@rpascene/core 0.30.8

package/dist/es/ai-model/prompt/llm-planning.mjs

@@ -0,0 +1,367 @@
+import { ifRpasceneLocatorField } from "../common.mjs";
+import { bboxDescription } from "./common.mjs";
+const vlCurrentLog = '"log": string, // Log your thoughts and what the next one action (ONLY ONE!) you can do according to the screenshot and the instruction. The log should contain the following information: "The user wants to do ... . According to the instruction and the previous logs, next step is to .... Now i am going to compose an action \'{ action-type }\' to do ....". If no action should be done, log the reason. Use the same language as the user\'s instruction.';
+const llmCurrentLog = '"log": string, // Log what the next actions you can do according to the screenshot and the instruction. The typical log looks like "Now i want to use action \'{ action-type }\' to do ..". If no action should be done, log the reason. ". Use the same language as the user\'s instruction.';
+const commonOutputFields = `"error"?: string, // Error messages about unexpected situations, if any. Only think it is an error when the situation is not foreseeable according to the instruction. Use the same language as the user's instruction.
+  "more_actions_needed_by_instruction": boolean, // Consider if there is still more action(s) to do after the action in "Log" is done, according to the instruction. If so, set this field to true. Otherwise, set it to false.`;
+const vlLocateParam = ()=>'{bbox: [number, number, number, number], prompt: string }';
+const llmLocateParam = ()=>'{"id": string, "prompt": string}';
+const descriptionForAction = (action, locatorSchemaTypeDescription)=>{
+    const tab = '  ';
+    const fields = [];
+    fields.push(`- type: "${action.name}"`);
+    if (action.paramSchema) {
+        const shape = action.paramSchema.shape;
+        if (!shape) console.warn(`action.paramSchema is not a ZodObject, may lead to unexpected behavior, action name: ${action.name}`);
+        const paramLines = [];
+        const getTypeName = (field)=>{
+            var _actualField__def;
+            const unwrapField = (f)=>{
+                if (!f._def) return f;
+                const typeName = f._def.typeName;
+                if ('ZodOptional' === typeName || 'ZodNullable' === typeName || 'ZodDefault' === typeName) return unwrapField(f._def.innerType);
+                if ('ZodEffects' === typeName) {
+                    if (f._def.schema) return unwrapField(f._def.schema);
+                }
+                return f;
+            };
+            const actualField = unwrapField(field);
+            const fieldTypeName = null == (_actualField__def = actualField._def) ? void 0 : _actualField__def.typeName;
+            if ('ZodString' === fieldTypeName) return 'string';
+            if ('ZodNumber' === fieldTypeName) return 'number';
+            if ('ZodBoolean' === fieldTypeName) return 'boolean';
+            if ('ZodArray' === fieldTypeName) return 'array';
+            if ('ZodObject' === fieldTypeName) {
+                if (ifRpasceneLocatorField(actualField)) return locatorSchemaTypeDescription;
+                return 'object';
+            }
+            if ('ZodEnum' === fieldTypeName) {
+                var _actualField__def_values, _actualField__def1;
+                const values = (null == (_actualField__def1 = actualField._def) ? void 0 : null == (_actualField__def_values = _actualField__def1.values) ? void 0 : _actualField__def_values.map((option)=>String(`'${option}'`)).join(', ')) ?? 'enum';
+                return `enum(${values})`;
+            }
+            if ('ZodUnion' === fieldTypeName) {
+                var _actualField__def2;
+                const options = null == (_actualField__def2 = actualField._def) ? void 0 : _actualField__def2.options;
+                if (options && options.length > 0) {
+                    const types = options.map((opt)=>getTypeName(opt));
+                    return types.join(' | ');
+                }
+                return 'union';
+            }
+            console.warn('failed to parse Zod type. This may lead to wrong params from the LLM.\n', actualField._def);
+            return actualField.toString();
+        };
+        const getDescription = (field)=>{
+            var _actualField__def;
+            const unwrapField = (f)=>{
+                if (!f._def) return f;
+                const typeName = f._def.typeName;
+                if ('ZodOptional' === typeName || 'ZodNullable' === typeName || 'ZodDefault' === typeName) return unwrapField(f._def.innerType);
+                if ('ZodEffects' === typeName) {
+                    if (f._def.schema) return unwrapField(f._def.schema);
+                }
+                return f;
+            };
+            if ("description" in field) return field.description || null;
+            const actualField = unwrapField(field);
+            if ("description" in actualField) return actualField.description || null;
+            if ((null == (_actualField__def = actualField._def) ? void 0 : _actualField__def.typeName) === 'ZodObject') {
+                if ('rpascene_location_field_flag' in actualField._def.shape()) return 'Location information for the target element';
+            }
+            return null;
+        };
+        for (const [key, field] of Object.entries(shape))if (field && 'object' == typeof field) {
+            const isOptional = 'function' == typeof field.isOptional && field.isOptional();
+            const keyWithOptional = isOptional ? `${key}?` : key;
+            const typeName = getTypeName(field);
+            const description = getDescription(field);
+            let paramLine = `${keyWithOptional}: ${typeName}`;
+            if (description) paramLine += ` // ${description}`;
+            paramLines.push(paramLine);
+        }
+        if (paramLines.length > 0) {
+            fields.push('- param:');
+            for (const paramLine of paramLines)fields.push(`  - ${paramLine}`);
+        }
+    }
+    return `- ${action.name}, ${action.description || "No description provided"}
+${tab}${fields.join(`\n${tab}`)}
+`.trim();
+};
+const systemTemplateOfVLPlanning = ({ actionSpace, vlMode })=>{
+    const actionNameList = actionSpace.map((action)=>action.name).join(', ');
+    const actionDescriptionList = actionSpace.map((action)=>descriptionForAction(action, vlLocateParam()));
+    const actionList = actionDescriptionList.join('\n');
+    return `
+Target: User will give you a latest screenshot, an instruction and some previous logs indicating what have been done. Please tell what the next one action is (or null if no action should be done) to do the tasks the instruction requires. 
+
+Restriction:
+- Don't give extra actions or plans beyond the instruction. ONLY plan for what the instruction requires. For example, don't try to submit the form if the instruction is only to fill something.
+- Always give ONLY ONE action in \`log\` field (or null if no action should be done), instead of multiple actions. Supported actions are ${actionNameList}.
+- Don't repeat actions in the previous logs.
+- Bbox is the bounding box of the element to be located. It's an array of 4 numbers, representing ${bboxDescription(vlMode)}.
+
+Supporting actions:
+${actionList}
+
+Field description:
+* The \`prompt\` field inside the \`locate\` field is a short description that could be used to locate the element.
+
+Return in JSON format:
+{
+  ${vlCurrentLog}
+  ${commonOutputFields}
+  "action": 
+    {
+      // one of the supporting actions
+    } | null,
+  ,
+  "sleep"?: number, // The sleep time after the action, in milliseconds.
+}
+
+For example, when the instruction is "click 'Confirm' button, and click 'Yes' in popup" and the previous log shows "The 'Confirm' button has been clicked", by viewing the screenshot and previous logs, you should consider: We have already clicked the 'Confirm' button, so next we should find and click 'Yes' in popup.
+
+this and output the JSON:
+
+{
+  "log": "The user wants to do click 'Confirm' button, and click 'Yes' in popup. According to the instruction and the previous logs, next step is to tap the 'Yes' button in the popup. Now i am going to compose an action 'Tap' to click 'Yes' in popup.",
+  "action": {
+    "type": "Tap",
+    "param": {
+      "locate": {
+        "bbox": [100, 100, 200, 200],
+        "prompt": "The 'Yes' button in popup"
+      }
+    }
+  },
+  "more_actions_needed_by_instruction": false,
+}
+`;
+};
+const systemTemplateOfLLM = ({ actionSpace })=>{
+    const actionNameList = actionSpace.map((action)=>action.name).join(' / ');
+    const actionDescriptionList = actionSpace.map((action)=>descriptionForAction(action, llmLocateParam()));
+    const actionList = actionDescriptionList.join('\n');
+    return `
+## Role
+
+You are a versatile professional in software UI automation. Your outstanding contributions will impact the user experience of billions of users.
+
+## Objective
+
+- Decompose the instruction user asked into a series of actions
+- Locate the target element if possible
+- If the instruction cannot be accomplished, give a further plan.
+
+## Workflow
+
+1. Receive the screenshot, element description of screenshot(if any), user's instruction and previous logs.
+2. Decompose the user's task into a sequence of feasible actions, and place it in the \`actions\` field. There are different types of actions (${actionNameList}). The "About the action" section below will give you more details.
+3. Consider whether the user's instruction will be accomplished after the actions you composed.
+- If the instruction is accomplished, set \`more_actions_needed_by_instruction\` to false.
+- If more actions are needed, set \`more_actions_needed_by_instruction\` to true. Get ready to hand over to the next talent people like you. Carefully log what have been done in the \`log\` field, he or she will continue the task according to your logs.
+4. If the task is not feasible on this page, set \`error\` field to the reason.
+
+## Constraints
+
+- All the actions you composed MUST be feasible, which means all the action fields can be filled with the page context information you get. If not, don't plan this action.
+- Trust the "What have been done" field about the task (if any), don't repeat actions in it.
+- Respond only with valid JSON. Do not write an introduction or summary or markdown prefix like \`\`\`json\`\`\`.
+- If the screenshot and the instruction are totally irrelevant, set reason in the \`error\` field.
+
+## About the \`actions\` field
+
+The \`locate\` param is commonly used in the \`param\` field of the action, means to locate the target element to perform the action, it conforms to the following scheme:
+
+type LocateParam = {
+  "id": string, // the id of the element found. It should either be the id marked with a rectangle in the screenshot or the id described in the description.
+  "prompt"?: string // the description of the element to find. It can only be omitted when locate is null.
+} | null // If it's not on the page, the LocateParam should be null
+
+## Supported actions
+
+Each action has a \`type\` and corresponding \`param\`. To be detailed:
+${actionList}
+
+`.trim();
+};
+const outputTemplate = `
+## Output JSON Format:
+
+The JSON format is as follows:
+
+{
+  "actions": [
+    // ... some actions
+  ],
+  ${llmCurrentLog}
+  ${commonOutputFields}
+}
+
+## Examples
+
+### Example: Decompose a task
+
+When you received the following information:
+
+* Instruction: 'Click the language switch button, wait 1s, click "English"'
+* Logs: null
+* Page Context (screenshot and description) shows: There is a language switch button, and the "English" option is not shown in the screenshot now.
+
+By viewing the page screenshot and description, you should consider this and output the JSON:
+
+* The user intent is: tap the switch button, sleep, and tap the 'English' option
+* The language switch button is shown in the screenshot, and can be located by the page description or the id marked with a rectangle. So we can plan a Tap action to do this.
+* Plan a Sleep action to wait for 1 second to ensure the language options are displayed.
+* The "English" option button is not shown in the screenshot now, it means it may only show after the previous actions are finished. So don't plan any action to do this.
+* Compose the log: The user wants to do click the language switch button, wait 1s, click "English". According to the instruction and the previous logs, next step is to tap the language switch button to open the language options. Now i am going to compose an action 'Tap' to click the language switch button.
+* The task cannot be accomplished (because the last tapping action is not finished yet), so the \`more_actions_needed_by_instruction\` field is true. The \`error\` field is null.
+
+{
+  "actions":[
+    {
+      "thought": "Click the language switch button to open the language options.",
+      "type": "Tap", 
+      "param": {
+        "locate": { id: "c81c4e9a33", prompt: "The language switch button" }
+      }
+    },
+    {
+      "thought": "Wait for 1 second to ensure the language options are displayed.",
+      "type": "Sleep",
+      "param": { "timeMs": 1000 },
+    }
+  ],
+  "error": null,
+  "more_actions_needed_by_instruction": true,
+  "log": "The user wants to do click the language switch button, wait 1s, click \"English\". According to the instruction and the previous logs, next step is to tap the language switch button to open the language options. Now i am going to compose an action 'Tap' to click the language switch button.",
+}
+
+### Example: What NOT to do
+Wrong output:
+{
+  "actions":[
+    {
+      "thought": "Click the language switch button to open the language options.",
+      "type": "Tap",
+      "param": {
+        "locate": { "id": "c81c4e9a33" } // WRONG: prompt is missing, this is not a valid LocateParam
+      }
+    },
+    {
+      "thought": "Click the English option",
+      "type": "Tap", 
+      "param": {
+        "locate": null // WRONG: if the element is not on the page, you should not plan this action
+      }
+    }
+  ],
+  "more_actions_needed_by_instruction": false, // WRONG: should be true
+  "log": "The user wants to do click the language switch button, wait 1s, click \"English\". According to the instruction and the previous logs, next step is to tap the language switch button to open the language options. Now i am going to compose an action 'Tap' to click the language switch button.",
+}
+`;
+async function systemPromptToTaskPlanning({ actionSpace, vlMode }) {
+    if (vlMode) return systemTemplateOfVLPlanning({
+        actionSpace,
+        vlMode
+    });
+    return `${systemTemplateOfLLM({
+        actionSpace
+    })}\n\n${outputTemplate}`;
+}
+const planSchema = {
+    type: 'json_schema',
+    json_schema: {
+        name: 'action_items',
+        strict: false,
+        schema: {
+            type: 'object',
+            strict: false,
+            properties: {
+                actions: {
+                    type: 'array',
+                    items: {
+                        type: 'object',
+                        strict: false,
+                        properties: {
+                            thought: {
+                                type: 'string',
+                                description: 'Reasons for generating this task, and why this task is feasible on this page'
+                            },
+                            type: {
+                                type: 'string',
+                                description: 'Type of action'
+                            },
+                            param: {
+                                anyOf: [
+                                    {
+                                        type: 'null'
+                                    },
+                                    {
+                                        type: 'object',
+                                        additionalProperties: true
+                                    }
+                                ],
+                                description: 'Parameter of the action'
+                            },
+                            locate: {
+                                type: [
+                                    'object',
+                                    'null'
+                                ],
+                                properties: {
+                                    id: {
+                                        type: 'string'
+                                    },
+                                    prompt: {
+                                        type: 'string'
+                                    }
+                                },
+                                required: [
+                                    'id',
+                                    'prompt'
+                                ],
+                                additionalProperties: false,
+                                description: 'Location information for the target element'
+                            }
+                        },
+                        required: [
+                            'thought',
+                            'type',
+                            'param',
+                            'locate'
+                        ],
+                        additionalProperties: false
+                    },
+                    description: 'List of actions to be performed'
+                },
+                more_actions_needed_by_instruction: {
+                    type: 'boolean',
+                    description: 'If all the actions described in the instruction have been covered by this action and logs, set this field to false.'
+                },
+                log: {
+                    type: 'string',
+                    description: 'Log what these planned actions do. Do not include further actions that have not been planned.'
+                },
+                error: {
+                    type: [
+                        'string',
+                        'null'
+                    ],
+                    description: 'Error messages about unexpected situations'
+                }
+            },
+            required: [
+                'actions',
+                'more_actions_needed_by_instruction',
+                'log',
+                'error'
+            ],
+            additionalProperties: false
+        }
+    }
+};
+export { descriptionForAction, planSchema, systemPromptToTaskPlanning };
+
+//# sourceMappingURL=llm-planning.mjs.map

package/dist/es/ai-model/prompt/llm-planning.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"ai-model\\prompt\\llm-planning.mjs","sources":["webpack://@rpascene/core/./src/ai-model/prompt/llm-planning.ts"],"sourcesContent":["import type { DeviceAction } from '@/types';\nimport type { TVlModeTypes } from '@rpascene/shared/env';\nimport type { ResponseFormatJSONSchema } from 'openai/resources/index';\nimport type { ZodObject, z } from 'zod';\nimport { ifRpasceneLocatorField } from '../common';\nimport { bboxDescription } from './common';\n\n// Note: put the log field first to trigger the CoT\n\nconst vlCurrentLog = `\"log\": string, // Log your thoughts and what the next one action (ONLY ONE!) you can do according to the screenshot and the instruction. The log should contain the following information: \"The user wants to do ... . According to the instruction and the previous logs, next step is to .... Now i am going to compose an action '{ action-type }' to do ....\". If no action should be done, log the reason. Use the same language as the user's instruction.`;\nconst llmCurrentLog = `\"log\": string, // Log what the next actions you can do according to the screenshot and the instruction. The typical log looks like \"Now i want to use action '{ action-type }' to do ..\". If no action should be done, log the reason. \". Use the same language as the user's instruction.`;\n\nconst commonOutputFields = `\"error\"?: string, // Error messages about unexpected situations, if any. Only think it is an error when the situation is not foreseeable according to the instruction. Use the same language as the user's instruction.\n  \"more_actions_needed_by_instruction\": boolean, // Consider if there is still more action(s) to do after the action in \"Log\" is done, according to the instruction. If so, set this field to true. Otherwise, set it to false.`;\nconst vlLocateParam = () =>\n  '{bbox: [number, number, number, number], prompt: string }';\nconst llmLocateParam = () => '{\"id\": string, \"prompt\": string}';\n\nexport const descriptionForAction = (\n  action: DeviceAction<any>,\n  locatorSchemaTypeDescription: string,\n) => {\n  const tab = '  ';\n  const fields: string[] = [];\n\n  // Add the action type field\n  fields.push(`- type: \"${action.name}\"`);\n\n  // Handle paramSchema if it exists\n  if (action.paramSchema) {\n    // Try to extract parameter information from the zod schema\n    // For zod object schemas, extract type information and descriptions\n    const shape = (action.paramSchema as ZodObject<any>).shape;\n\n    if (!shape) {\n      console.warn(\n        `action.paramSchema is not a ZodObject, may lead to unexpected behavior, action name: ${action.name}`,\n      );\n    }\n\n    const paramLines: string[] = [];\n\n    // Helper function to get type name from zod schema\n    const getTypeName = (field: any): string => {\n      // Recursively unwrap optional, nullable, and other wrapper types to get the actual inner type\n      const unwrapField = (f: any): any => {\n        if (!f._def) return f;\n\n        const typeName = f._def.typeName;\n\n        // Handle wrapper types that have innerType\n        if (\n          typeName === 'ZodOptional' ||\n          typeName === 'ZodNullable' ||\n          typeName === 'ZodDefault'\n        ) {\n          return unwrapField(f._def.innerType);\n        }\n\n        // Handle ZodEffects (transformations, refinements, preprocessors)\n        if (typeName === 'ZodEffects') {\n          // For ZodEffects, unwrap the schema field which contains the underlying type\n          if (f._def.schema) {\n            return unwrapField(f._def.schema);\n          }\n        }\n\n        return f;\n      };\n\n      const actualField = unwrapField(field);\n      const fieldTypeName = actualField._def?.typeName;\n\n      if (fieldTypeName === 'ZodString') return 'string';\n      if (fieldTypeName === 'ZodNumber') return 'number';\n      if (fieldTypeName === 'ZodBoolean') return 'boolean';\n      if (fieldTypeName === 'ZodArray') return 'array';\n      if (fieldTypeName === 'ZodObject') {\n        // Check if this is a passthrough object (like RpasceneLocation)\n        if (ifRpasceneLocatorField(actualField)) {\n          return locatorSchemaTypeDescription;\n        }\n        return 'object';\n      }\n      if (fieldTypeName === 'ZodEnum') {\n        const values =\n          (actualField._def?.values as unknown[] | undefined)\n            ?.map((option: unknown) => String(`'${option}'`))\n            .join(', ') ?? 'enum';\n\n        return `enum(${values})`;\n      }\n      // Handle ZodUnion by taking the first option (for display purposes)\n      if (fieldTypeName === 'ZodUnion') {\n        const options = actualField._def?.options as any[] | undefined;\n        if (options && options.length > 0) {\n          // For unions, list all types\n          const types = options.map((opt: any) => getTypeName(opt));\n          return types.join(' | ');\n        }\n        return 'union';\n      }\n\n      console.warn(\n        'failed to parse Zod type. This may lead to wrong params from the LLM.\\n',\n        actualField._def,\n      );\n      return actualField.toString();\n    };\n\n    // Helper function to get description from zod schema\n    const getDescription = (field: z.ZodTypeAny): string | null => {\n      // Recursively unwrap optional, nullable, and other wrapper types to get the actual inner type\n      const unwrapField = (f: any): any => {\n        if (!f._def) return f;\n\n        const typeName = f._def.typeName;\n\n        // Handle wrapper types that have innerType\n        if (\n          typeName === 'ZodOptional' ||\n          typeName === 'ZodNullable' ||\n          typeName === 'ZodDefault'\n        ) {\n          return unwrapField(f._def.innerType);\n        }\n\n        // Handle ZodEffects (transformations, refinements, preprocessors)\n        if (typeName === 'ZodEffects') {\n          // For ZodEffects, unwrap the schema field which contains the underlying type\n          if (f._def.schema) {\n            return unwrapField(f._def.schema);\n          }\n        }\n\n        return f;\n      };\n\n      // Check for direct description on the original field (wrapper may have description)\n      if ('description' in field) {\n        return field.description || null;\n      }\n\n      const actualField = unwrapField(field);\n\n      // Check for description on the unwrapped field\n      if ('description' in actualField) {\n        return actualField.description || null;\n      }\n\n      // Check for RpasceneLocation fields and add description\n      if (actualField._def?.typeName === 'ZodObject') {\n        if ('rpascene_location_field_flag' in actualField._def.shape()) {\n          return 'Location information for the target element';\n        }\n      }\n\n      return null;\n    };\n\n    for (const [key, field] of Object.entries(shape)) {\n      if (field && typeof field === 'object') {\n        // Check if field is optional\n        const isOptional =\n          typeof (field as any).isOptional === 'function' &&\n          (field as any).isOptional();\n        const keyWithOptional = isOptional ? `${key}?` : key;\n\n        // Get the type name\n        const typeName = getTypeName(field);\n\n        // Get description\n        const description = getDescription(field as z.ZodTypeAny);\n\n        // Build param line for this field\n        let paramLine = `${keyWithOptional}: ${typeName}`;\n        if (description) {\n          paramLine += ` // ${description}`;\n        }\n\n        paramLines.push(paramLine);\n      }\n    }\n\n    if (paramLines.length > 0) {\n      fields.push('- param:');\n      for (const paramLine of paramLines) {\n        fields.push(`  - ${paramLine}`);\n      }\n    }\n  }\n\n  return `- ${action.name}, ${action.description || 'No description provided'}\n${tab}${fields.join(`\\n${tab}`)}\n`.trim();\n};\n\nconst systemTemplateOfVLPlanning = ({\n  actionSpace,\n  vlMode,\n}: {\n  actionSpace: DeviceAction<any>[];\n  vlMode: TVlModeTypes | undefined;\n}) => {\n  const actionNameList = actionSpace.map((action) => action.name).join(', ');\n  const actionDescriptionList = actionSpace.map((action) => {\n    return descriptionForAction(action, vlLocateParam());\n  });\n  const actionList = actionDescriptionList.join('\\n');\n\n  return `\nTarget: User will give you a latest screenshot, an instruction and some previous logs indicating what have been done. Please tell what the next one action is (or null if no action should be done) to do the tasks the instruction requires. \n\nRestriction:\n- Don't give extra actions or plans beyond the instruction. ONLY plan for what the instruction requires. For example, don't try to submit the form if the instruction is only to fill something.\n- Always give ONLY ONE action in \\`log\\` field (or null if no action should be done), instead of multiple actions. Supported actions are ${actionNameList}.\n- Don't repeat actions in the previous logs.\n- Bbox is the bounding box of the element to be located. It's an array of 4 numbers, representing ${bboxDescription(vlMode)}.\n\nSupporting actions:\n${actionList}\n\nField description:\n* The \\`prompt\\` field inside the \\`locate\\` field is a short description that could be used to locate the element.\n\nReturn in JSON format:\n{\n  ${vlCurrentLog}\n  ${commonOutputFields}\n  \"action\": \n    {\n      // one of the supporting actions\n    } | null,\n  ,\n  \"sleep\"?: number, // The sleep time after the action, in milliseconds.\n}\n\nFor example, when the instruction is \"click 'Confirm' button, and click 'Yes' in popup\" and the previous log shows \"The 'Confirm' button has been clicked\", by viewing the screenshot and previous logs, you should consider: We have already clicked the 'Confirm' button, so next we should find and click 'Yes' in popup.\n\nthis and output the JSON:\n\n{\n  \"log\": \"The user wants to do click 'Confirm' button, and click 'Yes' in popup. According to the instruction and the previous logs, next step is to tap the 'Yes' button in the popup. Now i am going to compose an action 'Tap' to click 'Yes' in popup.\",\n  \"action\": {\n    \"type\": \"Tap\",\n    \"param\": {\n      \"locate\": {\n        \"bbox\": [100, 100, 200, 200],\n        \"prompt\": \"The 'Yes' button in popup\"\n      }\n    }\n  },\n  \"more_actions_needed_by_instruction\": false,\n}\n`;\n};\n\nconst systemTemplateOfLLM = ({\n  actionSpace,\n}: { actionSpace: DeviceAction<any>[] }) => {\n  const actionNameList = actionSpace.map((action) => action.name).join(' / ');\n  const actionDescriptionList = actionSpace.map((action) => {\n    return descriptionForAction(action, llmLocateParam());\n  });\n  const actionList = actionDescriptionList.join('\\n');\n\n  return `\n## Role\n\nYou are a versatile professional in software UI automation. Your outstanding contributions will impact the user experience of billions of users.\n\n## Objective\n\n- Decompose the instruction user asked into a series of actions\n- Locate the target element if possible\n- If the instruction cannot be accomplished, give a further plan.\n\n## Workflow\n\n1. Receive the screenshot, element description of screenshot(if any), user's instruction and previous logs.\n2. Decompose the user's task into a sequence of feasible actions, and place it in the \\`actions\\` field. There are different types of actions (${actionNameList}). The \"About the action\" section below will give you more details.\n3. Consider whether the user's instruction will be accomplished after the actions you composed.\n- If the instruction is accomplished, set \\`more_actions_needed_by_instruction\\` to false.\n- If more actions are needed, set \\`more_actions_needed_by_instruction\\` to true. Get ready to hand over to the next talent people like you. Carefully log what have been done in the \\`log\\` field, he or she will continue the task according to your logs.\n4. If the task is not feasible on this page, set \\`error\\` field to the reason.\n\n## Constraints\n\n- All the actions you composed MUST be feasible, which means all the action fields can be filled with the page context information you get. If not, don't plan this action.\n- Trust the \"What have been done\" field about the task (if any), don't repeat actions in it.\n- Respond only with valid JSON. Do not write an introduction or summary or markdown prefix like \\`\\`\\`json\\`\\`\\`.\n- If the screenshot and the instruction are totally irrelevant, set reason in the \\`error\\` field.\n\n## About the \\`actions\\` field\n\nThe \\`locate\\` param is commonly used in the \\`param\\` field of the action, means to locate the target element to perform the action, it conforms to the following scheme:\n\ntype LocateParam = {\n  \"id\": string, // the id of the element found. It should either be the id marked with a rectangle in the screenshot or the id described in the description.\n  \"prompt\"?: string // the description of the element to find. It can only be omitted when locate is null.\n} | null // If it's not on the page, the LocateParam should be null\n\n## Supported actions\n\nEach action has a \\`type\\` and corresponding \\`param\\`. To be detailed:\n${actionList}\n\n`.trim();\n};\n\nconst outputTemplate = `\n## Output JSON Format:\n\nThe JSON format is as follows:\n\n{\n  \"actions\": [\n    // ... some actions\n  ],\n  ${llmCurrentLog}\n  ${commonOutputFields}\n}\n\n## Examples\n\n### Example: Decompose a task\n\nWhen you received the following information:\n\n* Instruction: 'Click the language switch button, wait 1s, click \"English\"'\n* Logs: null\n* Page Context (screenshot and description) shows: There is a language switch button, and the \"English\" option is not shown in the screenshot now.\n\nBy viewing the page screenshot and description, you should consider this and output the JSON:\n\n* The user intent is: tap the switch button, sleep, and tap the 'English' option\n* The language switch button is shown in the screenshot, and can be located by the page description or the id marked with a rectangle. So we can plan a Tap action to do this.\n* Plan a Sleep action to wait for 1 second to ensure the language options are displayed.\n* The \"English\" option button is not shown in the screenshot now, it means it may only show after the previous actions are finished. So don't plan any action to do this.\n* Compose the log: The user wants to do click the language switch button, wait 1s, click \"English\". According to the instruction and the previous logs, next step is to tap the language switch button to open the language options. Now i am going to compose an action 'Tap' to click the language switch button.\n* The task cannot be accomplished (because the last tapping action is not finished yet), so the \\`more_actions_needed_by_instruction\\` field is true. The \\`error\\` field is null.\n\n{\n  \"actions\":[\n    {\n      \"thought\": \"Click the language switch button to open the language options.\",\n      \"type\": \"Tap\", \n      \"param\": {\n        \"locate\": { id: \"c81c4e9a33\", prompt: \"The language switch button\" }\n      }\n    },\n    {\n      \"thought\": \"Wait for 1 second to ensure the language options are displayed.\",\n      \"type\": \"Sleep\",\n      \"param\": { \"timeMs\": 1000 },\n    }\n  ],\n  \"error\": null,\n  \"more_actions_needed_by_instruction\": true,\n  \"log\": \"The user wants to do click the language switch button, wait 1s, click \\\"English\\\". According to the instruction and the previous logs, next step is to tap the language switch button to open the language options. Now i am going to compose an action 'Tap' to click the language switch button.\",\n}\n\n### Example: What NOT to do\nWrong output:\n{\n  \"actions\":[\n    {\n      \"thought\": \"Click the language switch button to open the language options.\",\n      \"type\": \"Tap\",\n      \"param\": {\n        \"locate\": { \"id\": \"c81c4e9a33\" } // WRONG: prompt is missing, this is not a valid LocateParam\n      }\n    },\n    {\n      \"thought\": \"Click the English option\",\n      \"type\": \"Tap\", \n      \"param\": {\n        \"locate\": null // WRONG: if the element is not on the page, you should not plan this action\n      }\n    }\n  ],\n  \"more_actions_needed_by_instruction\": false, // WRONG: should be true\n  \"log\": \"The user wants to do click the language switch button, wait 1s, click \\\"English\\\". According to the instruction and the previous logs, next step is to tap the language switch button to open the language options. Now i am going to compose an action 'Tap' to click the language switch button.\",\n}\n`;\n\nexport async function systemPromptToTaskPlanning({\n  actionSpace,\n  vlMode,\n}: {\n  actionSpace: DeviceAction<any>[];\n  vlMode: TVlModeTypes | undefined;\n}) {\n  if (vlMode) {\n    return systemTemplateOfVLPlanning({ actionSpace, vlMode });\n  }\n\n  return `${systemTemplateOfLLM({ actionSpace })}\\n\\n${outputTemplate}`;\n}\n\nexport const planSchema: ResponseFormatJSONSchema = {\n  type: 'json_schema',\n  json_schema: {\n    name: 'action_items',\n    strict: false,\n    schema: {\n      type: 'object',\n      strict: false,\n      properties: {\n        actions: {\n          type: 'array',\n          items: {\n            type: 'object',\n            strict: false,\n            properties: {\n              thought: {\n                type: 'string',\n                description:\n                  'Reasons for generating this task, and why this task is feasible on this page',\n              },\n              type: {\n                type: 'string',\n                description: 'Type of action',\n              },\n              param: {\n                anyOf: [\n                  { type: 'null' },\n                  {\n                    type: 'object',\n                    additionalProperties: true,\n                  },\n                ],\n                description: 'Parameter of the action',\n              },\n              locate: {\n                type: ['object', 'null'],\n                properties: {\n                  id: { type: 'string' },\n                  prompt: { type: 'string' },\n                },\n                required: ['id', 'prompt'],\n                additionalProperties: false,\n                description: 'Location information for the target element',\n              },\n            },\n            required: ['thought', 'type', 'param', 'locate'],\n            additionalProperties: false,\n          },\n          description: 'List of actions to be performed',\n        },\n        more_actions_needed_by_instruction: {\n          type: 'boolean',\n          description:\n            'If all the actions described in the instruction have been covered by this action and logs, set this field to false.',\n        },\n        log: {\n          type: 'string',\n          description:\n            'Log what these planned actions do. Do not include further actions that have not been planned.',\n        },\n        error: {\n          type: ['string', 'null'],\n          description: 'Error messages about unexpected situations',\n        },\n      },\n      required: [\n        'actions',\n        'more_actions_needed_by_instruction',\n        'log',\n        'error',\n      ],\n      additionalProperties: false,\n    },\n  },\n};\n"],"names":["vlCurrentLog","llmCurrentLog","commonOutputFields","vlLocateParam","llmLocateParam","descriptionForAction","action","locatorSchemaTypeDescription","tab","fields","shape","console","paramLines","getTypeName","field","_actualField__def","unwrapField","f","typeName","actualField","fieldTypeName","ifRpasceneLocatorField","_actualField__def_values","values","option","String","_actualField__def2","options","types","opt","getDescription","key","Object","isOptional","keyWithOptional","description","paramLine","systemTemplateOfVLPlanning","actionSpace","vlMode","actionNameList","actionDescriptionList","actionList","bboxDescription","systemTemplateOfLLM","outputTemplate","systemPromptToTaskPlanning","planSchema"],"mappings":";;AASA,MAAMA,eAAe;AACrB,MAAMC,gBAAgB;AAEtB,MAAMC,qBAAqB,CAAC;+NACmM,CAAC;AAChO,MAAMC,gBAAgB,IACpB;AACF,MAAMC,iBAAiB,IAAM;AAEtB,MAAMC,uBAAuB,CAClCC,QACAC;IAEA,MAAMC,MAAM;IACZ,MAAMC,SAAmB,EAAE;IAG3BA,OAAO,IAAI,CAAC,CAAC,SAAS,EAAEH,OAAO,IAAI,CAAC,CAAC,CAAC;IAGtC,IAAIA,OAAO,WAAW,EAAE;QAGtB,MAAMI,QAASJ,OAAO,WAAW,CAAoB,KAAK;QAE1D,IAAI,CAACI,OACHC,QAAQ,IAAI,CACV,CAAC,qFAAqF,EAAEL,OAAO,IAAI,EAAE;QAIzG,MAAMM,aAAuB,EAAE;QAG/B,MAAMC,cAAc,CAACC;gBA4BGC;YA1BtB,MAAMC,cAAc,CAACC;gBACnB,IAAI,CAACA,EAAE,IAAI,EAAE,OAAOA;gBAEpB,MAAMC,WAAWD,EAAE,IAAI,CAAC,QAAQ;gBAGhC,IACEC,AAAa,kBAAbA,YACAA,AAAa,kBAAbA,YACAA,AAAa,iBAAbA,UAEA,OAAOF,YAAYC,EAAE,IAAI,CAAC,SAAS;gBAIrC,IAAIC,AAAa,iBAAbA,UAEF;oBAAA,IAAID,EAAE,IAAI,CAAC,MAAM,EACf,OAAOD,YAAYC,EAAE,IAAI,CAAC,MAAM;gBAClC;gBAGF,OAAOA;YACT;YAEA,MAAME,cAAcH,YAAYF;YAChC,MAAMM,gBAAgB,QAAAL,CAAAA,oBAAAA,YAAY,IAAI,AAAD,IAAfA,KAAAA,IAAAA,kBAAkB,QAAQ;YAEhD,IAAIK,AAAkB,gBAAlBA,eAA+B,OAAO;YAC1C,IAAIA,AAAkB,gBAAlBA,eAA+B,OAAO;YAC1C,IAAIA,AAAkB,iBAAlBA,eAAgC,OAAO;YAC3C,IAAIA,AAAkB,eAAlBA,eAA8B,OAAO;YACzC,IAAIA,AAAkB,gBAAlBA,eAA+B;gBAEjC,IAAIC,uBAAuBF,cACzB,OAAOZ;gBAET,OAAO;YACT;YACA,IAAIa,AAAkB,cAAlBA,eAA6B;oBAE5BE,0BAAAA;gBADH,MAAMC,SACJ,AAAC,SAAAD,CAAAA,qBAAAA,YAAY,IAAI,AAAD,IAAfA,KAAAA,IAAAA,QAAAA,CAAAA,2BAAAA,mBAAkB,MAAM,AAAD,IAAvBA,KAAAA,IAAAA,yBACG,GAAG,CAAC,CAACE,SAAoBC,OAAO,CAAC,CAAC,EAAED,OAAO,CAAC,CAAC,GAC9C,IAAI,CAAC,KAAI,KAAK;gBAEnB,OAAO,CAAC,KAAK,EAAED,OAAO,CAAC,CAAC;YAC1B;YAEA,IAAIH,AAAkB,eAAlBA,eAA8B;oBAChBM;gBAAhB,MAAMC,UAAU,QAAAD,CAAAA,qBAAAA,YAAY,IAAI,AAAD,IAAfA,KAAAA,IAAAA,mBAAkB,OAAO;gBACzC,IAAIC,WAAWA,QAAQ,MAAM,GAAG,GAAG;oBAEjC,MAAMC,QAAQD,QAAQ,GAAG,CAAC,CAACE,MAAahB,YAAYgB;oBACpD,OAAOD,MAAM,IAAI,CAAC;gBACpB;gBACA,OAAO;YACT;YAEAjB,QAAQ,IAAI,CACV,2EACAQ,YAAY,IAAI;YAElB,OAAOA,YAAY,QAAQ;QAC7B;QAGA,MAAMW,iBAAiB,CAAChB;gBAwClBC;YAtCJ,MAAMC,cAAc,CAACC;gBACnB,IAAI,CAACA,EAAE,IAAI,EAAE,OAAOA;gBAEpB,MAAMC,WAAWD,EAAE,IAAI,CAAC,QAAQ;gBAGhC,IACEC,AAAa,kBAAbA,YACAA,AAAa,kBAAbA,YACAA,AAAa,iBAAbA,UAEA,OAAOF,YAAYC,EAAE,IAAI,CAAC,SAAS;gBAIrC,IAAIC,AAAa,iBAAbA,UAEF;oBAAA,IAAID,EAAE,IAAI,CAAC,MAAM,EACf,OAAOD,YAAYC,EAAE,IAAI,CAAC,MAAM;gBAClC;gBAGF,OAAOA;YACT;YAGA,IAAI,iBAAiBH,OACnB,OAAOA,MAAM,WAAW,IAAI;YAG9B,MAAMK,cAAcH,YAAYF;YAGhC,IAAI,iBAAiBK,aACnB,OAAOA,YAAY,WAAW,IAAI;YAIpC,IAAIJ,AAAAA,SAAAA,CAAAA,oBAAAA,YAAY,IAAI,AAAD,IAAfA,KAAAA,IAAAA,kBAAkB,QAAQ,AAAD,MAAM,aACjC;gBAAA,IAAI,kCAAkCI,YAAY,IAAI,CAAC,KAAK,IAC1D,OAAO;YACT;YAGF,OAAO;QACT;QAEA,KAAK,MAAM,CAACY,KAAKjB,MAAM,IAAIkB,OAAO,OAAO,CAACtB,OACxC,IAAII,SAAS,AAAiB,YAAjB,OAAOA,OAAoB;YAEtC,MAAMmB,aACJ,AAAqC,cAArC,OAAQnB,MAAc,UAAU,IAC/BA,MAAc,UAAU;YAC3B,MAAMoB,kBAAkBD,aAAa,GAAGF,IAAI,CAAC,CAAC,GAAGA;YAGjD,MAAMb,WAAWL,YAAYC;YAG7B,MAAMqB,cAAcL,eAAehB;YAGnC,IAAIsB,YAAY,GAAGF,gBAAgB,EAAE,EAAEhB,UAAU;YACjD,IAAIiB,aACFC,aAAa,CAAC,IAAI,EAAED,aAAa;YAGnCvB,WAAW,IAAI,CAACwB;QAClB;QAGF,IAAIxB,WAAW,MAAM,GAAG,GAAG;YACzBH,OAAO,IAAI,CAAC;YACZ,KAAK,MAAM2B,aAAaxB,WACtBH,OAAO,IAAI,CAAC,CAAC,IAAI,EAAE2B,WAAW;QAElC;IACF;IAEA,OAAO,CAAC,EAAE,EAAE9B,OAAO,IAAI,CAAC,EAAE,EAAEA,OAAO,WAAW,IAAI,0BAA0B;AAC9E,EAAEE,MAAMC,OAAO,IAAI,CAAC,CAAC,EAAE,EAAED,KAAK,EAAE;AAChC,CAAC,CAAC,IAAI;AACN;AAEA,MAAM6B,6BAA6B,CAAC,EAClCC,WAAW,EACXC,MAAM,EAIP;IACC,MAAMC,iBAAiBF,YAAY,GAAG,CAAC,CAAChC,SAAWA,OAAO,IAAI,EAAE,IAAI,CAAC;IACrE,MAAMmC,wBAAwBH,YAAY,GAAG,CAAC,CAAChC,SACtCD,qBAAqBC,QAAQH;IAEtC,MAAMuC,aAAaD,sBAAsB,IAAI,CAAC;IAE9C,OAAO,CAAC;;;;;yIAK+H,EAAED,eAAe;;kGAExD,EAAEG,gBAAgBJ,QAAQ;;;AAG5H,EAAEG,WAAW;;;;;;;EAOX,EAAE1C,aAAa;EACf,EAAEE,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;;;AA0BvB,CAAC;AACD;AAEA,MAAM0C,sBAAsB,CAAC,EAC3BN,WAAW,EAC0B;IACrC,MAAME,iBAAiBF,YAAY,GAAG,CAAC,CAAChC,SAAWA,OAAO,IAAI,EAAE,IAAI,CAAC;IACrE,MAAMmC,wBAAwBH,YAAY,GAAG,CAAC,CAAChC,SACtCD,qBAAqBC,QAAQF;IAEtC,MAAMsC,aAAaD,sBAAsB,IAAI,CAAC;IAE9C,OAAO,CAAC;;;;;;;;;;;;;;+IAcqI,EAAED,eAAe;;;;;;;;;;;;;;;;;;;;;;;;;AAyBhK,EAAEE,WAAW;;AAEb,CAAC,CAAC,IAAI;AACN;AAEA,MAAMG,iBAAiB,CAAC;;;;;;;;;EAStB,EAAE5C,cAAc;EAChB,EAAEC,mBAAmB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAgEvB,CAAC;AAEM,eAAe4C,2BAA2B,EAC/CR,WAAW,EACXC,MAAM,EAIP;IACC,IAAIA,QACF,OAAOF,2BAA2B;QAAEC;QAAaC;IAAO;IAG1D,OAAO,GAAGK,oBAAoB;QAAEN;IAAY,GAAG,IAAI,EAAEO,gBAAgB;AACvE;AAEO,MAAME,aAAuC;IAClD,MAAM;IACN,aAAa;QACX,MAAM;QACN,QAAQ;QACR,QAAQ;YACN,MAAM;YACN,QAAQ;YACR,YAAY;gBACV,SAAS;oBACP,MAAM;oBACN,OAAO;wBACL,MAAM;wBACN,QAAQ;wBACR,YAAY;4BACV,SAAS;gCACP,MAAM;gCACN,aACE;4BACJ;4BACA,MAAM;gCACJ,MAAM;gCACN,aAAa;4BACf;4BACA,OAAO;gCACL,OAAO;oCACL;wCAAE,MAAM;oCAAO;oCACf;wCACE,MAAM;wCACN,sBAAsB;oCACxB;iCACD;gCACD,aAAa;4BACf;4BACA,QAAQ;gCACN,MAAM;oCAAC;oCAAU;iCAAO;gCACxB,YAAY;oCACV,IAAI;wCAAE,MAAM;oCAAS;oCACrB,QAAQ;wCAAE,MAAM;oCAAS;gCAC3B;gCACA,UAAU;oCAAC;oCAAM;iCAAS;gCAC1B,sBAAsB;gCACtB,aAAa;4BACf;wBACF;wBACA,UAAU;4BAAC;4BAAW;4BAAQ;4BAAS;yBAAS;wBAChD,sBAAsB;oBACxB;oBACA,aAAa;gBACf;gBACA,oCAAoC;oBAClC,MAAM;oBACN,aACE;gBACJ;gBACA,KAAK;oBACH,MAAM;oBACN,aACE;gBACJ;gBACA,OAAO;oBACL,MAAM;wBAAC;wBAAU;qBAAO;oBACxB,aAAa;gBACf;YACF;YACA,UAAU;gBACR;gBACA;gBACA;gBACA;aACD;YACD,sBAAsB;QACxB;IACF;AACF"}

package/dist/es/ai-model/prompt/llm-section-locator.mjs

@@ -0,0 +1,47 @@
+import { PromptTemplate } from "@langchain/core/prompts";
+import { bboxDescription } from "./common.mjs";
+function systemPromptToLocateSection(vlMode) {
+    return `
+You goal is to find out one section containing the target element in the screenshot, put it in the \`bbox\` field. If the user describe the target element with some reference elements, you should also find the section containing the reference elements, put it in the \`references_bbox\` field.
+
+Usually, it should be approximately an area not more than 300x300px. Changes of the size are allowed if there are many elements to cover.
+
+return in this JSON format:
+\`\`\`json
+{
+  "bbox": [number, number, number, number],
+  "references_bbox"?: [
+    [number, number, number, number],
+    [number, number, number, number],
+    ...
+  ],
+  "error"?: string
+}
+\`\`\`
+
+In which, all the numbers in the \`bbox\` and \`references_bbox\` represent ${bboxDescription(vlMode)}.
+
+For example, if the user describe the target element as "the delete button on the second row with title 'Peter'", you should put the bounding box of the delete button in the \`bbox\` field, and the bounding box of the second row in the \`references_bbox\` field.
+
+the return value should be like this:
+\`\`\`json
+{
+  "bbox": [100, 100, 200, 200],
+  "references_bbox": [[100, 100, 200, 200]]
+}
+\`\`\`
+`;
+}
+const sectionLocatorInstruction = new PromptTemplate({
+    template: `Here is the target element user interested in:
+<targetDescription>
+{sectionDescription}
+</targetDescription>
+  `,
+    inputVariables: [
+        "sectionDescription"
+    ]
+});
+export { sectionLocatorInstruction, systemPromptToLocateSection };
+
+//# sourceMappingURL=llm-section-locator.mjs.map

package/dist/es/ai-model/prompt/llm-section-locator.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"ai-model\\prompt\\llm-section-locator.mjs","sources":["webpack://@rpascene/core/./src/ai-model/prompt/llm-section-locator.ts"],"sourcesContent":["import { PromptTemplate } from '@langchain/core/prompts';\nimport type { TVlModeTypes } from '@rpascene/shared/env';\nimport { bboxDescription } from './common';\n\nexport function systemPromptToLocateSection(vlMode: TVlModeTypes | undefined) {\n  return `\nYou goal is to find out one section containing the target element in the screenshot, put it in the \\`bbox\\` field. If the user describe the target element with some reference elements, you should also find the section containing the reference elements, put it in the \\`references_bbox\\` field.\n\nUsually, it should be approximately an area not more than 300x300px. Changes of the size are allowed if there are many elements to cover.\n\nreturn in this JSON format:\n\\`\\`\\`json\n{\n  \"bbox\": [number, number, number, number],\n  \"references_bbox\"?: [\n    [number, number, number, number],\n    [number, number, number, number],\n    ...\n  ],\n  \"error\"?: string\n}\n\\`\\`\\`\n\nIn which, all the numbers in the \\`bbox\\` and \\`references_bbox\\` represent ${bboxDescription(vlMode)}.\n\nFor example, if the user describe the target element as \"the delete button on the second row with title 'Peter'\", you should put the bounding box of the delete button in the \\`bbox\\` field, and the bounding box of the second row in the \\`references_bbox\\` field.\n\nthe return value should be like this:\n\\`\\`\\`json\n{\n  \"bbox\": [100, 100, 200, 200],\n  \"references_bbox\": [[100, 100, 200, 200]]\n}\n\\`\\`\\`\n`;\n}\n\nexport const sectionLocatorInstruction = new PromptTemplate({\n  template: `Here is the target element user interested in:\n<targetDescription>\n{sectionDescription}\n</targetDescription>\n  `,\n  inputVariables: ['sectionDescription'],\n});\n"],"names":["systemPromptToLocateSection","vlMode","bboxDescription","sectionLocatorInstruction","PromptTemplate"],"mappings":";;AAIO,SAASA,4BAA4BC,MAAgC;IAC1E,OAAO,CAAC;;;;;;;;;;;;;;;;;;4EAkBkE,EAAEC,gBAAgBD,QAAQ;;;;;;;;;;;AAWtG,CAAC;AACD;AAEO,MAAME,4BAA4B,IAAIC,eAAe;IAC1D,UAAU,CAAC;;;;EAIX,CAAC;IACD,gBAAgB;QAAC;KAAqB;AACxC"}

package/dist/es/ai-model/prompt/playwright-generator.mjs

@@ -0,0 +1,117 @@
+import { PLAYWRIGHT_EXAMPLE_CODE } from "@rpascene/shared/constants";
+import { AIActionType, callAI, callAIWithStringResponse } from "../index.mjs";
+import { createEventCounts, createMessageContent, extractInputDescriptions, filterEventsByType, getScreenshotsForLLM, prepareEventSummary, processEventsForLLM, validateEvents } from "./yaml-generator.mjs";
+const generatePlaywrightTest = async (events, options, modelConfig)=>{
+    validateEvents(events);
+    const summary = prepareEventSummary(events, {
+        testName: options.testName,
+        maxScreenshots: options.maxScreenshots || 3
+    });
+    const playwrightSummary = {
+        ...summary,
+        waitForNetworkIdle: false !== options.waitForNetworkIdle,
+        waitForNetworkIdleTimeout: options.waitForNetworkIdleTimeout || 2000,
+        viewportSize: options.viewportSize || {
+            width: 1280,
+            height: 800
+        }
+    };
+    const screenshots = getScreenshotsForLLM(events, options.maxScreenshots || 3);
+    const promptText = `Generate a Playwright test using @rpascene/web/playwright that reproduces this recorded browser session. The test should be based on the following events and follow the structure of the example provided. Make the test descriptive with appropriate assertions and validations.
+
+Event Summary:
+${JSON.stringify(playwrightSummary, null, 2)}
+
+Generated code should:
+1. Import required dependencies
+2. Set up the test with proper configuration
+3. Include a beforeEach hook to navigate to the starting URL
+4. Implement a test that uses Rpascene AI methods (aiTap, aiInput, aiAssert, etc.)
+5. Include appropriate assertions and validations
+6. Follow best practices for Playwright tests
+7. Be ready to execute without further modification
+
+Important: Return ONLY the raw Playwright test code. Do NOT wrap the response in markdown code blocks (no \`\`\`typescript, \`\`\`javascript or \`\`\`). Start directly with the code content.`;
+    const messageContent = createMessageContent(promptText, screenshots, false !== options.includeScreenshots);
+    const systemPrompt = `You are an expert test automation engineer specializing in Playwright and Rpascene. 
+Your task is to generate a complete, executable Playwright test using @rpascene/web/playwright that reproduces a recorded browser session.
+
+${PLAYWRIGHT_EXAMPLE_CODE}`;
+    const prompt = [
+        {
+            role: 'system',
+            content: systemPrompt
+        },
+        {
+            role: 'user',
+            content: messageContent
+        }
+    ];
+    const response = await callAIWithStringResponse(prompt, AIActionType.TEXT, modelConfig);
+    if ((null == response ? void 0 : response.content) && 'string' == typeof response.content) return response.content;
+    throw new Error('Failed to generate Playwright test code');
+};
+const generatePlaywrightTestStream = async (events, options, modelConfig)=>{
+    validateEvents(events);
+    const summary = prepareEventSummary(events, {
+        testName: options.testName,
+        maxScreenshots: options.maxScreenshots || 3
+    });
+    const playwrightSummary = {
+        ...summary,
+        waitForNetworkIdle: false !== options.waitForNetworkIdle,
+        waitForNetworkIdleTimeout: options.waitForNetworkIdleTimeout || 2000,
+        viewportSize: options.viewportSize || {
+            width: 1280,
+            height: 800
+        }
+    };
+    const screenshots = getScreenshotsForLLM(events, options.maxScreenshots || 3);
+    const promptText = `Generate a Playwright test using @rpascene/web/playwright that reproduces this recorded browser session. The test should be based on the following events and follow the structure of the example provided. Make the test descriptive with appropriate assertions and validations.
+
+Event Summary:
+${JSON.stringify(playwrightSummary, null, 2)}
+
+Generated code should:
+1. Import required dependencies
+2. Set up the test with proper configuration
+3. Include a beforeEach hook to navigate to the starting URL
+4. Implement a test that uses Rpascene AI methods (aiTap, aiInput, aiAssert, etc.)
+5. Include appropriate assertions and validations
+6. Follow best practices for Playwright tests
+7. Be ready to execute without further modification
+8. can't wrap this test code in markdown code block
+
+Important: Return ONLY the raw Playwright test code. Do NOT wrap the response in markdown code blocks (no \`\`\`typescript, \`\`\`javascript or \`\`\`). Start directly with the code content.`;
+    const messageContent = createMessageContent(promptText, screenshots, false !== options.includeScreenshots);
+    const systemPrompt = `You are an expert test automation engineer specializing in Playwright and Rpascene. 
+Your task is to generate a complete, executable Playwright test using @rpascene/web/playwright that reproduces a recorded browser session.
+
+${PLAYWRIGHT_EXAMPLE_CODE}`;
+    const prompt = [
+        {
+            role: 'system',
+            content: systemPrompt
+        },
+        {
+            role: 'user',
+            content: messageContent
+        }
+    ];
+    if (options.stream && options.onChunk) return await callAI(prompt, AIActionType.TEXT, modelConfig, {
+        stream: true,
+        onChunk: options.onChunk
+    });
+    {
+        const response = await callAIWithStringResponse(prompt, AIActionType.TEXT, modelConfig);
+        if ((null == response ? void 0 : response.content) && 'string' == typeof response.content) return {
+            content: response.content,
+            usage: response.usage,
+            isStreamed: false
+        };
+        throw new Error('Failed to generate Playwright test code');
+    }
+};
+export { createEventCounts, createMessageContent, extractInputDescriptions, filterEventsByType, generatePlaywrightTest, generatePlaywrightTestStream, getScreenshotsForLLM, prepareEventSummary, processEventsForLLM, validateEvents };
+
+//# sourceMappingURL=playwright-generator.mjs.map

package/dist/es/ai-model/prompt/playwright-generator.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"ai-model\\prompt\\playwright-generator.mjs","sources":["webpack://@rpascene/core/./src/ai-model/prompt/playwright-generator.ts"],"sourcesContent":["import type {\n  StreamingAIResponse,\n  StreamingCodeGenerationOptions,\n} from '@/types';\nimport { PLAYWRIGHT_EXAMPLE_CODE } from '@rpascene/shared/constants';\nimport type { IModelConfig } from '@rpascene/shared/env';\nimport type { ChatCompletionMessageParam } from 'openai/resources/index';\nimport { AIActionType, callAI, callAIWithStringResponse } from '../index';\n// Import shared utilities and types from yaml-generator\nimport {\n  type ChromeRecordedEvent,\n  type EventCounts,\n  type EventSummary,\n  type InputDescription,\n  type ProcessedEvent,\n  createEventCounts,\n  createMessageContent,\n  extractInputDescriptions,\n  filterEventsByType,\n  getScreenshotsForLLM,\n  prepareEventSummary,\n  processEventsForLLM,\n  validateEvents,\n} from './yaml-generator';\n\n// Playwright-specific interfaces\nexport interface PlaywrightGenerationOptions {\n  testName?: string;\n  includeScreenshots?: boolean;\n  includeTimestamps?: boolean;\n  maxScreenshots?: number;\n  description?: string;\n  viewportSize?: { width: number; height: number };\n  waitForNetworkIdle?: boolean;\n  waitForNetworkIdleTimeout?: number;\n}\n\n// Re-export shared types for backward compatibility\nexport type {\n  ChromeRecordedEvent,\n  EventCounts,\n  InputDescription,\n  ProcessedEvent,\n  EventSummary,\n};\n\n// Re-export shared utilities for backward compatibility\nexport {\n  getScreenshotsForLLM,\n  filterEventsByType,\n  createEventCounts,\n  extractInputDescriptions,\n  processEventsForLLM,\n  prepareEventSummary,\n  createMessageContent,\n  validateEvents,\n};\n\n/**\n * Generates Playwright test code from recorded events\n */\nexport const generatePlaywrightTest = async (\n  events: ChromeRecordedEvent[],\n  options: PlaywrightGenerationOptions,\n  modelConfig: IModelConfig,\n): Promise<string> => {\n  // Validate input\n  validateEvents(events);\n\n  // Prepare event summary using shared utilities\n  const summary = prepareEventSummary(events, {\n    testName: options.testName,\n    maxScreenshots: options.maxScreenshots || 3,\n  });\n\n  // Add Playwright-specific options to summary\n  const playwrightSummary = {\n    ...summary,\n    waitForNetworkIdle: options.waitForNetworkIdle !== false,\n    waitForNetworkIdleTimeout: options.waitForNetworkIdleTimeout || 2000,\n    viewportSize: options.viewportSize || { width: 1280, height: 800 },\n  };\n\n  // Get screenshots for visual context\n  const screenshots = getScreenshotsForLLM(events, options.maxScreenshots || 3);\n\n  // Create prompt text\n  const promptText = `Generate a Playwright test using @rpascene/web/playwright that reproduces this recorded browser session. The test should be based on the following events and follow the structure of the example provided. Make the test descriptive with appropriate assertions and validations.\n\nEvent Summary:\n${JSON.stringify(playwrightSummary, null, 2)}\n\nGenerated code should:\n1. Import required dependencies\n2. Set up the test with proper configuration\n3. Include a beforeEach hook to navigate to the starting URL\n4. Implement a test that uses Rpascene AI methods (aiTap, aiInput, aiAssert, etc.)\n5. Include appropriate assertions and validations\n6. Follow best practices for Playwright tests\n7. Be ready to execute without further modification\n\nImportant: Return ONLY the raw Playwright test code. Do NOT wrap the response in markdown code blocks (no \\`\\`\\`typescript, \\`\\`\\`javascript or \\`\\`\\`). Start directly with the code content.`;\n\n  // Create message content with screenshots\n  const messageContent = createMessageContent(\n    promptText,\n    screenshots,\n    options.includeScreenshots !== false,\n  );\n\n  // Create system prompt\n  const systemPrompt = `You are an expert test automation engineer specializing in Playwright and Rpascene. \nYour task is to generate a complete, executable Playwright test using @rpascene/web/playwright that reproduces a recorded browser session.\n\n${PLAYWRIGHT_EXAMPLE_CODE}`;\n\n  // Use LLM to generate the Playwright test code\n  const prompt: ChatCompletionMessageParam[] = [\n    {\n      role: 'system',\n      content: systemPrompt,\n    },\n    {\n      role: 'user',\n      content: messageContent,\n    },\n  ];\n\n  const response = await callAIWithStringResponse(\n    prompt,\n    AIActionType.TEXT,\n    modelConfig,\n  );\n\n  if (response?.content && typeof response.content === 'string') {\n    return response.content;\n  }\n\n  throw new Error('Failed to generate Playwright test code');\n};\n\n/**\n * Generates Playwright test code from recorded events with streaming support\n */\nexport const generatePlaywrightTestStream = async (\n  events: ChromeRecordedEvent[],\n  options: PlaywrightGenerationOptions & StreamingCodeGenerationOptions,\n  modelConfig: IModelConfig,\n): Promise<StreamingAIResponse> => {\n  // Validate input\n  validateEvents(events);\n\n  // Prepare event summary using shared utilities\n  const summary = prepareEventSummary(events, {\n    testName: options.testName,\n    maxScreenshots: options.maxScreenshots || 3,\n  });\n\n  // Add Playwright-specific options to summary\n  const playwrightSummary = {\n    ...summary,\n    waitForNetworkIdle: options.waitForNetworkIdle !== false,\n    waitForNetworkIdleTimeout: options.waitForNetworkIdleTimeout || 2000,\n    viewportSize: options.viewportSize || { width: 1280, height: 800 },\n  };\n\n  // Get screenshots for visual context\n  const screenshots = getScreenshotsForLLM(events, options.maxScreenshots || 3);\n\n  // Create prompt text\n  const promptText = `Generate a Playwright test using @rpascene/web/playwright that reproduces this recorded browser session. The test should be based on the following events and follow the structure of the example provided. Make the test descriptive with appropriate assertions and validations.\n\nEvent Summary:\n${JSON.stringify(playwrightSummary, null, 2)}\n\nGenerated code should:\n1. Import required dependencies\n2. Set up the test with proper configuration\n3. Include a beforeEach hook to navigate to the starting URL\n4. Implement a test that uses Rpascene AI methods (aiTap, aiInput, aiAssert, etc.)\n5. Include appropriate assertions and validations\n6. Follow best practices for Playwright tests\n7. Be ready to execute without further modification\n8. can't wrap this test code in markdown code block\n\nImportant: Return ONLY the raw Playwright test code. Do NOT wrap the response in markdown code blocks (no \\`\\`\\`typescript, \\`\\`\\`javascript or \\`\\`\\`). Start directly with the code content.`;\n\n  // Create message content with screenshots\n  const messageContent = createMessageContent(\n    promptText,\n    screenshots,\n    options.includeScreenshots !== false,\n  );\n\n  // Create system prompt\n  const systemPrompt = `You are an expert test automation engineer specializing in Playwright and Rpascene. \nYour task is to generate a complete, executable Playwright test using @rpascene/web/playwright that reproduces a recorded browser session.\n\n${PLAYWRIGHT_EXAMPLE_CODE}`;\n\n  // Use LLM to generate the Playwright test code with streaming\n  const prompt: ChatCompletionMessageParam[] = [\n    {\n      role: 'system',\n      content: systemPrompt,\n    },\n    {\n      role: 'user',\n      content: messageContent,\n    },\n  ];\n\n  if (options.stream && options.onChunk) {\n    // Use streaming\n    return await callAI(prompt, AIActionType.TEXT, modelConfig, {\n      stream: true,\n      onChunk: options.onChunk,\n    });\n  } else {\n    // Fallback to non-streaming\n    const response = await callAIWithStringResponse(\n      prompt,\n      AIActionType.TEXT,\n      modelConfig,\n    );\n\n    if (response?.content && typeof response.content === 'string') {\n      return {\n        content: response.content,\n        usage: response.usage,\n        isStreamed: false,\n      };\n    }\n\n    throw new Error('Failed to generate Playwright test code');\n  }\n};\n"],"names":["generatePlaywrightTest","events","options","modelConfig","validateEvents","summary","prepareEventSummary","playwrightSummary","screenshots","getScreenshotsForLLM","promptText","JSON","messageContent","createMessageContent","systemPrompt","PLAYWRIGHT_EXAMPLE_CODE","prompt","response","callAIWithStringResponse","AIActionType","Error","generatePlaywrightTestStream","callAI"],"mappings":";;;AA6DO,MAAMA,yBAAyB,OACpCC,QACAC,SACAC;IAGAC,eAAeH;IAGf,MAAMI,UAAUC,oBAAoBL,QAAQ;QAC1C,UAAUC,QAAQ,QAAQ;QAC1B,gBAAgBA,QAAQ,cAAc,IAAI;IAC5C;IAGA,MAAMK,oBAAoB;QACxB,GAAGF,OAAO;QACV,oBAAoBH,AAA+B,UAA/BA,QAAQ,kBAAkB;QAC9C,2BAA2BA,QAAQ,yBAAyB,IAAI;QAChE,cAAcA,QAAQ,YAAY,IAAI;YAAE,OAAO;YAAM,QAAQ;QAAI;IACnE;IAGA,MAAMM,cAAcC,qBAAqBR,QAAQC,QAAQ,cAAc,IAAI;IAG3E,MAAMQ,aAAa,CAAC;;;AAGtB,EAAEC,KAAK,SAAS,CAACJ,mBAAmB,MAAM,GAAG;;;;;;;;;;;8LAWiJ,CAAC;IAG7L,MAAMK,iBAAiBC,qBACrBH,YACAF,aACAN,AAA+B,UAA/BA,QAAQ,kBAAkB;IAI5B,MAAMY,eAAe,CAAC;;;AAGxB,EAAEC,yBAAyB;IAGzB,MAAMC,SAAuC;QAC3C;YACE,MAAM;YACN,SAASF;QACX;QACA;YACE,MAAM;YACN,SAASF;QACX;KACD;IAED,MAAMK,WAAW,MAAMC,yBACrBF,QACAG,aAAa,IAAI,EACjBhB;IAGF,IAAIc,AAAAA,CAAAA,QAAAA,WAAAA,KAAAA,IAAAA,SAAU,OAAO,AAAD,KAAK,AAA4B,YAA5B,OAAOA,SAAS,OAAO,EAC9C,OAAOA,SAAS,OAAO;IAGzB,MAAM,IAAIG,MAAM;AAClB;AAKO,MAAMC,+BAA+B,OAC1CpB,QACAC,SACAC;IAGAC,eAAeH;IAGf,MAAMI,UAAUC,oBAAoBL,QAAQ;QAC1C,UAAUC,QAAQ,QAAQ;QAC1B,gBAAgBA,QAAQ,cAAc,IAAI;IAC5C;IAGA,MAAMK,oBAAoB;QACxB,GAAGF,OAAO;QACV,oBAAoBH,AAA+B,UAA/BA,QAAQ,kBAAkB;QAC9C,2BAA2BA,QAAQ,yBAAyB,IAAI;QAChE,cAAcA,QAAQ,YAAY,IAAI;YAAE,OAAO;YAAM,QAAQ;QAAI;IACnE;IAGA,MAAMM,cAAcC,qBAAqBR,QAAQC,QAAQ,cAAc,IAAI;IAG3E,MAAMQ,aAAa,CAAC;;;AAGtB,EAAEC,KAAK,SAAS,CAACJ,mBAAmB,MAAM,GAAG;;;;;;;;;;;;8LAYiJ,CAAC;IAG7L,MAAMK,iBAAiBC,qBACrBH,YACAF,aACAN,AAA+B,UAA/BA,QAAQ,kBAAkB;IAI5B,MAAMY,eAAe,CAAC;;;AAGxB,EAAEC,yBAAyB;IAGzB,MAAMC,SAAuC;QAC3C;YACE,MAAM;YACN,SAASF;QACX;QACA;YACE,MAAM;YACN,SAASF;QACX;KACD;IAED,IAAIV,QAAQ,MAAM,IAAIA,QAAQ,OAAO,EAEnC,OAAO,MAAMoB,OAAON,QAAQG,aAAa,IAAI,EAAEhB,aAAa;QAC1D,QAAQ;QACR,SAASD,QAAQ,OAAO;IAC1B;IACK;QAEL,MAAMe,WAAW,MAAMC,yBACrBF,QACAG,aAAa,IAAI,EACjBhB;QAGF,IAAIc,AAAAA,CAAAA,QAAAA,WAAAA,KAAAA,IAAAA,SAAU,OAAO,AAAD,KAAK,AAA4B,YAA5B,OAAOA,SAAS,OAAO,EAC9C,OAAO;YACL,SAASA,SAAS,OAAO;YACzB,OAAOA,SAAS,KAAK;YACrB,YAAY;QACd;QAGF,MAAM,IAAIG,MAAM;IAClB;AACF"}