Skip to content

armgit / qm

Public Imported from GitHub
0

Multiplayer agent harness for work.

23.2 KB
Raw Blame
import { test } from "node:test";
import assert from "node:assert/strict";
import type { ToolActivity, WorkBlock } from "../src/core-bridge.ts";
import {
  messageWorkTimeline,
  currentTextPhase,
  streamedAnswer,
  streamingTextTail,
  buildTimeline,
  workTimelineSegments,
  toolCategory,
  toolRowKind,
  toolExecutionOutput,
  type ToolRowModel,
} from "../src/timeline.ts";

function act(seq: number, type: ToolActivity["type"], payload: unknown): ToolActivity {
  return { seq, parentSeq: null, type, payload, createdAt: seq };
}

function row(call: unknown, result?: unknown): ToolRowModel {
  return { call: act(1, "tool_call", call), result: result === undefined ? null : act(2, "tool_result", result) };
}

test("thinking, tool calls, and results stay in seq order — never bucketed by type", () => {
  const work: WorkBlock = {
    status: "complete",
    activity: [
      act(1, "thinking", { thinking: "first I'll look" }),
      act(2, "tool_call", { tool: "execute", command: "ls" }),
      act(3, "tool_result", { tool: "execute", code: 0 }),
      act(4, "thinking", { thinking: "now read it" }),
      act(5, "tool_call", { tool: "read", path: "/x" }),
      act(6, "tool_result", { tool: "read" }),
    ],
  };
  const items = buildTimeline(work);
  assert.deepEqual(
    items.map((i) => i.kind),
    ["thinking", "tool", "thinking", "tool"],
    "two thinking blocks interleave between the two tool rows, in arrival order",
  );
  assert.equal(
    items[0].kind === "thinking" && (items[0].activity.payload as { thinking: string }).thinking,
    "first I'll look",
  );
  assert.ok(
    items[1].kind === "tool" && items[1].row.call && items[1].row.result,
    "the first tool row pairs its call with its result",
  );
});

test("a paused command's approval marker attaches to its blocked tool row", () => {
  const work: WorkBlock = {
    status: "complete",
    activity: [
      act(1, "tool_call", { tool: "execute", command: "rm -rf x" }),
      act(2, "tool_result", { tool: "execute", blocked: "needs_approval", reason: "recursive delete" }),
    ],
    pendingApprovals: [{ requestId: "r1", command: "rm -rf x", reason: "recursive delete" }],
  };
  const items = buildTimeline(work);
  assert.equal(items.length, 1, "no separate approval item — the marker rides the blocked tool row");
  assert.ok(items[0].kind === "tool" && items[0].row.pending?.requestId === "r1");
});

test("a command approval's plain-English summary survives onto the timeline item for the renderer", () => {
  const work: WorkBlock = {
    status: "complete",
    activity: [
      act(1, "tool_call", { tool: "execute", command: "rm -rf x" }),
      act(2, "tool_result", { tool: "execute", blocked: "needs_approval", reason: "recursive delete" }),
    ],
    pendingApprovals: [
      {
        requestId: "r1",
        command: "rm -rf x",
        reason: "recursive delete",
        summary: "Deletes the x/ directory and everything inside it.",
      },
    ],
  };
  const items = buildTimeline(work);
  assert.ok(
    items[0].kind === "tool" && items[0].row.pending?.summary === "Deletes the x/ directory and everything inside it.",
  );
  assert.ok(items[0].kind === "tool" && items[0].row.pending?.reason === "recursive delete");
});

test("an approval whose tool_call isn't in the loaded activity lands at the end", () => {
  const work: WorkBlock = {
    status: "complete",
    activity: [
      act(1, "tool_call", { tool: "execute", command: "ls" }),
      act(2, "tool_result", { tool: "execute", code: 0 }),
    ],
    pendingApprovals: [{ requestId: "r9", command: "deploy prod", reason: "ship it" }],
  };
  const items = buildTimeline(work);
  assert.equal(items.length, 2, "the tool row plus a trailing fallback approval");
  assert.equal(items[1].kind, "approval");
  assert.ok(items[1].kind === "approval" && items[1].approval.requestId === "r9");
});

test("two distinct paused commands each get their own approval marker", () => {
  const work: WorkBlock = {
    status: "complete",
    activity: [
      act(1, "tool_call", { tool: "execute", command: "git clone a" }),
      act(2, "tool_result", { tool: "execute", blocked: "needs_approval", reason: "clone" }),
      act(3, "tool_call", { tool: "execute", command: "git clone b" }),
      act(4, "tool_result", { tool: "execute", blocked: "needs_approval", reason: "clone" }),
    ],
    pendingApprovals: [
      { requestId: "ra", command: "git clone a" },
      { requestId: "rb", command: "git clone b" },
    ],
  };
  const items = buildTimeline(work);
  assert.equal(items.length, 2, "two tool rows, each carrying its own marker — none left over at the end");
  assert.ok(items[0].kind === "tool" && items[0].row.pending?.requestId === "ra");
  assert.ok(items[1].kind === "tool" && items[1].row.pending?.requestId === "rb");
});

test("on a finished turn, retried identical orphan calls fold into one row with a count", () => {
  const work: WorkBlock = {
    status: "failed",
    activity: [
      act(1, "tool_call", { tool: "execute", command: "flaky" }),
      act(2, "tool_call", { tool: "execute", command: "flaky" }),
      act(3, "tool_call", { tool: "execute", command: "flaky" }),
    ],
  };
  const items = buildTimeline(work);
  assert.equal(items.length, 1);
  assert.ok(items[0].kind === "tool" && items[0].row.attempts === 3);
});

test("interleaved thinking breaks an orphan-call run so the rows aren't fused", () => {
  const work: WorkBlock = {
    status: "failed",
    activity: [
      act(1, "tool_call", { tool: "execute", command: "flaky" }),
      act(2, "thinking", { thinking: "hmm" }),
      act(3, "tool_call", { tool: "execute", command: "flaky" }),
    ],
  };
  const items = buildTimeline(work);
  assert.deepEqual(
    items.map((i) => i.kind),
    ["tool", "thinking", "tool"],
    "the thinking between the two identical calls keeps them as separate rows",
  );
});

test("interstitial narration (text) interleaves with thinking and tools by seq", () => {
  const work: WorkBlock = {
    status: "complete",
    activity: [
      act(1, "thinking", { thinking: "plan" }),
      act(2, "text", { text: "Let me check the config first." }),
      act(3, "tool_call", { tool: "read", path: "/cfg" }),
      act(4, "tool_result", { tool: "read" }),
    ],
  };
  const items = buildTimeline(work);
  assert.deepEqual(
    items.map((i) => i.kind),
    ["thinking", "text", "tool"],
    "narration sits between the thinking and the tool it precedes",
  );
  assert.ok(
    items[1].kind === "text" && (items[1].activity.payload as { text: string }).text.startsWith("Let me check"),
  );
});

test("results pair to their call by callId across batched (parallel) tool calls", () => {
  const work: WorkBlock = {
    status: "complete",
    activity: [
      act(3, "tool_call", { tool: "execute", command: "echo first", callId: "A" }),
      act(4, "tool_call", { tool: "execute", command: "echo second", callId: "B" }),
      act(5, "tool_call", { tool: "execute", command: "echo third", callId: "C" }),
      act(6, "tool_result", { tool: "execute", callId: "B", code: 0 }),
      act(9, "tool_call", { tool: "execute", command: "echo first", callId: "D" }),
      act(10, "tool_call", { tool: "execute", command: "echo third", callId: "E" }),
      act(11, "tool_result", { tool: "execute", callId: "D", code: 0 }),
      act(12, "tool_result", { tool: "execute", callId: "E", code: 0 }),
    ],
  };
  const items = buildTimeline(work);
  assert.equal(items.length, 5, "five tool rows, no stray orphan-result row");
  assert.ok(items.every((i) => i.kind === "tool"));
  const rows = items.map((i) => (i.kind === "tool" ? i.row : null));
  assert.ok(rows[0]!.call && !rows[0]!.result, "echo first (batch 1) stays resultless");
  assert.ok(
    rows[1]!.result && (rows[1]!.result.payload as { callId: string }).callId === "B",
    "echo second got result B, not C's",
  );
  assert.ok(rows[2]!.call && !rows[2]!.result, "echo third (batch 1) stays resultless");
  assert.ok(
    rows[3]!.result && (rows[3]!.result.payload as { callId: string }).callId === "D",
    "retried echo first got result D",
  );
  assert.ok(
    rows[4]!.result && (rows[4]!.result.payload as { callId: string }).callId === "E",
    "retried echo third got result E",
  );
});

test("toolRowKind: a successful result is `ok`", () => {
  assert.equal(
    toolRowKind(row({ tool: "publish", name: "site" }, { tool: "publish", url: "https://x" }), "complete"),
    "ok",
  );
  assert.equal(toolRowKind(row({ tool: "execute", command: "ls" }, { tool: "execute", code: 0 }), "complete"), "ok");
});

test("toolRowKind: an error result is `failed`, regardless of which tool", () => {
  assert.equal(
    toolRowKind(
      row({ tool: "publish", name: ":bad" }, { tool: "publish", error: "invalid deployment name" }),
      "complete",
    ),
    "failed",
  );
  assert.equal(
    toolRowKind(row({ tool: "write", path: "/x" }, { tool: "write", error: "EACCES" }), "complete"),
    "failed",
  );
});

test("toolRowKind: a policy-denied command is `failed`, and a nonzero exit is `ok`", () => {
  assert.equal(
    toolRowKind(
      row({ tool: "execute", command: "curl evil" }, { tool: "execute", denied: true, reason: "blocked host" }),
      "complete",
    ),
    "failed",
  );
  assert.equal(toolRowKind(row({ tool: "execute", command: "false" }, { tool: "execute", code: 1 }), "complete"), "ok");
});

test("toolRowKind: a needs-approval result is `approval`", () => {
  assert.equal(
    toolRowKind(
      row(
        { tool: "execute", command: "rm -rf x" },
        { tool: "execute", blocked: "needs_approval", reason: "recursive delete" },
      ),
      "working",
    ),
    "approval",
  );
});

test("toolRowKind: no result yet — `running` mid-turn, `attempted`/`failed` once the turn ends", () => {
  assert.equal(toolRowKind(row({ tool: "publish", name: "site" }), "working"), "running", "mid-turn, still in flight");
  assert.equal(
    toolRowKind(row({ tool: "publish", name: "site" }), "complete"),
    "attempted",
    "turn finished cleanly but we never heard back — orphan, not a failure",
  );
  assert.equal(
    toolRowKind(row({ tool: "publish", name: "site" }), "failed"),
    "failed",
    "turn itself failed, so the in-flight call counts as failed",
  );
});

function workWith(activity: ToolActivity[], status: WorkBlock["status"] = "working"): WorkBlock {
  return { status, activity };
}

test("buildTimeline is memoized: unchanged (activity, status, approvals) returns the identical array", () => {
  const work = workWith([act(1, "tool_call", { tool: "execute", command: "ls", callId: "a" })]);
  const first = buildTimeline(work);
  assert.equal(buildTimeline(work), first, "same inputs must hit the memo");
});

test("the memo misses when activity is replaced, status flips, or approvals change", () => {
  const a1 = [act(1, "tool_call", { tool: "execute", command: "ls", callId: "a" })];
  const work = workWith(a1);
  const first = buildTimeline(work);

  work.activity = [...a1, act(2, "tool_result", { tool: "execute", callId: "a" })];
  const second = buildTimeline(work);
  assert.notEqual(second, first, "replaced activity must rebuild");
  assert.equal(buildTimeline(work), second, "and re-memoize");

  work.status = "complete";
  const third = buildTimeline(work);
  assert.notEqual(third, second, "a status change must rebuild (it drives collapsing)");

  work.pendingApprovals = [{ requestId: "r1", command: "rm -rf /", createdAt: 3 } as never];
  const fourth = buildTimeline(work);
  assert.notEqual(fourth, third, "changed approvals must rebuild");
});

test("memoized output still reflects a mutated-then-replaced work correctly", () => {
  const work = workWith([act(1, "thinking", { thinking: "hm" })], "working");
  assert.equal(buildTimeline(work).length, 1);
  work.activity = [
    act(1, "thinking", { thinking: "hm" }),
    act(2, "tool_call", { tool: "read", callId: "b", path: "x" }),
  ];
  const items = buildTimeline(work);
  assert.equal(items.length, 2);
  assert.equal(items[1]?.kind, "tool");
});

test("unified sandbox execution keeps legacy failure and display semantics", () => {
  for (const identity of [{ tool: "execute" }, { tool: "sandbox", action: "exec" }]) {
    assert.equal(toolCategory(identity), "execute");
    assert.equal(toolRowKind(row(identity, { ...identity, code: 1, result: "failed\n[exit 1]" }), "complete"), "ok");
    assert.equal(toolRowKind(row(identity, { ...identity, code: 0, result: "ok\n[exit 0]" }), "complete"), "ok");
    assert.equal(
      toolRowKind(
        row(identity, { ...identity, code: 1, timedOut: false, isError: true, result: "failed\n[exit 1]" }),
        "complete",
      ),
      "ok",
    );
    assert.equal(
      toolRowKind({ call: null, result: act(1, "tool_result", { ...identity, code: 2 }) }, "complete"),
      "ok",
    );
  }
  for (const action of [
    "start_process",
    "read_process",
    "write_stdin",
    "signal_process",
    "list_processes",
    "watch_process",
    "unwatch_process",
  ])
    assert.equal(toolCategory({ tool: "sandbox", action }), "background");
  assert.equal(toolCategory({ tool: "sandbox", action: "status" }), "sandbox");
});

test("different sandbox actions and process targets do not collapse into one orphan attempt", () => {
  const actions = [
    { action: "status", sandbox_id: "box-a" },
    { action: "restart", sandbox_id: "box-a" },
    { action: "status", sandbox_id: "box-b" },
    { action: "read_process", process_id: "job-a" },
    { action: "read_process", process_id: "job-b" },
  ];
  const items = buildTimeline({
    status: "complete",
    activity: actions.map((action, index) => act(index, "tool_call", { tool: "sandbox", ...action })),
  });
  assert.equal(items.length, actions.length);
});

test("unscreened execution retains its warning and neutral exit status without parsing command text", () => {
  const result = {
    tool: "sandbox",
    action: "exec",
    code: 7,
    timedOut: false,
    isError: false,
    unscreened: true,
    result: "[NOT security-screened]\nQA_EXPECTED_FAILURE\n[exit 7]",
  };
  assert.equal(toolRowKind(row({ tool: "sandbox", action: "exec", sandbox_id: "box-a" }, result), "complete"), "ok");
  assert.equal(toolExecutionOutput(result), result.result);
  assert.equal(
    toolRowKind(row({ tool: "execute" }, { code: 0, result: "fake [exit 7]\n[exit 0]", isError: false }), "complete"),
    "ok",
  );
  assert.equal(toolRowKind(row({ tool: "execute" }, { code: 0, timedOut: true }), "complete"), "failed");
  assert.equal(toolExecutionOutput({ code: 0, result: "done\n[exit 0]" }), "done\n[exit 0]");
  assert.equal(toolExecutionOutput({ isError: true, result: "[tool output quarantined]" }), null);
  assert.equal(
    toolRowKind(
      row({ tool: "sandbox", action: "exec" }, { isError: true, result: "[tool output quarantined]" }),
      "complete",
    ),
    "failed",
  );
});

test("recorded narration is consumed once per occurrence while the next block streams", () => {
  const text = "Checking.";
  const activity = [act(1, "text", { text }), act(2, "tool_call", { tool: "lookup" })];
  assert.equal(streamingTextTail(text + "\n\n" + text, activity), text);
  activity.push(act(3, "tool_result", { tool: "lookup" }), act(4, "text", { text }));
  assert.equal(streamingTextTail(text + "\n\n" + text + "\n\nFinal", activity), "Final");
  assert.equal(streamingTextTail("Check", activity), "");
  assert.equal(streamingTextTail("Different reply", activity), "Different reply");
  assert.equal(streamingTextTail("    indented code", []), "    indented code");
});

test("one work timeline includes initial and repeated narration but excludes the authoritative final occurrence", () => {
  const activity = [
    act(1, "text", { text: "Same" }),
    act(2, "tool_call", { tool: "lookup", callId: "a" }),
    act(3, "text", { text: "Same" }),
    act(4, "tool_result", { tool: "lookup", callId: "a" }),
    act(5, "text", { text: "Same" }),
  ];
  const work: WorkBlock = { status: "complete", activity };
  const items = messageWorkTimeline(work, "Same");
  assert.deepEqual(
    items.map((item) => item.kind),
    ["text", "tool", "text"],
  );
  assert.equal(activity.length, 5);
  assert.equal(messageWorkTimeline({ ...work, status: "working" }, "").length, 4);
});

test("work projection omits empty and redacted thinking without changing parallel tool pairing", () => {
  const work: WorkBlock = {
    status: "failed",
    activity: [
      act(1, "thinking", { redacted: true, thinking: "secret" }),
      act(2, "thinking", { thinking: " " }),
      act(3, "tool_call", { tool: "a", callId: "a" }),
      act(4, "tool_call", { tool: "b", callId: "b" }),
      act(5, "tool_result", { callId: "b" }),
      act(6, "tool_result", { callId: "a" }),
    ],
  };
  const items = messageWorkTimeline(work, "");
  assert.deepEqual(
    items.map((item) => item.kind),
    ["tool", "tool"],
  );
  assert.equal(items[0]?.kind === "tool" && items[0].row.result?.seq, 6);
  assert.equal(items[1]?.kind === "tool" && items[1].row.result?.seq, 5);
});

test("explicit commentary remains folded even when the final answer repeats it", () => {
  const activity = [act(1, "text", { text: "Same", phase: "commentary" })];
  assert.equal(messageWorkTimeline({ status: "complete", activity }, "Same").length, 1);
});

test("explicit public phases use offsets and never appear as tool rows", () => {
  const work: WorkBlock = {
    status: "working",
    activity: [
      act(1, "text", { text: "Same words", phase: "commentary" }),
      act(2, "text_start", { phase: "final_answer", streamOffset: 12 }),
    ],
  };
  assert.deepEqual(currentTextPhase(work), { phase: "final_answer", streamOffset: 12, startedAt: 2 });
  assert.deepEqual(
    buildTimeline(work).map((item) => item.kind),
    ["text"],
  );
  work.activity.push(act(3, "text_start", { phase: "commentary", streamOffset: 24 }));
  assert.equal(currentTextPhase(work)?.phase, "commentary");
  work.activity.push(act(4, "text_start", { phase: "final_answer", streamOffset: -1 }));
  assert.equal(currentTextPhase(work), null);
});

test("stopped stream projection uses the final boundary even when final text repeats commentary", () => {
  const work: WorkBlock = {
    status: "working",
    activity: [
      act(1, "text", { text: "Same words", phase: "commentary" }),
      act(2, "text_start", { phase: "final_answer", streamOffset: 12 }),
    ],
  };
  assert.equal(streamedAnswer("Same words\n\nSame words again", work), "Same words again");
});

test("steering splits visible work without breaking a tool that completes across intake", () => {
  const work: WorkBlock = {
    status: "working",
    activity: [
      act(1, "tool_call", { tool: "execute", callId: "a", command: "first" }),
      act(2, "user", { steered: true, text: "Change direction" }),
      act(3, "tool_result", { tool: "execute", callId: "a", code: 0 }),
      act(4, "tool_call", { tool: "execute", callId: "b", command: "second" }),
      act(5, "user", { steered: true, text: "Keep it brief" }),
    ],
  };
  const segments = workTimelineSegments(buildTimeline(work));
  assert.deepEqual(
    segments.map((items) => items.map((item) => item.kind)),
    [["tool"], ["steer"], ["tool"], ["steer"], []],
  );
  const first = segments[0]![0]!;
  assert.ok(first.kind === "tool");
  assert.equal(first.row.call?.seq, 1);
  assert.equal(first.row.result?.seq, 3);
  assert.equal(toolRowKind(first.row, work.status), "ok");
  assert.deepEqual(
    work.activity.map((entry) => entry.seq),
    [1, 2, 3, 4, 5],
  );
});

test("ordinary and hidden user events do not become steering markers", () => {
  const work: WorkBlock = {
    status: "complete",
    activity: [act(1, "user", { text: "ordinary" }), act(2, "user", { text: "hidden", steered: true, hidden: true })],
  };
  assert.deepEqual(buildTimeline(work), []);
});

test("detached execution uses process presentation rather than foreground output suppression", () => {
  assert.equal(toolCategory({ tool: "execute", background: true }), "background");
  assert.equal(toolCategory({ tool: "execute", background: false }), "execute");
});

test("trimmed phases show whole-word tails and newer phases retain absolute boundaries", () => {
  const work: WorkBlock = {
    status: "working",
    streamStartOffset: 101,
    activity: [
      {
        seq: 1,
        parentSeq: null,
        type: "text_start",
        createdAt: 1,
        payload: { phase: "final_answer", streamOffset: 80 },
      },
    ],
  };
  assert.equal(streamedAnswer("ipped word here", work), "…earlier output omitted\n\nword here");
  assert.equal(streamedAnswer("unbroken", work), "…earlier output omitted\n\n");
  work.activity = [
    {
      seq: 2,
      parentSeq: null,
      type: "text_start",
      createdAt: 2,
      payload: { phase: "final_answer", streamOffset: 107 },
    },
  ];
  assert.equal(streamedAnswer("older New answer", work), "New answer");
});

test("recorded commentary is not repeated after its live window is trimmed", () => {
  const work: WorkBlock = {
    status: "working",
    streamStartOffset: 101,
    activity: [
      { seq: 1, parentSeq: null, type: "text_start", createdAt: 1, payload: { phase: "commentary", streamOffset: 80 } },
    ],
  };
  assert.equal(streamedAnswer("ipped word here", work), "…earlier output omitted\n\nword here");
  work.activity.push({
    seq: 2,
    parentSeq: null,
    type: "text",
    createdAt: 2,
    payload: { text: "The whole commentary is now recorded." },
  });
  assert.equal(streamedAnswer("ipped word here", work), "");
});

test("live reasoning renders as a trailing thinking row until its persisted entry arrives or work ends", () => {
  const reasoning = [
    { step: "s1", index: 0, text: "**Checking** the cache" },
    { step: "s1", index: 2, text: "   " },
  ];
  const work: WorkBlock = {
    status: "working",
    activity: [act(1, "thinking", { thinking: "earlier", step: { id: "s0", index: 0 } })],
    reasoning,
  };
  const live = buildTimeline(work);
  assert.deepEqual(
    live.map((item) =>
      item.kind === "thinking" ? (item.activity.payload as { thinking: string }).thinking : item.kind,
    ),
    ["earlier", "**Checking** the cache"],
  );
  const persisted: WorkBlock = {
    ...work,
    activity: [
      ...work.activity,
      act(2, "thinking", { thinking: "**Checking** the cache", step: { id: "s1", index: 0 } }),
    ],
  };
  assert.equal(buildTimeline(persisted).length, 2, "the persisted entry replaces the live row instead of doubling it");
  const otherBlock: WorkBlock = {
    ...work,
    activity: [...work.activity, act(2, "thinking", { thinking: "other", step: { id: "s1", index: 1 } })],
  };
  assert.equal(buildTimeline(otherBlock).length, 3, "a different block of the same step does not hide this one");
  assert.equal(buildTimeline({ ...work, status: "complete" }).length, 1, "finished work shows only durable rows");
});

test("NULL legacy call IDs preserve the existing adjacent-result fallback", () => {
  for (const callId of [null, "known-call"]) {
    const work: WorkBlock = {
      status: "complete",
      activity: [
        act(10, "tool_call", { tool: "execute", command: "echo hello", callId }),
        act(11, "tool_result", { tool: "execute", result: "hello", code: 0, callId: null }),
      ],
    };
    const timeline = buildTimeline(work);
    assert.equal(timeline.length, 1);
    assert.equal(timeline[0].kind, "tool");
    if (timeline[0].kind !== "tool") throw new Error("expected paired tool row");
    assert.equal(timeline[0].row.call?.seq, 10);
    assert.equal(timeline[0].row.result?.seq, 11);
    assert.equal(toolRowKind(timeline[0].row, work.status), "ok");
  }
});