@decohere/ci-tools 0.0.1-dev.109 → 0.0.1-dev.110

package/tests/bench/regression-issue.test.mjs

@@ -0,0 +1,50 @@
+import { describe, expect, it } from "vitest";
+import { issueBody, issueTitle } from "../../src/bench/regression-issue.mjs";
+
+const regression = {
+	id: "flux.load.cold",
+	label: "flux load, first import (cold process)",
+	baseline: 300,
+	percent: 33.3,
+	recent: [
+		{ date: "2026-10-22T03:17:00Z", sha: "cccccccccccccccc", median: 400 },
+		{ date: "2026-10-21T03:17:00Z", sha: "bbbbbbbbbbbbbbbb", median: 390 },
+	],
+	lastGood: { date: "2026-10-20T03:17:00Z", sha: "aaaaaaaaaaaaaaaa" },
+};
+const context = {
+	repoUrl: "https://github.com/typesmiths/decohere",
+	runUrl: "https://github.com/typesmiths/decohere/actions/runs/1",
+	machine: "gh-actions-node24, Test CPU × 4",
+	threshold: 0.2,
+};
+
+describe("regression issue", () => {
+	it("has one stable title per benchmark, so the nightly run finds and updates it", () => {
+		expect(issueTitle(regression)).toBe("perf(bench): flux.load.cold is slower");
+	});
+
+	it("gives the numbers, the commits to look at, and how to reproduce", () => {
+		const body = issueBody(regression, context);
+		expect(body).toContain(
+			"median 33% above its baseline, on each of the last 2 nightly runs",
+		);
+		expect(body).toContain("- Baseline (median of the runs before): **300.0 ms**");
+		expect(body).toContain(
+			"- 2026-10-22, [`ccccccc`](https://github.com/typesmiths/decohere/commit/cccccccccccccccc): **400.0 ms**",
+		);
+		expect(body).toContain(
+			"(https://github.com/typesmiths/decohere/compare/aaaaaaaaaaaaaaaa...cccccccccccccccc)",
+		);
+		expect(body).toContain("`pnpm bench:save`");
+		expect(body).toContain(context.runUrl);
+	});
+
+	it("says the range is unknown when the baseline run has no commit (converted old history)", () => {
+		const body = issueBody(
+			{ ...regression, lastGood: { date: "2026-08-12T00:00:00Z", sha: null } },
+			context,
+		);
+		expect(body).toContain("unknown (no baseline commit recorded)");
+	});
+});

package/tests/pages/bench-page.test.mjs

@@ -0,0 +1,106 @@
+import { describe, expect, it } from "vitest";
+import { buildBenchmarkHtml, trendSvg } from "../../src/pages/bench-page.mjs";
+
+const context = (benchEnv, cores) => ({
+	benchEnv,
+	os: { platform: "linux", arch: "x64" },
+	cpu: { model: "Test CPU", count: cores },
+	memoryGb: 16,
+	node: "v24.1.0",
+});
+
+const run = (date, sha, results) => ({ date, sha, results });
+const result = (id, median, label = id) => ({
+	id,
+	label,
+	unit: "ms",
+	median,
+	p90: median * 1.2,
+	min: median,
+	max: median,
+	samples: 5,
+});
+
+const history = {
+	version: 2,
+	contexts: {
+		ci: {
+			context: context("gh-actions-node24", 4),
+			runs: [
+				run("2026-10-20T03:17:00Z", "bbbbbbbbbbbb", [
+					result("flux.serve.start", 90, "spawn <to> ready"),
+					result("forge.build.debug", 9),
+				]),
+				run("2026-10-19T03:17:00Z", "aaaaaaaaaaaa", [
+					result("flux.serve.start", 100, "spawn <to> ready"),
+				]),
+			],
+		},
+		laptop: {
+			context: context("(local)", 24),
+			runs: [run("2026-10-01T10:00:00Z", "cccccccccccc", [result("flux.serve.start", 40)])],
+		},
+	},
+};
+
+describe("buildBenchmarkHtml", () => {
+	it("explains an empty history instead of rendering empty charts", () => {
+		const html = buildBenchmarkHtml({ version: 2, contexts: {} });
+		expect(html).toContain("No benchmark history yet");
+		expect(html).not.toContain("<svg");
+	});
+
+	it("draws one card per benchmark, grouped, with a line per machine that ran it", () => {
+		const html = buildBenchmarkHtml(history, { branch: "main", sha: "0123456789abcdef" });
+
+		expect(html.match(/<article class="card">/g)).toHaveLength(2);
+		expect(html.indexOf("flux serve (real process)")).toBeLessThan(
+			html.indexOf("<h2>forge build</h2>"),
+		);
+		const serveCard = html.slice(
+			html.indexOf("<code>flux.serve.start</code>"),
+			html.indexOf("<code>forge.build.debug</code>"),
+		);
+		expect(serveCard.match(/<polyline/g)).toHaveLength(2);
+		expect(html).toContain("built from main @ <code>0123456789ab</code>");
+	});
+
+	it("shows the latest median and the change from the previous run on the newest machine", () => {
+		const html = buildBenchmarkHtml(history);
+		expect(html).toContain("<strong>90.0 ms</strong> median · p90 108 ms");
+		expect(html).toContain('<span class="change better">10.0% faster</span>');
+	});
+
+	it("dates each result and links its commit, in the latest table and on each card", () => {
+		const html = buildBenchmarkHtml(history);
+		expect(html).toContain("<th>Run</th><th>Commit</th>");
+		expect(html).toContain("<td>2026-10-20 03:17 UTC</td>");
+		expect(html).toContain(
+			'<a href="https://github.com/typesmiths/decohere/commit/bbbbbbbbbbbb"><code>bbbbbbb</code></a>',
+		);
+		expect(html).toContain('<p class="when">Last run 2026-10-20 03:17 UTC</p>');
+	});
+
+	it("escapes labels from the history", () => {
+		const html = buildBenchmarkHtml(history);
+		expect(html).toContain("spawn &lt;to&gt; ready");
+		expect(html).not.toContain("spawn <to> ready");
+	});
+});
+
+describe("trendSvg", () => {
+	it("places a single point without dividing by a zero time range", () => {
+		const svg = trendSvg([
+			{
+				color: "red",
+				points: [{ time: 1, date: "2026-10-20", sha: "abc", median: 5, p90: 6 }],
+			},
+		]);
+		expect(svg).toContain("<circle");
+		expect(svg).not.toContain("NaN");
+	});
+
+	it("draws nothing when no machine has a point", () => {
+		expect(trendSvg([])).toBe("");
+	});
+});

package/src/pages/keep-prior-bench-badge.mjs

@@ -1,38 +0,0 @@
-#!/usr/bin/env node
-
-// Keep prior bench badge if no code changes; create fallback if none exists
-
-import { execSync } from "node:child_process";
-import { existsSync, writeFileSync } from "node:fs";
-import { mkdir } from "node:fs/promises";
-
-const BADGES_DIR = "projects/ci-tools/dist/ci/badges";
-const BADGE_FILE = `${BADGES_DIR}/bench.svg`;
-const PAGES_URL = "https://typesmiths.github.io/decohere/badges/bench.svg";
-
-const FALLBACK_BADGE = `<svg xmlns="http://www.w3.org/2000/svg" width="120" height="20" role="img" aria-label="bench: unavailable">
-  <title>bench: unavailable</title>
-  <rect width="120" height="20" fill="#555"/>
-  <rect x="46" width="74" height="20" fill="#9f9f9f"/>
-  <text x="24" y="14" fill="#fff" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" font-size="11" text-anchor="middle">bench</text>
-  <text x="83" y="14" fill="#fff" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" font-size="11" text-anchor="middle">unavailable</text>
-</svg>`;
-
-try {
-	if (!existsSync(BADGE_FILE)) {
-		await mkdir(BADGES_DIR, { recursive: true });
-
-		try {
-			execSync(`curl -fsSL "${PAGES_URL}" -o "${BADGE_FILE}"`, {
-				stdio: "pipe",
-			});
-			console.log("✓ Restored prior bench badge from Pages");
-		} catch {
-			console.warn("::warning::No prior bench badge found on Pages; creating fallback badge.");
-			writeFileSync(BADGE_FILE, FALLBACK_BADGE);
-		}
-	}
-} catch (error) {
-	console.error("Error keeping prior bench badge:", error.message);
-	process.exit(1);
-}