@blockrun/clawrouter 0.10.13 → 0.10.15

package/dist/index.d.ts

@@ -694,6 +694,13 @@ type BlockRunModel = {
     vision?: boolean;
     /** Models optimized for agentic workflows (multi-step autonomous tasks) */
     agentic?: boolean;
+    /**
+     * Model supports OpenAI-compatible structured function/tool calling.
+     * Models without this flag output tool invocations as plain text JSON,
+     * which leaks raw {"command":"..."} into visible chat messages.
+     * Default: false (must opt-in to prevent silent regressions on new models).
+     */
+    toolCalling?: boolean;
 };
 declare const BLOCKRUN_MODELS: BlockRunModel[];
 /**

package/dist/index.js

@@ -115,7 +115,8 @@ var BLOCKRUN_MODELS = [
     maxOutput: 128e3,
     reasoning: true,
     vision: true,
-    agentic: true
+    agentic: true,
+    toolCalling: true
   },
   {
     id: "openai/gpt-5-mini",
@@ -124,7 +125,8 @@ var BLOCKRUN_MODELS = [
     inputPrice: 0.25,
     outputPrice: 2,
     contextWindow: 2e5,
-    maxOutput: 65536
+    maxOutput: 65536,
+    toolCalling: true
   },
   {
     id: "openai/gpt-5-nano",
@@ -133,7 +135,8 @@ var BLOCKRUN_MODELS = [
     inputPrice: 0.05,
     outputPrice: 0.4,
     contextWindow: 128e3,
-    maxOutput: 32768
+    maxOutput: 32768,
+    toolCalling: true
   },
   {
     id: "openai/gpt-5.2-pro",
@@ -143,7 +146,8 @@ var BLOCKRUN_MODELS = [
     outputPrice: 168,
     contextWindow: 4e5,
     maxOutput: 128e3,
-    reasoning: true
+    reasoning: true,
+    toolCalling: true
   },
   // OpenAI Codex Family
   {
@@ -154,7 +158,8 @@ var BLOCKRUN_MODELS = [
     outputPrice: 14,
     contextWindow: 128e3,
     maxOutput: 32e3,
-    agentic: true
+    agentic: true,
+    toolCalling: true
   },
   // OpenAI GPT-4 Family
   {
@@ -165,7 +170,8 @@ var BLOCKRUN_MODELS = [
     outputPrice: 8,
     contextWindow: 128e3,
     maxOutput: 16384,
-    vision: true
+    vision: true,
+    toolCalling: true
   },
   {
     id: "openai/gpt-4.1-mini",
@@ -174,7 +180,8 @@ var BLOCKRUN_MODELS = [
     inputPrice: 0.4,
     outputPrice: 1.6,
     contextWindow: 128e3,
-    maxOutput: 16384
+    maxOutput: 16384,
+    toolCalling: true
   },
   {
     id: "openai/gpt-4.1-nano",
@@ -183,7 +190,8 @@ var BLOCKRUN_MODELS = [
     inputPrice: 0.1,
     outputPrice: 0.4,
     contextWindow: 128e3,
-    maxOutput: 16384
+    maxOutput: 16384,
+    toolCalling: true
   },
   {
     id: "openai/gpt-4o",
@@ -194,7 +202,8 @@ var BLOCKRUN_MODELS = [
     contextWindow: 128e3,
     maxOutput: 16384,
     vision: true,
-    agentic: true
+    agentic: true,
+    toolCalling: true
   },
   {
     id: "openai/gpt-4o-mini",
@@ -203,7 +212,8 @@ var BLOCKRUN_MODELS = [
     inputPrice: 0.15,
     outputPrice: 0.6,
     contextWindow: 128e3,
-    maxOutput: 16384
+    maxOutput: 16384,
+    toolCalling: true
   },
   // OpenAI O-series (Reasoning)
   {
@@ -214,7 +224,8 @@ var BLOCKRUN_MODELS = [
     outputPrice: 60,
     contextWindow: 2e5,
     maxOutput: 1e5,
-    reasoning: true
+    reasoning: true,
+    toolCalling: true
   },
   {
     id: "openai/o1-mini",
@@ -224,7 +235,8 @@ var BLOCKRUN_MODELS = [
     outputPrice: 4.4,
     contextWindow: 128e3,
     maxOutput: 65536,
-    reasoning: true
+    reasoning: true,
+    toolCalling: true
   },
   {
     id: "openai/o3",
@@ -234,7 +246,8 @@ var BLOCKRUN_MODELS = [
     outputPrice: 8,
     contextWindow: 2e5,
     maxOutput: 1e5,
-    reasoning: true
+    reasoning: true,
+    toolCalling: true
   },
   {
     id: "openai/o3-mini",
@@ -244,7 +257,8 @@ var BLOCKRUN_MODELS = [
     outputPrice: 4.4,
     contextWindow: 128e3,
     maxOutput: 65536,
-    reasoning: true
+    reasoning: true,
+    toolCalling: true
   },
   {
     id: "openai/o4-mini",
@@ -254,7 +268,8 @@ var BLOCKRUN_MODELS = [
     outputPrice: 4.4,
     contextWindow: 128e3,
     maxOutput: 65536,
-    reasoning: true
+    reasoning: true,
+    toolCalling: true
   },
   // Anthropic - all Claude models excel at agentic workflows
   // Use newest versions (4.6) with full provider prefix
@@ -266,7 +281,8 @@ var BLOCKRUN_MODELS = [
     outputPrice: 5,
     contextWindow: 2e5,
     maxOutput: 8192,
-    agentic: true
+    agentic: true,
+    toolCalling: true
   },
   {
     id: "anthropic/claude-sonnet-4.6",
@@ -277,7 +293,8 @@ var BLOCKRUN_MODELS = [
     contextWindow: 2e5,
     maxOutput: 64e3,
     reasoning: true,
-    agentic: true
+    agentic: true,
+    toolCalling: true
   },
   {
     id: "anthropic/claude-opus-4.6",
@@ -288,7 +305,8 @@ var BLOCKRUN_MODELS = [
     contextWindow: 2e5,
     maxOutput: 32e3,
     reasoning: true,
-    agentic: true
+    agentic: true,
+    toolCalling: true
   },
   // Google
   {
@@ -300,7 +318,8 @@ var BLOCKRUN_MODELS = [
     contextWindow: 105e4,
     maxOutput: 65536,
     reasoning: true,
-    vision: true
+    vision: true,
+    toolCalling: true
   },
   {
     id: "google/gemini-3-pro-preview",
@@ -311,7 +330,8 @@ var BLOCKRUN_MODELS = [
     contextWindow: 105e4,
     maxOutput: 65536,
     reasoning: true,
-    vision: true
+    vision: true,
+    toolCalling: true
   },
   {
     id: "google/gemini-3-flash-preview",
@@ -321,7 +341,8 @@ var BLOCKRUN_MODELS = [
     outputPrice: 3,
     contextWindow: 1e6,
     maxOutput: 65536,
-    vision: true
+    vision: true,
+    toolCalling: true
   },
   {
     id: "google/gemini-2.5-pro",
@@ -332,7 +353,8 @@ var BLOCKRUN_MODELS = [
     contextWindow: 105e4,
     maxOutput: 65536,
     reasoning: true,
-    vision: true
+    vision: true,
+    toolCalling: true
   },
   {
     id: "google/gemini-2.5-flash",
@@ -341,7 +363,8 @@ var BLOCKRUN_MODELS = [
     inputPrice: 0.3,
     outputPrice: 2.5,
     contextWindow: 1e6,
-    maxOutput: 65536
+    maxOutput: 65536,
+    toolCalling: true
   },
   {
     id: "google/gemini-2.5-flash-lite",
@@ -350,7 +373,8 @@ var BLOCKRUN_MODELS = [
     inputPrice: 0.1,
     outputPrice: 0.4,
     contextWindow: 1e6,
-    maxOutput: 65536
+    maxOutput: 65536,
+    toolCalling: true
   },
   // DeepSeek
   {
@@ -360,7 +384,8 @@ var BLOCKRUN_MODELS = [
     inputPrice: 0.28,
     outputPrice: 0.42,
     contextWindow: 128e3,
-    maxOutput: 8192
+    maxOutput: 8192,
+    toolCalling: true
   },
   {
     id: "deepseek/deepseek-reasoner",
@@ -370,7 +395,8 @@ var BLOCKRUN_MODELS = [
     outputPrice: 0.42,
     contextWindow: 128e3,
     maxOutput: 8192,
-    reasoning: true
+    reasoning: true,
+    toolCalling: true
   },
   // Moonshot / Kimi - optimized for agentic workflows
   {
@@ -383,7 +409,8 @@ var BLOCKRUN_MODELS = [
     maxOutput: 8192,
     reasoning: true,
     vision: true,
-    agentic: true
+    agentic: true,
+    toolCalling: true
   },
   // xAI / Grok
   {
@@ -394,7 +421,8 @@ var BLOCKRUN_MODELS = [
     outputPrice: 15,
     contextWindow: 131072,
     maxOutput: 16384,
-    reasoning: true
+    reasoning: true,
+    toolCalling: true
   },
   // grok-3-fast removed - too expensive ($5/$25), use grok-4-fast instead
   {
@@ -404,7 +432,8 @@ var BLOCKRUN_MODELS = [
     inputPrice: 0.3,
     outputPrice: 0.5,
     contextWindow: 131072,
-    maxOutput: 16384
+    maxOutput: 16384,
+    toolCalling: true
   },
   // xAI Grok 4 Family - Ultra-cheap fast models
   {
@@ -415,7 +444,8 @@ var BLOCKRUN_MODELS = [
     outputPrice: 0.5,
     contextWindow: 131072,
     maxOutput: 16384,
-    reasoning: true
+    reasoning: true,
+    toolCalling: true
   },
   {
     id: "xai/grok-4-fast-non-reasoning",
@@ -424,7 +454,8 @@ var BLOCKRUN_MODELS = [
     inputPrice: 0.2,
     outputPrice: 0.5,
     contextWindow: 131072,
-    maxOutput: 16384
+    maxOutput: 16384,
+    toolCalling: true
   },
   {
     id: "xai/grok-4-1-fast-reasoning",
@@ -434,7 +465,8 @@ var BLOCKRUN_MODELS = [
     outputPrice: 0.5,
     contextWindow: 131072,
     maxOutput: 16384,
-    reasoning: true
+    reasoning: true,
+    toolCalling: true
   },
   {
     id: "xai/grok-4-1-fast-non-reasoning",
@@ -443,7 +475,8 @@ var BLOCKRUN_MODELS = [
     inputPrice: 0.2,
     outputPrice: 0.5,
     contextWindow: 131072,
-    maxOutput: 16384
+    maxOutput: 16384,
+    toolCalling: true
   },
   {
     id: "xai/grok-code-fast-1",
@@ -452,9 +485,10 @@ var BLOCKRUN_MODELS = [
     inputPrice: 0.2,
     outputPrice: 1.5,
     contextWindow: 131072,
-    maxOutput: 16384,
-    agentic: true
-    // Good for coding tasks
+    maxOutput: 16384
+    // toolCalling intentionally omitted: outputs tool calls as plain text JSON,
+    // not OpenAI-compatible structured function calls. Will be skipped when
+    // request has tools to prevent the "talking to itself" bug.
   },
   {
     id: "xai/grok-4-0709",
@@ -464,7 +498,8 @@ var BLOCKRUN_MODELS = [
     outputPrice: 1.5,
     contextWindow: 131072,
     maxOutput: 16384,
-    reasoning: true
+    reasoning: true,
+    toolCalling: true
   },
   {
     id: "xai/grok-2-vision",
@@ -474,7 +509,8 @@ var BLOCKRUN_MODELS = [
     outputPrice: 10,
     contextWindow: 131072,
     maxOutput: 16384,
-    vision: true
+    vision: true,
+    toolCalling: true
   },
   // MiniMax
   {
@@ -486,7 +522,8 @@ var BLOCKRUN_MODELS = [
     contextWindow: 204800,
     maxOutput: 16384,
     reasoning: true,
-    agentic: true
+    agentic: true,
+    toolCalling: true
   },
   // NVIDIA - Free/cheap models
   {
@@ -497,6 +534,8 @@ var BLOCKRUN_MODELS = [
     outputPrice: 0,
     contextWindow: 128e3,
     maxOutput: 16384
+    // toolCalling intentionally omitted: free model, structured function
+    // calling support unverified. Excluded from tool-heavy routing paths.
   },
   {
     id: "nvidia/kimi-k2.5",
@@ -505,7 +544,8 @@ var BLOCKRUN_MODELS = [
     inputPrice: 0.55,
     outputPrice: 2.5,
     contextWindow: 262144,
-    maxOutput: 16384
+    maxOutput: 16384,
+    toolCalling: true
   }
 ];
 function toOpenClawModel(m) {
@@ -550,6 +590,11 @@ function isAgenticModel(modelId) {
 function getAgenticModels() {
   return BLOCKRUN_MODELS.filter((m) => m.agentic).map((m) => m.id);
 }
+function supportsToolCalling(modelId) {
+  const normalized = modelId.replace("blockrun/", "");
+  const model = BLOCKRUN_MODELS.find((m) => m.id === normalized);
+  return model?.toolCalling ?? false;
+}
 function getModelContextWindow(modelId) {
   const normalized = modelId.replace("blockrun/", "");
   const model = BLOCKRUN_MODELS.find((m) => m.id === normalized);
@@ -1061,7 +1106,8 @@ function classifyByRules(prompt, systemPrompt, estimatedTokens, config) {
       tier: "REASONING",
       confidence: Math.max(confidence2, 0.85),
       signals,
-      agenticScore
+      agenticScore,
+      dimensions
     };
   }
   const { simpleMedium, mediumComplex, complexReasoning } = config.tierBoundaries;
@@ -1085,9 +1131,9 @@ function classifyByRules(prompt, systemPrompt, estimatedTokens, config) {
   }
   const confidence = calibrateConfidence(distanceFromBoundary, config.confidenceSteepness);
   if (confidence < config.confidenceThreshold) {
-    return { score: weightedScore, tier: null, confidence, signals, agenticScore };
+    return { score: weightedScore, tier: null, confidence, signals, agenticScore, dimensions };
   }
-  return { score: weightedScore, tier, confidence, signals, agenticScore };
+  return { score: weightedScore, tier, confidence, signals, agenticScore, dimensions };
 }
 function calibrateConfidence(distance, steepness) {
   return 1 / (1 + Math.exp(-steepness * distance));
@@ -1143,6 +1189,11 @@ function calculateModelCost(model, modelPricing, estimatedInputTokens, maxOutput
   const savings = routingProfile === "premium" ? 0 : baselineCost > 0 ? Math.max(0, (baselineCost - costEstimate) / baselineCost) : 0;
   return { costEstimate, baselineCost, savings };
 }
+function filterByToolCalling(models, hasTools, supportsToolCalling2) {
+  if (!hasTools) return models;
+  const filtered = models.filter(supportsToolCalling2);
+  return filtered.length > 0 ? filtered : models;
+}
 function getFallbackChainFiltered(tier, tierConfigs, estimatedTotalTokens, getContextWindow) {
   const fullChain = getFallbackChain(tier, tierConfigs);
   const filtered = fullChain.filter((modelId) => {
@@ -2198,12 +2249,12 @@ var DEFAULT_ROUTING_CONFIG = {
       ]
     },
     MEDIUM: {
-      primary: "xai/grok-code-fast-1",
-      // Code specialist, $0.20/$1.50
+      primary: "moonshot/kimi-k2.5",
+      // $0.50/$2.40 - strong tool use, proper function call format
       fallback: [
+        "deepseek/deepseek-chat",
         "google/gemini-2.5-flash-lite",
         // 1M context, ultra cheap ($0.10/$0.40)
-        "deepseek/deepseek-chat",
         "xai/grok-4-1-fast-non-reasoning"
         // Upgraded Grok 4.1
       ]
@@ -2269,7 +2320,7 @@ var DEFAULT_ROUTING_CONFIG = {
       fallback: [
         "anthropic/claude-haiku-4.5",
         "google/gemini-2.5-flash-lite",
-        "xai/grok-code-fast-1"
+        "deepseek/deepseek-chat"
       ]
     },
     MEDIUM: {
@@ -2320,9 +2371,9 @@ var DEFAULT_ROUTING_CONFIG = {
       ]
     },
     MEDIUM: {
-      primary: "xai/grok-code-fast-1",
-      // Code specialist for agentic coding
-      fallback: ["moonshot/kimi-k2.5", "anthropic/claude-haiku-4.5", "claude-sonnet-4"]
+      primary: "moonshot/kimi-k2.5",
+      // $0.50/$2.40 - strong tool use, handles function calls correctly
+      fallback: ["anthropic/claude-haiku-4.5", "deepseek/deepseek-chat", "xai/grok-4-1-fast-non-reasoning"]
     },
     COMPLEX: {
       primary: "anthropic/claude-sonnet-4.6",
@@ -5182,6 +5233,7 @@ async function proxyRequest(req, res, apiBase, payFetch, options, routerOpts, de
   const originalContextSizeKB = Math.ceil(body.length / 1024);
   const debugMode = req.headers["x-clawrouter-debug"] !== "false";
   let routingDecision;
+  let hasTools = false;
   let isStreaming = false;
   let modelId = "";
   let maxTokens = 4096;
@@ -5196,10 +5248,11 @@ async function proxyRequest(req, res, apiBase, payFetch, options, routerOpts, de
       modelId = parsed.model || "";
       maxTokens = parsed.max_tokens || 4096;
       let bodyModified = false;
-      if (sessionId && Array.isArray(parsed.messages)) {
-        const messages = parsed.messages;
-        const lastUserMsg = [...messages].reverse().find((m) => m.role === "user");
-        const lastContent = typeof lastUserMsg?.content === "string" ? lastUserMsg.content : "";
+      const parsedMessages = Array.isArray(parsed.messages) ? parsed.messages : [];
+      const lastUserMsg = [...parsedMessages].reverse().find((m) => m.role === "user");
+      const lastContent = typeof lastUserMsg?.content === "string" ? lastUserMsg.content : "";
+      if (sessionId && parsedMessages.length > 0) {
+        const messages = parsedMessages;
         if (sessionJournal.needsContext(lastContent)) {
           const journalText = sessionJournal.format(sessionId);
           if (journalText) {
@@ -5220,6 +5273,106 @@ async function proxyRequest(req, res, apiBase, payFetch, options, routerOpts, de
           }
         }
       }
+      if (lastContent.startsWith("/debug")) {
+        const debugPrompt = lastContent.slice("/debug".length).trim() || "hello";
+        const messages = parsed.messages;
+        const systemMsg = messages?.find((m) => m.role === "system");
+        const systemPrompt = typeof systemMsg?.content === "string" ? systemMsg.content : void 0;
+        const fullText = `${systemPrompt ?? ""} ${debugPrompt}`;
+        const estimatedTokens = Math.ceil(fullText.length / 4);
+        const normalizedModel2 = typeof parsed.model === "string" ? parsed.model.trim().toLowerCase() : "";
+        const profileName = normalizedModel2.replace("blockrun/", "");
+        const debugProfile = ["free", "eco", "auto", "premium"].includes(profileName) ? profileName : "auto";
+        const scoring = classifyByRules(
+          debugPrompt,
+          systemPrompt,
+          estimatedTokens,
+          DEFAULT_ROUTING_CONFIG.scoring
+        );
+        const debugRouting = route(debugPrompt, systemPrompt, maxTokens, {
+          ...routerOpts,
+          routingProfile: debugProfile
+        });
+        const dimLines = (scoring.dimensions ?? []).map((d) => {
+          const nameStr = (d.name + ":").padEnd(24);
+          const scoreStr = d.score.toFixed(2).padStart(6);
+          const sigStr = d.signal ? `  [${d.signal}]` : "";
+          return `  ${nameStr}${scoreStr}${sigStr}`;
+        }).join("\n");
+        const sess = sessionId ? sessionStore.getSession(sessionId) : void 0;
+        const sessLine = sess ? `Session: ${sessionId.slice(0, 8)}... \u2192 pinned: ${sess.model} (${sess.requestCount} requests)` : sessionId ? `Session: ${sessionId.slice(0, 8)}... \u2192 no pinned model` : "Session: none";
+        const { simpleMedium, mediumComplex, complexReasoning } = DEFAULT_ROUTING_CONFIG.scoring.tierBoundaries;
+        const debugText = [
+          "ClawRouter Debug",
+          "",
+          `Profile: ${debugProfile} | Tier: ${debugRouting.tier} | Model: ${debugRouting.model}`,
+          `Confidence: ${debugRouting.confidence.toFixed(2)} | Cost: $${debugRouting.costEstimate.toFixed(4)} | Savings: ${(debugRouting.savings * 100).toFixed(0)}%`,
+          `Reasoning: ${debugRouting.reasoning}`,
+          "",
+          `Scoring (weighted: ${scoring.score.toFixed(3)})`,
+          dimLines,
+          "",
+          `Tier Boundaries: SIMPLE <${simpleMedium.toFixed(2)} | MEDIUM <${mediumComplex.toFixed(2)} | COMPLEX <${complexReasoning.toFixed(2)} | REASONING >=${complexReasoning.toFixed(2)}`,
+          "",
+          sessLine
+        ].join("\n");
+        const completionId = `chatcmpl-debug-${Date.now()}`;
+        const timestamp = Math.floor(Date.now() / 1e3);
+        const syntheticResponse = {
+          id: completionId,
+          object: "chat.completion",
+          created: timestamp,
+          model: "clawrouter/debug",
+          choices: [
+            {
+              index: 0,
+              message: { role: "assistant", content: debugText },
+              finish_reason: "stop"
+            }
+          ],
+          usage: { prompt_tokens: 0, completion_tokens: 0, total_tokens: 0 }
+        };
+        if (isStreaming) {
+          res.writeHead(200, {
+            "Content-Type": "text/event-stream",
+            "Cache-Control": "no-cache",
+            Connection: "keep-alive"
+          });
+          const sseChunk = {
+            id: completionId,
+            object: "chat.completion.chunk",
+            created: timestamp,
+            model: "clawrouter/debug",
+            choices: [
+              {
+                index: 0,
+                delta: { role: "assistant", content: debugText },
+                finish_reason: null
+              }
+            ]
+          };
+          const sseDone = {
+            id: completionId,
+            object: "chat.completion.chunk",
+            created: timestamp,
+            model: "clawrouter/debug",
+            choices: [{ index: 0, delta: {}, finish_reason: "stop" }]
+          };
+          res.write(`data: ${JSON.stringify(sseChunk)}
+
+`);
+          res.write(`data: ${JSON.stringify(sseDone)}
+
+`);
+          res.write("data: [DONE]\n\n");
+          res.end();
+        } else {
+          res.writeHead(200, { "Content-Type": "application/json" });
+          res.end(JSON.stringify(syntheticResponse));
+        }
+        console.log(`[ClawRouter] /debug command \u2192 ${debugRouting.tier} | ${debugRouting.model}`);
+        return;
+      }
       if (parsed.stream === true) {
         parsed.stream = false;
         bodyModified = true;
@@ -5274,20 +5427,20 @@ async function proxyRequest(req, res, apiBase, payFetch, options, routerOpts, de
             sessionStore.touchSession(sessionId2);
           } else {
             const messages = parsed.messages;
-            let lastUserMsg;
+            let lastUserMsg2;
             if (messages) {
               for (let i = messages.length - 1; i >= 0; i--) {
                 if (messages[i].role === "user") {
-                  lastUserMsg = messages[i];
+                  lastUserMsg2 = messages[i];
                   break;
                 }
               }
             }
             const systemMsg = messages?.find((m) => m.role === "system");
-            const prompt = typeof lastUserMsg?.content === "string" ? lastUserMsg.content : "";
+            const prompt = typeof lastUserMsg2?.content === "string" ? lastUserMsg2.content : "";
             const systemPrompt = typeof systemMsg?.content === "string" ? systemMsg.content : void 0;
             const tools = parsed.tools;
-            const hasTools = Array.isArray(tools) && tools.length > 0;
+            hasTools = Array.isArray(tools) && tools.length > 0;
             if (hasTools && tools) {
               console.log(
                 `[ClawRouter] Tools detected (${tools.length}), agentic mode via keywords`
@@ -5504,7 +5657,14 @@ async function proxyRequest(req, res, apiBase, payFetch, options, routerOpts, de
           `[ClawRouter] Context filter (~${estimatedTotalTokens} tokens): excluded ${contextExcluded.join(", ")}`
         );
       }
-      modelsToTry = contextFiltered.slice(0, MAX_FALLBACK_ATTEMPTS);
+      const toolFiltered = filterByToolCalling(contextFiltered, hasTools, supportsToolCalling);
+      const toolExcluded = contextFiltered.filter((m) => !toolFiltered.includes(m));
+      if (toolExcluded.length > 0) {
+        console.log(
+          `[ClawRouter] Tool-calling filter: excluded ${toolExcluded.join(", ")} (no structured function call support)`
+        );
+      }
+      modelsToTry = toolFiltered.slice(0, MAX_FALLBACK_ATTEMPTS);
       modelsToTry = prioritizeNonRateLimited(modelsToTry);
     } else {
       if (modelId && modelId !== FREE_MODEL) {