atlas-iac/testing/tests/test_hermes_hux_ui_activity.mjs

import assert from "node:assert/strict";
import { readFileSync } from "node:fs";
import test from "node:test";

import {
  ACTIVITY_TIMELINE_FLAG,
  FOUNDATION_FLAG,
  activityTimelineEnabled,
  mergeActivityEvents,
  normalizeCancellationReceipt,
  surfacePolicy,
} from "../../dockerfiles/hermes-webui-hux/activity/model.ts";
import {
  isOpaqueId,
  isRfc3339Utc,
  safeEvidenceLabel,
  safeText,
} from "../../dockerfiles/hermes-webui-hux/activity/security.ts";

const CONVERSATION = "conv_test1234";
const KINDS = JSON.parse(
  readFileSync("services/hermes/contracts/hux/event.schema.json"),
).properties.kind.enum;

function event(seq, kind = "tool.call", extra = {}) {
  return {
    schema: "hux.event.v1",
    id: `evt_test${String(seq).padStart(4, "0")}`,
    seq,
    ts: `2026-08-24T10:${String(Math.floor(seq / 60)).padStart(2, "0")}:${String(seq % 60).padStart(2, "0")}Z`,
    conversation_id: CONVERSATION,
    kind,
    summary: `Safe update ${seq}`,
    evidence: [],
    provenance: { surface: "chat" },
    sensitivity: "personal",
    redaction: { level: "none" },
    ...extra,
  };
}

test("feature remains default-off and depends on the foundation", () => {
  assert.equal(activityTimelineEnabled(), false);
  assert.equal(activityTimelineEnabled([]), false);
  assert.equal(activityTimelineEnabled([ACTIVITY_TIMELINE_FLAG]), false);
  assert.equal(activityTimelineEnabled([FOUNDATION_FLAG]), false);
  assert.equal(
    activityTimelineEnabled([FOUNDATION_FLAG, ACTIVITY_TIMELINE_FLAG]),
    true,
  );
});

test("Chat policy is intentionally calmer and every returned policy is isolated", () => {
  const chat = surfacePolicy("chat");
  const worker = surfacePolicy("worker");
  assert.ok(chat.maxItems < worker.maxItems);
  assert.ok(chat.initialItems < worker.initialItems);
  assert.ok(chat.evidenceLimit < worker.evidenceLimit);
  assert.equal(chat.expandRoutineEvents, false);
  assert.equal(worker.expandRoutineEvents, true);
  chat.maxItems = 1;
  assert.equal(surfacePolicy("chat").maxItems, 80);
  assert.deepEqual(surfacePolicy("api"), worker);
  assert.ok(
    surfacePolicy("voice").maxItems < surfacePolicy("telegram").maxItems,
  );
});

test("plain-text security redacts credential shapes without serializing objects", () => {
  assert.equal(safeText({ token: "nope" }, "fallback"), "fallback");
  assert.equal(safeText(" \n\t ", "fallback"), "fallback");
  assert.equal(safeText("abc\u0000def", "fallback"), "abc def");
  assert.equal(safeText("abcdef", "fallback", 4), "abc…");
  assert.equal(safeText("abcdef", "fallback", 0), "…");
  const unsafe = [
    "password=hunter2",
    "Authorization: Basic-abcd",
    "Bearer abc.DEF_123",
    "eyJabcdefgh.abcdefghijkl.abcdefghijkl",
    "https://user:pass@example.test/path",
    "Abcdefghijklmnopqrstuvwxyz1234567890ABCD",
    "-----BEGIN PRIVATE KEY----- top secret -----END PRIVATE KEY-----",
  ].join(" ");
  const scrubbed = safeText(unsafe, "fallback", 1000);
  assert.doesNotMatch(
    scrubbed,
    /hunter2|Basic-abcd|abc\.DEF_123|user:pass|top secret|1234567890ABCD/,
  );
  assert.match(scrubbed, /redacted/);
});

test("only allow-listed evidence labels cross the display boundary", () => {
  assert.deepEqual(safeEvidenceLabel("tool_result"), {
    kind: "tool_result",
    label: "Tool result",
  });
  assert.deepEqual(safeEvidenceLabel("url"), { kind: "url", label: "Website" });
  assert.equal(safeEvidenceLabel("raw_tool_arguments"), null);
  assert.equal(safeEvidenceLabel(7), null);
  assert.equal(isOpaqueId("evt_abcd", "evt"), true);
  assert.equal(isOpaqueId("evt_abcd"), true);
  assert.equal(isOpaqueId("mem_abcd", "evt"), false);
  assert.equal(isOpaqueId("not opaque"), false);
  assert.equal(isOpaqueId(7), false);
  assert.equal(isRfc3339Utc("2026-08-24T10:00:00.123Z"), true);
  assert.equal(isRfc3339Utc("today"), false);
  assert.equal(isRfc3339Utc("2026-99-99T10:00:00Z"), false);
  assert.equal(isRfc3339Utc(7), false);
});

test("every contract event maps to a compact semantic phase", () => {
  const records = KINDS.map((kind, seq) =>
    event(seq, kind, {
      parent_event_id: kind === "run.started" ? "evt_parent1234" : undefined,
    }),
  );
  const result = mergeActivityEvents([], records, CONVERSATION, "worker");
  assert.equal(result.invalid, 0);
  assert.equal(result.items.length, KINDS.length);
  const phases = new Set(result.items.map((item) => item.phase));
  assert.deepEqual(
    phases,
    new Set([
      "intent",
      "action",
      "evidence",
      "delegation",
      "decision",
      "completion",
      "failure",
    ]),
  );
  assert.equal(
    result.items.find((item) => item.kind === "run.failed").tone,
    "warning",
  );
  assert.equal(
    result.items.find((item) => item.kind === "run.completed").tone,
    "success",
  );
  assert.equal(
    result.items.find((item) => item.kind === "approval.requested").tone,
    "warning",
  );
  assert.equal(
    result.items.find((item) => item.kind === "tool.call").routine,
    true,
  );
  assert.equal(
    result.items.find((item) => item.kind === "message.user").routine,
    false,
  );
});

test("root runs are actions while child runs are explicit delegations", () => {
  const result = mergeActivityEvents(
    [],
    [
      event(1, "run.started"),
      event(2, "run.started", { parent_event_id: "evt_parent1234" }),
    ],
    CONVERSATION,
    "chat",
  );
  assert.deepEqual(
    result.items.map((item) => item.phase),
    ["action", "delegation"],
  );
  assert.deepEqual(
    result.items.map((item) => item.tone),
    ["active", "active"],
  );
});

test("normalization redacts summaries and never projects detail, IDs, or URIs", () => {
  const raw = event(1, "tool.result", {
    summary: "token=secret-value result is ready",
    run_id: "run-safe",
    detail: { arguments: "--token secret", result: "raw secret result" },
    evidence: [
      {
        kind: "tool_result",
        id: "secret-id",
        uri: "https://user:pass@example.test",
      },
      { kind: "file", id: "/secret/path" },
      { kind: "bogus", id: "raw" },
      { kind: "url", id: "https://secret" },
    ],
  });
  const item = mergeActivityEvents([], [raw], CONVERSATION, "chat").items[0];
  const serialized = JSON.stringify(item);
  assert.equal(item.runId, "run-safe");
  assert.deepEqual(
    item.evidence.map((entry) => entry.label),
    ["Tool result", "File", "Website"],
  );
  assert.doesNotMatch(
    serialized,
    /secret-value|raw secret result|secret-id|secret\/path|user:pass/,
  );
  assert.equal("detail" in item, false);

  const invalidRun = mergeActivityEvents(
    [],
    [event(2, "tool.call", { run_id: { raw: true } })],
    CONVERSATION,
    "chat",
  );
  assert.equal(invalidRun.items[0].runId, undefined);
});

test("full or restricted redaction ignores even an otherwise safe summary", () => {
  const result = mergeActivityEvents(
    [],
    [
      event(1, "tool.call", {
        summary: "do not show",
        redaction: { level: "full" },
      }),
      event(2, "run.failed", {
        summary: "health detail",
        sensitivity: "restricted",
      }),
      event(3, "run.completed", { summary: {}, redaction: null }),
      event(4, "citation.attached", {
        summary: "",
        redaction: { level: "partial" },
      }),
    ],
    CONVERSATION,
    "chat",
  );
  assert.deepEqual(
    result.items.map((item) => item.summary),
    [
      "Hermes used a tool",
      "The run failed; sensitive detail is hidden",
      "The run completed",
      "Activity recorded; sensitive detail is hidden",
    ],
  );
});

test("reconnect merge is ordered, idempotent, and fail-closed on conflicts", () => {
  const first = mergeActivityEvents(
    [],
    [event(3), event(1)],
    CONVERSATION,
    "chat",
  );
  const replay = { ...event(3), summary: "attempted rewrite" };
  const idConflict = { ...event(4), id: first.items[0].id };
  const seqConflict = { ...event(1), id: "evt_conflict1" };
  const result = mergeActivityEvents(
    first.items,
    [replay, idConflict, seqConflict, event(2)],
    CONVERSATION,
    "chat",
  );
  assert.deepEqual(
    result.items.map((item) => item.seq),
    [1, 2, 3],
  );
  assert.equal(result.items.at(-1).summary, "Safe update 3");
  assert.equal(result.replays, 1);
  assert.equal(result.conflicts, 2);
  assert.equal(result.afterSeq, 3);
  assert.equal(result.invalid, 0);
});

test("an initial same-sequence conflict resolves deterministically", () => {
  const higher = { ...event(7), id: "evt_zulu1234" };
  const lower = {
    ...event(7),
    id: "evt_alpha1234",
    summary: "Deterministic owner",
  };
  const forward = mergeActivityEvents(
    [],
    [higher, lower],
    CONVERSATION,
    "chat",
  );
  const reverse = mergeActivityEvents(
    [],
    [lower, higher],
    CONVERSATION,
    "chat",
  );
  assert.deepEqual(forward.items, reverse.items);
  assert.equal(forward.items[0].id, "evt_alpha1234");
  assert.equal(forward.conflicts, 1);
});

test("malformed and cross-conversation events are rejected", () => {
  const bad = [
    { ...event(1), schema: "hux.event.v0" },
    { ...event(2), id: "bad" },
    { ...event(3), seq: -1 },
    { ...event(4), ts: "later" },
    { ...event(5), conversation_id: "conv_other1234" },
    { ...event(6), kind: "tool.raw" },
  ];
  const result = mergeActivityEvents([], bad, CONVERSATION, "chat");
  assert.deepEqual(result.items, []);
  assert.equal(result.invalid, bad.length);
  assert.equal(result.afterSeq, -1);
});

test("retention is bounded per surface while preserving the resume watermark", () => {
  const records = Array.from({ length: 85 }, (_, seq) => event(seq));
  const result = mergeActivityEvents([], records, CONVERSATION, "chat");
  assert.equal(result.items.length, 80);
  assert.equal(result.items[0].seq, 5);
  assert.equal(result.afterSeq, 84);
  assert.equal(result.truncated, 5);
  const replayed = mergeActivityEvents(
    result.items,
    [event(84)],
    CONVERSATION,
    "chat",
  );
  assert.equal(replayed.replays, 1);
  assert.equal(replayed.truncated, 0);
});

test("evidence display and cancellation receipts obey surface bounds", () => {
  const evidence = Array.from({ length: 10 }, (_, index) => ({
    kind: index % 2 ? "file" : "source",
    id: `secret-${index}`,
    uri: `https://token-${index}`,
  }));
  assert.equal(
    mergeActivityEvents(
      [],
      [event(1, "citation.attached", { evidence })],
      CONVERSATION,
      "chat",
    ).items[0].evidence.length,
    3,
  );
  assert.equal(
    mergeActivityEvents(
      [],
      [event(1, "citation.attached", { evidence })],
      CONVERSATION,
      "worker",
    ).items[0].evidence.length,
    8,
  );

  const receipt = {
    schema: "hux.cancel_receipt.v1",
    id: "rcpt_test1234",
    run_id: "run_1",
    requested_at: "2026-08-24T10:00:00Z",
    acknowledged_at: "2026-08-24T10:00:01Z",
    completed_at: "not-a-time",
    outcome: "cancelled",
    side_effects: Array.from({ length: 7 }, (_, index) => ({
      description: index === 0 ? "token=do-not-show" : `Effect ${index}`,
      reverted: index % 2 === 0,
      evidence: { kind: "file", id: "secret" },
    })),
  };
  const safe = normalizeCancellationReceipt(receipt, "chat");
  assert.equal(safe.sideEffects.length, 4);
  assert.equal(safe.omittedEffects, 3);
  assert.equal(safe.completedAt, undefined);
  assert.equal(safe.acknowledgedAt, receipt.acknowledged_at);
  assert.doesNotMatch(JSON.stringify(safe), /do-not-show|"evidence"|"secret"/);
});

test("cancellation parsing rejects bad identity, time, run, and outcome", () => {
  const base = {
    schema: "hux.cancel_receipt.v1",
    id: "rcpt_test1234",
    run_id: "run_1",
    requested_at: "2026-08-24T10:00:00Z",
    outcome: "already_complete",
    side_effects: [{ description: {}, reverted: false }],
  };
  const safe = normalizeCancellationReceipt(base, "worker");
  assert.equal(safe.sideEffects[0].description, "A side effect was recorded");
  assert.equal(safe.completedAt, undefined);
  assert.equal(
    normalizeCancellationReceipt({ ...base, schema: "old" }, "chat"),
    null,
  );
  assert.equal(
    normalizeCancellationReceipt({ ...base, id: "bad" }, "chat"),
    null,
  );
  assert.equal(
    normalizeCancellationReceipt({ ...base, run_id: 3 }, "chat"),
    null,
  );
  assert.equal(
    normalizeCancellationReceipt({ ...base, requested_at: "bad" }, "chat"),
    null,
  );
  assert.equal(
    normalizeCancellationReceipt({ ...base, outcome: "unknown" }, "chat"),
    null,
  );
  assert.equal(
    normalizeCancellationReceipt({ ...base, side_effects: null }, "chat")
      .sideEffects.length,
    0,
  );
});