llmist 0.6.2 → 0.8.0

package/dist/cli.cjs

@@ -31,6 +31,20 @@ var __toESM = (mod, isNodeMode, target) => (target = mod != null ? __create(__ge
 ));
 var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: true }), mod);
 
+// src/core/constants.ts
+var GADGET_START_PREFIX, GADGET_END_PREFIX, DEFAULT_GADGET_OUTPUT_LIMIT, DEFAULT_GADGET_OUTPUT_LIMIT_PERCENT, CHARS_PER_TOKEN, FALLBACK_CONTEXT_WINDOW;
+var init_constants = __esm({
+  "src/core/constants.ts"() {
+    "use strict";
+    GADGET_START_PREFIX = "!!!GADGET_START:";
+    GADGET_END_PREFIX = "!!!GADGET_END";
+    DEFAULT_GADGET_OUTPUT_LIMIT = true;
+    DEFAULT_GADGET_OUTPUT_LIMIT_PERCENT = 15;
+    CHARS_PER_TOKEN = 4;
+    FALLBACK_CONTEXT_WINDOW = 128e3;
+  }
+});
+
 // src/core/model-shortcuts.ts
 function isKnownModelPattern(model) {
   const normalized = model.toLowerCase();
@@ -328,20 +342,6 @@ var init_registry = __esm({
   }
 });
 
-// src/core/constants.ts
-var GADGET_START_PREFIX, GADGET_END_PREFIX, DEFAULT_GADGET_OUTPUT_LIMIT, DEFAULT_GADGET_OUTPUT_LIMIT_PERCENT, CHARS_PER_TOKEN, FALLBACK_CONTEXT_WINDOW;
-var init_constants = __esm({
-  "src/core/constants.ts"() {
-    "use strict";
-    GADGET_START_PREFIX = "!!!GADGET_START:";
-    GADGET_END_PREFIX = "!!!GADGET_END";
-    DEFAULT_GADGET_OUTPUT_LIMIT = true;
-    DEFAULT_GADGET_OUTPUT_LIMIT_PERCENT = 15;
-    CHARS_PER_TOKEN = 4;
-    FALLBACK_CONTEXT_WINDOW = 128e3;
-  }
-});
-
 // src/core/prompt-config.ts
 function resolvePromptTemplate(template, defaultValue, context) {
   const resolved = template ?? defaultValue;
@@ -865,7 +865,7 @@ 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) {
@@ -923,6 +923,10 @@ function formatParamsAsYaml(params) {
   }
   return lines.join("\n");
 }
+function formatTomlInlineTable(obj) {
+  const entries = Object.entries(obj).map(([k, v]) => `${k} = ${formatTomlValue(v)}`);
+  return `{ ${entries.join(", ")} }`;
+}
 function formatTomlValue(value) {
   if (typeof value === "string") {
     if (value.includes("\n")) {
@@ -940,10 +944,17 @@ ${delimiter}`;
     return '""';
   }
   if (Array.isArray(value)) {
-    return JSON.stringify(value);
+    if (value.length === 0) return "[]";
+    const items = value.map((item) => {
+      if (typeof item === "object" && item !== null && !Array.isArray(item)) {
+        return formatTomlInlineTable(item);
+      }
+      return formatTomlValue(item);
+    });
+    return `[${items.join(", ")}]`;
   }
   if (typeof value === "object") {
-    return JSON.stringify(value);
+    return formatTomlInlineTable(value);
   }
   return JSON.stringify(value);
 }
@@ -961,7 +972,16 @@ 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.
@@ -1959,6 +1979,14 @@ function preprocessTomlHeredoc(tomlStr) {
   }
   return result.join("\n");
 }
+function stripMarkdownFences(content) {
+  let cleaned = content.trim();
+  const openingFence = /^```(?:toml|yaml|json)?\s*\n/i;
+  const closingFence = /\n?```\s*$/;
+  cleaned = cleaned.replace(openingFence, "");
+  cleaned = cleaned.replace(closingFence, "");
+  return cleaned.trim();
+}
 var yaml2, import_js_toml, globalInvocationCounter, StreamParser;
 var init_parser = __esm({
   "src/gadgets/parser.ts"() {
@@ -2014,35 +2042,36 @@ var init_parser = __esm({
        * Parse parameter string according to configured format
        */
       parseParameters(raw) {
+        const cleaned = stripMarkdownFences(raw);
         if (this.parameterFormat === "json") {
           try {
-            return { parameters: JSON.parse(raw) };
+            return { parameters: JSON.parse(cleaned) };
           } catch (error) {
             return { parseError: this.truncateParseError(error, "JSON") };
           }
         }
         if (this.parameterFormat === "yaml") {
           try {
-            return { parameters: yaml2.load(preprocessYaml(raw)) };
+            return { parameters: yaml2.load(preprocessYaml(cleaned)) };
           } catch (error) {
             return { parseError: this.truncateParseError(error, "YAML") };
           }
         }
         if (this.parameterFormat === "toml") {
           try {
-            return { parameters: (0, import_js_toml.load)(preprocessTomlHeredoc(raw)) };
+            return { parameters: (0, import_js_toml.load)(preprocessTomlHeredoc(cleaned)) };
           } catch (error) {
             return { parseError: this.truncateParseError(error, "TOML") };
           }
         }
         try {
-          return { parameters: JSON.parse(raw) };
+          return { parameters: JSON.parse(cleaned) };
         } catch {
           try {
-            return { parameters: (0, import_js_toml.load)(preprocessTomlHeredoc(raw)) };
+            return { parameters: (0, import_js_toml.load)(preprocessTomlHeredoc(cleaned)) };
           } catch {
             try {
-              return { parameters: yaml2.load(preprocessYaml(raw)) };
+              return { parameters: yaml2.load(preprocessYaml(cleaned)) };
             } catch (error) {
               return { parseError: this.truncateParseError(error, "auto") };
             }
@@ -2588,6 +2617,7 @@ var init_agent = __esm({
       gadgetEndPrefix;
       onHumanInputRequired;
       textOnlyHandler;
+      textWithGadgetsHandler;
       stopOnGadgetError;
       shouldContinueAfterError;
       defaultGadgetTimeoutMs;
@@ -2618,6 +2648,7 @@ var init_agent = __esm({
         this.gadgetEndPrefix = options.gadgetEndPrefix;
         this.onHumanInputRequired = options.onHumanInputRequired;
         this.textOnlyHandler = options.textOnlyHandler ?? "terminate";
+        this.textWithGadgetsHandler = options.textWithGadgetsHandler;
         this.stopOnGadgetError = options.stopOnGadgetError ?? true;
         this.shouldContinueAfterError = options.shouldContinueAfterError;
         this.defaultGadgetTimeoutMs = options.defaultGadgetTimeoutMs;
@@ -2805,6 +2836,17 @@ var init_agent = __esm({
               }
             }
             if (result.didExecuteGadgets) {
+              if (this.textWithGadgetsHandler) {
+                const textContent = result.outputs.filter((output) => output.type === "text").map((output) => output.content).join("");
+                if (textContent.trim()) {
+                  const { gadgetName, parameterMapping, resultMapping } = this.textWithGadgetsHandler;
+                  this.conversation.addGadgetCall(
+                    gadgetName,
+                    parameterMapping(textContent),
+                    resultMapping ? resultMapping(textContent) : textContent
+                  );
+                }
+              }
               for (const output of result.outputs) {
                 if (output.type === "gadget_result") {
                   const gadgetResult = output.result;
@@ -2816,7 +2858,13 @@ var init_agent = __esm({
                 }
               }
             } else {
-              this.conversation.addAssistantMessage(finalMessage);
+              if (finalMessage.trim()) {
+                this.conversation.addGadgetCall(
+                  "TellUser",
+                  { message: finalMessage, done: false, type: "info" },
+                  `\u2139\uFE0F  ${finalMessage}`
+                );
+              }
               const shouldBreak = await this.handleTextOnlyResponse(finalMessage);
               if (shouldBreak) {
                 break;
@@ -3011,7 +3059,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 3,
           output: 15,
-          cachedInput: 0.3
+          cachedInput: 0.3,
+          cacheWriteInput: 3.75
         },
         knowledgeCutoff: "2025-01",
         features: {
@@ -3035,7 +3084,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 1,
           output: 5,
-          cachedInput: 0.1
+          cachedInput: 0.1,
+          cacheWriteInput: 1.25
         },
         knowledgeCutoff: "2025-02",
         features: {
@@ -3059,7 +3109,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 3,
           output: 15,
-          cachedInput: 0.3
+          cachedInput: 0.3,
+          cacheWriteInput: 3.75
         },
         knowledgeCutoff: "2025-03",
         features: {
@@ -3083,7 +3134,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 3,
           output: 15,
-          cachedInput: 0.3
+          cachedInput: 0.3,
+          cacheWriteInput: 3.75
         },
         knowledgeCutoff: "2024-11",
         features: {
@@ -3107,7 +3159,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 15,
           output: 75,
-          cachedInput: 1.5
+          cachedInput: 1.5,
+          cacheWriteInput: 18.75
         },
         knowledgeCutoff: "2025-01",
         features: {
@@ -3131,7 +3184,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 15,
           output: 75,
-          cachedInput: 1.5
+          cachedInput: 1.5,
+          cacheWriteInput: 18.75
         },
         knowledgeCutoff: "2025-03",
         features: {
@@ -3154,7 +3208,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 0.8,
           output: 4,
-          cachedInput: 0.08
+          cachedInput: 0.08,
+          cacheWriteInput: 1
         },
         knowledgeCutoff: "2024-07",
         features: {
@@ -3177,7 +3232,8 @@ 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: {
@@ -3201,7 +3257,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 1,
           output: 5,
-          cachedInput: 0.1
+          cachedInput: 0.1,
+          cacheWriteInput: 1.25
         },
         knowledgeCutoff: "2025-02",
         features: {
@@ -3225,7 +3282,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 3,
           output: 15,
-          cachedInput: 0.3
+          cachedInput: 0.3,
+          cacheWriteInput: 3.75
         },
         knowledgeCutoff: "2025-01",
         features: {
@@ -3249,7 +3307,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 5,
           output: 25,
-          cachedInput: 0.5
+          cachedInput: 0.5,
+          cacheWriteInput: 6.25
         },
         knowledgeCutoff: "2025-03",
         features: {
@@ -3364,15 +3423,27 @@ 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" } } : {}
             }
           ]
         }));
@@ -3398,15 +3469,22 @@ 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
             };
@@ -3420,7 +3498,9 @@ 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 {
@@ -3501,6 +3581,7 @@ 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",
@@ -3509,8 +3590,11 @@ 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: {
@@ -3523,9 +3607,10 @@ 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",
@@ -3534,8 +3619,11 @@ 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: {
@@ -3548,9 +3636,10 @@ 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",
@@ -3559,8 +3648,10 @@ 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: {
@@ -3573,9 +3664,10 @@ 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",
@@ -3584,8 +3676,10 @@ 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: {
@@ -3597,9 +3691,10 @@ 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",
@@ -3608,8 +3703,10 @@ 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: {
@@ -3620,9 +3717,10 @@ 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",
@@ -3631,8 +3729,8 @@ 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: {
@@ -3643,7 +3741,7 @@ 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."
         }
       }
     ];
@@ -3813,7 +3911,9 @@ 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
         };
       }
       /**
@@ -3869,10 +3969,11 @@ 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: {
@@ -3892,34 +3993,7 @@ 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
         }
       },
@@ -3999,6 +4073,255 @@ 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
+        }
       }
     ];
   }
@@ -4079,7 +4402,8 @@ 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 };
@@ -4296,20 +4620,28 @@ 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"
@@ -4690,6 +5022,7 @@ var AgentBuilder;
 var init_builder = __esm({
   "src/agent/builder.ts"() {
     "use strict";
+    init_constants();
     init_model_shortcuts();
     init_registry();
     init_agent();
@@ -4711,6 +5044,7 @@ var init_builder = __esm({
       gadgetStartPrefix;
       gadgetEndPrefix;
       textOnlyHandler;
+      textWithGadgetsHandler;
       stopOnGadgetError;
       shouldContinueAfterError;
       defaultGadgetTimeoutMs;
@@ -4973,6 +5307,30 @@ var init_builder = __esm({
         this.textOnlyHandler = handler;
         return this;
       }
+      /**
+       * Set the handler for text content that appears alongside gadget calls.
+       *
+       * When set, text accompanying gadget responses will be wrapped as a
+       * synthetic gadget call before the actual gadget results in the
+       * conversation history.
+       *
+       * @param handler - Configuration for wrapping text
+       * @returns This builder for chaining
+       *
+       * @example
+       * ```typescript
+       * // Wrap text as TellUser gadget
+       * .withTextWithGadgetsHandler({
+       *   gadgetName: "TellUser",
+       *   parameterMapping: (text) => ({ message: text, done: false, type: "info" }),
+       *   resultMapping: (text) => `ℹ️  ${text}`,
+       * })
+       * ```
+       */
+      withTextWithGadgetsHandler(handler) {
+        this.textWithGadgetsHandler = handler;
+        return this;
+      }
       /**
        * Set whether to stop gadget execution on first error.
        *
@@ -5087,6 +5445,69 @@ var init_builder = __esm({
         this.gadgetOutputLimitPercent = percent;
         return this;
       }
+      /**
+       * Add a synthetic gadget call to the conversation history.
+       *
+       * This is useful for in-context learning - showing the LLM what "past self"
+       * did correctly so it mimics the pattern. The call is formatted with proper
+       * markers and parameter format.
+       *
+       * @param gadgetName - Name of the gadget
+       * @param parameters - Parameters passed to the gadget
+       * @param result - Result returned by the gadget
+       * @returns This builder for chaining
+       *
+       * @example
+       * ```typescript
+       * .withSyntheticGadgetCall(
+       *   'TellUser',
+       *   {
+       *     message: '👋 Hello!\n\nHere\'s what I can do:\n- Analyze code\n- Run commands',
+       *     done: false,
+       *     type: 'info'
+       *   },
+       *   'ℹ️  👋 Hello!\n\nHere\'s what I can do:\n- Analyze code\n- Run commands'
+       * )
+       * ```
+       */
+      withSyntheticGadgetCall(gadgetName, parameters, result) {
+        const startPrefix = this.gadgetStartPrefix ?? GADGET_START_PREFIX;
+        const endPrefix = this.gadgetEndPrefix ?? GADGET_END_PREFIX;
+        const format = this.parameterFormat ?? "yaml";
+        const paramStr = this.formatSyntheticParameters(parameters, format);
+        this.initialMessages.push({
+          role: "assistant",
+          content: `${startPrefix}${gadgetName}
+${paramStr}
+${endPrefix}`
+        });
+        this.initialMessages.push({
+          role: "user",
+          content: `Result: ${result}`
+        });
+        return this;
+      }
+      /**
+       * Format parameters for synthetic gadget calls.
+       * Uses heredoc for multiline string values.
+       */
+      formatSyntheticParameters(parameters, format) {
+        if (format === "json" || format === "auto") {
+          return JSON.stringify(parameters);
+        }
+        return Object.entries(parameters).map(([key, value]) => {
+          if (typeof value === "string" && value.includes("\n")) {
+            const separator = format === "yaml" ? ":" : " =";
+            return `${key}${separator} <<<EOF
+${value}
+EOF`;
+          }
+          if (format === "yaml") {
+            return typeof value === "string" ? `${key}: ${value}` : `${key}: ${JSON.stringify(value)}`;
+          }
+          return `${key} = ${JSON.stringify(value)}`;
+        }).join("\n");
+      }
       /**
        * Build and create the agent with the given user prompt.
        * Returns the Agent instance ready to run.
@@ -5129,6 +5550,7 @@ var init_builder = __esm({
           gadgetStartPrefix: this.gadgetStartPrefix,
           gadgetEndPrefix: this.gadgetEndPrefix,
           textOnlyHandler: this.textOnlyHandler,
+          textWithGadgetsHandler: this.textWithGadgetsHandler,
           stopOnGadgetError: this.stopOnGadgetError,
           shouldContinueAfterError: this.shouldContinueAfterError,
           defaultGadgetTimeoutMs: this.defaultGadgetTimeoutMs,
@@ -5230,6 +5652,7 @@ var init_builder = __esm({
           gadgetStartPrefix: this.gadgetStartPrefix,
           gadgetEndPrefix: this.gadgetEndPrefix,
           textOnlyHandler: this.textOnlyHandler,
+          textWithGadgetsHandler: this.textWithGadgetsHandler,
           stopOnGadgetError: this.stopOnGadgetError,
           shouldContinueAfterError: this.shouldContinueAfterError,
           defaultGadgetTimeoutMs: this.defaultGadgetTimeoutMs,
@@ -5265,7 +5688,8 @@ var OPTION_FLAGS = {
   logFile: "--log-file <path>",
   logReset: "--log-reset",
   noBuiltins: "--no-builtins",
-  noBuiltinInteraction: "--no-builtin-interaction"
+  noBuiltinInteraction: "--no-builtin-interaction",
+  quiet: "-q, --quiet"
 };
 var OPTION_DESCRIPTIONS = {
   model: "Model identifier, e.g. openai:gpt-5-nano or anthropic:claude-sonnet-4-5.",
@@ -5279,7 +5703,8 @@ var OPTION_DESCRIPTIONS = {
   logFile: "Path to log file. When set, logs are written to file instead of stderr.",
   logReset: "Reset (truncate) the log file at session start instead of appending.",
   noBuiltins: "Disable built-in gadgets (AskUser, TellUser).",
-  noBuiltinInteraction: "Disable interactive gadgets (AskUser) while keeping TellUser."
+  noBuiltinInteraction: "Disable interactive gadgets (AskUser) while keeping TellUser.",
+  quiet: "Suppress all output except content (text and TellUser messages)."
 };
 var SUMMARY_PREFIX = "[llmist]";
 
@@ -5289,7 +5714,7 @@ var import_commander3 = require("commander");
 // package.json
 var package_default = {
   name: "llmist",
-  version: "0.6.1",
+  version: "0.7.0",
   description: "Universal TypeScript LLM client with streaming-first agent framework. Works with any model - no structured outputs or native tool calling required. Implements its own flexible grammar for function calling.",
   type: "module",
   main: "dist/index.cjs",
@@ -5373,6 +5798,7 @@ var package_default = {
     "@google/genai": "^1.27.0",
     chalk: "^5.6.2",
     commander: "^12.1.0",
+    eta: "^4.4.1",
     "js-toml": "^1.0.2",
     "js-yaml": "^4.1.0",
     marked: "^15.0.12",
@@ -5437,7 +5863,7 @@ var tellUser = createGadget({
   name: "TellUser",
   description: "Tell the user something important. Set done=true when your work is complete and you want to end the conversation.",
   schema: import_zod2.z.object({
-    message: import_zod2.z.string().describe("The message to display to the user in Markdown"),
+    message: import_zod2.z.string().optional().describe("The message to display to the user in Markdown"),
     done: import_zod2.z.boolean().default(false).describe("Set to true to end the conversation, false to continue"),
     type: import_zod2.z.enum(["info", "success", "warning", "error"]).default("info").describe("Message type: info, success, warning, or error")
   }),
@@ -5457,9 +5883,20 @@ var tellUser = createGadget({
         done: false,
         type: "warning"
       }
+    },
+    {
+      comment: "Share detailed analysis with bullet points (use heredoc for multiline)",
+      params: {
+        message: "Here's what I found in the codebase:\n\n1. **Main entry point**: `src/index.ts` exports all public APIs\n2. **Core logic**: Located in `src/core/` with 5 modules\n3. **Tests**: Good coverage in `src/__tests__/`\n\nI'll continue exploring the core modules.",
+        done: false,
+        type: "info"
+      }
     }
   ],
   execute: ({ message, done, type }) => {
+    if (!message || message.trim() === "") {
+      return "\u26A0\uFE0F  TellUser was called without a message. Please provide content in the 'message' field.";
+    }
     const prefixes = {
       info: "\u2139\uFE0F  ",
       success: "\u2705 ",
@@ -5481,12 +5918,19 @@ var import_node_path2 = __toESM(require("path"), 1);
 var import_node_url = require("url");
 init_gadget();
 var PATH_PREFIXES = [".", "/", "~"];
+function isGadgetLike(value) {
+  if (typeof value !== "object" || value === null) {
+    return false;
+  }
+  const obj = value;
+  return typeof obj.execute === "function" && typeof obj.description === "string" && ("parameterSchema" in obj || "schema" in obj);
+}
 function isGadgetConstructor(value) {
   if (typeof value !== "function") {
     return false;
   }
   const prototype = value.prototype;
-  return Boolean(prototype) && prototype instanceof BaseGadget;
+  return Boolean(prototype) && (prototype instanceof BaseGadget || isGadgetLike(prototype));
 }
 function expandHomePath(input) {
   if (!input.startsWith("~")) {
@@ -5523,7 +5967,7 @@ function extractGadgetsFromModule(moduleExports) {
       return;
     }
     visited.add(value);
-    if (value instanceof BaseGadget) {
+    if (value instanceof BaseGadget || isGadgetLike(value)) {
       results.push(value);
       return;
     }
@@ -5648,8 +6092,14 @@ function renderSummary(metadata) {
     parts.push(import_chalk.default.magenta(metadata.model));
   }
   if (metadata.usage) {
-    const { inputTokens, outputTokens } = metadata.usage;
+    const { inputTokens, outputTokens, cachedInputTokens, cacheCreationInputTokens } = metadata.usage;
     parts.push(import_chalk.default.dim("\u2191") + import_chalk.default.yellow(` ${formatTokens(inputTokens)}`));
+    if (cachedInputTokens && cachedInputTokens > 0) {
+      parts.push(import_chalk.default.dim("\u27F3") + import_chalk.default.blue(` ${formatTokens(cachedInputTokens)}`));
+    }
+    if (cacheCreationInputTokens && cacheCreationInputTokens > 0) {
+      parts.push(import_chalk.default.dim("\u270E") + import_chalk.default.magenta(` ${formatTokens(cacheCreationInputTokens)}`));
+    }
     parts.push(import_chalk.default.dim("\u2193") + import_chalk.default.green(` ${formatTokens(outputTokens)}`));
   }
   if (metadata.elapsedSeconds !== void 0 && metadata.elapsedSeconds > 0) {
@@ -5741,53 +6191,6 @@ ${rendered}`;
 }
 
 // src/cli/utils.ts
-var RARE_EMOJI = [
-  "\u{1F531}",
-  "\u2697\uFE0F",
-  "\u{1F9FF}",
-  "\u{1F530}",
-  "\u269B\uFE0F",
-  "\u{1F3FA}",
-  "\u{1F9EB}",
-  "\u{1F52C}",
-  "\u2695\uFE0F",
-  "\u{1F5DD}\uFE0F",
-  "\u2696\uFE0F",
-  "\u{1F52E}",
-  "\u{1FAAC}",
-  "\u{1F9EC}",
-  "\u2699\uFE0F",
-  "\u{1F529}",
-  "\u{1FA9B}",
-  "\u26CF\uFE0F",
-  "\u{1FA83}",
-  "\u{1F3F9}",
-  "\u{1F6E1}\uFE0F",
-  "\u2694\uFE0F",
-  "\u{1F5E1}\uFE0F",
-  "\u{1FA93}",
-  "\u{1F5C3}\uFE0F",
-  "\u{1F4DC}",
-  "\u{1F4EF}",
-  "\u{1F3B4}",
-  "\u{1F004}",
-  "\u{1F3B2}"
-];
-function generateMarkers() {
-  const pick = (count) => {
-    const result = [];
-    const pool = [...RARE_EMOJI];
-    for (let i = 0; i < count && pool.length > 0; i++) {
-      const idx = Math.floor(Math.random() * pool.length);
-      result.push(pool.splice(idx, 1)[0]);
-    }
-    return result.join("");
-  };
-  return {
-    startPrefix: pick(5),
-    endPrefix: pick(5)
-  };
-}
 function createNumericParser({
   label,
   integer = false,
@@ -5865,6 +6268,9 @@ var StreamProgress = class {
   callOutputTokensEstimated = true;
   callOutputChars = 0;
   isStreaming = false;
+  // Cache token tracking for live cost estimation during streaming
+  callCachedInputTokens = 0;
+  callCacheCreationInputTokens = 0;
   // Cumulative stats (cumulative mode)
   totalStartTime = Date.now();
   totalTokens = 0;
@@ -5890,11 +6296,13 @@ var StreamProgress = class {
     this.callOutputTokensEstimated = true;
     this.callOutputChars = 0;
     this.isStreaming = false;
+    this.callCachedInputTokens = 0;
+    this.callCacheCreationInputTokens = 0;
     this.start();
   }
   /**
    * Ends the current LLM call. Updates cumulative stats and switches to cumulative mode.
-   * @param usage - Final token usage from the call
+   * @param usage - Final token usage from the call (including cached tokens if available)
    */
   endCall(usage) {
     this.iterations++;
@@ -5906,7 +6314,9 @@ var StreamProgress = class {
           const cost = this.modelRegistry.estimateCost(
             modelName,
             usage.inputTokens,
-            usage.outputTokens
+            usage.outputTokens,
+            usage.cachedInputTokens ?? 0,
+            usage.cacheCreationInputTokens ?? 0
           );
           if (cost) {
             this.totalCost += cost.totalCost;
@@ -5946,6 +6356,16 @@ var StreamProgress = class {
     this.callOutputTokens = tokens;
     this.callOutputTokensEstimated = estimated;
   }
+  /**
+   * Sets cached token counts for the current call (from stream metadata).
+   * Used for live cost estimation during streaming.
+   * @param cachedInputTokens - Number of tokens read from cache (cheaper)
+   * @param cacheCreationInputTokens - Number of tokens written to cache (more expensive)
+   */
+  setCachedTokens(cachedInputTokens, cacheCreationInputTokens) {
+    this.callCachedInputTokens = cachedInputTokens;
+    this.callCacheCreationInputTokens = cacheCreationInputTokens;
+  }
   /**
    * Get total elapsed time in seconds since the first call started.
    * @returns Elapsed time in seconds with 1 decimal place
@@ -6010,11 +6430,32 @@ var StreamProgress = class {
       parts.push(import_chalk2.default.dim("\u2193") + import_chalk2.default.green(` ${prefix}${formatTokens(outTokens)}`));
     }
     parts.push(import_chalk2.default.dim(`${elapsed}s`));
-    if (this.totalCost > 0) {
-      parts.push(import_chalk2.default.cyan(`$${formatCost(this.totalCost)}`));
+    const callCost = this.calculateCurrentCallCost(outTokens);
+    if (callCost > 0) {
+      parts.push(import_chalk2.default.cyan(`$${formatCost(callCost)}`));
     }
     this.target.write(`\r${parts.join(import_chalk2.default.dim(" | "))} ${import_chalk2.default.cyan(spinner)}`);
   }
+  /**
+   * Calculates live cost estimate for the current streaming call.
+   * Uses current input/output tokens and cached token counts.
+   */
+  calculateCurrentCallCost(outputTokens) {
+    if (!this.modelRegistry || !this.model) return 0;
+    try {
+      const modelName = this.model.includes(":") ? this.model.split(":")[1] : this.model;
+      const cost = this.modelRegistry.estimateCost(
+        modelName,
+        this.callInputTokens,
+        outputTokens,
+        this.callCachedInputTokens,
+        this.callCacheCreationInputTokens
+      );
+      return cost?.totalCost ?? 0;
+    } catch {
+      return 0;
+    }
+  }
   renderCumulativeMode(spinner) {
     const elapsed = ((Date.now() - this.totalStartTime) / 1e3).toFixed(1);
     const parts = [];
@@ -6163,7 +6604,7 @@ function addCompleteOptions(cmd, defaults) {
     OPTION_DESCRIPTIONS.maxTokens,
     createNumericParser({ label: "Max tokens", integer: true, min: 1 }),
     defaults?.["max-tokens"]
-  );
+  ).option(OPTION_FLAGS.quiet, OPTION_DESCRIPTIONS.quiet, defaults?.quiet);
 }
 function addAgentOptions(cmd, defaults) {
   const gadgetAccumulator = (value, previous = []) => [
@@ -6192,7 +6633,7 @@ function addAgentOptions(cmd, defaults) {
     OPTION_FLAGS.noBuiltinInteraction,
     OPTION_DESCRIPTIONS.noBuiltinInteraction,
     defaults?.["builtin-interaction"] !== false
-  );
+  ).option(OPTION_FLAGS.quiet, OPTION_DESCRIPTIONS.quiet, defaults?.quiet);
 }
 function configToCompleteOptions(config) {
   const result = {};
@@ -6200,6 +6641,7 @@ function configToCompleteOptions(config) {
   if (config.system !== void 0) result.system = config.system;
   if (config.temperature !== void 0) result.temperature = config.temperature;
   if (config["max-tokens"] !== void 0) result.maxTokens = config["max-tokens"];
+  if (config.quiet !== void 0) result.quiet = config.quiet;
   return result;
 }
 function configToAgentOptions(config) {
@@ -6213,6 +6655,11 @@ function configToAgentOptions(config) {
   if (config.builtins !== void 0) result.builtins = config.builtins;
   if (config["builtin-interaction"] !== void 0)
     result.builtinInteraction = config["builtin-interaction"];
+  if (config["gadget-start-prefix"] !== void 0)
+    result.gadgetStartPrefix = config["gadget-start-prefix"];
+  if (config["gadget-end-prefix"] !== void 0)
+    result.gadgetEndPrefix = config["gadget-end-prefix"];
+  if (config.quiet !== void 0) result.quiet = config.quiet;
   return result;
 }
 
@@ -6258,9 +6705,10 @@ async function executeAgent(promptArg, options, env) {
   const prompt = await resolvePrompt(promptArg, env);
   const client = env.createClient();
   const registry = new GadgetRegistry();
+  const stdinIsInteractive = isInteractive(env.stdin);
   if (options.builtins !== false) {
     for (const gadget of builtinGadgets) {
-      if (options.builtinInteraction === false && gadget.name === "AskUser") {
+      if (gadget.name === "AskUser" && (options.builtinInteraction === false || !stdinIsInteractive)) {
         continue;
       }
       registry.registerByClass(gadget);
@@ -6318,6 +6766,10 @@ async function executeAgent(promptArg, options, env) {
           if (context.usage.outputTokens) {
             progress.setOutputTokens(context.usage.outputTokens, false);
           }
+          progress.setCachedTokens(
+            context.usage.cachedInputTokens ?? 0,
+            context.usage.cacheCreationInputTokens ?? 0
+          );
         }
       },
       // onLLMCallComplete: Finalize metrics after each LLM call
@@ -6336,11 +6788,13 @@ async function executeAgent(promptArg, options, env) {
         let callCost;
         if (context.usage && client.modelRegistry) {
           try {
-            const modelName = options.model.includes(":") ? options.model.split(":")[1] : options.model;
+            const modelName = context.options.model.includes(":") ? context.options.model.split(":")[1] : context.options.model;
             const costResult = client.modelRegistry.estimateCost(
               modelName,
               context.usage.inputTokens,
-              context.usage.outputTokens
+              context.usage.outputTokens,
+              context.usage.cachedInputTokens ?? 0,
+              context.usage.cacheCreationInputTokens ?? 0
             );
             if (costResult) callCost = costResult.totalCost;
           } catch {
@@ -6348,7 +6802,7 @@ async function executeAgent(promptArg, options, env) {
         }
         const callElapsed = progress.getCallElapsedSeconds();
         progress.endCall(context.usage);
-        if (stderrTTY) {
+        if (!options.quiet) {
           const summary = renderSummary({
             iterations: context.iteration + 1,
             model: options.model,
@@ -6427,9 +6881,27 @@ Command rejected by user with message: "${response}"`
     builder.withGadgets(...gadgets);
   }
   builder.withParameterFormat(options.parameterFormat);
-  const markers = generateMarkers();
-  builder.withGadgetStartPrefix(markers.startPrefix);
-  builder.withGadgetEndPrefix(markers.endPrefix);
+  if (options.gadgetStartPrefix) {
+    builder.withGadgetStartPrefix(options.gadgetStartPrefix);
+  }
+  if (options.gadgetEndPrefix) {
+    builder.withGadgetEndPrefix(options.gadgetEndPrefix);
+  }
+  builder.withSyntheticGadgetCall(
+    "TellUser",
+    {
+      message: "\u{1F44B} Hello! I'm ready to help.\n\nHere's what I can do:\n- Analyze your codebase\n- Execute commands\n- Answer questions\n\nWhat would you like me to work on?",
+      done: false,
+      type: "info"
+    },
+    "\u2139\uFE0F  \u{1F44B} Hello! I'm ready to help.\n\nHere's what I can do:\n- Analyze your codebase\n- Execute commands\n- Answer questions\n\nWhat would you like me to work on?"
+  );
+  builder.withTextOnlyHandler("acknowledge");
+  builder.withTextWithGadgetsHandler({
+    gadgetName: "TellUser",
+    parameterMapping: (text) => ({ message: text, done: false, type: "info" }),
+    resultMapping: (text) => `\u2139\uFE0F  ${text}`
+  });
   const agent = builder.ask(prompt);
   for await (const event of agent.run()) {
     if (event.type === "text") {
@@ -6437,7 +6909,14 @@ Command rejected by user with message: "${response}"`
       printer.write(event.content);
     } else if (event.type === "gadget_result") {
       progress.pause();
-      if (stderrTTY) {
+      if (options.quiet) {
+        if (event.result.gadgetName === "TellUser" && event.result.parameters?.message) {
+          const message = String(event.result.parameters.message);
+          const rendered = renderMarkdown(message);
+          env.stdout.write(`${rendered}
+`);
+        }
+      } else {
         const tokenCount = await countGadgetOutputTokens(event.result.result);
         env.stderr.write(`${formatGadgetSummary({ ...event.result, tokenCount })}
 `);
@@ -6446,7 +6925,7 @@ Command rejected by user with message: "${response}"`
   }
   progress.complete();
   printer.ensureNewline();
-  if (stderrTTY && iterations > 1) {
+  if (!options.quiet && iterations > 1) {
     env.stderr.write(`${import_chalk3.default.dim("\u2500".repeat(40))}
 `);
     const summary = renderOverallSummary({
@@ -6519,7 +6998,7 @@ async function executeComplete(promptArg, options, env) {
   progress.endCall(usage);
   progress.complete();
   printer.ensureNewline();
-  if (stderrTTY) {
+  if (stderrTTY && !options.quiet) {
     const summary = renderSummary({ finishReason, usage, cost: progress.getTotalCost() });
     if (summary) {
       env.stderr.write(`${summary}
@@ -6540,9 +7019,102 @@ var import_node_fs3 = require("fs");
 var import_node_os = require("os");
 var import_node_path3 = require("path");
 var import_js_toml2 = require("js-toml");
+
+// src/cli/templates.ts
+var import_eta = require("eta");
+var TemplateError = class extends Error {
+  constructor(message, promptName, configPath) {
+    super(promptName ? `[prompts.${promptName}]: ${message}` : message);
+    this.promptName = promptName;
+    this.configPath = configPath;
+    this.name = "TemplateError";
+  }
+};
+function createTemplateEngine(prompts, configPath) {
+  const eta = new import_eta.Eta({
+    views: "/",
+    // Required but we use named templates
+    autoEscape: false,
+    // Don't escape - these are prompts, not HTML
+    autoTrim: false
+    // Preserve whitespace in prompts
+  });
+  for (const [name, template] of Object.entries(prompts)) {
+    try {
+      eta.loadTemplate(`@${name}`, template);
+    } catch (error) {
+      throw new TemplateError(
+        error instanceof Error ? error.message : String(error),
+        name,
+        configPath
+      );
+    }
+  }
+  return eta;
+}
+function resolveTemplate(eta, template, context = {}, configPath) {
+  try {
+    const fullContext = {
+      ...context,
+      env: process.env
+    };
+    return eta.renderString(template, fullContext);
+  } catch (error) {
+    throw new TemplateError(
+      error instanceof Error ? error.message : String(error),
+      void 0,
+      configPath
+    );
+  }
+}
+function validatePrompts(prompts, configPath) {
+  const eta = createTemplateEngine(prompts, configPath);
+  for (const [name, template] of Object.entries(prompts)) {
+    try {
+      eta.renderString(template, { env: {} });
+    } catch (error) {
+      throw new TemplateError(
+        error instanceof Error ? error.message : String(error),
+        name,
+        configPath
+      );
+    }
+  }
+}
+function validateEnvVars(template, promptName, configPath) {
+  const envVarPattern = /<%=\s*it\.env\.(\w+)\s*%>/g;
+  const matches = template.matchAll(envVarPattern);
+  for (const match of matches) {
+    const varName = match[1];
+    if (process.env[varName] === void 0) {
+      throw new TemplateError(
+        `Environment variable '${varName}' is not set`,
+        promptName,
+        configPath
+      );
+    }
+  }
+}
+function hasTemplateSyntax(str) {
+  return str.includes("<%");
+}
+
+// src/cli/config.ts
 var GLOBAL_CONFIG_KEYS = /* @__PURE__ */ new Set(["log-level", "log-file", "log-reset"]);
 var VALID_LOG_LEVELS = ["silly", "trace", "debug", "info", "warn", "error", "fatal"];
-var COMPLETE_CONFIG_KEYS = /* @__PURE__ */ new Set(["model", "system", "temperature", "max-tokens"]);
+var COMPLETE_CONFIG_KEYS = /* @__PURE__ */ new Set([
+  "model",
+  "system",
+  "temperature",
+  "max-tokens",
+  "quiet",
+  "inherits",
+  "log-level",
+  "log-file",
+  "log-reset",
+  "type"
+  // Allowed for inheritance compatibility, ignored for built-in commands
+]);
 var AGENT_CONFIG_KEYS = /* @__PURE__ */ new Set([
   "model",
   "system",
@@ -6551,16 +7123,22 @@ var AGENT_CONFIG_KEYS = /* @__PURE__ */ new Set([
   "gadget",
   "parameter-format",
   "builtins",
-  "builtin-interaction"
+  "builtin-interaction",
+  "gadget-start-prefix",
+  "gadget-end-prefix",
+  "quiet",
+  "inherits",
+  "log-level",
+  "log-file",
+  "log-reset",
+  "type"
+  // Allowed for inheritance compatibility, ignored for built-in commands
 ]);
 var CUSTOM_CONFIG_KEYS = /* @__PURE__ */ new Set([
   ...COMPLETE_CONFIG_KEYS,
   ...AGENT_CONFIG_KEYS,
   "type",
-  "description",
-  "log-level",
-  "log-file",
-  "log-reset"
+  "description"
 ]);
 var VALID_PARAMETER_FORMATS = ["json", "yaml", "toml", "auto"];
 function getConfigPath() {
@@ -6611,6 +7189,39 @@ function validateStringArray(value, key, section) {
   }
   return value;
 }
+function validateInherits(value, section) {
+  if (typeof value === "string") {
+    return value;
+  }
+  if (Array.isArray(value)) {
+    for (let i = 0; i < value.length; i++) {
+      if (typeof value[i] !== "string") {
+        throw new ConfigError(`[${section}].inherits[${i}] must be a string`);
+      }
+    }
+    return value;
+  }
+  throw new ConfigError(`[${section}].inherits must be a string or array of strings`);
+}
+function validateLoggingConfig(raw, section) {
+  const result = {};
+  if ("log-level" in raw) {
+    const level = validateString(raw["log-level"], "log-level", section);
+    if (!VALID_LOG_LEVELS.includes(level)) {
+      throw new ConfigError(
+        `[${section}].log-level must be one of: ${VALID_LOG_LEVELS.join(", ")}`
+      );
+    }
+    result["log-level"] = level;
+  }
+  if ("log-file" in raw) {
+    result["log-file"] = validateString(raw["log-file"], "log-file", section);
+  }
+  if ("log-reset" in raw) {
+    result["log-reset"] = validateBoolean(raw["log-reset"], "log-reset", section);
+  }
+  return result;
+}
 function validateBaseConfig(raw, section) {
   const result = {};
   if ("model" in raw) {
@@ -6625,6 +7236,9 @@ function validateBaseConfig(raw, section) {
       max: 2
     });
   }
+  if ("inherits" in raw) {
+    result.inherits = validateInherits(raw.inherits, section);
+  }
   return result;
 }
 function validateGlobalConfig(raw, section) {
@@ -6637,23 +7251,7 @@ function validateGlobalConfig(raw, section) {
       throw new ConfigError(`[${section}].${key} is not a valid option`);
     }
   }
-  const result = {};
-  if ("log-level" in rawObj) {
-    const level = validateString(rawObj["log-level"], "log-level", section);
-    if (!VALID_LOG_LEVELS.includes(level)) {
-      throw new ConfigError(
-        `[${section}].log-level must be one of: ${VALID_LOG_LEVELS.join(", ")}`
-      );
-    }
-    result["log-level"] = level;
-  }
-  if ("log-file" in rawObj) {
-    result["log-file"] = validateString(rawObj["log-file"], "log-file", section);
-  }
-  if ("log-reset" in rawObj) {
-    result["log-reset"] = validateBoolean(rawObj["log-reset"], "log-reset", section);
-  }
-  return result;
+  return validateLoggingConfig(rawObj, section);
 }
 function validateCompleteConfig(raw, section) {
   if (typeof raw !== "object" || raw === null) {
@@ -6665,13 +7263,19 @@ function validateCompleteConfig(raw, section) {
       throw new ConfigError(`[${section}].${key} is not a valid option`);
     }
   }
-  const result = { ...validateBaseConfig(rawObj, section) };
+  const result = {
+    ...validateBaseConfig(rawObj, section),
+    ...validateLoggingConfig(rawObj, section)
+  };
   if ("max-tokens" in rawObj) {
     result["max-tokens"] = validateNumber(rawObj["max-tokens"], "max-tokens", section, {
       integer: true,
       min: 1
     });
   }
+  if ("quiet" in rawObj) {
+    result.quiet = validateBoolean(rawObj.quiet, "quiet", section);
+  }
   return result;
 }
 function validateAgentConfig(raw, section) {
@@ -6684,7 +7288,10 @@ function validateAgentConfig(raw, section) {
       throw new ConfigError(`[${section}].${key} is not a valid option`);
     }
   }
-  const result = { ...validateBaseConfig(rawObj, section) };
+  const result = {
+    ...validateBaseConfig(rawObj, section),
+    ...validateLoggingConfig(rawObj, section)
+  };
   if ("max-iterations" in rawObj) {
     result["max-iterations"] = validateNumber(rawObj["max-iterations"], "max-iterations", section, {
       integer: true,
@@ -6713,6 +7320,23 @@ function validateAgentConfig(raw, section) {
       section
     );
   }
+  if ("gadget-start-prefix" in rawObj) {
+    result["gadget-start-prefix"] = validateString(
+      rawObj["gadget-start-prefix"],
+      "gadget-start-prefix",
+      section
+    );
+  }
+  if ("gadget-end-prefix" in rawObj) {
+    result["gadget-end-prefix"] = validateString(
+      rawObj["gadget-end-prefix"],
+      "gadget-end-prefix",
+      section
+    );
+  }
+  if ("quiet" in rawObj) {
+    result.quiet = validateBoolean(rawObj.quiet, "quiet", section);
+  }
   return result;
 }
 function validateCustomConfig(raw, section) {
@@ -6768,26 +7392,42 @@ function validateCustomConfig(raw, section) {
       section
     );
   }
+  if ("gadget-start-prefix" in rawObj) {
+    result["gadget-start-prefix"] = validateString(
+      rawObj["gadget-start-prefix"],
+      "gadget-start-prefix",
+      section
+    );
+  }
+  if ("gadget-end-prefix" in rawObj) {
+    result["gadget-end-prefix"] = validateString(
+      rawObj["gadget-end-prefix"],
+      "gadget-end-prefix",
+      section
+    );
+  }
   if ("max-tokens" in rawObj) {
     result["max-tokens"] = validateNumber(rawObj["max-tokens"], "max-tokens", section, {
       integer: true,
       min: 1
     });
   }
-  if ("log-level" in rawObj) {
-    const level = validateString(rawObj["log-level"], "log-level", section);
-    if (!VALID_LOG_LEVELS.includes(level)) {
-      throw new ConfigError(
-        `[${section}].log-level must be one of: ${VALID_LOG_LEVELS.join(", ")}`
-      );
-    }
-    result["log-level"] = level;
+  if ("quiet" in rawObj) {
+    result.quiet = validateBoolean(rawObj.quiet, "quiet", section);
   }
-  if ("log-file" in rawObj) {
-    result["log-file"] = validateString(rawObj["log-file"], "log-file", section);
+  Object.assign(result, validateLoggingConfig(rawObj, section));
+  return result;
+}
+function validatePromptsConfig(raw, section) {
+  if (typeof raw !== "object" || raw === null) {
+    throw new ConfigError(`[${section}] must be a table`);
   }
-  if ("log-reset" in rawObj) {
-    result["log-reset"] = validateBoolean(rawObj["log-reset"], "log-reset", section);
+  const result = {};
+  for (const [key, value] of Object.entries(raw)) {
+    if (typeof value !== "string") {
+      throw new ConfigError(`[${section}].${key} must be a string`);
+    }
+    result[key] = value;
   }
   return result;
 }
@@ -6805,6 +7445,8 @@ function validateConfig(raw, configPath) {
         result.complete = validateCompleteConfig(value, key);
       } else if (key === "agent") {
         result.agent = validateAgentConfig(value, key);
+      } else if (key === "prompts") {
+        result.prompts = validatePromptsConfig(value, key);
       } else {
         result[key] = validateCustomConfig(value, key);
       }
@@ -6840,12 +7482,119 @@ function loadConfig() {
       configPath
     );
   }
-  return validateConfig(raw, configPath);
+  const validated = validateConfig(raw, configPath);
+  const inherited = resolveInheritance(validated, configPath);
+  return resolveTemplatesInConfig(inherited, configPath);
 }
 function getCustomCommandNames(config) {
-  const reserved = /* @__PURE__ */ new Set(["global", "complete", "agent"]);
+  const reserved = /* @__PURE__ */ new Set(["global", "complete", "agent", "prompts"]);
   return Object.keys(config).filter((key) => !reserved.has(key));
 }
+function resolveTemplatesInConfig(config, configPath) {
+  const prompts = config.prompts ?? {};
+  const hasPrompts = Object.keys(prompts).length > 0;
+  let hasTemplates = false;
+  for (const [sectionName, section] of Object.entries(config)) {
+    if (sectionName === "global" || sectionName === "prompts") continue;
+    if (!section || typeof section !== "object") continue;
+    const sectionObj = section;
+    if (typeof sectionObj.system === "string" && hasTemplateSyntax(sectionObj.system)) {
+      hasTemplates = true;
+      break;
+    }
+  }
+  for (const template of Object.values(prompts)) {
+    if (hasTemplateSyntax(template)) {
+      hasTemplates = true;
+      break;
+    }
+  }
+  if (!hasPrompts && !hasTemplates) {
+    return config;
+  }
+  try {
+    validatePrompts(prompts, configPath);
+  } catch (error) {
+    if (error instanceof TemplateError) {
+      throw new ConfigError(error.message, configPath);
+    }
+    throw error;
+  }
+  for (const [name, template] of Object.entries(prompts)) {
+    try {
+      validateEnvVars(template, name, configPath);
+    } catch (error) {
+      if (error instanceof TemplateError) {
+        throw new ConfigError(error.message, configPath);
+      }
+      throw error;
+    }
+  }
+  const eta = createTemplateEngine(prompts, configPath);
+  const result = { ...config };
+  for (const [sectionName, section] of Object.entries(config)) {
+    if (sectionName === "global" || sectionName === "prompts") continue;
+    if (!section || typeof section !== "object") continue;
+    const sectionObj = section;
+    if (typeof sectionObj.system === "string" && hasTemplateSyntax(sectionObj.system)) {
+      try {
+        validateEnvVars(sectionObj.system, void 0, configPath);
+      } catch (error) {
+        if (error instanceof TemplateError) {
+          throw new ConfigError(`[${sectionName}].system: ${error.message}`, configPath);
+        }
+        throw error;
+      }
+      try {
+        const resolved = resolveTemplate(eta, sectionObj.system, {}, configPath);
+        result[sectionName] = {
+          ...sectionObj,
+          system: resolved
+        };
+      } catch (error) {
+        if (error instanceof TemplateError) {
+          throw new ConfigError(`[${sectionName}].system: ${error.message}`, configPath);
+        }
+        throw error;
+      }
+    }
+  }
+  return result;
+}
+function resolveInheritance(config, configPath) {
+  const resolved = {};
+  const resolving = /* @__PURE__ */ new Set();
+  function resolveSection(name) {
+    if (name in resolved) {
+      return resolved[name];
+    }
+    if (resolving.has(name)) {
+      throw new ConfigError(`Circular inheritance detected: ${name}`, configPath);
+    }
+    const section = config[name];
+    if (section === void 0 || typeof section !== "object") {
+      throw new ConfigError(`Cannot inherit from unknown section: ${name}`, configPath);
+    }
+    resolving.add(name);
+    const sectionObj = section;
+    const inheritsRaw = sectionObj.inherits;
+    const inheritsList = inheritsRaw ? Array.isArray(inheritsRaw) ? inheritsRaw : [inheritsRaw] : [];
+    let merged = {};
+    for (const parent of inheritsList) {
+      const parentResolved = resolveSection(parent);
+      merged = { ...merged, ...parentResolved };
+    }
+    const { inherits: _inherits, ...ownValues } = sectionObj;
+    merged = { ...merged, ...ownValues };
+    resolving.delete(name);
+    resolved[name] = merged;
+    return merged;
+  }
+  for (const name of Object.keys(config)) {
+    resolveSection(name);
+  }
+  return resolved;
+}
 
 // src/cli/models-command.ts
 var import_chalk4 = __toESM(require("chalk"), 1);