File size: 15,359 Bytes
c453128 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 | import { describe, it, expect } from "vitest";
import { buildOverviewTurns } from "../activityOverviewModel";
import type { ActivityEvent } from "../types";
function ev(
event_type: ActivityEvent["event_type"],
timestamp = "2026-06-01T12:00:00.000Z",
detail = "",
): ActivityEvent {
return {
timestamp,
event_type,
icon: "",
title: "",
detail,
tool_name: "",
is_error: false,
files_touched: [],
};
}
describe("buildOverviewTurns", () => {
it("charts a single turn when Hermes batch-flushes identical timestamps", () => {
const ts = "2026-06-01T12:00:00.000Z";
const turns = buildOverviewTurns(
[
ev("user_message", ts, "Build a CNN"),
ev("tool_call", ts),
ev("tool_result", ts),
ev("message", ts),
],
{
startTime: ts,
endTime: Date.parse("2026-06-01T13:00:00.000Z") / 1000,
},
);
expect(turns.length).toBe(1);
expect(turns[0].end).toBeGreaterThan(turns[0].start);
expect(turns[0].counts.tool).toBe(1);
});
it("records agent events in order so the overview can time-resolve them", () => {
const ts = "2026-06-01T12:00:00.000Z";
const turns = buildOverviewTurns(
[
ev("user_message", ts, "Build a CNN"),
ev("thinking_start", ts),
ev("tool_result", ts),
ev("message", ts),
],
{ startTime: ts, endTime: Date.parse("2026-06-01T13:00:00.000Z") / 1000 },
);
// Sequence preserves order (reasoning → tool → generating) and excludes the
// user_message that opened the turn.
expect(turns[0].seq).toEqual(["reasoning", "tool", "generating"]);
});
it("uses taskContent when the DB has agent events but no user_message yet", () => {
const turns = buildOverviewTurns(
[ev("tool_call"), ev("tool_result"), ev("message")],
{
taskContent: "Train MNIST classifier",
startTime: "2026-06-01T10:00:00.000Z",
endTime: Date.parse("2026-06-01T11:00:00.000Z") / 1000,
},
);
expect(turns.length).toBe(1);
expect(turns[0].task).toContain("Train MNIST");
expect(turns[0].end).toBeGreaterThan(turns[0].start);
});
it("spreads multiple turns that share one flush timestamp", () => {
const ts = "2026-06-01T12:00:00.000Z";
const turns = buildOverviewTurns(
[
ev("user_message", ts, "Task A"),
ev("tool_call", ts),
ev("user_message", ts, "Task B"),
ev("message", ts),
],
{
startTime: "2026-06-01T11:00:00.000Z",
endTime: Date.parse("2026-06-01T13:00:00.000Z") / 1000,
},
);
expect(turns.length).toBe(2);
expect(turns[0].end).toBeGreaterThan(turns[0].start);
expect(turns[1].end).toBeGreaterThan(turns[1].start);
});
it("spreads across full desk span when flush timestamps cluster at resume time", () => {
const flush = "2026-06-05T15:43:00.000Z";
const deskStart = "2026-06-05T13:43:00.000Z";
const deskEnd = "2026-06-05T16:00:00.000Z";
const turns = buildOverviewTurns(
[
ev("user_message", flush, "Task A"),
ev("tool_call", flush),
ev("user_message", flush, "Task B"),
ev("message", flush),
],
{ startTime: deskStart, deskEndTime: deskEnd },
);
expect(turns.length).toBe(2);
const span = turns[turns.length - 1].end - turns[0].start;
expect(span).toBeGreaterThan(3600);
// Clustered timestamps are redistributed onto a synthetic axis — per-event
// times aren't trustworthy, so the overview keeps its synthetic layout.
expect(turns.every((t) => !t.timed)).toBe(true);
});
// Regression: a single long agentic turn (one user message, many events over
// an hour) used to collapse into a ~Nx1s sliver because per-event times were
// ignored. With recorded emit-times it must span the real wall-clock instead.
it("places events at their real recorded times across one long turn", () => {
const start = "2026-06-23T12:48:11.000Z";
const startSec = Date.parse(start) / 1000;
const at = (type: ActivityEvent["event_type"], secOffset: number): ActivityEvent => ({
...ev(type, new Date((startSec + secOffset) * 1000).toISOString()),
time_exact: true,
});
const turns = buildOverviewTurns(
[
{ ...ev("user_message", start, "Train a model"), time_exact: true },
at("thinking_start", 5),
at("tool_result", 60), // 1 min in
at("message", 1800), // 30 min in
at("tool_result", 3600), // 60 min in
],
{ startTime: start, deskEndTime: new Date((startSec + 3900) * 1000).toISOString() },
);
expect(turns.length).toBe(1);
expect(turns[0].timed).toBe(true);
expect(turns[0].times.length).toBe(turns[0].seq.length);
// Times are anchored to the real offsets, monotonic, and span ~the full hour.
expect(turns[0].times[0]).toBeCloseTo(startSec + 5, 0);
expect(turns[0].times[turns[0].times.length - 1]).toBeCloseTo(startSec + 3600, 0);
const span = turns[0].times[turns[0].times.length - 1] - turns[0].times[0];
expect(span).toBeGreaterThan(3000);
});
it("interpolates approximate events between exact anchors (monotonic)", () => {
const start = "2026-06-23T12:00:00.000Z";
const startSec = Date.parse(start) / 1000;
const exact = (type: ActivityEvent["event_type"], secOffset: number): ActivityEvent => ({
...ev(type, new Date((startSec + secOffset) * 1000).toISOString()),
time_exact: true,
});
const approx = (type: ActivityEvent["event_type"]): ActivityEvent => ({
...ev(type, start),
time_exact: false,
});
const turns = buildOverviewTurns(
[
{ ...ev("user_message", start, "Task"), time_exact: true },
exact("tool_result", 100),
approx("message"), // no real time — interpolated to ~200s
approx("message"), // ~300s
exact("tool_result", 400),
],
{ startTime: start, deskEndTime: new Date((startSec + 500) * 1000).toISOString() },
);
expect(turns[0].timed).toBe(true);
const t = turns[0].times;
// Non-decreasing, with the two approximate events landing strictly between
// their bracketing anchors (100s and 400s).
for (let i = 1; i < t.length; i++) expect(t[i]).toBeGreaterThanOrEqual(t[i - 1]);
expect(t[1]).toBeGreaterThan(startSec + 100);
expect(t[2]).toBeLessThan(startSec + 400);
});
it("a single turn with only approximate times stays synthetic (timed:false)", () => {
const ts = "2026-06-01T12:00:00.000Z";
const approx = (type: ActivityEvent["event_type"]): ActivityEvent => ({ ...ev(type, ts), time_exact: false });
const turns = buildOverviewTurns(
[approx("user_message"), approx("tool_result"), approx("message")],
{ startTime: ts, endTime: Date.parse("2026-06-01T13:00:00.000Z") / 1000 },
);
expect(turns.length).toBe(1);
expect(turns[0].timed).toBe(false);
});
// Regression (recover-desk-timestamps companion): a recovered desk has REAL,
// well-spread per-event times written back into state.db, but they arrive
// exact:false because the in-memory time_exact markers were lost on the server
// restart. The overview must honour those spread DB times (so the chart spans
// the real ~45-min run) instead of collapsing to the synthetic 1s/event sliver.
it("honours spread DB times even when events are approximate (recovered desk)", () => {
const start = "2026-06-23T12:48:13.000Z";
const startSec = Date.parse(start) / 1000;
// approximate (time_exact:false) but each at its own real wall-clock offset —
// exactly what recover_timestamps.py leaves in the DB after a restart.
const recovered = (type: ActivityEvent["event_type"], secOffset: number): ActivityEvent => ({
...ev(type, new Date((startSec + secOffset) * 1000).toISOString()),
time_exact: false,
});
const turns = buildOverviewTurns(
[
{ ...ev("user_message", start, "Train a model"), time_exact: false },
recovered("tool_result", 5),
recovered("message", 900), // 15 min in
recovered("tool_result", 1800), // 30 min in
recovered("message", 2748), // ~45 min in (real run end)
],
{ startTime: start, deskEndTime: new Date((startSec + 2748) * 1000).toISOString() },
);
expect(turns.length).toBe(1);
expect(turns[0].timed).toBe(true);
expect(turns[0].times.length).toBe(turns[0].seq.length);
// Times track the real offsets and span the full run, not a ~Nx1s sliver.
expect(turns[0].times[0]).toBeCloseTo(startSec + 5, 0);
expect(turns[0].times[turns[0].times.length - 1]).toBeCloseTo(startSec + 2748, 0);
expect(turns[0].times[turns[0].times.length - 1] - turns[0].times[0]).toBeGreaterThan(2000);
});
// Regression (recover-overview, multi-turn): the single-turn recovered case above
// dodges the redistribution path because `timestampsClustered` requires >1 turn. A
// real recovered desk has SEVERAL user turns, and when viewed long after it ran the
// model pins `end` to `now`, inflating deskSpan so the genuine ~45-min eventSpan
// looks "clustered" (< 15% of deskSpan) — which used to trip redistribution and
// throw away the recovered times, collapsing the run to the synthetic sliver. The
// spread DB times must win: every turn stays timed and tracks its real offsets.
it("honours spread DB times across multiple turns viewed long after the run", () => {
const start = "2026-06-23T12:48:13.000Z";
const startSec = Date.parse(start) / 1000;
const recovered = (type: ActivityEvent["event_type"], secOffset: number, detail = ""): ActivityEvent => ({
...ev(type, new Date((startSec + secOffset) * 1000).toISOString(), detail),
time_exact: false,
});
const turns = buildOverviewTurns(
[
recovered("user_message", 0, "Task A"),
recovered("tool_result", 5),
recovered("message", 900), // 15 min in
recovered("user_message", 1800, "Task B"), // 30 min in
recovered("tool_result", 1805),
recovered("message", 2748), // ~45 min in (real run end)
],
// deskEndTime is the REAL end (in the past); `end` still snaps to now() in the
// model, so deskSpan dwarfs eventSpan — the condition that used to mis-fire.
{ startTime: start, deskEndTime: new Date((startSec + 2748) * 1000).toISOString() },
);
expect(turns.length).toBe(2);
// Both turns keep their real per-event times instead of a synthetic redistribution.
expect(turns.every((t) => t.timed)).toBe(true);
// Turn A spans its first ~15 min; turn B ends at the real ~45-min mark.
expect(turns[0].times[0]).toBeCloseTo(startSec + 5, 0);
expect(turns[1].times[turns[1].times.length - 1]).toBeCloseTo(startSec + 2748, 0);
// The run as a whole spans the real ~45 min, not a few-second sliver.
expect(turns[1].times[turns[1].times.length - 1] - turns[0].times[0]).toBeGreaterThan(2000);
});
// Regression (persisted markers, viewed long after the run): a forward desk born
// under the marker-persistence path has REAL per-event times (time_exact:true) for
// EVERY turn — they now survive restart. Viewed days later, `end` snaps to now() so
// deskSpan dwarfs the genuine ~45-min eventSpan, which used to trip the
// `timestampsClustered` redistribution and throw the real times away (collapsing
// the run to the synthetic 1s/event sliver). The exact times must win: every turn
// stays timed and tracks its real offsets, regardless of how late it's viewed.
it("honours exact per-event times across multiple turns viewed long after the run", () => {
const start = "2026-06-23T12:48:13.000Z";
const startSec = Date.parse(start) / 1000;
const exact = (type: ActivityEvent["event_type"], secOffset: number, detail = ""): ActivityEvent => ({
...ev(type, new Date((startSec + secOffset) * 1000).toISOString(), detail),
time_exact: true,
});
const turns = buildOverviewTurns(
[
exact("user_message", 0, "Task A"),
exact("tool_result", 5),
exact("message", 900), // 15 min in
exact("user_message", 1800, "Task B"), // 30 min in
exact("tool_result", 1805),
exact("message", 2748), // ~45 min in (real run end)
],
// Real end is in the past; the model still snaps `end` to now(), so deskSpan
// dwarfs eventSpan — the condition that used to mis-fire redistribution.
{ startTime: start, deskEndTime: new Date((startSec + 2748) * 1000).toISOString() },
);
expect(turns.length).toBe(2);
// No redistribution: both turns keep their real per-event times.
expect(turns.every((t) => t.timed)).toBe(true);
expect(turns[0].times[0]).toBeCloseTo(startSec + 5, 0);
expect(turns[1].times[turns[1].times.length - 1]).toBeCloseTo(startSec + 2748, 0);
// The run as a whole spans the real ~45 min, not a few-second sliver.
expect(turns[1].times[turns[1].times.length - 1] - turns[0].times[0]).toBeGreaterThan(2000);
});
// Regression: real per-event times must win even when the turn *boundaries* are
// degenerate. If several user messages share one DB timestamp, the same-start
// clause of `needsEvenSpread` fires; without gating it on real-times, that
// redistributed onto the synthetic axis and dropped the exact times. The last
// turn (which has room to span) must still resolve its real recorded times.
it("keeps exact times even when turn boundaries are degenerate (needsEvenSpread)", () => {
const base = "2026-06-23T12:00:00.000Z";
const baseSec = Date.parse(base) / 1000;
const at = (type: ActivityEvent["event_type"], sec: number): ActivityEvent => ({
...ev(type, new Date((baseSec + sec) * 1000).toISOString()),
time_exact: true,
});
const turns = buildOverviewTurns(
[
{ ...ev("user_message", base, "A"), time_exact: true },
at("tool_result", 10), at("message", 200),
{ ...ev("user_message", base, "B"), time_exact: true }, // identical start ts
at("tool_result", 210), at("message", 400),
],
{ startTime: base, deskEndTime: new Date((baseSec + 410) * 1000).toISOString() },
);
expect(turns.length).toBe(2);
// Real per-event placement survives the degenerate (identical) turn starts —
// without the gate, redistribution would force every turn to timed:false.
expect(turns.some((t) => t.timed)).toBe(true);
});
// Guard the gate: approximate times that are CLUSTERED (all within a second —
// the normal batch-flush case, no recovery) must NOT be trusted; the overview
// keeps its synthetic layout rather than reading meaning into flush times.
it("still ignores clustered approximate times (no false positive)", () => {
const ts = "2026-06-01T12:00:00.000Z";
const approx = (type: ActivityEvent["event_type"]): ActivityEvent => ({ ...ev(type, ts), time_exact: false });
const turns = buildOverviewTurns(
[approx("user_message"), approx("thinking_start"), approx("tool_result"), approx("message")],
{ startTime: ts, endTime: Date.parse("2026-06-01T13:00:00.000Z") / 1000 },
);
expect(turns.length).toBe(1);
expect(turns[0].timed).toBe(false);
});
});
|