reeval 0.6.3 → 0.8.0

package/dist/{cli-BIhzl_v_.js → cli-BMjTKY-D.js}

@@ -5,25 +5,25 @@ import { o as redactSecrets, s as registerSecret } from "./redact-DJ9JCShu.js";
 import { o as PACKAGE_VERSION } from "./root-CvdRlRZX.js";
 import { T as workspaceLabel, u as environmentSandbox } from "./environment-DiCljXMD.js";
 import { a as fromCwd, i as evalSlug, n as attemptKey, o as newRunId } from "./ids-Bp6ut7ZD.js";
-import { I as LABEL, X as imageInfo } from "./src-zOeHvrzR.js";
-import { d as sandboxId, s as harnessId } from "./options-CWYB613o.js";
-import { c as hostCapacity, d as missingCredentials, f as resolveConcurrency, i as configPathFor, l as loadConfig, p as selectProfile, r as configLabel, s as findConfig, u as malformedCredentials } from "./config-7-gXOgQf.js";
+import { I as LABEL, X as imageInfo } from "./src-TdJuDYqB.js";
+import { d as sandboxId, s as harnessId } from "./options-DPUlm4FI.js";
+import { c as hostCapacity, d as missingCredentials, f as resolveConcurrency, i as configPathFor, l as loadConfig, p as selectProfile, r as configLabel, s as findConfig, u as malformedCredentials } from "./config-B-E0xRkf.js";
 import { s as debugTime } from "./debug-D24NOQH1.js";
 import { i as sandboxWarningLine, n as clip, r as formatDuration } from "./text-gEcKjAyp.js";
 import { t as PrepareError } from "./provider-CmAYTQQs.js";
-import { B as cellContrast, G as effortLabel, L as judgeRatesLabel, M as cellJudgeRates, N as cellUsage, P as collapsedJudgeLabel, _ as hasUngatedCells, b as isUngatedMiss, h as formatUsage, w as runOutcome } from "./model-CtrJdFi6.js";
-import { n as LIVE_FILE, o as liveCounts } from "./live-DZkCtmNP.js";
-import { a as liftLabel, i as gateLabel, n as answersLabel } from "./summary-CahgSiVt.js";
-import { n as FLAGS, o as helpText, r as GLOBAL, t as COMMANDS } from "./cli-spec-DUeyuSIQ.js";
-import { i as runtimeInfo, n as gitInfo, r as runnerInfo, t as author } from "./provenance--cVNo80L.js";
+import { H as cellContrast, L as judgeRatesLabel, M as cellJudgeRates, N as cellUsage, P as collapsedJudgeLabel, _ as hasUngatedCells, b as isUngatedMiss, h as formatUsage, q as effortLabel, w as runOutcome } from "./model-BndnmcLm.js";
+import { a as sampleReasons } from "./pairs-DDsjgm5F.js";
+import { a as liftLabel, i as gateLabel, n as answersLabel } from "./summary-k3mhWWXv.js";
+import { c as writeLiveProgress, i as prepareResume, o as resolveResume, r as loadPrevious, s as writeCurrent, t as completedKeysFor } from "./resume-DZ9AmfPv.js";
+import { n as FLAGS, o as helpText, r as GLOBAL, t as COMMANDS } from "./cli-spec-DSvKsOK4.js";
+import { i as runtimeInfo, n as gitInfo, r as runnerInfo, t as author } from "./provenance-CsX6PHDR.js";
 import { t as budgetMarker } from "./budget-DBUrH4jX.js";
-import { n as prepareAll, o as writePrepared, r as prepareTargets, t as RunStopError } from "./preflight-r0DgibQV.js";
-import { r as imageHarness } from "./prepare-CTLg3nuk.js";
-import { n as createProviders } from "./providers-DNLQzVnc.js";
-import { a as redactArgv, r as commandOf, t as logInvocation } from "./invocations-DHC946b-.js";
+import { n as prepareAll, o as writePrepared, r as prepareTargets, t as RunStopError } from "./preflight-CddycK92.js";
+import { r as imageHarness } from "./prepare-CsP-057p.js";
+import { n as createProviders } from "./providers-DkwiLKDS.js";
+import { a as redactArgv, r as commandOf, t as logInvocation } from "./invocations-nPvrlZPj.js";
 import { n as stripAnsi, t as colors } from "./colors-QV9dFVhN.js";
-import { existsSync, mkdirSync, readdirSync, renameSync, rmSync, writeFileSync } from "node:fs";
-import { hostname } from "node:os";
+import { existsSync, mkdirSync, readdirSync, writeFileSync } from "node:fs";
 import { dirname, join } from "node:path";
 import { appendFile, readFile, rm } from "node:fs/promises";
 import { inspect } from "node:util";
@@ -32,7 +32,7 @@ import { Command, Option } from "commander";
 //#region src/cli-parse.ts
 const OPTIONS = Object.fromEntries(Object.entries(FLAGS).map(([name, spec]) => {
 	const short = "short" in spec ? `-${spec.short}, ` : "";
-	const value = spec.type === "string" ? " <value>" : "";
+	const value = spec.type === "string" ? "optional" in spec && spec.optional ? " [value]" : " <value>" : "";
 	return [name, () => {
 		const option = new Option(`${short}--${name}${value}`);
 		if ("multiple" in spec && spec.multiple) option.argParser((v, prev = []) => [...prev, spec.type === "string" ? v : true]);
@@ -108,7 +108,10 @@ function valuesOf(command, flags) {
 		const value = option && opts[option.attributeName()];
 		if (option?.negate) {
 			if (value === false) values[flag] = true;
-		} else if (value !== void 0) values[flag] = value;
+		} else if (value !== void 0) {
+			const spec = FLAGS[flag];
+			values[flag] = spec.type === "string" && "optional" in spec && spec.optional && value === true ? "" : value;
+		}
 	}
 	return values;
 }
@@ -127,118 +130,6 @@ function usageError(error, command) {
 	}
 	return new UsageError(first.charAt(0).toUpperCase() + first.slice(1));
 }
-function finished(result) {
-	const outcome = result.meta?.outcome;
-	if (result.state === "skipped") return outcome === "fail" || outcome === "infra_error" ? outcome : "skipped";
-	if (outcome === "infra_error") return "infra_error";
-	return result.state === "passed" ? "pass" : "fail";
-}
-/** Sharing-violation codes Windows uses while a reader has `to` open. */
-const BUSY = /* @__PURE__ */ new Set([
-	"EPERM",
-	"EACCES",
-	"EBUSY"
-]);
-/** How long a busy rename is retried. A studio poll holds the file for a few milliseconds. */
-const BUSY_MS = 1e3;
-function pause(ms) {
-	Atomics.wait(new Int32Array(new SharedArrayBuffer(4)), 0, 0, ms);
-}
-/**
-* Renames `from` over `to`. Windows refuses while a reader (`reeval studio`) has `to` open, for a
-* moment: try again for up to a second rather than drop the update.
-*/
-function replace(from, to, rename = renameSync, hooks = {}) {
-	const now = hooks.now ?? Date.now;
-	const wait = hooks.wait ?? pause;
-	const deadline = now() + (hooks.busyMs ?? BUSY_MS);
-	for (let attempt = 1;; attempt++) try {
-		rename(from, to);
-		return;
-	} catch (error) {
-		const code = error.code;
-		if (!code || !BUSY.has(code) || now() >= deadline) throw error;
-		wait(Math.min(attempt, 32));
-	}
-}
-/** Writes the file at once, then on every change and heartbeat. Write errors never fail the run. */
-function writeLiveProgress(options) {
-	const now = options.now ?? (() => /* @__PURE__ */ new Date());
-	const file = join(options.dir, LIVE_FILE);
-	const tmp = `${file}.${process.pid}.tmp`;
-	const attempts = new Map(options.planned.map((p) => [p.key, {
-		key: p.key,
-		evalId: p.id,
-		name: p.attempt.name,
-		file: p.file,
-		cell: p.attempt.cell.key,
-		n: p.attempt.n,
-		status: p.attempt.plan.skip ? "skipped" : "queued",
-		startedAt: null,
-		durationMs: null
-	}]));
-	let phase = "running";
-	let stopped = false;
-	const read = () => {
-		const list = [...attempts.values()];
-		return {
-			schemaVersion: 1,
-			runId: options.runId,
-			startedAt: options.startedAt.toISOString(),
-			pid: process.pid,
-			host: hostname(),
-			heartbeatAt: now().toISOString(),
-			phase,
-			planned: list.length,
-			counts: liveCounts(list),
-			attempts: list
-		};
-	};
-	const write = () => {
-		if (stopped) return;
-		try {
-			writeFileSync(tmp, JSON.stringify(read()));
-			replace(tmp, file);
-		} catch {}
-	};
-	try {
-		mkdirSync(options.dir, { recursive: true });
-	} catch {}
-	write();
-	const heartbeat = setInterval(write, options.heartbeatMs ?? 2e3);
-	heartbeat.unref();
-	const stop = () => {
-		stopped = true;
-		clearInterval(heartbeat);
-	};
-	return {
-		start(key, at) {
-			const a = attempts.get(key);
-			if (!a) return;
-			a.status = "running";
-			a.startedAt = new Date(at).toISOString();
-			write();
-		},
-		end(result) {
-			const a = attempts.get(result.key);
-			if (!a) return;
-			a.status = finished(result);
-			a.durationMs = result.durationMs;
-			write();
-		},
-		phase(next) {
-			phase = next;
-			write();
-		},
-		stop,
-		remove() {
-			stop();
-			rmSync(file, { force: true });
-			rmSync(tmp, { force: true });
-		},
-		read
-	};
-}
 //#endregion
 //#region src/preflight.ts
 /** How the terminal names a sandbox to prepare: its cell id, `+ install`, and a harness other than Cursor. */
@@ -578,9 +469,9 @@ function cellRows(c, e, indent) {
 }
 /** One row per compare question: its win rates, or why it has none. They never gate. */
 function questionRows(c, e, indent) {
-	return (e.answers ?? []).map((a) => {
+	return (e.answers ?? []).flatMap((a) => {
 		const label = answersLabel(a);
-		return `${indent}${c.dim("question")}  ${a.error !== void 0 ? c.yellow(label) : label}`;
+		return [`${indent}${c.dim("question")}  ${a.error !== void 0 ? c.yellow(label) : label}`, ...sampleReasons(a.votes).map((reason) => `${indent}  ${c.dim(reason)}`)];
 	});
 }
 /** The run's evals as a file › group tree, one row per eval and per cell of a matrix. */
@@ -793,7 +684,7 @@ var AgentRenderer = class {
 				for (const w of e.sandboxWarnings ?? []) this.#write(sandboxWarningLine(w));
 				return;
 			case "run-start":
-				this.#write(`reeval run: ${e.files} eval file${e.files === 1 ? "" : "s"}`);
+				this.#write(`reeval run: ${e.runId} · ${e.files} eval file${e.files === 1 ? "" : "s"}`);
 				return;
 			case "collected":
 				this.#tags.addFile(e.evals);
@@ -839,7 +730,10 @@ var AgentRenderer = class {
 					if (outcome === "cancelled" && !this.#printed.has(result.key)) this.#failure(result, attempt.cell, attempt.n);
 				}
 				for (const error of e.unhandledErrors) this.#write(`ERROR ${error}`);
-				for (const entry of e.record.evals) for (const a of entry.answers ?? []) this.#write(`question ${entry.id}: ${answersLabel(a)}`);
+				for (const entry of e.record.evals) for (const a of entry.answers ?? []) {
+					this.#write(`question ${entry.id}: ${answersLabel(a)}`);
+					for (const reason of sampleReasons(a.votes)) this.#write(`  ${reason}`);
+				}
 				const s = e.record.summary;
 				const cancelled = failing.filter((f) => f.outcome === "cancelled").length;
 				const parts = [
@@ -946,7 +840,7 @@ var PlainRenderer = class {
 				const files = `${e.files} eval file${e.files === 1 ? "" : "s"}`;
 				const order = `${this.#o.output.logOrder} logs`;
 				const setup = `harness ${e.harness} · sandbox ${e.sandbox} · concurrency ${e.concurrency}`;
-				this.#write(c.dim(`reeval ${e.version} · ${files} · ${setup} · ${order}`));
+				this.#write(c.dim(`reeval ${e.version} · run ${e.runId} · ${files} · ${setup} · ${order}`));
 				return;
 			}
 			case "plan":
@@ -1226,7 +1120,7 @@ async function run(argv) {
 		return values.help ? EXIT.ok : EXIT.usage;
 	}
 	if (command === "init") {
-		const { initCommand } = await import("./init-DDEpamaf.js");
+		const { initCommand } = await import("./init-DBHyi92L.js");
 		return initCommand(values, filters, {
 			out: (s) => console.log(s),
 			input: process.stdin,
@@ -1235,7 +1129,7 @@ async function run(argv) {
 		});
 	}
 	if (command === "doctor") {
-		const { doctorCommand } = await import("./doctor-C06kYs2r.js");
+		const { doctorCommand } = await import("./doctor-B6oYTPoR.js");
 		return doctorCommand(values, {
 			out: (s) => console.log(s),
 			color: !!process.stdout.isTTY && !process.env.NO_COLOR,
@@ -1243,7 +1137,7 @@ async function run(argv) {
 		});
 	}
 	if (command === "report") {
-		const { reportCommand } = await import("./command-KoOJ0ntD.js");
+		const { reportCommand } = await import("./command-JPGNzFsA.js");
 		return reportCommand(values, filters, {
 			out: (s) => console.log(s),
 			err: (s) => console.error(dim(s)),
@@ -1258,7 +1152,7 @@ async function run(argv) {
 	}
 	if (command === "studio") {
 		const config = await loadConfig(configPath, "studio");
-		const { studioCommand } = await import("./command-fJpZK205.js");
+		const { studioCommand } = await import("./command-BdstSEHu.js");
 		return studioCommand({
 			config,
 			flags: values,
@@ -1269,7 +1163,7 @@ async function run(argv) {
 	if (command === "mcp") {
 		if (!values.http) console.log = console.info = console.debug = console.error;
 		const config = await loadConfig(configPath, "mcp");
-		const { mcpCommand } = await import("./command-B8tcwUJk.js");
+		const { mcpCommand } = await import("./command-BK4kOgkH.js");
 		return mcpCommand({
 			config,
 			flags: values,
@@ -1278,7 +1172,7 @@ async function run(argv) {
 	}
 	if (command in HISTORY_COMMANDS) {
 		const config = await loadConfig(configPath, command);
-		const commands = await import("./commands-DE604ldB.js");
+		const commands = await import("./commands-Duo2ur3r.js");
 		const ctx = {
 			config,
 			flags: values,
@@ -1291,7 +1185,7 @@ async function run(argv) {
 			case "compare": return commands.compareCommand(ctx, filters);
 			case "checkout": return commands.checkoutCommand(ctx, filters);
 			case "query": {
-				const { queryCommand } = await import("./query-BWc_5Lhe.js");
+				const { queryCommand } = await import("./query-noGbekGF.js");
 				return queryCommand({
 					...ctx,
 					err: (s) => console.error(dim(s))
@@ -1299,7 +1193,7 @@ async function run(argv) {
 			}
 			case "analyze": {
 				loadEnvFiles(configPath);
-				const { analyzeCommand } = await import("./command-CwuwuKJD.js");
+				const { analyzeCommand } = await import("./command-CTVGuH8y.js");
 				return analyzeCommand({
 					...ctx,
 					err: (s) => console.error(dim(s))
@@ -1338,8 +1232,8 @@ async function cleanCommand(flags, configPath) {
 	const where = fromCwd(debugRoot(config.root)) || ".";
 	console.log(runs.length ? `Removed debug data for ${runs.length} run${runs.length === 1 ? "" : "s"} (${formatBytes(bytes)}) from ${where}` : `No debug data to remove in ${where}`);
 	if (flags.images) {
-		const { pruneImages } = await import("./src-CDiGyoDt.js");
-		const { requireDocker } = await import("./providers-B2cSfewm.js");
+		const { pruneImages } = await import("./src-BT4dB5CO.js");
+		const { requireDocker } = await import("./providers-yXZZcPfx.js");
 		await requireDocker();
 		const removed = await pruneImages();
 		console.log(removed ? `Removed ${removed} image${removed === 1 ? "" : "s"} reeval built that nothing uses` : "No leftover images to remove");
@@ -1357,7 +1251,7 @@ function verbosityOf(flags) {
 * does. Without a config, just reeval's image.
 */
 async function imageBuildCommand(flags, configPath, filters) {
-	const { buildImages, imagePlanLines, imageTargets } = await import("./images-B8Mp_VXZ.js");
+	const { buildImages, imagePlanLines, imageTargets } = await import("./images-CscddrJU.js");
 	const { docker } = await import("./sandbox.js");
 	let root = process.cwd();
 	const { fakeHarness } = await import("./harness.js");
@@ -1444,6 +1338,10 @@ async function runCommand(flags, configPath, filters) {
 	const setup = await prepareRun(flags, configPath);
 	const ctx = await startRun(setup);
 	invocation.runId = ctx.runId;
+	const githubOut = process.env.GITHUB_OUTPUT;
+	if (githubOut) await appendFile(githubOut, `run-id=${ctx.runId}\n`).catch(() => {});
+	const stepStart = process.env.GITHUB_STEP_SUMMARY;
+	if (stepStart) await appendFile(stepStart, `run ${ctx.runId}\n`).catch(() => {});
 	const { engineVersions } = await import("./node-Co7EXge0.js");
 	const renderer = createRenderer({
 		out: setup.json ? { write: () => true } : process.stdout,
@@ -1457,7 +1355,7 @@ async function runCommand(flags, configPath, filters) {
 		executed = await executeRun(ctx, setup, filters, renderer);
 	} catch (error) {
 		renderer.close();
-		await rm(ctx.stagingDir, {
+		if (setup.resume === void 0) await rm(ctx.stagingDir, {
 			recursive: true,
 			force: true
 		});
@@ -1466,14 +1364,14 @@ async function runCommand(flags, configPath, filters) {
 	const { engineRun, plan, sandboxVersions, budgetReached, live, notes, questions } = executed;
 	invocation.unhandledErrors = engineRun.unhandledErrors;
 	if ((interrupted || engineRun.cancelled) && plan?.sandboxes.some((s) => s.kind === "docker")) {
-		const { removeRunContainers } = await import("./docker-CiNVc7di.js");
+		const { removeRunContainers } = await import("./docker-D6t23fZl.js");
 		await removeRunContainers(ctx.runId);
 	}
 	if (notes) {
 		const cancelled = interrupted || engineRun.cancelled;
 		const written = await notes.finish(cancelled);
 		if (!cancelled && written.length) {
-			const { notesLine } = await import("./agent-DKuFnZPM.js");
+			const { notesLine } = await import("./agent-B72lZ6w5.js");
 			const text = notesLine(ctx.config.root, ctx.runId, written);
 			renderer.onEvent({
 				type: "notice",
@@ -1490,8 +1388,8 @@ async function runCommand(flags, configPath, filters) {
 	}
 	const stopped = interrupted || engineRun.cancelled;
 	const answers = questions && !stopped ? await questions.settle() : void 0;
-	const { lockfiles } = await import("./state-BaHjYXr1.js");
-	const { writeScratchRun } = await import("./results-CrFs7_In.js");
+	const { lockfiles } = await import("./state-Bf_Vq_pV.js");
+	const { writeScratchRun } = await import("./results-BcSHWbgB.js");
 	const run = await writeScratchRun({
 		ctx,
 		author: ctx.author,
@@ -1509,13 +1407,15 @@ async function runCommand(flags, configPath, filters) {
 		worktreePatch: ctx.worktreePatch,
 		lockfiles: await lockfiles(ctx.config.root),
 		git: ctx.git,
+		previous: ctx.previous,
+		previousArtifacts: ctx.previousArtifacts,
 		runner: runnerInfo(),
 		...answers?.size ? { answers } : {}
 	}, () => live?.stop());
 	const { record } = run;
 	if (notes && !stopped) {
-		const { summarizeRun } = await import("./agent-DKuFnZPM.js");
-		const { resolveAnalysisModel } = await import("./model-BeoRxCcJ.js");
+		const { summarizeRun } = await import("./agent-B72lZ6w5.js");
+		const { resolveAnalysisModel } = await import("./model-DMOraCcm.js");
 		const model = resolveAnalysisModel(ctx.config, "agent");
 		const { level, text } = await summarizeRun(ctx.config, record.runId, model);
 		renderer.onEvent({
@@ -1542,7 +1442,7 @@ async function runCommand(flags, configPath, filters) {
 	const { config } = ctx;
 	const mode = setup.record;
 	if ((flags.record || mode === "always" || mode === "ci" && inCI()) && record.status !== "interrupted") {
-		const { record: recordCmd } = await import("./commands-DE604ldB.js");
+		const { record: recordCmd } = await import("./commands-Duo2ur3r.js");
 		const allow = record.status === "budget" ? { "allow-incomplete": true } : {};
 		console.log(await recordCmd(config, record.runId, {
 			...flags,
@@ -1580,7 +1480,7 @@ async function prepareRun(flags, configPath) {
 	const sandboxes = [environmentSandbox(config.environment, config.sandbox), ...profile?.defaults.environment ? [environmentSandbox(profile.defaults.environment, config.sandbox)] : []];
 	const concurrency = await resolveConcurrency(config, flags.concurrency, async () => {
 		if (!sandboxes.some((s) => s.kind === "docker")) return hostCapacity();
-		const { dockerCapacity } = await import("./src-CDiGyoDt.js");
+		const { dockerCapacity } = await import("./src-BT4dB5CO.js");
 		return await dockerCapacity() ?? hostCapacity();
 	});
 	return {
@@ -1598,19 +1498,52 @@ async function prepareRun(flags, configPath) {
 		verbosity: verbosityOf(flags),
 		configFile: configPath && configLabel(configPath),
 		/** Named in the plan line when the project root isn't the current directory. */
-		configElsewhere: !!configPath && dirname(configPath) !== process.cwd()
+		configElsewhere: !!configPath && dirname(configPath) !== process.cwd(),
+		/** `undefined` is a new run; `''` is newest resumable; a string is a run id or prefix. */
+		resume: flags.resume
 	};
 }
 /** The run's id and the state it starts from: commit, uncommitted changes, author. */
-async function startRun({ config, profile, runs, tags, concurrency, verbosity, configFile }) {
+async function startRun({ config, profile, runs, tags, concurrency, verbosity, configFile, resume, dryRun }) {
 	const startedAt = /* @__PURE__ */ new Date();
 	const git = await gitInfo(config.root, [config.history.dir]);
-	const { worktreePatch } = await import("./state-BaHjYXr1.js");
+	const { worktreePatch } = await import("./state-Bf_Vq_pV.js");
 	const patch = git?.dirty ? await worktreePatch(config.root, [config.history.dir]) : null;
 	const who = await author(config.root);
-	const { stagingDir, sweepStaging } = await import("./results-CrFs7_In.js");
-	const { scratchDir } = await import("./files-Cs8tbYQY.js");
+	const { stagingDir, sweepStaging } = await import("./results-BcSHWbgB.js");
+	const { scratchDir } = await import("./files-BpYgF2z0.js");
 	const scratch = scratchDir(config.root);
+	if (resume !== void 0) {
+		const target = await resolveResume(config.root, resume || void 0);
+		if (!dryRun) await prepareResume(target, config.root);
+		await sweepStaging(config.root, target.runId);
+		const aside = dryRun ? {
+			record: target.previous,
+			artifacts: target.previousArtifacts
+		} : await loadPrevious(stagingDir(config.root, target.runId));
+		if (!dryRun) try {
+			const { removeRunContainers } = await import("./docker-D6t23fZl.js");
+			await removeRunContainers(target.runId);
+		} catch {}
+		return {
+			runId: target.runId,
+			stagingDir: stagingDir(config.root, target.runId),
+			config,
+			runs,
+			tags,
+			concurrency,
+			profile,
+			verbosity,
+			configFile,
+			startedAt: target.startedAt,
+			git,
+			worktreePatch: patch,
+			author: who,
+			scratch,
+			previous: aside.record ?? target.previous,
+			previousArtifacts: aside.artifacts ?? target.previousArtifacts
+		};
+	}
 	const stamps = new Set([scratch, config.history.dir].flatMap((dir) => existsSync(dir) ? readdirSync(dir).map((name) => name.slice(0, 16)) : []));
 	let runId = newRunId(startedAt, git?.sha ?? null, who);
 	for (let n = 1; stamps.has(runId.slice(0, 16)); n++) runId = newRunId(new Date(startedAt.getTime() + n * 1e3), git?.sha ?? null, who);
@@ -1629,7 +1562,9 @@ async function startRun({ config, profile, runs, tags, concurrency, verbosity, c
 		git,
 		worktreePatch: patch,
 		author: who,
-		scratch
+		scratch,
+		previous: null,
+		previousArtifacts: null
 	};
 }
 function noMatch(config, filters, tags) {
@@ -1646,8 +1581,20 @@ function unmatchedFilter(config, parts) {
 */
 async function executeRun(ctx, setup, filters, renderer) {
 	const { runId, stagingDir, config, runs, tags, concurrency, profile, verbosity, configFile } = ctx;
+	const run = JSON.parse(JSON.stringify({
+		runId,
+		stagingDir,
+		config,
+		runs,
+		tags,
+		concurrency,
+		profile,
+		verbosity,
+		configFile,
+		completed: []
+	}));
 	const { runEngine } = await import("./node-Co7EXge0.js");
-	const { buildPlan, checkAttemptBudget, planDocument, planLine, planTable } = await import("./plan-CsPE24pp.js");
+	const { buildPlan, checkAttemptBudget, planDocument, planLine, planTable } = await import("./plan-PPc4-6xd.js");
 	let plan;
 	let sandboxVersions = { images: [] };
 	let preflightStop;
@@ -1660,11 +1607,23 @@ async function executeRun(ctx, setup, filters, renderer) {
 	const onPlan = async ({ planned, excluded, matched, unmatched }) => {
 		if (unmatched.length) throw unmatchedFilter(config, unmatched);
 		if (matched === 0) throw noMatch(config, filters, tags);
-		toRun = planned;
-		for (const p of planned) for (const name of p.attempt.secretVars) registerSecret(process.env[name]);
-		const models = planned.some((p) => !p.attempt.plan.skip && p.attempt.plan.questions?.length) ? await import("./model-BeoRxCcJ.js") : void 0;
+		const completed = completedKeysFor(planned, stagingDir, ctx.previous);
+		const skipped = planned.map((p) => completed.has(p.key) ? {
+			...p,
+			attempt: {
+				...p.attempt,
+				skip: "already completed in this run",
+				plan: {
+					...p.attempt.plan,
+					skip: "already completed in this run"
+				}
+			}
+		} : p);
+		toRun = skipped;
+		for (const p of skipped) for (const name of p.attempt.secretVars) registerSecret(process.env[name]);
+		const models = skipped.some((p) => !p.attempt.plan.skip && p.attempt.plan.questions?.length) ? await import("./model-DMOraCcm.js") : void 0;
 		plan = await buildPlan({
-			planned,
+			planned: skipped,
 			profile,
 			budget: setup.budget,
 			concurrency,
@@ -1672,8 +1631,14 @@ async function executeRun(ctx, setup, filters, renderer) {
 			historyDir: config.history.dir,
 			questionModel: models?.questionModelLabel(config) ?? null
 		});
-		const { imageStages, imageTargets } = await import("./images-B8Mp_VXZ.js");
-		const targets = imageTargets(planned.filter((p) => !p.attempt.plan.skip).map((p) => ({
+		plan.runId = runId;
+		if (completed.size) plan.resume = {
+			runId,
+			completed: completed.size
+		};
+		run.completed = [...completed];
+		const { imageStages, imageTargets } = await import("./images-CscddrJU.js");
+		const targets = imageTargets(skipped.filter((p) => !p.attempt.plan.skip).map((p) => ({
 			id: p.id,
 			harness: p.attempt.harness,
 			sandbox: p.attempt.sandbox
@@ -1685,16 +1650,16 @@ async function executeRun(ctx, setup, filters, renderer) {
 				evals: t.evals,
 				stages: await imageStages(t, config.root)
 			})));
-			const doc = planDocument(plan, planned, excluded, images);
+			const doc = planDocument(plan, skipped, excluded, images);
 			process.stdout.write(`${JSON.stringify(doc, null, 2)}\n`);
 			return false;
 		}
-		const lines = [planLine(plan, setup.configElsewhere ? configFile : void 0)];
+		const lines = [planLine(plan, setup.configElsewhere ? configFile : void 0, runId)];
 		if (setup.dryRun) {
 			const todos = excluded.filter((e) => e.mode === "todo").map((e) => e.id);
 			lines.push(...planTable(plan, todos));
-			const { imageStatusLines } = await import("./images-B8Mp_VXZ.js");
-			const { dockerAvailable } = await import("./src-CDiGyoDt.js");
+			const { imageStatusLines } = await import("./images-CscddrJU.js");
+			const { dockerAvailable } = await import("./src-BT4dB5CO.js");
 			if (targets.length && await dockerAvailable()) lines.push(...await imageStatusLines(targets, config.root));
 			else if (targets.length) lines.push("images  not checked: Docker is not running");
 		}
@@ -1708,7 +1673,7 @@ async function executeRun(ctx, setup, filters, renderer) {
 		});
 		if (setup.dryRun) return false;
 		checkAttemptBudget(plan, setup.force, config.file);
-		const analysis = config.analysis.agent ? await import("./model-BeoRxCcJ.js") : null;
+		const analysis = config.analysis.agent ? await import("./model-DMOraCcm.js") : null;
 		analysis?.analysisModelName(config, "agent");
 		const questionModels = /* @__PURE__ */ new Set();
 		const asked = [];
@@ -1716,7 +1681,8 @@ async function executeRun(ctx, setup, filters, renderer) {
 			if (row.plan.skip || !models) continue;
 			const questions = (row.plan.questions ?? []).map((q) => ({
 				question: q.question,
-				model: models.questionModel(config, q, row.id)
+				model: models.questionModel(config, q, row.id),
+				...q.across ? { across: q.across } : {}
 			}));
 			for (const q of questions) if (typeof q.model === "string") questionModels.add(q.model);
 			if (questions.length) asked.push({
@@ -1737,13 +1703,13 @@ async function executeRun(ctx, setup, filters, renderer) {
 		});
 		if (malformed.length) throw new UsageError(`Credentials with whitespace or a line break in them: ${malformed.join(", ")}. A key copied from a wrapped terminal line picks these up; set it again as one line.`);
 		if (analysis) {
-			const { startNotes } = await import("./agent-DKuFnZPM.js");
+			const { startNotes } = await import("./agent-B72lZ6w5.js");
 			notes = startNotes({
 				config,
 				model: analysis.resolveAnalysisModel(config, "agent"),
 				runId,
 				stagingDir,
-				planned,
+				planned: skipped,
 				warn: (text) => renderer.onEvent({
 					type: "notice",
 					at: Date.now(),
@@ -1753,7 +1719,7 @@ async function executeRun(ctx, setup, filters, renderer) {
 			});
 		}
 		if (asked.length) {
-			const { questionRunner } = await import("./questions-ZdT3JDyQ.js");
+			const { questionRunner } = await import("./questions-BEPj98fZ.js");
 			questions = questionRunner({
 				evals: asked,
 				stagingDir,
@@ -1770,8 +1736,8 @@ async function executeRun(ctx, setup, filters, renderer) {
 				}
 			});
 		}
-		const { prepareWorkspaces, workspaceKey } = await import("./workspace-U6M1gKzo.js");
-		const live = planned.filter((p) => !p.attempt.plan.skip);
+		const { prepareWorkspaces, workspaceKey } = await import("./workspace-DzNL-3iO.js");
+		const live = skipped.filter((p) => !p.attempt.plan.skip);
 		const outcomes = await prepareWorkspaces(live.flatMap(({ attempt: a }) => a.workspace ? [{
 			workspace: a.workspace,
 			environment: a.cell.environment
@@ -1816,7 +1782,7 @@ async function executeRun(ctx, setup, filters, renderer) {
 			});
 		}
 		try {
-			sandboxVersions = await preflight(planned.filter((p) => !p.attempt.plan.skip).map((p) => ({
+			sandboxVersions = await preflight(skipped.filter((p) => !p.attempt.plan.skip).map((p) => ({
 				harness: p.attempt.harness,
 				sandbox: p.attempt.sandbox,
 				setBy: p.attempt.setBy
@@ -1855,20 +1821,11 @@ async function executeRun(ctx, setup, filters, renderer) {
 	let engineRun;
 	try {
 		engineRun = await runEngine({
-			run: JSON.parse(JSON.stringify({
-				runId,
-				stagingDir,
-				config,
-				runs,
-				tags,
-				concurrency,
-				profile,
-				verbosity,
-				configFile
-			})),
+			run,
 			filters,
 			onPlan,
 			onRunStart: (files) => {
+				writeCurrent(setup.config.root, runId);
 				live = writeLiveProgress({
 					dir: stagingDir,
 					runId,
@@ -1882,7 +1839,8 @@ async function executeRun(ctx, setup, filters, renderer) {
 					files,
 					harness: (plan?.harnesses ?? config.harness.map(harnessId)).join(", "),
 					sandbox: sandboxId(config.sandbox),
-					concurrency
+					concurrency,
+					runId
 				});
 			},
 			onEvent: (e) => {
@@ -1926,9 +1884,9 @@ async function executeRun(ctx, setup, filters, renderer) {
 		throw error;
 	} finally {
 		clearTimeout(budgetTimer);
-		const { removePrepared } = await import("./preflight-DKZjb7ZH.js");
+		const { removePrepared } = await import("./preflight-Zumy0adT.js");
 		removePrepared(config.root, runId);
-		const { removePreparedWorkspaces } = await import("./workspace-U6M1gKzo.js");
+		const { removePreparedWorkspaces } = await import("./workspace-DzNL-3iO.js");
 		removePreparedWorkspaces(config.root, runId);
 	}
 	if (preflightStop) throw preflightStop;
@@ -2011,4 +1969,4 @@ run(process.argv.slice(2)).finally(() => {
 //#endregion
 export {};
 
-//# sourceMappingURL=cli-BIhzl_v_.js.map
+//# sourceMappingURL=cli-BMjTKY-D.js.map