@prestyj/ai 5.9.0 → 5.11.0

package/dist/index.d.cts

@@ -2,7 +2,9 @@ import { z } from 'zod';
 import Anthropic from '@anthropic-ai/sdk';
 import OpenAI from 'openai';
 
-type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "sakana" | "xai" | "palsu";
+type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "sakana" | "xai" | "palsu"
+/** Locally hosted OpenAI-compatible server (Ollama, LM Studio, llama.cpp, vLLM). */
+ | "local";
 type ThinkingLevel = "low" | "medium" | "high" | "xhigh" | "max" | "ultra";
 type CacheRetention = "none" | "short" | "long";
 interface TextContent {
@@ -75,19 +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[];
 }
@@ -159,7 +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?: {
@@ -293,6 +310,14 @@ declare class StreamResult implements AsyncIterable<StreamEvent> {
     then<TResult1 = StreamResponse, TResult2 = never>(onfulfilled?: ((value: StreamResponse) => TResult1 | PromiseLike<TResult1>) | null, onrejected?: ((reason: unknown) => TResult2 | PromiseLike<TResult2>) | null): Promise<TResult1 | TResult2>;
 }
 
+/**
+ * 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
+ * `formatLocalModelId`/`parseLocalModelId`.
+ */
+declare function localWireModelId(id: string): string;
 /**
  * Unified streaming entry point. Returns a StreamResult that is both
  * an async iterable (for streaming events) and thenable (await for
@@ -480,6 +505,21 @@ 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;
+/**
+ * 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 gg-agent's
  * setStreamDiagnostic — the host app wires a callback (typically writing to
@@ -507,6 +547,8 @@ declare function toOpenAIMessages(messages: Message[], options?: {
     provider?: string;
     thinking?: boolean;
     supportsImages?: boolean;
+    /** Wire name for reasoning on assistant messages. Defaults to `reasoning_content`. */
+    reasoningField?: string;
 }): OpenAI.ChatCompletionMessageParam[];
 
 /**
@@ -600,4 +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, 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

@@ -2,7 +2,9 @@ import { z } from 'zod';
 import Anthropic from '@anthropic-ai/sdk';
 import OpenAI from 'openai';
 
-type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "sakana" | "xai" | "palsu";
+type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "sakana" | "xai" | "palsu"
+/** Locally hosted OpenAI-compatible server (Ollama, LM Studio, llama.cpp, vLLM). */
+ | "local";
 type ThinkingLevel = "low" | "medium" | "high" | "xhigh" | "max" | "ultra";
 type CacheRetention = "none" | "short" | "long";
 interface TextContent {
@@ -75,19 +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[];
 }
@@ -159,7 +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?: {
@@ -293,6 +310,14 @@ declare class StreamResult implements AsyncIterable<StreamEvent> {
     then<TResult1 = StreamResponse, TResult2 = never>(onfulfilled?: ((value: StreamResponse) => TResult1 | PromiseLike<TResult1>) | null, onrejected?: ((reason: unknown) => TResult2 | PromiseLike<TResult2>) | null): Promise<TResult1 | TResult2>;
 }
 
+/**
+ * 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
+ * `formatLocalModelId`/`parseLocalModelId`.
+ */
+declare function localWireModelId(id: string): string;
 /**
  * Unified streaming entry point. Returns a StreamResult that is both
  * an async iterable (for streaming events) and thenable (await for
@@ -480,6 +505,21 @@ 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;
+/**
+ * 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 gg-agent's
  * setStreamDiagnostic — the host app wires a callback (typically writing to
@@ -507,6 +547,8 @@ declare function toOpenAIMessages(messages: Message[], options?: {
     provider?: string;
     thinking?: boolean;
     supportsImages?: boolean;
+    /** Wire name for reasoning on assistant messages. Defaults to `reasoning_content`. */
+    reasoningField?: string;
 }): OpenAI.ChatCompletionMessageParam[];
 
 /**
@@ -600,4 +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, 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

@@ -501,6 +501,39 @@ function normalizeRootForAnthropic(schema) {
   return out;
 }
 
+// src/providers/reasoning-field.ts
+var REASONING_FIELD_ALIASES = [
+  "reasoning_content",
+  "reasoning",
+  "reasoning_text"
+];
+var DEFAULT_REASONING_FIELD = REASONING_FIELD_ALIASES[0];
+function readReasoning(obj) {
+  if (!obj) return void 0;
+  for (const field of REASONING_FIELD_ALIASES) {
+    const value = obj[field];
+    if (typeof value === "string" && value) return { field, text: value };
+  }
+  return void 0;
+}
+function reasoningFieldKey(provider, baseUrl, model) {
+  return `${provider}|${baseUrl ?? ""}|${model}`;
+}
+var MAX_REMEMBERED_ENDPOINTS = 64;
+var detectedFields = /* @__PURE__ */ new Map();
+function rememberReasoningField(key, field) {
+  if (detectedFields.get(key) === field) return;
+  detectedFields.set(key, field);
+  while (detectedFields.size > MAX_REMEMBERED_ENDPOINTS) {
+    const oldest = detectedFields.keys().next();
+    if (oldest.done) break;
+    detectedFields.delete(oldest.value);
+  }
+}
+function getReasoningField(key) {
+  return detectedFields.get(key) ?? DEFAULT_REASONING_FIELD;
+}
+
 // src/providers/transform.ts
 function hasValidThinkingSignature(part) {
   return typeof part.signature === "string" && part.signature.trim().length > 0;
@@ -885,12 +918,12 @@ function toAnthropicToolChoice(choice) {
   return { type: "tool", name: choice.name };
 }
 function isAdaptiveThinkingModel(model) {
-  return /opus-4[-.]8|opus-4[-.]7|opus-4[-.]6|sonnet-5|fable-5|mythos-5/.test(model);
+  return /opus-5|opus-4[-.]8|opus-4[-.]7|opus-4[-.]6|sonnet-5|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/.test(model)) {
+    if (effort === "xhigh" && !/opus-5|opus-4-8|opus-4-7/.test(model)) {
       effort = "high";
     }
     return {
@@ -921,6 +954,7 @@ function remapToolCallId(id, idMap) {
   return mapped;
 }
 function toOpenAIMessages(messages, options) {
+  const reasoningField = options?.reasoningField || DEFAULT_REASONING_FIELD;
   const out = [];
   const idMap = /* @__PURE__ */ new Map();
   const mergeToolResultText = options?.provider === "glm";
@@ -987,9 +1021,9 @@ function toOpenAIMessages(messages, options) {
         ...hasToolCalls ? { tool_calls: toolCalls } : {}
       };
       if (thinkingParts) {
-        assistantMsg.reasoning_content = thinkingParts;
+        assistantMsg[reasoningField] = thinkingParts;
       } else if (options?.thinking && hasToolCalls && options.provider !== "glm") {
-        assistantMsg.reasoning_content = " ";
+        assistantMsg[reasoningField] = " ";
       }
       out.push(assistantMsg);
       continue;
@@ -1067,6 +1101,10 @@ function toOpenAIToolChoice(choice) {
   if (choice === "required") return "required";
   return { type: "function", function: { name: choice.name } };
 }
+function toLocalReasoningEffort(level) {
+  if (level === "max" || level === "ultra" || level === "xhigh") return "max";
+  return level;
+}
 function toOpenAIReasoningEffort(level, model) {
   const effort = level === "max" || level === "ultra" ? "xhigh" : level;
   if (model.startsWith("fugu") && (effort === "low" || effort === "medium")) {
@@ -1863,7 +1901,9 @@ function streamOpenAI(options) {
 async function* runStream2(options) {
   const providerName = options.provider ?? "openai";
   const useStreaming = options.streaming !== false;
+  const endpointKey = reasoningFieldKey(providerName, options.baseUrl, options.model);
   const client = createClient2(options);
+  const isLocal = options.provider === "local";
   const isKimiK3 = options.provider === "moonshot" && options.model === "kimi-k3";
   const isManagedKimiK3 = isKimiK3 && options.baseUrl?.replace(/\/+$/, "").endsWith("/coding/v1") === true;
   const k3Effort = options.thinking ? toKimiK3Effort(options.thinking) : void 0;
@@ -1886,7 +1926,8 @@ async function* runStream2(options) {
     // disabled K3 must NOT carry placeholder reasoning_content (mirrors the
     // official CLI: reasoning is preserved only while thinking is enabled).
     thinking: isKimiK27 || !!options.thinking,
-    supportsImages: options.supportsImages
+    supportsImages: options.supportsImages,
+    reasoningField: getReasoningField(endpointKey)
   });
   const defaultTemp = options.provider === "glm" ? 0.6 : void 0;
   const effectiveTemp = options.temperature ?? defaultTemp;
@@ -1898,7 +1939,7 @@ async function* runStream2(options) {
     ...effectiveTemp != null && !options.thinking && !hasFixedKimiSampling ? { temperature: effectiveTemp } : {},
     ...options.topP != null && !hasFixedKimiSampling ? { top_p: options.topP } : {},
     ...options.stop ? { stop: options.stop } : {},
-    ...options.thinking && !usesThinkingParam && !isKimiK3 && !isKimiK27 ? { reasoning_effort: toOpenAIReasoningEffort(options.thinking, options.model) } : {},
+    ...options.thinking && !usesThinkingParam && !isKimiK3 && !isKimiK27 && !isLocal ? { reasoning_effort: toOpenAIReasoningEffort(options.thinking, options.model) } : {},
     ...options.tools?.length ? { tools: toOpenAITools(options.tools) } : {},
     ...options.toolChoice && options.tools?.length ? { tool_choice: toOpenAIToolChoice(options.toolChoice) } : {},
     ...useStreaming ? { stream_options: { include_usage: true } } : {}
@@ -1912,6 +1953,11 @@ async function* runStream2(options) {
       paramsAny.prompt_cache_retention = "24h";
     }
   }
+  if (isLocal && options.thinking) {
+    params.reasoning_effort = toLocalReasoningEffort(
+      options.thinking
+    );
+  }
   if (options.provider === "openai" && options.serviceTier) {
     params.service_tier = options.serviceTier;
   }
@@ -1948,8 +1994,8 @@ async function* runStream2(options) {
       const completion = await client.chat.completions.create(params, {
         signal: options.signal ?? void 0
       });
-      yield* synthesizeEventsFromCompletion(completion, !!options.thinking);
-      return completionToResponse(completion);
+      yield* synthesizeEventsFromCompletion(completion, !!options.thinking, endpointKey);
+      return completionToResponse(completion, endpointKey);
     } catch (err) {
       throw toError2(err, providerName);
     }
@@ -1984,11 +2030,12 @@ async function* runStream2(options) {
         finishReason = choice.finish_reason;
       }
       const delta = choice.delta;
-      const reasoningContent = delta.reasoning_content;
-      if (typeof reasoningContent === "string" && reasoningContent) {
-        thinkingAccum += reasoningContent;
+      const reasoning = readReasoning(delta);
+      if (reasoning) {
+        rememberReasoningField(endpointKey, reasoning.field);
+        thinkingAccum += reasoning.text;
         if (options.thinking) {
-          yield { type: "thinking_delta", text: reasoningContent };
+          yield { type: "thinking_delta", text: reasoning.text };
         }
       }
       if (delta.content) {
@@ -2073,16 +2120,17 @@ async function* runStream2(options) {
   yield { type: "done", stopReason };
   return response;
 }
-function* synthesizeEventsFromCompletion(completion, thinkingEnabled) {
+function* synthesizeEventsFromCompletion(completion, thinkingEnabled, endpointKey) {
   const choice = completion.choices?.[0];
   if (!choice) {
     yield { type: "done", stopReason: normalizeOpenAIStopReason(null) };
     return;
   }
   const msg = choice.message;
-  const reasoning = msg.reasoning_content;
-  if (typeof reasoning === "string" && reasoning && thinkingEnabled) {
-    yield { type: "thinking_delta", text: reasoning };
+  const reasoning = readReasoning(msg);
+  if (reasoning) {
+    rememberReasoningField(endpointKey, reasoning.field);
+    if (thinkingEnabled) yield { type: "thinking_delta", text: reasoning.text };
   }
   if (typeof msg.content === "string" && msg.content) {
     yield { type: "text_delta", text: msg.content };
@@ -2110,15 +2158,16 @@ function* synthesizeEventsFromCompletion(completion, thinkingEnabled) {
   }
   yield { type: "done", stopReason: normalizeOpenAIStopReason(choice.finish_reason ?? null) };
 }
-function completionToResponse(completion) {
+function completionToResponse(completion, endpointKey) {
   const choice = completion.choices?.[0];
   const contentParts = [];
   let textAccum = "";
   if (choice) {
     const msg = choice.message;
-    const reasoning = msg.reasoning_content;
-    if (typeof reasoning === "string" && reasoning) {
-      contentParts.push({ type: "thinking", text: reasoning });
+    const reasoning = readReasoning(msg);
+    if (reasoning) {
+      rememberReasoningField(endpointKey, reasoning.field);
+      contentParts.push({ type: "thinking", text: reasoning.text });
     }
     if (typeof msg.content === "string" && msg.content) {
       textAccum = msg.content;
@@ -3270,14 +3319,16 @@ 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);
@@ -3333,6 +3384,7 @@ 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",
@@ -3342,6 +3394,7 @@ async function* runStream4(options) {
     usage: {
       inputTokens: adjustedInputTokens,
       outputTokens,
+      ...reasoningTokens > 0 ? { reasoningTokens } : {},
       ...cacheRead > 0 ? { cacheRead } : {}
     }
   };
@@ -3390,6 +3443,140 @@ 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/${process.env.KIMI_CODE_VERSION ?? "1.0.11"}`;
@@ -3479,6 +3666,28 @@ providerRegistry.register("minimax", {
     serverTools: void 0
   })
 });
+function localWireModelId(id) {
+  const match = /^local\/[^/]+\/(.+)$/.exec(id);
+  return match?.[1] ?? id;
+}
+providerRegistry.register("local", {
+  // Locally hosted OpenAI-compatible servers (Ollama, LM Studio, llama.cpp,
+  // vLLM). There is no default endpoint: the baseUrl comes from the endpoint
+  // credential the discovery layer wrote, so a missing one is a wiring bug, not
+  // something to paper over with a guess at someone else's port.
+  stream: (options) => {
+    if (!options.baseUrl) {
+      throw new EZCoderAIError(
+        "Local provider requires a baseUrl (e.g. http://127.0.0.1:11434/v1). No local endpoint was resolved for this model \u2014 re-scan for local models."
+      );
+    }
+    return streamOpenAI({
+      ...options,
+      model: localWireModelId(options.model),
+      webSearch: false
+    });
+  }
+});
 function stream(options) {
   const entry = providerRegistry.get(options.provider);
   if (!entry) {
@@ -3489,13 +3698,25 @@ 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;
@@ -3880,8 +4101,10 @@ export {
   environmentSecrets,
   formatError,
   formatErrorForDisplay,
+  hasLoneSurrogate,
   isHardBillingMessage,
   isUsageLimitError,
+  localWireModelId,
   palsuAssistantMessage,
   palsuText,
   palsuThinking,
@@ -3891,9 +4114,13 @@ export {
   redactText,
   redactValue,
   registerPalsuProvider,
+  sanitizeMessagesForWire,
   setProviderDiagnostic,
+  sliceHead,
+  sliceTail,
   stream,
   toAnthropicMessages,
-  toOpenAIMessages
+  toOpenAIMessages,
+  toWellFormedText
 };
 //# sourceMappingURL=index.js.map