@kenkaiiii/gg-ai 5.74.3 → 5.75.1

package/dist/index.js

@@ -443,10 +443,22 @@ function zodToJsonSchema(schema) {
   if (cached) return cached;
   const jsonSchema = z.toJSONSchema(schema);
   const { $schema: _schema, ...rest } = jsonSchema;
-  const normalized = normalizeRootForAnthropic(rest);
+  const normalized = normalizeRootForAnthropic(stripNoOpKeywords(rest));
   schemaCache.set(schema, normalized);
   return normalized;
 }
+function stripNoOpKeywords(node) {
+  if (Array.isArray(node)) return node.map(stripNoOpKeywords);
+  if (node === null || typeof node !== "object") return node;
+  const out = {};
+  for (const [key, value] of Object.entries(node)) {
+    if (key === "maximum" && value === Number.MAX_SAFE_INTEGER) continue;
+    if (key === "minimum" && value === Number.MIN_SAFE_INTEGER) continue;
+    if (key === "additionalProperties" && value === false) continue;
+    out[key] = stripNoOpKeywords(value);
+  }
+  return out;
+}
 function resolveToolSchema(tool) {
   const schema = tool.rawInputSchema ?? zodToJsonSchema(tool.parameters);
   if (schema.type === "object" && schema.properties === void 0) {
@@ -1221,16 +1233,23 @@ function toAnthropicToolChoice(choice) {
   return { type: "tool", name: choice.name };
 }
 function isAdaptiveThinkingModel(model) {
-  return /opus-5|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|haiku-5|fable-5|mythos-5/.test(model);
 }
 function toAnthropicThinking(level, maxTokens, model) {
   if (isAdaptiveThinkingModel(model)) {
     let effort = level;
-    if (effort === "xhigh" && !/opus-5|opus-4-8|opus-4-7|sonnet-5[-.]5/.test(model)) {
+    if (effort === "xhigh" && !/opus-5|opus-4-8|opus-4-7|sonnet-5[-.]5|haiku-5[-.]5/.test(model)) {
       effort = "high";
     }
     return {
-      thinking: { type: "adaptive" },
+      // "omitted" (what Claude Code sends): the model still thinks, but
+      // thinking blocks stream empty (signature only). The app never renders
+      // reasoning text, and "summarized" costs speed on Haiku 5.5: bench/h2h
+      // measured output tok/s over each whole request at 151 (summarized) vs
+      // 191 (omitted); Claude Code 195. Sonnet/Opus 5.5 measured even either way.
+      // The UI's "Thinking…" timer keys off the empty thinking_delta emitted
+      // when a thinking block opens, so it is unaffected.
+      thinking: { type: "adaptive", display: "omitted" },
       maxTokens,
       outputConfig: { effort }
     };
@@ -2792,7 +2811,7 @@ function extractRequestIdFromMessage(message) {
 
 // src/providers/openai-codex.ts
 var DEFAULT_BASE_URL = "https://chatgpt.com/backend-api";
-var CODEX_CLIENT_VERSION = "0.159.1";
+var CODEX_CLIENT_VERSION = "0.161.0";
 var CODEX_REQUEST_COMPRESSION_MIN_BYTES = 16 * 1024;
 var zstdInitPromise;
 async function encodeCodexRequest(body) {
@@ -2980,6 +2999,7 @@ async function* runStream3(options, retriedWithoutReasoning = false) {
   let outputTokens = 0;
   let cacheRead = 0;
   let cacheWrite = 0;
+  let thinkingTextEmitted = false;
   const diagStart = Date.now();
   const diagSeen = /* @__PURE__ */ new Set();
   for await (const event of parseSSE(response.body)) {
@@ -3057,9 +3077,17 @@ async function* runStream3(options, retriedWithoutReasoning = false) {
         }
       }
     }
+    if (type === "response.reasoning_summary_part.added") {
+      if (options.thinking && thinkingTextEmitted) {
+        yield { type: "thinking_delta", text: "\n\n" };
+      }
+    }
     if (type === "response.reasoning_summary_text.delta" || type === "response.reasoning_summary.delta" || type === "response.reasoning_text.delta" || type === "response.reasoning.delta") {
       const delta = event.delta;
-      if (options.thinking) yield { type: "thinking_delta", text: delta };
+      if (options.thinking && delta) {
+        thinkingTextEmitted = true;
+        yield { type: "thinking_delta", text: delta };
+      }
     }
     if (type === "response.output_item.added") {
       const item = event.item;
@@ -4220,7 +4248,7 @@ 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
+  // Hub repo paths ("moonshotai/Kimi-K2.7-Code"), 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({