* feat: add design subagent * refactor: extract shared subagent prompt constants * refactor: localize explorer reminder constant --------- Co-authored-by: nicoalbanese <49612682+nicoalbanese@users.noreply.github.com>
102 lines
2.9 KiB
TypeScript
102 lines
2.9 KiB
TypeScript
import type { LanguageModel } from "ai";
|
|
import { gateway, stepCountIs, ToolLoopAgent } from "ai";
|
|
import { z } from "zod";
|
|
import { bashTool } from "../tools/bash";
|
|
import { globTool } from "../tools/glob";
|
|
import { grepTool } from "../tools/grep";
|
|
import { readFileTool } from "../tools/read";
|
|
import { editFileTool, writeFileTool } from "../tools/write";
|
|
import type { SandboxExecutionContext } from "../types";
|
|
import {
|
|
SUBAGENT_BASH_RULES,
|
|
SUBAGENT_COMPLETE_TASK_RULES,
|
|
SUBAGENT_NO_QUESTIONS_RULES,
|
|
SUBAGENT_REMINDER,
|
|
SUBAGENT_RESPONSE_FORMAT,
|
|
SUBAGENT_STEP_LIMIT,
|
|
SUBAGENT_VALIDATE_RULES,
|
|
SUBAGENT_WORKING_DIR,
|
|
} from "./constants";
|
|
|
|
const EXECUTOR_SYSTEM_PROMPT = `You are an executor agent - a fire-and-forget subagent that completes specific, well-defined implementation tasks autonomously.
|
|
|
|
Think of yourself as a productive engineer who cannot ask follow-up questions once started.
|
|
|
|
## CRITICAL RULES
|
|
|
|
${SUBAGENT_NO_QUESTIONS_RULES}
|
|
|
|
${SUBAGENT_COMPLETE_TASK_RULES}
|
|
|
|
${SUBAGENT_RESPONSE_FORMAT}
|
|
|
|
Example final response:
|
|
---
|
|
**Summary**: I created the new user authentication module with JWT validation. I added the auth middleware, updated the routes, and created unit tests.
|
|
|
|
**Answer**: The authentication system is now implemented:
|
|
- \`src/middleware/auth.ts\` - JWT validation middleware
|
|
- \`src/routes/auth.ts\` - Login/logout endpoints
|
|
- \`src/tests/auth.test.ts\` - Unit tests (all passing)
|
|
---
|
|
|
|
${SUBAGENT_VALIDATE_RULES}
|
|
|
|
## TOOLS
|
|
You have full access to file operations (read, write, edit, grep, glob) and bash commands. Use them to complete your task.
|
|
|
|
${SUBAGENT_BASH_RULES}`;
|
|
|
|
const callOptionsSchema = z.object({
|
|
task: z.string().describe("Short description of the task"),
|
|
instructions: z.string().describe("Detailed instructions for the task"),
|
|
sandbox: z
|
|
.custom<SandboxExecutionContext["sandbox"]>()
|
|
.describe("Sandbox for file system and shell operations"),
|
|
model: z.custom<LanguageModel>().describe("Language model for this subagent"),
|
|
});
|
|
|
|
export type ExecutorCallOptions = z.infer<typeof callOptionsSchema>;
|
|
|
|
export const executorSubagent = new ToolLoopAgent({
|
|
model: gateway("anthropic/claude-haiku-4.5"),
|
|
instructions: EXECUTOR_SYSTEM_PROMPT,
|
|
tools: {
|
|
read: readFileTool(),
|
|
write: writeFileTool(),
|
|
edit: editFileTool(),
|
|
grep: grepTool(),
|
|
glob: globTool(),
|
|
bash: bashTool(),
|
|
},
|
|
stopWhen: stepCountIs(SUBAGENT_STEP_LIMIT),
|
|
callOptionsSchema,
|
|
prepareCall: ({ options, ...settings }) => {
|
|
if (!options) {
|
|
throw new Error("Executor subagent requires task call options.");
|
|
}
|
|
|
|
const sandbox = options.sandbox;
|
|
const model = options.model ?? settings.model;
|
|
return {
|
|
...settings,
|
|
model,
|
|
instructions: `${EXECUTOR_SYSTEM_PROMPT}
|
|
|
|
${SUBAGENT_WORKING_DIR}
|
|
|
|
## Your Task
|
|
${options.task}
|
|
|
|
## Detailed Instructions
|
|
${options.instructions}
|
|
|
|
${SUBAGENT_REMINDER}`,
|
|
experimental_context: {
|
|
sandbox,
|
|
model,
|
|
},
|
|
};
|
|
},
|
|
});
|