@prestyj/ai 5.12.0 → 5.16.0

package/dist/index.d.cts

@@ -3,6 +3,8 @@ import Anthropic from '@anthropic-ai/sdk';
 import OpenAI from 'openai';
 
 type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "sakana" | "xai" | "palsu"
+/** Hugging Face Inference Providers router (OpenAI-compatible). */
+ | "huggingface"
 /** Locally hosted OpenAI-compatible server (Ollama, LM Studio, llama.cpp, vLLM). */
  | "local";
 type ThinkingLevel = "low" | "medium" | "high" | "xhigh" | "max" | "ultra";
@@ -530,6 +532,47 @@ type ProviderDiagnosticFn = (phase: string, data?: Record<string, unknown>) => v
 /** Register a diagnostic callback for provider-level tracing. */
 declare function setProviderDiagnostic(fn: ProviderDiagnosticFn | null): void;
 
+/**
+ * Converts a Zod schema to a JSON Schema object suitable for provider tool
+ * parameter definitions.
+ *
+ * Anthropic's `input_schema` validator is strict in two ways:
+ *
+ *   1. The root must be `type: "object"`. Returns 400 with
+ *      `tools.N.custom.input_schema.type: Field required` otherwise.
+ *
+ *   2. The root must NOT contain `oneOf`, `anyOf`, or `allOf`. Returns 400 with
+ *      `input_schema does not support oneOf, allOf, or anyOf at the top level`.
+ *
+ * Both rules trip whenever a tool's parameters are defined via
+ * `z.discriminatedUnion(...)` or `z.union(...)` — Zod 4's
+ * `z.toJSONSchema` emits `{oneOf: [...]}` at the root with no `type`.
+ *
+ * The fix is to collapse the union into a single flat object schema:
+ *
+ *   - properties = union of all branch properties (later branches win on
+ *     conflict; that's fine because the model only uses these for hints —
+ *     Zod's actual `tool.parameters.parse(args)` is the real validator)
+ *   - required = intersection of branch `required` arrays (a field is only
+ *     required if EVERY branch requires it)
+ *   - if the union has a discriminator field (every branch has the same
+ *     property as a `const`), we replace the discriminator's per-branch
+ *     `const` with an `enum` listing every literal — the model gets a clear
+ *     hint of the valid action values without needing oneOf
+ *
+ * The flattening is lossy for *schema-level* constraints (e.g. "if action=X,
+ * then field Y is required") — Zod still enforces those at parse time. For
+ * the model's purposes this is identical to a single object with optional
+ * fields and a discriminator enum, which is exactly how Anthropic-supported
+ * tools are typically authored anyway.
+ */
+type JsonSchema = Record<string, unknown>;
+/**
+ * Resolve a tool's JSON Schema for provider tool definitions: prefer the
+ * tool's pre-built `rawInputSchema`, otherwise convert its Zod `parameters`.
+ */
+declare function resolveToolSchema(tool: Tool): JsonSchema;
+
 /**
  * Cap historical images before provider dispatch, removing the oldest first.
  * The persisted/live conversation is never mutated; only modified messages and
@@ -642,4 +685,4 @@ 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 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 };
+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, resolveToolSchema, sanitizeMessagesForWire, setProviderDiagnostic, sliceHead, sliceTail, stream, toAnthropicMessages, toOpenAIMessages, toWellFormedText };

package/dist/index.d.ts

@@ -3,6 +3,8 @@ import Anthropic from '@anthropic-ai/sdk';
 import OpenAI from 'openai';
 
 type Provider = "anthropic" | "xiaomi" | "openai" | "gemini" | "glm" | "moonshot" | "minimax" | "deepseek" | "openrouter" | "sakana" | "xai" | "palsu"
+/** Hugging Face Inference Providers router (OpenAI-compatible). */
+ | "huggingface"
 /** Locally hosted OpenAI-compatible server (Ollama, LM Studio, llama.cpp, vLLM). */
  | "local";
 type ThinkingLevel = "low" | "medium" | "high" | "xhigh" | "max" | "ultra";
@@ -530,6 +532,47 @@ type ProviderDiagnosticFn = (phase: string, data?: Record<string, unknown>) => v
 /** Register a diagnostic callback for provider-level tracing. */
 declare function setProviderDiagnostic(fn: ProviderDiagnosticFn | null): void;
 
+/**
+ * Converts a Zod schema to a JSON Schema object suitable for provider tool
+ * parameter definitions.
+ *
+ * Anthropic's `input_schema` validator is strict in two ways:
+ *
+ *   1. The root must be `type: "object"`. Returns 400 with
+ *      `tools.N.custom.input_schema.type: Field required` otherwise.
+ *
+ *   2. The root must NOT contain `oneOf`, `anyOf`, or `allOf`. Returns 400 with
+ *      `input_schema does not support oneOf, allOf, or anyOf at the top level`.
+ *
+ * Both rules trip whenever a tool's parameters are defined via
+ * `z.discriminatedUnion(...)` or `z.union(...)` — Zod 4's
+ * `z.toJSONSchema` emits `{oneOf: [...]}` at the root with no `type`.
+ *
+ * The fix is to collapse the union into a single flat object schema:
+ *
+ *   - properties = union of all branch properties (later branches win on
+ *     conflict; that's fine because the model only uses these for hints —
+ *     Zod's actual `tool.parameters.parse(args)` is the real validator)
+ *   - required = intersection of branch `required` arrays (a field is only
+ *     required if EVERY branch requires it)
+ *   - if the union has a discriminator field (every branch has the same
+ *     property as a `const`), we replace the discriminator's per-branch
+ *     `const` with an `enum` listing every literal — the model gets a clear
+ *     hint of the valid action values without needing oneOf
+ *
+ * The flattening is lossy for *schema-level* constraints (e.g. "if action=X,
+ * then field Y is required") — Zod still enforces those at parse time. For
+ * the model's purposes this is identical to a single object with optional
+ * fields and a discriminator enum, which is exactly how Anthropic-supported
+ * tools are typically authored anyway.
+ */
+type JsonSchema = Record<string, unknown>;
+/**
+ * Resolve a tool's JSON Schema for provider tool definitions: prefer the
+ * tool's pre-built `rawInputSchema`, otherwise convert its Zod `parameters`.
+ */
+declare function resolveToolSchema(tool: Tool): JsonSchema;
+
 /**
  * Cap historical images before provider dispatch, removing the oldest first.
  * The persisted/live conversation is never mutated; only modified messages and
@@ -642,4 +685,4 @@ 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 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 };
+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, resolveToolSchema, sanitizeMessagesForWire, setProviderDiagnostic, sliceHead, sliceTail, stream, toAnthropicMessages, toOpenAIMessages, toWellFormedText };

package/dist/index.js

@@ -59,6 +59,7 @@ var PROVIDER_DISPLAY = {
   openrouter: "OpenRouter",
   sakana: "Sakana",
   xai: "xAI (Grok)",
+  huggingface: "Hugging Face",
   xiaomi: "Xiaomi (MiMo)",
   minimax: "MiniMax"
 };
@@ -126,7 +127,7 @@ function formatError(err) {
         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 via the model selector \u2014 same underlying model, generally available."
+        guidance: "Request access via your Anthropic account team (see platform.claude.com/docs/en/about-claude/models/overview), or switch to Claude Fable 5.1 via the model selector \u2014 same underlying model, generally available."
       };
     }
     if (isUsageLimitError(err)) {
@@ -1105,6 +1106,9 @@ function toLocalReasoningEffort(level) {
   if (level === "max" || level === "ultra" || level === "xhigh") return "max";
   return level;
 }
+function toGlmReasoningEffort(level) {
+  return level === "ultra" ? "max" : level;
+}
 function toOpenAIReasoningEffort(level, model) {
   const effort = level === "max" || level === "ultra" ? "xhigh" : level;
   if (model.startsWith("fugu") && (effort === "low" || effort === "medium")) {
@@ -1551,7 +1555,15 @@ async function* runStream(options) {
           yield keepalive;
           break;
         }
-        // message_stop — loop exits naturally
+        // message_stop — loop exits naturally.
+        //
+        // Deliberately NOT breaking early here. Breaking makes the SDK iterator
+        // run `if (!done) controller.abort()` in its `finally`
+        // (core/streaming.js:97), which tears the connection down instead of
+        // returning it to the keep-alive pool — every turn would then pay a
+        // fresh TLS handshake. Draining to the end is what every other Anthropic
+        // client does, and the stall it guards against is handled by the agent
+        // loop's idle timeout.
         default:
           yield keepalive;
           break;
@@ -1974,6 +1986,11 @@ async function* runStream2(options) {
   if (usesThinkingParam) {
     if (options.thinking) {
       params.thinking = { type: "enabled" };
+      if (options.provider === "glm") {
+        params.reasoning_effort = toGlmReasoningEffort(
+          options.thinking
+        );
+      }
     } else {
       params.thinking = { type: "disabled" };
     }
@@ -2956,6 +2973,7 @@ var CODE_ASSIST_SUPPORTED_MODELS = /* @__PURE__ */ new Set([
   "gemini-3.5-flash",
   "gemini-3-flash",
   "gemini-3.1-flash-lite",
+  "gemini-3.7-flash",
   "gemini-2.5-pro",
   "gemini-2.5-flash",
   "gemma-4-31b-it",
@@ -2978,13 +2996,14 @@ var ACCOUNT_GATED_MODELS = /* @__PURE__ */ new Set([
   "gemini-3-flash",
   "gemini-3.5-flash",
   "gemini-3.1-pro-preview",
-  "gemini-3.1-pro-preview-customtools"
+  "gemini-3.1-pro-preview-customtools",
+  "gemini-3.7-flash"
 ]);
 function accountGatedMessage(model) {
   return `Your Google account isn't entitled to "${model}" over Gemini Code Assist OAuth, so the API reports it as not found. This is an account-access limit, not a ezcoder bug.`;
 }
 function accountGatedHint() {
-  return `Newer Gemini models (3.5 Flash, 3.1 Pro Preview) are available only to Code Assist Standard/Enterprise accounts with preview/GA access enabled by a cloud admin \u2014 free/personal accounts usually can't call them. Switch to Gemini 3.1 Flash Lite (it works on this account) with /model, or sign in with a Code Assist Standard/Enterprise account that has preview access.`;
+  return `Newer Gemini models (3.7 Flash, 3.5 Flash, 3.1 Pro Preview) are available only to Code Assist Standard/Enterprise accounts with preview/GA access enabled by a cloud admin \u2014 free/personal accounts usually can't call them. Switch to Gemini 3.1 Flash Lite (it works on this account) with /model, or sign in with a Code Assist Standard/Enterprise account that has preview access.`;
 }
 function formatErrorMessage(status, body, model) {
   if (status === 404 && !CODE_ASSIST_SUPPORTED_MODELS.has(model)) {
@@ -3668,6 +3687,18 @@ providerRegistry.register("openrouter", {
     baseUrl: options.baseUrl ?? "https://openrouter.ai/api/v1"
   })
 });
+providerRegistry.register("huggingface", {
+  // Hugging Face Inference Providers router — one HF token (hf.co/settings/tokens,
+  // "Make calls to Inference Providers" permission) routes to whichever hosted
+  // backend serves each open model. Chat Completions-compatible; model ids are
+  // Hub repo paths ("Qwen/Qwen3-Coder-480B-A35B-Instruct"), optionally with an
+  // ":auto"/":fastest"/":cheapest" provider-selection suffix. Billing follows
+  // each backend's per-token rates on the HF account (small free tier).
+  stream: (options) => streamOpenAI({
+    ...options,
+    baseUrl: options.baseUrl ?? "https://router.huggingface.co/v1"
+  })
+});
 providerRegistry.register("sakana", {
   // Sakana Fugu is a multi-agent system exposed as a standard LLM through the
   // OpenAI-compatible Sakana API. We ride the Chat Completions transport (the
@@ -4162,6 +4193,7 @@ export {
   redactText,
   redactValue,
   registerPalsuProvider,
+  resolveToolSchema,
   sanitizeMessagesForWire,
   setProviderDiagnostic,
   sliceHead,