@theokit/sdk 2.11.3 → 2.13.0

package/dist/{cron-BRfFPEKY.d.ts → cron-CL_9nfhQ.d.ts}

@@ -760,6 +760,17 @@ interface ProviderRoute {
     capability: ProviderCapability;
     provider: string;
     model?: string;
+    /**
+     * Opt-in leaked-dialect safe-parse for this route's provider (theokit#58
+     * follow-up). When `true`, a `chat_completions` finish that carries ZERO
+     * native `tool_calls` has its assistant text scanned for the Hermes
+     * `<function=…></tool_call>` dialect, and any recovered calls are surfaced as
+     * real `tool_calls` so the loop executes them — for models (qwen3-coder via
+     * OpenRouter) that intermittently leak tool calls as text. Default `false`;
+     * fail-open (a partial/unclosed block never fabricates a call). Scoped to the
+     * resolved chat chain, so a non-leaking route is unaffected.
+     */
+    extractToolCallsFromContent?: boolean;
 }
 /**
  * Provider routing configuration accepted by `Agent.create()` via
@@ -873,6 +884,13 @@ interface ProviderProfile {
     fallbackModels: ReadonlyArray<string>;
     extraHeaders?: Record<string, string>;
     bodyOverrides?: Record<string, unknown>;
+    /**
+     * Opt-in leaked-dialect safe-parse (theokit#58 follow-up). When `true`, a chat_completions finish
+     * with ZERO native `tool_calls` has its assistant content scanned for the Hermes
+     * `<function=…></tool_call>` dialect and any recovered calls are surfaced as real `tool_calls`.
+     * Default off — only enable for routes/models known to leak (e.g. a qwen3-coder profile variant).
+     */
+    extractToolCallsFromContent?: boolean;
 }
 
 type HookName = "pre_tool_call" | "post_tool_call" | "pre_llm_call" | "post_llm_call" | "on_session_start" | "on_session_end" | "transform_tool_result" | "transform_llm_output" | "pre_user_send" | "post_assistant_reply";

package/dist/{cron-DOw-Hqol.d.cts → cron-CpxLdAXc.d.cts}

@@ -760,6 +760,17 @@ interface ProviderRoute {
     capability: ProviderCapability;
     provider: string;
     model?: string;
+    /**
+     * Opt-in leaked-dialect safe-parse for this route's provider (theokit#58
+     * follow-up). When `true`, a `chat_completions` finish that carries ZERO
+     * native `tool_calls` has its assistant text scanned for the Hermes
+     * `<function=…></tool_call>` dialect, and any recovered calls are surfaced as
+     * real `tool_calls` so the loop executes them — for models (qwen3-coder via
+     * OpenRouter) that intermittently leak tool calls as text. Default `false`;
+     * fail-open (a partial/unclosed block never fabricates a call). Scoped to the
+     * resolved chat chain, so a non-leaking route is unaffected.
+     */
+    extractToolCallsFromContent?: boolean;
 }
 /**
  * Provider routing configuration accepted by `Agent.create()` via
@@ -873,6 +884,13 @@ interface ProviderProfile {
     fallbackModels: ReadonlyArray<string>;
     extraHeaders?: Record<string, string>;
     bodyOverrides?: Record<string, unknown>;
+    /**
+     * Opt-in leaked-dialect safe-parse (theokit#58 follow-up). When `true`, a chat_completions finish
+     * with ZERO native `tool_calls` has its assistant content scanned for the Hermes
+     * `<function=…></tool_call>` dialect and any recovered calls are surfaced as real `tool_calls`.
+     * Default off — only enable for routes/models known to leak (e.g. a qwen3-coder profile variant).
+     */
+    extractToolCallsFromContent?: boolean;
 }
 
 type HookName = "pre_tool_call" | "post_tool_call" | "pre_llm_call" | "post_llm_call" | "on_session_start" | "on_session_end" | "transform_tool_result" | "transform_llm_output" | "pre_user_send" | "post_assistant_reply";

package/dist/cron.cjs

@@ -10956,6 +10956,35 @@ function toOllamaTools(tools) {
   }));
 }
 
+// src/internal/llm/hermes-tool-extract.ts
+var HERMES_BLOCK = /<function=\s*([^>\s]+)\s*>([\s\S]*?)<\/tool_call>/g;
+var HERMES_PARAM = /<parameter=\s*([^>\s]+)\s*>([\s\S]*?)<\/parameter>/g;
+function extractHermesToolCalls(content, makeId) {
+  const toolCalls = [];
+  for (const block of content.matchAll(HERMES_BLOCK)) {
+    const name = (block[1] ?? "").trim();
+    if (name.length === 0) continue;
+    toolCalls.push({
+      type: "tool_use",
+      id: makeId(),
+      name,
+      input: parseHermesParams(block[2] ?? "")
+    });
+  }
+  const residualText = toolCalls.length === 0 ? content : content.replace(HERMES_BLOCK, "").trim();
+  return { toolCalls, residualText };
+}
+function parseHermesParams(inner) {
+  const input = {};
+  for (const param of inner.matchAll(HERMES_PARAM)) {
+    const key = param[1];
+    const value = param[2];
+    if (key === void 0 || value === void 0) continue;
+    input[key.trim()] = value;
+  }
+  return input;
+}
+
 // src/internal/llm/openai.ts
 var OpenAIClient = class {
   constructor(options) {
@@ -10967,6 +10996,14 @@ var OpenAIClient = class {
   name = "openai";
   baseUrl;
   fetchImpl;
+  /**
+   * Whether this client recovers leaked Hermes tool-call dialect from assistant
+   * text (theokit#58 follow-up). Observability for the route→client wiring of
+   * `extractToolCallsFromContent`. @internal
+   */
+  get recoversLeakedToolCalls() {
+    return this.options.extractToolCallsFromContent ?? false;
+  }
   // biome-ignore lint/complexity/noExcessiveCognitiveComplexity: HTTP+SSE handshake + accumulator is intentionally one block
   async *stream(request, signal) {
     const headers = {
@@ -11017,7 +11054,10 @@ var OpenAIClient = class {
         endpoint: "/v1/chat/completions"
       });
     }
-    const accumulator = new OpenAIStreamAccumulator();
+    const accumulator = new OpenAIStreamAccumulator(
+      this.options.extractToolCallsFromContent ?? false,
+      providerId
+    );
     for await (const record of parseSseStream(response.body, signal)) {
       if (record.data === "[DONE]") break;
       let chunk;
@@ -11043,6 +11083,16 @@ var OpenAIClient = class {
   }
 };
 var OpenAIStreamAccumulator = class {
+  /**
+   * @param extractFromContent opt-in leaked-dialect safe-parse (theokit#58). Default false.
+   * @param providerName provider id, used only to label the recovery log line.
+   */
+  constructor(extractFromContent = false, providerName = "openai") {
+    this.extractFromContent = extractFromContent;
+    this.providerName = providerName;
+  }
+  extractFromContent;
+  providerName;
   text = "";
   stopReason = "end_turn";
   inputTokens;
@@ -11108,9 +11158,26 @@ var OpenAIStreamAccumulator = class {
       const input = parseToolArguments(call.args);
       toolCalls.push({ type: "tool_use", id: call.id, name: call.name, input });
     }
+    let text = this.text;
+    let stopReason = this.stopReason;
+    if (this.extractFromContent && toolCalls.length === 0) {
+      const recovered = extractHermesToolCalls(
+        this.text,
+        () => `hermes-${globalThis.crypto.randomUUID()}`
+      );
+      if (recovered.toolCalls.length > 0) {
+        toolCalls.push(...recovered.toolCalls);
+        text = recovered.residualText;
+        stopReason = "tool_use";
+        process.stderr.write(
+          `[theokit-sdk] recovered ${recovered.toolCalls.length} leaked tool call(s) from assistant content (provider="${this.providerName}", names=${recovered.toolCalls.map((c) => c.name).join(",")})
+`
+        );
+      }
+    }
     return makeLlmFinish({
-      stopReason: this.stopReason,
-      text: this.text,
+      stopReason,
+      text,
       toolCalls,
       inputTokens: this.inputTokens,
       outputTokens: this.outputTokens,
@@ -11587,8 +11654,9 @@ function buildChain(options) {
   return clients;
 }
 function buildClient(name, routerOptions) {
-  const profile = getProviderProfile(name);
-  if (profile === void 0) return void 0;
+  const baseProfile = getProviderProfile(name);
+  if (baseProfile === void 0) return void 0;
+  const profile = routerOptions.extractToolCallsFromContent === true && baseProfile.extractToolCallsFromContent !== true ? { ...baseProfile, extractToolCallsFromContent: true } : baseProfile;
   const ambient = currentCredentialPool(name);
   if (ambient !== void 0) {
     return new PoolAwareLlmClient(ambient, (apiKey) => selectTransport(profile, apiKey));
@@ -11690,6 +11758,9 @@ function selectTransport(profile, apiKey) {
     const opts = { apiKey };
     opts.baseUrl = profile.baseUrl;
     opts.providerName = profile.name;
+    if (profile.extractToolCallsFromContent === true) {
+      opts.extractToolCallsFromContent = true;
+    }
     if (profile.name === "openai" && process.env.OPENAI_ORGANIZATION !== void 0) {
       opts.organization = process.env.OPENAI_ORGANIZATION;
     }
@@ -11986,11 +12057,13 @@ function buildLoopInputs(options, runId, userText) {
   const fallback = options.agentOptions.providers?.fallback;
   const apiKeys = options.agentOptions.providers?.apiKeys;
   const credentialPoolStrategy = options.agentOptions.providers?.credentialPoolStrategy;
+  const extractToolCallsFromContent = options.agentOptions.providers?.routes?.[0]?.extractToolCallsFromContent;
   const chain = resolveProviderChain({
     primary,
     ...fallback !== void 0 ? { fallback } : {},
     ...apiKeys !== void 0 ? { apiKeys } : {},
-    ...credentialPoolStrategy !== void 0 ? { credentialPoolStrategy } : {}
+    ...credentialPoolStrategy !== void 0 ? { credentialPoolStrategy } : {},
+    ...extractToolCallsFromContent === true ? { extractToolCallsFromContent: true } : {}
   });
   const llm = chain.length === 1 ? chain[0] : new FallbackLlmClient(chain);
   return {