reeval 0.11.0 → 0.13.0

package/dist/{cli-ClpPT8kn.js → cli-BCE1xlm4.js}

@@ -1,28 +1,30 @@
 import { n as isOneOf } from "./object-CLf2-tNq.js";
 import { n as UsageError } from "./errors-B6BSMs03.js";
 import { p as parseDuration } from "./shared-CA61BuPJ.js";
-import { f as sandboxId, s as harnessId } from "./options-C0d0AzeR.js";
-import { A as workspaceLabel, p as environmentSandbox } from "./environment-Cca6a4wB.js";
+import { p as sandboxId, s as harnessId, u as plannedSandboxes } from "./options-CyAyLjgJ.js";
+import { M as workspaceLabel, T as profileSandbox, p as environmentSandbox } from "./environment-B4IuURK1.js";
 import { a as fromCwd, o as newRunId } from "./ids-Bp6ut7ZD.js";
-import { _ as selectProfile, a as configLabel, d as hostCapacity, f as loadConfig, g as resolveConcurrency, h as requireConfig, l as findConfig, m as missingCredentials, o as configPathFor, p as malformedCredentials, u as hasAnalysisModel, v as inCI, y as loadEnvFiles } from "./config-DrVSEVN0.js";
-import { o as redactSecrets, s as registerSecret } from "./redact-DJ9JCShu.js";
-import { o as PACKAGE_VERSION } from "./root-DICDHCtp.js";
-import { I as LABEL, X as imageInfo } from "./src-ftu5wPSt.js";
+import { c as registerSecret, s as redactSecrets } from "./redact-CRGdfV2s.js";
+import { a as configPathFor, c as findConfig, d as loadConfig, f as malformedCredentials, g as selectProfile, h as resolveConcurrency, i as configLabel, l as hasAnalysisModel, m as requireConfig, p as missingCredentials, u as hostCapacity } from "./config-D7TFvUpL.js";
+import { n as loadEnvFiles, t as inCI } from "./env-BQqj93gu.js";
+import { W as LABEL, at as imageInfo } from "./src-CIpKL45U.js";
+import { o as PACKAGE_VERSION } from "./root-DNuLhaA5.js";
+import { C as pluginBuildsLabel } from "./plugins-P-uh54VF.js";
 import { s as debugTime } from "./debug-D24NOQH1.js";
-import { i as sandboxWarningLine, n as clip, r as formatDuration } from "./text-gEcKjAyp.js";
+import { i as formatDuration, n as clip, o as sandboxWarningLine } from "./text-CrVy5ioL.js";
 import { t as PrepareError } from "./provider-CmAYTQQs.js";
-import { T as runOutcome, g as formatUsage } from "./model-DaX4bNLZ.js";
-import { s as suiteResult } from "./suite-Cw_7G2RM.js";
+import { T as runOutcome, g as formatUsage } from "./model-UgHDMhI1.js";
+import { c as suiteResult } from "./suite-DmuF_ApO.js";
+import { a as commandHelpText, n as FLAGS, r as GLOBAL, s as helpText, t as COMMANDS } from "./cli-spec-BXIBUlaK.js";
+import { p as render } from "./summary-BFMHMTz2.js";
 import { n as stripAnsi, t as colors } from "./colors-QV9dFVhN.js";
-import { a as render } from "./summary-BUKfU-9l.js";
-import { c as writeLiveProgress, i as prepareResume, o as resolveResume, r as loadPrevious, s as writeCurrent, t as completedKeysFor } from "./resume-Ca1c8vix.js";
-import { a as commandHelpText, n as FLAGS, r as GLOBAL, s as helpText, t as COMMANDS } from "./cli-spec-qmEnQ25_.js";
-import { i as runtimeInfo, n as gitInfo, r as runnerInfo, t as author } from "./provenance-DQ0AXn6c.js";
+import { i as runtimeInfo, n as gitInfo, r as runnerInfo, t as author } from "./provenance-C_-8y-x3.js";
 import { t as budgetMarker } from "./budget-DBUrH4jX.js";
-import { n as prepareAll, o as writePrepared, r as prepareTargets, t as RunStopError } from "./preflight-BRzD30L5.js";
-import { n as imageHarness } from "./prepare-B_y-AFMv.js";
-import { r as createProviders, t as attemptImage } from "./providers-DoTq1jHW.js";
-import { a as redactArgv, r as commandOf, t as logInvocation } from "./invocations-BYDPe3fV.js";
+import { c as writeLiveProgress, i as prepareResume, o as resolveResume, r as loadPrevious, s as writeCurrent, t as completedKeysFor } from "./resume-8-nnW8iO.js";
+import { n as prepareAll, o as writePrepared, r as prepareTargets, t as RunStopError } from "./preflight-D3HvKeAx.js";
+import { n as imageHarness } from "./prepare-Dx7dOnJB.js";
+import { i as createProviders, n as attemptImage } from "./providers-CF5UEaP8.js";
+import { a as redactArgv, r as commandOf, t as logInvocation } from "./invocations-D_ePcwyL.js";
 import { existsSync, mkdirSync, readdirSync, writeFileSync } from "node:fs";
 import { dirname, join } from "node:path";
 import { inspect } from "node:util";
@@ -161,12 +163,15 @@ function usageError(error, command) {
 }
 //#endregion
 //#region src/preflight.ts
-/** How the terminal names a sandbox to prepare: its cell id, `+ install`, and a harness other than Cursor. */
-function prepareLabel({ sandbox, harness }) {
+/**
+* How the terminal names a sandbox to prepare: its cell id, `+ install`, `+ plugin builds`, and a
+* harness other than Cursor.
+*/
+function prepareLabel({ sandbox, harness, plugins }) {
 	if (sandbox.kind !== "docker") return sandboxId(sandbox);
 	const baked = imageHarness(harness);
 	const install = sandbox.install?.length ? " + install" : "";
-	return `${sandboxId(sandbox)}${install}${baked === "cursor" ? "" : ` · ${baked}`}`;
+	return `${sandboxId(sandbox)}${install}${pluginBuildsLabel(plugins)}${baked === "cursor" ? "" : ` · ${baked}`}`;
 }
 /** A run that stopped in preflight on an infra failure: exit 3. */
 var PreflightInfraError = class extends Error {
@@ -384,12 +389,12 @@ function resultLines(c, e, verbose) {
 			`results ${e.summaryPath}`
 		].filter(Boolean).join(" · "))
 	];
-	return [...e.unhandledErrors.flatMap((error) => [c.red(`error: ${error}`), ""]), ...render(suiteResult(record), null, {
+	return render(suiteResult(record), e.insights ?? null, {
 		color: c.enabled,
 		verbose,
-		notes,
+		notes: [...(e.insightLines ?? []).map((l) => c.yellow(l)), ...notes],
 		hint: e.hint
-	})];
+	});
 }
 //#endregion
 //#region src/ui/tags.ts
@@ -552,6 +557,8 @@ var PlainRenderer = class {
 	#evals = /* @__PURE__ */ new Set();
 	#tries = /* @__PURE__ */ new Set();
 	#ended = /* @__PURE__ */ new Set();
+	/** Insights after the last attempt, once they've started: they take the status line. */
+	#insights = null;
 	constructor(options) {
 		this.#o = options;
 		this.#c = options.colors;
@@ -568,11 +575,16 @@ var PlainRenderer = class {
 		this.#o.out.write(this.#status ? `${CLEAR}${line}${this.#status}` : line);
 		this.#lastWrite = this.#now();
 	}
-	/** `7 evals · 12/28 tries`: what's collected so far and how many tries ended. */
+	/** `7 evals · 12/28 tries`: what's collected so far and how many tries ended; then `insights 3/7…`. */
 	#drawStatus() {
-		if (!this.#live || !this.#tries.size) return;
-		const evals = `${this.#evals.size} eval${this.#evals.size === 1 ? "" : "s"}`;
-		const status = this.#c.dim(`${evals} · ${this.#ended.size}/${this.#tries.size} tries`);
+		if (!this.#live) return;
+		let status;
+		if (this.#insights) status = this.#c.dim(`insights ${this.#insights.done}/${this.#insights.total}…`);
+		else {
+			if (!this.#tries.size) return;
+			const evals = `${this.#evals.size} eval${this.#evals.size === 1 ? "" : "s"}`;
+			status = this.#c.dim(`${evals} · ${this.#ended.size}/${this.#tries.size} tries`);
+		}
 		if (status === this.#status) return;
 		this.#o.out.write(`${CLEAR}${status}`);
 		this.#status = status;
@@ -700,6 +712,14 @@ var PlainRenderer = class {
 					this.#write(`::error title=${title}::${ghEscape(failure.message)}`);
 				}
 				return;
+			case "insights-progress":
+				this.#insights = {
+					done: e.done,
+					total: e.total,
+					startedAt: this.#insights?.startedAt ?? e.at
+				};
+				this.#drawStatus();
+				return;
 			case "notice": {
 				const paint = {
 					info: c.dim,
@@ -760,12 +780,19 @@ var PlainRenderer = class {
 			this.#write(`::error file=${ghEscape(result.file, true)}${where},title=${title}::${message}`);
 		}
 	}
-	/** Prints the heartbeat if nothing was written for `heartbeatMs` and evals are running. */
+	/** Prints the heartbeat if nothing was written for `heartbeatMs` and evals or insights are running. */
 	tick(now = this.#now()) {
 		const { heartbeatMs } = this.#o.output;
-		if (this.#live || heartbeatMs <= 0 || this.#running.size === 0) return;
+		if (this.#live || heartbeatMs <= 0) return;
 		if (now - this.#lastWrite < heartbeatMs) return;
 		const c = this.#c;
+		const insights = this.#insights;
+		if (insights) {
+			const elapsed = formatDuration(now - insights.startedAt);
+			if (insights.done < insights.total) this.#write(c.dim(`… insights ${insights.done}/${insights.total} after ${elapsed}`));
+			return;
+		}
+		if (this.#running.size === 0) return;
 		const running = [...this.#running.values()].map((a) => {
 			const progress = a.last?.type === "eval-progress" ? a.last.progress : void 0;
 			return `${this.#prefix(a)} ${phaseOf(progress)} ${c.dim(formatDuration(now - a.startedAt))}`;
@@ -899,6 +926,11 @@ const invocation = {
 let interrupted = false;
 /** The last signal that cancelled the run: 130 after SIGINT, 143 after SIGTERM. */
 let stopCode = EXIT.interrupted;
+/**
+* Set while insights run, after the attempts and the exit code: Ctrl-C or SIGTERM then stops the
+* insights instead of cancelling the run, which has nothing left to cancel.
+*/
+let onInsightsSignal = null;
 async function run(argv) {
 	const { values, positionals } = parseCli(argv);
 	invocation.flags = values;
@@ -916,7 +948,7 @@ async function run(argv) {
 		return values.help ? EXIT.ok : EXIT.usage;
 	}
 	if (command === "init") {
-		const { initCommand } = await import("./init-fhOXED8F.js");
+		const { initCommand } = await import("./init-kMkTgTZc.js");
 		return initCommand(values, filters, {
 			out: (s) => console.log(s),
 			input: process.stdin,
@@ -925,7 +957,7 @@ async function run(argv) {
 		});
 	}
 	if (command === "doctor") {
-		const { doctorCommand } = await import("./doctor-D2MbNO8L.js");
+		const { doctorCommand } = await import("./doctor-De_QlOTM.js");
 		return doctorCommand(values, {
 			out: (s) => console.log(s),
 			color: !!process.stdout.isTTY && !process.env.NO_COLOR,
@@ -933,7 +965,7 @@ async function run(argv) {
 		});
 	}
 	if (command === "report") {
-		const { reportCommand } = await import("./command-DbAe1Rzb.js");
+		const { reportCommand } = await import("./command-CANun0m0.js");
 		return reportCommand(values, filters, {
 			out: (s) => console.log(s),
 			err: (s) => console.error(dim(s)),
@@ -948,7 +980,7 @@ async function run(argv) {
 	}
 	if (command === "studio") {
 		const config = await loadConfig(configPath, "studio");
-		const { studioCommand } = await import("./command-BlM0y8jN.js");
+		const { studioCommand } = await import("./command-Bh3yhN_b.js");
 		return studioCommand({
 			config,
 			flags: values,
@@ -959,7 +991,7 @@ 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-TgbzPizb.js");
+		const { mcpCommand } = await import("./command-D_kEnYcX.js");
 		return mcpCommand({
 			config,
 			flags: values,
@@ -968,7 +1000,7 @@ async function run(argv) {
 	}
 	if (command in HISTORY_COMMANDS) {
 		const config = await loadConfig(configPath, command);
-		const commands = await import("./commands-CthqwkAn.js");
+		const commands = await import("./commands-BKGzBMgb.js");
 		const ctx = {
 			config,
 			flags: values,
@@ -981,7 +1013,7 @@ 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-CRQfsw0e.js");
+				const { queryCommand } = await import("./query-D_1ARx5q.js");
 				return queryCommand({
 					...ctx,
 					err: (s) => console.error(dim(s))
@@ -989,7 +1021,7 @@ async function run(argv) {
 			}
 			case "analyze": {
 				loadEnvFiles(configPath);
-				const { analyzeCommand } = await import("./command-C-cVjlny.js");
+				const { analyzeCommand } = await import("./command-CBaHukyl.js");
 				return analyzeCommand({
 					...ctx,
 					err: (s) => console.error(dim(s))
@@ -1028,8 +1060,8 @@ 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 ${where}`);
 	if (flags.images) {
-		const { dockerAvailable, pruneImages } = await import("./src-DYO3DdnQ.js");
-		const { requireDocker } = await import("./providers-B9yTdIqF.js");
+		const { dockerAvailable, pruneImages } = await import("./src-D9fSTxlc.js");
+		const { requireDocker } = await import("./providers-CTBYvMxu.js");
 		const engines = [];
 		for (const engine of ["docker", "podman"]) if (await dockerAvailable(engine)) engines.push(engine);
 		if (!engines.length) await requireDocker();
@@ -1050,7 +1082,7 @@ function verbosityOf(flags) {
 * does. Needs a config, as `reeval run` does: there is no default image.
 */
 async function imageBuildCommand(flags, configPath, filters) {
-	const { buildImages, imagePlanLines, imageTargets, plannedImageUses } = await import("./images-Bi3aGxb9.js");
+	const { buildImages, imagePlanLines, imageTargets, plannedImageUses } = await import("./images-m784ueGK.js");
 	const { runId, stagingDir, config, runs, tags, concurrency, profile } = await startRun(await prepareRun({
 		...flags,
 		"dry-run": true
@@ -1110,7 +1142,7 @@ async function runCommand(flags, configPath, filters) {
 	if (githubOut) await appendFile(githubOut, `run-id=${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-B3ydHo6t.js");
+	const { engineVersions } = await import("./node-CGZYh4Fj.js");
 	const renderer = createRenderer({
 		out: setup.json ? setup.dryRun ? { write: () => true } : process.stderr : process.stdout,
 		output: setup.output,
@@ -1128,27 +1160,13 @@ async function runCommand(flags, configPath, filters) {
 		});
 		throw error;
 	}
-	const { engineRun, plan, sandboxVersions, budgetReached, live, notes, questions } = executed;
+	const { engineRun, plan, sandboxVersions, budgetReached, live, questions } = executed;
 	invocation.unhandledErrors = engineRun.unhandledErrors;
 	const engines = [...new Set(plan?.sandboxes.flatMap((s) => s.kind === "docker" ? [s.engine ?? "docker"] : []))];
 	if ((interrupted || engineRun.cancelled) && engines.length) {
-		const { removeRunContainers } = await import("./docker-DNmMPOBb.js");
+		const { removeRunContainers } = await import("./docker-Dac6Zalz.js");
 		await removeRunContainers(ctx.runId, engines);
 	}
-	if (notes) {
-		const cancelled = interrupted || engineRun.cancelled;
-		const written = await notes.finish(cancelled);
-		if (!cancelled && written.length) {
-			const { notesLine } = await import("./agent-D483ru7D.js");
-			const text = notesLine(ctx.config.root, ctx.runId, written);
-			renderer.onEvent({
-				type: "notice",
-				at: Date.now(),
-				level: "info",
-				text
-			});
-		}
-	}
 	if (engineRun.stopped || !plan) {
 		live?.remove();
 		renderer.close();
@@ -1156,8 +1174,8 @@ 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-DpAxYyED.js");
-	const { writeScratchRun } = await import("./results-BE2f-1XS.js");
+	const { lockfiles } = await import("./state-BGWGAwdo.js");
+	const { writeScratchRun } = await import("./results-D-naai_Q.js");
 	const run = await writeScratchRun({
 		ctx,
 		author: ctx.author,
@@ -1181,17 +1199,56 @@ async function runCommand(flags, configPath, filters) {
 		...answers?.size ? { answers } : {}
 	}, () => live?.stop());
 	const { record } = run;
-	if (notes && !stopped) {
-		const { summarizeRun } = await import("./agent-D483ru7D.js");
-		const { resolveAnalysisModel } = await import("./model-CXhrziBe.js");
-		const model = resolveAnalysisModel(ctx.config, "agent");
-		const { level, text } = await summarizeRun(ctx.config, record.runId, model);
-		renderer.onEvent({
-			type: "notice",
-			at: Date.now(),
-			level,
-			text
-		});
+	const outcome = runOutcome(record);
+	const code = stopped ? stopCode : outcome === "fail" ? EXIT.failed : outcome === "infra_error" ? EXIT.infra : EXIT.ok;
+	let insights = run.insights;
+	let insightLines = [];
+	const { runInsights, wantsInsights } = await import("./run-Bel7Vde4.js");
+	const wanted = !stopped && wantsInsights(ctx.config, !flags["no-insights"]);
+	if (wanted && interrupted) insightLines = [`insights skipped (${stopCode === EXIT.terminated ? "SIGTERM" : "Ctrl-C"}): reeval analyze ${record.runId} --insights to write them`];
+	else if (wanted) {
+		const { scratchDir } = await import("./files-CtjEXWXm.js");
+		const stop = new AbortController();
+		let stopSignal = "SIGINT";
+		onInsightsSignal = (signal) => {
+			if (stop.signal.aborted) {
+				renderer.close();
+				process.exit(code);
+			}
+			stopSignal = signal;
+			stop.abort(/* @__PURE__ */ new Error(signal === "SIGTERM" ? "SIGTERM" : "Ctrl-C"));
+		};
+		try {
+			const done = await runInsights({
+				config: ctx.config,
+				record,
+				runDir: run.artifactDir ?? join(scratchDir(ctx.config.root), record.runId),
+				remarks: run.insights,
+				verbose: (ctx.verbosity ?? 0) > 0,
+				signal: stop.signal,
+				stopSignal: () => stopSignal,
+				onStart: (text) => {
+					renderer.onEvent({
+						type: "notice",
+						at: Date.now(),
+						level: "info",
+						text
+					});
+				},
+				onProgress: (n, total) => {
+					renderer.onEvent({
+						type: "insights-progress",
+						at: Date.now(),
+						done: n,
+						total
+					});
+				}
+			});
+			insights = done.insights ?? insights;
+			insightLines = done.lines;
+		} finally {
+			onInsightsSignal = null;
+		}
 	}
 	const { pruneDebug } = await import("./files-BAt0i00X.js");
 	await pruneDebug(ctx.config.root, ctx.config.debug.keep).catch(() => {});
@@ -1202,8 +1259,9 @@ async function runCommand(flags, configPath, filters) {
 		record,
 		summaryPath: fromCwd(summary),
 		results: engineRun.evals,
-		unhandledErrors: engineRun.unhandledErrors,
-		hint: !hasAnalysisModel(ctx.config)
+		hint: !hasAnalysisModel(ctx.config),
+		insights,
+		insightLines
 	});
 	renderer.close();
 	const stepSummary = process.env.GITHUB_STEP_SUMMARY;
@@ -1211,7 +1269,7 @@ 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-CthqwkAn.js");
+		const { record: recordCmd } = await import("./commands-BKGzBMgb.js");
 		const allow = record.status === "budget" ? { "allow-incomplete": true } : {};
 		const line = await recordCmd(config, record.runId, {
 			...flags,
@@ -1221,13 +1279,10 @@ async function runCommand(flags, configPath, filters) {
 		else console.log(line);
 	}
 	if (setup.json) {
-		const { suiteResult } = await import("./suite-DamAEnNF.js");
+		const { suiteResult } = await import("./suite-mn9g4veR.js");
 		process.stdout.write(`${JSON.stringify(suiteResult(record), null, 2)}\n`);
 	}
-	if (interrupted || engineRun.cancelled) return stopCode;
-	if (engineRun.unhandledErrors.length > 0) return EXIT.failed;
-	const outcome = runOutcome(record);
-	return outcome === "fail" ? EXIT.failed : outcome === "infra_error" ? EXIT.infra : EXIT.ok;
+	return code;
 }
 /** Everything that can fail before the evals are collected: output mode, config, profile, flags. */
 async function prepareRun(flags, configPath) {
@@ -1253,11 +1308,11 @@ async function prepareRun(flags, configPath) {
 		runs = Number(flags.runs);
 		if (!(Number.isInteger(runs) && runs >= 1)) throw new UsageError(`--runs must be a whole number of at least 1 (got ${flags.runs})`);
 	}
-	const sandboxes = [environmentSandbox(config.environment, config.sandbox), ...profile?.defaults.environment ? [environmentSandbox(profile.defaults.environment, config.sandbox)] : []];
+	const sandboxes = [environmentSandbox(config.environment, config.sandbox), ...profile ? [profileSandbox(config, profile.defaults)] : []];
 	const concurrency = await resolveConcurrency(config, flags.concurrency, async () => {
 		const container = sandboxes.find((s) => s.kind === "docker");
 		if (!container) return hostCapacity();
-		const { dockerCapacity } = await import("./src-DYO3DdnQ.js");
+		const { dockerCapacity } = await import("./src-D9fSTxlc.js");
 		return await dockerCapacity(container.engine) ?? hostCapacity();
 	});
 	return {
@@ -1272,6 +1327,8 @@ async function prepareRun(flags, configPath) {
 		dryRun: !!flags["dry-run"],
 		json: !!flags.json,
 		force: !!flags.force,
+		/** `--no-insights` turns them off for this run. */
+		insights: !flags["no-insights"],
 		verbosity: verbosityOf(flags),
 		configFile: configLabel(path),
 		/** Named in the plan line when the project root isn't the current directory. */
@@ -1284,11 +1341,11 @@ async function prepareRun(flags, configPath) {
 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-DpAxYyED.js");
+	const { worktreePatch } = await import("./state-BGWGAwdo.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-BE2f-1XS.js");
-	const { scratchDir } = await import("./files-Bw8lWEjg.js");
+	const { stagingDir, sweepStaging } = await import("./results-D-naai_Q.js");
+	const { scratchDir } = await import("./files-CtjEXWXm.js");
 	const scratch = scratchDir(config.root);
 	if (resume !== void 0) {
 		const target = await resolveResume(config.root, resume || void 0);
@@ -1299,7 +1356,7 @@ async function startRun({ config, profile, runs, tags, concurrency, verbosity, c
 			artifacts: target.previousArtifacts
 		} : await loadPrevious(stagingDir(config.root, target.runId));
 		if (!dryRun) try {
-			const { removeRunContainers } = await import("./docker-DNmMPOBb.js");
+			const { removeRunContainers } = await import("./docker-Dac6Zalz.js");
 			await removeRunContainers(target.runId, ["docker", "podman"]);
 		} catch {}
 		return {
@@ -1370,8 +1427,8 @@ async function executeRun(ctx, setup, filters, renderer) {
 		configFile,
 		completed: []
 	}));
-	const { runEngine } = await import("./node-B3ydHo6t.js");
-	const { buildPlan, checkAttemptBudget, planDocument, planLine, planTable } = await import("./plan-DgQxp9yy.js");
+	const { runEngine } = await import("./node-CGZYh4Fj.js");
+	const { buildPlan, checkAttemptBudget, planDocument, planInsights, planLine, planTable } = await import("./plan-CZYgXnZh.js");
 	let plan;
 	let sandboxVersions = { images: [] };
 	let preflightStop;
@@ -1379,7 +1436,6 @@ async function executeRun(ctx, setup, filters, renderer) {
 	let budgetTimer;
 	let toRun = [];
 	let live;
-	let notes;
 	let questions;
 	const onPlan = async ({ planned, excluded, matched, unmatched }) => {
 		if (unmatched.length) throw unmatchedFilter(config, unmatched);
@@ -1398,7 +1454,7 @@ async function executeRun(ctx, setup, filters, renderer) {
 		} : 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-CXhrziBe.js") : void 0;
+		const models = skipped.some((p) => !p.attempt.plan.skip && p.attempt.plan.questions?.length) ? await import("./model-B8H3gL69.js") : void 0;
 		plan = await buildPlan({
 			planned: skipped,
 			profile,
@@ -1406,19 +1462,26 @@ async function executeRun(ctx, setup, filters, renderer) {
 			concurrency,
 			root: config.root,
 			historyDir: config.history.dir,
-			questionModel: models?.questionModelLabel(config) ?? null
+			questionModel: models?.questionModelLabel(config) ?? null,
+			insights: await planInsights(config, setup.insights)
 		});
 		plan.runId = runId;
+		if (plan.insights) {
+			const { analysisKeyProblem } = await import("./model-B8H3gL69.js");
+			const problem = analysisKeyProblem(config);
+			if (problem) plan.warnings.push(`insights will be skipped: ${problem}`);
+		}
 		if (completed.size) plan.resume = {
 			runId,
 			completed: completed.size
 		};
 		run.completed = [...completed];
-		const { imageStages, imageTargets } = await import("./images-Bi3aGxb9.js");
+		const { imageStages, imageTargets } = await import("./images-m784ueGK.js");
 		const targets = imageTargets(skipped.filter((p) => !p.attempt.plan.skip).map((p) => ({
 			id: p.id,
 			harness: p.attempt.harness,
-			sandbox: p.attempt.sandbox
+			sandbox: p.attempt.sandbox,
+			plugins: p.attempt.pluginBuilds ?? []
 		})));
 		if (setup.json && setup.dryRun) {
 			const images = await Promise.all(targets.map(async (t) => ({
@@ -1435,8 +1498,8 @@ async function executeRun(ctx, setup, filters, renderer) {
 		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-Bi3aGxb9.js");
-			const { dockerAvailable, engineSpec } = await import("./src-DYO3DdnQ.js");
+			const { imageStatusLines } = await import("./images-m784ueGK.js");
+			const { dockerAvailable, engineSpec } = await import("./src-D9fSTxlc.js");
 			const engines = [...new Set(targets.map((t) => t.sandbox.engine ?? "docker"))];
 			const down = [];
 			for (const engine of engines) if (!await dockerAvailable(engine)) down.push(engine);
@@ -1446,6 +1509,10 @@ async function executeRun(ctx, setup, filters, renderer) {
 				lines.push(`images  not checked: ${names} ${down.length === 1 ? "is" : "are"} not running`);
 			}
 		}
+		if (setup.dryRun && plan.insights) {
+			const { formatTokens } = await import("./usage-BDo2q_yu.js");
+			lines.push(`insights  ${plan.insights.model} · up to ${formatTokens(plan.insights.maxTokens)} tokens`);
+		}
 		lines.push(...plan.warnings.map((w) => `warning: ${w}`));
 		lines.push(...plan.hints.map((h) => `hint: ${h}`));
 		renderer.onEvent({
@@ -1456,8 +1523,6 @@ 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-CXhrziBe.js") : null;
-		analysis?.analysisModelName(config, "agent");
 		const questionModels = /* @__PURE__ */ new Set();
 		const asked = [];
 		for (const row of plan.rows) {
@@ -1480,31 +1545,15 @@ async function executeRun(ctx, setup, filters, renderer) {
 			profile,
 			questionModels: [...questionModels]
 		});
-		if (missing.length) throw new UsageError(`Missing credentials: ${missing.join(", ")}. Set them in the environment or .env.local.`);
+		if (missing.length) throw new UsageError(`Missing credentials: ${missing.join("; ")}. Set them in the environment or .env.local.`);
 		const malformed = malformedCredentials(config, process.env, {
 			harnesses: plan.harnessDescriptors,
 			profile,
 			questionModels: [...questionModels]
 		});
 		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-D483ru7D.js");
-			notes = startNotes({
-				config,
-				model: analysis.resolveAnalysisModel(config, "agent"),
-				runId,
-				stagingDir,
-				planned: skipped,
-				warn: (text) => renderer.onEvent({
-					type: "notice",
-					at: Date.now(),
-					level: "warn",
-					text
-				})
-			});
-		}
 		if (asked.length) {
-			const { questionRunner } = await import("./questions-C2dym2ce.js");
+			const { questionRunner } = await import("./questions-KENpWnfV.js");
 			questions = questionRunner({
 				evals: asked,
 				stagingDir,
@@ -1521,7 +1570,7 @@ async function executeRun(ctx, setup, filters, renderer) {
 				}
 			});
 		}
-		const { prepareWorkspaces, workspaceKey } = await import("./workspace-C_X0tdv6.js");
+		const { prepareWorkspaces, sameTreeArms, workspaceKey } = await import("./workspace-B1eGO46M.js");
 		const live = skipped.filter((p) => !p.attempt.plan.skip);
 		const outcomes = await prepareWorkspaces(live.flatMap(({ attempt: a }) => a.workspace ? [{
 			workspace: a.workspace,
@@ -1551,6 +1600,7 @@ async function executeRun(ctx, setup, filters, renderer) {
 				label: `workspace ${o.label}`,
 				sandbox: user.attempt.sandbox,
 				harness: user.attempt.harness,
+				plugins: [],
 				setBy: o.setBy,
 				runWide: false
 			};
@@ -1566,10 +1616,17 @@ async function executeRun(ctx, setup, filters, renderer) {
 				}
 			});
 		}
+		for (const text of sameTreeArms(live, outcomes)) renderer.onEvent({
+			type: "notice",
+			at: Date.now(),
+			level: "warn",
+			text
+		});
 		try {
 			sandboxVersions = await preflight(skipped.filter((p) => !p.attempt.plan.skip).map((p) => ({
 				harness: p.attempt.harness,
 				sandbox: p.attempt.sandbox,
+				plugins: p.attempt.pluginBuilds ?? [],
 				setBy: p.attempt.setBy
 			})), {
 				root: config.root,
@@ -1623,7 +1680,7 @@ async function executeRun(ctx, setup, filters, renderer) {
 					version,
 					files,
 					harness: (plan?.harnesses ?? config.harness.map(harnessId)).join(", "),
-					sandbox: sandboxId(config.sandbox),
+					sandbox: plannedSandboxes(plan?.sandboxes ?? [], profileSandbox(config, profile?.defaults)),
 					concurrency,
 					runId
 				});
@@ -1635,7 +1692,6 @@ async function executeRun(ctx, setup, filters, renderer) {
 				}
 				if (e.type === "eval-end") {
 					live?.end(e.result);
-					notes?.onResult(e.result);
 					questions?.onEvalEnd(e.result);
 				}
 				if (interrupted && e.type === "eval-end" && e.result.state === "skipped") return;
@@ -1643,6 +1699,7 @@ async function executeRun(ctx, setup, filters, renderer) {
 			},
 			onCancellable: (cancel) => {
 				const onSignal = (signal) => {
+					if (onInsightsSignal) return onInsightsSignal(signal === "SIGTERM" ? "SIGTERM" : "SIGINT");
 					stopCode = signal === "SIGTERM" ? EXIT.terminated : EXIT.interrupted;
 					if (interrupted) {
 						renderer.close();
@@ -1669,9 +1726,9 @@ async function executeRun(ctx, setup, filters, renderer) {
 		throw error;
 	} finally {
 		clearTimeout(budgetTimer);
-		const { removePrepared } = await import("./preflight-DlQvbySb.js");
+		const { removePrepared } = await import("./preflight-BQlPTHVu.js");
 		removePrepared(config.root, runId);
-		const { removePreparedWorkspaces } = await import("./workspace-C_X0tdv6.js");
+		const { removePreparedWorkspaces } = await import("./workspace-B1eGO46M.js");
 		removePreparedWorkspaces(config.root, runId);
 	}
 	if (preflightStop) throw preflightStop;
@@ -1684,7 +1741,6 @@ async function executeRun(ctx, setup, filters, renderer) {
 		sandboxVersions,
 		budgetReached: () => reached,
 		live,
-		notes,
 		questions
 	};
 }
@@ -1754,4 +1810,4 @@ run(process.argv.slice(2)).finally(() => {
 //#endregion
 export {};
 
-//# sourceMappingURL=cli-ClpPT8kn.js.map
+//# sourceMappingURL=cli-BCE1xlm4.js.map