llmist 0.7.0 → 1.0.0

package/dist/cli.cjs

@@ -32,12 +32,13 @@ 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 GADGET_START_PREFIX, GADGET_END_PREFIX, GADGET_ARG_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";
+    GADGET_ARG_PREFIX = "!!!ARG:";
     DEFAULT_GADGET_OUTPUT_LIMIT = true;
     DEFAULT_GADGET_OUTPUT_LIMIT_PERCENT = 15;
     CHARS_PER_TOKEN = 4;
@@ -370,17 +371,12 @@ var init_prompt_config = __esm({
         "EACH MARKER MUST START WITH A NEWLINE."
       ].join("\n"),
       criticalUsage: "INVOKE gadgets using the markers - do not describe what you want to do.",
-      formatDescriptionYaml: "Parameters in YAML format (one per line)",
-      formatDescriptionJson: "Parameters in JSON format (valid JSON object)",
-      formatDescriptionToml: "Parameters in TOML format (key = value pairs, use heredoc for multiline: key = <<<EOF ... EOF)",
+      formatDescription: (ctx) => `Parameters using ${ctx.argPrefix}name markers (value on next line(s), no escaping needed)`,
       rules: () => [
         "Output ONLY plain text with the exact markers - never use function/tool calling",
         "You can invoke multiple gadgets in a single response",
         "For dependent gadgets, invoke the first one and wait for the result"
       ],
-      schemaLabelJson: "\n\nInput Schema (JSON):",
-      schemaLabelYaml: "\n\nInput Schema (YAML):",
-      schemaLabelToml: "\n\nInput Schema (TOML):",
       customExamples: null
     };
   }
@@ -397,6 +393,7 @@ var init_messages = __esm({
       messages = [];
       startPrefix = GADGET_START_PREFIX;
       endPrefix = GADGET_END_PREFIX;
+      argPrefix = GADGET_ARG_PREFIX;
       promptConfig;
       constructor(promptConfig) {
         this.promptConfig = promptConfig ?? {};
@@ -405,26 +402,32 @@ var init_messages = __esm({
        * Set custom prefixes for gadget markers.
        * Used to configure history builder to match system prompt markers.
        */
-      withPrefixes(startPrefix, endPrefix) {
+      withPrefixes(startPrefix, endPrefix, argPrefix) {
         this.startPrefix = startPrefix;
         this.endPrefix = endPrefix;
+        if (argPrefix) {
+          this.argPrefix = argPrefix;
+        }
         return this;
       }
       addSystem(content, metadata) {
         this.messages.push({ role: "system", content, metadata });
         return this;
       }
-      addGadgets(gadgets, parameterFormat = "json", options) {
+      addGadgets(gadgets, options) {
         if (options?.startPrefix) {
           this.startPrefix = options.startPrefix;
         }
         if (options?.endPrefix) {
           this.endPrefix = options.endPrefix;
         }
+        if (options?.argPrefix) {
+          this.argPrefix = options.argPrefix;
+        }
         const context = {
-          parameterFormat,
           startPrefix: this.startPrefix,
           endPrefix: this.endPrefix,
+          argPrefix: this.argPrefix,
           gadgetCount: gadgets.length,
           gadgetNames: gadgets.map((g) => g.name ?? g.constructor.name)
         };
@@ -435,26 +438,19 @@ var init_messages = __esm({
           context
         );
         parts.push(mainInstruction);
-        parts.push(this.buildGadgetsSection(gadgets, parameterFormat));
-        parts.push(this.buildUsageSection(parameterFormat, context));
+        parts.push(this.buildGadgetsSection(gadgets));
+        parts.push(this.buildUsageSection(context));
         this.messages.push({ role: "system", content: parts.join("") });
         return this;
       }
-      buildGadgetsSection(gadgets, parameterFormat) {
+      buildGadgetsSection(gadgets) {
         const parts = [];
         parts.push("\n\nAVAILABLE GADGETS");
         parts.push("\n=================\n");
         for (const gadget of gadgets) {
           const gadgetName = gadget.name ?? gadget.constructor.name;
-          const instruction = gadget.getInstruction(parameterFormat);
-          const schemaMarkers = {
-            yaml: "\n\nInput Schema (YAML):",
-            json: "\n\nInput Schema (JSON):",
-            toml: "\n\nInput Schema (TOML):",
-            auto: "\n\nInput Schema (JSON):"
-            // auto defaults to JSON schema display
-          };
-          const schemaMarker = schemaMarkers[parameterFormat];
+          const instruction = gadget.getInstruction(this.argPrefix);
+          const schemaMarker = "\n\nInput Schema (BLOCK):";
           const schemaIndex = instruction.indexOf(schemaMarker);
           const description = (schemaIndex !== -1 ? instruction.substring(0, schemaIndex) : instruction).trim();
           const schema = schemaIndex !== -1 ? instruction.substring(schemaIndex + schemaMarker.length).trim() : "";
@@ -465,35 +461,20 @@ ${description}`);
           if (schema) {
             parts.push(`
 
-PARAMETERS (${parameterFormat.toUpperCase()}):
+PARAMETERS (BLOCK):
 ${schema}`);
           }
           parts.push("\n\n---");
         }
         return parts.join("");
       }
-      buildUsageSection(parameterFormat, context) {
+      buildUsageSection(context) {
         const parts = [];
-        const formatDescriptionMap = {
-          yaml: {
-            config: this.promptConfig.formatDescriptionYaml,
-            defaultValue: DEFAULT_PROMPTS.formatDescriptionYaml
-          },
-          json: {
-            config: this.promptConfig.formatDescriptionJson,
-            defaultValue: DEFAULT_PROMPTS.formatDescriptionJson
-          },
-          toml: {
-            config: this.promptConfig.formatDescriptionToml,
-            defaultValue: DEFAULT_PROMPTS.formatDescriptionToml
-          },
-          auto: {
-            config: this.promptConfig.formatDescriptionJson,
-            defaultValue: DEFAULT_PROMPTS.formatDescriptionJson
-          }
-        };
-        const { config, defaultValue } = formatDescriptionMap[parameterFormat];
-        const formatDescription = resolvePromptTemplate(config, defaultValue, context);
+        const formatDescription = resolvePromptTemplate(
+          this.promptConfig.formatDescription,
+          DEFAULT_PROMPTS.formatDescription,
+          context
+        );
         parts.push("\n\nHOW TO INVOKE GADGETS");
         parts.push("\n=====================\n");
         const criticalUsage = resolvePromptTemplate(
@@ -511,124 +492,90 @@ CRITICAL: ${criticalUsage}
   2. ${formatDescription}`);
         parts.push(`
   3. End marker: ${this.endPrefix}`);
-        parts.push(this.buildExamplesSection(parameterFormat, context));
+        parts.push(this.buildExamplesSection(context));
         parts.push(this.buildRulesSection(context));
         parts.push("\n");
         return parts.join("");
       }
-      buildExamplesSection(parameterFormat, context) {
+      buildExamplesSection(context) {
         if (this.promptConfig.customExamples) {
           return this.promptConfig.customExamples(context);
         }
         const parts = [];
-        const singleExamples = {
-          yaml: `${this.startPrefix}translate
-from: English
-to: Polish
-content: "Paris is the capital of France: a beautiful city."
-${this.endPrefix}`,
-          json: `${this.startPrefix}translate
-{"from": "English", "to": "Polish", "content": "Paris is the capital of France: a beautiful city."}
-${this.endPrefix}`,
-          toml: `${this.startPrefix}translate
-from = "English"
-to = "Polish"
-content = "Paris is the capital of France: a beautiful city."
-${this.endPrefix}`,
-          auto: `${this.startPrefix}translate
-{"from": "English", "to": "Polish", "content": "Paris is the capital of France: a beautiful city."}
-${this.endPrefix}`
-        };
+        const singleExample = `${this.startPrefix}translate
+${this.argPrefix}from
+English
+${this.argPrefix}to
+Polish
+${this.argPrefix}content
+Paris is the capital of France: a beautiful city.
+${this.endPrefix}`;
         parts.push(`
 
 EXAMPLE (Single Gadget):
 
-${singleExamples[parameterFormat]}`);
-        const multipleExamples = {
-          yaml: `${this.startPrefix}translate
-from: English
-to: Polish
-content: "Paris is the capital of France: a beautiful city."
-${this.endPrefix}
-${this.startPrefix}analyze
-type: economic_analysis
-matter: "Polish Economy"
-question: <<<EOF
-Analyze the following:
-- Polish arms exports 2025
-- Economic implications
-EOF
-${this.endPrefix}`,
-          json: `${this.startPrefix}translate
-{"from": "English", "to": "Polish", "content": "Paris is the capital of France: a beautiful city."}
-${this.endPrefix}
-${this.startPrefix}analyze
-{"type": "economic_analysis", "matter": "Polish Economy", "question": "Analyze the following: Polish arms exports 2025, economic implications"}
-${this.endPrefix}`,
-          toml: `${this.startPrefix}translate
-from = "English"
-to = "Polish"
-content = "Paris is the capital of France: a beautiful city."
+${singleExample}`);
+        const multipleExample = `${this.startPrefix}translate
+${this.argPrefix}from
+English
+${this.argPrefix}to
+Polish
+${this.argPrefix}content
+Paris is the capital of France: a beautiful city.
 ${this.endPrefix}
 ${this.startPrefix}analyze
-type = "economic_analysis"
-matter = "Polish Economy"
-question = <<<EOF
+${this.argPrefix}type
+economic_analysis
+${this.argPrefix}matter
+Polish Economy
+${this.argPrefix}question
 Analyze the following:
 - Polish arms exports 2025
 - Economic implications
-EOF
-${this.endPrefix}`,
-          auto: `${this.startPrefix}translate
-{"from": "English", "to": "Polish", "content": "Paris is the capital of France: a beautiful city."}
-${this.endPrefix}
-${this.startPrefix}analyze
-{"type": "economic_analysis", "matter": "Polish Economy", "question": "Analyze the following: Polish arms exports 2025, economic implications"}
-${this.endPrefix}`
-        };
+${this.endPrefix}`;
         parts.push(`
 
 EXAMPLE (Multiple Gadgets):
 
-${multipleExamples[parameterFormat]}`);
-        if (parameterFormat === "yaml") {
-          parts.push(`
-
-YAML HEREDOC SYNTAX:
-For string values with multiple lines, use heredoc syntax (<<<DELIMITER...DELIMITER):
-
-filePath: "README.md"
-content: <<<EOF
-# Project Title
+${multipleExample}`);
+        parts.push(`
 
-This content can contain:
-- Markdown lists
-- Special characters: # : -
-- Multiple paragraphs
-EOF
+BLOCK FORMAT SYNTAX:
+Block format uses ${this.argPrefix}name markers. Values are captured verbatim until the next marker.
 
-The delimiter (EOF) can be any identifier. The closing delimiter must be on its own line.
-No indentation is required for the content.`);
-        } else if (parameterFormat === "toml") {
-          parts.push(`
+${this.argPrefix}filename
+calculator.ts
+${this.argPrefix}code
+class Calculator {
+  private history: string[] = [];
 
-TOML HEREDOC SYNTAX:
-For string values with multiple lines, use heredoc syntax (<<<DELIMITER...DELIMITER):
+  add(a: number, b: number): number {
+    const result = a + b;
+    this.history.push(\`\${a} + \${b} = \${result}\`);
+    return result;
+  }
+}
 
-filePath = "README.md"
-content = <<<EOF
-# Project Title
+BLOCK FORMAT RULES:
+- Each parameter starts with ${this.argPrefix}parameterName on its own line
+- The value starts on the NEXT line after the marker
+- Value ends when the next ${this.argPrefix} or ${this.endPrefix} appears
+- NO escaping needed - write values exactly as they should appear
+- Perfect for code, JSON, markdown, or any content with special characters
 
-This content can contain:
-- Markdown lists
-- Special characters: # : -
-- Multiple paragraphs
-EOF
+NESTED OBJECTS (use / separator):
+${this.argPrefix}config/timeout
+30
+${this.argPrefix}config/retries
+3
+Produces: { "config": { "timeout": "30", "retries": "3" } }
 
-The delimiter (EOF) can be any identifier. The closing delimiter must be on its own line.
-IMPORTANT: Content inside heredoc is LITERAL - do NOT escape backticks, dollar signs, or any characters.
-NEVER use TOML triple-quote strings ("""). ALWAYS use heredoc syntax (<<<EOF...EOF) for multiline content.`);
-        }
+ARRAYS (use numeric indices):
+${this.argPrefix}items/0
+first
+${this.argPrefix}items/1
+second
+Produces: { "items": ["first", "second"] }`);
         return parts.join("");
       }
       buildRulesSection(context) {
@@ -649,8 +596,8 @@ NEVER use TOML triple-quote strings ("""). ALWAYS use heredoc syntax (<<<EOF...E
         this.messages.push({ role: "assistant", content, metadata });
         return this;
       }
-      addGadgetCall(gadget, parameters, result, parameterFormat = "json") {
-        const paramStr = this.formatParameters(parameters, parameterFormat);
+      addGadgetCall(gadget, parameters, result) {
+        const paramStr = this.formatBlockParameters(parameters, "");
         this.messages.push({
           role: "assistant",
           content: `${this.startPrefix}${gadget}
@@ -663,26 +610,32 @@ ${this.endPrefix}`
         });
         return this;
       }
-      formatParameters(parameters, format) {
-        if (format === "yaml") {
-          return Object.entries(parameters).map(([key, value]) => {
-            if (typeof value === "string") {
-              return `${key}: ${value}`;
-            }
-            return `${key}: ${JSON.stringify(value)}`;
-          }).join("\n");
-        }
-        if (format === "toml") {
-          return Object.entries(parameters).map(([key, value]) => {
-            if (typeof value === "string" && value.includes("\n")) {
-              return `${key} = <<<EOF
-${value}
-EOF`;
-            }
-            return `${key} = ${JSON.stringify(value)}`;
-          }).join("\n");
+      /**
+       * Format parameters as Block format with JSON Pointer paths.
+       * Uses the configured argPrefix for consistency with system prompt.
+       */
+      formatBlockParameters(params, prefix) {
+        const lines = [];
+        for (const [key, value] of Object.entries(params)) {
+          const fullPath = prefix ? `${prefix}/${key}` : key;
+          if (Array.isArray(value)) {
+            value.forEach((item, index) => {
+              const itemPath = `${fullPath}/${index}`;
+              if (typeof item === "object" && item !== null) {
+                lines.push(this.formatBlockParameters(item, itemPath));
+              } else {
+                lines.push(`${this.argPrefix}${itemPath}`);
+                lines.push(String(item));
+              }
+            });
+          } else if (typeof value === "object" && value !== null) {
+            lines.push(this.formatBlockParameters(value, fullPath));
+          } else {
+            lines.push(`${this.argPrefix}${fullPath}`);
+            lines.push(String(value));
+          }
         }
-        return JSON.stringify(parameters);
+        return lines.join("\n");
       }
       build() {
         return [...this.messages];
@@ -854,125 +807,72 @@ var init_schema_to_json = __esm({
 });
 
 // src/gadgets/gadget.ts
-function findSafeDelimiter(content) {
-  const lines = content.split("\n");
-  for (const delimiter of HEREDOC_DELIMITERS) {
-    const regex = new RegExp(`^${delimiter}\\s*$`);
-    const isUsed = lines.some((line) => regex.test(line));
-    if (!isUsed) {
-      return delimiter;
-    }
-  }
-  let counter = 1;
-  while (counter < 1e3) {
-    const delimiter = `HEREDOC_${counter}`;
-    const regex = new RegExp(`^${delimiter}\\s*$`);
-    const isUsed = lines.some((line) => regex.test(line));
-    if (!isUsed) {
-      return delimiter;
-    }
-    counter++;
-  }
-  return "HEREDOC_FALLBACK";
-}
-function formatYamlValue(value, indent = "") {
-  if (typeof value === "string") {
-    const lines = value.split("\n");
-    if (lines.length === 1 && !value.includes(":") && !value.startsWith("-")) {
-      return value;
-    }
-    const delimiter = findSafeDelimiter(value);
-    return `<<<${delimiter}
-${value}
-${delimiter}`;
-  }
-  if (typeof value === "number" || typeof value === "boolean") {
-    return String(value);
-  }
-  if (value === null || value === void 0) {
-    return "null";
-  }
-  if (Array.isArray(value)) {
-    if (value.length === 0) return "[]";
-    const items = value.map((item) => `${indent}- ${formatYamlValue(item, indent + "  ")}`);
-    return "\n" + items.join("\n");
-  }
-  if (typeof value === "object") {
-    const entries = Object.entries(value);
-    if (entries.length === 0) return "{}";
-    const lines = entries.map(([k, v]) => {
-      const formattedValue = formatYamlValue(v, indent + "  ");
-      if (formattedValue.startsWith("\n") || formattedValue.startsWith("|")) {
-        return `${indent}${k}: ${formattedValue}`;
-      }
-      return `${indent}${k}: ${formattedValue}`;
-    });
-    return "\n" + lines.join("\n");
-  }
-  return yaml.dump(value).trimEnd();
-}
-function formatParamsAsYaml(params) {
+function formatParamsAsBlock(params, prefix = "", argPrefix = GADGET_ARG_PREFIX) {
   const lines = [];
   for (const [key, value] of Object.entries(params)) {
-    const formattedValue = formatYamlValue(value, "");
-    if (formattedValue.startsWith("\n")) {
-      lines.push(`${key}:${formattedValue}`);
+    const fullPath = prefix ? `${prefix}/${key}` : key;
+    if (Array.isArray(value)) {
+      value.forEach((item, index) => {
+        const itemPath = `${fullPath}/${index}`;
+        if (typeof item === "object" && item !== null) {
+          lines.push(formatParamsAsBlock(item, itemPath, argPrefix));
+        } else {
+          lines.push(`${argPrefix}${itemPath}`);
+          lines.push(String(item));
+        }
+      });
+    } else if (typeof value === "object" && value !== null) {
+      lines.push(formatParamsAsBlock(value, fullPath, argPrefix));
     } else {
-      lines.push(`${key}: ${formattedValue}`);
+      lines.push(`${argPrefix}${fullPath}`);
+      lines.push(String(value));
     }
   }
   return lines.join("\n");
 }
-function formatTomlInlineTable(obj) {
-  const entries = Object.entries(obj).map(([k, v]) => `${k} = ${formatTomlValue(v)}`);
-  return `{ ${entries.join(", ")} }`;
-}
-function formatTomlValue(value) {
-  if (typeof value === "string") {
-    if (value.includes("\n")) {
-      const delimiter = findSafeDelimiter(value);
-      return `<<<${delimiter}
-${value}
-${delimiter}`;
-    }
-    return JSON.stringify(value);
-  }
-  if (typeof value === "number" || typeof value === "boolean") {
-    return String(value);
-  }
-  if (value === null || value === void 0) {
-    return '""';
-  }
-  if (Array.isArray(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 `[${items.join(", ")}]`;
-  }
-  if (typeof value === "object") {
-    return formatTomlInlineTable(value);
-  }
-  return JSON.stringify(value);
-}
-function formatParamsAsToml(params) {
+function formatSchemaAsPlainText(schema, indent = "") {
   const lines = [];
-  for (const [key, value] of Object.entries(params)) {
-    lines.push(`${key} = ${formatTomlValue(value)}`);
+  const properties = schema.properties || {};
+  const required = schema.required || [];
+  for (const [key, prop] of Object.entries(properties)) {
+    const propObj = prop;
+    const type = propObj.type;
+    const description = propObj.description;
+    const isRequired = required.includes(key);
+    const enumValues = propObj.enum;
+    let line = `${indent}- ${key}`;
+    if (type === "array") {
+      const items = propObj.items;
+      const itemType = items?.type || "any";
+      line += ` (array of ${itemType})`;
+    } else if (type === "object" && propObj.properties) {
+      line += " (object)";
+    } else {
+      line += ` (${type})`;
+    }
+    if (isRequired) {
+      line += " [required]";
+    }
+    if (description) {
+      line += `: ${description}`;
+    }
+    if (enumValues) {
+      line += ` - one of: ${enumValues.map((v) => `"${v}"`).join(", ")}`;
+    }
+    lines.push(line);
+    if (type === "object" && propObj.properties) {
+      lines.push(formatSchemaAsPlainText(propObj, indent + "  "));
+    }
   }
   return lines.join("\n");
 }
-var yaml, HEREDOC_DELIMITERS, BaseGadget;
+var BaseGadget;
 var init_gadget = __esm({
   "src/gadgets/gadget.ts"() {
     "use strict";
-    yaml = __toESM(require("js-yaml"), 1);
+    init_constants();
     init_schema_to_json();
     init_schema_validator();
-    HEREDOC_DELIMITERS = ["EOF", "END", "DOC", "CONTENT", "TEXT", "HEREDOC", "DATA", "BLOCK"];
     BaseGadget = class {
       /**
        * The name of the gadget. Used for identification when LLM calls it.
@@ -1003,19 +903,19 @@ var init_gadget = __esm({
       /**
        * Auto-generated instruction text for the LLM.
        * Combines name, description, and parameter schema into a formatted instruction.
-       * @deprecated Use getInstruction(format) instead for format-specific schemas
+       * @deprecated Use getInstruction() instead
        */
       get instruction() {
-        return this.getInstruction("yaml");
+        return this.getInstruction();
       }
       /**
-       * Generate instruction text for the LLM with format-specific schema.
+       * Generate instruction text for the LLM.
        * Combines name, description, and parameter schema into a formatted instruction.
        *
-       * @param format - Format for the schema representation ('json' | 'yaml' | 'toml' | 'auto')
+       * @param argPrefix - Optional custom argument prefix for block format examples
        * @returns Formatted instruction string
        */
-      getInstruction(format = "json") {
+      getInstruction(argPrefix) {
         const parts = [];
         parts.push(this.description);
         if (this.parameterSchema) {
@@ -1024,20 +924,12 @@ var init_gadget = __esm({
           const jsonSchema = schemaToJSONSchema(this.parameterSchema, {
             target: "draft-7"
           });
-          if (format === "json" || format === "auto") {
-            parts.push("\n\nInput Schema (JSON):");
-            parts.push(JSON.stringify(jsonSchema, null, 2));
-          } else if (format === "toml") {
-            parts.push("\n\nInput Schema (TOML):");
-            parts.push(JSON.stringify(jsonSchema, null, 2));
-          } else {
-            const yamlSchema = yaml.dump(jsonSchema).trimEnd();
-            parts.push("\n\nInput Schema (YAML):");
-            parts.push(yamlSchema);
-          }
+          parts.push("\n\nParameters:");
+          parts.push(formatSchemaAsPlainText(jsonSchema));
         }
         if (this.examples && this.examples.length > 0) {
           parts.push("\n\nExamples:");
+          const effectiveArgPrefix = argPrefix ?? GADGET_ARG_PREFIX;
           this.examples.forEach((example, index) => {
             if (index > 0) {
               parts.push("");
@@ -1046,13 +938,7 @@ var init_gadget = __esm({
               parts.push(`# ${example.comment}`);
             }
             parts.push("Input:");
-            if (format === "json" || format === "auto") {
-              parts.push(JSON.stringify(example.params, null, 2));
-            } else if (format === "toml") {
-              parts.push(formatParamsAsToml(example.params));
-            } else {
-              parts.push(formatParamsAsYaml(example.params));
-            }
+            parts.push(formatParamsAsBlock(example.params, "", effectiveArgPrefix));
             if (example.output !== void 0) {
               parts.push("Output:");
               parts.push(example.output);
@@ -1329,14 +1215,12 @@ var init_conversation_manager = __esm({
       baseMessages;
       initialMessages;
       historyBuilder;
-      parameterFormat;
       constructor(baseMessages, initialMessages, options = {}) {
         this.baseMessages = baseMessages;
         this.initialMessages = initialMessages;
-        this.parameterFormat = options.parameterFormat ?? "json";
         this.historyBuilder = new LLMMessageBuilder();
         if (options.startPrefix && options.endPrefix) {
-          this.historyBuilder.withPrefixes(options.startPrefix, options.endPrefix);
+          this.historyBuilder.withPrefixes(options.startPrefix, options.endPrefix, options.argPrefix);
         }
       }
       addUserMessage(content) {
@@ -1346,7 +1230,7 @@ var init_conversation_manager = __esm({
         this.historyBuilder.addAssistant(content);
       }
       addGadgetCall(gadgetName, parameters, result) {
-        this.historyBuilder.addGadgetCall(gadgetName, parameters, result, this.parameterFormat);
+        this.historyBuilder.addGadgetCall(gadgetName, parameters, result);
       }
       getMessages() {
         return [...this.baseMessages, ...this.initialMessages, ...this.historyBuilder.build()];
@@ -1369,7 +1253,7 @@ async function runWithHandlers(agentGenerator, handlers) {
           await handlers.onGadgetCall({
             gadgetName: event.call.gadgetName,
             parameters: event.call.parameters,
-            parametersYaml: event.call.parametersYaml
+            parametersRaw: event.call.parametersRaw
           });
         }
         break;
@@ -1646,7 +1530,7 @@ var init_executor = __esm({
             this.logger.error("Gadget parameter parse error", {
               gadgetName: call.gadgetName,
               parseError: call.parseError,
-              rawParameters: call.parametersYaml
+              rawParameters: call.parametersRaw
             });
             return {
               gadgetName: call.gadgetName,
@@ -1808,168 +1692,107 @@ var init_executor = __esm({
   }
 });
 
-// src/gadgets/parser.ts
-function preprocessYaml(yamlStr) {
-  const lines = yamlStr.split("\n");
-  const result = [];
-  let i = 0;
-  while (i < lines.length) {
-    const line = lines[i];
-    const heredocMatch = line.match(/^(\s*)([\w-]+):\s*<<<([A-Za-z_][A-Za-z0-9_]*)\s*$/);
-    if (heredocMatch) {
-      const [, indent, key, delimiter] = heredocMatch;
-      const bodyLines = [];
-      i++;
-      const closingRegex = new RegExp(`^${delimiter}\\s*$`);
-      while (i < lines.length && !closingRegex.test(lines[i])) {
-        bodyLines.push(lines[i]);
-        i++;
-      }
-      if (i < lines.length) {
-        i++;
-      }
-      result.push(`${indent}${key}: |`);
-      for (const bodyLine of bodyLines) {
-        result.push(`${indent}  ${bodyLine}`);
+// src/gadgets/block-params.ts
+function parseBlockParams(content, options) {
+  const argPrefix = options?.argPrefix ?? GADGET_ARG_PREFIX;
+  const result = {};
+  const seenPointers = /* @__PURE__ */ new Set();
+  const parts = content.split(argPrefix);
+  for (let i = 1; i < parts.length; i++) {
+    const part = parts[i];
+    const newlineIndex = part.indexOf("\n");
+    if (newlineIndex === -1) {
+      const pointer2 = part.trim();
+      if (pointer2) {
+        if (seenPointers.has(pointer2)) {
+          throw new Error(`Duplicate pointer: ${pointer2}`);
+        }
+        seenPointers.add(pointer2);
+        setByPointer(result, pointer2, "");
       }
       continue;
     }
-    const match = line.match(/^(\s*)([\w-]+):\s+(.+)$/);
-    if (match) {
-      const [, indent, key, value] = match;
-      if (value === "|" || value === ">" || value === "|-" || value === ">-") {
-        result.push(line);
-        i++;
-        const keyIndentLen2 = indent.length;
-        const blockLines = [];
-        let minContentIndent = Infinity;
-        while (i < lines.length) {
-          const blockLine = lines[i];
-          const blockIndentMatch = blockLine.match(/^(\s*)/);
-          const blockIndentLen = blockIndentMatch ? blockIndentMatch[1].length : 0;
-          if (blockLine.trim() === "") {
-            blockLines.push({ content: "", originalIndent: 0 });
-            i++;
-            continue;
-          }
-          if (blockIndentLen > keyIndentLen2) {
-            const content = blockLine.substring(blockIndentLen);
-            blockLines.push({ content, originalIndent: blockIndentLen });
-            if (content.trim().length > 0) {
-              minContentIndent = Math.min(minContentIndent, blockIndentLen);
-            }
-            i++;
-          } else {
-            break;
-          }
-        }
-        const targetIndent = keyIndentLen2 + 2;
-        for (const blockLine of blockLines) {
-          if (blockLine.content === "") {
-            result.push("");
-          } else {
-            result.push(" ".repeat(targetIndent) + blockLine.content);
-          }
-        }
-        continue;
-      }
-      if (value.startsWith('"') || value.startsWith("'") || value === "true" || value === "false" || /^-?\d+(\.\d+)?$/.test(value)) {
-        result.push(line);
-        i++;
-        continue;
-      }
-      const keyIndentLen = indent.length;
-      const continuationLines = [];
-      let j = i + 1;
-      while (j < lines.length) {
-        const nextLine = lines[j];
-        if (nextLine.trim() === "") {
-          continuationLines.push(nextLine);
-          j++;
-          continue;
-        }
-        const nextIndentMatch = nextLine.match(/^(\s*)/);
-        const nextIndentLen = nextIndentMatch ? nextIndentMatch[1].length : 0;
-        if (nextIndentLen > keyIndentLen) {
-          continuationLines.push(nextLine);
-          j++;
-        } else {
-          break;
-        }
-      }
-      if (continuationLines.length > 0 && continuationLines.some((l) => l.trim().length > 0)) {
-        result.push(`${indent}${key}: |`);
-        result.push(`${indent}  ${value}`);
-        for (const contLine of continuationLines) {
-          if (contLine.trim() === "") {
-            result.push("");
-          } else {
-            const contIndentMatch = contLine.match(/^(\s*)/);
-            const contIndent = contIndentMatch ? contIndentMatch[1] : "";
-            const contContent = contLine.substring(contIndent.length);
-            result.push(`${indent}  ${contContent}`);
-          }
-        }
-        i = j;
-        continue;
-      }
-      if (value.includes(": ") || value.endsWith(":")) {
-        const escaped = value.replace(/\\/g, "\\\\").replace(/"/g, '\\"');
-        result.push(`${indent}${key}: "${escaped}"`);
-        i++;
-        continue;
-      }
+    const pointer = part.substring(0, newlineIndex).trim();
+    let value = part.substring(newlineIndex + 1);
+    if (value.endsWith("\n")) {
+      value = value.slice(0, -1);
     }
-    result.push(line);
-    i++;
+    if (!pointer) {
+      continue;
+    }
+    if (seenPointers.has(pointer)) {
+      throw new Error(`Duplicate pointer: ${pointer}`);
+    }
+    seenPointers.add(pointer);
+    setByPointer(result, pointer, value);
   }
-  return result.join("\n");
+  return result;
 }
-function unescapeHeredocContent(content) {
-  return content.replace(/\\`/g, "`").replace(/\\\$/g, "$").replace(/\\{/g, "{").replace(/\\}/g, "}");
+function coerceValue(value) {
+  if (value.includes("\n")) {
+    return value;
+  }
+  const trimmed = value.trim();
+  if (trimmed === "true") return true;
+  if (trimmed === "false") return false;
+  if (trimmed !== "" && /^-?\d+(\.\d+)?$/.test(trimmed)) {
+    const num = Number(trimmed);
+    if (!isNaN(num) && isFinite(num)) {
+      return num;
+    }
+  }
+  return value;
 }
-function preprocessTomlHeredoc(tomlStr) {
-  const lines = tomlStr.split("\n");
-  const result = [];
-  let i = 0;
-  const heredocStartRegex = /^(\s*)([\w-]+)\s*=\s*<<<([A-Za-z_][A-Za-z0-9_]*)\s*$/;
-  while (i < lines.length) {
-    const line = lines[i];
-    const match = line.match(heredocStartRegex);
-    if (match) {
-      const [, indent, key, delimiter] = match;
-      const bodyLines = [];
-      i++;
-      const closingRegex = new RegExp(`^${delimiter}\\s*$`);
-      let foundClosing = false;
-      while (i < lines.length) {
-        const bodyLine = lines[i];
-        if (closingRegex.test(bodyLine)) {
-          foundClosing = true;
-          i++;
-          break;
-        }
-        bodyLines.push(bodyLine);
-        i++;
-      }
-      if (bodyLines.length === 0) {
-        result.push(`${indent}${key} = ''''''`);
-      } else {
-        result.push(`${indent}${key} = '''`);
-        for (let j = 0; j < bodyLines.length - 1; j++) {
-          result.push(unescapeHeredocContent(bodyLines[j]));
-        }
-        result.push(`${unescapeHeredocContent(bodyLines[bodyLines.length - 1])}'''`);
-      }
-      if (!foundClosing) {
+function setByPointer(obj, pointer, value) {
+  const segments = pointer.split("/");
+  let current = obj;
+  for (let i = 0; i < segments.length - 1; i++) {
+    const segment = segments[i];
+    const nextSegment = segments[i + 1];
+    const nextIsArrayIndex = /^\d+$/.test(nextSegment);
+    if (Array.isArray(current)) {
+      const index = parseInt(segment, 10);
+      if (isNaN(index) || index < 0) {
+        throw new Error(`Invalid array index: ${segment}`);
+      }
+      if (index > current.length) {
+        throw new Error(`Array index gap: expected ${current.length}, got ${index}`);
+      }
+      if (current[index] === void 0) {
+        current[index] = nextIsArrayIndex ? [] : {};
+      }
+      current = current[index];
+    } else {
+      const rec = current;
+      if (rec[segment] === void 0) {
+        rec[segment] = nextIsArrayIndex ? [] : {};
       }
-      continue;
+      current = rec[segment];
     }
-    result.push(line);
-    i++;
   }
-  return result.join("\n");
+  const lastSegment = segments[segments.length - 1];
+  const coercedValue = coerceValue(value);
+  if (Array.isArray(current)) {
+    const index = parseInt(lastSegment, 10);
+    if (isNaN(index) || index < 0) {
+      throw new Error(`Invalid array index: ${lastSegment}`);
+    }
+    if (index > current.length) {
+      throw new Error(`Array index gap: expected ${current.length}, got ${index}`);
+    }
+    current[index] = coercedValue;
+  } else {
+    current[lastSegment] = coercedValue;
+  }
 }
+var init_block_params = __esm({
+  "src/gadgets/block-params.ts"() {
+    "use strict";
+    init_constants();
+  }
+});
+
+// src/gadgets/parser.ts
 function stripMarkdownFences(content) {
   let cleaned = content.trim();
   const openingFence = /^```(?:toml|yaml|json)?\s*\n/i;
@@ -1978,24 +1801,23 @@ function stripMarkdownFences(content) {
   cleaned = cleaned.replace(closingFence, "");
   return cleaned.trim();
 }
-var yaml2, import_js_toml, globalInvocationCounter, StreamParser;
+var globalInvocationCounter, StreamParser;
 var init_parser = __esm({
   "src/gadgets/parser.ts"() {
     "use strict";
-    yaml2 = __toESM(require("js-yaml"), 1);
-    import_js_toml = require("js-toml");
     init_constants();
+    init_block_params();
     globalInvocationCounter = 0;
     StreamParser = class {
       buffer = "";
       lastReportedTextLength = 0;
       startPrefix;
       endPrefix;
-      parameterFormat;
+      argPrefix;
       constructor(options = {}) {
         this.startPrefix = options.startPrefix ?? GADGET_START_PREFIX;
         this.endPrefix = options.endPrefix ?? GADGET_END_PREFIX;
-        this.parameterFormat = options.parameterFormat ?? "json";
+        this.argPrefix = options.argPrefix ?? GADGET_ARG_PREFIX;
       }
       takeTextUntil(index) {
         if (index <= this.lastReportedTextLength) {
@@ -2030,43 +1852,14 @@ var init_parser = __esm({
         return `${firstLine.slice(0, maxLen)}... (${message.length} chars total)`;
       }
       /**
-       * Parse parameter string according to configured format
+       * Parse parameter string using block format
        */
       parseParameters(raw) {
         const cleaned = stripMarkdownFences(raw);
-        if (this.parameterFormat === "json") {
-          try {
-            return { parameters: JSON.parse(cleaned) };
-          } catch (error) {
-            return { parseError: this.truncateParseError(error, "JSON") };
-          }
-        }
-        if (this.parameterFormat === "yaml") {
-          try {
-            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(cleaned)) };
-          } catch (error) {
-            return { parseError: this.truncateParseError(error, "TOML") };
-          }
-        }
         try {
-          return { parameters: JSON.parse(cleaned) };
-        } catch {
-          try {
-            return { parameters: (0, import_js_toml.load)(preprocessTomlHeredoc(cleaned)) };
-          } catch {
-            try {
-              return { parameters: yaml2.load(preprocessYaml(cleaned)) };
-            } catch (error) {
-              return { parseError: this.truncateParseError(error, "auto") };
-            }
-          }
+          return { parameters: parseBlockParams(cleaned, { argPrefix: this.argPrefix }) };
+        } catch (error) {
+          return { parseError: this.truncateParseError(error, "block") };
         }
       }
       // Feed a chunk of text and get parsed events
@@ -2125,8 +1918,7 @@ var init_parser = __esm({
             call: {
               gadgetName: actualGadgetName,
               invocationId,
-              parametersYaml: parametersRaw,
-              // Keep property name for backward compatibility
+              parametersRaw,
               parameters,
               parseError
             }
@@ -2160,7 +1952,7 @@ var init_parser = __esm({
               call: {
                 gadgetName: actualGadgetName,
                 invocationId,
-                parametersYaml: parametersRaw,
+                parametersRaw,
                 parameters,
                 parseError
               }
@@ -2211,9 +2003,9 @@ var init_stream_processor = __esm({
         this.stopOnGadgetError = options.stopOnGadgetError ?? true;
         this.shouldContinueAfterError = options.shouldContinueAfterError;
         this.parser = new StreamParser({
-          parameterFormat: options.parameterFormat,
           startPrefix: options.gadgetStartPrefix,
-          endPrefix: options.gadgetEndPrefix
+          endPrefix: options.gadgetEndPrefix,
+          argPrefix: options.gadgetArgPrefix
         });
         this.executor = new GadgetExecutor(
           options.registry,
@@ -2367,7 +2159,7 @@ var init_stream_processor = __esm({
           this.logger.warn("Gadget has parse error", {
             gadgetName: call.gadgetName,
             error: call.parseError,
-            rawParameters: call.parametersYaml
+            rawParameters: call.parametersRaw
           });
           const shouldContinue = await this.checkContinueAfterError(
             call.parseError,
@@ -2603,9 +2395,9 @@ var init_agent = __esm({
       hooks;
       conversation;
       registry;
-      parameterFormat;
       gadgetStartPrefix;
       gadgetEndPrefix;
+      gadgetArgPrefix;
       onHumanInputRequired;
       textOnlyHandler;
       textWithGadgetsHandler;
@@ -2634,9 +2426,9 @@ var init_agent = __esm({
         this.temperature = options.temperature;
         this.logger = options.logger ?? createLogger({ name: "llmist:agent" });
         this.registry = options.registry;
-        this.parameterFormat = options.parameterFormat ?? "json";
         this.gadgetStartPrefix = options.gadgetStartPrefix;
         this.gadgetEndPrefix = options.gadgetEndPrefix;
+        this.gadgetArgPrefix = options.gadgetArgPrefix;
         this.onHumanInputRequired = options.onHumanInputRequired;
         this.textOnlyHandler = options.textOnlyHandler ?? "terminate";
         this.textWithGadgetsHandler = options.textWithGadgetsHandler;
@@ -2658,9 +2450,10 @@ var init_agent = __esm({
         if (options.systemPrompt) {
           baseBuilder.addSystem(options.systemPrompt);
         }
-        baseBuilder.addGadgets(this.registry.getAll(), this.parameterFormat, {
+        baseBuilder.addGadgets(this.registry.getAll(), {
           startPrefix: options.gadgetStartPrefix,
-          endPrefix: options.gadgetEndPrefix
+          endPrefix: options.gadgetEndPrefix,
+          argPrefix: options.gadgetArgPrefix
         });
         const baseMessages = baseBuilder.build();
         const initialMessages = (options.initialMessages ?? []).map((message) => ({
@@ -2668,9 +2461,9 @@ var init_agent = __esm({
           content: message.content
         }));
         this.conversation = new ConversationManager(baseMessages, initialMessages, {
-          parameterFormat: this.parameterFormat,
           startPrefix: options.gadgetStartPrefix,
-          endPrefix: options.gadgetEndPrefix
+          endPrefix: options.gadgetEndPrefix,
+          argPrefix: options.gadgetArgPrefix
         });
         this.userPromptProvided = !!options.userPrompt;
         if (options.userPrompt) {
@@ -2763,9 +2556,9 @@ var init_agent = __esm({
             const processor = new StreamProcessor({
               iteration: currentIteration,
               registry: this.registry,
-              parameterFormat: this.parameterFormat,
               gadgetStartPrefix: this.gadgetStartPrefix,
               gadgetEndPrefix: this.gadgetEndPrefix,
+              gadgetArgPrefix: this.gadgetArgPrefix,
               hooks: this.hooks,
               logger: this.logger.getSubLogger({ name: "stream-processor" }),
               onHumanInputRequired: this.onHumanInputRequired,
@@ -3050,7 +2843,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 3,
           output: 15,
-          cachedInput: 0.3
+          cachedInput: 0.3,
+          cacheWriteInput: 3.75
         },
         knowledgeCutoff: "2025-01",
         features: {
@@ -3074,7 +2868,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 1,
           output: 5,
-          cachedInput: 0.1
+          cachedInput: 0.1,
+          cacheWriteInput: 1.25
         },
         knowledgeCutoff: "2025-02",
         features: {
@@ -3098,7 +2893,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 3,
           output: 15,
-          cachedInput: 0.3
+          cachedInput: 0.3,
+          cacheWriteInput: 3.75
         },
         knowledgeCutoff: "2025-03",
         features: {
@@ -3122,7 +2918,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 3,
           output: 15,
-          cachedInput: 0.3
+          cachedInput: 0.3,
+          cacheWriteInput: 3.75
         },
         knowledgeCutoff: "2024-11",
         features: {
@@ -3146,7 +2943,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 15,
           output: 75,
-          cachedInput: 1.5
+          cachedInput: 1.5,
+          cacheWriteInput: 18.75
         },
         knowledgeCutoff: "2025-01",
         features: {
@@ -3170,7 +2968,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 15,
           output: 75,
-          cachedInput: 1.5
+          cachedInput: 1.5,
+          cacheWriteInput: 18.75
         },
         knowledgeCutoff: "2025-03",
         features: {
@@ -3193,7 +2992,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 0.8,
           output: 4,
-          cachedInput: 0.08
+          cachedInput: 0.08,
+          cacheWriteInput: 1
         },
         knowledgeCutoff: "2024-07",
         features: {
@@ -3216,7 +3016,8 @@ 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: {
@@ -3240,7 +3041,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 1,
           output: 5,
-          cachedInput: 0.1
+          cachedInput: 0.1,
+          cacheWriteInput: 1.25
         },
         knowledgeCutoff: "2025-02",
         features: {
@@ -3264,7 +3066,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 3,
           output: 15,
-          cachedInput: 0.3
+          cachedInput: 0.3,
+          cacheWriteInput: 3.75
         },
         knowledgeCutoff: "2025-01",
         features: {
@@ -3288,7 +3091,8 @@ var init_anthropic_models = __esm({
         pricing: {
           input: 5,
           output: 25,
-          cachedInput: 0.5
+          cachedInput: 0.5,
+          cacheWriteInput: 6.25
         },
         knowledgeCutoff: "2025-03",
         features: {
@@ -3403,15 +3207,27 @@ 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" } } : {}
             }
           ]
         }));
@@ -3437,15 +3253,22 @@ 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
             };
@@ -3459,7 +3282,9 @@ 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 {
@@ -3540,6 +3365,7 @@ 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",
@@ -3548,8 +3374,11 @@ 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: {
@@ -3562,9 +3391,10 @@ 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",
@@ -3573,8 +3403,11 @@ 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: {
@@ -3587,9 +3420,10 @@ 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",
@@ -3598,8 +3432,10 @@ 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: {
@@ -3612,9 +3448,10 @@ 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",
@@ -3623,8 +3460,10 @@ 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: {
@@ -3636,9 +3475,10 @@ 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",
@@ -3647,8 +3487,10 @@ 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: {
@@ -3659,9 +3501,10 @@ 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",
@@ -3670,8 +3513,8 @@ 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: {
@@ -3682,7 +3525,7 @@ 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."
         }
       }
     ];
@@ -3852,7 +3695,9 @@ 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
         };
       }
       /**
@@ -3908,10 +3753,11 @@ 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: {
@@ -3931,34 +3777,7 @@ 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
         }
       },
@@ -4038,87 +3857,337 @@ var init_openai_models = __esm({
           notes: "Fastest, most cost-efficient version for well-defined tasks",
           supportsTemperature: false
         }
-      }
-    ];
-  }
-});
-
-// src/providers/openai.ts
-function sanitizeExtra(extra, allowTemperature) {
-  if (!extra) {
-    return void 0;
-  }
-  if (allowTemperature || !Object.hasOwn(extra, "temperature")) {
-    return extra;
-  }
-  return Object.fromEntries(Object.entries(extra).filter(([key]) => key !== "temperature"));
-}
-function createOpenAIProviderFromEnv() {
-  return createProviderFromEnv("OPENAI_API_KEY", import_openai.default, OpenAIChatProvider);
-}
-var import_openai, import_tiktoken, ROLE_MAP, OpenAIChatProvider;
-var init_openai = __esm({
-  "src/providers/openai.ts"() {
-    "use strict";
-    import_openai = __toESM(require("openai"), 1);
-    import_tiktoken = require("tiktoken");
-    init_base_provider();
-    init_constants2();
-    init_openai_models();
-    init_utils();
-    ROLE_MAP = {
-      system: "system",
-      user: "user",
-      assistant: "assistant"
-    };
-    OpenAIChatProvider = class extends BaseProviderAdapter {
-      providerId = "openai";
-      supports(descriptor) {
-        return descriptor.provider === this.providerId;
-      }
-      getModelSpecs() {
-        return OPENAI_MODELS;
-      }
-      buildRequestPayload(options, descriptor, spec, messages) {
-        const { maxTokens, temperature, topP, stopSequences, extra } = options;
-        const supportsTemperature = spec?.metadata?.supportsTemperature !== false;
-        const shouldIncludeTemperature = typeof temperature === "number" && supportsTemperature;
-        const sanitizedExtra = sanitizeExtra(extra, shouldIncludeTemperature);
-        return {
-          model: descriptor.name,
-          messages: messages.map((message) => ({
-            role: ROLE_MAP[message.role],
-            content: message.content,
-            name: message.name
-          })),
-          // Only set max_completion_tokens if explicitly provided
-          // Otherwise let the API use "as much as fits" in the context window
-          ...maxTokens !== void 0 ? { max_completion_tokens: maxTokens } : {},
-          top_p: topP,
-          stop: stopSequences,
-          stream: true,
-          stream_options: { include_usage: true },
-          ...sanitizedExtra ?? {},
-          ...shouldIncludeTemperature ? { temperature } : {}
-        };
-      }
-      async executeStreamRequest(payload) {
-        const client = this.client;
-        const stream2 = await client.chat.completions.create(payload);
-        return stream2;
-      }
-      async *wrapStream(iterable) {
-        const stream2 = iterable;
-        for await (const chunk of stream2) {
-          const text = chunk.choices.map((choice) => choice.delta?.content ?? "").join("");
-          if (text) {
-            yield { text, rawEvent: chunk };
-          }
-          const finishReason = chunk.choices.find((choice) => choice.finish_reason)?.finish_reason;
+      },
+      {
+        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
+        }
+      }
+    ];
+  }
+});
+
+// src/providers/openai.ts
+function sanitizeExtra(extra, allowTemperature) {
+  if (!extra) {
+    return void 0;
+  }
+  if (allowTemperature || !Object.hasOwn(extra, "temperature")) {
+    return extra;
+  }
+  return Object.fromEntries(Object.entries(extra).filter(([key]) => key !== "temperature"));
+}
+function createOpenAIProviderFromEnv() {
+  return createProviderFromEnv("OPENAI_API_KEY", import_openai.default, OpenAIChatProvider);
+}
+var import_openai, import_tiktoken, ROLE_MAP, OpenAIChatProvider;
+var init_openai = __esm({
+  "src/providers/openai.ts"() {
+    "use strict";
+    import_openai = __toESM(require("openai"), 1);
+    import_tiktoken = require("tiktoken");
+    init_base_provider();
+    init_constants2();
+    init_openai_models();
+    init_utils();
+    ROLE_MAP = {
+      system: "system",
+      user: "user",
+      assistant: "assistant"
+    };
+    OpenAIChatProvider = class extends BaseProviderAdapter {
+      providerId = "openai";
+      supports(descriptor) {
+        return descriptor.provider === this.providerId;
+      }
+      getModelSpecs() {
+        return OPENAI_MODELS;
+      }
+      buildRequestPayload(options, descriptor, spec, messages) {
+        const { maxTokens, temperature, topP, stopSequences, extra } = options;
+        const supportsTemperature = spec?.metadata?.supportsTemperature !== false;
+        const shouldIncludeTemperature = typeof temperature === "number" && supportsTemperature;
+        const sanitizedExtra = sanitizeExtra(extra, shouldIncludeTemperature);
+        return {
+          model: descriptor.name,
+          messages: messages.map((message) => ({
+            role: ROLE_MAP[message.role],
+            content: message.content,
+            name: message.name
+          })),
+          // Only set max_completion_tokens if explicitly provided
+          // Otherwise let the API use "as much as fits" in the context window
+          ...maxTokens !== void 0 ? { max_completion_tokens: maxTokens } : {},
+          top_p: topP,
+          stop: stopSequences,
+          stream: true,
+          stream_options: { include_usage: true },
+          ...sanitizedExtra ?? {},
+          ...shouldIncludeTemperature ? { temperature } : {}
+        };
+      }
+      async executeStreamRequest(payload) {
+        const client = this.client;
+        const stream2 = await client.chat.completions.create(payload);
+        return stream2;
+      }
+      async *wrapStream(iterable) {
+        const stream2 = iterable;
+        for await (const chunk of stream2) {
+          const text = chunk.choices.map((choice) => choice.delta?.content ?? "").join("");
+          if (text) {
+            yield { text, rawEvent: chunk };
+          }
+          const finishReason = chunk.choices.find((choice) => choice.finish_reason)?.finish_reason;
           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 };
@@ -4335,20 +4404,28 @@ 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"
@@ -4747,9 +4824,9 @@ var init_builder = __esm({
       gadgets = [];
       initialMessages = [];
       onHumanInputRequired;
-      parameterFormat;
       gadgetStartPrefix;
       gadgetEndPrefix;
+      gadgetArgPrefix;
       textOnlyHandler;
       textWithGadgetsHandler;
       stopOnGadgetError;
@@ -4936,21 +5013,6 @@ var init_builder = __esm({
         this.onHumanInputRequired = handler;
         return this;
       }
-      /**
-       * Set the parameter format for gadget calls.
-       *
-       * @param format - Parameter format ("json" or "xml")
-       * @returns This builder for chaining
-       *
-       * @example
-       * ```typescript
-       * .withParameterFormat("xml")
-       * ```
-       */
-      withParameterFormat(format) {
-        this.parameterFormat = format;
-        return this;
-      }
       /**
        * Set custom gadget marker prefix.
        *
@@ -4981,6 +5043,21 @@ var init_builder = __esm({
         this.gadgetEndPrefix = suffix;
         return this;
       }
+      /**
+       * Set custom argument prefix for block format parameters.
+       *
+       * @param prefix - Custom prefix for argument markers (default: "!!!ARG:")
+       * @returns This builder for chaining
+       *
+       * @example
+       * ```typescript
+       * .withGadgetArgPrefix("<<ARG>>")
+       * ```
+       */
+      withGadgetArgPrefix(prefix) {
+        this.gadgetArgPrefix = prefix;
+        return this;
+      }
       /**
        * Set the text-only handler strategy.
        *
@@ -5180,8 +5257,7 @@ var init_builder = __esm({
       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);
+        const paramStr = this.formatBlockParameters(parameters, "");
         this.initialMessages.push({
           role: "assistant",
           content: `${startPrefix}${gadgetName}
@@ -5195,25 +5271,31 @@ ${endPrefix}`
         return this;
       }
       /**
-       * Format parameters for synthetic gadget calls.
-       * Uses heredoc for multiline string values.
+       * Format parameters as block format with JSON Pointer paths.
        */
-      formatSyntheticParameters(parameters, format) {
-        if (format === "json" || format === "auto") {
-          return JSON.stringify(parameters);
-        }
-        return Object.entries(parameters).map(([key, value]) => {
-          if (typeof value === "string" && value.includes("\n")) {
-            const separator = format === "yaml" ? ":" : " =";
-            return `${key}${separator} <<<EOF
-${value}
-EOF`;
-          }
-          if (format === "yaml") {
-            return typeof value === "string" ? `${key}: ${value}` : `${key}: ${JSON.stringify(value)}`;
+      formatBlockParameters(params, prefix) {
+        const lines = [];
+        const argPrefix = this.gadgetArgPrefix ?? GADGET_ARG_PREFIX;
+        for (const [key, value] of Object.entries(params)) {
+          const fullPath = prefix ? `${prefix}/${key}` : key;
+          if (Array.isArray(value)) {
+            value.forEach((item, index) => {
+              const itemPath = `${fullPath}/${index}`;
+              if (typeof item === "object" && item !== null) {
+                lines.push(this.formatBlockParameters(item, itemPath));
+              } else {
+                lines.push(`${argPrefix}${itemPath}`);
+                lines.push(String(item));
+              }
+            });
+          } else if (typeof value === "object" && value !== null) {
+            lines.push(this.formatBlockParameters(value, fullPath));
+          } else {
+            lines.push(`${argPrefix}${fullPath}`);
+            lines.push(String(value));
           }
-          return `${key} = ${JSON.stringify(value)}`;
-        }).join("\n");
+        }
+        return lines.join("\n");
       }
       /**
        * Build and create the agent with the given user prompt.
@@ -5253,9 +5335,9 @@ EOF`;
           promptConfig: this.promptConfig,
           initialMessages: this.initialMessages,
           onHumanInputRequired: this.onHumanInputRequired,
-          parameterFormat: this.parameterFormat,
           gadgetStartPrefix: this.gadgetStartPrefix,
           gadgetEndPrefix: this.gadgetEndPrefix,
+          gadgetArgPrefix: this.gadgetArgPrefix,
           textOnlyHandler: this.textOnlyHandler,
           textWithGadgetsHandler: this.textWithGadgetsHandler,
           stopOnGadgetError: this.stopOnGadgetError,
@@ -5355,9 +5437,9 @@ EOF`;
           promptConfig: this.promptConfig,
           initialMessages: this.initialMessages,
           onHumanInputRequired: this.onHumanInputRequired,
-          parameterFormat: this.parameterFormat,
           gadgetStartPrefix: this.gadgetStartPrefix,
           gadgetEndPrefix: this.gadgetEndPrefix,
+          gadgetArgPrefix: this.gadgetArgPrefix,
           textOnlyHandler: this.textOnlyHandler,
           textWithGadgetsHandler: this.textWithGadgetsHandler,
           stopOnGadgetError: this.stopOnGadgetError,
@@ -5382,7 +5464,6 @@ var COMMANDS = {
 };
 var LOG_LEVELS = ["silly", "trace", "debug", "info", "warn", "error", "fatal"];
 var DEFAULT_MODEL = "openai:gpt-5-nano";
-var DEFAULT_PARAMETER_FORMAT = "toml";
 var OPTION_FLAGS = {
   model: "-m, --model <identifier>",
   systemPrompt: "-s, --system <prompt>",
@@ -5390,12 +5471,14 @@ var OPTION_FLAGS = {
   maxTokens: "--max-tokens <count>",
   maxIterations: "-i, --max-iterations <count>",
   gadgetModule: "-g, --gadget <module>",
-  parameterFormat: "--parameter-format <format>",
   logLevel: "--log-level <level>",
   logFile: "--log-file <path>",
   logReset: "--log-reset",
+  logLlmRequests: "--log-llm-requests [dir]",
+  logLlmResponses: "--log-llm-responses [dir]",
   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.",
@@ -5404,22 +5487,24 @@ var OPTION_DESCRIPTIONS = {
   maxTokens: "Maximum number of output tokens requested from the model.",
   maxIterations: "Maximum number of agent loop iterations before exiting.",
   gadgetModule: "Path or module specifier for a gadget export. Repeat to register multiple gadgets.",
-  parameterFormat: "Format for gadget parameter schemas: 'json', 'yaml', 'toml', or 'auto'.",
   logLevel: "Log level: silly, trace, debug, info, warn, error, fatal.",
   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.",
+  logLlmRequests: "Save raw LLM requests as plain text. Optional dir, defaults to ~/.llmist/logs/requests/",
+  logLlmResponses: "Save raw LLM responses as plain text. Optional dir, defaults to ~/.llmist/logs/responses/",
   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]";
 
 // src/cli/program.ts
-var import_commander3 = require("commander");
+var import_commander2 = require("commander");
 
 // package.json
 var package_default = {
   name: "llmist",
-  version: "0.6.2",
+  version: "0.8.0",
   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",
@@ -5503,6 +5588,7 @@ 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",
@@ -5531,7 +5617,7 @@ var package_default = {
 };
 
 // src/cli/agent-command.ts
-var import_promises = require("readline/promises");
+var import_promises2 = require("readline/promises");
 var import_chalk3 = __toESM(require("chalk"), 1);
 init_builder();
 init_registry();
@@ -5622,12 +5708,19 @@ 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("~")) {
@@ -5664,7 +5757,7 @@ function extractGadgetsFromModule(moduleExports) {
       return;
     }
     visited.add(value);
-    if (value instanceof BaseGadget) {
+    if (value instanceof BaseGadget || isGadgetLike(value)) {
       results.push(value);
       return;
     }
@@ -5713,8 +5806,33 @@ async function loadGadgets(specifiers, cwd, importer = (specifier) => import(spe
   return gadgets;
 }
 
-// src/cli/option-helpers.ts
-var import_commander2 = require("commander");
+// src/cli/llm-logging.ts
+var import_promises = require("fs/promises");
+var import_node_os = require("os");
+var import_node_path3 = require("path");
+var DEFAULT_LLM_LOG_DIR = (0, import_node_path3.join)((0, import_node_os.homedir)(), ".llmist", "logs");
+function resolveLogDir(option, subdir) {
+  if (option === true) {
+    return (0, import_node_path3.join)(DEFAULT_LLM_LOG_DIR, subdir);
+  }
+  if (typeof option === "string") {
+    return option;
+  }
+  return void 0;
+}
+function formatLlmRequest(messages) {
+  const lines = [];
+  for (const msg of messages) {
+    lines.push(`=== ${msg.role.toUpperCase()} ===`);
+    lines.push(msg.content ?? "");
+    lines.push("");
+  }
+  return lines.join("\n");
+}
+async function writeLogFile(dir, filename, content) {
+  await (0, import_promises.mkdir)(dir, { recursive: true });
+  await (0, import_promises.writeFile)((0, import_node_path3.join)(dir, filename), content, "utf-8");
+}
 
 // src/cli/utils.ts
 var import_chalk2 = __toESM(require("chalk"), 1);
@@ -5758,9 +5876,29 @@ function ensureMarkedConfigured() {
 }
 function renderMarkdown(text) {
   ensureMarkedConfigured();
-  const rendered = import_marked.marked.parse(text);
+  let rendered = import_marked.marked.parse(text);
+  rendered = rendered.replace(/\*\*(.+?)\*\*/g, (_, content) => import_chalk.default.bold(content)).replace(/(?<!\*)\*(\S[^*]*)\*(?!\*)/g, (_, content) => import_chalk.default.italic(content));
   return rendered.trimEnd();
 }
+function createRainbowSeparator() {
+  const colors = [import_chalk.default.red, import_chalk.default.yellow, import_chalk.default.green, import_chalk.default.cyan, import_chalk.default.blue, import_chalk.default.magenta];
+  const char = "\u2500";
+  const width = process.stdout.columns || 80;
+  let result = "";
+  for (let i = 0; i < width; i++) {
+    result += colors[i % colors.length](char);
+  }
+  return result;
+}
+function renderMarkdownWithSeparators(text) {
+  const rendered = renderMarkdown(text);
+  const separator = createRainbowSeparator();
+  return `
+${separator}
+${rendered}
+${separator}
+`;
+}
 function formatTokens(tokens) {
   return tokens >= 1e3 ? `${(tokens / 1e3).toFixed(1)}k` : `${tokens}`;
 }
@@ -5789,8 +5927,14 @@ 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(` ${formatTokens(outputTokens)}`));
   }
   if (metadata.elapsedSeconds !== void 0 && metadata.elapsedSeconds > 0) {
@@ -5874,7 +6018,7 @@ function formatGadgetSummary(result) {
   const summaryLine = `${icon} ${gadgetLabel}${paramsLabel} ${import_chalk.default.dim("\u2192")} ${outputLabel} ${timeLabel}`;
   if (result.gadgetName === "TellUser" && result.parameters?.message) {
     const message = String(result.parameters.message);
-    const rendered = renderMarkdown(message);
+    const rendered = renderMarkdownWithSeparators(message);
     return `${summaryLine}
 ${rendered}`;
   }
@@ -5959,6 +6103,9 @@ 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;
@@ -5984,11 +6131,13 @@ 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++;
@@ -6000,7 +6149,9 @@ 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;
@@ -6040,6 +6191,16 @@ 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
@@ -6104,11 +6265,32 @@ 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 = [];
@@ -6143,7 +6325,7 @@ var StreamProgress = class {
     }
     this.isRunning = false;
     if (this.hasRendered) {
-      this.target.write("\r\x1B[K");
+      this.target.write("\r\x1B[K\x1B[0G");
       this.hasRendered = false;
     }
   }
@@ -6236,16 +6418,6 @@ async function executeAction(action, env) {
 }
 
 // src/cli/option-helpers.ts
-var PARAMETER_FORMAT_VALUES = ["json", "yaml", "toml", "auto"];
-function parseParameterFormat(value) {
-  const normalized = value.toLowerCase();
-  if (!PARAMETER_FORMAT_VALUES.includes(normalized)) {
-    throw new import_commander2.InvalidArgumentError(
-      `Parameter format must be one of: ${PARAMETER_FORMAT_VALUES.join(", ")}`
-    );
-  }
-  return normalized;
-}
 function addCompleteOptions(cmd, defaults) {
   return cmd.option(OPTION_FLAGS.model, OPTION_DESCRIPTIONS.model, defaults?.model ?? DEFAULT_MODEL).option(OPTION_FLAGS.systemPrompt, OPTION_DESCRIPTIONS.systemPrompt, defaults?.system).option(
     OPTION_FLAGS.temperature,
@@ -6257,7 +6429,7 @@ 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).option(OPTION_FLAGS.logLlmRequests, OPTION_DESCRIPTIONS.logLlmRequests, defaults?.["log-llm-requests"]).option(OPTION_FLAGS.logLlmResponses, OPTION_DESCRIPTIONS.logLlmResponses, defaults?.["log-llm-responses"]);
 }
 function addAgentOptions(cmd, defaults) {
   const gadgetAccumulator = (value, previous = []) => [
@@ -6277,16 +6449,11 @@ function addAgentOptions(cmd, defaults) {
     defaults?.["max-iterations"]
   ).option(OPTION_FLAGS.gadgetModule, OPTION_DESCRIPTIONS.gadgetModule, gadgetAccumulator, [
     ...defaultGadgets
-  ]).option(
-    OPTION_FLAGS.parameterFormat,
-    OPTION_DESCRIPTIONS.parameterFormat,
-    parseParameterFormat,
-    defaults?.["parameter-format"] ?? DEFAULT_PARAMETER_FORMAT
-  ).option(OPTION_FLAGS.noBuiltins, OPTION_DESCRIPTIONS.noBuiltins, defaults?.builtins !== false).option(
+  ]).option(OPTION_FLAGS.noBuiltins, OPTION_DESCRIPTIONS.noBuiltins, defaults?.builtins !== false).option(
     OPTION_FLAGS.noBuiltinInteraction,
     OPTION_DESCRIPTIONS.noBuiltinInteraction,
     defaults?.["builtin-interaction"] !== false
-  );
+  ).option(OPTION_FLAGS.quiet, OPTION_DESCRIPTIONS.quiet, defaults?.quiet).option(OPTION_FLAGS.logLlmRequests, OPTION_DESCRIPTIONS.logLlmRequests, defaults?.["log-llm-requests"]).option(OPTION_FLAGS.logLlmResponses, OPTION_DESCRIPTIONS.logLlmResponses, defaults?.["log-llm-responses"]);
 }
 function configToCompleteOptions(config) {
   const result = {};
@@ -6294,6 +6461,9 @@ 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;
+  if (config["log-llm-requests"] !== void 0) result.logLlmRequests = config["log-llm-requests"];
+  if (config["log-llm-responses"] !== void 0) result.logLlmResponses = config["log-llm-responses"];
   return result;
 }
 function configToAgentOptions(config) {
@@ -6303,7 +6473,6 @@ function configToAgentOptions(config) {
   if (config.temperature !== void 0) result.temperature = config.temperature;
   if (config["max-iterations"] !== void 0) result.maxIterations = config["max-iterations"];
   if (config.gadget !== void 0) result.gadget = config.gadget;
-  if (config["parameter-format"] !== void 0) result.parameterFormat = config["parameter-format"];
   if (config.builtins !== void 0) result.builtins = config.builtins;
   if (config["builtin-interaction"] !== void 0)
     result.builtinInteraction = config["builtin-interaction"];
@@ -6311,12 +6480,17 @@ function configToAgentOptions(config) {
     result.gadgetStartPrefix = config["gadget-start-prefix"];
   if (config["gadget-end-prefix"] !== void 0)
     result.gadgetEndPrefix = config["gadget-end-prefix"];
+  if (config["gadget-arg-prefix"] !== void 0)
+    result.gadgetArgPrefix = config["gadget-arg-prefix"];
+  if (config.quiet !== void 0) result.quiet = config.quiet;
+  if (config["log-llm-requests"] !== void 0) result.logLlmRequests = config["log-llm-requests"];
+  if (config["log-llm-responses"] !== void 0) result.logLlmResponses = config["log-llm-responses"];
   return result;
 }
 
 // src/cli/agent-command.ts
 async function promptApproval(env, prompt) {
-  const rl = (0, import_promises.createInterface)({ input: env.stdin, output: env.stderr });
+  const rl = (0, import_promises2.createInterface)({ input: env.stdin, output: env.stderr });
   try {
     const answer = await rl.question(prompt);
     return answer.trim();
@@ -6331,10 +6505,10 @@ function createHumanInputHandler(env, progress) {
   }
   return async (question) => {
     progress.pause();
-    const rl = (0, import_promises.createInterface)({ input: env.stdin, output: env.stdout });
+    const rl = (0, import_promises2.createInterface)({ input: env.stdin, output: env.stdout });
     try {
       const questionLine = question.trim() ? `
-${renderMarkdown(question.trim())}` : "";
+${renderMarkdownWithSeparators(question.trim())}` : "";
       let isFirst = true;
       while (true) {
         const statsPrompt = progress.formatPrompt();
@@ -6356,9 +6530,10 @@ 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);
@@ -6376,6 +6551,9 @@ async function executeAgent(promptArg, options, env) {
   const progress = new StreamProgress(env.stderr, stderrTTY, client.modelRegistry);
   let usage;
   let iterations = 0;
+  const llmRequestsDir = resolveLogDir(options.logLlmRequests, "requests");
+  const llmResponsesDir = resolveLogDir(options.logLlmResponses, "responses");
+  let llmCallCounter = 0;
   const countMessagesTokens = async (model, messages) => {
     try {
       return await client.countTokens(model, messages);
@@ -6398,12 +6576,18 @@ async function executeAgent(promptArg, options, env) {
       // onLLMCallStart: Start progress indicator for each LLM call
       // This showcases how to react to agent lifecycle events
       onLLMCallStart: async (context) => {
+        llmCallCounter++;
         const inputTokens = await countMessagesTokens(
           context.options.model,
           context.options.messages
         );
         progress.startCall(context.options.model, inputTokens);
         progress.setInputTokens(inputTokens, false);
+        if (llmRequestsDir) {
+          const filename = `${Date.now()}_call_${llmCallCounter}.request.txt`;
+          const content = formatLlmRequest(context.options.messages);
+          await writeLogFile(llmRequestsDir, filename, content);
+        }
       },
       // onStreamChunk: Real-time updates as LLM generates tokens
       // This enables responsive UIs that show progress during generation
@@ -6416,6 +6600,10 @@ 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
@@ -6434,11 +6622,13 @@ 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 {
@@ -6446,7 +6636,7 @@ 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,
@@ -6460,6 +6650,10 @@ async function executeAgent(promptArg, options, env) {
 `);
           }
         }
+        if (llmResponsesDir) {
+          const filename = `${Date.now()}_call_${llmCallCounter}.response.txt`;
+          await writeLogFile(llmResponsesDir, filename, context.rawResponse);
+        }
       }
     },
     // SHOWCASE: Controller-based approval gating for dangerous gadgets
@@ -6524,13 +6718,15 @@ Command rejected by user with message: "${response}"`
   if (gadgets.length > 0) {
     builder.withGadgets(...gadgets);
   }
-  builder.withParameterFormat(options.parameterFormat);
   if (options.gadgetStartPrefix) {
     builder.withGadgetStartPrefix(options.gadgetStartPrefix);
   }
   if (options.gadgetEndPrefix) {
     builder.withGadgetEndPrefix(options.gadgetEndPrefix);
   }
+  if (options.gadgetArgPrefix) {
+    builder.withGadgetArgPrefix(options.gadgetArgPrefix);
+  }
   builder.withSyntheticGadgetCall(
     "TellUser",
     {
@@ -6547,22 +6743,38 @@ Command rejected by user with message: "${response}"`
     resultMapping: (text) => `\u2139\uFE0F  ${text}`
   });
   const agent = builder.ask(prompt);
+  let textBuffer = "";
+  const flushTextBuffer = () => {
+    if (textBuffer) {
+      const output = options.quiet ? textBuffer : renderMarkdownWithSeparators(textBuffer);
+      printer.write(output);
+      textBuffer = "";
+    }
+  };
   for await (const event of agent.run()) {
     if (event.type === "text") {
       progress.pause();
-      printer.write(event.content);
+      textBuffer += event.content;
     } else if (event.type === "gadget_result") {
+      flushTextBuffer();
       progress.pause();
-      if (stderrTTY) {
+      if (options.quiet) {
+        if (event.result.gadgetName === "TellUser" && event.result.parameters?.message) {
+          const message = String(event.result.parameters.message);
+          env.stdout.write(`${message}
+`);
+        }
+      } else {
         const tokenCount = await countGadgetOutputTokens(event.result.result);
         env.stderr.write(`${formatGadgetSummary({ ...event.result, tokenCount })}
 `);
       }
     }
   }
+  flushTextBuffer();
   progress.complete();
   printer.ensureNewline();
-  if (stderrTTY && iterations > 1) {
+  if (!options.quiet && iterations > 1) {
     env.stderr.write(`${import_chalk3.default.dim("\u2500".repeat(40))}
 `);
     const summary = renderOverallSummary({
@@ -6598,9 +6810,18 @@ async function executeComplete(promptArg, options, env) {
     builder.addSystem(options.system);
   }
   builder.addUser(prompt);
+  const messages = builder.build();
+  const llmRequestsDir = resolveLogDir(options.logLlmRequests, "requests");
+  const llmResponsesDir = resolveLogDir(options.logLlmResponses, "responses");
+  const timestamp = Date.now();
+  if (llmRequestsDir) {
+    const filename = `${timestamp}_complete.request.txt`;
+    const content = formatLlmRequest(messages);
+    await writeLogFile(llmRequestsDir, filename, content);
+  }
   const stream2 = client.stream({
     model,
-    messages: builder.build(),
+    messages,
     temperature: options.temperature,
     maxTokens: options.maxTokens
   });
@@ -6611,7 +6832,7 @@ async function executeComplete(promptArg, options, env) {
   progress.startCall(model, estimatedInputTokens);
   let finishReason;
   let usage;
-  let totalChars = 0;
+  let accumulatedResponse = "";
   for await (const chunk of stream2) {
     if (chunk.usage) {
       usage = chunk.usage;
@@ -6624,8 +6845,8 @@ async function executeComplete(promptArg, options, env) {
     }
     if (chunk.text) {
       progress.pause();
-      totalChars += chunk.text.length;
-      progress.update(totalChars);
+      accumulatedResponse += chunk.text;
+      progress.update(accumulatedResponse.length);
       printer.write(chunk.text);
     }
     if (chunk.finishReason !== void 0) {
@@ -6635,7 +6856,11 @@ async function executeComplete(promptArg, options, env) {
   progress.endCall(usage);
   progress.complete();
   printer.ensureNewline();
-  if (stderrTTY) {
+  if (llmResponsesDir) {
+    const filename = `${timestamp}_complete.response.txt`;
+    await writeLogFile(llmResponsesDir, filename, accumulatedResponse);
+  }
+  if (stderrTTY && !options.quiet) {
     const summary = renderSummary({ finishReason, usage, cost: progress.getTotalCost() });
     if (summary) {
       env.stderr.write(`${summary}
@@ -6653,36 +6878,136 @@ function registerCompleteCommand(program, env, config) {
 
 // src/cli/config.ts
 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");
+var import_node_os2 = require("os");
+var import_node_path4 = require("path");
+var import_js_toml = 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",
+  "log-llm-requests",
+  "log-llm-responses",
+  "type"
+  // Allowed for inheritance compatibility, ignored for built-in commands
+]);
 var AGENT_CONFIG_KEYS = /* @__PURE__ */ new Set([
   "model",
   "system",
   "temperature",
   "max-iterations",
   "gadget",
-  "parameter-format",
   "builtins",
   "builtin-interaction",
   "gadget-start-prefix",
-  "gadget-end-prefix"
+  "gadget-end-prefix",
+  "gadget-arg-prefix",
+  "quiet",
+  "inherits",
+  "log-level",
+  "log-file",
+  "log-reset",
+  "log-llm-requests",
+  "log-llm-responses",
+  "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() {
-  return (0, import_node_path3.join)((0, import_node_os.homedir)(), ".llmist", "cli.toml");
+  return (0, import_node_path4.join)((0, import_node_os2.homedir)(), ".llmist", "cli.toml");
 }
 var ConfigError = class extends Error {
   constructor(message, path2) {
@@ -6729,6 +7054,39 @@ 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: ${VALID_LOG_LEVELS.join(", ")}`
+      );
+    }
+    result["log-level"] = level;
+  }
+  if ("log-file" in raw) {
+    result["log-file"] = validateString(raw["log-file"], "log-file", section);
+  }
+  if ("log-reset" in raw) {
+    result["log-reset"] = validateBoolean(raw["log-reset"], "log-reset", section);
+  }
+  return result;
+}
 function validateBaseConfig(raw, section) {
   const result = {};
   if ("model" in raw) {
@@ -6743,6 +7101,9 @@ function validateBaseConfig(raw, section) {
       max: 2
     });
   }
+  if ("inherits" in raw) {
+    result.inherits = validateInherits(raw.inherits, section);
+  }
   return result;
 }
 function validateGlobalConfig(raw, section) {
@@ -6755,23 +7116,7 @@ function validateGlobalConfig(raw, section) {
       throw new ConfigError(`[${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: ${VALID_LOG_LEVELS.join(", ")}`
-      );
-    }
-    result["log-level"] = level;
-  }
-  if ("log-file" in rawObj) {
-    result["log-file"] = validateString(rawObj["log-file"], "log-file", section);
-  }
-  if ("log-reset" in rawObj) {
-    result["log-reset"] = validateBoolean(rawObj["log-reset"], "log-reset", section);
-  }
-  return result;
+  return validateLoggingConfig(rawObj, section);
 }
 function validateCompleteConfig(raw, section) {
   if (typeof raw !== "object" || raw === null) {
@@ -6783,13 +7128,33 @@ function validateCompleteConfig(raw, section) {
       throw new ConfigError(`[${section}].${key} is not a valid option`);
     }
   }
-  const result = { ...validateBaseConfig(rawObj, section) };
+  const result = {
+    ...validateBaseConfig(rawObj, section),
+    ...validateLoggingConfig(rawObj, section)
+  };
   if ("max-tokens" in rawObj) {
     result["max-tokens"] = validateNumber(rawObj["max-tokens"], "max-tokens", section, {
       integer: true,
       min: 1
     });
   }
+  if ("quiet" in rawObj) {
+    result.quiet = validateBoolean(rawObj.quiet, "quiet", section);
+  }
+  if ("log-llm-requests" in rawObj) {
+    result["log-llm-requests"] = validateStringOrBoolean(
+      rawObj["log-llm-requests"],
+      "log-llm-requests",
+      section
+    );
+  }
+  if ("log-llm-responses" in rawObj) {
+    result["log-llm-responses"] = validateStringOrBoolean(
+      rawObj["log-llm-responses"],
+      "log-llm-responses",
+      section
+    );
+  }
   return result;
 }
 function validateAgentConfig(raw, section) {
@@ -6802,7 +7167,10 @@ function validateAgentConfig(raw, section) {
       throw new ConfigError(`[${section}].${key} is not a valid option`);
     }
   }
-  const result = { ...validateBaseConfig(rawObj, section) };
+  const result = {
+    ...validateBaseConfig(rawObj, section),
+    ...validateLoggingConfig(rawObj, section)
+  };
   if ("max-iterations" in rawObj) {
     result["max-iterations"] = validateNumber(rawObj["max-iterations"], "max-iterations", section, {
       integer: true,
@@ -6812,15 +7180,6 @@ function validateAgentConfig(raw, section) {
   if ("gadget" in rawObj) {
     result.gadget = validateStringArray(rawObj.gadget, "gadget", section);
   }
-  if ("parameter-format" in rawObj) {
-    const format = validateString(rawObj["parameter-format"], "parameter-format", section);
-    if (!VALID_PARAMETER_FORMATS.includes(format)) {
-      throw new ConfigError(
-        `[${section}].parameter-format must be one of: ${VALID_PARAMETER_FORMATS.join(", ")}`
-      );
-    }
-    result["parameter-format"] = format;
-  }
   if ("builtins" in rawObj) {
     result.builtins = validateBoolean(rawObj.builtins, "builtins", section);
   }
@@ -6845,8 +7204,38 @@ function validateAgentConfig(raw, section) {
       section
     );
   }
+  if ("gadget-arg-prefix" in rawObj) {
+    result["gadget-arg-prefix"] = validateString(
+      rawObj["gadget-arg-prefix"],
+      "gadget-arg-prefix",
+      section
+    );
+  }
+  if ("quiet" in rawObj) {
+    result.quiet = validateBoolean(rawObj.quiet, "quiet", section);
+  }
+  if ("log-llm-requests" in rawObj) {
+    result["log-llm-requests"] = validateStringOrBoolean(
+      rawObj["log-llm-requests"],
+      "log-llm-requests",
+      section
+    );
+  }
+  if ("log-llm-responses" in rawObj) {
+    result["log-llm-responses"] = validateStringOrBoolean(
+      rawObj["log-llm-responses"],
+      "log-llm-responses",
+      section
+    );
+  }
   return result;
 }
+function validateStringOrBoolean(value, field, section) {
+  if (typeof value === "string" || typeof value === "boolean") {
+    return value;
+  }
+  throw new ConfigError(`[${section}].${field} must be a string or boolean`);
+}
 function validateCustomConfig(raw, section) {
   if (typeof raw !== "object" || raw === null) {
     throw new ConfigError(`[${section}] must be a table`);
@@ -6881,15 +7270,6 @@ function validateCustomConfig(raw, section) {
   if ("gadget" in rawObj) {
     result.gadget = validateStringArray(rawObj.gadget, "gadget", section);
   }
-  if ("parameter-format" in rawObj) {
-    const format = validateString(rawObj["parameter-format"], "parameter-format", section);
-    if (!VALID_PARAMETER_FORMATS.includes(format)) {
-      throw new ConfigError(
-        `[${section}].parameter-format must be one of: ${VALID_PARAMETER_FORMATS.join(", ")}`
-      );
-    }
-    result["parameter-format"] = format;
-  }
   if ("builtins" in rawObj) {
     result.builtins = validateBoolean(rawObj.builtins, "builtins", section);
   }
@@ -6914,26 +7294,35 @@ function validateCustomConfig(raw, section) {
       section
     );
   }
+  if ("gadget-arg-prefix" in rawObj) {
+    result["gadget-arg-prefix"] = validateString(
+      rawObj["gadget-arg-prefix"],
+      "gadget-arg-prefix",
+      section
+    );
+  }
   if ("max-tokens" in rawObj) {
     result["max-tokens"] = validateNumber(rawObj["max-tokens"], "max-tokens", section, {
       integer: true,
       min: 1
     });
   }
-  if ("log-level" in rawObj) {
-    const level = validateString(rawObj["log-level"], "log-level", section);
-    if (!VALID_LOG_LEVELS.includes(level)) {
-      throw new ConfigError(
-        `[${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;
 }
@@ -6951,6 +7340,8 @@ 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);
       }
@@ -6979,19 +7370,126 @@ function loadConfig() {
   }
   let raw;
   try {
-    raw = (0, import_js_toml2.load)(content);
+    raw = (0, import_js_toml.load)(content);
   } catch (error) {
     throw new ConfigError(
       `Invalid TOML syntax: ${error instanceof Error ? error.message : "Unknown error"}`,
       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);
@@ -7288,12 +7786,12 @@ function registerCustomCommand(program, name, config, env) {
 function parseLogLevel2(value) {
   const normalized = value.toLowerCase();
   if (!LOG_LEVELS.includes(normalized)) {
-    throw new import_commander3.InvalidArgumentError(`Log level must be one of: ${LOG_LEVELS.join(", ")}`);
+    throw new import_commander2.InvalidArgumentError(`Log level must be one of: ${LOG_LEVELS.join(", ")}`);
   }
   return normalized;
 }
 function createProgram(env, config) {
-  const program = new import_commander3.Command();
+  const program = new import_commander2.Command();
   program.name(CLI_NAME).description(CLI_DESCRIPTION).version(package_default.version).option(OPTION_FLAGS.logLevel, OPTION_DESCRIPTIONS.logLevel, parseLogLevel2).option(OPTION_FLAGS.logFile, OPTION_DESCRIPTIONS.logFile).option(OPTION_FLAGS.logReset, OPTION_DESCRIPTIONS.logReset).configureOutput({
     writeOut: (str) => env.stdout.write(str),
     writeErr: (str) => env.stderr.write(str)
@@ -7314,7 +7812,7 @@ async function runCLI(overrides = {}) {
   const opts = "env" in overrides || "config" in overrides ? overrides : { env: overrides };
   const config = opts.config !== void 0 ? opts.config : loadConfig();
   const envOverrides = opts.env ?? {};
-  const preParser = new import_commander3.Command();
+  const preParser = new import_commander2.Command();
   preParser.option(OPTION_FLAGS.logLevel, OPTION_DESCRIPTIONS.logLevel, parseLogLevel2).option(OPTION_FLAGS.logFile, OPTION_DESCRIPTIONS.logFile).option(OPTION_FLAGS.logReset, OPTION_DESCRIPTIONS.logReset).allowUnknownOption().allowExcessArguments().helpOption(false);
   preParser.parse(process.argv);
   const globalOpts = preParser.opts();