@prestyj/ai 5.10.0 → 5.12.0

package/dist/index.d.cts

@@ -77,19 +77,31 @@ interface RawContent {
     data: Record<string, unknown>;
 }
 type ContentPart = TextContent | ThinkingContent | ImageContent | VideoContent | ToolCall | ServerToolCall | ServerToolResult | RawContent;
-interface SystemMessage {
+type MessageProvenanceSource = "human" | "agent" | "runtime";
+type MessageProvenanceKind = "prompt" | "steering" | "notification" | "completion_gate" | "review_follow_up" | "continuation" | "model_switch" | "automation" | "compaction_summary" | "compaction_ack";
+type MessageProvenanceVisibility = "transcript" | "hidden" | "summary";
+/** Internal message metadata. `stream()` removes it before provider dispatch. */
+interface MessageProvenance {
+    source: MessageProvenanceSource;
+    kind: MessageProvenanceKind;
+    visibility: MessageProvenanceVisibility;
+}
+interface MessageMetadata {
+    provenance?: MessageProvenance;
+}
+interface SystemMessage extends MessageMetadata {
     role: "system";
     content: string;
 }
-interface UserMessage {
+interface UserMessage extends MessageMetadata {
     role: "user";
     content: string | (TextContent | ImageContent | VideoContent)[];
 }
-interface AssistantMessage {
+interface AssistantMessage extends MessageMetadata {
     role: "assistant";
     content: string | ContentPart[];
 }
-interface ToolResultMessage {
+interface ToolResultMessage extends MessageMetadata {
     role: "tool";
     content: ToolResult[];
 }
@@ -161,7 +173,10 @@ interface StreamResponse {
 }
 interface Usage {
     inputTokens: number;
+    /** Total billed output tokens, including reasoning tokens when the provider reports them separately. */
     outputTokens: number;
+    /** Reasoning/thinking-token subset of outputTokens. */
+    reasoningTokens?: number;
     cacheRead?: number;
     cacheWrite?: number;
     serverToolUse?: {
@@ -299,7 +314,7 @@ declare class StreamResult implements AsyncIterable<StreamEvent> {
  * Local model ids are namespaced by endpoint (`local/<endpointId>/<rawId>`) so
  * the same model name served by two machines stays distinct in the registry.
  * The server only knows the raw id, so strip the routing prefix here — at the
- * one place that talks to the wire. Counterpart to gg-core's
+ * one place that talks to the wire. Counterpart to @prestyj/core's
  * `formatLocalModelId`/`parseLocalModelId`.
  */
 declare function localWireModelId(id: string): string;
@@ -368,7 +383,7 @@ declare class ProviderRegistryImpl {
 declare const providerRegistry: ProviderRegistryImpl;
 
 /**
- * Error model for gg-ai and downstream consumers.
+ * Error model for @prestyj/ai and downstream consumers.
  *
  * Every error users see should answer one question: "is this me or them?"
  * That answer drives whether they retry, switch model, log in, or report a
@@ -435,7 +450,7 @@ declare class ProviderError extends EZCoderAIError {
  * transient per-minute throttle)? These don't clear with a quick retry — the
  * user has to wait for the window to reset — so callers must surface them as a
  * hard stop, not silently retry for minutes. Detected from the canonical
- * "usage limit reached" message gg-ai stamps onto the ProviderError.
+ * "usage limit reached" message @prestyj/ai stamps onto the ProviderError.
  */
 declare function isUsageLimitError(err: unknown): boolean;
 /**
@@ -490,8 +505,23 @@ declare function redactText(text: string, options?: RedactionOptions): string;
  */
 declare function redactValue<T>(value: T, options?: RedactionOptions): T;
 
+/** True when the string contains at least one unpaired surrogate. */
+declare function hasLoneSurrogate(text: string): boolean;
+/** Replace unpaired surrogates with U+FFFD; returns the input when already valid. */
+declare function toWellFormedText(text: string): string;
+/** `text.slice(0, chars)` that never cuts an astral character in half. */
+declare function sliceHead(text: string, chars: number): string;
+/** `text.slice(-chars)` that never cuts an astral character in half. */
+declare function sliceTail(text: string, chars: number): string;
 /**
- * Provider-level diagnostic hook. Mirrors the pattern used by gg-agent's
+ * Strip unpaired surrogates from everything headed for the wire. Returns the
+ * same array (and same message objects) when the history is already valid, so
+ * the clean path stays allocation-free.
+ */
+declare function sanitizeMessagesForWire(messages: Message[]): Message[];
+
+/**
+ * Provider-level diagnostic hook. Mirrors the pattern used by @prestyj/agent's
  * setStreamDiagnostic — the host app wires a callback (typically writing to
  * a debug log) and providers call `providerDiag(...)` to record interesting
  * lifecycle events (e.g. raw SSE event types and timings).
@@ -612,4 +642,4 @@ 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, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, 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, clampProviderContextImages, classifyProviderError, environmentSecrets, formatError, formatErrorForDisplay, isHardBillingMessage, isUsageLimitError, localWireModelId, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, 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 MessageProvenance, type MessageProvenanceKind, type MessageProvenanceSource, type MessageProvenanceVisibility, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, 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, clampProviderContextImages, classifyProviderError, environmentSecrets, formatError, formatErrorForDisplay, hasLoneSurrogate, isHardBillingMessage, isUsageLimitError, localWireModelId, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, sanitizeMessagesForWire, setProviderDiagnostic, sliceHead, sliceTail, stream, toAnthropicMessages, toOpenAIMessages, toWellFormedText };

package/dist/index.d.ts

@@ -77,19 +77,31 @@ interface RawContent {
     data: Record<string, unknown>;
 }
 type ContentPart = TextContent | ThinkingContent | ImageContent | VideoContent | ToolCall | ServerToolCall | ServerToolResult | RawContent;
-interface SystemMessage {
+type MessageProvenanceSource = "human" | "agent" | "runtime";
+type MessageProvenanceKind = "prompt" | "steering" | "notification" | "completion_gate" | "review_follow_up" | "continuation" | "model_switch" | "automation" | "compaction_summary" | "compaction_ack";
+type MessageProvenanceVisibility = "transcript" | "hidden" | "summary";
+/** Internal message metadata. `stream()` removes it before provider dispatch. */
+interface MessageProvenance {
+    source: MessageProvenanceSource;
+    kind: MessageProvenanceKind;
+    visibility: MessageProvenanceVisibility;
+}
+interface MessageMetadata {
+    provenance?: MessageProvenance;
+}
+interface SystemMessage extends MessageMetadata {
     role: "system";
     content: string;
 }
-interface UserMessage {
+interface UserMessage extends MessageMetadata {
     role: "user";
     content: string | (TextContent | ImageContent | VideoContent)[];
 }
-interface AssistantMessage {
+interface AssistantMessage extends MessageMetadata {
     role: "assistant";
     content: string | ContentPart[];
 }
-interface ToolResultMessage {
+interface ToolResultMessage extends MessageMetadata {
     role: "tool";
     content: ToolResult[];
 }
@@ -161,7 +173,10 @@ interface StreamResponse {
 }
 interface Usage {
     inputTokens: number;
+    /** Total billed output tokens, including reasoning tokens when the provider reports them separately. */
     outputTokens: number;
+    /** Reasoning/thinking-token subset of outputTokens. */
+    reasoningTokens?: number;
     cacheRead?: number;
     cacheWrite?: number;
     serverToolUse?: {
@@ -299,7 +314,7 @@ declare class StreamResult implements AsyncIterable<StreamEvent> {
  * Local model ids are namespaced by endpoint (`local/<endpointId>/<rawId>`) so
  * the same model name served by two machines stays distinct in the registry.
  * The server only knows the raw id, so strip the routing prefix here — at the
- * one place that talks to the wire. Counterpart to gg-core's
+ * one place that talks to the wire. Counterpart to @prestyj/core's
  * `formatLocalModelId`/`parseLocalModelId`.
  */
 declare function localWireModelId(id: string): string;
@@ -368,7 +383,7 @@ declare class ProviderRegistryImpl {
 declare const providerRegistry: ProviderRegistryImpl;
 
 /**
- * Error model for gg-ai and downstream consumers.
+ * Error model for @prestyj/ai and downstream consumers.
  *
  * Every error users see should answer one question: "is this me or them?"
  * That answer drives whether they retry, switch model, log in, or report a
@@ -435,7 +450,7 @@ declare class ProviderError extends EZCoderAIError {
  * transient per-minute throttle)? These don't clear with a quick retry — the
  * user has to wait for the window to reset — so callers must surface them as a
  * hard stop, not silently retry for minutes. Detected from the canonical
- * "usage limit reached" message gg-ai stamps onto the ProviderError.
+ * "usage limit reached" message @prestyj/ai stamps onto the ProviderError.
  */
 declare function isUsageLimitError(err: unknown): boolean;
 /**
@@ -490,8 +505,23 @@ declare function redactText(text: string, options?: RedactionOptions): string;
  */
 declare function redactValue<T>(value: T, options?: RedactionOptions): T;
 
+/** True when the string contains at least one unpaired surrogate. */
+declare function hasLoneSurrogate(text: string): boolean;
+/** Replace unpaired surrogates with U+FFFD; returns the input when already valid. */
+declare function toWellFormedText(text: string): string;
+/** `text.slice(0, chars)` that never cuts an astral character in half. */
+declare function sliceHead(text: string, chars: number): string;
+/** `text.slice(-chars)` that never cuts an astral character in half. */
+declare function sliceTail(text: string, chars: number): string;
 /**
- * Provider-level diagnostic hook. Mirrors the pattern used by gg-agent's
+ * Strip unpaired surrogates from everything headed for the wire. Returns the
+ * same array (and same message objects) when the history is already valid, so
+ * the clean path stays allocation-free.
+ */
+declare function sanitizeMessagesForWire(messages: Message[]): Message[];
+
+/**
+ * Provider-level diagnostic hook. Mirrors the pattern used by @prestyj/agent's
  * setStreamDiagnostic — the host app wires a callback (typically writing to
  * a debug log) and providers call `providerDiag(...)` to record interesting
  * lifecycle events (e.g. raw SSE event types and timings).
@@ -612,4 +642,4 @@ 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, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, 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, clampProviderContextImages, classifyProviderError, environmentSecrets, formatError, formatErrorForDisplay, isHardBillingMessage, isUsageLimitError, localWireModelId, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, 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 MessageProvenance, type MessageProvenanceKind, type MessageProvenanceSource, type MessageProvenanceVisibility, type PalsuModelConfig, type PalsuModelHandle, type PalsuProviderConfig, type PalsuProviderHandle, type PalsuProviderState, type PalsuResponse, type PalsuResponseFactory, type Provider, type ProviderDiagnosticFn, type ProviderEntry, ProviderError, type ProviderStreamFn, REDACTED as REDACTION_MARKER, type RawContent, type RedactionOptions, 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, clampProviderContextImages, classifyProviderError, environmentSecrets, formatError, formatErrorForDisplay, hasLoneSurrogate, isHardBillingMessage, isUsageLimitError, localWireModelId, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, sanitizeMessagesForWire, setProviderDiagnostic, sliceHead, sliceTail, stream, toAnthropicMessages, toOpenAIMessages, toWellFormedText };

package/dist/index.js

@@ -1160,7 +1160,7 @@ function parseToolArguments(argsJson) {
 var NON_STREAMING_TIMEOUT_MS = 60 * 60 * 1e3;
 var anthropicClientCache = /* @__PURE__ */ new Map();
 function fineGrainedToolStreamingEnabled() {
-  const raw = process.env.GG_FINE_GRAINED_TOOL_STREAMING ?? process.env.CLAUDE_CODE_ENABLE_FINE_GRAINED_TOOL_STREAMING;
+  const raw = process.env.EZ_FINE_GRAINED_TOOL_STREAMING ?? process.env.CLAUDE_CODE_ENABLE_FINE_GRAINED_TOOL_STREAMING;
   if (!raw) return false;
   const v = raw.trim().toLowerCase();
   return v === "1" || v === "true" || v === "yes" || v === "on";
@@ -2025,7 +2025,15 @@ async function* runStream2(options) {
       if (chunk.usage) {
         ({ inputTokens, outputTokens, cacheRead, cacheWrite } = extractOpenAIUsage(chunk.usage));
       }
-      if (!choice) continue;
+      if (!choice) {
+        const gatewayError = classifyChoicelessFrame(chunk);
+        if (gatewayError) {
+          throw new ProviderError(providerName, gatewayError.message, {
+            statusCode: gatewayError.statusCode
+          });
+        }
+        continue;
+      }
       if (choice.finish_reason) {
         finishReason = choice.finish_reason;
       }
@@ -2209,6 +2217,31 @@ function completionToResponse(completion, endpointKey) {
     }
   };
 }
+function classifyChoicelessFrame(frame) {
+  if (!frame || typeof frame !== "object" || Array.isArray(frame)) return null;
+  const rec = frame;
+  if (Array.isArray(rec.choices)) return null;
+  const statusOf = (value) => {
+    const n = typeof value === "string" ? Number(value) : value;
+    return typeof n === "number" && Number.isFinite(n) && n >= 400 && n <= 599 ? n : void 0;
+  };
+  const statusCode = statusOf(rec.status) ?? statusOf(rec.statusCode) ?? statusOf(rec.code);
+  const typeIsError = typeof rec.type === "string" && rec.type.toLowerCase() === "error";
+  let detailText;
+  const detail = rec.detail;
+  if (typeof detail === "string" && detail.trim()) {
+    detailText = detail.trim();
+  } else if (Array.isArray(detail)) {
+    const parts = detail.map(
+      (d) => d && typeof d === "object" && typeof d.msg === "string" ? d.msg : typeof d === "string" ? d : ""
+    ).filter(Boolean);
+    if (parts.length) detailText = parts.join("; ");
+  }
+  if (statusCode === void 0 && !typeIsError && !detailText) return null;
+  const rawMessage = (typeof rec.message === "string" && rec.message.trim() ? rec.message.trim() : void 0) ?? detailText ?? (typeof rec.error === "string" && rec.error.trim() ? rec.error.trim() : void 0) ?? (statusCode !== void 0 ? `Gateway returned status ${statusCode}.` : "Gateway error.");
+  const message = rawMessage.slice(0, 500);
+  return { message, statusCode };
+}
 function classifyOpenAICompatLimit(args) {
   const { status, code, type, message } = args;
   const codeType = `${code ?? ""} ${type ?? ""}`.toLowerCase();
@@ -2220,6 +2253,7 @@ function classifyOpenAICompatLimit(args) {
   return null;
 }
 function toError2(err, provider = "openai") {
+  if (err instanceof ProviderError) return err;
   if (err instanceof OpenAI.APIError) {
     const body = err.error;
     const bodyMessage = typeof body?.message === "string" && body.message.trim() ? body.message.trim() : void 0;
@@ -3319,14 +3353,16 @@ async function* runStream4(options) {
   let thinkingAccum = "";
   let stopReason = "end_turn";
   let inputTokens = 0;
-  let outputTokens = 0;
+  let candidateTokens = 0;
+  let reasoningTokens = 0;
   let cacheRead = 0;
   let toolIndex = 0;
   const handleResponse = function* (chunk) {
     const usage = usageFromResponse(chunk);
     if (usage) {
       inputTokens = usage.promptTokenCount ?? inputTokens;
-      outputTokens = usage.candidatesTokenCount ?? outputTokens;
+      candidateTokens = usage.candidatesTokenCount ?? candidateTokens;
+      reasoningTokens = usage.thoughtsTokenCount ?? reasoningTokens;
       cacheRead = usage.cachedContentTokenCount ?? cacheRead;
     }
     const reason = finishReasonFromResponse(chunk);
@@ -3382,6 +3418,7 @@ async function* runStream4(options) {
   }
   if (pendingToolCalls.length > 0) stopReason = "tool_use";
   const adjustedInputTokens = Math.max(0, inputTokens - cacheRead);
+  const outputTokens = candidateTokens + reasoningTokens;
   const streamResponse = {
     message: {
       role: "assistant",
@@ -3391,6 +3428,7 @@ async function* runStream4(options) {
     usage: {
       inputTokens: adjustedInputTokens,
       outputTokens,
+      ...reasoningTokens > 0 ? { reasoningTokens } : {},
       ...cacheRead > 0 ? { cacheRead } : {}
     }
   };
@@ -3439,9 +3477,145 @@ var ProviderRegistryImpl = class {
 };
 var providerRegistry = new ProviderRegistryImpl();
 
+// src/utils/well-formed.ts
+var LONE_SURROGATE = /[\uD800-\uDBFF](?![\uDC00-\uDFFF])|(?<![\uD800-\uDBFF])[\uDC00-\uDFFF]/;
+var LONE_SURROGATE_GLOBAL = new RegExp(LONE_SURROGATE, "g");
+var REPLACEMENT = "\uFFFD";
+function hasLoneSurrogate(text) {
+  const isWellFormed = text.isWellFormed;
+  if (typeof isWellFormed === "function") return !isWellFormed.call(text);
+  return LONE_SURROGATE.test(text);
+}
+function toWellFormedText(text) {
+  if (!hasLoneSurrogate(text)) return text;
+  const toWellFormed = text.toWellFormed;
+  if (typeof toWellFormed === "function") return toWellFormed.call(text);
+  return text.replace(LONE_SURROGATE_GLOBAL, REPLACEMENT);
+}
+function isHighSurrogate(code) {
+  return code !== void 0 && code >= 55296 && code <= 56319;
+}
+function isLowSurrogate(code) {
+  return code !== void 0 && code >= 56320 && code <= 57343;
+}
+function sliceHead(text, chars) {
+  if (chars <= 0) return "";
+  if (chars >= text.length) return text;
+  const end = isHighSurrogate(text.charCodeAt(chars - 1)) ? chars - 1 : chars;
+  return text.slice(0, end);
+}
+function sliceTail(text, chars) {
+  if (chars <= 0) return "";
+  if (chars >= text.length) return text;
+  const start = text.length - chars;
+  return text.slice(isLowSurrogate(text.charCodeAt(start)) ? start + 1 : start);
+}
+function sanitizeJsonValue(value) {
+  if (typeof value === "string") return toWellFormedText(value);
+  if (Array.isArray(value)) {
+    let changed = false;
+    const next = value.map((item) => {
+      const sanitized = sanitizeJsonValue(item);
+      if (sanitized !== item) changed = true;
+      return sanitized;
+    });
+    return changed ? next : value;
+  }
+  if (value !== null && typeof value === "object") {
+    let changed = false;
+    const next = {};
+    for (const [key, item] of Object.entries(value)) {
+      const sanitizedKey = toWellFormedText(key);
+      const sanitized = sanitizeJsonValue(item);
+      if (sanitizedKey !== key || sanitized !== item) changed = true;
+      next[sanitizedKey] = sanitized;
+    }
+    return changed ? next : value;
+  }
+  return value;
+}
+function sanitizeRecord(value) {
+  return sanitizeJsonValue(value);
+}
+function sanitizePart(part) {
+  switch (part.type) {
+    case "text":
+    case "thinking": {
+      const text = toWellFormedText(part.text);
+      return text === part.text ? part : { ...part, text };
+    }
+    case "tool_call": {
+      const args = sanitizeRecord(part.args);
+      return args === part.args ? part : { ...part, args };
+    }
+    case "server_tool_call": {
+      const input = sanitizeJsonValue(part.input);
+      return input === part.input ? part : { ...part, input };
+    }
+    case "server_tool_result": {
+      const data = sanitizeJsonValue(part.data);
+      return data === part.data ? part : { ...part, data };
+    }
+    case "raw": {
+      const data = sanitizeRecord(part.data);
+      return data === part.data ? part : { ...part, data };
+    }
+    default:
+      return part;
+  }
+}
+function sanitizeParts(parts) {
+  let changed = false;
+  const next = parts.map((part) => {
+    const sanitized = sanitizePart(part);
+    if (sanitized !== part) changed = true;
+    return sanitized;
+  });
+  return changed ? next : parts;
+}
+function sanitizeToolResultContent(content) {
+  if (typeof content === "string") return toWellFormedText(content);
+  return sanitizeParts(content);
+}
+function sanitizeToolResults(results) {
+  let changed = false;
+  const next = results.map((result) => {
+    const content = sanitizeToolResultContent(result.content);
+    if (content === result.content) return result;
+    changed = true;
+    return { ...result, content };
+  });
+  return changed ? next : results;
+}
+function sanitizeMessage(message) {
+  if (message.role === "tool") {
+    const content2 = sanitizeToolResults(message.content);
+    return content2 === message.content ? message : { ...message, content: content2 };
+  }
+  if (typeof message.content === "string") {
+    const content2 = toWellFormedText(message.content);
+    return content2 === message.content ? message : { ...message, content: content2 };
+  }
+  const content = sanitizeParts(message.content);
+  return content === message.content ? message : { ...message, content };
+}
+function sanitizeMessagesForWire(messages) {
+  let sanitized;
+  for (let index = 0; index < messages.length; index++) {
+    const message = messages[index];
+    const next = sanitizeMessage(message);
+    if (next === message) continue;
+    sanitized ??= messages.slice();
+    sanitized[index] = next;
+  }
+  return sanitized ?? messages;
+}
+
 // src/stream.ts
 var GLM_CODING_BASE_URL = "https://api.z.ai/api/coding/paas/v4";
 var KIMI_CODE_USER_AGENT = `kimi-code-cli/${process.env.KIMI_CODE_VERSION ?? "1.0.11"}`;
+var GROK_CLI_PROXY_HOST = "cli-chat-proxy.grok.com";
+var GROK_CLI_VERSION = process.env.GROK_CLI_VERSION ?? "0.2.101";
 providerRegistry.register("anthropic", {
   stream: (options) => streamAnthropic(options)
 });
@@ -3508,13 +3682,25 @@ providerRegistry.register("xai", {
   // xAI's public API (console.x.ai key) is OpenAI-compatible — ride the Chat
   // Completions transport like Moonshot/DeepSeek. Grok reasoning models take
   // top-level `reasoning_effort` (low/medium/high), which the shared thinking
-  // path already sends. xAI's OAuth path exists but only via the Grok CLI's
-  // private Responses proxy (cli-chat-proxy.grok.com) with reverse-engineered
-  // attribution headers and account-tier gating — intentionally not wired.
-  stream: (options) => streamOpenAI({
-    ...options,
-    baseUrl: options.baseUrl ?? "https://api.x.ai/v1"
-  })
+  // path already sends.
+  //
+  // Subscription OAuth (SuperGrok / X Premium) routes to the Grok CLI chat proxy
+  // instead, which speaks the same Chat Completions wire but gates on Grok-CLI
+  // client identity. Inject those headers centrally here — exactly as the Kimi
+  // endpoint above — so EVERY stream (agent loop, compaction, title-gen,
+  // sub-agents) is accepted rather than depending on each call site to thread
+  // headers. Caller-provided headers still win on collision.
+  stream: (options) => {
+    const baseUrl = options.baseUrl ?? "https://api.x.ai/v1";
+    const defaultHeaders = baseUrl.includes(GROK_CLI_PROXY_HOST) ? {
+      "X-XAI-Token-Auth": "xai-grok-cli",
+      "x-grok-client-version": GROK_CLI_VERSION,
+      "x-grok-client-identifier": "ezcoder",
+      "x-grok-model-override": options.model,
+      ...options.defaultHeaders
+    } : options.defaultHeaders;
+    return streamOpenAI({ ...options, baseUrl, defaultHeaders });
+  }
 });
 providerRegistry.register("minimax", {
   stream: (options) => streamAnthropic({
@@ -3560,13 +3746,25 @@ function stream(options) {
   if (options.supportsVideo !== true && messagesContainVideo(options.messages)) {
     throw new VideoUnsupportedError();
   }
+  const wireMessages = stripMessageProvenance(options.messages);
   const messages = clampProviderContextImages(
-    options.messages,
+    sanitizeMessagesForWire(wireMessages),
     options.provider,
     options.supportsImages
   );
   return entry.stream(messages === options.messages ? options : { ...options, messages });
 }
+function stripMessageProvenance(messages) {
+  let stripped;
+  for (let index = 0; index < messages.length; index++) {
+    const message = messages[index];
+    if (!message.provenance) continue;
+    stripped ??= messages.slice();
+    const { provenance: _provenance, ...wireMessage } = message;
+    stripped[index] = wireMessage;
+  }
+  return stripped ?? messages;
+}
 function messagesContainVideo(messages) {
   for (const msg of messages) {
     if (typeof msg.content === "string" || !Array.isArray(msg.content)) continue;
@@ -3951,6 +4149,7 @@ export {
   environmentSecrets,
   formatError,
   formatErrorForDisplay,
+  hasLoneSurrogate,
   isHardBillingMessage,
   isUsageLimitError,
   localWireModelId,
@@ -3963,9 +4162,13 @@ export {
   redactText,
   redactValue,
   registerPalsuProvider,
+  sanitizeMessagesForWire,
   setProviderDiagnostic,
+  sliceHead,
+  sliceTail,
   stream,
   toAnthropicMessages,
-  toOpenAIMessages
+  toOpenAIMessages,
+  toWellFormedText
 };
 //# sourceMappingURL=index.js.map