1604,51 +1675,62 @@ async function* runStream2(options) {
   let outputTokens = 0;
   let cacheRead = 0;
   let finishReason = null;
-  for await (const chunk of stream2) {
-    const choice = chunk.choices?.[0];
-    if (chunk.usage) {
-      ({ inputTokens, outputTokens, cacheRead } = extractOpenAIUsage(chunk.usage));
-    }
-    if (!choice) continue;
-    if (choice.finish_reason) {
-      finishReason = choice.finish_reason;
-    }
-    const delta = choice.delta;
-    const reasoningContent = delta.reasoning_content;
-    if (typeof reasoningContent === "string" && reasoningContent) {
-      thinkingAccum += reasoningContent;
-      if (options.thinking) {
-        yield { type: "thinking_delta", text: reasoningContent };
+  let receivedAnyChunk = false;
+  try {
+    for await (const chunk of stream2) {
+      receivedAnyChunk = true;
+      const choice = chunk.choices?.[0];
+      if (chunk.usage) {
+        ({ inputTokens, outputTokens, cacheRead } = extractOpenAIUsage(chunk.usage));
       }
-    }
-    if (delta.content) {
-      textAccum += delta.content;
-      yield { type: "text_delta", text: delta.content };
-    }
-    if (delta.tool_calls) {
-      for (const tc of delta.tool_calls) {
-        let accum = toolCallAccum.get(tc.index);
-        if (!accum) {
-          accum = {
-            id: tc.id ?? "",
-            name: tc.function?.name ?? "",
-            argsJson: ""
-          };
-          toolCallAccum.set(tc.index, accum);
+      if (!choice) continue;
+      if (choice.finish_reason) {
+        finishReason = choice.finish_reason;
+      }
+      const delta = choice.delta;
+      const reasoningContent = delta.reasoning_content;
+      if (typeof reasoningContent === "string" && reasoningContent) {
+        thinkingAccum += reasoningContent;
+        if (options.thinking) {
+          yield { type: "thinking_delta", text: reasoningContent };
         }
-        if (tc.id) accum.id = tc.id;
-        if (tc.function?.name) accum.name = tc.function.name;
-        if (tc.function?.arguments) {
-          accum.argsJson += tc.function.arguments;
-          yield {
-            type: "toolcall_delta",
-            id: accum.id,
-            name: accum.name,
-            argsJson: tc.function.arguments
-          };
+      }
+      if (delta.content) {
+        textAccum += delta.content;
+        yield { type: "text_delta", text: delta.content };
+      }
+      if (delta.tool_calls) {
+        for (const tc of delta.tool_calls) {
+          let accum = toolCallAccum.get(tc.index);
+          if (!accum) {
+            accum = {
+              id: tc.id ?? "",
+              name: tc.function?.name ?? "",
+              argsJson: ""
+            };
+            toolCallAccum.set(tc.index, accum);
+          }
+          if (tc.id) accum.id = tc.id;
+          if (tc.function?.name) accum.name = tc.function.name;
+          if (tc.function?.arguments) {
+            accum.argsJson += tc.function.arguments;
+            yield {
+              type: "toolcall_delta",
+              id: accum.id,
+              name: accum.name,
+              argsJson: tc.function.arguments
+            };
+          }
         }
       }
     }
+  } catch (err) {
+    throw toError2(err, providerName);
+  }
+  if (!receivedAnyChunk) {
+    throw new ProviderError(providerName, "Stream ended without producing any chunks.", {
+      statusCode: 504
+    });
   }
   if (thinkingAccum) {
     contentParts.push({ type: "thinking", text: thinkingAccum });
@@ -1891,7 +1973,7 @@ function isVisibleOutputItem(itemType) {
   return itemType === "message";
 }
 function streamOpenAICodex(options) {
-  return new StreamResult(runStream3(options));
+  return new StreamResult(runStream3(options), options.signal);
 }
 async function* runStream3(options) {
   const baseUrl = (options.baseUrl || DEFAULT_BASE_URL).replace(/\/+$/, "");
@@ -2221,10 +2303,16 @@ async function* parseSSE(body) {
   }
 }
 function remapCodexId(id, idMap) {
-  if (id.startsWith("fc_") || id.startsWith("fc-")) return id;
   const existing = idMap.get(id);
   if (existing) return existing;
-  const mapped = `fc_${id.replace(/^toolu_/, "")}`;
+  const withPrefix = id.startsWith("fc_") || id.startsWith("fc-") ? id : `fc_${id.replace(/^toolu_/, "")}`;
+  const sanitized = withPrefix.replace(/[^A-Za-z0-9_-]/g, "_");
+  let mapped = sanitized;
+  let suffix = 2;
+  const used = new Set(idMap.values());
+  while (used.has(mapped)) {
+    mapped = `${sanitized}_${suffix++}`;
+  }
   idMap.set(id, mapped);
   return mapped;
 }
@@ -2747,7 +2835,7 @@ async function fetchCodeAssistWithRetry(plan, options) {
   throw lastError ?? new ProviderError("gemini", "Gemini Code Assist request failed.");
 }
 function streamGemini(options) {
-  return new StreamResult(runStream4(options));
+  return new StreamResult(runStream4(options), options.signal);
 }
 async function* runStream4(options) {
   const useStreaming = options.streaming !== false;
@@ -2967,6 +3055,96 @@ function messagesContainVideo(messages) {
   return false;
 }
 
+// src/error-classification.ts
+var CONTEXT_OVERFLOW_PATTERNS = [
+  /context_length_exceeded/i,
+  /context length exceeded/i,
+  /context window/i,
+  // OpenAI Codex / Responses
+  /maximum context length/i,
+  // OpenAI / OpenRouter / Mistral
+  /prompt is too long/i,
+  // Anthropic
+  /request_too_large/i,
+  // Anthropic HTTP 413
+  /input is too long/i,
+  // Bedrock
+  /input token count.*exceeds the maximum/i,
+  // Gemini
+  /maximum prompt length/i,
+  // xAI / Grok
+  /reduce the length of the messages/i,
+  // Groq
+  /too large for model/i,
+  // Mistral
+  /token limit/i
+  // generic
+];
+var RATE_LIMIT_PATTERNS = [
+  /rate[ _-]?limit/i,
+  /\b429\b/,
+  /too many requests/i,
+  /tokens per minute/i,
+  /requests per minute/i
+];
+var PROVIDER_TRANSIENT_PATTERNS = [
+  /\b5\d\d\b/,
+  /api_error/i,
+  /server_error/i,
+  /internal server error/i,
+  /bad gateway/i,
+  /service unavailable/i,
+  /gateway timeout/i,
+  /overloaded/i,
+  /\b529\b/
+];
+var BILLING_PATTERNS = [
+  /payment required/i,
+  /\b402\b/,
+  /quota_exceeded/i,
+  // underscore variant not in isHardBillingMessage
+  /credit balance/i
+];
+var AUTH_PATTERNS = [
+  /invalid[ _]api[ _]key/i,
+  /unauthorized/i,
+  /\b401\b/,
+  /authentication[ _]failed/i,
+  /please run \/login/i
+  // Anthropic Claude Code-style hint
+];
+function matchesAny(message, patterns) {
+  return patterns.some((p) => p.test(message));
+}
+function classifyProviderError(message) {
+  if (matchesAny(message, CONTEXT_OVERFLOW_PATTERNS)) {
+    return `[context_overflow] Worker context window exceeded \u2014 the conversation is too large to continue. Recovery: call reset_worker(project) to wipe history, then re-prompt with the task. Re-prompting WITHOUT reset will fail the same way.
+
+Original: ${message}`;
+  }
+  if (isHardBillingMessage(message) || matchesAny(message, BILLING_PATTERNS)) {
+    return `[billing] Provider billing/quota issue. Recovery: surface to the user \u2014 they need to top up or switch providers. Do NOT retry.
+
+Original: ${message}`;
+  }
+  if (matchesAny(message, AUTH_PATTERNS)) {
+    return `[auth] Provider authentication failed. Recovery: surface to the user \u2014 they need to re-login. Do NOT retry.
+
+Original: ${message}`;
+  }
+  if (matchesAny(message, RATE_LIMIT_PATTERNS)) {
+    return `[rate_limited] Provider rate limit hit. Recovery: wait ~30s, then re-prompt the same worker (no reset needed).
+
+Original: ${message}`;
+  }
+  if (matchesAny(message, PROVIDER_TRANSIENT_PATTERNS)) {
+    return `[provider_transient] Provider server-side/transient error. Recovery: wait briefly, then re-prompt the same worker (no reset needed). If it keeps happening, switch models/providers or check provider status.
+
+Original: ${message}`;
+  }
+  return message;
+}
+
 // src/providers/palsu.ts
 function palsuText(text) {
   return { role: "assistant", content: text ? [{ type: "text", text }] : [] };
@@ -3113,7 +3291,7 @@ function registerPalsuProvider(config) {
         const stopReason = explicitStop ?? (hasToolCalls ? "tool_use" : "end_turn");
         return yield* simulateStream(message, stopReason, options.signal, cacheUsage);
       })();
-      return new StreamResult(gen);
+      return new StreamResult(gen, options.signal);
     }
   });
   return handle;
@@ -3123,6 +3301,7 @@ export {
   EventStream,
   ProviderError,
   StreamResult,
+  classifyProviderError,
   formatError,
   formatErrorForDisplay,
   isHardBillingMessage,