mirror of
https://github.com/langchain-ai/deepagentsjs.git
synced 2026-07-22 12:25:36 -04:00
Hunter Lovell
454fa26804
* feat(quickjs): add @langchain/quickjs — sandboxed JS/TS REPL with PTC, env isolation, and RLM support * chore: split changesets into quickjs minor + deepagents patch * refactor(quickjs): remove env/secrets, lazy runtime, buffered writes, fix thread_id propagation - Remove entire env/secrets system (EnvVarConfig, EnvConfig, secretRef, injectEnv, blockEnvLeaks, checkEnvAccess, generateEnvPrompt) - Lazy-load QuickJS WASM runtime on first eval via ensureStarted() - Buffer file writes in pendingWrites[], flush after eval via flushWrites(backend) - Make ReplSession.getOrCreate() synchronous (runtime init deferred) - Add toJSON()/fromJSON() for session serialization - Fix thread_id propagation: move session creation from wrapModelCall (which lacks configurable) to tool handler (which has config.configurable.thread_id) - Remove dead starting map from ReplSession - Add integration test verifying REPL state persists across js_eval calls - Add thread_id propagation unit tests - Remove secret-env-agent example - 68 unit tests + 1 integration test passing, types clean * cr * cr * docs(quickjs): add README and LICENSE, fix lint errors * cr * cr
51 lines
1.6 KiB
TypeScript
51 lines
1.6 KiB
TypeScript
/**
|
|
* Data Analysis Agent Example
|
|
*
|
|
* Demonstrates the QuickJS REPL as a computational scratch pad.
|
|
* The agent reads data from the VFS, processes it in sandboxed JavaScript,
|
|
* and writes results back — all without network access or Node.js APIs.
|
|
*
|
|
* This is useful for tasks where LLMs typically hallucinate:
|
|
* - Arithmetic and statistical calculations
|
|
* - Sorting, filtering, grouping data
|
|
* - JSON transformation and restructuring
|
|
* - Multi-step logic with intermediate state
|
|
*/
|
|
import "dotenv/config";
|
|
import dedent from "dedent";
|
|
import { HumanMessage } from "@langchain/core/messages";
|
|
import { ChatAnthropic } from "@langchain/anthropic";
|
|
import { createDeepAgent } from "deepagents";
|
|
import { createQuickJSMiddleware } from "@langchain/quickjs";
|
|
|
|
const model = new ChatAnthropic({
|
|
model: "claude-sonnet-4-5",
|
|
temperature: 0,
|
|
});
|
|
|
|
const agent = createDeepAgent({
|
|
model,
|
|
systemPrompt: dedent`
|
|
You are a data analyst. Use the js_eval REPL to perform calculations
|
|
and data transformations. Always show your work in code — never guess
|
|
at arithmetic or statistics.
|
|
`,
|
|
middleware: [createQuickJSMiddleware()],
|
|
});
|
|
|
|
const result = await agent.invoke({
|
|
messages: [
|
|
new HumanMessage(dedent`
|
|
I have sales data in /data/sales.json. Parse it, calculate the total
|
|
revenue per region, find the top-performing region, and write a
|
|
summary report to /reports/sales-summary.md.
|
|
`),
|
|
],
|
|
});
|
|
|
|
const last = result.messages[result.messages.length - 1];
|
|
// eslint-disable-next-line no-console
|
|
console.log(
|
|
typeof last.content === "string" ? last.content.slice(0, 500) : last.content,
|
|
);
|