agent-r/test/orchestrator.test.ts

import { mkdtempSync, readFileSync } from "node:fs";
import { tmpdir } from "node:os";
import path from "node:path";
import { describe, expect, test } from "vitest";
import { ArtifactManager } from "../src/artifacts.js";
import { loadConfig } from "../src/config.js";
import { AgentROrchestrator } from "../src/orchestrator.js";
import { RunStore } from "../src/store.js";
import type { AgentInvocation, AgentInvocationResult, RawAgentRunner } from "../src/types.js";

class ScriptedRunner implements RawAgentRunner {
  public readonly prompts: string[] = [];

  constructor(
    private readonly outputs: string[],
  ) {}

  async invoke<T>(request: AgentInvocation<T>): Promise<AgentInvocationResult<T>> {
    const next = this.outputs.shift();
    if (!next) {
      throw new Error("No scripted output remaining.");
    }

    this.prompts.push(request.prompt);
    readFileSync(request.artifacts.schemaPath, "utf8");
    const parsed = JSON.parse(next) as Record<string, unknown>;
    if (typeof parsed.taskId === "string" && !parsed.taskId) {
      const match = request.prompt.match(/exact task id `([^`]+)`/);
      if (match) {
        parsed.taskId = match[1];
      }
    }
    return {
      sessionId: request.sessionId ?? `session-${this.prompts.length}`,
      output: parsed as T,
      rawMessage: JSON.stringify(parsed),
      rawEvents: ['{"type":"turn.completed"}'],
      stderr: "",
      artifacts: request.artifacts,
    };
  }
}

describe("AgentROrchestrator", () => {
  test("runs through plan, implementation, self-check, and independent check", async () => {
    const repoPath = mkdtempSync(path.join(tmpdir(), "agent-r-orchestrator-"));
    const config = loadConfig(repoPath);
    const store = new RunStore(config.databasePath);
    const artifacts = new ArtifactManager(config.runsDir);
    const runner = new ScriptedRunner([
      JSON.stringify({
        decision: "continue",
        summary: "Implement the CLI skeleton.",
        rationale: "The project is empty and needs a first vertical slice.",
        goalProgress: "No implementation exists yet.",
        risks: [],
        tasks: [
          {
            title: "Create CLI entrypoint",
            objective: "Build the initial command surface.",
            acceptanceCriteria: ["A run command exists."],
            verificationSteps: ["Run the help command."],
            allowedPaths: ["src", "package.json"],
          },
        ],
      }),
      JSON.stringify({
        taskId: "",
        status: "completed",
        summary: "Added the CLI skeleton.",
        changes: ["Created an entrypoint."],
        verification: [{ command: "agent-r --help", outcome: "passed", details: "Help output rendered." }],
        followUps: [],
        touchedFiles: ["src/index.ts", "package.json"],
        blockers: [],
      }),
      JSON.stringify({
        decision: "done",
        summary: "The requested slice is complete.",
        rationale: "The only planned task is complete and verified.",
        goalProgress: "The vertical slice exists.",
        risks: [],
        tasks: [],
      }),
      JSON.stringify({
        readyForIndependentCheck: true,
        summary: "The run is ready for independent review.",
        rationale: "Implementation and verification are present.",
        evidence: ["CLI entrypoint exists."],
        remainingGaps: [],
      }),
      JSON.stringify({
        verdict: "approved",
        summary: "The goal is satisfied.",
        rationale: "The requested slice exists and is verified.",
        evidence: ["CLI entrypoint present."],
        remainingTasks: [],
      }),
    ]);

    const orchestrator = new AgentROrchestrator(config, store, artifacts, runner);
    const run = orchestrator.createRun("Create a CLI skeleton");

    const firstStatus = await orchestrator.runUntilStable(run.id, 10);
    const snapshot = store.buildSnapshot(run.id);

    expect(firstStatus.status).toBe("done");
    expect(snapshot.completedTasks).toHaveLength(1);
    expect(snapshot.approvals).toHaveLength(2);
    expect(snapshot.pendingTasks).toHaveLength(0);
    expect(runner.prompts.at(1)).toContain("You must echo the exact task id");
  });

  test("strategy prompt explicitly permits read-only repository inspection", async () => {
    const repoPath = mkdtempSync(path.join(tmpdir(), "agent-r-strategy-"));
    const config = loadConfig(repoPath);
    const store = new RunStore(config.databasePath);
    const artifacts = new ArtifactManager(config.runsDir);
    const runner = new ScriptedRunner([
      JSON.stringify({
        decision: "blocked",
        summary: "Cannot continue.",
        rationale: "Test stop.",
        goalProgress: "None.",
        risks: [],
        tasks: [],
        blockedReason: "Stop immediately.",
      }),
    ]);

    const orchestrator = new AgentROrchestrator(config, store, artifacts, runner);
    const run = orchestrator.createRun("Inspect repo");
    await orchestrator.runUntilStable(run.id, 1);

    expect(runner.prompts[0]).toContain("You may inspect the repository directly");
    expect(runner.prompts[0]).toContain("Do not edit files");
  });
});