Declarative AI Workflow Engine

[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8,"_9":10},"actionData","errors","root",{"_153":154,"_155":156},"routes/_app.x.$handle.$name",{"_95":94,"_32":45,"_103":85,"_27":104,"_13":14,"_105":-5,"_11":106,"_107":108,"_109":110,"_111":112,"_113":114,"_20":-5,"_115":-5,"_116":-5,"_117":25,"_118":119,"_120":57,"_121":-5,"_122":-5,"_123":124},"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",7,"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.tsx","latestPath","fileId","424b236e-73ee-11f0-8547-0224a6c84d84","name","fileType","http","content","// @ts-ignore\nimport { OpenAI } from \"https://esm.town/v/std/openai?v=4\";\n// @ts-ignore\nimport { Hono } from \"npm:hono@4.4.12\";\n// @ts-ignore\nimport { cors } from \"npm:hono@4.4.12/cors\";\n// @ts-ignore\nimport Parser from \"npm:rss-parser@3.13.0\";\n\n// --- ACTION LIBRARY (PRIMITIVE CAPABILITIES) ---\nconst ACTIONS = {\n FETCH_NEWS: async (params) => {\n const parser = new Parser();\n const RSS_URL = `https://news.google.com/rss/search?q=${\n encodeURIComponent(params.query)\n } when:30d&hl=en-US&gl=US&ceid=US:en`;\n try {\n const feed = await parser.parseURL(RSS_URL);\n return feed.items.slice(0, 7).map((item) => ({\n title: item.title,\n snippet: item.contentSnippet,\n }));\n } catch {\n return [{ title: \"Could not fetch news via RSS.\" }];\n }\n },\n FETCH_REDDIT_SENTIMENT: async (params) => {\n const allPosts = [];\n const subreddits = Array.isArray(params.subreddits)\n ? params.subreddits\n : [params.subreddits || \"-\"];\n for (const subreddit of subreddits) {\n try {\n const url = `https://www.reddit.com/r/${subreddit}/search.json?q=${\n encodeURIComponent(params.query)\n }&sort=relevance&t=month&restrict_sr=on`;\n const r = await fetch(url, {\n headers: { \"User-Agent\": \"Mozilla/5.0\" },\n });\n const d = await r.json();\n allPosts.push(\n ...(d.data?.children?.map((p) => ({\n title: p.data.title,\n subreddit: p.data.subreddit,\n })) || []),\n );\n } catch {}\n }\n return allPosts.slice(0, 15);\n },\n // MODIFIED: This action now expects fully resolved prompt text.\n INVOKE_AI: async (params) => {\n if (!params.user_prompt) {\n throw new Error(\"INVOKE_AI action requires a 'user' prompt.\");\n }\n const openai = new OpenAI();\n const completion = await openai.chat.completions.create({\n model: \"gpt-4o\",\n messages: [\n {\n role: \"system\",\n content: params.system_prompt || \"You are a helpful AI assistant.\",\n },\n { role: \"user\", content: params.user_prompt },\n ],\n temperature: parseFloat(params.temperature) || 0.4,\n });\n return completion.choices[0].message.content;\n },\n};\n\n// --- CORE ENGINE ---\n\n// BUG FIX: The original parser was flawed. This new version correctly handles the declarative structure\n// and the new nested prompt format with 'system' and 'user' keys.\nfunction parseDeclarativePlan(text) {\n const lines = text.split(\"\\n\").filter((line) => !line.trim().startsWith(\"#\"));\n const plan = { VARS: {}, PROMPTS: {}, WORKFLOW: [] };\n let currentSection = null;\n let currentPromptKey = null;\n let currentPromptPart = null;\n let currentWorkflowStep = null;\n\n for (const line of lines) {\n const indent = line.match(/^\\s*/)[0].length;\n const trimmed = line.trim();\n if (!trimmed) continue;\n\n if (indent === 0) {\n currentSection = trimmed.replace(\":\", \"\");\n currentPromptKey = null;\n currentWorkflowStep = null;\n continue;\n }\n\n switch (currentSection) {\n case \"VARS\":\n const [varKey, ...varValueParts] = trimmed.split(\":\");\n plan.VARS[varKey.trim()] = varValueParts.join(\":\").trim();\n break;\n\n case \"PROMPTS\":\n if (indent === 2) {\n currentPromptKey = trimmed.replace(\":\", \"\");\n plan.PROMPTS[currentPromptKey] = {};\n currentPromptPart = null;\n } else if (indent === 4 && currentPromptKey) {\n const partKey = trimmed.split(\":\")[0].trim();\n if (partKey === \"system\" || partKey === \"user\") {\n currentPromptPart = partKey;\n plan.PROMPTS[currentPromptKey][currentPromptPart] = trimmed\n .substring(partKey.length + 1).replace(/>$/, \"\").trim();\n }\n } else if (indent > 4 && currentPromptKey && currentPromptPart) {\n plan.PROMPTS[currentPromptKey][currentPromptPart] += \"\\n\" + trimmed;\n }\n break;\n\n case \"WORKFLOW\":\n if (trimmed.startsWith(\"- id:\")) {\n if (currentWorkflowStep) plan.WORKFLOW.push(currentWorkflowStep);\n currentWorkflowStep = { params: {} };\n const [wfKey, ...wfValueParts] = trimmed.substring(2).split(\":\");\n currentWorkflowStep[wfKey.trim()] = wfValueParts.join(\":\").trim();\n } else if (currentWorkflowStep) {\n const [paramKey, ...paramValueParts] = trimmed.split(\":\");\n const value = paramValueParts.join(\":\").trim();\n if (\n paramKey.trim() === \"action\" || paramKey.trim() === \"prompt_key\" ||\n paramKey.trim() === \"temperature\"\n ) {\n currentWorkflowStep[paramKey.trim()] = value;\n } else {\n currentWorkflowStep.params[paramKey.trim()] = value;\n }\n }\n break;\n }\n }\n if (currentWorkflowStep) plan.WORKFLOW.push(currentWorkflowStep);\n return plan;\n}\n\n// Resolves {{...}} variables from the state\nfunction resolveVariables(template, state) {\n if (typeof template !== \"string\") return template;\n return template.replace(/\\{\\{(.*?)\\}\\}/g, (match, key) => {\n const keys = key.trim().split(\".\");\n let value = state;\n for (const k of keys) {\n value = value?.[k];\n }\n if (typeof value === \"object\") {\n return JSON.stringify(value, null, 2);\n }\n return value !== undefined ? value : match;\n });\n}\n\n// BUG FIX: The execution engine is now more robust in how it resolves and passes parameters.\nasync function executePlan(plan) {\n const state = { VARS: plan.VARS, PROMPTS: plan.PROMPTS };\n let finalOutput =\n \"Workflow completed, but the final step produced no output.\";\n\n for (const step of plan.WORKFLOW) {\n const action = ACTIONS[step.action];\n if (!action) throw new Error(`Unknown action: ${step.action}`);\n\n let resolvedParams = {};\n\n if (step.action === \"INVOKE_AI\") {\n const promptTemplate = state.PROMPTS[step.prompt_key];\n if (!promptTemplate) {\n throw new Error(\n `Prompt key not found in PROMPTS section: '${step.prompt_key}'`,\n );\n }\n\n resolvedParams = {\n system_prompt: resolveVariables(promptTemplate.system, state),\n user_prompt: resolveVariables(promptTemplate.user, state),\n temperature: step.temperature || \"0.4\",\n };\n } else {\n for (const key in step.params) {\n resolvedParams[key] = resolveVariables(step.params[key], state);\n }\n }\n\n console.log(`Executing step: ${step.id} with action: ${step.action}`);\n const result = await action(resolvedParams);\n state[step.id] = { result };\n finalOutput = result;\n }\n\n return finalOutput;\n}\n\n// --- HTML & FRONTEND ---\nfunction generateHtml() {\n const examplePlan = `# A declarative plan for the AI Workflow Engine\n\nVARS:\n product_idea: \"A new AI-powered mobile app for personalized meal planning that adjusts to your fitness tracker data.\"\n competitors: \"MyFitnessPal, PlateJoy, Noom\"\n\n# A library of reusable prompt templates for this plan\nPROMPTS:\n summarize_news_prompt:\n system: >\n You are a concise market analyst. Your task is to summarize the key themes, \n opportunities, and threats from the following news articles about the\n digital health and meal planning market. Focus on market saturation, \n user acquisition costs, and recent funding events.\n user: >\n ARTICLES:\n {{FETCH_NEWS_STEP.result}}\n\n critique_idea_prompt:\n system: >\n You are a skeptical Venture Capitalist from a top-tier firm like Sequoia or \n Andreessen Horowitz. Your firm is considering an investment in the product idea described\n below. Given the provided market summary, your task is to provide a sharp, data-driven\n critique of the product's viability. Identify the 3 most significant risks.\n Ground every point in the provided market summary.\n user: >\n MARKET_SUMMARY:\n {{SUMMARIZE_NEWS_STEP.result}}\n\n PRODUCT_IDEA:\n {{VARS.product_idea}}\n\n# The sequence of actions to execute\nWORKFLOW:\n - id: FETCH_NEWS_STEP\n action: FETCH_NEWS\n params:\n query: \"{{VARS.competitors}} app review\"\n\n - id: SUMMARIZE_NEWS_STEP\n action: INVOKE_AI\n prompt_key: summarize_news_prompt\n temperature: 0.2\n\n - id: CRITIQUE_IDEA_STEP\n action: INVOKE_AI\n prompt_key: critique_idea_prompt\n temperature: 0.5\n`;\n\n return `\n\n\n \n \n Declarative AI Workflow Engine\n \n \n \n \n\n\n\n

\n \n

\n\n\n`;\n}\n\n// --- HONO ROUTER & MAIN HANDLER ---\nconst app = new Hono();\napp.use(\"*\", cors());\napp.get(\"/\", (c) => c.html(generateHtml()));\n\napp.post(\"/api/execute\", async (c) => {\n try {\n const planText = await c.req.text();\n const plan = parseDeclarativePlan(planText);\n\n if (!plan.WORKFLOW || plan.WORKFLOW.length === 0) {\n return c.json({\n error:\n \"Invalid or empty workflow. Make sure to define steps under the 'WORKFLOW:' section.\",\n }, 400);\n }\n\n const finalResult = await executePlan(plan);\n\n return c.text(finalResult);\n } catch (error) {\n console.error(\"Workflow Execution Error:\", error);\n return c.json({\n error: error.message || \"An internal server error occurred.\",\n }, 500);\n }\n});\n\nACTIONS.FETCH_REDDIT_SENTIMENT = async (params) => {\n const allPosts = [];\n const subreddits = Array.isArray(params.subreddits)\n ? params.subreddits\n : [params.subreddits || \"-\"];\n for (const subreddit of subreddits) {\n try {\n const url = `https://www.reddit.com/r/${subreddit}/search.json?q=${\n encodeURIComponent(params.query)\n }&sort=relevance&t=month&restrict_sr=on`;\n const r = await fetch(url, { headers: { \"User-Agent\": \"Mozilla/5.0\" } });\n const d = await r.json();\n allPosts.push(\n ...(d.data?.children?.map((p) => ({\n title: p.data.title,\n subreddit: p.data.subreddit,\n })) || []),\n );\n } catch {}\n }\n return allPosts.slice(0, 15);\n};\n\nexport default app.fetch;","Of course. I will make the example workflow significantly more complex to showcase the full power of the declarative engine.\nThis new plan simulates a multi-stage \"Go/No-Go\" analysis for launching a new video game. It utilizes every available feature, including multiple data-fetching steps and a chain of four distinct AI invocations, each with a different persona and using the output of the previous steps as its input.\nThis demonstrates how you can orchestrate a complex research and analysis pipeline, including adversarial \"Red Team\" and supportive \"Blue Team\" viewpoints, culminating in a final executive decision memo.\nI have updated the default text in the application to this more advanced workflow.\nThe Declarative AI Workflow Engine (with Complex Example)\n// @ts-ignore\nimport { OpenAI } from \"https://esm.town/v/std/openai?v=4\";\n// @ts-ignore\nimport { Hono } from \"npm:hono@4.4.12\";\n// @ts-ignore\nimport { cors } from \"npm:hono@4.4.12/cors\";\n// @ts-ignore\nimport Parser from \"npm:rss-parser@3.13.0\";\n\n// --- ACTION LIBRARY (PRIMITIVE CAPABILITIES) ---\nconst ACTIONS = {\n FETCH_NEWS: async (params) => {\n const parser = new Parser();\n const RSS_URL = `https://news.google.com/rss/search?q=${encodeURIComponent(params.query)} when:30d&hl=en-US&gl=US&ceid=US:en`;\n try {\n const feed = await parser.parseURL(RSS_URL);\n return feed.items.slice(0, 7).map(item => ({ title: item.title, snippet: item.contentSnippet }));\n } catch {\n return [{ title: \"Could not fetch news via RSS.\" }];\n }\n },\n FETCH_REDDIT_SENTIMENT: async (params) => {\n const allPosts = [];\n const subreddits = Array.isArray(params.subreddits) ? params.subreddits : [params.subreddits || \"-\"];\n for (const subreddit of subreddits) {\n try {\n const url = `https://www.reddit.com/r/${subreddit}/search.json?q=${encodeURIComponent(params.query)}&sort=relevance&t=month&restrict_sr=on`;\n const r = await fetch(url, { headers: { \"User-Agent\": \"Mozilla/5.0\" } });\n const d = await r.json();\n allPosts.push(...(d.data?.children?.map(p => ({ title: p.data.title, subreddit: p.data.subreddit, text: p.data.selftext?.substring(0, 200) })) || []));\n } catch {}\n }\n return allPosts.slice(0, 15);\n },\n INVOKE_AI: async (params) => {\n if (!params.user_prompt) throw new Error(\"INVOKE_AI action requires a 'user' prompt.\");\n const openai = new OpenAI();\n const completion = await openai.chat.completions.create({\n model: \"gpt-4o\",\n messages: [\n { role: \"system\", content: params.system_prompt || \"You are a helpful AI assistant.\" },\n { role: \"user\", content: params.user_prompt },\n ],\n temperature: parseFloat(params.temperature) || 0.4,\n });\n return completion.choices[0].message.content;\n },\n};\n\n// --- CORE ENGINE ---\nfunction parseDeclarativePlan(text) {\n const lines = text.split('\\n').filter(line => !line.trim().startsWith('#'));\n const plan = { VARS: {}, PROMPTS: {}, WORKFLOW: [] };\n let currentSection = null;\n let currentPromptKey = null;\n let currentPromptPart = null;\n let currentWorkflowStep = null;\n\n for (const line of lines) {\n const indent = line.match(/^\\s*/)[0].length;\n const trimmed = line.trim();\n if (!trimmed) continue;\n\n if (indent === 0) {\n currentSection = trimmed.replace(':', '');\n currentPromptKey = null;\n currentWorkflowStep = null;\n continue;\n }\n\n switch (currentSection) {\n case 'VARS':\n const [varKey, ...varValueParts] = trimmed.split(':');\n plan.VARS[varKey.trim()] = varValueParts.join(':').trim();\n break;\n \n case 'PROMPTS':\n if (indent === 2) {\n currentPromptKey = trimmed.replace(':', '');\n plan.PROMPTS[currentPromptKey] = {};\n currentPromptPart = null;\n } else if (indent === 4 && currentPromptKey) {\n const partKey = trimmed.split(':')[0].trim();\n if (partKey === 'system' || partKey === 'user') {\n currentPromptPart = partKey;\n plan.PROMPTS[currentPromptKey][currentPromptPart] = trimmed.substring(partKey.length + 1).replace(/>$/, '').trim();\n }\n } else if (indent > 4 && currentPromptKey && currentPromptPart) {\n plan.PROMPTS[currentPromptKey][currentPromptPart] += '\\n' + trimmed;\n }\n break;\n \n case 'WORKFLOW':\n if (trimmed.startsWith('- id:')) {\n if (currentWorkflowStep) plan.WORKFLOW.push(currentWorkflowStep);\n currentWorkflowStep = { params: {} };\n const [wfKey, ...wfValueParts] = trimmed.substring(2).split(':');\n currentWorkflowStep[wfKey.trim()] = wfValueParts.join(':').trim();\n } else if (currentWorkflowStep) {\n const [paramKey, ...paramValueParts] = trimmed.split(':');\n const value = paramValueParts.join(':').trim();\n if (paramKey.trim() === 'action' || paramKey.trim() === 'prompt_key' || paramKey.trim() === 'temperature') {\n currentWorkflowStep[paramKey.trim()] = value;\n } else {\n currentWorkflowStep.params[paramKey.trim()] = value;\n }\n }\n break;\n }\n }\n if (currentWorkflowStep) plan.WORKFLOW.push(currentWorkflowStep);\n return plan;\n}\n\nfunction resolveVariables(template, state) {\n if (typeof template !== 'string') return template;\n return template.replace(/\\{\\{(.*?)\\}\\}/g, (match, key) => {\n const keys = key.trim().split('.');\n let value = state;\n for (const k of keys) {\n value = value?.[k];\n }\n if (typeof value === 'object') {\n return JSON.stringify(value, null, 2);\n }\n return value !== undefined ? value : match;\n });\n}\n\nasync function executePlan(plan) {\n const state = { VARS: plan.VARS, PROMPTS: plan.PROMPTS };\n let finalOutput = \"Workflow completed, but the final step produced no output.\";\n\n for (const step of plan.WORKFLOW) {\n const action = ACTIONS[step.action];\n if (!action) throw new Error(`Unknown action: ${step.action}`);\n \n let resolvedParams = {};\n\n if (step.action === 'INVOKE_AI') {\n const promptTemplate = state.PROMPTS[step.prompt_key];\n if (!promptTemplate) throw new Error(`Prompt key not found in PROMPTS section: '${step.prompt_key}'`);\n \n resolvedParams = {\n system_prompt: resolveVariables(promptTemplate.system, state),\n user_prompt: resolveVariables(promptTemplate.user, state),\n temperature: step.temperature || \"0.4\",\n };\n } else {\n for (const key in step.params) {\n resolvedParams[key] = resolveVariables(step.params[key], state);\n }\n }\n \n const result = await action(resolvedParams);\n state[step.id] = { result };\n finalOutput = result;\n }\n\n return finalOutput;\n}\n\n// --- HTML & FRONTEND ---\nfunction generateHtml() {\n // MODIFIED: A much more complex and comprehensive example workflow.\n const examplePlan = `# Plan: 'Wickerbrook Vale' Go/No-Go Analysis\n\nVARS:\n game_idea: \"A cozy life-simulation game called 'Wickerbrook Vale', similar to Stardew Valley but with a witch-themed narrative and potion-crafting system.\"\n genre: \"Cozy life-simulation game\"\n platform: \"Nintendo Switch\"\n competitors: \"Stardew Valley, Animal Crossing, Wylde Flowers\"\n\nPROMPTS:\n market_synthesis_prompt:\n system: >\n You are a Senior Market Analyst for a major video game publisher. Your task is to synthesize the provided news articles and raw player sentiment from Reddit into a concise \"State of the Market\" report for the cozy game genre.\n Identify:\n 1. Key opportunities (e.g., features players are begging for).\n 2. Major threats (e.g., market saturation, player burnout).\n user: >\n RECENT NEWS ARTICLES:\n {{FETCH_MARKET_NEWS.result}}\n\n RAW PLAYER SENTIMENT FROM REDDIT:\n {{FETCH_PLAYER_SENTIMENT.result}}\n\n red_team_critique_prompt:\n system: >\n You are a skeptical, data-driven Executive Producer (Red Team). Your job is to find reasons this game will fail. Using the provided market analysis, write a sharp critique of the game idea. Connect every weakness you identify directly back to a threat or a lack of opportunity mentioned in the market data.\n user: >\n GAME IDEA:\n {{VARS.game_idea}}\n\n STATE OF THE MARKET REPORT:\n {{SYNTHESIZE_MARKET.result}}\n\n blue_team_pitch_prompt:\n system: >\n You are an optimistic, player-focused Marketing Lead (Blue Team). Your job is to create a compelling pitch for this game. Using the provided market analysis, write an enthusiastic pitch that shows how the game's features perfectly align with the opportunities and address the player complaints found in the market data.\n user: >\n GAME IDEA:\n {{VARS.game_idea}}\n\n STATE OF THE MARKET REPORT:\n {{SYNTHESIZE_MARKET.result}}\n\n final_memo_prompt:\n system: >\n You are the Head of Studio. You have just read the Red Team critique and the Blue Team pitch. Your task is to write the final Go/No-Go decision memo.\n 1. Briefly summarize the core conflict between the Red and Blue team perspectives.\n 2. Make a final \"Go\" or \"No-Go\" decision.\n 3. If \"Go\", list the 2-3 critical success factors that must be prioritized during development to mitigate the risks raised by the Red Team.\n user: >\n RED TEAM CRITIQUE:\n {{RED_TEAM_ANALYSIS.result}}\n\n BLUE TEAM PITCH:\n {{BLUE_TEAM_PITCH.result}}\n\nWORKFLOW:\n - id: FETCH_MARKET_NEWS\n action: FETCH_NEWS\n params:\n query: \"{{VARS.genre}} on {{VARS.platform}}\"\n\n - id: FETCH_PLAYER_SENTIMENT\n action: FETCH_REDDIT_SENTIMENT\n params:\n query: \"{{VARS.competitors}}\"\n subreddits: \"CozyGamers, NintendoSwitch\"\n\n - id: SYNTHESIZE_MARKET\n action: INVOKE_AI\n prompt_key: market_synthesis_prompt\n temperature: 0.2\n\n - id: RED_TEAM_ANALYSIS\n action: INVOKE_AI\n prompt_key: red_team_critique_prompt\n temperature: 0.6\n\n - id: BLUE_TEAM_PITCH\n action: INVOKE_AI\n prompt_key: blue_team_pitch_prompt\n temperature: 0.6\n\n - id: FINAL_RECOMMENDATION\n action: INVOKE_AI\n prompt_key: final_memo_prompt\n temperature: 0.3\n`;\n\n return `\n\n\n \n \n Declarative AI Workflow Engine\n \n \n \n \n\n\n\n

\n \n

Diff · v7