[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8,"_9":10},"actionData","errors","root",{"_152":153,"_154":155},"routes/_app.x.$handle.$name",{"_94":93,"_32":45,"_102":84,"_27":103,"_13":14,"_104":-5,"_11":105,"_106":107,"_108":109,"_110":111,"_112":113,"_20":-5,"_114":-5,"_115":-5,"_116":25,"_117":118,"_119":57,"_120":-5,"_121":-5,"_122":123},"routes/_app.x.$handle.$name.branch.$branch.version.$version.diff",{"_11":12,"_13":14,"_15":16,"_17":-5,"_18":-5},"project",{"_43":44,"_32":45,"_38":46,"_47":-5,"_48":25,"_49":50,"_51":50,"_52":-5,"_53":-5,"_18":54,"_55":31,"_56":57,"_58":57,"_59":-5,"_60":61,"_62":63,"_64":-5,"_65":-5,"_66":-5,"_67":-5,"_68":69,"_70":69,"_71":72,"_73":72,"_74":72,"_75":69,"_76":69,"_77":78,"_79":80,"_81":82,"_83":25},"version",4,"changes",[19],"aiChatInfo","userId",{"_20":21,"_22":23,"_24":25,"_26":25,"_27":28,"_29":28},"file",{"_30":31,"_32":28,"_33":34,"_35":37,"_38":39,"_40":14,"_41":42,"_24":25},"previous",{"_30":31,"_32":28,"_33":34,"_35":36},"deleted",false,"deletedHere","path","main.ts","latestPath","fileId","3dada5de-b3d1-11f0-a3a2-0224a6c84d84","name","fileType","http","content","// @ts-ignore\nimport { OpenAI } from \"https://esm.town/v/std/openai?v=4\";\n// @ts-ignore\nimport { blob } from \"https://esm.town/v/std/blob?v=11\";\nimport { Hono } from \"npm:hono@4.4.12\";\n// No import for Queue is needed.\n\n/*\n * =============================================================================\n * AGENT TASK EXECUTION FRAMEWORK (ATEF)\n *\n * This Val Town application orchestrates a five-agent LLM team to execute\n * complex user requests. It uses a custom-built queue for asynchronous,\n * long-running task processing and blob storage for persistent state management.\n *\n * The novel application of OpenAI is the agentic framework itself:\n * 1. Strategist: Deconstructs the request into a Blueprint.\n * 2. Researcher: Gathers Raw Materials based on the Blueprint.\n * 3. Creator: Builds a Version 1.0 Draft from the materials.\n * 4. Refiner: \"Red Teams\" the draft against the Blueprint to create a Pristine Output.\n * 5. Communicator: Packages the Pristine Output for the target Audience.\n *\n * This composition of specialized, chain-dependent agents is the\n * cornerstone of the application's value, moving beyond single-prompt\n_ * execution to a robust, high-quality generation pipeline.\n * =============================================================================\n */\n\n// --- CONSTANTS: AI AGENT SYSTEM PROMPTS ---\n\nconst SYSTEM_PROMPT_STRATEGIST = `\nYou are The Strategist, the first agent in an LLM team, and your purpose is to create the \"Blueprint\" for any given task by deconstructing a high-level user request.\n\nSteps\n1. Analyze the Request: Receive the [USER_REQUEST] and thoroughly analyze it to identify the core objective.\n2. Deconstruct and Define: Systematically deconstruct the request into its four core components: Intent (The 'Why'), Audience (The 'Who'), Success Criteria (The 'What'), and Constraints (The 'Box').\n3. Formulate Reasoning: Document your reasoning for how you derived each of the four components from the user request. This step must be completed before generating the final output.\n4. Construct Blueprint: Assemble the defined components and your reasoning into the specified JSON output format. Do not execute the task itself.\n\nOutput Format\nYou must provide your analysis as a single JSON object. The output must contain a reasoning key and a blueprint object.\n\n{\n \"reasoning\": \"[Your detailed justification for how the Intent, Audience, Success Criteria, and Constraints were derived from the user's request. Explain your thought process here.]\",\n \"blueprint\": {\n \"intent\": \"[The 'Why': A concise statement of the task's core purpose.]\",\n \"audience\": \"[The 'Who': A description of the end-user, including details on tone, complexity, and format.]\",\n \"success_criteria\": \"[The 'What': A bulleted or numbered list of tangible, measurable definitions for success.]\",\n \"constraints\": \"[The 'Box': A bulleted or numbered list of all non-negotiable boundaries.]\"\n }\n}\n`;\n\nconst SYSTEM_PROMPT_RESEARCHER = `\nYou are the Researcher, the second agent in an LLM team, and your purpose is to gather the \"Raw Materials\" required for the task based only on the \"Blueprint\" provided by the Strategist.\n\nSteps\n1. Receive and Parse Blueprint: Ingest the [BLUEPRINT_JSON] provided by the Strategist.\n2. Analyze Blueprint: Thoroughly analyze the intent, audience, success_criteria, and constraints within the blueprint.\n3. Identify Required Materials: Based only on the blueprint, identify the specific \"Raw Materials\" needed to successfully execute the task. Categorize these materials into: Data & Information, Context, and Precedents.\n4. Formulate Reasoning: For each category, document your reasoning for why these specific materials are necessary to meet the blueprint's success_criteria and respect its constraints. This step must be completed before generating the final output.\n5. Construct Package: Assemble the identified materials and your reasoning into the specified JSON output format.\n\nOutput Format\nYou must provide your analysis as a single JSON object. The output must contain a reasoning key and a raw_materials object.\n\n{\n \"reasoning\": {\n \"data_and_information\": \"[Your justification for why this specific data is required based on the blueprint's intent and success criteria.]\",\n \"context\": \"[Your justification for why this contextual information is needed to satisfy the audience and intent.]\",\n \"precedents\": \"[Your justification for why these precedents are necessary to meet the required format and quality standards.]\"\n },\n \"raw_materials\": {\n \"data_and_information\": [\n \"[List of raw facts, figures, reports, and feedback needed.]\"\n ],\n \"context\": [\n \"[List of surrounding information needed (e.g., market conditions, historical data for comparison, competitor actions).]\"\n ],\n \"precedents\": [\n \"[List of existing templates, 'state of the art' examples, best practices, or previous versions.]\"\n ]\n }\n}\n`;\n\nconst SYSTEM_PROMPT_CREATOR = `\nYou are The Creator, the third agent in an LLM team. You receive the \"Blueprint\" from the Strategist and the \"Raw Materials\" from the Researcher, and your purpose is to execute the \"Process\" phase and build the \"Version 1.0 Draft.\"\n\nSteps\n1. Analyze Inputs: Receive and parse the [BLUEPRINT_JSON] and [RAW_MATERIALS_JSON].\n2. Verify Adherence: Cross-reference the raw_materials with the blueprint to ensure all success_criteria and constraints can be met.\n3. Formulate Creation Plan (Reasoning): Before writing the draft, you must articulate your plan. This plan serves as your reasoning and must detail: The Structure (Skeleton), The Content (Muscle), and The Presentation (Skin).\n4. Execute Draft: Construct the \"Version 1.0 Draft\" based only on your articulated plan, the blueprint, and the raw_materials.\n\nOutput Format\nYou must provide your response in two distinct parts, separated by a horizontal rule (---).\n\n## Creation Plan (Reasoning)\n\n**Skeleton:** [Your outline or structure for the draft.]\n**Muscle:** [Your plan for how the raw materials will be used to create the core content.]\n**Skin:** [Your plan for formatting and presentation based on the blueprint's audience and constraints.]\n\n---\n\n## Version 1.0 Draft\n\n[Your complete, first-draft output (e.g., the summary, the email, the code) goes here. This output must strictly adhere to the plan above and the blueprint's requirements.]\n`;\n\nconst SYSTEM_PROMPT_REFINER = `\nYou are The Refiner, the fourth agent, also known as the \"Red Team.\" Your purpose is Quality Assurance, and you must relentlessly test the \"Version 1.0 Draft\" against the original \"Blueprint\" to create a \"Pristine Output.\"\n\nSteps\n1. Receive and Parse Inputs: Ingest the [VERSION_1_DRAFT] from the Creator and the [BLUEPRINT_JSON] from the Strategist.\n2. Analyze and Compare: Systematically compare the [VERSION_1_DRAFT] against every item in the blueprint.success_criteria, blueprint.audience, and blueprint.constraints.\n3. Execute Red Team Analysis (Reasoning): Before creating the final output, you must generate a detailed analysis by performing the following three checks. This analysis is your reasoning.\n * Error Checking: Identify all factual, logical, mathematical, grammatical, or technical errors.\n * Edge Case & Scenario Analysis: Stress test the output. Ask: \"How could this be misinterpreted?\"\n * Feedback Loop Simulation: Act as the target blueprint.audience. Read the draft from their perspective and note any confusion, ambiguity, or inappropriate tone.\n4. Implement Polishing: Correct all issues identified during your Red Team Analysis.\n5. Construct Final Output: Assemble your analysis and the final, polished content into the specified JSON output format.\n\nOutput Format\nYou must provide your response as a single JSON object. The output must contain a red_team_analysis key (your reasoning) and a pristine_output key (the final, corrected content).\n\n{\n \"red_team_analysis\": [\n {\n \"finding_type\": \"[Error Checking | Edge Case Analysis | Feedback Loop Simulation]\",\n \"issue_identified\": \"[A brief description of the error, ambiguity, or missed constraint found in the Version 1.0 Draft.]\",\n \"correction_made\": \"[A brief description of how you fixed the issue to create the Pristine Output.]\"\n }\n ],\n \"pristine_output\": \"[The complete, final, and corrected content, formatted exactly as required by the blueprint.]\"\n}\n`;\n\nconst SYSTEM_PROMPT_COMMUNICATOR = `\nYou are The Communicator, the final agent in an LLM team. Your purpose is the \"Handoff\": to package the \"Pristine Output\" specifically for the \"Audience\" defined in the \"Blueprint.\"\n\nSteps\n1. Receive and Parse Inputs: Ingest the [PRISTINE_OUTPUT] from the Refiner and the [BLUEPRINT_JSON] from the Strategist.\n2. Analyze Blueprint: Thoroughly analyze the blueprint.intent and, most importantly, the blueprint.audience (including its specified tone, complexity, and format requirements).\n3. Formulate Handoff Plan (Reasoning): Before constructing the package, you must articulate your plan. This reasoning must define:\n * Packaging: Your decision for the final, correct format (e.g., a formal PDF, an interactive web component, a pull request, a concise email, or a JSON object for a delivery system) based on the blueprint.audience and blueprint.constraints.\n * Documentation & Context: The plan for any necessary accompanying text (e.g., the summary email, the README, the chart caption) that explains the intent and how to use the output.\n4. Execute Packaging: Construct the \"Final Delivery Package\" based only on your articulated plan, the blueprint, and the pristine_output.\n\nOutput Format\nYou must provide your response as a single JSON object. The output must contain a handoff_plan key (your reasoning) and a final_delivery_package key.\n\n{\n \"handoff_plan\": {\n \"packaging\": \"[Your reasoning for selecting the final output format, based explicitly on the blueprint's audience and constraints.]\",\n \"documentation\": \"[Your reasoning for how you will provide context, such as a summary, README, or instructions, to explain the intent to the user.]\"\n },\n \"final_delivery_package\": {\n \"type\": \"[The format you selected, e.g., 'email', 'json_payload', 'markdown_document']\",\n \"content\": \"[The complete, final, and packaged output. This could be a string, a nested JSON object, or other format as determined by your packaging plan.]\"\n }\n}\n`;\n\n// --- TYPESCRIPT INTERFACES ---\n\ninterface TaskStep {\n timestamp: string;\n message: string;\n}\n\ninterface TaskState {\n id: string;\n status: \"pending\" | \"running\" | \"complete\" | \"error\";\n request: string;\n steps: TaskStep[];\n strategistReasoning?: any;\n blueprint?: any;\n researcherReasoning?: any;\n rawMaterials?: any;\n creatorReasoning?: string;\n draft?: string;\n red_team_analysis?: any[];\n pristine_output?: string;\n handoff_plan?: any;\n final_delivery_package?: any;\n}\n\n// --- QUEUE & STORAGE INITIALIZATION ---\n\n/**\n * This is our custom, dependency-free queue implementation.\n * It uses the Val Town V1 Run API to asynchronously invoke\n * the exported worker function ('atef_taskQueue_v3') in this same val.\n */\nconst taskQueue = {\n async push(payload: { taskId: string; userRequest: string }) {\n // 1. Get the required environment variables.\n // \"valtown\" is the API key, \"VT_OWNER\" and \"VT_VAL_NAME\" identify this val.\n const token = Deno.env.get(\"valtown\");\n const author = Deno.env.get(\"VT_OWNER\");\n const valName = Deno.env.get(\"VT_VAL_NAME\");\n\n if (!token || !author || !valName) {\n const errorMsg =\n \"Queue configuration error: Missing valtown environment variables.\";\n console.error(errorMsg);\n // This is a critical failure. Update the task state to error.\n await updateTaskState(payload.taskId, errorMsg, { status: \"error\" });\n return; // Stop execution\n }\n\n // 2. Define the target function (the exported worker)\n const functionName = \"atef_taskQueue_v3\";\n\n // 3. Construct the Val Town API URL\n const apiUrl =\n `https://api.val.town/v1/run/@${author}.${valName}.${functionName}`;\n\n // 4. Make the asynchronous, authenticated fetch call.\n // We do NOT await this call's completion in the request handler.\n // We \"fire and forget\" and use .then/.catch to handle the\n // result of the *enqueue* action itself.\n fetch(apiUrl, {\n method: \"POST\",\n headers: {\n \"Authorization\": `Bearer ${token}`,\n \"Content-Type\": \"application/json\",\n },\n body: JSON.stringify({\n args: [payload], // The payload becomes the 'job' argument for the worker\n }),\n })\n .then(async (res) => {\n // This block executes after the API request finishes,\n // *not* when the worker job finishes.\n if (!res.ok) {\n const errorText = await res.text();\n const errorMsg =\n `Error: Failed to enqueue job. Status: ${res.status}. Response: ${errorText}`;\n console.error(errorMsg);\n await updateTaskState(payload.taskId, errorMsg, {\n status: \"error\",\n });\n } else {\n console.log(`Successfully enqueued task: ${payload.taskId}`);\n // The 'pending' state is already set, this confirms\n // the job was accepted by the API.\n }\n })\n .catch((err) => {\n // This catches network errors during the fetch itself.\n const errorMsg =\n `Error: Network failure during enqueue. ${err.message}`;\n console.error(errorMsg, err);\n // We must update the state to \"error\"\n updateTaskState(payload.taskId, errorMsg, {\n status: \"error\",\n }).catch((e) =>\n console.error(\"Failed to update task state to error:\", e)\n );\n });\n\n // 5. Return immediately.\n // The Hono handler awaiting this will now resolve,\n // allowing it to send the 202 response to the user.\n return Promise.resolve();\n },\n};\n\n// --- CORE AI AGENT RUNNERS ---\n\n/**\n * Robustly extracts a JSON object from an AI's text response.\n * This adheres to the \"Sanitize All Output\" principle.\n */\nfunction extractJson(text: string, agentName: string): any {\n // Find the first '{' and the last '}' to capture the JSON block\n const startIndex = text.indexOf(\"{\");\n const endIndex = text.lastIndexOf(\"}\");\n\n if (startIndex === -1 || endIndex === -1 || endIndex < startIndex) {\n throw new Error(\n `Agent [${agentName}] failed to return a valid JSON structure. Output: ${text}`,\n );\n }\n\n const jsonString = text.substring(startIndex, endIndex + 1);\n\n try {\n return JSON.parse(jsonString);\n } catch (e) {\n throw new Error(\n `Agent [${agentName}] failed to parse extracted JSON. Error: ${e.message}. Extracted: ${jsonString}`,\n );\n }\n}\n\n/**\n * Generic function to run an AI agent.\n */\nasync function runAgent(\n openai: OpenAI,\n systemPrompt: string,\n userMessage: string,\n): Promise {\n const completion = await openai.chat.completions.create({\n model: \"gpt-4o\",\n messages: [\n { role: \"system\", content: systemPrompt },\n { role: \"user\", content: userMessage },\n ],\n });\n return completion.choices[0].message.content;\n}\n\n// Agent 1: Strategist\nasync function runStrategist(openai: OpenAI, userRequest: string) {\n const content = await runAgent(\n openai,\n SYSTEM_PROMPT_STRATEGIST,\n userRequest,\n );\n return extractJson(content, \"Strategist\");\n}\n\n// Agent 2: Researcher\nasync function runResearcher(openai: OpenAI, blueprint: any) {\n const userMessage = JSON.stringify({ blueprint });\n const content = await runAgent(\n openai,\n SYSTEM_PROMPT_RESEARCHER,\n userMessage,\n );\n return extractJson(content, \"Researcher\");\n}\n\n// Agent 3: Creator\nasync function runCreator(openai: OpenAI, blueprint: any, rawMaterials: any) {\n const userMessage = JSON.stringify({\n blueprint,\n raw_materials: rawMaterials,\n });\n const content = await runAgent(\n openai,\n SYSTEM_PROMPT_CREATOR,\n userMessage,\n );\n\n // The Creator has a custom format: Reasoning --- Draft\n const separator = \"\\n---\\n\";\n const separatorIndex = content.indexOf(separator);\n\n if (separatorIndex === -1) {\n throw new Error(\n \"Agent [Creator] output missing '---' separator. Cannot parse draft.\",\n );\n }\n\n const reasoning = content.substring(0, separatorIndex).trim();\n const draft = content.substring(separatorIndex + separator.length).trim();\n return { reasoning, draft };\n}\n\n// Agent 4: Refiner\nasync function runRefiner(openai: OpenAI, blueprint: any, draft: string) {\n const userMessage = JSON.stringify({\n blueprint: blueprint,\n version_1_draft: draft,\n });\n const content = await runAgent(\n openai,\n SYSTEM_PROMPT_REFINER,\n userMessage,\n );\n return extractJson(content, \"Refiner\");\n}\n\n// Agent 5: Communicator\nasync function runCommunicator(\n openai: OpenAI,\n blueprint: any,\n pristine_output: string,\n) {\n const userMessage = JSON.stringify({\n blueprint: blueprint,\n pristine_output: pristine_output,\n });\n const content = await runAgent(\n openai,\n SYSTEM_PROMPT_COMMUNICATOR,\n userMessage,\n );\n return extractJson(content, \"Communicator\");\n}\n\n// --- STATE MANAGEMENT ---\n\n/**\n * Asynchronously updates the task's state in blob storage.\n */\nasync function updateTaskState(\n taskId: string,\n message: string,\n dataToMerge: Partial,\n) {\n let currentState = await blob.getJSON(taskId) as TaskState;\n\n if (!currentState) {\n // This should not happen if the job was enqueued correctly\n currentState = {\n id: taskId,\n status: \"error\",\n request: \"Unknown\",\n steps: [],\n };\n }\n\n const newStep: TaskStep = {\n timestamp: new Date().toISOString(),\n message: message,\n };\n\n const newState: TaskState = {\n ...currentState,\n ...dataToMerge,\n steps: [...currentState.steps, newStep],\n };\n\n await blob.setJSON(taskId, newState);\n}\n\n// --- QUEUE TASK EXECUTOR ---\n\n/**\n * This is the long-running queue handler that orchestrates the agent chain.\n * It is EXPORTED so the Val Town Run API can invoke it.\n */\nexport async function atef_taskQueue_v3(job: {\n taskId: string;\n userRequest: string;\n}) {\n const { taskId, userRequest } = job;\n const openai = new OpenAI();\n\n try {\n // 1. STRATEGIST\n await updateTaskState(taskId, \"Agent [Strategist] is running...\", {\n status: \"running\",\n });\n const strategistResponse = await runStrategist(openai, userRequest);\n const blueprint = strategistResponse.blueprint;\n if (!blueprint || !strategistResponse.reasoning) {\n throw new Error(\"Strategist failed to return valid blueprint.\");\n }\n\n // 2. RESEARCHER\n await updateTaskState(taskId, \"Agent [Researcher] is running...\", {\n strategistReasoning: strategistResponse.reasoning,\n blueprint: blueprint,\n });\n const researcherResponse = await runResearcher(openai, blueprint);\n const rawMaterials = researcherResponse.raw_materials;\n if (!rawMaterials || !researcherResponse.reasoning) {\n throw new Error(\"Researcher failed to return valid raw_materials.\");\n }\n\n // 3. CREATOR\n await updateTaskState(taskId, \"Agent [Creator] is running...\", {\n researcherReasoning: researcherResponse.reasoning,\n rawMaterials: rawMaterials,\n });\n const creatorResponse = await runCreator(openai, blueprint, rawMaterials);\n const draft = creatorResponse.draft;\n if (!draft || !creatorResponse.reasoning) {\n throw new Error(\"Creator failed to return a valid draft.\");\n }\n\n // 4. REFINER\n await updateTaskState(taskId, \"Agent [Refiner] is running...\", {\n creatorReasoning: creatorResponse.reasoning,\n draft: draft,\n });\n const refinerResponse = await runRefiner(openai, blueprint, draft);\n const pristine_output = refinerResponse.pristine_output;\n if (pristine_output === undefined || !refinerResponse.red_team_analysis) {\n throw new Error(\"Refiner failed to return a valid pristine_output.\");\n }\n\n // 5. COMMUNICATOR\n await updateTaskState(taskId, \"Agent [Communicator] is running...\", {\n red_team_analysis: refinerResponse.red_team_analysis,\n pristine_output: pristine_output,\n });\n const communicatorResponse = await runCommunicator(\n openai,\n blueprint,\n pristine_output,\n );\n const final_delivery_package = communicatorResponse.final_delivery_package;\n if (!final_delivery_package || !communicatorResponse.handoff_plan) {\n throw new Error(\"Communicator failed to return a valid final_package.\");\n }\n\n // FINAL: COMPLETE\n await updateTaskState(taskId, \"Task complete.\", {\n handoff_plan: communicatorResponse.handoff_plan,\n final_delivery_package: final_delivery_package,\n status: \"complete\",\n });\n } catch (error) {\n console.error(`[ATEF Error] Task ${taskId}:`, error);\n await updateTaskState(taskId, `Error: ${error.message}`, {\n status: \"error\",\n });\n }\n}\n\n// --- HTML & FRONTEND ---\n\nfunction generateHtml(sourceUrl: string) {\n const baseUrl = `${sourceUrl.endsWith(\"/\") ? sourceUrl : sourceUrl + \"/\"}`;\n return `\n\n\n\n \n \n Agent Task Execution Framework (ATEF)\n \n \n \n\n\n\n

Agentic Framework

An AI assembly line for complex task execution.

\n\n

\n \n

\n\n

\n Task is running...\n

\n\n

\n\n \n

\n\n\n\n\n `;\n}\n\n// --- HTTP ROUTER (HONO) ---\n\nconst app = new Hono();\n\n// API: Start a new task\napp.post(\"/api/v1/tasks\", async (c) => {\n try {\n const { request } = await c.req.json();\n if (typeof request !== \"string\" || !request.trim()) {\n return c.json({ error: \"Invalid 'request' field in body.\" }, 400);\n }\n\n const taskId = crypto.randomUUID();\n const initialState: TaskState = {\n id: taskId,\n status: \"pending\",\n request: request,\n steps: [{\n timestamp: new Date().toISOString(),\n message: \"Task created and queued.\",\n }],\n };\n\n // 1. Store the initial state\n await blob.setJSON(taskId, initialState);\n\n // 2. Push the job to our custom queue.\n // This await resolves *immediately* once the fetch\n // is fired, not when the job is done.\n await taskQueue.push({ taskId, userRequest: request });\n\n // 3. Return the task ID immediately\n return c.json({ taskId: taskId }, 202); // 202 Accepted\n } catch (error) {\n console.error(\"Task creation error:\", error);\n return c.json({ error: \"Failed to create task.\" }, 500);\n }\n});\n\n// API: Get task status and results\napp.get(\"/api/v1/tasks/:id\", async (c) => {\n const taskId = c.req.param(\"id\");\n try {\n const taskData = await blob.getJSON(taskId);\n if (!taskData) {\n return c.json({ error: \"Task not found.\" }, 404);\n }\n return c.json(taskData);\n } catch (error) {\n console.error(\"Task fetch error:\", error);\n return c.json({ error: \"Failed to retrieve task data.\" }, 500);\n }\n});\n\n// Frontend: Serve the HTML\napp.get(\"/\", (c) => {\n const sourceUrl = new URL(c.req.url).origin;\n return c.html(generateHtml(sourceUrl));\n});\n\nexport default app.fetch;","// @ts-ignore\nimport { OpenAI } from \"https://esm.town/v/std/openai?v=4\";\n// @ts-ignore\nimport { blob } from \"https://esm.town/v/std/blob?v=11\";\nimport { Hono } from \"npm:hono@4.4.12\";\n// No import for Queue is needed.\n\n/*\n * =============================================================================\n * AGENT TASK EXECUTION FRAMEWORK (ATEF)\n *\n * This Val Town application orchestrates a five-agent LLM team to execute\n * complex user requests. It uses a custom-built queue for asynchronous,\n * long-running task processing and blob storage for persistent state management.\n *\n * The novel application of OpenAI is the agentic framework itself:\n * 1. Strategist: Deconstructs the request into a Blueprint.\n * 2. Researcher: Gathers Raw Materials based on the Blueprint.\n * 3. Creator: Builds a Version 1.0 Draft from the materials.\n * 4. Refiner: \"Red Teams\" the draft against the Blueprint to create a Pristine Output.\n * 5. Communicator: Packages the Pristine Output for the target Audience.\n *\n * This composition of specialized, chain-dependent agents is the\n * cornerstone of the application's value, moving beyond single-prompt\n_ * execution to a robust, high-quality generation pipeline.\n * =============================================================================\n */\n\n// --- CONSTANTS: AI AGENT SYSTEM PROMPTS ---\n\nconst SYSTEM_PROMPT_STRATEGIST = `\nYou are The Strategist, the first agent in an LLM team, and your purpose is to create the \"Blueprint\" for any given task by deconstructing a high-level user request.\n\nSteps\n1. Analyze the Request: Receive the [USER_REQUEST] and thoroughly analyze it to identify the core objective.\n2. Deconstruct and Define: Systematically deconstruct the request into its four core components: Intent (The 'Why'), Audience (The 'Who'), Success Criteria (The 'What'), and Constraints (The 'Box').\n3. Formulate Reasoning: Document your reasoning for how you derived each of the four components from the user request. This step must be completed before generating the final output.\n4. Construct Blueprint: Assemble the defined components and your reasoning into the specified JSON output format. Do not execute the task itself.\n\nOutput Format\nYou must provide your analysis as a single JSON object. The output must contain a reasoning key and a blueprint object.\n\n{\n \"reasoning\": \"[Your detailed justification for how the Intent, Audience, Success Criteria, and Constraints were derived from the user's request. Explain your thought process here.]\",\n \"blueprint\": {\n \"intent\": \"[The 'Why': A concise statement of the task's core purpose.]\",\n \"audience\": \"[The 'Who': A description of the end-user, including details on tone, complexity, and format.]\",\n \"success_criteria\": \"[The 'What': A bulleted or numbered list of tangible, measurable definitions for success.]\",\n \"constraints\": \"[The 'Box': A bulleted or numbered list of all non-negotiable boundaries.]\"\n }\n}\n`;\n\nconst SYSTEM_PROMPT_RESEARCHER = `\nYou are the Researcher, the second agent in an LLM team, and your purpose is to gather the \"Raw Materials\" required for the task based only on the \"Blueprint\" provided by the Strategist.\n\nSteps\n1. Receive and Parse Blueprint: Ingest the [BLUEPRINT_JSON] provided by the Strategist.\n2. Analyze Blueprint: Thoroughly analyze the intent, audience, success_criteria, and constraints within the blueprint.\n3. Identify Required Materials: Based only on the blueprint, identify the specific \"Raw Materials\" needed to successfully execute the task. Categorize these materials into: Data & Information, Context, and Precedents.\n4. Formulate Reasoning: For each category, document your reasoning for why these specific materials are necessary to meet the blueprint's success_criteria and respect its constraints. This step must be completed before generating the final output.\n5. Construct Package: Assemble the identified materials and your reasoning into the specified JSON output format.\n\nOutput Format\nYou must provide your analysis as a single JSON object. The output must contain a reasoning key and a raw_materials object.\n\n{\n \"reasoning\": {\n \"data_and_information\": \"[Your justification for why this specific data is required based on the blueprint's intent and success criteria.]\",\n \"context\": \"[Your justification for why this contextual information is needed to satisfy the audience and intent.]\",\n \"precedents\": \"[Your justification for why these precedents are necessary to meet the required format and quality standards.]\"\n },\n \"raw_materials\": {\n \"data_and_information\": [\n \"[List of raw facts, figures, reports, and feedback needed.]\"\n ],\n \"context\": [\n \"[List of surrounding information needed (e.g., market conditions, historical data for comparison, competitor actions).]\"\n ],\n \"precedents\": [\n \"[List of existing templates, 'state of the art' examples, best practices, or previous versions.]\"\n ]\n }\n}\n`;\n\nconst SYSTEM_PROMPT_CREATOR = `\nYou are The Creator, the third agent in an LLM team. You receive the \"Blueprint\" from the Strategist and the \"Raw Materials\" from the Researcher, and your purpose is to execute the \"Process\" phase and build the \"Version 1.0 Draft.\"\n\nSteps\n1. Analyze Inputs: Receive and parse the [BLUEPRINT_JSON] and [RAW_MATERIALS_JSON].\n2. Verify Adherence: Cross-reference the raw_materials with the blueprint to ensure all success_criteria and constraints can be met.\n3. Formulate Creation Plan (Reasoning): Before writing the draft, you must articulate your plan. This plan serves as your reasoning and must detail: The Structure (Skeleton), The Content (Muscle), and The Presentation (Skin).\n4. Execute Draft: Construct the \"Version 1.0 Draft\" based only on your articulated plan, the blueprint, and the raw_materials.\n\nOutput Format\nYou must provide your response in two distinct parts, separated by a horizontal rule (---).\n\n## Creation Plan (Reasoning)\n\n**Skeleton:** [Your outline or structure for the draft.]\n**Muscle:** [Your plan for how the raw materials will be used to create the core content.]\n**Skin:** [Your plan for formatting and presentation based on the blueprint's audience and constraints.]\n\n---\n\n## Version 1.0 Draft\n\n[Your complete, first-draft output (e.g., the summary, the email, the code) goes here. This output must strictly adhere to the plan above and the blueprint's requirements.]\n`;\n\nconst SYSTEM_PROMPT_REFINER = `\nYou are The Refiner, the fourth agent, also known as the \"Red Team.\" Your purpose is Quality Assurance, and you must relentlessly test the \"Version 1.0 Draft\" against the original \"Blueprint\" to create a \"Pristine Output.\"\n\nSteps\n1. Receive and Parse Inputs: Ingest the [VERSION_1_DRAFT] from the Creator and the [BLUEPRINT_JSON] from the Strategist.\n2. Analyze and Compare: Systematically compare the [VERSION_1_DRAFT] against every item in the blueprint.success_criteria, blueprint.audience, and blueprint.constraints.\n3. Execute Red Team Analysis (Reasoning): Before creating the final output, you must generate a detailed analysis by performing the following three checks. This analysis is your reasoning.\n * Error Checking: Identify all factual, logical, mathematical, grammatical, or technical errors.\n * Edge Case & Scenario Analysis: Stress test the output. Ask: \"How could this be misinterpreted?\"\n * Feedback Loop Simulation: Act as the target blueprint.audience. Read the draft from their perspective and note any confusion, ambiguity, or inappropriate tone.\n4. Implement Polishing: Correct all issues identified during your Red Team Analysis.\n5. Construct Final Output: Assemble your analysis and the final, polished content into the specified JSON output format.\n\nOutput Format\nYou must provide your response as a single JSON object. The output must contain a red_team_analysis key (your reasoning) and a pristine_output key (the final, corrected content).\n\n{\n \"red_team_analysis\": [\n {\n \"finding_type\": \"[Error Checking | Edge Case Analysis | Feedback Loop Simulation]\",\n \"issue_identified\": \"[A brief description of the error, ambiguity, or missed constraint found in the Version 1.0 Draft.]\",\n \"correction_made\": \"[A brief description of how you fixed the issue to create the Pristine Output.]\"\n }\n ],\n \"pristine_output\": \"[The complete, final, and corrected content, formatted exactly as required by the blueprint.]\"\n}\n`;\n\nconst SYSTEM_PROMPT_COMMUNICATOR = `\nYou are The Communicator, the final agent in an LLM team. Your purpose is the \"Handoff\": to package the \"Pristine Output\" specifically for the \"Audience\" defined in the \"Blueprint.\"\n\nSteps\n1. Receive and Parse Inputs: Ingest the [PRISTINE_OUTPUT] from the Refiner and the [BLUEPRINT_JSON] from the Strategist.\n2. Analyze Blueprint: Thoroughly analyze the blueprint.intent and, most importantly, the blueprint.audience (including its specified tone, complexity, and format requirements).\n3. Formulate Handoff Plan (Reasoning): Before constructing the package, you must articulate your plan. This reasoning must define:\n * Packaging: Your decision for the final, correct format (e.g., a formal PDF, an interactive web component, a pull request, a concise email, or a JSON object for a delivery system) based on the blueprint.audience and blueprint.constraints.\n * Documentation & Context: The plan for any necessary accompanying text (e.g., the summary email, the README, the chart caption) that explains the intent and how to use the output.\n4. Execute Packaging: Construct the \"Final Delivery Package\" based only on your articulated plan, the blueprint, and the pristine_output.\n\nOutput Format\nYou must provide your response as a single JSON object. The output must contain a handoff_plan key (your reasoning) and a final_delivery_package key.\n\n{\n \"handoff_plan\": {\n \"packaging\": \"[Your reasoning for selecting the final output format, based explicitly on the blueprint's audience and constraints.]\",\n \"documentation\": \"[Your reasoning for how you will provide context, such as a summary, README, or instructions, to explain the intent to the user.]\"\n },\n \"final_delivery_package\": {\n \"type\": \"[The format you selected, e.g., 'email', 'json_payload', 'markdown_document']\",\n \"content\": \"[The complete, final, and packaged output. This could be a string, a nested JSON object, or other format as determined by your packaging plan.]\"\n }\n}\n`;\n\n// --- TYPESCRIPT INTERFACES ---\n\ninterface TaskStep {\n timestamp: string;\n message: string;\n}\n\ninterface TaskState {\n id: string;\n status: \"pending\" | \"running\" | \"complete\" | \"error\";\n request: string;\n steps: TaskStep[];\n strategistReasoning?: any;\n blueprint?: any;\n researcherReasoning?: any;\n rawMaterials?: any;\n creatorReasoning?: string;\n draft?: string;\n red_team_analysis?: any[];\n pristine_output?: string;\n handoff_plan?: any;\n final_delivery_package?: any;\n}\n\n// --- QUEUE & STORAGE INITIALIZATION ---\n\n/**\n * This is our custom, dependency-free queue implementation.\n * It uses the Val Town V1 Run API to asynchronously invoke\n * the exported worker function ('atef_taskQueue_v3') in this same val.\n */\nconst taskQueue = {\n async push(payload: { taskId: string; userRequest: string }) {\n // 1. Get the required environment variables.\n // \"valtown\" is the API key, \"VT_OWNER\" and \"VT_VAL_NAME\" identify this val.\n const token = Deno.env.get(\"valtown\");\n const author = Deno.env.get(\"VT_OWNER\");\n const valName = Deno.env.get(\"VT_VAL_NAME\");\n\n if (!token || !author || !valName) {\n const errorMsg =\n \"Queue configuration error: Missing valtown environment variables.\";\n console.error(errorMsg);\n // This is a critical failure. Update the task state to error.\n await updateTaskState(payload.taskId, errorMsg, { status: \"error\" });\n return; // Stop execution\n }\n\n // 2. Define the target function (the exported worker)\n const functionName = \"atef_taskQueue_v3\";\n\n // 3. Construct the Val Town API URL\n const apiUrl =\n `https://api.val.town/v1/run/@${author}.${valName}.${functionName}`;\n\n // 4. Make the asynchronous, authenticated fetch call.\n // We do NOT await this call's completion in the request handler.\n // We \"fire and forget\" and use .then/.catch to handle the\n // result of the *enqueue* action itself.\n fetch(apiUrl, {\n method: \"POST\",\n headers: {\n \"Authorization\": `Bearer ${token}`,\n \"Content-Type\": \"application/json\",\n },\n body: JSON.stringify({\n args: [payload], // The payload becomes the 'job' argument for the worker\n }),\n })\n .then(async (res) => {\n // This block executes after the API request finishes,\n // *not* when the worker job finishes.\n if (!res.ok) {\n const errorText = await res.text();\n const errorMsg =\n `Error: Failed to enqueue job. Status: ${res.status}. Response: ${errorText}`;\n console.error(errorMsg);\n await updateTaskState(payload.taskId, errorMsg, {\n status: \"error\",\n });\n } else {\n console.log(`Successfully enqueued task: ${payload.taskId}`);\n // The 'pending' state is already set, this confirms\n // the job was accepted by the API.\n }\n })\n .catch((err) => {\n // This catches network errors during the fetch itself.\n const errorMsg =\n `Error: Network failure during enqueue. ${err.message}`;\n console.error(errorMsg, err);\n // We must update the state to \"error\"\n updateTaskState(payload.taskId, errorMsg, {\n status: \"error\",\n }).catch((e) =>\n console.error(\"Failed to update task state to error:\", e)\n );\n });\n\n // 5. Return immediately.\n // The Hono handler awaiting this will now resolve,\n // allowing it to send the 202 response to the user.\n return Promise.resolve();\n },\n};\n\n// --- CORE AI AGENT RUNNERS ---\n\n/**\n * Robustly extracts a JSON object from an AI's text response.\n * This adheres to the \"Sanitize All Output\" principle.\n */\nfunction extractJson(text: string, agentName: string): any {\n // Find the first '{' and the last '}' to capture the JSON block\n const startIndex = text.indexOf(\"{\");\n const endIndex = text.lastIndexOf(\"}\");\n\n if (startIndex === -1 || endIndex === -1 || endIndex < startIndex) {\n throw new Error(\n `Agent [${agentName}] failed to return a valid JSON structure. Output: ${text}`,\n );\n }\n\n const jsonString = text.substring(startIndex, endIndex + 1);\n\n try {\n return JSON.parse(jsonString);\n } catch (e) {\n throw new Error(\n `Agent [${agentName}] failed to parse extracted JSON. Error: ${e.message}. Extracted: ${jsonString}`,\n );\n }\n}\n\n/**\n * Generic function to run an AI agent.\n */\nasync function runAgent(\n openai: OpenAI,\n systemPrompt: string,\n userMessage: string,\n): Promise {\n const completion = await openai.chat.completions.create({\n model: \"gpt-4o\",\n messages: [\n { role: \"system\", content: systemPrompt },\n { role: \"user\", content: userMessage },\n ],\n });\n return completion.choices[0].message.content;\n}\n\n// Agent 1: Strategist\nasync function runStrategist(openai: OpenAI, userRequest: string) {\n const content = await runAgent(\n openai,\n SYSTEM_PROMPT_STRATEGIST,\n userRequest,\n );\n return extractJson(content, \"Strategist\");\n}\n\n// Agent 2: Researcher\nasync function runResearcher(openai: OpenAI, blueprint: any) {\n const userMessage = JSON.stringify({ blueprint });\n const content = await runAgent(\n openai,\n SYSTEM_PROMPT_RESEARCHER,\n userMessage,\n );\n return extractJson(content, \"Researcher\");\n}\n\n// Agent 3: Creator\nasync function runCreator(openai: OpenAI, blueprint: any, rawMaterials: any) {\n const userMessage = JSON.stringify({\n blueprint,\n raw_materials: rawMaterials,\n });\n const content = await runAgent(\n openai,\n SYSTEM_PROMPT_CREATOR,\n userMessage,\n );\n\n // The Creator has a custom format: Reasoning --- Draft\n const separator = \"\\n---\\n\";\n const separatorIndex = content.indexOf(separator);\n\n if (separatorIndex === -1) {\n throw new Error(\n \"Agent [Creator] output missing '---' separator. Cannot parse draft.\",\n );\n }\n\n const reasoning = content.substring(0, separatorIndex).trim();\n const draft = content.substring(separatorIndex + separator.length).trim();\n return { reasoning, draft };\n}\n\n// Agent 4: Refiner\nasync function runRefiner(openai: OpenAI, blueprint: any, draft: string) {\n const userMessage = JSON.stringify({\n blueprint: blueprint,\n version_1_draft: draft,\n });\n const content = await runAgent(\n openai,\n SYSTEM_PROMPT_REFINER,\n userMessage,\n );\n return extractJson(content, \"Refiner\");\n}\n\n// Agent 5: Communicator\nasync function runCommunicator(\n openai: OpenAI,\n blueprint: any,\n pristine_output: string,\n) {\n const userMessage = JSON.stringify({\n blueprint: blueprint,\n pristine_output: pristine_output,\n });\n const content = await runAgent(\n openai,\n SYSTEM_PROMPT_COMMUNICATOR,\n userMessage,\n );\n return extractJson(content, \"Communicator\");\n}\n\n// --- STATE MANAGEMENT ---\n\n/**\n * Asynchronously updates the task's state in blob storage.\n */\nasync function updateTaskState(\n taskId: string,\n message: string,\n dataToMerge: Partial,\n) {\n let currentState = await blob.getJSON(taskId) as TaskState;\n\n if (!currentState) {\n // This should not happen if the job was enqueued correctly\n currentState = {\n id: taskId,\n status: \"error\",\n request: \"Unknown\",\n steps: [],\n };\n }\n\n const newStep: TaskStep = {\n timestamp: new Date().toISOString(),\n message: message,\n };\n\n const newState: TaskState = {\n ...currentState,\n ...dataToMerge,\n steps: [...currentState.steps, newStep],\n };\n\n await blob.setJSON(taskId, newState);\n}\n\n// --- QUEUE TASK EXECUTOR ---\n\n/**\n * This is the long-running queue handler that orchestrates the agent chain.\n * It is EXPORTED so the Val Town Run API can invoke it.\n */\nexport async function atef_taskQueue_v3(job: {\n taskId: string;\n userRequest: string;\n}) {\n const { taskId, userRequest } = job;\n const openai = new OpenAI();\n\n try {\n // 1. STRATEGIST\n await updateTaskState(taskId, \"Agent [Strategist] is running...\", {\n status: \"running\",\n });\n const strategistResponse = await runStrategist(openai, userRequest);\n const blueprint = strategistResponse.blueprint;\n if (!blueprint || !strategistResponse.reasoning) {\n throw new Error(\"Strategist failed to return valid blueprint.\");\n }\n\n // 2. RESEARCHER\n await updateTaskState(taskId, \"Agent [Researcher] is running...\", {\n strategistReasoning: strategistResponse.reasoning,\n blueprint: blueprint,\n });\n const researcherResponse = await runResearcher(openai, blueprint);\n const rawMaterials = researcherResponse.raw_materials;\n if (!rawMaterials || !researcherResponse.reasoning) {\n throw new Error(\"Researcher failed to return valid raw_materials.\");\n }\n\n // 3. CREATOR\n await updateTaskState(taskId, \"Agent [Creator] is running...\", {\n researcherReasoning: researcherResponse.reasoning,\n rawMaterials: rawMaterials,\n });\n const creatorResponse = await runCreator(openai, blueprint, rawMaterials);\n const draft = creatorResponse.draft;\n if (!draft || !creatorResponse.reasoning) {\n throw new Error(\"Creator failed to return a valid draft.\");\n }\n\n // 4. REFINER\n await updateTaskState(taskId, \"Agent [Refiner] is running...\", {\n creatorReasoning: creatorResponse.reasoning,\n draft: draft,\n });\n const refinerResponse = await runRefiner(openai, blueprint, draft);\n const pristine_output = refinerResponse.pristine_output;\n if (pristine_output === undefined || !refinerResponse.red_team_analysis) {\n throw new Error(\"Refiner failed to return a valid pristine_output.\");\n }\n\n // 5. COMMUNICATOR\n await updateTaskState(taskId, \"Agent [Communicator] is running...\", {\n red_team_analysis: refinerResponse.red_team_analysis,\n pristine_output: pristine_output,\n });\n const communicatorResponse = await runCommunicator(\n openai,\n blueprint,\n pristine_output,\n );\n const final_delivery_package = communicatorResponse.final_delivery_package;\n if (!final_delivery_package || !communicatorResponse.handoff_plan) {\n throw new Error(\"Communicator failed to return a valid final_package.\");\n }\n\n // FINAL: COMPLETE\n await updateTaskState(taskId, \"Task complete.\", {\n handoff_plan: communicatorResponse.handoff_plan,\n final_delivery_package: final_delivery_package,\n status: \"complete\",\n });\n } catch (error) {\n console.error(`[ATEF Error] Task ${taskId}:`, error);\n await updateTaskState(taskId, `Error: ${error.message}`, {\n status: \"error\",\n });\n }\n}\n\n// --- HTML & FRONTEND ---\n\nfunction generateHtml(sourceUrl: string) {\n const baseUrl = `${sourceUrl.endsWith(\"/\") ? sourceUrl : sourceUrl + \"/\"}`;\n return `\n\n\n\n \n \n Agent Task Execution Framework (ATEF)\n \n \n \n\n\n\n

Agentic Framework

An AI assembly line for complex task execution.

\n\n

\n \n

\n\n

\n Task is running...\n

\n\n

\n\n \n

Diff · v4