@kenkaiiii/gg-ai 5.68.1 → 5.70.0

package/dist/index.d.cts

@@ -594,6 +594,11 @@ 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`.
+ *
+ * A no-argument tool may declare a bare `{ "type": "object" }` root (common
+ * for MCP servers). OpenAI rejects that root with HTTP 400 "object schema
+ * missing properties", failing the whole request, so an object root without
+ * `properties` gets an explicit empty one. JSON Schema meaning is unchanged.
  */
 declare function resolveToolSchema(tool: Tool): JsonSchema;
 
@@ -604,9 +609,70 @@ declare function resolveToolSchema(tool: Tool): JsonSchema;
  * tool results are cloned for the outgoing request.
  */
 declare function clampProviderContextImages(messages: Message[], provider: Provider, supportsImages: boolean | undefined): Message[];
+/**
+ * What a provider accepts as a replayed tool-call name. `pattern` already
+ * encodes the length limit where the provider documents one; `maxLength` is
+ * checked separately so the generic rule can cap length without a charset.
+ */
+interface ToolCallNameRule {
+    /** Short identifier, for diagnostics and tests. */
+    id: "anthropic" | "openai-chat" | "openai-responses" | "gemini" | "generic";
+    maxLength: number;
+    pattern?: RegExp;
+}
+/**
+ * Per-transport tool-name rules. A legitimate call always carries the name of a
+ * tool we declared, and tool declarations are validated against these same
+ * rules, so only model-invented names (blank, invocation text stuffed into the
+ * name slot, …) can fail them.
+ */
+declare const TOOL_CALL_NAME_RULES: Record<ToolCallNameRule["id"], ToolCallNameRule>;
+/** True when `name` is a tool-call name the provider described by `rule` will accept. */
+declare function isValidToolCallName(name: unknown, rule: ToolCallNameRule): boolean;
+/**
+ * Pick the tool-name rule for the transport a request will actually use. Mirrors
+ * the routing in `stream.ts`: `openai` with an `accountId` is the Codex
+ * (Responses) endpoint, MiniMax rides the Anthropic transport but is a third
+ * party, and an `openai` pointed at a custom baseUrl is an unknown compatible
+ * server.
+ */
+declare function toolCallNameRuleFor(provider: Provider | string, options?: {
+    accountId?: string;
+    baseUrl?: string;
+}): ToolCallNameRule;
+/**
+ * Drop assistant tool calls whose name (or id) the target provider rejects,
+ * together with the tool results that answer them.
+ *
+ * One malformed call — a blank name, a whole invocation string written into the
+ * name slot, a name over the provider's length limit — would otherwise sit in
+ * history and 400 every later request (`tool_use.name: String should match
+ * pattern`, `string_above_max_length`, …), wedging the session.
+ *
+ * Pairing is positional per window: each assistant turn opens a window, and its
+ * results (role `tool`) are matched FIFO per id until the next assistant or user
+ * message, so a reused id never consumes another turn's result. Valid sibling
+ * calls, text and thinking are kept in place. An assistant turn left with
+ * nothing replayable is removed whole, as is a tool message left empty — every
+ * transport accepts the resulting adjacent user turns (the Anthropic API
+ * combines consecutive same-role turns). A Codex reasoning item orphaned by the
+ * drop is removed too. If the drop would leave the request ENDING on that
+ * pruned assistant turn (its only call was the bad one, so no results follow),
+ * the turn is removed as well: a trailing assistant message is an assistant
+ * prefill, which current Claude models reject and other APIs treat as a
+ * continuation rather than a fresh turn.
+ *
+ * Returns the input array untouched when nothing is malformed.
+ */
+declare function dropInvalidToolCalls(messages: Message[], rule: ToolCallNameRule): Message[];
 declare function toAnthropicMessages(messages: Message[], cacheControl?: {
     type: "ephemeral";
     ttl?: "1h";
+}, options?: {
+    /** Replay server-side refusal `fallback` blocks in place. Only set when the
+     * request carries the server-side-fallback beta; otherwise they are
+     * stripped (after applying the replay rules). Default false. */
+    fallbackBlocks?: boolean;
 }): {
     system: Anthropic.TextBlockParam[] | undefined;
     messages: Anthropic.MessageParam[];
@@ -710,4 +776,4 @@ interface PalsuProviderConfig {
  */
 declare function registerPalsuProvider(config?: PalsuProviderConfig): PalsuProviderHandle;
 
-export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, GGAIError, 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 PreparedContext, 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 };
+export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, GGAIError, 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 PreparedContext, 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, TOOL_CALL_NAME_RULES, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolCallNameRule, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, clampProviderContextImages, classifyProviderError, dropInvalidToolCalls, environmentSecrets, formatError, formatErrorForDisplay, hasLoneSurrogate, isHardBillingMessage, isUsageLimitError, isValidToolCallName, localWireModelId, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, resolveToolSchema, sanitizeMessagesForWire, setProviderDiagnostic, sliceHead, sliceTail, stream, toAnthropicMessages, toOpenAIMessages, toWellFormedText, toolCallNameRuleFor };

package/dist/index.d.ts

@@ -594,6 +594,11 @@ 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`.
+ *
+ * A no-argument tool may declare a bare `{ "type": "object" }` root (common
+ * for MCP servers). OpenAI rejects that root with HTTP 400 "object schema
+ * missing properties", failing the whole request, so an object root without
+ * `properties` gets an explicit empty one. JSON Schema meaning is unchanged.
  */
 declare function resolveToolSchema(tool: Tool): JsonSchema;
 
@@ -604,9 +609,70 @@ declare function resolveToolSchema(tool: Tool): JsonSchema;
  * tool results are cloned for the outgoing request.
  */
 declare function clampProviderContextImages(messages: Message[], provider: Provider, supportsImages: boolean | undefined): Message[];
+/**
+ * What a provider accepts as a replayed tool-call name. `pattern` already
+ * encodes the length limit where the provider documents one; `maxLength` is
+ * checked separately so the generic rule can cap length without a charset.
+ */
+interface ToolCallNameRule {
+    /** Short identifier, for diagnostics and tests. */
+    id: "anthropic" | "openai-chat" | "openai-responses" | "gemini" | "generic";
+    maxLength: number;
+    pattern?: RegExp;
+}
+/**
+ * Per-transport tool-name rules. A legitimate call always carries the name of a
+ * tool we declared, and tool declarations are validated against these same
+ * rules, so only model-invented names (blank, invocation text stuffed into the
+ * name slot, …) can fail them.
+ */
+declare const TOOL_CALL_NAME_RULES: Record<ToolCallNameRule["id"], ToolCallNameRule>;
+/** True when `name` is a tool-call name the provider described by `rule` will accept. */
+declare function isValidToolCallName(name: unknown, rule: ToolCallNameRule): boolean;
+/**
+ * Pick the tool-name rule for the transport a request will actually use. Mirrors
+ * the routing in `stream.ts`: `openai` with an `accountId` is the Codex
+ * (Responses) endpoint, MiniMax rides the Anthropic transport but is a third
+ * party, and an `openai` pointed at a custom baseUrl is an unknown compatible
+ * server.
+ */
+declare function toolCallNameRuleFor(provider: Provider | string, options?: {
+    accountId?: string;
+    baseUrl?: string;
+}): ToolCallNameRule;
+/**
+ * Drop assistant tool calls whose name (or id) the target provider rejects,
+ * together with the tool results that answer them.
+ *
+ * One malformed call — a blank name, a whole invocation string written into the
+ * name slot, a name over the provider's length limit — would otherwise sit in
+ * history and 400 every later request (`tool_use.name: String should match
+ * pattern`, `string_above_max_length`, …), wedging the session.
+ *
+ * Pairing is positional per window: each assistant turn opens a window, and its
+ * results (role `tool`) are matched FIFO per id until the next assistant or user
+ * message, so a reused id never consumes another turn's result. Valid sibling
+ * calls, text and thinking are kept in place. An assistant turn left with
+ * nothing replayable is removed whole, as is a tool message left empty — every
+ * transport accepts the resulting adjacent user turns (the Anthropic API
+ * combines consecutive same-role turns). A Codex reasoning item orphaned by the
+ * drop is removed too. If the drop would leave the request ENDING on that
+ * pruned assistant turn (its only call was the bad one, so no results follow),
+ * the turn is removed as well: a trailing assistant message is an assistant
+ * prefill, which current Claude models reject and other APIs treat as a
+ * continuation rather than a fresh turn.
+ *
+ * Returns the input array untouched when nothing is malformed.
+ */
+declare function dropInvalidToolCalls(messages: Message[], rule: ToolCallNameRule): Message[];
 declare function toAnthropicMessages(messages: Message[], cacheControl?: {
     type: "ephemeral";
     ttl?: "1h";
+}, options?: {
+    /** Replay server-side refusal `fallback` blocks in place. Only set when the
+     * request carries the server-side-fallback beta; otherwise they are
+     * stripped (after applying the replay rules). Default false. */
+    fallbackBlocks?: boolean;
 }): {
     system: Anthropic.TextBlockParam[] | undefined;
     messages: Anthropic.MessageParam[];
@@ -710,4 +776,4 @@ interface PalsuProviderConfig {
  */
 declare function registerPalsuProvider(config?: PalsuProviderConfig): PalsuProviderHandle;
 
-export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, GGAIError, 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 PreparedContext, 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 };
+export { type AssistantMessage, type CacheRetention, type ContentPart, type DoneEvent, type ErrorEvent, type ErrorSource, EventStream, type FormattedError, GGAIError, 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 PreparedContext, 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, TOOL_CALL_NAME_RULES, type TextContent, type TextDeltaEvent, type ThinkingContent, type ThinkingDeltaEvent, type ThinkingLevel, type Tool, type ToolCall, type ToolCallDeltaEvent, type ToolCallDoneEvent, type ToolCallNameRule, type ToolChoice, type ToolResult, type ToolResultContent, type ToolResultMessage, type Usage, type UserMessage, type VideoContent, clampProviderContextImages, classifyProviderError, dropInvalidToolCalls, environmentSecrets, formatError, formatErrorForDisplay, hasLoneSurrogate, isHardBillingMessage, isUsageLimitError, isValidToolCallName, localWireModelId, palsuAssistantMessage, palsuText, palsuThinking, palsuToolCall, prewarmAnthropicCache, providerRegistry, redactText, redactValue, registerPalsuProvider, resolveToolSchema, sanitizeMessagesForWire, setProviderDiagnostic, sliceHead, sliceTail, stream, toAnthropicMessages, toOpenAIMessages, toWellFormedText, toolCallNameRuleFor };

package/dist/index.js

@@ -448,7 +448,11 @@ function zodToJsonSchema(schema) {
   return normalized;
 }
 function resolveToolSchema(tool) {
-  return tool.rawInputSchema ?? zodToJsonSchema(tool.parameters);
+  const schema = tool.rawInputSchema ?? zodToJsonSchema(tool.parameters);
+  if (schema.type === "object" && schema.properties === void 0) {
+    return { ...schema, properties: {} };
+  }
+  return schema;
 }
 function normalizeRootForAnthropic(schema) {
   const branches = schema.oneOf ?? schema.anyOf;
@@ -655,6 +659,10 @@ var ANTHROPIC_INPUT_BLOCK_TYPES = /* @__PURE__ */ new Set([
   "connector_text",
   "container_upload",
   "document",
+  // Server-side refusal fallback marker. Only reaches the wire when the request
+  // carries the server-side-fallback beta; otherwise applyServerFallbackReplay
+  // strips it first (see toAnthropicMessages' `fallbackBlocks` option).
+  "fallback",
   "image",
   "mid_conv_system",
   "redacted_thinking",
@@ -676,6 +684,42 @@ function isPositionSensitiveThinking(part) {
   if (part.type === "thinking") return hasValidThinkingSignature(part);
   return isRawThinking(part);
 }
+function isServerFallbackBlock(part) {
+  return part.type === "raw" && part.data.type === "fallback";
+}
+function applyServerFallbackReplay(content, keepMarkers, droppedToolCallIds) {
+  let lastFallbackIdx = -1;
+  content.forEach((part, idx) => {
+    if (isServerFallbackBlock(part)) lastFallbackIdx = idx;
+  });
+  if (lastFallbackIdx === -1) return content;
+  const resultIds = /* @__PURE__ */ new Set();
+  for (const part of content) {
+    if (part.type === "server_tool_result") resultIds.add(part.toolUseId);
+    else if (part.type === "raw" && typeof part.data.tool_use_id === "string")
+      resultIds.add(part.data.tool_use_id);
+  }
+  const out = [];
+  content.forEach((part, idx) => {
+    if (isServerFallbackBlock(part)) {
+      if (keepMarkers) out.push(part);
+      return;
+    }
+    if (idx < lastFallbackIdx) {
+      if (part.type === "thinking" || isRawThinking(part)) return;
+      if (part.type === "raw" && part.data.type === "connector_text") return;
+      if (part.type === "tool_call") {
+        droppedToolCallIds?.add(part.id);
+        return;
+      }
+      if (part.type === "server_tool_call" && !resultIds.has(part.id)) return;
+      if (part.type === "raw" && part.data.type === "server_tool_use" && !resultIds.has(String(part.data.id)))
+        return;
+    }
+    out.push(part);
+  });
+  return out.every(isServerFallbackBlock) ? [] : out;
+}
 function toAnthropicAssistantPart(part, idMap) {
   if (part.type === "text") return { type: "text", text: part.text };
   if (part.type === "thinking") {
@@ -787,6 +831,144 @@ function clampProviderContextImages(messages, provider, supportsImages) {
     return message;
   });
 }
+var TOOL_CALL_NAME_RULES = {
+  // Anthropic Messages API: tool / tool_use name `^[a-zA-Z0-9_-]{1,128}$`.
+  anthropic: { id: "anthropic", maxLength: 128, pattern: /^[a-zA-Z0-9_-]{1,128}$/ },
+  // OpenAI Chat Completions: function name a-z A-Z 0-9 _ -, max length 64.
+  "openai-chat": { id: "openai-chat", maxLength: 64, pattern: /^[a-zA-Z0-9_-]{1,64}$/ },
+  // OpenAI Responses (Codex): same charset; replayed `input[N].name` over 128
+  // chars is rejected with `string_above_max_length`.
+  "openai-responses": {
+    id: "openai-responses",
+    maxLength: 128,
+    pattern: /^[a-zA-Z0-9_-]{1,128}$/
+  },
+  // Gemini FunctionDeclaration name: starts with a letter or underscore, then
+  // a-z A-Z 0-9 _ . : -, max length 128.
+  gemini: { id: "gemini", maxLength: 128, pattern: /^[a-zA-Z_][a-zA-Z0-9_.:-]{0,127}$/ },
+  // OpenAI-compatible third parties (GLM, Kimi, DeepSeek, OpenRouter, xAI,
+  // local servers, MiniMax, …) don't share one documented charset, so only
+  // reject what no declared tool name can contain: blank, >128 chars, or any
+  // whitespace / control character (the signature of invocation text).
+  generic: { id: "generic", maxLength: 128 }
+};
+var TOOL_NAME_FORBIDDEN_CHARS = /[\s\p{Cc}]/u;
+function isValidToolCallName(name, rule) {
+  if (typeof name !== "string" || name.trim().length === 0) return false;
+  if (name.length > rule.maxLength) return false;
+  if (TOOL_NAME_FORBIDDEN_CHARS.test(name)) return false;
+  return rule.pattern ? rule.pattern.test(name) : true;
+}
+function isDefaultOrHost(baseUrl, host) {
+  if (!baseUrl) return true;
+  try {
+    return new URL(baseUrl).hostname === host;
+  } catch {
+    return false;
+  }
+}
+function toolCallNameRuleFor(provider, options) {
+  switch (provider) {
+    case "anthropic":
+      return TOOL_CALL_NAME_RULES.anthropic;
+    case "openai":
+      if (options?.accountId) return TOOL_CALL_NAME_RULES["openai-responses"];
+      return isDefaultOrHost(options?.baseUrl, "api.openai.com") ? TOOL_CALL_NAME_RULES["openai-chat"] : TOOL_CALL_NAME_RULES.generic;
+    case "gemini":
+      return TOOL_CALL_NAME_RULES.gemini;
+    default:
+      return TOOL_CALL_NAME_RULES.generic;
+  }
+}
+function isValidToolCallId(id) {
+  return typeof id === "string" && id.trim().length > 0;
+}
+function isRawReasoning(part) {
+  return part.type === "raw" && part.data.type === "reasoning";
+}
+function isDanglingReasoning(parts, idx) {
+  for (let i = idx + 1; i < parts.length; i++) {
+    const next = parts[i];
+    if (isRawReasoning(next)) return true;
+    if (next.type === "text" || next.type === "tool_call") return false;
+  }
+  return true;
+}
+function hasReplayableAssistantContent(parts) {
+  return parts.some((part) => {
+    if (part.type === "text") return part.text.length > 0;
+    if (part.type === "thinking") return false;
+    if (part.type === "raw") {
+      const t = part.data.type;
+      return !(t === "thinking" || t === "redacted_thinking" || t === "reasoning" || t === "fallback");
+    }
+    return true;
+  });
+}
+function dropInvalidToolCalls(messages, rule) {
+  const isInvalid = (part) => part.type === "tool_call" && (!isValidToolCallName(part.name, rule) || !isValidToolCallId(part.id));
+  const hasInvalid = messages.some(
+    (m) => m.role === "assistant" && Array.isArray(m.content) && m.content.some(isInvalid)
+  );
+  if (!hasInvalid) return messages;
+  const out = [];
+  let pending = /* @__PURE__ */ new Map();
+  let prunedAssistant = null;
+  for (const msg of messages) {
+    if (msg.role === "user" || msg.role === "system") {
+      if (msg.role === "user") pending = /* @__PURE__ */ new Map();
+      out.push(msg);
+      continue;
+    }
+    if (msg.role === "assistant") {
+      pending = /* @__PURE__ */ new Map();
+      if (typeof msg.content === "string") {
+        out.push(msg);
+        continue;
+      }
+      const original = msg.content;
+      let dropped = false;
+      const kept = [];
+      for (const part of original) {
+        if (part.type === "tool_call") {
+          const keep = !isInvalid(part);
+          const queue = pending.get(part.id) ?? [];
+          queue.push(keep);
+          pending.set(part.id, queue);
+          if (!keep) {
+            dropped = true;
+            continue;
+          }
+        }
+        kept.push(part);
+      }
+      if (!dropped) {
+        out.push(msg);
+        continue;
+      }
+      const content = kept.filter(
+        (part, idx) => !isRawReasoning(part) || !isDanglingReasoning(kept, idx) || isDanglingReasoning(original, original.indexOf(part))
+      );
+      if (hasReplayableAssistantContent(content)) {
+        const pruned = { ...msg, content };
+        out.push(pruned);
+        prunedAssistant = pruned;
+      }
+      continue;
+    }
+    let changed = false;
+    const results = msg.content.filter((result) => {
+      const queue = pending.get(result.toolCallId);
+      const keep = queue && queue.length > 0 ? queue.shift() : true;
+      if (!keep) changed = true;
+      return keep;
+    });
+    if (!changed) out.push(msg);
+    else if (results.length > 0) out.push({ ...msg, content: results });
+  }
+  if (prunedAssistant && out[out.length - 1] === prunedAssistant) out.pop();
+  return out;
+}
 var NON_VISION_USER_IMAGE_PLACEHOLDER = "(image omitted: model does not support images)";
 var NON_VISION_TOOL_IMAGE_PLACEHOLDER = "(tool image omitted: model does not support images)";
 var NON_VIDEO_USER_PLACEHOLDER = "(video omitted: model does not support video)";
@@ -893,10 +1075,12 @@ function remapAnthropicToolCallId(id, idMap) {
   idMap.set(id, mapped);
   return mapped;
 }
-function toAnthropicMessages(messages, cacheControl) {
+function toAnthropicMessages(messages, cacheControl, options) {
   let systemText;
   const out = [];
   const idMap = /* @__PURE__ */ new Map();
+  const keepFallbackBlocks = options?.fallbackBlocks === true;
+  const droppedToolCallIds = /* @__PURE__ */ new Set();
   const trajectoryStartIdx = messages.reduce(
     (last, m, i) => m.role === "user" ? i : last,
     -1
@@ -942,17 +1126,23 @@ function toAnthropicMessages(messages, cacheControl) {
     }
     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);
+      const content = typeof msg.content === "string" ? msg.content : toAnthropicAssistantContent(
+        applyServerFallbackReplay(msg.content, keepFallbackBlocks, droppedToolCallIds),
+        msgIdx > trajectoryStartIdx,
+        idMap
+      );
       if (Array.isArray(content) && content.length === 0) continue;
       out.push({ role: "assistant", content });
       continue;
     }
     if (msg.role === "tool") {
+      const results = droppedToolCallIds.size ? msg.content.filter((r) => !droppedToolCallIds.has(r.toolCallId)) : msg.content;
+      if (results.length === 0) continue;
       out.push({
         role: "user",
         // Cast covers the video block (used by the Anthropic-compatible MiniMax
         // API), which isn't in the first-party Anthropic tool_result types.
-        content: msg.content.map((result) => ({
+        content: results.map((result) => ({
           type: "tool_result",
           tool_use_id: remapAnthropicToolCallId(result.toolCallId, idMap),
           content: toAnthropicToolResultContent(result.content),
@@ -1292,6 +1482,48 @@ function fineGrainedToolStreamingEnabled() {
   const v = raw.trim().toLowerCase();
   return v === "1" || v === "true" || v === "yes" || v === "on";
 }
+var SERVER_FALLBACK_BETA = "server-side-fallback-2026-07-01";
+var anthropicServerFallback = {
+  disabled: /* @__PURE__ */ new Set(),
+  reset() {
+    this.disabled.clear();
+  }
+};
+function serverFallbackKey(options) {
+  const auth = options.apiKey?.startsWith("sk-ant-oat") ? "oauth" : "key";
+  return `${options.baseUrl ?? process.env.ANTHROPIC_BASE_URL ?? ""}|${auth}`;
+}
+function isDirectAnthropicApi(baseUrl) {
+  const effective = baseUrl ?? process.env.ANTHROPIC_BASE_URL;
+  if (!effective) return true;
+  try {
+    const url = new URL(effective);
+    return url.protocol === "https:" && url.hostname === "api.anthropic.com";
+  } catch {
+    return false;
+  }
+}
+function isServerFallbackRejection(err) {
+  const status = err?.status;
+  if (status !== 400) return false;
+  const message = err instanceof Error ? err.message : String(err);
+  return /fallbacks|server-side-fallback/i.test(message);
+}
+function sumUsageIterations(usage) {
+  const iterations = usage?.iterations;
+  if (!Array.isArray(iterations) || iterations.length === 0) return null;
+  const num = (v) => typeof v === "number" && Number.isFinite(v) ? v : 0;
+  const totals = { inputTokens: 0, outputTokens: 0, cacheRead: 0, cacheWrite: 0 };
+  for (const it of iterations) {
+    if (!it || typeof it !== "object") continue;
+    const rec = it;
+    totals.inputTokens += num(rec.input_tokens);
+    totals.outputTokens += num(rec.output_tokens);
+    totals.cacheRead += num(rec.cache_read_input_tokens);
+    totals.cacheWrite += num(rec.cache_creation_input_tokens);
+  }
+  return totals;
+}
 function createClient(options) {
   const isOAuth = options.apiKey?.startsWith("sk-ant-oat");
   const userAgent = isOAuth ? options.userAgent ?? "claude-cli/2.1.75 (external, cli)" : "";
@@ -1388,7 +1620,11 @@ async function* runStream(options) {
   const supportsFirstPartyToolExtras = !options.baseUrl || options.baseUrl.includes("api.anthropic.com");
   const downgradedImages = downgradeUnsupportedImages(options.messages, options.supportsImages);
   const downgradedMessages = downgradeUnsupportedVideos(downgradedImages, options.supportsVideo);
-  const { system: rawSystem, messages } = toAnthropicMessages(downgradedMessages, cacheControl);
+  const fallbackKey = serverFallbackKey(options);
+  const useServerFallback = options.provider !== "minimax" && isDirectAnthropicApi(options.baseUrl) && !anthropicServerFallback.disabled.has(fallbackKey);
+  const { system: rawSystem, messages } = toAnthropicMessages(downgradedMessages, cacheControl, {
+    fallbackBlocks: useServerFallback
+  });
   const system = isOAuth ? [
     {
       type: "text",
@@ -1458,6 +1694,7 @@ async function* runStream(options) {
       ];
       return contextEdits.length ? { context_management: { edits: contextEdits } } : {};
     })(),
+    ...useServerFallback ? { fallbacks: "default" } : {},
     stream: useStreaming
   };
   const hasAdaptiveThinking = isAdaptiveThinkingModel(options.model);
@@ -1474,20 +1711,41 @@ async function* runStream(options) {
     // it Anthropic silently ignores ttl:"1h" and falls back to the 5-min default,
     // so a pre-warmed cache expires before the user's first turn. cacheControl.ttl
     // is only "1h" on the first-party endpoint (see toAnthropicCacheControl).
-    ...cacheControl?.ttl === "1h" ? ["extended-cache-ttl-2025-04-11"] : []
+    ...cacheControl?.ttl === "1h" ? ["extended-cache-ttl-2025-04-11"] : [],
+    ...useServerFallback ? [SERVER_FALLBACK_BETA] : []
   ];
-  const requestOptions = {
+  const toRequestOptions = (betas) => ({
     signal: options.signal ?? void 0,
-    ...betaHeaders.length ? { headers: { "anthropic-beta": betaHeaders.join(",") } } : {}
+    ...betas.length ? { headers: { "anthropic-beta": betas.join(",") } } : {}
+  });
+  const requestOptions = toRequestOptions(betaHeaders);
+  const send = async (create) => {
+    try {
+      return await create(params, requestOptions);
+    } catch (err) {
+      if (!useServerFallback || !isServerFallbackRejection(err)) throw err;
+      anthropicServerFallback.disabled.add(fallbackKey);
+      const { fallbacks: _dropped, ...rest } = params;
+      const retryParams = {
+        ...rest,
+        messages: toAnthropicMessages(downgradedMessages, cacheControl, { fallbackBlocks: false }).messages
+      };
+      return create(
+        retryParams,
+        toRequestOptions(betaHeaders.filter((b) => b !== SERVER_FALLBACK_BETA))
+      );
+    }
   };
   if (!useStreaming) {
     try {
       const nonStreamingClient = client.withOptions({
         timeout: NON_STREAMING_REQUEST_TIMEOUT_MS
       });
-      const message = await nonStreamingClient.messages.create(
-        { ...params, stream: false },
-        requestOptions
+      const message = await send(
+        (p, o) => nonStreamingClient.messages.create(
+          { ...p, stream: false },
+          o
+        )
       );
       yield* synthesizeEventsFromMessage(message);
       return messageToResponse(message);
@@ -1505,9 +1763,8 @@ async function* runStream(options) {
   const keepalive = { type: "keepalive" };
   let receivedAnyEvent = false;
   try {
-    const stream2 = await client.messages.create(
-      params,
-      requestOptions
+    const stream2 = await send(
+      (p, o) => client.messages.create(p, o)
     );
     for await (const event of stream2) {
       receivedAnyEvent = true;
@@ -1685,6 +1942,13 @@ async function* runStream(options) {
           if (usage?.output_tokens != null) {
             outputTokens = usage.output_tokens;
           }
+          const totals = sumUsageIterations(usage);
+          if (totals) {
+            inputTokens = totals.inputTokens;
+            outputTokens = totals.outputTokens;
+            cacheRead = totals.cacheRead;
+            cacheWrite = totals.cacheWrite;
+          }
           yield keepalive;
           break;
         }
@@ -1814,10 +2078,11 @@ function messageToResponse(message) {
     }
   }
   const usage = message.usage;
-  const inputTokens = usage.input_tokens ?? 0;
-  const outputTokens = usage.output_tokens ?? 0;
-  const cacheRead = usage.cache_read_input_tokens;
-  const cacheWrite = usage.cache_creation_input_tokens;
+  const totals = sumUsageIterations(usage);
+  const inputTokens = totals?.inputTokens ?? usage.input_tokens ?? 0;
+  const outputTokens = totals?.outputTokens ?? usage.output_tokens ?? 0;
+  const cacheRead = totals?.cacheRead ?? usage.cache_read_input_tokens;
+  const cacheWrite = totals?.cacheWrite ?? usage.cache_creation_input_tokens;
   return {
     message: {
       role: "assistant",
@@ -4025,15 +4290,23 @@ function stream(options) {
     throw new VideoUnsupportedError();
   }
   const wireMessages = stripMessageProvenance(options.messages);
-  const messages = clampProviderContextImages(
-    sanitizeMessagesForWire(wireMessages),
-    options.provider,
-    options.supportsImages
+  const sanitized = sanitizeMessagesForWire(wireMessages);
+  const replayable = dropInvalidToolCalls(
+    sanitized,
+    toolCallNameRuleFor(options.provider, {
+      accountId: options.accountId,
+      baseUrl: options.baseUrl
+    })
   );
+  const messages = clampProviderContextImages(replayable, options.provider, options.supportsImages);
   if (options.onContextPrepared) {
     try {
       options.onContextPrepared(
-        observePreparedContext(wireMessages, messages, options.tools ?? [])
+        observePreparedContext(
+          replayable === sanitized ? wireMessages : replayable,
+          messages,
+          options.tools ?? []
+        )
       );
     } catch {
     }
@@ -4448,14 +4721,17 @@ export {
   ProviderError,
   REDACTED as REDACTION_MARKER,
   StreamResult,
+  TOOL_CALL_NAME_RULES,
   clampProviderContextImages,
   classifyProviderError,
+  dropInvalidToolCalls,
   environmentSecrets,
   formatError,
   formatErrorForDisplay,
   hasLoneSurrogate,
   isHardBillingMessage,
   isUsageLimitError,
+  isValidToolCallName,
   localWireModelId,
   palsuAssistantMessage,
   palsuText,
@@ -4474,6 +4750,7 @@ export {
   stream,
   toAnthropicMessages,
   toOpenAIMessages,
-  toWellFormedText
+  toWellFormedText,
+  toolCallNameRuleFor
 };
 //# sourceMappingURL=index.js.map