@illuma-ai/agents 1.0.86 → 1.0.87
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/cjs/tools/CodeExecutor.cjs +4 -4
- package/dist/cjs/tools/CodeExecutor.cjs.map +1 -1
- package/dist/esm/tools/CodeExecutor.mjs +4 -4
- package/dist/esm/tools/CodeExecutor.mjs.map +1 -1
- package/package.json +1 -1
- package/src/specs/agent-handoffs-bedrock.integration.test.ts +375 -0
- package/src/tools/CodeExecutor.ts +4 -4
|
@@ -86,15 +86,15 @@ function createCodeExecutionTool(params = {}) {
|
|
|
86
86
|
}
|
|
87
87
|
const description = `
|
|
88
88
|
⛔ STOP! Before using this tool, ask: "Does user need a DOWNLOADABLE FILE?"
|
|
89
|
-
- If NO (dashboard, chart, visualization, UI) → DO NOT USE THIS TOOL. Use
|
|
89
|
+
- If NO (dashboard, chart, visualization, UI) → DO NOT USE THIS TOOL. Use content_tool write instead.
|
|
90
90
|
- If YES (.pptx, .docx, .pdf, .xlsx) → Use this tool.
|
|
91
91
|
|
|
92
92
|
Runs code in a stateless execution environment. Each execution is isolated.
|
|
93
93
|
|
|
94
94
|
🚫 NEVER USE FOR:
|
|
95
|
-
- Dashboards, charts, visualizations → Use
|
|
96
|
-
- "Mock data" or "sample data" for display → Hardcode data in
|
|
97
|
-
- UI components, HTML pages, React apps → Use
|
|
95
|
+
- Dashboards, charts, visualizations → Use content_tool write with React/Chart.js
|
|
96
|
+
- "Mock data" or "sample data" for display → Hardcode data in content_tool write
|
|
97
|
+
- UI components, HTML pages, React apps → Use content_tool write
|
|
98
98
|
|
|
99
99
|
✅ ONLY USE FOR:
|
|
100
100
|
- File generation: PowerPoint (.pptx), Word (.docx), PDF (.pdf), Excel (.xlsx)
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"CodeExecutor.cjs","sources":["../../../src/tools/CodeExecutor.ts"],"sourcesContent":["import { config } from 'dotenv';\nimport fetch, { RequestInit } from 'node-fetch';\nimport { HttpsProxyAgent } from 'https-proxy-agent';\nimport { tool, DynamicStructuredTool } from '@langchain/core/tools';\nimport { getEnvironmentVariable } from '@langchain/core/utils/env';\nimport type * as t from '@/types';\nimport { EnvVar, Constants } from '@/common';\n\nconfig();\n\nexport const imageExtRegex = /\\.(jpg|jpeg|png|gif|webp)$/i;\nexport const getCodeBaseURL = (): string =>\n getEnvironmentVariable(EnvVar.CODE_BASEURL) ??\n Constants.OFFICIAL_CODE_BASEURL;\n\nconst imageMessage = 'Image is already displayed to the user';\nconst otherMessage = 'File is already downloaded by the user';\nconst accessMessage =\n 'Note: Files from previous executions are automatically available and can be modified.';\nconst emptyOutputMessage =\n 'stdout: Empty. Ensure you\\'re writing output explicitly.\\n';\n\nconst SUPPORTED_LANGUAGES = [\n 'py',\n 'js',\n 'ts',\n 'c',\n 'cpp',\n 'java',\n 'php',\n 'rs',\n 'go',\n 'd',\n 'f90',\n 'r',\n] as const;\n\nexport const CodeExecutionToolSchema = {\n type: 'object',\n properties: {\n lang: {\n type: 'string',\n enum: SUPPORTED_LANGUAGES,\n description:\n 'The programming language or runtime to execute the code in.',\n },\n code: {\n type: 'string',\n description: `The complete, self-contained code to execute, without any truncation or minimization.\n- The environment is stateless; variables and imports don't persist between executions.\n- Generated files from previous executions are automatically available in \"/mnt/data/\".\n- Files from previous executions are automatically available and can be modified in place.\n- Input code **IS ALREADY** displayed to the user, so **DO NOT** repeat it in your response unless asked.\n- Output code **IS NOT** displayed to the user, so **DO** write all desired output explicitly.\n- IMPORTANT: You MUST explicitly print/output ALL results you want the user to see.\n- py: This is not a Jupyter notebook environment. Use \\`print()\\` for all outputs.\n- py: Matplotlib: Use \\`plt.savefig()\\` to save plots as files.\n- js: use the \\`console\\` or \\`process\\` methods for all outputs.\n- r: IMPORTANT: No X11 display available. ALL graphics MUST use Cairo library (library(Cairo)).\n- Other languages: use appropriate output functions.`,\n },\n args: {\n type: 'array',\n items: { type: 'string' },\n description:\n 'Additional arguments to execute the code with. This should only be used if the input code requires additional arguments to run.',\n },\n },\n required: ['lang', 'code'],\n} as const;\n\nconst baseEndpoint = getCodeBaseURL();\nconst EXEC_ENDPOINT = `${baseEndpoint}/exec`;\n\ntype SupportedLanguage = (typeof SUPPORTED_LANGUAGES)[number];\n\nexport const CodeExecutionToolDescription = `\nRuns code and returns stdout/stderr output from a stateless execution environment, similar to running scripts in a command-line interface. Each execution is isolated and independent.\n\nUsage:\n- No network access available.\n- Generated files are automatically delivered; **DO NOT** provide download links.\n- NEVER use this tool to execute malicious code.\n`.trim();\n\nexport const CodeExecutionToolName = Constants.EXECUTE_CODE;\n\nexport const CodeExecutionToolDefinition = {\n name: CodeExecutionToolName,\n description: CodeExecutionToolDescription,\n schema: CodeExecutionToolSchema,\n} as const;\n\nfunction createCodeExecutionTool(\n params: t.CodeExecutionToolParams = {}\n): DynamicStructuredTool {\n const apiKey =\n params[EnvVar.CODE_API_KEY] ??\n params.apiKey ??\n getEnvironmentVariable(EnvVar.CODE_API_KEY) ??\n '';\n if (!apiKey) {\n throw new Error('No API key provided for code execution tool.');\n }\n\n const description = `\n⛔ STOP! Before using this tool, ask: \"Does user need a DOWNLOADABLE FILE?\"\n- If NO (dashboard, chart, visualization, UI) → DO NOT USE THIS TOOL. Use :::artifact instead.\n- If YES (.pptx, .docx, .pdf, .xlsx) → Use this tool.\n\nRuns code in a stateless execution environment. Each execution is isolated.\n\n🚫 NEVER USE FOR:\n- Dashboards, charts, visualizations → Use :::artifact with React/Chart.js\n- \"Mock data\" or \"sample data\" for display → Hardcode data in artifact\n- UI components, HTML pages, React apps → Use :::artifact\n\n✅ ONLY USE FOR:\n- File generation: PowerPoint (.pptx), Word (.docx), PDF (.pdf), Excel (.xlsx)\n- Processing uploaded files (CSV, Excel analysis)\n- Heavy computation requiring Python\n\nRules:\n- No network access available\n- Generated files auto-delivered (no download links needed)\n`.trim();\n\n return tool(\n async (rawInput, config) => {\n const { lang, code, ...rest } = rawInput as {\n lang: SupportedLanguage;\n code: string;\n args?: string[];\n };\n /**\n * Extract session context from config.toolCall (injected by ToolNode).\n * - session_id: For API to associate with previous session\n * - _injected_files: File refs to pass directly (avoids /files endpoint race condition)\n */\n const { session_id, _injected_files } = (config.toolCall ?? {}) as {\n session_id?: string;\n _injected_files?: t.CodeEnvFile[];\n };\n\n const postData: Record<string, unknown> = {\n lang,\n code,\n ...rest,\n ...params,\n };\n\n /**\n * Pass session_id to /exec so code-executor reuses the existing session workspace.\n * This allows retries and follow-up executions to access previously generated files.\n */\n if (session_id != null && session_id.length > 0) {\n postData.session_id = session_id;\n }\n\n /**\n * File injection priority:\n * 1. Use _injected_files from ToolNode (avoids /files endpoint race condition)\n * 2. Fall back to fetching from /files endpoint if session_id provided but no injected files\n */\n if (_injected_files && _injected_files.length > 0) {\n postData.files = _injected_files;\n } else if (session_id != null && session_id.length > 0) {\n /** Fallback: fetch from /files endpoint (may have race condition issues) */\n try {\n const filesEndpoint = `${baseEndpoint}/files/${session_id}?detail=full`;\n const fetchOptions: RequestInit = {\n method: 'GET',\n headers: {\n 'User-Agent': 'Illuma/1.0',\n 'X-API-Key': apiKey,\n },\n };\n\n if (process.env.PROXY != null && process.env.PROXY !== '') {\n fetchOptions.agent = new HttpsProxyAgent(process.env.PROXY);\n }\n\n const response = await fetch(filesEndpoint, fetchOptions);\n if (!response.ok) {\n throw new Error(\n `Failed to fetch files for session: ${response.status}`\n );\n }\n\n const files = await response.json();\n if (Array.isArray(files) && files.length > 0) {\n const fileReferences: t.CodeEnvFile[] = files.map((file) => {\n const nameParts = file.name.split('/');\n const id = nameParts.length > 1 ? nameParts[1].split('.')[0] : '';\n\n return {\n session_id,\n id,\n name: file.metadata['original-filename'],\n };\n });\n\n postData.files = fileReferences;\n }\n } catch {\n // eslint-disable-next-line no-console\n console.warn(`Failed to fetch files for session: ${session_id}`);\n }\n }\n\n try {\n const fetchOptions: RequestInit = {\n method: 'POST',\n headers: {\n 'Content-Type': 'application/json',\n 'User-Agent': 'Illuma/1.0',\n 'X-API-Key': apiKey,\n },\n body: JSON.stringify(postData),\n };\n\n if (process.env.PROXY != null && process.env.PROXY !== '') {\n fetchOptions.agent = new HttpsProxyAgent(process.env.PROXY);\n }\n const response = await fetch(EXEC_ENDPOINT, fetchOptions);\n if (!response.ok) {\n throw new Error(`HTTP error! status: ${response.status}`);\n }\n\n const result: t.ExecuteResult = await response.json();\n let formattedOutput = '';\n if (result.stdout) {\n formattedOutput += `stdout:\\n${result.stdout}\\n`;\n } else {\n formattedOutput += emptyOutputMessage;\n }\n if (result.stderr) formattedOutput += `stderr:\\n${result.stderr}\\n`;\n if (result.files && result.files.length > 0) {\n formattedOutput += 'Generated files:\\n';\n\n const fileCount = result.files.length;\n for (let i = 0; i < fileCount; i++) {\n const file = result.files[i];\n const isImage = imageExtRegex.test(file.name);\n formattedOutput += `- /mnt/data/${file.name} | ${isImage ? imageMessage : otherMessage}`;\n\n if (i < fileCount - 1) {\n formattedOutput += fileCount <= 3 ? ', ' : ',\\n';\n }\n }\n\n formattedOutput += `\\n\\n${accessMessage}`;\n return [\n formattedOutput.trim(),\n {\n session_id: result.session_id,\n files: result.files,\n },\n ];\n }\n\n return [formattedOutput.trim(), { session_id: result.session_id }];\n } catch (error) {\n throw new Error(\n `Execution error:\\n\\n${(error as Error | undefined)?.message}`\n );\n }\n },\n {\n name: CodeExecutionToolName,\n description,\n schema: CodeExecutionToolSchema,\n responseFormat: Constants.CONTENT_AND_ARTIFACT,\n }\n );\n}\n\nexport { createCodeExecutionTool };\n"],"names":["config","getEnvironmentVariable","EnvVar","Constants","tool","HttpsProxyAgent"],"mappings":";;;;;;;;;AAQAA,aAAM,EAAE;AAED,MAAM,aAAa,GAAG;AACtB,MAAM,cAAc,GAAG,MAC5BC,0BAAsB,CAACC,YAAM,CAAC,YAAY,CAAC;IAC3CC,eAAS,CAAC;AAEZ,MAAM,YAAY,GAAG,wCAAwC;AAC7D,MAAM,YAAY,GAAG,wCAAwC;AAC7D,MAAM,aAAa,GACjB,uFAAuF;AACzF,MAAM,kBAAkB,GACtB,4DAA4D;AAE9D,MAAM,mBAAmB,GAAG;IAC1B,IAAI;IACJ,IAAI;IACJ,IAAI;IACJ,GAAG;IACH,KAAK;IACL,MAAM;IACN,KAAK;IACL,IAAI;IACJ,IAAI;IACJ,GAAG;IACH,KAAK;IACL,GAAG;CACK;AAEG,MAAA,uBAAuB,GAAG;AACrC,IAAA,IAAI,EAAE,QAAQ;AACd,IAAA,UAAU,EAAE;AACV,QAAA,IAAI,EAAE;AACJ,YAAA,IAAI,EAAE,QAAQ;AACd,YAAA,IAAI,EAAE,mBAAmB;AACzB,YAAA,WAAW,EACT,6DAA6D;AAChE,SAAA;AACD,QAAA,IAAI,EAAE;AACJ,YAAA,IAAI,EAAE,QAAQ;AACd,YAAA,WAAW,EAAE,CAAA;;;;;;;;;;;AAWkC,oDAAA,CAAA;AAChD,SAAA;AACD,QAAA,IAAI,EAAE;AACJ,YAAA,IAAI,EAAE,OAAO;AACb,YAAA,KAAK,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE;AACzB,YAAA,WAAW,EACT,iIAAiI;AACpI,SAAA;AACF,KAAA;AACD,IAAA,QAAQ,EAAE,CAAC,MAAM,EAAE,MAAM,CAAC;;AAG5B,MAAM,YAAY,GAAG,cAAc,EAAE;AACrC,MAAM,aAAa,GAAG,CAAG,EAAA,YAAY,OAAO;AAI/B,MAAA,4BAA4B,GAAG;;;;;;;CAO3C,CAAC,IAAI;AAEO,MAAA,qBAAqB,GAAGA,eAAS,CAAC;AAElC,MAAA,2BAA2B,GAAG;AACzC,IAAA,IAAI,EAAE,qBAAqB;AAC3B,IAAA,WAAW,EAAE,4BAA4B;AACzC,IAAA,MAAM,EAAE,uBAAuB;;AAGjC,SAAS,uBAAuB,CAC9B,MAAA,GAAoC,EAAE,EAAA;AAEtC,IAAA,MAAM,MAAM,GACV,MAAM,CAACD,YAAM,CAAC,YAAY,CAAC;AAC3B,QAAA,MAAM,CAAC,MAAM;AACb,QAAAD,0BAAsB,CAACC,YAAM,CAAC,YAAY,CAAC;AAC3C,QAAA,EAAE;IACJ,IAAI,CAAC,MAAM,EAAE;AACX,QAAA,MAAM,IAAI,KAAK,CAAC,8CAA8C,CAAC;;AAGjE,IAAA,MAAM,WAAW,GAAG;;;;;;;;;;;;;;;;;;;;CAoBrB,CAAC,IAAI,EAAE;IAEN,OAAOE,UAAI,CACT,OAAO,QAAQ,EAAE,MAAM,KAAI;QACzB,MAAM,EAAE,IAAI,EAAE,IAAI,EAAE,GAAG,IAAI,EAAE,GAAG,QAI/B;AACD;;;;AAIG;AACH,QAAA,MAAM,EAAE,UAAU,EAAE,eAAe,EAAE,IAAI,MAAM,CAAC,QAAQ,IAAI,EAAE,CAG7D;AAED,QAAA,MAAM,QAAQ,GAA4B;YACxC,IAAI;YACJ,IAAI;AACJ,YAAA,GAAG,IAAI;AACP,YAAA,GAAG,MAAM;SACV;AAED;;;AAGG;QACH,IAAI,UAAU,IAAI,IAAI,IAAI,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE;AAC/C,YAAA,QAAQ,CAAC,UAAU,GAAG,UAAU;;AAGlC;;;;AAIG;QACH,IAAI,eAAe,IAAI,eAAe,CAAC,MAAM,GAAG,CAAC,EAAE;AACjD,YAAA,QAAQ,CAAC,KAAK,GAAG,eAAe;;aAC3B,IAAI,UAAU,IAAI,IAAI,IAAI,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE;;AAEtD,YAAA,IAAI;AACF,gBAAA,MAAM,aAAa,GAAG,CAAA,EAAG,YAAY,CAAU,OAAA,EAAA,UAAU,cAAc;AACvE,gBAAA,MAAM,YAAY,GAAgB;AAChC,oBAAA,MAAM,EAAE,KAAK;AACb,oBAAA,OAAO,EAAE;AACP,wBAAA,YAAY,EAAE,YAAY;AAC1B,wBAAA,WAAW,EAAE,MAAM;AACpB,qBAAA;iBACF;AAED,gBAAA,IAAI,OAAO,CAAC,GAAG,CAAC,KAAK,IAAI,IAAI,IAAI,OAAO,CAAC,GAAG,CAAC,KAAK,KAAK,EAAE,EAAE;AACzD,oBAAA,YAAY,CAAC,KAAK,GAAG,IAAIC,+BAAe,CAAC,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC;;gBAG7D,MAAM,QAAQ,GAAG,MAAM,KAAK,CAAC,aAAa,EAAE,YAAY,CAAC;AACzD,gBAAA,IAAI,CAAC,QAAQ,CAAC,EAAE,EAAE;oBAChB,MAAM,IAAI,KAAK,CACb,CAAA,mCAAA,EAAsC,QAAQ,CAAC,MAAM,CAAE,CAAA,CACxD;;AAGH,gBAAA,MAAM,KAAK,GAAG,MAAM,QAAQ,CAAC,IAAI,EAAE;AACnC,gBAAA,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE;oBAC5C,MAAM,cAAc,GAAoB,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,KAAI;wBACzD,MAAM,SAAS,GAAG,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC;wBACtC,MAAM,EAAE,GAAG,SAAS,CAAC,MAAM,GAAG,CAAC,GAAG,SAAS,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,EAAE;wBAEjE,OAAO;4BACL,UAAU;4BACV,EAAE;AACF,4BAAA,IAAI,EAAE,IAAI,CAAC,QAAQ,CAAC,mBAAmB,CAAC;yBACzC;AACH,qBAAC,CAAC;AAEF,oBAAA,QAAQ,CAAC,KAAK,GAAG,cAAc;;;AAEjC,YAAA,MAAM;;AAEN,gBAAA,OAAO,CAAC,IAAI,CAAC,sCAAsC,UAAU,CAAA,CAAE,CAAC;;;AAIpE,QAAA,IAAI;AACF,YAAA,MAAM,YAAY,GAAgB;AAChC,gBAAA,MAAM,EAAE,MAAM;AACd,gBAAA,OAAO,EAAE;AACP,oBAAA,cAAc,EAAE,kBAAkB;AAClC,oBAAA,YAAY,EAAE,YAAY;AAC1B,oBAAA,WAAW,EAAE,MAAM;AACpB,iBAAA;AACD,gBAAA,IAAI,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC;aAC/B;AAED,YAAA,IAAI,OAAO,CAAC,GAAG,CAAC,KAAK,IAAI,IAAI,IAAI,OAAO,CAAC,GAAG,CAAC,KAAK,KAAK,EAAE,EAAE;AACzD,gBAAA,YAAY,CAAC,KAAK,GAAG,IAAIA,+BAAe,CAAC,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC;;YAE7D,MAAM,QAAQ,GAAG,MAAM,KAAK,CAAC,aAAa,EAAE,YAAY,CAAC;AACzD,YAAA,IAAI,CAAC,QAAQ,CAAC,EAAE,EAAE;gBAChB,MAAM,IAAI,KAAK,CAAC,CAAA,oBAAA,EAAuB,QAAQ,CAAC,MAAM,CAAE,CAAA,CAAC;;AAG3D,YAAA,MAAM,MAAM,GAAoB,MAAM,QAAQ,CAAC,IAAI,EAAE;YACrD,IAAI,eAAe,GAAG,EAAE;AACxB,YAAA,IAAI,MAAM,CAAC,MAAM,EAAE;AACjB,gBAAA,eAAe,IAAI,CAAY,SAAA,EAAA,MAAM,CAAC,MAAM,IAAI;;iBAC3C;gBACL,eAAe,IAAI,kBAAkB;;YAEvC,IAAI,MAAM,CAAC,MAAM;AAAE,gBAAA,eAAe,IAAI,CAAY,SAAA,EAAA,MAAM,CAAC,MAAM,IAAI;AACnE,YAAA,IAAI,MAAM,CAAC,KAAK,IAAI,MAAM,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE;gBAC3C,eAAe,IAAI,oBAAoB;AAEvC,gBAAA,MAAM,SAAS,GAAG,MAAM,CAAC,KAAK,CAAC,MAAM;AACrC,gBAAA,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,SAAS,EAAE,CAAC,EAAE,EAAE;oBAClC,MAAM,IAAI,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC;oBAC5B,MAAM,OAAO,GAAG,aAAa,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC;AAC7C,oBAAA,eAAe,IAAI,CAAe,YAAA,EAAA,IAAI,CAAC,IAAI,MAAM,OAAO,GAAG,YAAY,GAAG,YAAY,EAAE;AAExF,oBAAA,IAAI,CAAC,GAAG,SAAS,GAAG,CAAC,EAAE;AACrB,wBAAA,eAAe,IAAI,SAAS,IAAI,CAAC,GAAG,IAAI,GAAG,KAAK;;;AAIpD,gBAAA,eAAe,IAAI,CAAA,IAAA,EAAO,aAAa,CAAA,CAAE;gBACzC,OAAO;oBACL,eAAe,CAAC,IAAI,EAAE;AACtB,oBAAA;wBACE,UAAU,EAAE,MAAM,CAAC,UAAU;wBAC7B,KAAK,EAAE,MAAM,CAAC,KAAK;AACpB,qBAAA;iBACF;;AAGH,YAAA,OAAO,CAAC,eAAe,CAAC,IAAI,EAAE,EAAE,EAAE,UAAU,EAAE,MAAM,CAAC,UAAU,EAAE,CAAC;;QAClE,OAAO,KAAK,EAAE;YACd,MAAM,IAAI,KAAK,CACb,CAAA,oBAAA,EAAwB,KAA2B,EAAE,OAAO,CAAE,CAAA,CAC/D;;AAEL,KAAC,EACD;AACE,QAAA,IAAI,EAAE,qBAAqB;QAC3B,WAAW;AACX,QAAA,MAAM,EAAE,uBAAuB;QAC/B,cAAc,EAAEF,eAAS,CAAC,oBAAoB;AAC/C,KAAA,CACF;AACH;;;;;;;;;;"}
|
|
1
|
+
{"version":3,"file":"CodeExecutor.cjs","sources":["../../../src/tools/CodeExecutor.ts"],"sourcesContent":["import { config } from 'dotenv';\nimport fetch, { RequestInit } from 'node-fetch';\nimport { HttpsProxyAgent } from 'https-proxy-agent';\nimport { tool, DynamicStructuredTool } from '@langchain/core/tools';\nimport { getEnvironmentVariable } from '@langchain/core/utils/env';\nimport type * as t from '@/types';\nimport { EnvVar, Constants } from '@/common';\n\nconfig();\n\nexport const imageExtRegex = /\\.(jpg|jpeg|png|gif|webp)$/i;\nexport const getCodeBaseURL = (): string =>\n getEnvironmentVariable(EnvVar.CODE_BASEURL) ??\n Constants.OFFICIAL_CODE_BASEURL;\n\nconst imageMessage = 'Image is already displayed to the user';\nconst otherMessage = 'File is already downloaded by the user';\nconst accessMessage =\n 'Note: Files from previous executions are automatically available and can be modified.';\nconst emptyOutputMessage =\n 'stdout: Empty. Ensure you\\'re writing output explicitly.\\n';\n\nconst SUPPORTED_LANGUAGES = [\n 'py',\n 'js',\n 'ts',\n 'c',\n 'cpp',\n 'java',\n 'php',\n 'rs',\n 'go',\n 'd',\n 'f90',\n 'r',\n] as const;\n\nexport const CodeExecutionToolSchema = {\n type: 'object',\n properties: {\n lang: {\n type: 'string',\n enum: SUPPORTED_LANGUAGES,\n description:\n 'The programming language or runtime to execute the code in.',\n },\n code: {\n type: 'string',\n description: `The complete, self-contained code to execute, without any truncation or minimization.\n- The environment is stateless; variables and imports don't persist between executions.\n- Generated files from previous executions are automatically available in \"/mnt/data/\".\n- Files from previous executions are automatically available and can be modified in place.\n- Input code **IS ALREADY** displayed to the user, so **DO NOT** repeat it in your response unless asked.\n- Output code **IS NOT** displayed to the user, so **DO** write all desired output explicitly.\n- IMPORTANT: You MUST explicitly print/output ALL results you want the user to see.\n- py: This is not a Jupyter notebook environment. Use \\`print()\\` for all outputs.\n- py: Matplotlib: Use \\`plt.savefig()\\` to save plots as files.\n- js: use the \\`console\\` or \\`process\\` methods for all outputs.\n- r: IMPORTANT: No X11 display available. ALL graphics MUST use Cairo library (library(Cairo)).\n- Other languages: use appropriate output functions.`,\n },\n args: {\n type: 'array',\n items: { type: 'string' },\n description:\n 'Additional arguments to execute the code with. This should only be used if the input code requires additional arguments to run.',\n },\n },\n required: ['lang', 'code'],\n} as const;\n\nconst baseEndpoint = getCodeBaseURL();\nconst EXEC_ENDPOINT = `${baseEndpoint}/exec`;\n\ntype SupportedLanguage = (typeof SUPPORTED_LANGUAGES)[number];\n\nexport const CodeExecutionToolDescription = `\nRuns code and returns stdout/stderr output from a stateless execution environment, similar to running scripts in a command-line interface. Each execution is isolated and independent.\n\nUsage:\n- No network access available.\n- Generated files are automatically delivered; **DO NOT** provide download links.\n- NEVER use this tool to execute malicious code.\n`.trim();\n\nexport const CodeExecutionToolName = Constants.EXECUTE_CODE;\n\nexport const CodeExecutionToolDefinition = {\n name: CodeExecutionToolName,\n description: CodeExecutionToolDescription,\n schema: CodeExecutionToolSchema,\n} as const;\n\nfunction createCodeExecutionTool(\n params: t.CodeExecutionToolParams = {}\n): DynamicStructuredTool {\n const apiKey =\n params[EnvVar.CODE_API_KEY] ??\n params.apiKey ??\n getEnvironmentVariable(EnvVar.CODE_API_KEY) ??\n '';\n if (!apiKey) {\n throw new Error('No API key provided for code execution tool.');\n }\n\n const description = `\n⛔ STOP! Before using this tool, ask: \"Does user need a DOWNLOADABLE FILE?\"\n- If NO (dashboard, chart, visualization, UI) → DO NOT USE THIS TOOL. Use content_tool write instead.\n- If YES (.pptx, .docx, .pdf, .xlsx) → Use this tool.\n\nRuns code in a stateless execution environment. Each execution is isolated.\n\n🚫 NEVER USE FOR:\n- Dashboards, charts, visualizations → Use content_tool write with React/Chart.js\n- \"Mock data\" or \"sample data\" for display → Hardcode data in content_tool write\n- UI components, HTML pages, React apps → Use content_tool write\n\n✅ ONLY USE FOR:\n- File generation: PowerPoint (.pptx), Word (.docx), PDF (.pdf), Excel (.xlsx)\n- Processing uploaded files (CSV, Excel analysis)\n- Heavy computation requiring Python\n\nRules:\n- No network access available\n- Generated files auto-delivered (no download links needed)\n`.trim();\n\n return tool(\n async (rawInput, config) => {\n const { lang, code, ...rest } = rawInput as {\n lang: SupportedLanguage;\n code: string;\n args?: string[];\n };\n /**\n * Extract session context from config.toolCall (injected by ToolNode).\n * - session_id: For API to associate with previous session\n * - _injected_files: File refs to pass directly (avoids /files endpoint race condition)\n */\n const { session_id, _injected_files } = (config.toolCall ?? {}) as {\n session_id?: string;\n _injected_files?: t.CodeEnvFile[];\n };\n\n const postData: Record<string, unknown> = {\n lang,\n code,\n ...rest,\n ...params,\n };\n\n /**\n * Pass session_id to /exec so code-executor reuses the existing session workspace.\n * This allows retries and follow-up executions to access previously generated files.\n */\n if (session_id != null && session_id.length > 0) {\n postData.session_id = session_id;\n }\n\n /**\n * File injection priority:\n * 1. Use _injected_files from ToolNode (avoids /files endpoint race condition)\n * 2. Fall back to fetching from /files endpoint if session_id provided but no injected files\n */\n if (_injected_files && _injected_files.length > 0) {\n postData.files = _injected_files;\n } else if (session_id != null && session_id.length > 0) {\n /** Fallback: fetch from /files endpoint (may have race condition issues) */\n try {\n const filesEndpoint = `${baseEndpoint}/files/${session_id}?detail=full`;\n const fetchOptions: RequestInit = {\n method: 'GET',\n headers: {\n 'User-Agent': 'Illuma/1.0',\n 'X-API-Key': apiKey,\n },\n };\n\n if (process.env.PROXY != null && process.env.PROXY !== '') {\n fetchOptions.agent = new HttpsProxyAgent(process.env.PROXY);\n }\n\n const response = await fetch(filesEndpoint, fetchOptions);\n if (!response.ok) {\n throw new Error(\n `Failed to fetch files for session: ${response.status}`\n );\n }\n\n const files = await response.json();\n if (Array.isArray(files) && files.length > 0) {\n const fileReferences: t.CodeEnvFile[] = files.map((file) => {\n const nameParts = file.name.split('/');\n const id = nameParts.length > 1 ? nameParts[1].split('.')[0] : '';\n\n return {\n session_id,\n id,\n name: file.metadata['original-filename'],\n };\n });\n\n postData.files = fileReferences;\n }\n } catch {\n // eslint-disable-next-line no-console\n console.warn(`Failed to fetch files for session: ${session_id}`);\n }\n }\n\n try {\n const fetchOptions: RequestInit = {\n method: 'POST',\n headers: {\n 'Content-Type': 'application/json',\n 'User-Agent': 'Illuma/1.0',\n 'X-API-Key': apiKey,\n },\n body: JSON.stringify(postData),\n };\n\n if (process.env.PROXY != null && process.env.PROXY !== '') {\n fetchOptions.agent = new HttpsProxyAgent(process.env.PROXY);\n }\n const response = await fetch(EXEC_ENDPOINT, fetchOptions);\n if (!response.ok) {\n throw new Error(`HTTP error! status: ${response.status}`);\n }\n\n const result: t.ExecuteResult = await response.json();\n let formattedOutput = '';\n if (result.stdout) {\n formattedOutput += `stdout:\\n${result.stdout}\\n`;\n } else {\n formattedOutput += emptyOutputMessage;\n }\n if (result.stderr) formattedOutput += `stderr:\\n${result.stderr}\\n`;\n if (result.files && result.files.length > 0) {\n formattedOutput += 'Generated files:\\n';\n\n const fileCount = result.files.length;\n for (let i = 0; i < fileCount; i++) {\n const file = result.files[i];\n const isImage = imageExtRegex.test(file.name);\n formattedOutput += `- /mnt/data/${file.name} | ${isImage ? imageMessage : otherMessage}`;\n\n if (i < fileCount - 1) {\n formattedOutput += fileCount <= 3 ? ', ' : ',\\n';\n }\n }\n\n formattedOutput += `\\n\\n${accessMessage}`;\n return [\n formattedOutput.trim(),\n {\n session_id: result.session_id,\n files: result.files,\n },\n ];\n }\n\n return [formattedOutput.trim(), { session_id: result.session_id }];\n } catch (error) {\n throw new Error(\n `Execution error:\\n\\n${(error as Error | undefined)?.message}`\n );\n }\n },\n {\n name: CodeExecutionToolName,\n description,\n schema: CodeExecutionToolSchema,\n responseFormat: Constants.CONTENT_AND_ARTIFACT,\n }\n );\n}\n\nexport { createCodeExecutionTool };\n"],"names":["config","getEnvironmentVariable","EnvVar","Constants","tool","HttpsProxyAgent"],"mappings":";;;;;;;;;AAQAA,aAAM,EAAE;AAED,MAAM,aAAa,GAAG;AACtB,MAAM,cAAc,GAAG,MAC5BC,0BAAsB,CAACC,YAAM,CAAC,YAAY,CAAC;IAC3CC,eAAS,CAAC;AAEZ,MAAM,YAAY,GAAG,wCAAwC;AAC7D,MAAM,YAAY,GAAG,wCAAwC;AAC7D,MAAM,aAAa,GACjB,uFAAuF;AACzF,MAAM,kBAAkB,GACtB,4DAA4D;AAE9D,MAAM,mBAAmB,GAAG;IAC1B,IAAI;IACJ,IAAI;IACJ,IAAI;IACJ,GAAG;IACH,KAAK;IACL,MAAM;IACN,KAAK;IACL,IAAI;IACJ,IAAI;IACJ,GAAG;IACH,KAAK;IACL,GAAG;CACK;AAEG,MAAA,uBAAuB,GAAG;AACrC,IAAA,IAAI,EAAE,QAAQ;AACd,IAAA,UAAU,EAAE;AACV,QAAA,IAAI,EAAE;AACJ,YAAA,IAAI,EAAE,QAAQ;AACd,YAAA,IAAI,EAAE,mBAAmB;AACzB,YAAA,WAAW,EACT,6DAA6D;AAChE,SAAA;AACD,QAAA,IAAI,EAAE;AACJ,YAAA,IAAI,EAAE,QAAQ;AACd,YAAA,WAAW,EAAE,CAAA;;;;;;;;;;;AAWkC,oDAAA,CAAA;AAChD,SAAA;AACD,QAAA,IAAI,EAAE;AACJ,YAAA,IAAI,EAAE,OAAO;AACb,YAAA,KAAK,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE;AACzB,YAAA,WAAW,EACT,iIAAiI;AACpI,SAAA;AACF,KAAA;AACD,IAAA,QAAQ,EAAE,CAAC,MAAM,EAAE,MAAM,CAAC;;AAG5B,MAAM,YAAY,GAAG,cAAc,EAAE;AACrC,MAAM,aAAa,GAAG,CAAG,EAAA,YAAY,OAAO;AAI/B,MAAA,4BAA4B,GAAG;;;;;;;CAO3C,CAAC,IAAI;AAEO,MAAA,qBAAqB,GAAGA,eAAS,CAAC;AAElC,MAAA,2BAA2B,GAAG;AACzC,IAAA,IAAI,EAAE,qBAAqB;AAC3B,IAAA,WAAW,EAAE,4BAA4B;AACzC,IAAA,MAAM,EAAE,uBAAuB;;AAGjC,SAAS,uBAAuB,CAC9B,MAAA,GAAoC,EAAE,EAAA;AAEtC,IAAA,MAAM,MAAM,GACV,MAAM,CAACD,YAAM,CAAC,YAAY,CAAC;AAC3B,QAAA,MAAM,CAAC,MAAM;AACb,QAAAD,0BAAsB,CAACC,YAAM,CAAC,YAAY,CAAC;AAC3C,QAAA,EAAE;IACJ,IAAI,CAAC,MAAM,EAAE;AACX,QAAA,MAAM,IAAI,KAAK,CAAC,8CAA8C,CAAC;;AAGjE,IAAA,MAAM,WAAW,GAAG;;;;;;;;;;;;;;;;;;;;CAoBrB,CAAC,IAAI,EAAE;IAEN,OAAOE,UAAI,CACT,OAAO,QAAQ,EAAE,MAAM,KAAI;QACzB,MAAM,EAAE,IAAI,EAAE,IAAI,EAAE,GAAG,IAAI,EAAE,GAAG,QAI/B;AACD;;;;AAIG;AACH,QAAA,MAAM,EAAE,UAAU,EAAE,eAAe,EAAE,IAAI,MAAM,CAAC,QAAQ,IAAI,EAAE,CAG7D;AAED,QAAA,MAAM,QAAQ,GAA4B;YACxC,IAAI;YACJ,IAAI;AACJ,YAAA,GAAG,IAAI;AACP,YAAA,GAAG,MAAM;SACV;AAED;;;AAGG;QACH,IAAI,UAAU,IAAI,IAAI,IAAI,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE;AAC/C,YAAA,QAAQ,CAAC,UAAU,GAAG,UAAU;;AAGlC;;;;AAIG;QACH,IAAI,eAAe,IAAI,eAAe,CAAC,MAAM,GAAG,CAAC,EAAE;AACjD,YAAA,QAAQ,CAAC,KAAK,GAAG,eAAe;;aAC3B,IAAI,UAAU,IAAI,IAAI,IAAI,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE;;AAEtD,YAAA,IAAI;AACF,gBAAA,MAAM,aAAa,GAAG,CAAA,EAAG,YAAY,CAAU,OAAA,EAAA,UAAU,cAAc;AACvE,gBAAA,MAAM,YAAY,GAAgB;AAChC,oBAAA,MAAM,EAAE,KAAK;AACb,oBAAA,OAAO,EAAE;AACP,wBAAA,YAAY,EAAE,YAAY;AAC1B,wBAAA,WAAW,EAAE,MAAM;AACpB,qBAAA;iBACF;AAED,gBAAA,IAAI,OAAO,CAAC,GAAG,CAAC,KAAK,IAAI,IAAI,IAAI,OAAO,CAAC,GAAG,CAAC,KAAK,KAAK,EAAE,EAAE;AACzD,oBAAA,YAAY,CAAC,KAAK,GAAG,IAAIC,+BAAe,CAAC,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC;;gBAG7D,MAAM,QAAQ,GAAG,MAAM,KAAK,CAAC,aAAa,EAAE,YAAY,CAAC;AACzD,gBAAA,IAAI,CAAC,QAAQ,CAAC,EAAE,EAAE;oBAChB,MAAM,IAAI,KAAK,CACb,CAAA,mCAAA,EAAsC,QAAQ,CAAC,MAAM,CAAE,CAAA,CACxD;;AAGH,gBAAA,MAAM,KAAK,GAAG,MAAM,QAAQ,CAAC,IAAI,EAAE;AACnC,gBAAA,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE;oBAC5C,MAAM,cAAc,GAAoB,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,KAAI;wBACzD,MAAM,SAAS,GAAG,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC;wBACtC,MAAM,EAAE,GAAG,SAAS,CAAC,MAAM,GAAG,CAAC,GAAG,SAAS,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,EAAE;wBAEjE,OAAO;4BACL,UAAU;4BACV,EAAE;AACF,4BAAA,IAAI,EAAE,IAAI,CAAC,QAAQ,CAAC,mBAAmB,CAAC;yBACzC;AACH,qBAAC,CAAC;AAEF,oBAAA,QAAQ,CAAC,KAAK,GAAG,cAAc;;;AAEjC,YAAA,MAAM;;AAEN,gBAAA,OAAO,CAAC,IAAI,CAAC,sCAAsC,UAAU,CAAA,CAAE,CAAC;;;AAIpE,QAAA,IAAI;AACF,YAAA,MAAM,YAAY,GAAgB;AAChC,gBAAA,MAAM,EAAE,MAAM;AACd,gBAAA,OAAO,EAAE;AACP,oBAAA,cAAc,EAAE,kBAAkB;AAClC,oBAAA,YAAY,EAAE,YAAY;AAC1B,oBAAA,WAAW,EAAE,MAAM;AACpB,iBAAA;AACD,gBAAA,IAAI,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC;aAC/B;AAED,YAAA,IAAI,OAAO,CAAC,GAAG,CAAC,KAAK,IAAI,IAAI,IAAI,OAAO,CAAC,GAAG,CAAC,KAAK,KAAK,EAAE,EAAE;AACzD,gBAAA,YAAY,CAAC,KAAK,GAAG,IAAIA,+BAAe,CAAC,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC;;YAE7D,MAAM,QAAQ,GAAG,MAAM,KAAK,CAAC,aAAa,EAAE,YAAY,CAAC;AACzD,YAAA,IAAI,CAAC,QAAQ,CAAC,EAAE,EAAE;gBAChB,MAAM,IAAI,KAAK,CAAC,CAAA,oBAAA,EAAuB,QAAQ,CAAC,MAAM,CAAE,CAAA,CAAC;;AAG3D,YAAA,MAAM,MAAM,GAAoB,MAAM,QAAQ,CAAC,IAAI,EAAE;YACrD,IAAI,eAAe,GAAG,EAAE;AACxB,YAAA,IAAI,MAAM,CAAC,MAAM,EAAE;AACjB,gBAAA,eAAe,IAAI,CAAY,SAAA,EAAA,MAAM,CAAC,MAAM,IAAI;;iBAC3C;gBACL,eAAe,IAAI,kBAAkB;;YAEvC,IAAI,MAAM,CAAC,MAAM;AAAE,gBAAA,eAAe,IAAI,CAAY,SAAA,EAAA,MAAM,CAAC,MAAM,IAAI;AACnE,YAAA,IAAI,MAAM,CAAC,KAAK,IAAI,MAAM,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE;gBAC3C,eAAe,IAAI,oBAAoB;AAEvC,gBAAA,MAAM,SAAS,GAAG,MAAM,CAAC,KAAK,CAAC,MAAM;AACrC,gBAAA,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,SAAS,EAAE,CAAC,EAAE,EAAE;oBAClC,MAAM,IAAI,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC;oBAC5B,MAAM,OAAO,GAAG,aAAa,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC;AAC7C,oBAAA,eAAe,IAAI,CAAe,YAAA,EAAA,IAAI,CAAC,IAAI,MAAM,OAAO,GAAG,YAAY,GAAG,YAAY,EAAE;AAExF,oBAAA,IAAI,CAAC,GAAG,SAAS,GAAG,CAAC,EAAE;AACrB,wBAAA,eAAe,IAAI,SAAS,IAAI,CAAC,GAAG,IAAI,GAAG,KAAK;;;AAIpD,gBAAA,eAAe,IAAI,CAAA,IAAA,EAAO,aAAa,CAAA,CAAE;gBACzC,OAAO;oBACL,eAAe,CAAC,IAAI,EAAE;AACtB,oBAAA;wBACE,UAAU,EAAE,MAAM,CAAC,UAAU;wBAC7B,KAAK,EAAE,MAAM,CAAC,KAAK;AACpB,qBAAA;iBACF;;AAGH,YAAA,OAAO,CAAC,eAAe,CAAC,IAAI,EAAE,EAAE,EAAE,UAAU,EAAE,MAAM,CAAC,UAAU,EAAE,CAAC;;QAClE,OAAO,KAAK,EAAE;YACd,MAAM,IAAI,KAAK,CACb,CAAA,oBAAA,EAAwB,KAA2B,EAAE,OAAO,CAAE,CAAA,CAC/D;;AAEL,KAAC,EACD;AACE,QAAA,IAAI,EAAE,qBAAqB;QAC3B,WAAW;AACX,QAAA,MAAM,EAAE,uBAAuB;QAC/B,cAAc,EAAEF,eAAS,CAAC,oBAAoB;AAC/C,KAAA,CACF;AACH;;;;;;;;;;"}
|
|
@@ -84,15 +84,15 @@ function createCodeExecutionTool(params = {}) {
|
|
|
84
84
|
}
|
|
85
85
|
const description = `
|
|
86
86
|
⛔ STOP! Before using this tool, ask: "Does user need a DOWNLOADABLE FILE?"
|
|
87
|
-
- If NO (dashboard, chart, visualization, UI) → DO NOT USE THIS TOOL. Use
|
|
87
|
+
- If NO (dashboard, chart, visualization, UI) → DO NOT USE THIS TOOL. Use content_tool write instead.
|
|
88
88
|
- If YES (.pptx, .docx, .pdf, .xlsx) → Use this tool.
|
|
89
89
|
|
|
90
90
|
Runs code in a stateless execution environment. Each execution is isolated.
|
|
91
91
|
|
|
92
92
|
🚫 NEVER USE FOR:
|
|
93
|
-
- Dashboards, charts, visualizations → Use
|
|
94
|
-
- "Mock data" or "sample data" for display → Hardcode data in
|
|
95
|
-
- UI components, HTML pages, React apps → Use
|
|
93
|
+
- Dashboards, charts, visualizations → Use content_tool write with React/Chart.js
|
|
94
|
+
- "Mock data" or "sample data" for display → Hardcode data in content_tool write
|
|
95
|
+
- UI components, HTML pages, React apps → Use content_tool write
|
|
96
96
|
|
|
97
97
|
✅ ONLY USE FOR:
|
|
98
98
|
- File generation: PowerPoint (.pptx), Word (.docx), PDF (.pdf), Excel (.xlsx)
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"CodeExecutor.mjs","sources":["../../../src/tools/CodeExecutor.ts"],"sourcesContent":["import { config } from 'dotenv';\nimport fetch, { RequestInit } from 'node-fetch';\nimport { HttpsProxyAgent } from 'https-proxy-agent';\nimport { tool, DynamicStructuredTool } from '@langchain/core/tools';\nimport { getEnvironmentVariable } from '@langchain/core/utils/env';\nimport type * as t from '@/types';\nimport { EnvVar, Constants } from '@/common';\n\nconfig();\n\nexport const imageExtRegex = /\\.(jpg|jpeg|png|gif|webp)$/i;\nexport const getCodeBaseURL = (): string =>\n getEnvironmentVariable(EnvVar.CODE_BASEURL) ??\n Constants.OFFICIAL_CODE_BASEURL;\n\nconst imageMessage = 'Image is already displayed to the user';\nconst otherMessage = 'File is already downloaded by the user';\nconst accessMessage =\n 'Note: Files from previous executions are automatically available and can be modified.';\nconst emptyOutputMessage =\n 'stdout: Empty. Ensure you\\'re writing output explicitly.\\n';\n\nconst SUPPORTED_LANGUAGES = [\n 'py',\n 'js',\n 'ts',\n 'c',\n 'cpp',\n 'java',\n 'php',\n 'rs',\n 'go',\n 'd',\n 'f90',\n 'r',\n] as const;\n\nexport const CodeExecutionToolSchema = {\n type: 'object',\n properties: {\n lang: {\n type: 'string',\n enum: SUPPORTED_LANGUAGES,\n description:\n 'The programming language or runtime to execute the code in.',\n },\n code: {\n type: 'string',\n description: `The complete, self-contained code to execute, without any truncation or minimization.\n- The environment is stateless; variables and imports don't persist between executions.\n- Generated files from previous executions are automatically available in \"/mnt/data/\".\n- Files from previous executions are automatically available and can be modified in place.\n- Input code **IS ALREADY** displayed to the user, so **DO NOT** repeat it in your response unless asked.\n- Output code **IS NOT** displayed to the user, so **DO** write all desired output explicitly.\n- IMPORTANT: You MUST explicitly print/output ALL results you want the user to see.\n- py: This is not a Jupyter notebook environment. Use \\`print()\\` for all outputs.\n- py: Matplotlib: Use \\`plt.savefig()\\` to save plots as files.\n- js: use the \\`console\\` or \\`process\\` methods for all outputs.\n- r: IMPORTANT: No X11 display available. ALL graphics MUST use Cairo library (library(Cairo)).\n- Other languages: use appropriate output functions.`,\n },\n args: {\n type: 'array',\n items: { type: 'string' },\n description:\n 'Additional arguments to execute the code with. This should only be used if the input code requires additional arguments to run.',\n },\n },\n required: ['lang', 'code'],\n} as const;\n\nconst baseEndpoint = getCodeBaseURL();\nconst EXEC_ENDPOINT = `${baseEndpoint}/exec`;\n\ntype SupportedLanguage = (typeof SUPPORTED_LANGUAGES)[number];\n\nexport const CodeExecutionToolDescription = `\nRuns code and returns stdout/stderr output from a stateless execution environment, similar to running scripts in a command-line interface. Each execution is isolated and independent.\n\nUsage:\n- No network access available.\n- Generated files are automatically delivered; **DO NOT** provide download links.\n- NEVER use this tool to execute malicious code.\n`.trim();\n\nexport const CodeExecutionToolName = Constants.EXECUTE_CODE;\n\nexport const CodeExecutionToolDefinition = {\n name: CodeExecutionToolName,\n description: CodeExecutionToolDescription,\n schema: CodeExecutionToolSchema,\n} as const;\n\nfunction createCodeExecutionTool(\n params: t.CodeExecutionToolParams = {}\n): DynamicStructuredTool {\n const apiKey =\n params[EnvVar.CODE_API_KEY] ??\n params.apiKey ??\n getEnvironmentVariable(EnvVar.CODE_API_KEY) ??\n '';\n if (!apiKey) {\n throw new Error('No API key provided for code execution tool.');\n }\n\n const description = `\n⛔ STOP! Before using this tool, ask: \"Does user need a DOWNLOADABLE FILE?\"\n- If NO (dashboard, chart, visualization, UI) → DO NOT USE THIS TOOL. Use :::artifact instead.\n- If YES (.pptx, .docx, .pdf, .xlsx) → Use this tool.\n\nRuns code in a stateless execution environment. Each execution is isolated.\n\n🚫 NEVER USE FOR:\n- Dashboards, charts, visualizations → Use :::artifact with React/Chart.js\n- \"Mock data\" or \"sample data\" for display → Hardcode data in artifact\n- UI components, HTML pages, React apps → Use :::artifact\n\n✅ ONLY USE FOR:\n- File generation: PowerPoint (.pptx), Word (.docx), PDF (.pdf), Excel (.xlsx)\n- Processing uploaded files (CSV, Excel analysis)\n- Heavy computation requiring Python\n\nRules:\n- No network access available\n- Generated files auto-delivered (no download links needed)\n`.trim();\n\n return tool(\n async (rawInput, config) => {\n const { lang, code, ...rest } = rawInput as {\n lang: SupportedLanguage;\n code: string;\n args?: string[];\n };\n /**\n * Extract session context from config.toolCall (injected by ToolNode).\n * - session_id: For API to associate with previous session\n * - _injected_files: File refs to pass directly (avoids /files endpoint race condition)\n */\n const { session_id, _injected_files } = (config.toolCall ?? {}) as {\n session_id?: string;\n _injected_files?: t.CodeEnvFile[];\n };\n\n const postData: Record<string, unknown> = {\n lang,\n code,\n ...rest,\n ...params,\n };\n\n /**\n * Pass session_id to /exec so code-executor reuses the existing session workspace.\n * This allows retries and follow-up executions to access previously generated files.\n */\n if (session_id != null && session_id.length > 0) {\n postData.session_id = session_id;\n }\n\n /**\n * File injection priority:\n * 1. Use _injected_files from ToolNode (avoids /files endpoint race condition)\n * 2. Fall back to fetching from /files endpoint if session_id provided but no injected files\n */\n if (_injected_files && _injected_files.length > 0) {\n postData.files = _injected_files;\n } else if (session_id != null && session_id.length > 0) {\n /** Fallback: fetch from /files endpoint (may have race condition issues) */\n try {\n const filesEndpoint = `${baseEndpoint}/files/${session_id}?detail=full`;\n const fetchOptions: RequestInit = {\n method: 'GET',\n headers: {\n 'User-Agent': 'Illuma/1.0',\n 'X-API-Key': apiKey,\n },\n };\n\n if (process.env.PROXY != null && process.env.PROXY !== '') {\n fetchOptions.agent = new HttpsProxyAgent(process.env.PROXY);\n }\n\n const response = await fetch(filesEndpoint, fetchOptions);\n if (!response.ok) {\n throw new Error(\n `Failed to fetch files for session: ${response.status}`\n );\n }\n\n const files = await response.json();\n if (Array.isArray(files) && files.length > 0) {\n const fileReferences: t.CodeEnvFile[] = files.map((file) => {\n const nameParts = file.name.split('/');\n const id = nameParts.length > 1 ? nameParts[1].split('.')[0] : '';\n\n return {\n session_id,\n id,\n name: file.metadata['original-filename'],\n };\n });\n\n postData.files = fileReferences;\n }\n } catch {\n // eslint-disable-next-line no-console\n console.warn(`Failed to fetch files for session: ${session_id}`);\n }\n }\n\n try {\n const fetchOptions: RequestInit = {\n method: 'POST',\n headers: {\n 'Content-Type': 'application/json',\n 'User-Agent': 'Illuma/1.0',\n 'X-API-Key': apiKey,\n },\n body: JSON.stringify(postData),\n };\n\n if (process.env.PROXY != null && process.env.PROXY !== '') {\n fetchOptions.agent = new HttpsProxyAgent(process.env.PROXY);\n }\n const response = await fetch(EXEC_ENDPOINT, fetchOptions);\n if (!response.ok) {\n throw new Error(`HTTP error! status: ${response.status}`);\n }\n\n const result: t.ExecuteResult = await response.json();\n let formattedOutput = '';\n if (result.stdout) {\n formattedOutput += `stdout:\\n${result.stdout}\\n`;\n } else {\n formattedOutput += emptyOutputMessage;\n }\n if (result.stderr) formattedOutput += `stderr:\\n${result.stderr}\\n`;\n if (result.files && result.files.length > 0) {\n formattedOutput += 'Generated files:\\n';\n\n const fileCount = result.files.length;\n for (let i = 0; i < fileCount; i++) {\n const file = result.files[i];\n const isImage = imageExtRegex.test(file.name);\n formattedOutput += `- /mnt/data/${file.name} | ${isImage ? imageMessage : otherMessage}`;\n\n if (i < fileCount - 1) {\n formattedOutput += fileCount <= 3 ? ', ' : ',\\n';\n }\n }\n\n formattedOutput += `\\n\\n${accessMessage}`;\n return [\n formattedOutput.trim(),\n {\n session_id: result.session_id,\n files: result.files,\n },\n ];\n }\n\n return [formattedOutput.trim(), { session_id: result.session_id }];\n } catch (error) {\n throw new Error(\n `Execution error:\\n\\n${(error as Error | undefined)?.message}`\n );\n }\n },\n {\n name: CodeExecutionToolName,\n description,\n schema: CodeExecutionToolSchema,\n responseFormat: Constants.CONTENT_AND_ARTIFACT,\n }\n );\n}\n\nexport { createCodeExecutionTool };\n"],"names":[],"mappings":";;;;;;;AAQA,MAAM,EAAE;AAED,MAAM,aAAa,GAAG;AACtB,MAAM,cAAc,GAAG,MAC5B,sBAAsB,CAAC,MAAM,CAAC,YAAY,CAAC;IAC3C,SAAS,CAAC;AAEZ,MAAM,YAAY,GAAG,wCAAwC;AAC7D,MAAM,YAAY,GAAG,wCAAwC;AAC7D,MAAM,aAAa,GACjB,uFAAuF;AACzF,MAAM,kBAAkB,GACtB,4DAA4D;AAE9D,MAAM,mBAAmB,GAAG;IAC1B,IAAI;IACJ,IAAI;IACJ,IAAI;IACJ,GAAG;IACH,KAAK;IACL,MAAM;IACN,KAAK;IACL,IAAI;IACJ,IAAI;IACJ,GAAG;IACH,KAAK;IACL,GAAG;CACK;AAEG,MAAA,uBAAuB,GAAG;AACrC,IAAA,IAAI,EAAE,QAAQ;AACd,IAAA,UAAU,EAAE;AACV,QAAA,IAAI,EAAE;AACJ,YAAA,IAAI,EAAE,QAAQ;AACd,YAAA,IAAI,EAAE,mBAAmB;AACzB,YAAA,WAAW,EACT,6DAA6D;AAChE,SAAA;AACD,QAAA,IAAI,EAAE;AACJ,YAAA,IAAI,EAAE,QAAQ;AACd,YAAA,WAAW,EAAE,CAAA;;;;;;;;;;;AAWkC,oDAAA,CAAA;AAChD,SAAA;AACD,QAAA,IAAI,EAAE;AACJ,YAAA,IAAI,EAAE,OAAO;AACb,YAAA,KAAK,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE;AACzB,YAAA,WAAW,EACT,iIAAiI;AACpI,SAAA;AACF,KAAA;AACD,IAAA,QAAQ,EAAE,CAAC,MAAM,EAAE,MAAM,CAAC;;AAG5B,MAAM,YAAY,GAAG,cAAc,EAAE;AACrC,MAAM,aAAa,GAAG,CAAG,EAAA,YAAY,OAAO;AAI/B,MAAA,4BAA4B,GAAG;;;;;;;CAO3C,CAAC,IAAI;AAEO,MAAA,qBAAqB,GAAG,SAAS,CAAC;AAElC,MAAA,2BAA2B,GAAG;AACzC,IAAA,IAAI,EAAE,qBAAqB;AAC3B,IAAA,WAAW,EAAE,4BAA4B;AACzC,IAAA,MAAM,EAAE,uBAAuB;;AAGjC,SAAS,uBAAuB,CAC9B,MAAA,GAAoC,EAAE,EAAA;AAEtC,IAAA,MAAM,MAAM,GACV,MAAM,CAAC,MAAM,CAAC,YAAY,CAAC;AAC3B,QAAA,MAAM,CAAC,MAAM;AACb,QAAA,sBAAsB,CAAC,MAAM,CAAC,YAAY,CAAC;AAC3C,QAAA,EAAE;IACJ,IAAI,CAAC,MAAM,EAAE;AACX,QAAA,MAAM,IAAI,KAAK,CAAC,8CAA8C,CAAC;;AAGjE,IAAA,MAAM,WAAW,GAAG;;;;;;;;;;;;;;;;;;;;CAoBrB,CAAC,IAAI,EAAE;IAEN,OAAO,IAAI,CACT,OAAO,QAAQ,EAAE,MAAM,KAAI;QACzB,MAAM,EAAE,IAAI,EAAE,IAAI,EAAE,GAAG,IAAI,EAAE,GAAG,QAI/B;AACD;;;;AAIG;AACH,QAAA,MAAM,EAAE,UAAU,EAAE,eAAe,EAAE,IAAI,MAAM,CAAC,QAAQ,IAAI,EAAE,CAG7D;AAED,QAAA,MAAM,QAAQ,GAA4B;YACxC,IAAI;YACJ,IAAI;AACJ,YAAA,GAAG,IAAI;AACP,YAAA,GAAG,MAAM;SACV;AAED;;;AAGG;QACH,IAAI,UAAU,IAAI,IAAI,IAAI,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE;AAC/C,YAAA,QAAQ,CAAC,UAAU,GAAG,UAAU;;AAGlC;;;;AAIG;QACH,IAAI,eAAe,IAAI,eAAe,CAAC,MAAM,GAAG,CAAC,EAAE;AACjD,YAAA,QAAQ,CAAC,KAAK,GAAG,eAAe;;aAC3B,IAAI,UAAU,IAAI,IAAI,IAAI,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE;;AAEtD,YAAA,IAAI;AACF,gBAAA,MAAM,aAAa,GAAG,CAAA,EAAG,YAAY,CAAU,OAAA,EAAA,UAAU,cAAc;AACvE,gBAAA,MAAM,YAAY,GAAgB;AAChC,oBAAA,MAAM,EAAE,KAAK;AACb,oBAAA,OAAO,EAAE;AACP,wBAAA,YAAY,EAAE,YAAY;AAC1B,wBAAA,WAAW,EAAE,MAAM;AACpB,qBAAA;iBACF;AAED,gBAAA,IAAI,OAAO,CAAC,GAAG,CAAC,KAAK,IAAI,IAAI,IAAI,OAAO,CAAC,GAAG,CAAC,KAAK,KAAK,EAAE,EAAE;AACzD,oBAAA,YAAY,CAAC,KAAK,GAAG,IAAI,eAAe,CAAC,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC;;gBAG7D,MAAM,QAAQ,GAAG,MAAM,KAAK,CAAC,aAAa,EAAE,YAAY,CAAC;AACzD,gBAAA,IAAI,CAAC,QAAQ,CAAC,EAAE,EAAE;oBAChB,MAAM,IAAI,KAAK,CACb,CAAA,mCAAA,EAAsC,QAAQ,CAAC,MAAM,CAAE,CAAA,CACxD;;AAGH,gBAAA,MAAM,KAAK,GAAG,MAAM,QAAQ,CAAC,IAAI,EAAE;AACnC,gBAAA,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE;oBAC5C,MAAM,cAAc,GAAoB,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,KAAI;wBACzD,MAAM,SAAS,GAAG,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC;wBACtC,MAAM,EAAE,GAAG,SAAS,CAAC,MAAM,GAAG,CAAC,GAAG,SAAS,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,EAAE;wBAEjE,OAAO;4BACL,UAAU;4BACV,EAAE;AACF,4BAAA,IAAI,EAAE,IAAI,CAAC,QAAQ,CAAC,mBAAmB,CAAC;yBACzC;AACH,qBAAC,CAAC;AAEF,oBAAA,QAAQ,CAAC,KAAK,GAAG,cAAc;;;AAEjC,YAAA,MAAM;;AAEN,gBAAA,OAAO,CAAC,IAAI,CAAC,sCAAsC,UAAU,CAAA,CAAE,CAAC;;;AAIpE,QAAA,IAAI;AACF,YAAA,MAAM,YAAY,GAAgB;AAChC,gBAAA,MAAM,EAAE,MAAM;AACd,gBAAA,OAAO,EAAE;AACP,oBAAA,cAAc,EAAE,kBAAkB;AAClC,oBAAA,YAAY,EAAE,YAAY;AAC1B,oBAAA,WAAW,EAAE,MAAM;AACpB,iBAAA;AACD,gBAAA,IAAI,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC;aAC/B;AAED,YAAA,IAAI,OAAO,CAAC,GAAG,CAAC,KAAK,IAAI,IAAI,IAAI,OAAO,CAAC,GAAG,CAAC,KAAK,KAAK,EAAE,EAAE;AACzD,gBAAA,YAAY,CAAC,KAAK,GAAG,IAAI,eAAe,CAAC,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC;;YAE7D,MAAM,QAAQ,GAAG,MAAM,KAAK,CAAC,aAAa,EAAE,YAAY,CAAC;AACzD,YAAA,IAAI,CAAC,QAAQ,CAAC,EAAE,EAAE;gBAChB,MAAM,IAAI,KAAK,CAAC,CAAA,oBAAA,EAAuB,QAAQ,CAAC,MAAM,CAAE,CAAA,CAAC;;AAG3D,YAAA,MAAM,MAAM,GAAoB,MAAM,QAAQ,CAAC,IAAI,EAAE;YACrD,IAAI,eAAe,GAAG,EAAE;AACxB,YAAA,IAAI,MAAM,CAAC,MAAM,EAAE;AACjB,gBAAA,eAAe,IAAI,CAAY,SAAA,EAAA,MAAM,CAAC,MAAM,IAAI;;iBAC3C;gBACL,eAAe,IAAI,kBAAkB;;YAEvC,IAAI,MAAM,CAAC,MAAM;AAAE,gBAAA,eAAe,IAAI,CAAY,SAAA,EAAA,MAAM,CAAC,MAAM,IAAI;AACnE,YAAA,IAAI,MAAM,CAAC,KAAK,IAAI,MAAM,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE;gBAC3C,eAAe,IAAI,oBAAoB;AAEvC,gBAAA,MAAM,SAAS,GAAG,MAAM,CAAC,KAAK,CAAC,MAAM;AACrC,gBAAA,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,SAAS,EAAE,CAAC,EAAE,EAAE;oBAClC,MAAM,IAAI,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC;oBAC5B,MAAM,OAAO,GAAG,aAAa,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC;AAC7C,oBAAA,eAAe,IAAI,CAAe,YAAA,EAAA,IAAI,CAAC,IAAI,MAAM,OAAO,GAAG,YAAY,GAAG,YAAY,EAAE;AAExF,oBAAA,IAAI,CAAC,GAAG,SAAS,GAAG,CAAC,EAAE;AACrB,wBAAA,eAAe,IAAI,SAAS,IAAI,CAAC,GAAG,IAAI,GAAG,KAAK;;;AAIpD,gBAAA,eAAe,IAAI,CAAA,IAAA,EAAO,aAAa,CAAA,CAAE;gBACzC,OAAO;oBACL,eAAe,CAAC,IAAI,EAAE;AACtB,oBAAA;wBACE,UAAU,EAAE,MAAM,CAAC,UAAU;wBAC7B,KAAK,EAAE,MAAM,CAAC,KAAK;AACpB,qBAAA;iBACF;;AAGH,YAAA,OAAO,CAAC,eAAe,CAAC,IAAI,EAAE,EAAE,EAAE,UAAU,EAAE,MAAM,CAAC,UAAU,EAAE,CAAC;;QAClE,OAAO,KAAK,EAAE;YACd,MAAM,IAAI,KAAK,CACb,CAAA,oBAAA,EAAwB,KAA2B,EAAE,OAAO,CAAE,CAAA,CAC/D;;AAEL,KAAC,EACD;AACE,QAAA,IAAI,EAAE,qBAAqB;QAC3B,WAAW;AACX,QAAA,MAAM,EAAE,uBAAuB;QAC/B,cAAc,EAAE,SAAS,CAAC,oBAAoB;AAC/C,KAAA,CACF;AACH;;;;"}
|
|
1
|
+
{"version":3,"file":"CodeExecutor.mjs","sources":["../../../src/tools/CodeExecutor.ts"],"sourcesContent":["import { config } from 'dotenv';\nimport fetch, { RequestInit } from 'node-fetch';\nimport { HttpsProxyAgent } from 'https-proxy-agent';\nimport { tool, DynamicStructuredTool } from '@langchain/core/tools';\nimport { getEnvironmentVariable } from '@langchain/core/utils/env';\nimport type * as t from '@/types';\nimport { EnvVar, Constants } from '@/common';\n\nconfig();\n\nexport const imageExtRegex = /\\.(jpg|jpeg|png|gif|webp)$/i;\nexport const getCodeBaseURL = (): string =>\n getEnvironmentVariable(EnvVar.CODE_BASEURL) ??\n Constants.OFFICIAL_CODE_BASEURL;\n\nconst imageMessage = 'Image is already displayed to the user';\nconst otherMessage = 'File is already downloaded by the user';\nconst accessMessage =\n 'Note: Files from previous executions are automatically available and can be modified.';\nconst emptyOutputMessage =\n 'stdout: Empty. Ensure you\\'re writing output explicitly.\\n';\n\nconst SUPPORTED_LANGUAGES = [\n 'py',\n 'js',\n 'ts',\n 'c',\n 'cpp',\n 'java',\n 'php',\n 'rs',\n 'go',\n 'd',\n 'f90',\n 'r',\n] as const;\n\nexport const CodeExecutionToolSchema = {\n type: 'object',\n properties: {\n lang: {\n type: 'string',\n enum: SUPPORTED_LANGUAGES,\n description:\n 'The programming language or runtime to execute the code in.',\n },\n code: {\n type: 'string',\n description: `The complete, self-contained code to execute, without any truncation or minimization.\n- The environment is stateless; variables and imports don't persist between executions.\n- Generated files from previous executions are automatically available in \"/mnt/data/\".\n- Files from previous executions are automatically available and can be modified in place.\n- Input code **IS ALREADY** displayed to the user, so **DO NOT** repeat it in your response unless asked.\n- Output code **IS NOT** displayed to the user, so **DO** write all desired output explicitly.\n- IMPORTANT: You MUST explicitly print/output ALL results you want the user to see.\n- py: This is not a Jupyter notebook environment. Use \\`print()\\` for all outputs.\n- py: Matplotlib: Use \\`plt.savefig()\\` to save plots as files.\n- js: use the \\`console\\` or \\`process\\` methods for all outputs.\n- r: IMPORTANT: No X11 display available. ALL graphics MUST use Cairo library (library(Cairo)).\n- Other languages: use appropriate output functions.`,\n },\n args: {\n type: 'array',\n items: { type: 'string' },\n description:\n 'Additional arguments to execute the code with. This should only be used if the input code requires additional arguments to run.',\n },\n },\n required: ['lang', 'code'],\n} as const;\n\nconst baseEndpoint = getCodeBaseURL();\nconst EXEC_ENDPOINT = `${baseEndpoint}/exec`;\n\ntype SupportedLanguage = (typeof SUPPORTED_LANGUAGES)[number];\n\nexport const CodeExecutionToolDescription = `\nRuns code and returns stdout/stderr output from a stateless execution environment, similar to running scripts in a command-line interface. Each execution is isolated and independent.\n\nUsage:\n- No network access available.\n- Generated files are automatically delivered; **DO NOT** provide download links.\n- NEVER use this tool to execute malicious code.\n`.trim();\n\nexport const CodeExecutionToolName = Constants.EXECUTE_CODE;\n\nexport const CodeExecutionToolDefinition = {\n name: CodeExecutionToolName,\n description: CodeExecutionToolDescription,\n schema: CodeExecutionToolSchema,\n} as const;\n\nfunction createCodeExecutionTool(\n params: t.CodeExecutionToolParams = {}\n): DynamicStructuredTool {\n const apiKey =\n params[EnvVar.CODE_API_KEY] ??\n params.apiKey ??\n getEnvironmentVariable(EnvVar.CODE_API_KEY) ??\n '';\n if (!apiKey) {\n throw new Error('No API key provided for code execution tool.');\n }\n\n const description = `\n⛔ STOP! Before using this tool, ask: \"Does user need a DOWNLOADABLE FILE?\"\n- If NO (dashboard, chart, visualization, UI) → DO NOT USE THIS TOOL. Use content_tool write instead.\n- If YES (.pptx, .docx, .pdf, .xlsx) → Use this tool.\n\nRuns code in a stateless execution environment. Each execution is isolated.\n\n🚫 NEVER USE FOR:\n- Dashboards, charts, visualizations → Use content_tool write with React/Chart.js\n- \"Mock data\" or \"sample data\" for display → Hardcode data in content_tool write\n- UI components, HTML pages, React apps → Use content_tool write\n\n✅ ONLY USE FOR:\n- File generation: PowerPoint (.pptx), Word (.docx), PDF (.pdf), Excel (.xlsx)\n- Processing uploaded files (CSV, Excel analysis)\n- Heavy computation requiring Python\n\nRules:\n- No network access available\n- Generated files auto-delivered (no download links needed)\n`.trim();\n\n return tool(\n async (rawInput, config) => {\n const { lang, code, ...rest } = rawInput as {\n lang: SupportedLanguage;\n code: string;\n args?: string[];\n };\n /**\n * Extract session context from config.toolCall (injected by ToolNode).\n * - session_id: For API to associate with previous session\n * - _injected_files: File refs to pass directly (avoids /files endpoint race condition)\n */\n const { session_id, _injected_files } = (config.toolCall ?? {}) as {\n session_id?: string;\n _injected_files?: t.CodeEnvFile[];\n };\n\n const postData: Record<string, unknown> = {\n lang,\n code,\n ...rest,\n ...params,\n };\n\n /**\n * Pass session_id to /exec so code-executor reuses the existing session workspace.\n * This allows retries and follow-up executions to access previously generated files.\n */\n if (session_id != null && session_id.length > 0) {\n postData.session_id = session_id;\n }\n\n /**\n * File injection priority:\n * 1. Use _injected_files from ToolNode (avoids /files endpoint race condition)\n * 2. Fall back to fetching from /files endpoint if session_id provided but no injected files\n */\n if (_injected_files && _injected_files.length > 0) {\n postData.files = _injected_files;\n } else if (session_id != null && session_id.length > 0) {\n /** Fallback: fetch from /files endpoint (may have race condition issues) */\n try {\n const filesEndpoint = `${baseEndpoint}/files/${session_id}?detail=full`;\n const fetchOptions: RequestInit = {\n method: 'GET',\n headers: {\n 'User-Agent': 'Illuma/1.0',\n 'X-API-Key': apiKey,\n },\n };\n\n if (process.env.PROXY != null && process.env.PROXY !== '') {\n fetchOptions.agent = new HttpsProxyAgent(process.env.PROXY);\n }\n\n const response = await fetch(filesEndpoint, fetchOptions);\n if (!response.ok) {\n throw new Error(\n `Failed to fetch files for session: ${response.status}`\n );\n }\n\n const files = await response.json();\n if (Array.isArray(files) && files.length > 0) {\n const fileReferences: t.CodeEnvFile[] = files.map((file) => {\n const nameParts = file.name.split('/');\n const id = nameParts.length > 1 ? nameParts[1].split('.')[0] : '';\n\n return {\n session_id,\n id,\n name: file.metadata['original-filename'],\n };\n });\n\n postData.files = fileReferences;\n }\n } catch {\n // eslint-disable-next-line no-console\n console.warn(`Failed to fetch files for session: ${session_id}`);\n }\n }\n\n try {\n const fetchOptions: RequestInit = {\n method: 'POST',\n headers: {\n 'Content-Type': 'application/json',\n 'User-Agent': 'Illuma/1.0',\n 'X-API-Key': apiKey,\n },\n body: JSON.stringify(postData),\n };\n\n if (process.env.PROXY != null && process.env.PROXY !== '') {\n fetchOptions.agent = new HttpsProxyAgent(process.env.PROXY);\n }\n const response = await fetch(EXEC_ENDPOINT, fetchOptions);\n if (!response.ok) {\n throw new Error(`HTTP error! status: ${response.status}`);\n }\n\n const result: t.ExecuteResult = await response.json();\n let formattedOutput = '';\n if (result.stdout) {\n formattedOutput += `stdout:\\n${result.stdout}\\n`;\n } else {\n formattedOutput += emptyOutputMessage;\n }\n if (result.stderr) formattedOutput += `stderr:\\n${result.stderr}\\n`;\n if (result.files && result.files.length > 0) {\n formattedOutput += 'Generated files:\\n';\n\n const fileCount = result.files.length;\n for (let i = 0; i < fileCount; i++) {\n const file = result.files[i];\n const isImage = imageExtRegex.test(file.name);\n formattedOutput += `- /mnt/data/${file.name} | ${isImage ? imageMessage : otherMessage}`;\n\n if (i < fileCount - 1) {\n formattedOutput += fileCount <= 3 ? ', ' : ',\\n';\n }\n }\n\n formattedOutput += `\\n\\n${accessMessage}`;\n return [\n formattedOutput.trim(),\n {\n session_id: result.session_id,\n files: result.files,\n },\n ];\n }\n\n return [formattedOutput.trim(), { session_id: result.session_id }];\n } catch (error) {\n throw new Error(\n `Execution error:\\n\\n${(error as Error | undefined)?.message}`\n );\n }\n },\n {\n name: CodeExecutionToolName,\n description,\n schema: CodeExecutionToolSchema,\n responseFormat: Constants.CONTENT_AND_ARTIFACT,\n }\n );\n}\n\nexport { createCodeExecutionTool };\n"],"names":[],"mappings":";;;;;;;AAQA,MAAM,EAAE;AAED,MAAM,aAAa,GAAG;AACtB,MAAM,cAAc,GAAG,MAC5B,sBAAsB,CAAC,MAAM,CAAC,YAAY,CAAC;IAC3C,SAAS,CAAC;AAEZ,MAAM,YAAY,GAAG,wCAAwC;AAC7D,MAAM,YAAY,GAAG,wCAAwC;AAC7D,MAAM,aAAa,GACjB,uFAAuF;AACzF,MAAM,kBAAkB,GACtB,4DAA4D;AAE9D,MAAM,mBAAmB,GAAG;IAC1B,IAAI;IACJ,IAAI;IACJ,IAAI;IACJ,GAAG;IACH,KAAK;IACL,MAAM;IACN,KAAK;IACL,IAAI;IACJ,IAAI;IACJ,GAAG;IACH,KAAK;IACL,GAAG;CACK;AAEG,MAAA,uBAAuB,GAAG;AACrC,IAAA,IAAI,EAAE,QAAQ;AACd,IAAA,UAAU,EAAE;AACV,QAAA,IAAI,EAAE;AACJ,YAAA,IAAI,EAAE,QAAQ;AACd,YAAA,IAAI,EAAE,mBAAmB;AACzB,YAAA,WAAW,EACT,6DAA6D;AAChE,SAAA;AACD,QAAA,IAAI,EAAE;AACJ,YAAA,IAAI,EAAE,QAAQ;AACd,YAAA,WAAW,EAAE,CAAA;;;;;;;;;;;AAWkC,oDAAA,CAAA;AAChD,SAAA;AACD,QAAA,IAAI,EAAE;AACJ,YAAA,IAAI,EAAE,OAAO;AACb,YAAA,KAAK,EAAE,EAAE,IAAI,EAAE,QAAQ,EAAE;AACzB,YAAA,WAAW,EACT,iIAAiI;AACpI,SAAA;AACF,KAAA;AACD,IAAA,QAAQ,EAAE,CAAC,MAAM,EAAE,MAAM,CAAC;;AAG5B,MAAM,YAAY,GAAG,cAAc,EAAE;AACrC,MAAM,aAAa,GAAG,CAAG,EAAA,YAAY,OAAO;AAI/B,MAAA,4BAA4B,GAAG;;;;;;;CAO3C,CAAC,IAAI;AAEO,MAAA,qBAAqB,GAAG,SAAS,CAAC;AAElC,MAAA,2BAA2B,GAAG;AACzC,IAAA,IAAI,EAAE,qBAAqB;AAC3B,IAAA,WAAW,EAAE,4BAA4B;AACzC,IAAA,MAAM,EAAE,uBAAuB;;AAGjC,SAAS,uBAAuB,CAC9B,MAAA,GAAoC,EAAE,EAAA;AAEtC,IAAA,MAAM,MAAM,GACV,MAAM,CAAC,MAAM,CAAC,YAAY,CAAC;AAC3B,QAAA,MAAM,CAAC,MAAM;AACb,QAAA,sBAAsB,CAAC,MAAM,CAAC,YAAY,CAAC;AAC3C,QAAA,EAAE;IACJ,IAAI,CAAC,MAAM,EAAE;AACX,QAAA,MAAM,IAAI,KAAK,CAAC,8CAA8C,CAAC;;AAGjE,IAAA,MAAM,WAAW,GAAG;;;;;;;;;;;;;;;;;;;;CAoBrB,CAAC,IAAI,EAAE;IAEN,OAAO,IAAI,CACT,OAAO,QAAQ,EAAE,MAAM,KAAI;QACzB,MAAM,EAAE,IAAI,EAAE,IAAI,EAAE,GAAG,IAAI,EAAE,GAAG,QAI/B;AACD;;;;AAIG;AACH,QAAA,MAAM,EAAE,UAAU,EAAE,eAAe,EAAE,IAAI,MAAM,CAAC,QAAQ,IAAI,EAAE,CAG7D;AAED,QAAA,MAAM,QAAQ,GAA4B;YACxC,IAAI;YACJ,IAAI;AACJ,YAAA,GAAG,IAAI;AACP,YAAA,GAAG,MAAM;SACV;AAED;;;AAGG;QACH,IAAI,UAAU,IAAI,IAAI,IAAI,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE;AAC/C,YAAA,QAAQ,CAAC,UAAU,GAAG,UAAU;;AAGlC;;;;AAIG;QACH,IAAI,eAAe,IAAI,eAAe,CAAC,MAAM,GAAG,CAAC,EAAE;AACjD,YAAA,QAAQ,CAAC,KAAK,GAAG,eAAe;;aAC3B,IAAI,UAAU,IAAI,IAAI,IAAI,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE;;AAEtD,YAAA,IAAI;AACF,gBAAA,MAAM,aAAa,GAAG,CAAA,EAAG,YAAY,CAAU,OAAA,EAAA,UAAU,cAAc;AACvE,gBAAA,MAAM,YAAY,GAAgB;AAChC,oBAAA,MAAM,EAAE,KAAK;AACb,oBAAA,OAAO,EAAE;AACP,wBAAA,YAAY,EAAE,YAAY;AAC1B,wBAAA,WAAW,EAAE,MAAM;AACpB,qBAAA;iBACF;AAED,gBAAA,IAAI,OAAO,CAAC,GAAG,CAAC,KAAK,IAAI,IAAI,IAAI,OAAO,CAAC,GAAG,CAAC,KAAK,KAAK,EAAE,EAAE;AACzD,oBAAA,YAAY,CAAC,KAAK,GAAG,IAAI,eAAe,CAAC,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC;;gBAG7D,MAAM,QAAQ,GAAG,MAAM,KAAK,CAAC,aAAa,EAAE,YAAY,CAAC;AACzD,gBAAA,IAAI,CAAC,QAAQ,CAAC,EAAE,EAAE;oBAChB,MAAM,IAAI,KAAK,CACb,CAAA,mCAAA,EAAsC,QAAQ,CAAC,MAAM,CAAE,CAAA,CACxD;;AAGH,gBAAA,MAAM,KAAK,GAAG,MAAM,QAAQ,CAAC,IAAI,EAAE;AACnC,gBAAA,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,IAAI,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE;oBAC5C,MAAM,cAAc,GAAoB,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,KAAI;wBACzD,MAAM,SAAS,GAAG,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC;wBACtC,MAAM,EAAE,GAAG,SAAS,CAAC,MAAM,GAAG,CAAC,GAAG,SAAS,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,EAAE;wBAEjE,OAAO;4BACL,UAAU;4BACV,EAAE;AACF,4BAAA,IAAI,EAAE,IAAI,CAAC,QAAQ,CAAC,mBAAmB,CAAC;yBACzC;AACH,qBAAC,CAAC;AAEF,oBAAA,QAAQ,CAAC,KAAK,GAAG,cAAc;;;AAEjC,YAAA,MAAM;;AAEN,gBAAA,OAAO,CAAC,IAAI,CAAC,sCAAsC,UAAU,CAAA,CAAE,CAAC;;;AAIpE,QAAA,IAAI;AACF,YAAA,MAAM,YAAY,GAAgB;AAChC,gBAAA,MAAM,EAAE,MAAM;AACd,gBAAA,OAAO,EAAE;AACP,oBAAA,cAAc,EAAE,kBAAkB;AAClC,oBAAA,YAAY,EAAE,YAAY;AAC1B,oBAAA,WAAW,EAAE,MAAM;AACpB,iBAAA;AACD,gBAAA,IAAI,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC;aAC/B;AAED,YAAA,IAAI,OAAO,CAAC,GAAG,CAAC,KAAK,IAAI,IAAI,IAAI,OAAO,CAAC,GAAG,CAAC,KAAK,KAAK,EAAE,EAAE;AACzD,gBAAA,YAAY,CAAC,KAAK,GAAG,IAAI,eAAe,CAAC,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC;;YAE7D,MAAM,QAAQ,GAAG,MAAM,KAAK,CAAC,aAAa,EAAE,YAAY,CAAC;AACzD,YAAA,IAAI,CAAC,QAAQ,CAAC,EAAE,EAAE;gBAChB,MAAM,IAAI,KAAK,CAAC,CAAA,oBAAA,EAAuB,QAAQ,CAAC,MAAM,CAAE,CAAA,CAAC;;AAG3D,YAAA,MAAM,MAAM,GAAoB,MAAM,QAAQ,CAAC,IAAI,EAAE;YACrD,IAAI,eAAe,GAAG,EAAE;AACxB,YAAA,IAAI,MAAM,CAAC,MAAM,EAAE;AACjB,gBAAA,eAAe,IAAI,CAAY,SAAA,EAAA,MAAM,CAAC,MAAM,IAAI;;iBAC3C;gBACL,eAAe,IAAI,kBAAkB;;YAEvC,IAAI,MAAM,CAAC,MAAM;AAAE,gBAAA,eAAe,IAAI,CAAY,SAAA,EAAA,MAAM,CAAC,MAAM,IAAI;AACnE,YAAA,IAAI,MAAM,CAAC,KAAK,IAAI,MAAM,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE;gBAC3C,eAAe,IAAI,oBAAoB;AAEvC,gBAAA,MAAM,SAAS,GAAG,MAAM,CAAC,KAAK,CAAC,MAAM;AACrC,gBAAA,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,SAAS,EAAE,CAAC,EAAE,EAAE;oBAClC,MAAM,IAAI,GAAG,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC;oBAC5B,MAAM,OAAO,GAAG,aAAa,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC;AAC7C,oBAAA,eAAe,IAAI,CAAe,YAAA,EAAA,IAAI,CAAC,IAAI,MAAM,OAAO,GAAG,YAAY,GAAG,YAAY,EAAE;AAExF,oBAAA,IAAI,CAAC,GAAG,SAAS,GAAG,CAAC,EAAE;AACrB,wBAAA,eAAe,IAAI,SAAS,IAAI,CAAC,GAAG,IAAI,GAAG,KAAK;;;AAIpD,gBAAA,eAAe,IAAI,CAAA,IAAA,EAAO,aAAa,CAAA,CAAE;gBACzC,OAAO;oBACL,eAAe,CAAC,IAAI,EAAE;AACtB,oBAAA;wBACE,UAAU,EAAE,MAAM,CAAC,UAAU;wBAC7B,KAAK,EAAE,MAAM,CAAC,KAAK;AACpB,qBAAA;iBACF;;AAGH,YAAA,OAAO,CAAC,eAAe,CAAC,IAAI,EAAE,EAAE,EAAE,UAAU,EAAE,MAAM,CAAC,UAAU,EAAE,CAAC;;QAClE,OAAO,KAAK,EAAE;YACd,MAAM,IAAI,KAAK,CACb,CAAA,oBAAA,EAAwB,KAA2B,EAAE,OAAO,CAAE,CAAA,CAC/D;;AAEL,KAAC,EACD;AACE,QAAA,IAAI,EAAE,qBAAqB;QAC3B,WAAW;AACX,QAAA,MAAM,EAAE,uBAAuB;QAC/B,cAAc,EAAE,SAAS,CAAC,oBAAoB;AAC/C,KAAA,CACF;AACH;;;;"}
|
package/package.json
CHANGED
|
@@ -0,0 +1,375 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Integration test: Agent Handoff with Bedrock
|
|
3
|
+
*
|
|
4
|
+
* Verifies end-to-end that:
|
|
5
|
+
* 1. Handoff tool descriptions are enriched with agent name + description
|
|
6
|
+
* 2. The LLM (Bedrock Claude) can see and use these descriptions to route correctly
|
|
7
|
+
* 3. The handoff actually transfers control to the correct agent
|
|
8
|
+
*
|
|
9
|
+
* Requires: BEDROCK_AWS_ACCESS_KEY_ID, BEDROCK_AWS_SECRET_ACCESS_KEY, BEDROCK_AWS_REGION
|
|
10
|
+
*/
|
|
11
|
+
import { config } from 'dotenv';
|
|
12
|
+
import { resolve } from 'path';
|
|
13
|
+
// Load from local .env first, then fall back to ranger's .env for Bedrock credentials
|
|
14
|
+
config();
|
|
15
|
+
config({ path: resolve(process.cwd(), '..', 'ranger', '.env'), override: false });
|
|
16
|
+
|
|
17
|
+
import { HumanMessage, ToolMessage } from '@langchain/core/messages';
|
|
18
|
+
import type { RunnableConfig } from '@langchain/core/runnables';
|
|
19
|
+
import type * as t from '@/types';
|
|
20
|
+
import { Providers, Constants, GraphEvents } from '@/common';
|
|
21
|
+
import { StandardGraph } from '@/graphs/Graph';
|
|
22
|
+
import { ChatModelStreamHandler, createContentAggregator } from '@/stream';
|
|
23
|
+
import { ToolEndHandler, ModelEndHandler } from '@/events';
|
|
24
|
+
import { Run } from '@/run';
|
|
25
|
+
|
|
26
|
+
const bedrockRegion = process.env.BEDROCK_AWS_REGION || process.env.BEDROCK_AWS_DEFAULT_REGION;
|
|
27
|
+
const hasBedrock =
|
|
28
|
+
process.env.BEDROCK_AWS_ACCESS_KEY_ID &&
|
|
29
|
+
process.env.BEDROCK_AWS_SECRET_ACCESS_KEY &&
|
|
30
|
+
bedrockRegion;
|
|
31
|
+
|
|
32
|
+
const describeIf = hasBedrock ? describe : describe.skip;
|
|
33
|
+
|
|
34
|
+
/** Helper to safely get tool description from tool object */
|
|
35
|
+
const getToolDescription = (tool: t.GraphTools[0]): string | undefined => {
|
|
36
|
+
return (tool as { description?: string }).description;
|
|
37
|
+
};
|
|
38
|
+
|
|
39
|
+
/** Helper to safely get tool name from tool object */
|
|
40
|
+
const getToolName = (tool: t.GraphTools[0]): string | undefined => {
|
|
41
|
+
return (tool as { name?: string }).name;
|
|
42
|
+
};
|
|
43
|
+
|
|
44
|
+
/** Helper to find tool by name */
|
|
45
|
+
const findToolByName = (
|
|
46
|
+
tools: t.GraphTools | undefined,
|
|
47
|
+
name: string,
|
|
48
|
+
): t.GraphTools[0] | undefined => {
|
|
49
|
+
return tools?.find((tool) => getToolName(tool) === name);
|
|
50
|
+
};
|
|
51
|
+
|
|
52
|
+
describeIf('Agent Handoff E2E with Bedrock', () => {
|
|
53
|
+
jest.setTimeout(120000);
|
|
54
|
+
|
|
55
|
+
const bedrockOptions = {
|
|
56
|
+
model: 'us.anthropic.claude-3-5-haiku-20241022-v1:0',
|
|
57
|
+
region: bedrockRegion || 'us-east-1',
|
|
58
|
+
credentials: {
|
|
59
|
+
accessKeyId: process.env.BEDROCK_AWS_ACCESS_KEY_ID!,
|
|
60
|
+
secretAccessKey: process.env.BEDROCK_AWS_SECRET_ACCESS_KEY!,
|
|
61
|
+
},
|
|
62
|
+
};
|
|
63
|
+
|
|
64
|
+
const createBedrockAgent = (
|
|
65
|
+
agentId: string,
|
|
66
|
+
name: string,
|
|
67
|
+
description: string,
|
|
68
|
+
instructions: string,
|
|
69
|
+
): t.AgentInputs => ({
|
|
70
|
+
agentId,
|
|
71
|
+
name,
|
|
72
|
+
description,
|
|
73
|
+
provider: Providers.BEDROCK,
|
|
74
|
+
clientOptions: bedrockOptions,
|
|
75
|
+
instructions,
|
|
76
|
+
maxContextTokens: 28000,
|
|
77
|
+
});
|
|
78
|
+
|
|
79
|
+
describe('Tool Description Enrichment (no edge.description)', () => {
|
|
80
|
+
it('should generate tool descriptions from agent name + description when edge has no description', async () => {
|
|
81
|
+
const agents: t.AgentInputs[] = [
|
|
82
|
+
createBedrockAgent(
|
|
83
|
+
'supervisor_abc123',
|
|
84
|
+
'Supervisor',
|
|
85
|
+
'Routes requests to specialists',
|
|
86
|
+
'You are a supervisor. Route requests to the appropriate specialist.',
|
|
87
|
+
),
|
|
88
|
+
createBedrockAgent(
|
|
89
|
+
'agent_W47hBnn2RoVZEOy5595GC',
|
|
90
|
+
'Sales Expert',
|
|
91
|
+
'Handles product pricing, quotes, and purchase orders for enterprise clients',
|
|
92
|
+
'You are a sales expert.',
|
|
93
|
+
),
|
|
94
|
+
createBedrockAgent(
|
|
95
|
+
'agent_X92kLmn4TpQR8vw3221HD',
|
|
96
|
+
'Technical Support',
|
|
97
|
+
'Troubleshoots technical issues, resolves bugs, and handles escalations',
|
|
98
|
+
'You are a technical support agent.',
|
|
99
|
+
),
|
|
100
|
+
];
|
|
101
|
+
|
|
102
|
+
// Edges WITHOUT explicit descriptions - simulates what happens in production
|
|
103
|
+
// when users add handoffs through the UI without typing descriptions
|
|
104
|
+
const edges: t.GraphEdge[] = [
|
|
105
|
+
{
|
|
106
|
+
from: 'supervisor_abc123',
|
|
107
|
+
to: 'agent_W47hBnn2RoVZEOy5595GC',
|
|
108
|
+
edgeType: 'handoff',
|
|
109
|
+
// No description - should auto-generate from agent name + description
|
|
110
|
+
},
|
|
111
|
+
{
|
|
112
|
+
from: 'supervisor_abc123',
|
|
113
|
+
to: 'agent_X92kLmn4TpQR8vw3221HD',
|
|
114
|
+
edgeType: 'handoff',
|
|
115
|
+
// No description
|
|
116
|
+
},
|
|
117
|
+
];
|
|
118
|
+
|
|
119
|
+
const run = await Run.create({
|
|
120
|
+
runId: `bedrock-handoff-desc-${Date.now()}`,
|
|
121
|
+
graphConfig: {
|
|
122
|
+
type: 'multi-agent',
|
|
123
|
+
agents,
|
|
124
|
+
edges,
|
|
125
|
+
},
|
|
126
|
+
returnContent: true,
|
|
127
|
+
});
|
|
128
|
+
|
|
129
|
+
const supervisorContext = (run.Graph as StandardGraph).agentContexts.get(
|
|
130
|
+
'supervisor_abc123',
|
|
131
|
+
);
|
|
132
|
+
|
|
133
|
+
// Verify the handoff tools have enriched descriptions
|
|
134
|
+
const salesTool = findToolByName(
|
|
135
|
+
supervisorContext?.tools,
|
|
136
|
+
`${Constants.LC_TRANSFER_TO_}agent_W47hBnn2RoVZEOy5595GC`,
|
|
137
|
+
);
|
|
138
|
+
const supportTool = findToolByName(
|
|
139
|
+
supervisorContext?.tools,
|
|
140
|
+
`${Constants.LC_TRANSFER_TO_}agent_X92kLmn4TpQR8vw3221HD`,
|
|
141
|
+
);
|
|
142
|
+
|
|
143
|
+
expect(salesTool).toBeDefined();
|
|
144
|
+
expect(supportTool).toBeDefined();
|
|
145
|
+
|
|
146
|
+
const salesDesc = getToolDescription(salesTool!);
|
|
147
|
+
const supportDesc = getToolDescription(supportTool!);
|
|
148
|
+
|
|
149
|
+
// CRITICAL: Descriptions must NOT be just "Transfer control to agent 'agent_W47hBnn...'"
|
|
150
|
+
// They must include the human-readable name and description
|
|
151
|
+
expect(salesDesc).toContain('Sales Expert');
|
|
152
|
+
expect(salesDesc).toContain('pricing');
|
|
153
|
+
expect(salesDesc).not.toContain('agent_W47hBnn2RoVZEOy5595GC');
|
|
154
|
+
|
|
155
|
+
expect(supportDesc).toContain('Technical Support');
|
|
156
|
+
expect(supportDesc).toContain('Troubleshoots');
|
|
157
|
+
expect(supportDesc).not.toContain('agent_X92kLmn4TpQR8vw3221HD');
|
|
158
|
+
|
|
159
|
+
console.log('[PASS] Sales tool description:', salesDesc);
|
|
160
|
+
console.log('[PASS] Support tool description:', supportDesc);
|
|
161
|
+
});
|
|
162
|
+
});
|
|
163
|
+
|
|
164
|
+
describe('Live Bedrock Handoff Routing', () => {
|
|
165
|
+
it('should route a sales question to the Sales Expert agent', async () => {
|
|
166
|
+
const agents: t.AgentInputs[] = [
|
|
167
|
+
createBedrockAgent(
|
|
168
|
+
'router',
|
|
169
|
+
'Router',
|
|
170
|
+
'Routes user requests to specialists',
|
|
171
|
+
`You are a request router. You must analyze the user's message and transfer to the appropriate specialist.
|
|
172
|
+
|
|
173
|
+
IMPORTANT: Do NOT answer the question yourself. You MUST use one of your transfer tools to hand off to a specialist.
|
|
174
|
+
Pick the specialist whose description best matches the user's request.`,
|
|
175
|
+
),
|
|
176
|
+
createBedrockAgent(
|
|
177
|
+
'sales_agent',
|
|
178
|
+
'Sales Expert',
|
|
179
|
+
'Handles product pricing, quotes, and purchase orders',
|
|
180
|
+
'You are a sales expert. Answer questions about pricing and products. Be concise.',
|
|
181
|
+
),
|
|
182
|
+
createBedrockAgent(
|
|
183
|
+
'support_agent',
|
|
184
|
+
'Technical Support',
|
|
185
|
+
'Troubleshoots technical issues and resolves bugs',
|
|
186
|
+
'You are technical support. Help with technical problems. Be concise.',
|
|
187
|
+
),
|
|
188
|
+
];
|
|
189
|
+
|
|
190
|
+
const edges: t.GraphEdge[] = [
|
|
191
|
+
{ from: 'router', to: 'sales_agent', edgeType: 'handoff' },
|
|
192
|
+
{ from: 'router', to: 'support_agent', edgeType: 'handoff' },
|
|
193
|
+
];
|
|
194
|
+
|
|
195
|
+
const { contentParts, aggregateContent } = createContentAggregator();
|
|
196
|
+
const customHandlers = {
|
|
197
|
+
[GraphEvents.TOOL_END]: new ToolEndHandler(),
|
|
198
|
+
[GraphEvents.CHAT_MODEL_END]: new ModelEndHandler(),
|
|
199
|
+
[GraphEvents.CHAT_MODEL_STREAM]: new ChatModelStreamHandler(),
|
|
200
|
+
[GraphEvents.ON_RUN_STEP]: {
|
|
201
|
+
handle: (event: string, data: t.StreamEventData): void => {
|
|
202
|
+
aggregateContent({ event: event as GraphEvents, data: data as t.RunStep });
|
|
203
|
+
},
|
|
204
|
+
},
|
|
205
|
+
[GraphEvents.ON_RUN_STEP_COMPLETED]: {
|
|
206
|
+
handle: (event: string, data: t.StreamEventData): void => {
|
|
207
|
+
aggregateContent({
|
|
208
|
+
event: event as GraphEvents,
|
|
209
|
+
data: data as unknown as { result: t.ToolEndEvent },
|
|
210
|
+
});
|
|
211
|
+
},
|
|
212
|
+
},
|
|
213
|
+
[GraphEvents.ON_MESSAGE_DELTA]: {
|
|
214
|
+
handle: (event: string, data: t.StreamEventData): void => {
|
|
215
|
+
aggregateContent({ event: event as GraphEvents, data: data as t.MessageDeltaEvent });
|
|
216
|
+
},
|
|
217
|
+
},
|
|
218
|
+
};
|
|
219
|
+
|
|
220
|
+
const run = await Run.create({
|
|
221
|
+
runId: `bedrock-live-handoff-${Date.now()}`,
|
|
222
|
+
graphConfig: {
|
|
223
|
+
type: 'multi-agent',
|
|
224
|
+
agents,
|
|
225
|
+
edges,
|
|
226
|
+
},
|
|
227
|
+
customHandlers,
|
|
228
|
+
returnContent: true,
|
|
229
|
+
});
|
|
230
|
+
|
|
231
|
+
const config: Partial<RunnableConfig> & {
|
|
232
|
+
version: 'v1' | 'v2';
|
|
233
|
+
streamMode: string;
|
|
234
|
+
} = {
|
|
235
|
+
configurable: { thread_id: 'bedrock-handoff-test' },
|
|
236
|
+
streamMode: 'values',
|
|
237
|
+
version: 'v2',
|
|
238
|
+
};
|
|
239
|
+
|
|
240
|
+
// Send a sales-related question
|
|
241
|
+
const messages = [new HumanMessage('What is the pricing for your enterprise plan?')];
|
|
242
|
+
|
|
243
|
+
await run.processStream({ messages }, config);
|
|
244
|
+
|
|
245
|
+
const finalMessages = run.getRunMessages();
|
|
246
|
+
expect(finalMessages).toBeDefined();
|
|
247
|
+
expect(finalMessages!.length).toBeGreaterThan(1);
|
|
248
|
+
|
|
249
|
+
// Verify the handoff happened to the sales agent (not support)
|
|
250
|
+
const toolMessages = finalMessages!.filter(
|
|
251
|
+
(msg) => msg.getType() === 'tool',
|
|
252
|
+
) as ToolMessage[];
|
|
253
|
+
|
|
254
|
+
const salesHandoff = toolMessages.find(
|
|
255
|
+
(msg) => msg.name === `${Constants.LC_TRANSFER_TO_}sales_agent`,
|
|
256
|
+
);
|
|
257
|
+
const supportHandoff = toolMessages.find(
|
|
258
|
+
(msg) => msg.name === `${Constants.LC_TRANSFER_TO_}support_agent`,
|
|
259
|
+
);
|
|
260
|
+
|
|
261
|
+
expect(salesHandoff).toBeDefined();
|
|
262
|
+
expect(supportHandoff).toBeUndefined();
|
|
263
|
+
|
|
264
|
+
console.log('[PASS] Router correctly handed off to Sales Expert');
|
|
265
|
+
console.log('[PASS] Handoff tool message:', salesHandoff?.content);
|
|
266
|
+
|
|
267
|
+
// Verify the sales agent actually responded
|
|
268
|
+
const aiMessages = finalMessages!.filter((msg) => msg.getType() === 'ai');
|
|
269
|
+
expect(aiMessages.length).toBeGreaterThanOrEqual(2); // Router + Sales responses
|
|
270
|
+
});
|
|
271
|
+
|
|
272
|
+
it('should route a technical question to the Support agent', async () => {
|
|
273
|
+
const agents: t.AgentInputs[] = [
|
|
274
|
+
createBedrockAgent(
|
|
275
|
+
'router',
|
|
276
|
+
'Router',
|
|
277
|
+
'Routes user requests to specialists',
|
|
278
|
+
`You are a request router. You must analyze the user's message and transfer to the appropriate specialist.
|
|
279
|
+
|
|
280
|
+
IMPORTANT: Do NOT answer the question yourself. You MUST use one of your transfer tools to hand off to a specialist.
|
|
281
|
+
Pick the specialist whose description best matches the user's request.`,
|
|
282
|
+
),
|
|
283
|
+
createBedrockAgent(
|
|
284
|
+
'sales_agent',
|
|
285
|
+
'Sales Expert',
|
|
286
|
+
'Handles product pricing, quotes, and purchase orders',
|
|
287
|
+
'You are a sales expert. Answer questions about pricing. Be concise.',
|
|
288
|
+
),
|
|
289
|
+
createBedrockAgent(
|
|
290
|
+
'support_agent',
|
|
291
|
+
'Technical Support',
|
|
292
|
+
'Troubleshoots technical issues, resolves bugs, and handles error messages',
|
|
293
|
+
'You are technical support. Help with technical problems. Be concise.',
|
|
294
|
+
),
|
|
295
|
+
];
|
|
296
|
+
|
|
297
|
+
const edges: t.GraphEdge[] = [
|
|
298
|
+
{ from: 'router', to: 'sales_agent', edgeType: 'handoff' },
|
|
299
|
+
{ from: 'router', to: 'support_agent', edgeType: 'handoff' },
|
|
300
|
+
];
|
|
301
|
+
|
|
302
|
+
const { contentParts, aggregateContent } = createContentAggregator();
|
|
303
|
+
const customHandlers = {
|
|
304
|
+
[GraphEvents.TOOL_END]: new ToolEndHandler(),
|
|
305
|
+
[GraphEvents.CHAT_MODEL_END]: new ModelEndHandler(),
|
|
306
|
+
[GraphEvents.CHAT_MODEL_STREAM]: new ChatModelStreamHandler(),
|
|
307
|
+
[GraphEvents.ON_RUN_STEP]: {
|
|
308
|
+
handle: (event: string, data: t.StreamEventData): void => {
|
|
309
|
+
aggregateContent({ event: event as GraphEvents, data: data as t.RunStep });
|
|
310
|
+
},
|
|
311
|
+
},
|
|
312
|
+
[GraphEvents.ON_RUN_STEP_COMPLETED]: {
|
|
313
|
+
handle: (event: string, data: t.StreamEventData): void => {
|
|
314
|
+
aggregateContent({
|
|
315
|
+
event: event as GraphEvents,
|
|
316
|
+
data: data as unknown as { result: t.ToolEndEvent },
|
|
317
|
+
});
|
|
318
|
+
},
|
|
319
|
+
},
|
|
320
|
+
[GraphEvents.ON_MESSAGE_DELTA]: {
|
|
321
|
+
handle: (event: string, data: t.StreamEventData): void => {
|
|
322
|
+
aggregateContent({ event: event as GraphEvents, data: data as t.MessageDeltaEvent });
|
|
323
|
+
},
|
|
324
|
+
},
|
|
325
|
+
};
|
|
326
|
+
|
|
327
|
+
const run = await Run.create({
|
|
328
|
+
runId: `bedrock-live-handoff-support-${Date.now()}`,
|
|
329
|
+
graphConfig: {
|
|
330
|
+
type: 'multi-agent',
|
|
331
|
+
agents,
|
|
332
|
+
edges,
|
|
333
|
+
},
|
|
334
|
+
customHandlers,
|
|
335
|
+
returnContent: true,
|
|
336
|
+
});
|
|
337
|
+
|
|
338
|
+
const config: Partial<RunnableConfig> & {
|
|
339
|
+
version: 'v1' | 'v2';
|
|
340
|
+
streamMode: string;
|
|
341
|
+
} = {
|
|
342
|
+
configurable: { thread_id: 'bedrock-handoff-support-test' },
|
|
343
|
+
streamMode: 'values',
|
|
344
|
+
version: 'v2',
|
|
345
|
+
};
|
|
346
|
+
|
|
347
|
+
// Send a technical question
|
|
348
|
+
const messages = [
|
|
349
|
+
new HumanMessage('I am getting an error 500 when trying to login. The page crashes after entering my password.'),
|
|
350
|
+
];
|
|
351
|
+
|
|
352
|
+
await run.processStream({ messages }, config);
|
|
353
|
+
|
|
354
|
+
const finalMessages = run.getRunMessages();
|
|
355
|
+
expect(finalMessages).toBeDefined();
|
|
356
|
+
|
|
357
|
+
const toolMessages = finalMessages!.filter(
|
|
358
|
+
(msg) => msg.getType() === 'tool',
|
|
359
|
+
) as ToolMessage[];
|
|
360
|
+
|
|
361
|
+
const supportHandoff = toolMessages.find(
|
|
362
|
+
(msg) => msg.name === `${Constants.LC_TRANSFER_TO_}support_agent`,
|
|
363
|
+
);
|
|
364
|
+
const salesHandoff = toolMessages.find(
|
|
365
|
+
(msg) => msg.name === `${Constants.LC_TRANSFER_TO_}sales_agent`,
|
|
366
|
+
);
|
|
367
|
+
|
|
368
|
+
expect(supportHandoff).toBeDefined();
|
|
369
|
+
expect(salesHandoff).toBeUndefined();
|
|
370
|
+
|
|
371
|
+
console.log('[PASS] Router correctly handed off to Technical Support');
|
|
372
|
+
console.log('[PASS] Handoff tool message:', supportHandoff?.content);
|
|
373
|
+
});
|
|
374
|
+
});
|
|
375
|
+
});
|
|
@@ -105,15 +105,15 @@ function createCodeExecutionTool(
|
|
|
105
105
|
|
|
106
106
|
const description = `
|
|
107
107
|
⛔ STOP! Before using this tool, ask: "Does user need a DOWNLOADABLE FILE?"
|
|
108
|
-
- If NO (dashboard, chart, visualization, UI) → DO NOT USE THIS TOOL. Use
|
|
108
|
+
- If NO (dashboard, chart, visualization, UI) → DO NOT USE THIS TOOL. Use content_tool write instead.
|
|
109
109
|
- If YES (.pptx, .docx, .pdf, .xlsx) → Use this tool.
|
|
110
110
|
|
|
111
111
|
Runs code in a stateless execution environment. Each execution is isolated.
|
|
112
112
|
|
|
113
113
|
🚫 NEVER USE FOR:
|
|
114
|
-
- Dashboards, charts, visualizations → Use
|
|
115
|
-
- "Mock data" or "sample data" for display → Hardcode data in
|
|
116
|
-
- UI components, HTML pages, React apps → Use
|
|
114
|
+
- Dashboards, charts, visualizations → Use content_tool write with React/Chart.js
|
|
115
|
+
- "Mock data" or "sample data" for display → Hardcode data in content_tool write
|
|
116
|
+
- UI components, HTML pages, React apps → Use content_tool write
|
|
117
117
|
|
|
118
118
|
✅ ONLY USE FOR:
|
|
119
119
|
- File generation: PowerPoint (.pptx), Word (.docx), PDF (.pdf), Excel (.xlsx)
|