cascade-ai 0.12.7 → 0.12.9

package/dist/index.cjs

@@ -77,7 +77,7 @@ var cron__default = /*#__PURE__*/_interopDefault(cron);
 
 
 // src/constants.ts
-var CASCADE_VERSION = "0.12.7";
+var CASCADE_VERSION = "0.12.9";
 var CASCADE_CONFIG_DIR = ".cascade";
 var CASCADE_MD_FILE = "CASCADE.md";
 var CASCADE_IGNORE_FILE = ".cascadeignore";
@@ -477,6 +477,12 @@ var BaseProvider = class {
 };
 
 // src/providers/anthropic.ts
+function anthropicThinkingParam(modelId, maxTokens) {
+  if (!/claude-(opus|sonnet)-4/i.test(modelId)) return {};
+  const budget = Math.min(8e3, maxTokens - 1024);
+  if (budget < 1024) return {};
+  return { thinking: { type: "enabled", budget_tokens: budget } };
+}
 var AnthropicProvider = class extends BaseProvider {
   client;
   constructor(config, model) {
@@ -506,13 +512,18 @@ var AnthropicProvider = class extends BaseProvider {
     let fullContent = "";
     let inputTokens = 0;
     let outputTokens = 0;
+    const maxTokens = options.maxTokens ?? this.model.maxOutputTokens;
+    const thinkParam = anthropicThinkingParam(this.model.id, maxTokens);
+    const useThinking = !!thinkParam.thinking;
     const stream = this.client.messages.stream({
       model: this.model.id,
-      max_tokens: options.maxTokens ?? this.model.maxOutputTokens,
-      temperature: options.temperature ?? 0.7,
+      max_tokens: maxTokens,
+      // Extended thinking requires temperature = 1; otherwise honor the request.
+      temperature: useThinking ? 1 : options.temperature ?? 0.7,
       system: options.systemPrompt,
       messages,
-      tools: tools?.length ? tools : void 0
+      tools: tools?.length ? tools : void 0,
+      ...thinkParam
     }, { signal: options.signal });
     let isThinking = false;
     for await (const event of stream) {
@@ -1516,6 +1527,7 @@ var ModelSelector = class {
       if (lower.includes("claude")) providerStr = "anthropic";
       else if (lower.startsWith("gpt") || lower.startsWith("o1") || lower.startsWith("o3")) providerStr = "openai";
       else if (lower.includes("gemini")) providerStr = "gemini";
+      else if ((lower.endsWith(".gguf") || actualId.includes("/") || actualId.includes("\\")) && this.availableProviders.has("openai-compatible")) providerStr = "openai-compatible";
       else if (this.availableProviders.has("ollama")) providerStr = "ollama";
       else if (this.availableProviders.has("openai-compatible")) providerStr = "openai-compatible";
       else if (this.availableProviders.size === 1) providerStr = Array.from(this.availableProviders)[0];
@@ -2203,6 +2215,11 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter__default.default {
     if (availableProviders.has("ollama")) {
       await this.discoverOllamaModels(ollamaCfg);
     }
+    if (availableProviders.has("openai-compatible")) {
+      await Promise.all(
+        config.providers.filter((p) => p.type === "openai-compatible").map((cfg) => this.discoverOpenAICompatibleModels(cfg))
+      );
+    }
     for (const tier of ["T1", "T2", "T3"]) {
       const override = tier === "T1" ? config.models.t1 : tier === "T2" ? config.models.t2 : config.models.t3;
       if (!override || override === "auto") continue;
@@ -2634,6 +2651,14 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter__default.default {
   getModelsForProvider(provider) {
     return this.selector.getAvailableModelsForProvider(provider);
   }
+  /**
+   * Every model available across the configured + reachable providers, after
+   * discovery (Ollama tags, OpenAI-compatible/llama.cpp models, cloud catalog).
+   * Used to populate the desktop model pickers with the user's real models.
+   */
+  getAvailableModels() {
+    return this.selector?.getAllAvailableModels() ?? [];
+  }
   // ── Private ──────────────────────────────────
   async detectAvailableProviders(configs) {
     const available = /* @__PURE__ */ new Set();
@@ -2664,6 +2689,28 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter__default.default {
     } catch {
     }
   }
+  async discoverOpenAICompatibleModels(cfg) {
+    try {
+      const seed = {
+        id: "openai-compatible",
+        name: "openai-compatible",
+        provider: "openai-compatible",
+        contextWindow: 32e3,
+        isVisionCapable: false,
+        inputCostPer1kTokens: 0,
+        outputCostPer1kTokens: 0,
+        maxOutputTokens: 4e3,
+        supportsStreaming: true,
+        isLocal: false
+      };
+      const provider = new OpenAICompatibleProvider(cfg, seed);
+      const models = await provider.listModels();
+      for (const m of models) {
+        this.selector.addDynamicModel(m);
+      }
+    } catch {
+    }
+  }
   ensureProvider(model, configs) {
     const key = `${model.provider}:${model.id}`;
     if (this.providers.has(key)) return;
@@ -2693,7 +2740,23 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter__default.default {
     }
   }
   getAnyModelForProvider(type) {
-    return Object.values(MODELS).find((m) => m.provider === type);
+    const fromCatalog = Object.values(MODELS).find((m) => m.provider === type);
+    if (fromCatalog) return fromCatalog;
+    if (type === "openai-compatible" || type === "azure") {
+      return {
+        id: type,
+        name: type,
+        provider: type,
+        contextWindow: 32e3,
+        isVisionCapable: false,
+        inputCostPer1kTokens: 0,
+        outputCostPer1kTokens: 0,
+        maxOutputTokens: 4e3,
+        supportsStreaming: true,
+        isLocal: false
+      };
+    }
+    return void 0;
   }
   recordStats(tier, model, usage) {
     this.stats.totalTokens += usage.totalTokens;
@@ -8834,7 +8897,11 @@ ${last.partialOutput}` : "");
   looksLikeConversational(prompt) {
     const LOW_COMPLEXITY = [
       /^(?:hi|hello|hey|thanks|thank you|ok|okay|yes|no|sure|got it|sounds good)\b/i,
-      /^(?:what is|what are|list|show me|tell me|who is|where is|when is|how do i)\b/i,
+      /^(?:what is|what are|what'?s|list|show me|tell me|who is|who are|who'?re|where is|when is|how do i)\b/i,
+      // Self-identity / capability questions ("who are you", "what can you do",
+      // "who made you") are pure conversation — never a multi-agent build.
+      /^(?:who|what)\b.*\byou\b/i,
+      /^what can you\b/i,
       /\b(?:simple|quick|brief|small|single|one-line|typo|rename)\b/i
     ];
     const wordCount = prompt.trim().split(/\s+/).length;
@@ -8932,10 +8999,16 @@ ${prompt}` : prompt;
         temperature: 0
       });
       const content = result.content.trim();
-      const firstWord = (content.split(/[\s—–-]+/)[0] ?? "").toLowerCase();
+      const match = content.toLowerCase().match(/\b(simple|moderate|complex)\b/);
       const reason = content.replace(/^\S+\s*[—–-]*\s*/, "").trim();
-      const verdict = firstWord.includes("simple") ? "Simple" : firstWord.includes("moderate") ? "Moderate" : "Complex";
-      this.recordDecision("complexity", `${verdict} \u2014 classifier: ${reason || "no reason given"}`);
+      let verdict;
+      if (match) {
+        verdict = match[1] === "simple" ? "Simple" : match[1] === "moderate" ? "Moderate" : "Complex";
+        this.recordDecision("complexity", `${verdict} \u2014 classifier: ${reason || "no reason given"}`);
+      } else {
+        verdict = prompt.trim().split(/\s+/).length <= 12 ? "Simple" : "Moderate";
+        this.recordDecision("complexity", `${verdict} \u2014 classifier output unparseable; defaulted by length`);
+      }
       return verdict;
     } catch {
       const followUpPrompt = /^(proceed|continue|go ahead|do it|yes|yep|ok|okay|carry on)$/i.test(prompt.trim());