zesdex/src/features/agent/application/turn_service.test.ts

import { describe, expect, it } from "bun:test";
import {
  adaptiveMaxTokens,
  conversationChars,
  ErrorTracker,
  truncateToolOutput,
  bashExitCode,
  isToolFailure,
  deriveVerifyCommand,
  AgentTurnServiceImpl,
  takeRecentTail,
  reduceMessagesToDigest,
  compactMessagesWithAi,
} from "./turn_service.ts";
import {
  type ChatMessage,
  newConversation,
  systemMessage,
  userMessage,
  assistantMessage,
  toolResultMessage,
} from "@zesdex/domain";
import type { AgentTurnParams } from "@zesdex/agent";
import type { ToolExecutor } from "./index.ts";
import type { ProviderService } from "./ports.ts";

describe("truncateToolOutput", () => {
  it("short output is unchanged", () => {
    expect(truncateToolOutput("short")).toBe("short");
  });

  it("long output preserves head and marks cut", () => {
    const long = "x".repeat(12_000 + 500);
    const truncated = truncateToolOutput(long);
    expect(truncated.length).toBeLessThan(long.length);
    expect(truncated).toContain("...[truncated");
    expect(truncated.startsWith("xxx")).toBe(true);
  });
});

describe("adaptiveMaxTokens", () => {
  it("scales with request length", () => {
    expect(adaptiveMaxTokens(10)).toBe(800);
    expect(adaptiveMaxTokens(200)).toBe(1600);
    expect(adaptiveMaxTokens(5000)).toBe(4096);
  });
});

describe("ErrorTracker", () => {
  it("injects recovery note after repeated errors", () => {
    const tracker = new ErrorTracker();
    const messages: ChatMessage[] = [];
    tracker.record("read", "Error: File not found", messages);
    tracker.record("read", "Error: File not found", messages);
    expect(tracker.shouldStop()).toBe(false);
    tracker.record("read", "Error: File not found", messages);
    expect(messages.some((m) => m.content?.includes("[System note]"))).toBe(true);
  });

  it("stops after too many errors", () => {
    const tracker = new ErrorTracker();
    const messages: ChatMessage[] = [];
    for (let i = 0; i < 8; i++) {
      tracker.record("bash", `Error: boom ${i}`, messages);
    }
    expect(tracker.shouldStop()).toBe(true);
  });
});

describe("conversationChars", () => {
  it("sums content only", () => {
    const conv = newConversation("sys", "s1");
    conv.messages.push(systemMessage("sys"), userMessage("hello world"), toolResultMessage("id", "output"));
    expect(conversationChars(conv.messages)).toBe(3 + 11 + 6);
  });
});
/* ── New flow helpers ─────────────────────────────────────────────── */

describe("bashExitCode / isToolFailure", () => {
  it("extracts a non-zero exit code", () => {
    expect(bashExitCode("boom\n\nExit code: 1 (1s)")).toBe(1);
    expect(bashExitCode("done\n\nExit code: 0 (0.5s)")).toBe(0);
  });
  it("returns null when no exit-code line exists", () => {
    expect(bashExitCode("plain output")).toBeNull();
  });
  it("isToolFailure flags bash non-zero exits as failures", () => {
    expect(isToolFailure("bash", "nope\n\nExit code: 2 (1s)")).toBe(true);
    expect(isToolFailure("bash", "ok\n\nExit code: 0 (1s)")).toBe(false);
  });
  it("isToolFailure still flags Error: prefixes for other tools", () => {
    expect(isToolFailure("read", "Error: no such file")).toBe(true);
  });
});

describe("deriveVerifyCommand", () => {
  it("defaults to bun check+test for a Bun repo", () => {
    // No real manifests in a control dir we guarantee to not exist.
    expect(deriveVerifyCommand("/nonexistent-zesdex-dir")).toContain("bun");
  });
});

/* ── runTurn flow tests (fakes, no network) ───────────────────────── */

interface ScriptedStep {
  content?: string | null;
  tools?: Array<{ name: string; args?: string; id?: string }>;
}

/** A ProviderService that replays scripted chatStream responses. */
function makeScriptedProvider(script: ScriptedStep[]): {
  provider: ProviderService;
  seen: Array<ChatMessage[]>;
  chatCalls: Array<ChatMessage[]>;
} {
  const seen: Array<ChatMessage[]> = [];
  const chatCalls: Array<ChatMessage[]> = [];
  let step = 0;
  const chatResponses: string[] = [];
  const provider: ProviderService = {
    async chat(messages, _tools, _maxTokens, _temperature) {
      // Used by compaction and review. For review tests we script below.
      chatCalls.push([...messages]);
      const text = chatResponses.shift() ?? "";
      return { message: { role: "assistant", content: text || null }, usage: [10, 5] };
    },
    async chatStream(messages, _tools, _max, _temp, onEvent) {
      // Snapshot a copy: runTurn freely mutates the live array (splice/compact).
      seen.push([...messages]);
      const s = script[Math.min(step, script.length - 1)] ?? { content: "done" };
      step += 1;
      if (s.content !== undefined && s.content !== null) onEvent({ kind: "token", content: s.content });
      if (s.tools && s.tools.length > 0) {
        const msg = assistantMessage(null) as ChatMessage;
        msg.tool_calls = s.tools.map((t, i) => ({
          id: t.id ?? `call-${i}`,
          type: "function",
          function: { name: t.name, arguments: t.args ?? "{}" },
        }));
        return { message: msg, usage: [10, 2] };
      }
      return {
        message: { role: "assistant", content: s.content ?? null },
        usage: [10, 2],
      };
    },
    // expose a way for tests to script chat replies
  } as ProviderService;
  (provider as unknown as { setChatReply: (s: string) => void }).setChatReply = (text: string) => {
    chatResponses.push(text);
  };
  return { provider, seen, chatCalls };
}

/** A ToolExecutor that echoes fixed outputs per tool. */
function fixedExecutor(outs: Record<string, string>): ToolExecutor {
  return {
    async execute(name) {
      return outs[name] ?? "ok";
    },
    isParallelSafe() {
      return false;
    },
  };
}

function buildParams(userText: string): {
  params: AgentTurnParams;
  events: Array<any>;
} {
  const events: Array<any> = [];
  const params: AgentTurnParams = {
    messages: [userMessage(userText)],
    session_dir: "/tmp/zesdex-flow-test",
    workspace_roots: ["/nonexistent-zesdex-dir"], // no AGENTS.md/package.json side effects
    turn_events: { push: (e) => events.push(e), drain: () => [] },
    in_flight: { value: false },
    abort: new AbortController(),
    api_key: "k",
    model: "m",
    api_base: "https://x",
  };
  return { params, events };
}

describe("runTurn", () => {
  it("runs a simple one-shot prompt to completion", async () => {
    const { provider } = makeScriptedProvider([{ content: "here is the answer" }]);
    const svc = new AgentTurnServiceImpl(provider as never, fixedExecutor({}), [] as never);
    const { params } = buildParams("hello");
    await svc.runTurn(params);
    expect(params.in_flight.value).toBe(false);
  });
});

describe("runTurn verify-after-edit", () => {
  it("injects a verify nudge after a write, cleared after a successful bash", async () => {
    const { provider, seen } = makeScriptedProvider([
      { tools: [{ name: "write" }] },
      { tools: [{ name: "bash" }] },
      { content: "done" },
    ]);
    const executor = fixedExecutor({ write: "wrote it", bash: "ok\n\nExit code: 0 (1s)" });
    const svc = new AgentTurnServiceImpl(provider as never, executor, [] as never);
    const { params } = buildParams("add a comment");
    await svc.runTurn(params);
    // The LLM call after the write should carry the verify nudge.
    const second = seen[1] ?? [];
    expect(second.some((m) => (m.content ?? "").includes("Run the verify command"))).toBe(true);
    // The successful bash exits 0, which clears pendingVerify — so the nudge is
    // injected exactly once, not repeated on every later call.
    const finalBatch = seen[seen.length - 1] ?? [];
    const nudgeCount = finalBatch.filter((m) => (m.content ?? "").includes("Run the verify command")).length;
    expect(nudgeCount).toBe(1);
  });
});

describe("runTurn convergence guard", () => {
  it("stops after repeated identical read calls", async () => {
    const script: ScriptedStep[] = [];
    for (let i = 0; i < 6; i++) script.push({ tools: [{ name: "read" }] });
    const { provider, seen } = makeScriptedProvider(script);
    const executor = fixedExecutor({ read: "same content" });
    const svc = new AgentTurnServiceImpl(provider as never, executor, [] as never);
    const { params, events } = buildParams("read something");
    await svc.runTurn(params);
    const warned = events.some((e) => e.kind === "system_note" && /without any progress|repeated/.test(e.message ?? ""));
    expect(warned).toBe(true);
    // Not every call got issued — the guard broke early.
    expect(seen.length).toBeLessThan(script.length);
  });
});

describe("runTurn self-review", () => {
  it("feeds a reviewer critique back as a system message after a mutator", async () => {
    const { provider, seen } = makeScriptedProvider([
      { tools: [{ name: "edit" }] },
      { content: "fixed" },
    ]);
    const withChat = provider as unknown as { setChatReply(s: string): void };
    withChat.setChatReply("- [PRIORITY: high] handle empty input in parse()");
    const executor = fixedExecutor({ edit: "edited" });
    const svc = new AgentTurnServiceImpl(provider as never, executor, [] as never);
    const { params, events } = buildParams("fix parse()");
    await svc.runTurn(params);
    // The reviewer critique must have reached the next LLM call's history.
    const second = seen[1] ?? [];
    expect(second.some((m) => (m.content ?? "").includes("[Reviewer]"))).toBe(true);
    // review_usage emitted.
    expect(events.some((e) => e.kind === "review_usage")).toBe(true);
  });

  it("passes the work-in-progress to the reviewer (not an empty slate)", async () => {
    const { provider, chatCalls } = makeScriptedProvider([
      { tools: [{ name: "edit" }] },
      { content: "fixed" },
    ]);
    (provider as unknown as { setChatReply(s: string): void }).setChatReply(
      "- [PRIORITY: medium] add a null check",
    );
    const svc = new AgentTurnServiceImpl(provider as never, fixedExecutor({ edit: "changed" }), [] as never);
    const { params } = buildParams("harden the parser");
    await svc.runTurn(params);
    // The REVIEWER chat call carries the digest in its user message.
    const reviewerCall = chatCalls.find((c) =>
      c.some((m) => (m.content ?? "").includes("Review the work-in-progress below")),
    );
    expect(reviewerCall).toBeTruthy();
    expect(reviewerCall!.some((m) => (m.content ?? "").includes("harden the parser"))).toBe(true);
  });
});

/* ── Compaction: efficient + accurate ─────────────────────────────── */

describe("takeRecentTail", () => {
  it("keeps the most recent messages and evicts an oversized older blob", () => {
    const messages = [
      toolResultMessage("t0", "y".repeat(5000)), // huge old tool output
      userMessage("old request"),
      assistantMessage("recent reply A"),
      userMessage("recent reply B"),
    ];
    const { tail, evicted } = takeRecentTail(messages, 25, 2);
    expect(evicted.length).toBe(2); // the blob + the old request
    expect(tail.map((m) => m.content)).toEqual(["recent reply A", "recent reply B"]);
  });

  it("gives the whole history as tail when it is small enough", () => {
    const messages = [userMessage("a"), assistantMessage("b")];
    const { tail, evicted } = takeRecentTail(messages, 1000, 6);
    expect(evicted).toEqual([]);
    expect(tail.length).toBe(2);
  });

  it("does not let a huge tool output blow the tail past the min count", () => {
    // minTail 2: newest two kept regardless; the huge tool body stays evicted.
    const messages = [
      toolResultMessage("t0", "x".repeat(50000)),
      userMessage("keep1"),
      assistantMessage("keep2"),
    ];
    const { tail, evicted } = takeRecentTail(messages, 8000, 2);
    expect(tail.map((m) => m.content)).toEqual(["keep1", "keep2"]);
    expect(evicted.length).toBe(1);
  });
});

describe("reduceMessagesToDigest", () => {
  it("stubs tool bodies entirely and previews text", () => {
    const digest = reduceMessagesToDigest([
      toolResultMessage("t0", "y".repeat(5000)),
      userMessage("short request"),
    ]);
    expect(digest).toContain("- tool");
    expect(digest).not.toContain("yyyy");
    expect(digest).toContain("short request");
  });

  it("keeps user previews well above the shorter assistant cap", () => {
    const long = "w".repeat(2000);
    // User budget is 1500 chars — larger than the 400-char assistant cap, so
    // a long user request survives far more of its body than a long assistant
    // reply would (requests matter most for summary accuracy).
    const digestUser = reduceMessagesToDigest([userMessage(long)]);
    const digestAssistant = reduceMessagesToDigest([assistantMessage(long)]);
    expect(digestUser.length).toBeGreaterThan(800);
    expect(digestAssistant.length).toBeLessThan(500);
  });

  it("respects the hard input cap", () => {
    const many = [];
    for (let i = 0; i < 200; i++) many.push(userMessage("r".repeat(300)));
    const digest = reduceMessagesToDigest(many);
    expect(digest.length).toBeLessThanOrEqual(20_000 + 40);
  });
});

describe("compactMessagesWithAi", () => {
  it("keeps the recent tail and prepends an AI summary", async () => {
    // Long enough to bypass the min-size guard, with an oversized OLD tool blob
    // (should be evicted + stubbed, not kept verbatim).
    const messages: ChatMessage[] = [
      userMessage("old request one"),
      assistantMessage("old response"),
      toolResultMessage("t0", "big".repeat(4000)),
    ];
    for (let i = 0; i < 7; i++) messages.push(userMessage(`mid ${i}`));
    messages.push(toolResultMessage("t1", "tail-result"));
    messages.push(userMessage("newest request"));
    messages.push(assistantMessage("newest response"));

    const provider: ProviderService = {
      async chat() {
        return { message: { role: "assistant", content: "## Requests\n- old request one" }, usage: null };
      },
      async chatStream() {
        return { message: { role: "assistant", content: null }, usage: null };
      },
    };
    await compactMessagesWithAi(messages, provider);
    // Summary first, recent tail preserved verbatim, tool blobs gone.
    expect(messages[0]?.content).toContain("[AI Summary of Previous Conversation]");
    expect(messages.some((m) => m.content === "newest request")).toBe(true);
    expect(messages.some((m) => m.content === "newest response")).toBe(true);
    expect(messages.some((m) => (m.content ?? "").includes("big".repeat(10)))).toBe(false);
  });

  it("degrades gracefully on summarizer failure, preserving the tail", async () => {
    const messages = [
      userMessage("old"),
      toolResultMessage("t0", "x".repeat(100)),
      userMessage("recent request"),
    ];
    const failing: ProviderService = {
      async chat() {
        throw new Error("network down");
      },
      async chatStream() {
        return { message: { role: "assistant", content: null }, usage: null };
      },
    };
    await compactMessagesWithAi(messages, failing);
    // The recent message survives even when the LLM call fails.
    expect(messages.some((m) => m.content === "recent request")).toBe(true);
  });
});