@prestyj/ai 5.8.0 → 5.10.0

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 {
@@ -41,6 +43,21 @@ interface ToolResult {
     toolCallId: string;
     content: ToolResultContent;
     isError?: boolean;
+    /**
+     * Set when the agent loop trimmed `content` to fit a per-result or per-turn
+     * budget. The provider (model input) and the persistent transcript both see
+     * the trimmed `content`, but the live `tool_call_end` event carried the FULL
+     * preview — so this marker makes that divergence explicit and reconcilable.
+     * Internal metadata only: it is never serialized onto the provider wire.
+     */
+    capped?: {
+        /** Length of the original, untrimmed string content. */
+        originalChars: number;
+        /** Length of the trimmed content actually sent to the model. */
+        keptChars: number;
+        /** Which budget triggered the trim. */
+        scope: "per-result" | "per-turn";
+    };
 }
 interface ServerToolCall {
     type: "server_tool_call";
@@ -278,6 +295,14 @@ 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
@@ -492,6 +517,8 @@ 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[];
 
 /**
@@ -585,4 +612,4 @@ 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 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 };

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 {
@@ -41,6 +43,21 @@ interface ToolResult {
     toolCallId: string;
     content: ToolResultContent;
     isError?: boolean;
+    /**
+     * Set when the agent loop trimmed `content` to fit a per-result or per-turn
+     * budget. The provider (model input) and the persistent transcript both see
+     * the trimmed `content`, but the live `tool_call_end` event carried the FULL
+     * preview — so this marker makes that divergence explicit and reconcilable.
+     * Internal metadata only: it is never serialized onto the provider wire.
+     */
+    capped?: {
+        /** Length of the original, untrimmed string content. */
+        originalChars: number;
+        /** Length of the trimmed content actually sent to the model. */
+        keptChars: number;
+        /** Which budget triggered the trim. */
+        scope: "per-result" | "per-turn";
+    };
 }
 interface ServerToolCall {
     type: "server_tool_call";
@@ -278,6 +295,14 @@ 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
@@ -492,6 +517,8 @@ 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[];
 
 /**
@@ -585,4 +612,4 @@ 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 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 };

package/dist/index.js

@@ -264,6 +264,9 @@ function providerGuidance(provider, message, statusCode) {
   if (lower.includes("context_length_exceeded") || lower.includes("prompt is too long")) {
     return `Context window for this ${name} model is full. Compact the conversation to shrink history, or start a new session.`;
   }
+  if (lower.includes("many-image request") || lower.includes("image dimensions") && lower.includes("max allowed size")) {
+    return `An image in conversation history exceeds ${name}'s many-image limit. Restart EZ Coder so restored images are resized, then retry; if it persists, start a new session.`;
+  }
   if (statusCode === 413 || lower.includes("request_too_large") || lower.includes("request exceeds the maximum size")) {
     return `The request to ${name} is too large. Compact the conversation to shrink history, or start a new session.`;
   }
@@ -498,6 +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;
@@ -760,9 +796,14 @@ function toAnthropicMessages(messages, cacheControl) {
       continue;
     }
     if (msg.role === "user") {
+      if (typeof msg.content === "string") {
+        if (msg.content === "") continue;
+      } else if (!msg.content.some((p) => !(p.type === "text" && p.text === ""))) {
+        continue;
+      }
       out.push({
         role: "user",
-        content: typeof msg.content === "string" ? msg.content : msg.content.map((part) => {
+        content: typeof msg.content === "string" ? msg.content : msg.content.filter((part) => !(part.type === "text" && part.text === "")).map((part) => {
           if (part.type === "text") return { type: "text", text: part.text };
           if (part.type === "video") {
             return {
@@ -787,6 +828,7 @@ function toAnthropicMessages(messages, cacheControl) {
       continue;
     }
     if (msg.role === "assistant") {
+      if (typeof msg.content === "string" && msg.content === "") continue;
       const content = typeof msg.content === "string" ? msg.content : toAnthropicAssistantContent(msg.content, msgIdx > trajectoryStartIdx, idMap);
       if (Array.isArray(content) && content.length === 0) continue;
       out.push({ role: "assistant", content });
@@ -876,12 +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 {
@@ -907,11 +949,12 @@ function remapToolCallId(id, idMap) {
   if (!id.startsWith("toolu_")) return id;
   const existing = idMap.get(id);
   if (existing) return existing;
-  const mapped = `call_${id.slice(5)}`;
+  const mapped = `call_${id.slice(6)}`;
   idMap.set(id, mapped);
   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";
@@ -978,9 +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;
@@ -1058,6 +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")) {
@@ -1518,6 +1565,12 @@ async function* runStream(options) {
       statusCode: 504
     });
   }
+  if (stopReason === null) {
+    throw new ProviderError("anthropic", "Stream ended before completion (no stop_reason).", {
+      statusCode: 504,
+      cause: { partialContent: contentParts, outputTokens }
+    });
+  }
   const normalizedStop = normalizeAnthropicStopReason(stopReason);
   const response = {
     message: {
@@ -1784,6 +1837,17 @@ function getEnvironment() {
 }
 
 // src/providers/openai.ts
+function toKimiK3Effort(level) {
+  switch (level) {
+    case "low":
+      return "low";
+    case "medium":
+    case "high":
+      return "high";
+    default:
+      return "max";
+  }
+}
 function extractOpenAIUsage(usage) {
   let cacheRead = 0;
   let cacheWrite = 0;
@@ -1837,9 +1901,12 @@ 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;
   const isKimiK27 = options.provider === "moonshot" && options.model.startsWith("kimi-k2.7-code");
   const hasFixedKimiSampling = isKimiK3 || isKimiK27;
   const usesThinkingParam = options.provider === "glm" || options.provider === "moonshot" && !isKimiK3 && !isKimiK27 || options.provider === "xiaomi";
@@ -1854,10 +1921,13 @@ async function* runStream2(options) {
   }
   const messages = toOpenAIMessages(downgradedMessages, {
     provider: options.provider,
-    // K3 and K2.7 preserve reasoning even when the user hides thinking in the
-    // UI; keep assistant tool-call history wire-valid in that display mode.
-    thinking: isKimiK3 || isKimiK27 || !!options.thinking,
-    supportsImages: options.supportsImages
+    // K2.7 preserves reasoning even when the user hides thinking in the UI;
+    // keep assistant tool-call history wire-valid in that display mode. A
+    // 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,
+    reasoningField: getReasoningField(endpointKey)
   });
   const defaultTemp = options.provider === "glm" ? 0.6 : void 0;
   const effectiveTemp = options.temperature ?? defaultTemp;
@@ -1869,7 +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 } } : {}
@@ -1883,15 +1953,22 @@ 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;
   }
   if (isKimiK3) {
     const paramsAny = params;
     if (isManagedKimiK3) {
-      paramsAny.thinking = { type: "enabled", effort: "max", keep: "all" };
+      paramsAny.thinking = k3Effort ? { type: "enabled", effort: k3Effort, keep: "all" } : { type: "disabled" };
+    } else if (k3Effort) {
+      paramsAny.reasoning_effort = k3Effort;
     } else {
-      paramsAny.reasoning_effort = "max";
+      paramsAny.thinking = { type: "disabled" };
     }
   }
   if (usesThinkingParam) {
@@ -1917,8 +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);
     }
@@ -1953,11 +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) {
@@ -1997,6 +2075,12 @@ async function* runStream2(options) {
       statusCode: 504
     });
   }
+  if (finishReason === null) {
+    throw new ProviderError(providerName, "Stream ended before completion (no finish_reason).", {
+      statusCode: 504,
+      cause: { partialText: textAccum, outputTokens }
+    });
+  }
   if (thinkingAccum) {
     contentParts.push({ type: "thinking", text: thinkingAccum });
   }
@@ -2036,16 +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 };
@@ -2073,15 +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;
@@ -3442,6 +3528,28 @@ 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) {
@@ -3845,6 +3953,7 @@ export {
   formatErrorForDisplay,
   isHardBillingMessage,
   isUsageLimitError,
+  localWireModelId,
   palsuAssistantMessage,
   palsuText,
   palsuThinking,