@loopstack/code-agent

A module for the Loopstack AI automation framework.

AI-powered code exploration for Loopstack workflows. Bundles a sub-agent workflow that searches, reads, and synthesises findings from a remote workspace, plus a tool that lets any parent workflow delegate exploration as a single call.

Overview

The Code Agent module is the go-to primitive when a workflow needs to “figure something out” about a codebase before acting — for example, locating a function, summarising a feature, or answering a free-form question about existing code. The sub-agent runs its own plan-then-tool-call loop against glob, grep, and read from @loopstack/remote-client, and returns a plain-text response.

By using this module you’ll get:

ExploreAgentWorkflow — a self-contained Claude-powered agent that iteratively explores a remote workspace and returns a synthesised answer
ExploreTask — a tool that launches ExploreAgentWorkflow as a sub-workflow, so a parent workflow can delegate exploration with one call

Installation


npm install @loopstack/code-agent


import { CodeAgentModule } from '@loopstack/code-agent';
 
@Module({
  imports: [CodeAgentModule /* ... */],
})
export class AppModule {}

CodeAgentModule depends on ClaudeModule, LoopCoreModule, and RemoteClientModule — make sure those are configured in your app (typically done in the project scaffold).

How It Works

Using `ExploreTask` from a parent workflow

Inject the tool via the constructor and call it with a natural-language instruction. The tool returns the workflowId of the launched sub-agent and resolves with the synthesised text response when the sub-agent completes:


import { ExploreTask } from '@loopstack/code-agent';
import { BaseWorkflow, InjectTool, Transition, Workflow } from '@loopstack/common';
 
@Workflow({ uiConfig: __dirname + '/my.ui.yaml' })
export class MyWorkflow extends BaseWorkflow {
  constructor(private readonly explore: ExploreTask) {
    super();
  }
 
  @Transition({ from: 'ready', to: 'done' })
  async investigate() {
    const result = await this.explore.call({
      instructions: 'Find the NestJS module where user authentication is wired up.',
    });
    // result.data.response contains the agent's summary
  }
}

What the sub-agent does

ExploreAgentWorkflow runs a Claude model with three tools exposed: glob, grep, and read. On each turn it either calls a tool (and transitions to awaiting_tools) or produces a final answer (@Final). The agent’s system prompt instructs it to prefer grep over reading whole files — see src/workflows/explore-agent.workflow.ts for the prompt.

Running the workflow directly

You can also run ExploreAgentWorkflow standalone without going through the tool:


import { ExploreAgentWorkflow } from '@loopstack/code-agent';
 
// Inject and call .run({ instructions: '...' }) from any service or workflow.

Public API

Workflow: ExploreAgentWorkflow
Tool: ExploreTask
Module: CodeAgentModule

Dependencies

@loopstack/common — core decorators and BaseTool / BaseWorkflow
@loopstack/core — LoopCoreModule
@loopstack/llm-provider-module — LlmGenerateTextTool, LlmDelegateToolCallsTool, LlmMessageDocument
@loopstack/remote-client — GlobTool, GrepTool, ReadTool (executed on the remote workspace)

About

Author: Jakob Klippel

License: MIT

Additional Resources

Loopstack Documentation
Find more Loopstack modules in the Loopstack Registry