@kenkaiiii/gg-ai 5.71.0 → 5.72.0

package/dist/index.d.cts

@@ -223,6 +223,15 @@ interface StreamOptions {
     promptCacheKey?: string;
     /** OpenAI service tier for latency-sensitive requests. Only sent to first-party OpenAI API calls. */
     serviceTier?: "auto" | "default" | "flex" | "priority";
+    /** Codex endpoint only: send the Responses-Lite request shape (lite header,
+     *  `parallel_tool_calls: false`, all-turns reasoning context). Unset follows
+     *  the model family (on for gpt-5.6/gpt-6 models). Lite allows only one tool
+     *  call per response, so turning it off lets the model batch calls. */
+    responsesLite?: boolean;
+    /** OpenAI only (Codex and API-key routes): send tools with strict
+     *  (grammar-constrained) schemas. Default true. Strict calls carry every
+     *  optional field as null, which costs output tokens and latency per call. */
+    strictTools?: boolean;
     /** OpenAI ChatGPT account ID (from OAuth JWT) for codex endpoint */
     accountId?: string;
     /** Stable conversation identity for Codex transport headers. This is distinct from
@@ -685,6 +694,8 @@ declare function toOpenAIMessages(messages: Message[], options?: {
     reasoningField?: string;
 }): OpenAI.ChatCompletionMessageParam[];
 
+declare function usesResponsesLite(model: string): boolean;
+
 /**
  * Fire a minimal `max_tokens: 1` request that populates the Anthropic prompt
  * cache with the system prompt + tools prefix, so the first real user turn is
@@ -776,4 +787,4 @@ 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, 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 };
+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, usesResponsesLite };

package/dist/index.d.ts

@@ -223,6 +223,15 @@ interface StreamOptions {
     promptCacheKey?: string;
     /** OpenAI service tier for latency-sensitive requests. Only sent to first-party OpenAI API calls. */
     serviceTier?: "auto" | "default" | "flex" | "priority";
+    /** Codex endpoint only: send the Responses-Lite request shape (lite header,
+     *  `parallel_tool_calls: false`, all-turns reasoning context). Unset follows
+     *  the model family (on for gpt-5.6/gpt-6 models). Lite allows only one tool
+     *  call per response, so turning it off lets the model batch calls. */
+    responsesLite?: boolean;
+    /** OpenAI only (Codex and API-key routes): send tools with strict
+     *  (grammar-constrained) schemas. Default true. Strict calls carry every
+     *  optional field as null, which costs output tokens and latency per call. */
+    strictTools?: boolean;
     /** OpenAI ChatGPT account ID (from OAuth JWT) for codex endpoint */
     accountId?: string;
     /** Stable conversation identity for Codex transport headers. This is distinct from
@@ -685,6 +694,8 @@ declare function toOpenAIMessages(messages: Message[], options?: {
     reasoningField?: string;
 }): OpenAI.ChatCompletionMessageParam[];
 
+declare function usesResponsesLite(model: string): boolean;
+
 /**
  * Fire a minimal `max_tokens: 1` request that populates the Anthropic prompt
  * cache with the system prompt + tools prefix, so the first real user turn is
@@ -776,4 +787,4 @@ 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, 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 };
+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, usesResponsesLite };

package/dist/index.js

@@ -2352,7 +2352,7 @@ async function* runStream2(options) {
     ...options.thinking && !usesThinkingParam && !isKimiK3 && !isKimiK27 && !isLocal ? { reasoning_effort: toOpenAIReasoningEffort(options.thinking, options.model) } : {},
     ...options.tools?.length ? {
       tools: toOpenAITools(options.tools, {
-        strict: supportsStrictToolSampling(options.provider)
+        strict: supportsStrictToolSampling(options.provider) && (options.strictTools ?? true)
       })
     } : {},
     ...options.toolChoice && options.tools?.length ? { tool_choice: toOpenAIToolChoice(options.toolChoice) } : {},
@@ -2867,6 +2867,7 @@ async function* runStream3(options, retriedWithoutReasoning = false) {
   const downgraded = downgradeUnsupportedVideos(downgradedImages, options.supportsVideo);
   const { system, input } = toCodexInput(downgraded, { supportsImages: options.supportsImages });
   const responsesLite = usesResponsesLite(options.model);
+  const liteShape = options.responsesLite ?? responsesLite;
   const body = {
     model: options.model,
     store: false,
@@ -2874,11 +2875,11 @@ async function* runStream3(options, retriedWithoutReasoning = false) {
     instructions: system,
     input,
     tool_choice: toCodexToolChoice(options.toolChoice, options.tools),
-    parallel_tool_calls: !responsesLite,
+    parallel_tool_calls: !liteShape,
     include: ["reasoning.encrypted_content"]
   };
   if (options.tools?.length) {
-    body.tools = toCodexTools(options.tools);
+    body.tools = toCodexTools(options.tools, options.strictTools ?? true);
   }
   body.prompt_cache_key = normalizePromptCacheKey(options.promptCacheKey ?? "ggcoder");
   if (options.temperature != null && !options.thinking) {
@@ -2890,7 +2891,7 @@ async function* runStream3(options, retriedWithoutReasoning = false) {
     // `ultra` is a client orchestration preset, not a Codex API effort.
     effort: options.thinking === "ultra" ? "max" : options.thinking ?? (responsesLite ? "low" : "none"),
     summary: "auto",
-    ...responsesLite ? { context: "all_turns" } : {}
+    ...liteShape ? { context: "all_turns" } : {}
   };
   if (responsesLite) {
     body.text = { verbosity: "low" };
@@ -2902,10 +2903,8 @@ async function* runStream3(options, retriedWithoutReasoning = false) {
     "OpenAI-Beta": "responses=experimental",
     originator: responsesLite ? "codex_cli_rs" : "ggcoder",
     "User-Agent": responsesLite ? `codex_cli_rs/${CODEX_CLIENT_VERSION}` : `ggcoder (${os.platform()} ${os.release()}; ${os.arch()})`,
-    ...responsesLite ? {
-      version: CODEX_CLIENT_VERSION,
-      "X-OpenAI-Internal-Codex-Responses-Lite": "true"
-    } : {}
+    ...responsesLite ? { version: CODEX_CLIENT_VERSION } : {},
+    ...liteShape ? { "X-OpenAI-Internal-Codex-Responses-Lite": "true" } : {}
   };
   if (options.accountId) {
     headers["chatgpt-account-id"] = options.accountId;
@@ -3369,15 +3368,17 @@ function toCodexInput(messages, options) {
   }
   return { system, input };
 }
-function toCodexTools(tools) {
+function toCodexTools(tools, strictTools) {
   return tools.map((tool) => {
     let parameters = resolveToolSchema(tool);
     let strict = null;
-    try {
-      parameters = makeStrictToolSchema(parameters);
-      strict = true;
-    } catch (error) {
-      if (!(error instanceof UnsupportedStrictSchemaError)) throw error;
+    if (strictTools) {
+      try {
+        parameters = makeStrictToolSchema(parameters);
+        strict = true;
+      } catch (error) {
+        if (!(error instanceof UnsupportedStrictSchemaError)) throw error;
+      }
     }
     return {
       type: "function",
@@ -3614,6 +3615,17 @@ function stripUnsupportedSchemaFields(value) {
   }
   delete value.$schema;
   delete value.additionalProperties;
+  for (const [key, target, step] of [
+    ["exclusiveMinimum", "minimum", 1],
+    ["exclusiveMaximum", "maximum", -1]
+  ]) {
+    const bound = value[key];
+    if (bound === void 0) continue;
+    delete value[key];
+    if (typeof bound === "number" && value[target] === void 0) {
+      value[target] = value.type === "integer" ? bound + step : bound;
+    }
+  }
   for (const item of Object.values(value)) {
     if (isJsonObject(item) || Array.isArray(item)) {
       stripUnsupportedSchemaFields(item);
@@ -4751,6 +4763,7 @@ export {
   toAnthropicMessages,
   toOpenAIMessages,
   toWellFormedText,
-  toolCallNameRuleFor
+  toolCallNameRuleFor,
+  usesResponsesLite
 };
 //# sourceMappingURL=index.js.map