llmist 0.7.0 → 0.8.0

package/dist/index.cjs

@@ -881,7 +881,7 @@ function findSafeDelimiter(content) {
   }
   let counter = 1;
   while (counter < 1e3) {
-    const delimiter = `HEREDOC_${counter}`;
+    const delimiter = `__GADGET_PARAM_${counter}__`;
     const regex = new RegExp(`^${delimiter}\\s*$`);
     const isUsed = lines.some((line) => regex.test(line));
     if (!isUsed) {
@@ -988,7 +988,16 @@ var init_gadget = __esm({
     yaml = __toESM(require("js-yaml"), 1);
     init_schema_to_json();
     init_schema_validator();
-    HEREDOC_DELIMITERS = ["EOF", "END", "DOC", "CONTENT", "TEXT", "HEREDOC", "DATA", "BLOCK"];
+    HEREDOC_DELIMITERS = [
+      "__GADGET_PARAM_EOF__",
+      "__GADGET_PARAM_END__",
+      "__GADGET_PARAM_DOC__",
+      "__GADGET_PARAM_CONTENT__",
+      "__GADGET_PARAM_TEXT__",
+      "__GADGET_PARAM_HEREDOC__",
+      "__GADGET_PARAM_DATA__",
+      "__GADGET_PARAM_BLOCK__"
+    ];
     BaseGadget = class {
       /**
        * The name of the gadget. Used for identification when LLM calls it.
@@ -3096,7 +3105,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 3,
           output: 15,
-          cachedInput: 0.3
+          cachedInput: 0.3,
+          cacheWriteInput: 3.75
         },
         knowledgeCutoff: "2025-01",
         features: {
@@ -3120,7 +3130,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 1,
           output: 5,
-          cachedInput: 0.1
+          cachedInput: 0.1,
+          cacheWriteInput: 1.25
         },
         knowledgeCutoff: "2025-02",
         features: {
@@ -3144,7 +3155,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 3,
           output: 15,
-          cachedInput: 0.3
+          cachedInput: 0.3,
+          cacheWriteInput: 3.75
         },
         knowledgeCutoff: "2025-03",
         features: {
@@ -3168,7 +3180,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 3,
           output: 15,
-          cachedInput: 0.3
+          cachedInput: 0.3,
+          cacheWriteInput: 3.75
         },
         knowledgeCutoff: "2024-11",
         features: {
@@ -3192,7 +3205,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 15,
           output: 75,
-          cachedInput: 1.5
+          cachedInput: 1.5,
+          cacheWriteInput: 18.75
         },
         knowledgeCutoff: "2025-01",
         features: {
@@ -3216,7 +3230,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 15,
           output: 75,
-          cachedInput: 1.5
+          cachedInput: 1.5,
+          cacheWriteInput: 18.75
         },
         knowledgeCutoff: "2025-03",
         features: {
@@ -3239,7 +3254,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 0.8,
           output: 4,
-          cachedInput: 0.08
+          cachedInput: 0.08,
+          cacheWriteInput: 1
         },
         knowledgeCutoff: "2024-07",
         features: {
@@ -3262,7 +3278,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 0.25,
           output: 1.25,
-          cachedInput: 0.025
+          cachedInput: 0.025,
+          cacheWriteInput: 0.3125
         },
         knowledgeCutoff: "2023-08",
         features: {
@@ -3286,7 +3303,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 1,
           output: 5,
-          cachedInput: 0.1
+          cachedInput: 0.1,
+          cacheWriteInput: 1.25
         },
         knowledgeCutoff: "2025-02",
         features: {
@@ -3310,7 +3328,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 3,
           output: 15,
-          cachedInput: 0.3
+          cachedInput: 0.3,
+          cacheWriteInput: 3.75
         },
         knowledgeCutoff: "2025-01",
         features: {
@@ -3334,7 +3353,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 5,
           output: 25,
-          cachedInput: 0.5
+          cachedInput: 0.5,
+          cacheWriteInput: 6.25
         },
         knowledgeCutoff: "2025-03",
         features: {
@@ -3449,15 +3469,27 @@ var init_anthropic = __esm({
       }
       buildRequestPayload(options, descriptor, spec, messages) {
         const systemMessages = messages.filter((message) => message.role === "system");
-        const system = systemMessages.length > 0 ? systemMessages.map((m) => m.content).join("\n\n") : void 0;
-        const conversation = messages.filter(
+        const system = systemMessages.length > 0 ? systemMessages.map((m, index) => ({
+          type: "text",
+          text: m.content,
+          // Add cache_control to the LAST system message block
+          ...index === systemMessages.length - 1 ? { cache_control: { type: "ephemeral" } } : {}
+        })) : void 0;
+        const nonSystemMessages = messages.filter(
           (message) => message.role !== "system"
-        ).map((message) => ({
+        );
+        const lastUserIndex = nonSystemMessages.reduce(
+          (lastIdx, msg, idx) => msg.role === "user" ? idx : lastIdx,
+          -1
+        );
+        const conversation = nonSystemMessages.map((message, index) => ({
           role: message.role,
           content: [
             {
               type: "text",
-              text: message.content
+              text: message.content,
+              // Add cache_control to the LAST user message
+              ...message.role === "user" && index === lastUserIndex ? { cache_control: { type: "ephemeral" } } : {}
             }
           ]
         }));
@@ -3483,15 +3515,22 @@ var init_anthropic = __esm({
       async *wrapStream(iterable) {
         const stream2 = iterable;
         let inputTokens = 0;
+        let cachedInputTokens = 0;
+        let cacheCreationInputTokens = 0;
         for await (const event of stream2) {
           if (event.type === "message_start") {
-            inputTokens = event.message.usage.input_tokens;
+            const usage = event.message.usage;
+            cachedInputTokens = usage.cache_read_input_tokens ?? 0;
+            cacheCreationInputTokens = usage.cache_creation_input_tokens ?? 0;
+            inputTokens = usage.input_tokens + cachedInputTokens + cacheCreationInputTokens;
             yield {
               text: "",
               usage: {
                 inputTokens,
                 outputTokens: 0,
-                totalTokens: inputTokens
+                totalTokens: inputTokens,
+                cachedInputTokens,
+                cacheCreationInputTokens
               },
               rawEvent: event
             };
@@ -3505,7 +3544,9 @@ var init_anthropic = __esm({
             const usage = event.usage ? {
               inputTokens,
               outputTokens: event.usage.output_tokens,
-              totalTokens: inputTokens + event.usage.output_tokens
+              totalTokens: inputTokens + event.usage.output_tokens,
+              cachedInputTokens,
+              cacheCreationInputTokens
             } : void 0;
             if (event.delta.stop_reason || usage) {
               yield {
@@ -3586,6 +3627,7 @@ var init_gemini_models = __esm({
   "src/providers/gemini-models.ts"() {
     "use strict";
     GEMINI_MODELS = [
+      // Gemini 3 Pro (Preview)
       {
         provider: "gemini",
         modelId: "gemini-3-pro-preview",
@@ -3594,8 +3636,11 @@ var init_gemini_models = __esm({
         maxOutputTokens: 65536,
         pricing: {
           input: 2,
+          // $2.00 for prompts <= 200k, $4.00 for > 200k (using lower tier)
           output: 12,
+          // $12.00 for prompts <= 200k, $18.00 for > 200k
           cachedInput: 0.2
+          // $0.20 for prompts <= 200k
         },
         knowledgeCutoff: "2025-01",
         features: {
@@ -3608,9 +3653,10 @@ var init_gemini_models = __esm({
         metadata: {
           family: "Gemini 3",
           releaseDate: "2025-11-18",
-          notes: "Most advanced model. 1501 Elo LMArena, 91.9% GPQA Diamond, 76.2% SWE-bench. Deep Think mode available."
+          notes: "Best model for multimodal understanding, agentic and vibe-coding. Deep Think mode available."
         }
       },
+      // Gemini 2.5 Pro
       {
         provider: "gemini",
         modelId: "gemini-2.5-pro",
@@ -3619,8 +3665,11 @@ var init_gemini_models = __esm({
         maxOutputTokens: 65536,
         pricing: {
           input: 1.25,
+          // $1.25 for prompts <= 200k, $2.50 for > 200k
           output: 10,
+          // $10.00 for prompts <= 200k, $15.00 for > 200k
           cachedInput: 0.125
+          // $0.125 for prompts <= 200k
         },
         knowledgeCutoff: "2025-01",
         features: {
@@ -3633,9 +3682,10 @@ var init_gemini_models = __esm({
         metadata: {
           family: "Gemini 2.5",
           releaseDate: "2025-06",
-          notes: "Balanced multimodal model with 1M context. Best for complex agents and reasoning."
+          notes: "State-of-the-art multipurpose model. Excels at coding and complex reasoning."
         }
       },
+      // Gemini 2.5 Flash
       {
         provider: "gemini",
         modelId: "gemini-2.5-flash",
@@ -3644,8 +3694,10 @@ var init_gemini_models = __esm({
         maxOutputTokens: 65536,
         pricing: {
           input: 0.3,
+          // $0.30 for text/image/video, $1.00 for audio
           output: 2.5,
           cachedInput: 0.03
+          // $0.03 for text/image/video
         },
         knowledgeCutoff: "2025-01",
         features: {
@@ -3658,9 +3710,10 @@ var init_gemini_models = __esm({
         metadata: {
           family: "Gemini 2.5",
           releaseDate: "2025-06",
-          notes: "Best price-performance ratio with thinking enabled by default"
+          notes: "First hybrid reasoning model with 1M context and thinking budgets."
         }
       },
+      // Gemini 2.5 Flash-Lite
       {
         provider: "gemini",
         modelId: "gemini-2.5-flash-lite",
@@ -3669,8 +3722,10 @@ var init_gemini_models = __esm({
         maxOutputTokens: 65536,
         pricing: {
           input: 0.1,
+          // $0.10 for text/image/video, $0.30 for audio
           output: 0.4,
           cachedInput: 0.01
+          // $0.01 for text/image/video
         },
         knowledgeCutoff: "2025-01",
         features: {
@@ -3682,9 +3737,10 @@ var init_gemini_models = __esm({
         metadata: {
           family: "Gemini 2.5",
           releaseDate: "2025-06",
-          notes: "Fastest and most cost-efficient model for high-volume, low-latency tasks"
+          notes: "Smallest and most cost effective model, built for at scale usage."
         }
       },
+      // Gemini 2.0 Flash
       {
         provider: "gemini",
         modelId: "gemini-2.0-flash",
@@ -3693,8 +3749,10 @@ var init_gemini_models = __esm({
         maxOutputTokens: 8192,
         pricing: {
           input: 0.1,
+          // $0.10 for text/image/video, $0.70 for audio
           output: 0.4,
-          cachedInput: 0.01
+          cachedInput: 0.025
+          // $0.025 for text/image/video
         },
         knowledgeCutoff: "2024-08",
         features: {
@@ -3705,9 +3763,10 @@ var init_gemini_models = __esm({
         },
         metadata: {
           family: "Gemini 2.0",
-          notes: "Previous generation with 1M context and multimodal capabilities"
+          notes: "Balanced multimodal model with 1M context, built for the era of Agents."
         }
       },
+      // Gemini 2.0 Flash-Lite
       {
         provider: "gemini",
         modelId: "gemini-2.0-flash-lite",
@@ -3716,8 +3775,8 @@ var init_gemini_models = __esm({
         maxOutputTokens: 8192,
         pricing: {
           input: 0.075,
-          output: 0.3,
-          cachedInput: 75e-4
+          output: 0.3
+          // No context caching available for 2.0-flash-lite
         },
         knowledgeCutoff: "2024-08",
         features: {
@@ -3728,7 +3787,7 @@ var init_gemini_models = __esm({
         },
         metadata: {
           family: "Gemini 2.0",
-          notes: "Lightweight previous generation model for cost-sensitive applications"
+          notes: "Smallest and most cost effective 2.0 model for at scale usage."
         }
       }
     ];
@@ -3898,7 +3957,9 @@ var init_gemini = __esm({
         return {
           inputTokens: usageMetadata.promptTokenCount ?? 0,
           outputTokens: usageMetadata.candidatesTokenCount ?? 0,
-          totalTokens: usageMetadata.totalTokenCount ?? 0
+          totalTokens: usageMetadata.totalTokenCount ?? 0,
+          // Gemini returns cached token count in cachedContentTokenCount
+          cachedInputTokens: usageMetadata.cachedContentTokenCount ?? 0
         };
       }
       /**
@@ -3954,10 +4015,11 @@ var init_openai_models = __esm({
   "src/providers/openai-models.ts"() {
     "use strict";
     OPENAI_MODELS = [
+      // GPT-5 Family
       {
         provider: "openai",
         modelId: "gpt-5.1",
-        displayName: "GPT-5.1 Instant",
+        displayName: "GPT-5.1",
         contextWindow: 128e3,
         maxOutputTokens: 32768,
         pricing: {
@@ -3977,34 +4039,7 @@ var init_openai_models = __esm({
         metadata: {
           family: "GPT-5",
           releaseDate: "2025-11-12",
-          notes: "Warmer, more intelligent, better instruction following. 2-3x faster than GPT-5.",
-          supportsTemperature: false
-        }
-      },
-      {
-        provider: "openai",
-        modelId: "gpt-5.1-thinking",
-        displayName: "GPT-5.1 Thinking",
-        contextWindow: 196e3,
-        maxOutputTokens: 32768,
-        pricing: {
-          input: 1.25,
-          output: 10,
-          cachedInput: 0.125
-        },
-        knowledgeCutoff: "2024-09-30",
-        features: {
-          streaming: true,
-          functionCalling: true,
-          vision: true,
-          reasoning: true,
-          structuredOutputs: true,
-          fineTuning: true
-        },
-        metadata: {
-          family: "GPT-5",
-          releaseDate: "2025-11-12",
-          notes: "Advanced reasoning with thinking levels: Light, Standard, Extended, Heavy. Best for complex tasks.",
+          notes: "Latest GPT-5 with improved instruction following. 2-3x faster than GPT-5.",
           supportsTemperature: false
         }
       },
@@ -4084,6 +4119,255 @@ var init_openai_models = __esm({
           notes: "Fastest, most cost-efficient version for well-defined tasks",
           supportsTemperature: false
         }
+      },
+      {
+        provider: "openai",
+        modelId: "gpt-5-pro",
+        displayName: "GPT-5 Pro",
+        contextWindow: 272e3,
+        maxOutputTokens: 128e3,
+        pricing: {
+          input: 15,
+          output: 120
+          // No cached input pricing for gpt-5-pro
+        },
+        knowledgeCutoff: "2024-09-30",
+        features: {
+          streaming: true,
+          functionCalling: true,
+          vision: true,
+          reasoning: true,
+          structuredOutputs: true
+        },
+        metadata: {
+          family: "GPT-5",
+          notes: "Premium tier with enhanced capabilities. Does not support prompt caching.",
+          supportsTemperature: false
+        }
+      },
+      // GPT-4.1 Family
+      {
+        provider: "openai",
+        modelId: "gpt-4.1",
+        displayName: "GPT-4.1",
+        contextWindow: 128e3,
+        maxOutputTokens: 32768,
+        pricing: {
+          input: 2,
+          output: 8,
+          cachedInput: 0.5
+        },
+        knowledgeCutoff: "2024-04-01",
+        features: {
+          streaming: true,
+          functionCalling: true,
+          vision: true,
+          structuredOutputs: true,
+          fineTuning: true
+        },
+        metadata: {
+          family: "GPT-4.1",
+          notes: "Improved GPT-4 with better instruction following"
+        }
+      },
+      {
+        provider: "openai",
+        modelId: "gpt-4.1-mini",
+        displayName: "GPT-4.1 Mini",
+        contextWindow: 128e3,
+        maxOutputTokens: 32768,
+        pricing: {
+          input: 0.4,
+          output: 1.6,
+          cachedInput: 0.1
+        },
+        knowledgeCutoff: "2024-04-01",
+        features: {
+          streaming: true,
+          functionCalling: true,
+          vision: true,
+          structuredOutputs: true,
+          fineTuning: true
+        },
+        metadata: {
+          family: "GPT-4.1",
+          notes: "Cost-efficient GPT-4.1 variant"
+        }
+      },
+      {
+        provider: "openai",
+        modelId: "gpt-4.1-nano",
+        displayName: "GPT-4.1 Nano",
+        contextWindow: 128e3,
+        maxOutputTokens: 32768,
+        pricing: {
+          input: 0.1,
+          output: 0.4,
+          cachedInput: 0.025
+        },
+        knowledgeCutoff: "2024-04-01",
+        features: {
+          streaming: true,
+          functionCalling: true,
+          vision: true,
+          structuredOutputs: true,
+          fineTuning: true
+        },
+        metadata: {
+          family: "GPT-4.1",
+          notes: "Fastest GPT-4.1 variant for simple tasks"
+        }
+      },
+      // GPT-4o Family
+      {
+        provider: "openai",
+        modelId: "gpt-4o",
+        displayName: "GPT-4o",
+        contextWindow: 128e3,
+        maxOutputTokens: 16384,
+        pricing: {
+          input: 2.5,
+          output: 10,
+          cachedInput: 1.25
+        },
+        knowledgeCutoff: "2024-04-01",
+        features: {
+          streaming: true,
+          functionCalling: true,
+          vision: true,
+          structuredOutputs: true,
+          fineTuning: true
+        },
+        metadata: {
+          family: "GPT-4o",
+          notes: "Multimodal model optimized for speed"
+        }
+      },
+      {
+        provider: "openai",
+        modelId: "gpt-4o-mini",
+        displayName: "GPT-4o Mini",
+        contextWindow: 128e3,
+        maxOutputTokens: 16384,
+        pricing: {
+          input: 0.15,
+          output: 0.6,
+          cachedInput: 0.075
+        },
+        knowledgeCutoff: "2024-04-01",
+        features: {
+          streaming: true,
+          functionCalling: true,
+          vision: true,
+          structuredOutputs: true,
+          fineTuning: true
+        },
+        metadata: {
+          family: "GPT-4o",
+          notes: "Fast and affordable multimodal model"
+        }
+      },
+      // o-series (Reasoning models)
+      {
+        provider: "openai",
+        modelId: "o1",
+        displayName: "o1",
+        contextWindow: 2e5,
+        maxOutputTokens: 1e5,
+        pricing: {
+          input: 15,
+          output: 60,
+          cachedInput: 7.5
+        },
+        knowledgeCutoff: "2024-12-01",
+        features: {
+          streaming: true,
+          functionCalling: true,
+          vision: true,
+          reasoning: true,
+          structuredOutputs: true
+        },
+        metadata: {
+          family: "o-series",
+          notes: "Advanced reasoning model with chain-of-thought",
+          supportsTemperature: false
+        }
+      },
+      {
+        provider: "openai",
+        modelId: "o3",
+        displayName: "o3",
+        contextWindow: 2e5,
+        maxOutputTokens: 1e5,
+        pricing: {
+          input: 2,
+          output: 8,
+          cachedInput: 0.5
+        },
+        knowledgeCutoff: "2025-01-01",
+        features: {
+          streaming: true,
+          functionCalling: true,
+          vision: true,
+          reasoning: true,
+          structuredOutputs: true
+        },
+        metadata: {
+          family: "o-series",
+          notes: "Next-gen reasoning model, more efficient than o1",
+          supportsTemperature: false
+        }
+      },
+      {
+        provider: "openai",
+        modelId: "o4-mini",
+        displayName: "o4 Mini",
+        contextWindow: 2e5,
+        maxOutputTokens: 1e5,
+        pricing: {
+          input: 1.1,
+          output: 4.4,
+          cachedInput: 0.275
+        },
+        knowledgeCutoff: "2025-04-01",
+        features: {
+          streaming: true,
+          functionCalling: true,
+          vision: true,
+          reasoning: true,
+          structuredOutputs: true,
+          fineTuning: true
+        },
+        metadata: {
+          family: "o-series",
+          notes: "Cost-efficient reasoning model",
+          supportsTemperature: false
+        }
+      },
+      {
+        provider: "openai",
+        modelId: "o3-mini",
+        displayName: "o3 Mini",
+        contextWindow: 2e5,
+        maxOutputTokens: 1e5,
+        pricing: {
+          input: 1.1,
+          output: 4.4,
+          cachedInput: 0.55
+        },
+        knowledgeCutoff: "2025-01-01",
+        features: {
+          streaming: true,
+          functionCalling: true,
+          vision: true,
+          reasoning: true,
+          structuredOutputs: true
+        },
+        metadata: {
+          family: "o-series",
+          notes: "Compact reasoning model for cost-sensitive applications",
+          supportsTemperature: false
+        }
       }
     ];
   }
@@ -4164,7 +4448,8 @@ var init_openai = __esm({
           const usage = chunk.usage ? {
             inputTokens: chunk.usage.prompt_tokens,
             outputTokens: chunk.usage.completion_tokens,
-            totalTokens: chunk.usage.total_tokens
+            totalTokens: chunk.usage.total_tokens,
+            cachedInputTokens: chunk.usage.prompt_tokens_details?.cached_tokens ?? 0
           } : void 0;
           if (finishReason || usage) {
             yield { text: "", finishReason, usage, rawEvent: chunk };
@@ -4381,20 +4666,28 @@ var init_model_registry = __esm({
       /**
        * Estimate API cost for a given model and token usage
        * @param modelId - Full model identifier
-       * @param inputTokens - Number of input tokens
+       * @param inputTokens - Number of input tokens (total, including cached and cache creation)
        * @param outputTokens - Number of output tokens
-       * @param useCachedInput - Whether to use cached input pricing (if supported by provider)
+       * @param cachedInputTokens - Number of cached input tokens (subset of inputTokens)
+       * @param cacheCreationInputTokens - Number of cache creation tokens (subset of inputTokens, Anthropic only)
        * @returns CostEstimate if model found, undefined otherwise
        */
-      estimateCost(modelId, inputTokens, outputTokens, useCachedInput = false) {
+      estimateCost(modelId, inputTokens, outputTokens, cachedInputTokens = 0, cacheCreationInputTokens = 0) {
         const spec = this.getModelSpec(modelId);
         if (!spec) return void 0;
-        const inputRate = useCachedInput && spec.pricing.cachedInput !== void 0 ? spec.pricing.cachedInput : spec.pricing.input;
-        const inputCost = inputTokens / 1e6 * inputRate;
+        const cachedRate = spec.pricing.cachedInput ?? spec.pricing.input;
+        const cacheWriteRate = spec.pricing.cacheWriteInput ?? spec.pricing.input;
+        const uncachedInputTokens = inputTokens - cachedInputTokens - cacheCreationInputTokens;
+        const uncachedInputCost = uncachedInputTokens / 1e6 * spec.pricing.input;
+        const cachedInputCost = cachedInputTokens / 1e6 * cachedRate;
+        const cacheCreationCost = cacheCreationInputTokens / 1e6 * cacheWriteRate;
+        const inputCost = uncachedInputCost + cachedInputCost + cacheCreationCost;
         const outputCost = outputTokens / 1e6 * spec.pricing.output;
         const totalCost = inputCost + outputCost;
         return {
           inputCost,
+          cachedInputCost,
+          cacheCreationCost,
           outputCost,
           totalCost,
           currency: "USD"