@tangle-network/agent-eval 0.202.1 → 0.203.0

package/dist/{report-command-DzPpFgjB.js → report-command-D0XL2Cvi.js}

@@ -1,3 +1,4 @@
+import { a as hashCanonical } from "./canonical-BsZVgBhW.js";
 import { o as summarizeNumberSeries } from "./descriptive-D9LUFk1-.js";
 import { a as parseJson, c as workerSourceKey, d as SUPERVISOR_RUN_SCHEMA, f as isUnavailable, i as asRecord, l as NO_SOURCE_LIMITS, m as unavailable, o as parseJsonl, p as showMeasured, r as readTerminalRecord, s as parseSupervisorTree, u as SUPERVISOR_RUN_ROLLUP_SCHEMA } from "./terminal-record-BHocwFw1.js";
 import { basename, join, resolve } from "node:path";
@@ -322,7 +323,18 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
 			unknownNodes: closeUsdUnreported && rootId !== null ? [rootId] : []
 		}
 	};
+	const noteCaptureGaps = (owner, captures) => {
+		if (captures === null || captures?.length === 0) gaps.push(`${owner} workspaceCapture: no retained output capture evidence`);
+		else if (captures !== void 0) for (const [index, capture] of captures.entries()) {
+			const source = `${owner} workspaceCapture[${index + 1}]`;
+			if (capture.projectionGap !== null) gaps.push(`${source}: ${capture.projectionGap}`);
+			if (capture.coverageComplete === false) gaps.push(`${source}: ${capture.incompleteReason ?? "provider reported incomplete coverage"}`);
+			if (capture.coverageComplete === null && capture.projectionGap === null) gaps.push(`${source}: capture coverage was not recorded`);
+		}
+	};
+	noteCaptureGaps(`root ${rootId ?? src.runRef}`, src.rootWorkspaceCaptures);
 	const perWorker = (src.workers ?? []).map((w) => {
+		noteCaptureGaps(`worker ${w.workerId ?? w.label}`, w.workspaceCaptures);
 		const f = tree.workerLogs.get(workerSourceKey(w));
 		const matchingSpawns = spawnsForSource(w);
 		const spawn = spawnForSource(w);
@@ -347,7 +359,8 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
 			usd: usdLimit !== null ? null : w.workerId === void 0 ? workerUsdByLabel.get(w.label) ?? null : workerUsdById.get(w.workerId) ?? null,
 			patchBytes: w.patchBytes ?? f?.finishedPatchBytes ?? null,
 			passed,
-			score: close?.score ?? f?.score ?? null
+			score: close?.score ?? f?.score ?? null,
+			workspaceCaptures: w.workspaceCaptures ?? null
 		};
 	});
 	const wallDistribution = summarizeNumberSeries(perWorker.map((w) => w.wallMs).filter((w) => w !== null));
@@ -408,6 +421,7 @@ function analyzeSupervisorRunSources(src, now = Date.now) {
 		decision,
 		economics,
 		outcome,
+		rootWorkspaceCaptures: src.rootWorkspaceCaptures ?? null,
 		gaps,
 		traceCommand: src.traceCommand ?? "npx --yes @tangle-network/traces@latest analyze --harness opencode --cwd <worker-clone-cwd>"
 	};
@@ -750,7 +764,7 @@ function renderSupervisorRollupMarkdown(rollup, title = "Round rollup") {
 *
 * Transcripts come from Runtime's own records too: `root-stream.jsonl` is the root's, each
 * worker's turn outputs and native-session receipt are read from the journal, and every blob
-* a record names is checked on disk (see `workerTranscripts`).
+* a record names is checked on disk (see `nodeTranscripts`).
 *
 * The run's terminal record is Runtime's own: `result.json` is the
 * `SupervisedResult` that `supervise()` returned, verbatim, and its `kind`
@@ -1043,13 +1057,87 @@ function runtimeBeginState(root, startedAt) {
 function blobFile(runDir, ref) {
 	return typeof ref === "string" && /^sha256:[0-9a-f]{64}$/u.test(ref) ? join(runDir, BLOB_DIR, `${ref.replace(":", "-")}.json`) : null;
 }
+async function workspaceCapturesFromOutput(runDir, ref) {
+	const outRef = typeof ref === "string" ? ref : null;
+	const blobPath = blobFile(runDir, ref);
+	const missing = (projectionGap) => ({
+		outRef,
+		blobPath,
+		receiptPointer: null,
+		receipt: null,
+		attempts: null,
+		coverageComplete: null,
+		incompleteReason: null,
+		projectionGap
+	});
+	if (blobPath === null) return [missing("Runtime output has no valid content address")];
+	if (!await isFile(blobPath)) return [missing("Runtime output blob is missing or not a file")];
+	const raw = await readMaybe(blobPath);
+	if (raw === null) return [missing("Runtime output blob is missing")];
+	let output;
+	try {
+		output = JSON.parse(raw);
+	} catch {
+		return [missing("Runtime output blob is not JSON")];
+	}
+	let addressGap = null;
+	try {
+		if (hashCanonical(output) !== outRef) addressGap = "Runtime output content address does not match output blob";
+	} catch {
+		addressGap = "Runtime output cannot be canonically hashed";
+	}
+	const project = (rawReceipt, receiptPointer) => {
+		const receipt = record(rawReceipt);
+		if (receipt === null) return {
+			...missing(`workspaceCapture receipt is absent or malformed${addressGap === null ? "" : `; ${addressGap}`}`),
+			receiptPointer,
+			receipt: rawReceipt
+		};
+		const provenance = record(receipt.provenance);
+		const attempts = Array.isArray(provenance?.attempts) ? provenance.attempts : null;
+		const coverageComplete = typeof receipt.coverageComplete === "boolean" ? receipt.coverageComplete : null;
+		const incompleteReason = typeof receipt.incompleteReason === "string" ? receipt.incompleteReason : null;
+		const gaps = [];
+		if (addressGap !== null) gaps.push(addressGap);
+		if (coverageComplete === null) gaps.push("workspaceCapture.coverageComplete is absent");
+		if (record(receipt.snapshot) === null) gaps.push("workspaceCapture.snapshot is absent");
+		if (provenance === null) gaps.push("workspaceCapture.provenance is absent");
+		if (attempts === null) gaps.push("workspaceCapture.provenance.attempts is absent");
+		else if (attempts.length === 0 || attempts.some((attempt) => record(attempt) === null)) gaps.push("workspaceCapture.provenance.attempts is empty or malformed");
+		return {
+			outRef,
+			blobPath,
+			receiptPointer,
+			receipt,
+			attempts,
+			coverageComplete,
+			incompleteReason,
+			projectionGap: gaps.length > 0 ? gaps.join("; ") : null
+		};
+	};
+	const result = record(output);
+	const captures = [];
+	for (const key of Object.keys(result ?? {})) if (key === "workspaceCapture") captures.push(project(result?.[key], "/workspaceCapture"));
+	else if (key === "workspaceCaptures") {
+		const receipts = result?.[key];
+		if (Array.isArray(receipts) && receipts.length > 0) for (const [index, receipt] of receipts.entries()) captures.push(project(receipt, `/workspaceCaptures/${index}`));
+		else captures.push({
+			...missing(`workspaceCaptures is empty or malformed${addressGap === null ? "" : `; ${addressGap}`}`),
+			receiptPointer: "/workspaceCaptures",
+			receipt: receipts ?? null
+		});
+	}
+	return captures.length > 0 ? captures : [project(null, null)];
+}
 /**
-* What Runtime retained of each worker's decisions, from its own records.
+* What Runtime retained of each node's decisions, from its own records.
 *
 * - A turn is one `execution-admitted` event in the `dispatched` phase. Its `execution-result`
 *   names an output blob holding that turn's provider event stream (reasoning, tool calls and
 *   results as the harness reported them). A dispatched turn with no retained output lost its
 *   record, which is how a worker that went down mid-turn shows up.
+* - Every event's `outRef` enters the same collector, once per node and address. Terminal
+*   records retain ordinary and steerable results; they are not counted as dispatched turns.
 * - The terminal event's `harnessTranscript` receipt covers the harness's native session files.
 *   Those are the only record of the harness's own subagents, so an unavailable receipt is a
 *   separate gap from a missing turn, and its reason is Runtime's, verbatim.
@@ -1057,14 +1145,15 @@ function blobFile(runDir, ref) {
 * Blob files are checked on disk: a receipt that names a blob the directory no longer holds is
 * not a retained transcript.
 */
-async function workerTranscripts(runDir, events, workerIds) {
+async function nodeTranscripts(runDir, events, nodeIds) {
 	const facts = /* @__PURE__ */ new Map();
 	const entry = (id) => {
 		let found = facts.get(id);
 		if (found === void 0) {
 			found = {
 				dispatched: 0,
-				outputFiles: [],
+				outputRefs: [],
+				outputPositions: /* @__PURE__ */ new Map(),
 				receipt: null
 			};
 			facts.set(id, found);
@@ -1073,12 +1162,23 @@ async function workerTranscripts(runDir, events, workerIds) {
 	};
 	for (const event of events) {
 		const id = nonEmptyString(event.id);
-		if (id === null || !workerIds.has(id)) continue;
+		if (id === null || !nodeIds.has(id)) continue;
+		if (Object.hasOwn(event, "outRef")) {
+			const fact = entry(id);
+			const ref = event.outRef;
+			const position = typeof ref === "string" ? fact.outputPositions.get(ref) : void 0;
+			if (position !== void 0) {
+				if (event.kind === "execution-result") fact.outputRefs[position].turn = true;
+			} else {
+				if (typeof ref === "string") fact.outputPositions.set(ref, fact.outputRefs.length);
+				fact.outputRefs.push({
+					value: ref,
+					turn: event.kind === "execution-result"
+				});
+			}
+		}
 		if (event.kind === "execution-admitted" && record(event.admission)?.phase === "dispatched") entry(id).dispatched += 1;
-		else if (event.kind === "execution-result") {
-			const file = blobFile(runDir, event.outRef);
-			if (file !== null) entry(id).outputFiles.push(file);
-		} else if (event.kind === "settled" || event.kind === "cancelled") {
+		else if (event.kind === "settled" || event.kind === "cancelled") {
 			const receipt = record(event.harnessTranscript);
 			if (receipt !== null) entry(id).receipt = receipt;
 		}
@@ -1086,7 +1186,16 @@ async function workerTranscripts(runDir, events, workerIds) {
 	const out = /* @__PURE__ */ new Map();
 	for (const [id, fact] of facts) {
 		const retained = [];
-		for (const file of fact.outputFiles) if (await isFile(file)) retained.push(file);
+		let retainedTurns = 0;
+		const workspaceCaptures = [];
+		for (const { value: ref, turn } of fact.outputRefs) {
+			const file = blobFile(runDir, ref);
+			if (file !== null && await isFile(file)) {
+				retained.push(file);
+				if (turn) retainedTurns += 1;
+			}
+			workspaceCaptures.push(...await workspaceCapturesFromOutput(runDir, ref));
+		}
 		let nativeSession = null;
 		if (fact.receipt !== null) {
 			if (fact.receipt.status === "available") {
@@ -1107,9 +1216,10 @@ async function workerTranscripts(runDir, events, workerIds) {
 			transcriptRef: nativeSession?.status === "available" ? nativeSession.ref : retained.at(-1) ?? null,
 			turns: {
 				dispatched: fact.dispatched,
-				retained: retained.length
+				retained: retainedTurns
 			},
-			nativeSession
+			nativeSession,
+			workspaceCaptures: workspaceCaptures.length > 0 ? workspaceCaptures : null
 		});
 	}
 	return out;
@@ -1153,6 +1263,7 @@ function absentRuntimeSupervisorRun(runDir, resultText, failureText, instanceId)
 		harnessMissingReason: "Runtime FileRunContext has no external worker-token join",
 		limits: NO_SOURCE_LIMITS,
 		rootTranscriptRef: null,
+		rootWorkspaceCaptures: null,
 		traceCommand: "unavailable — Runtime FileRunContext records no provider-session trace identity"
 	};
 }
@@ -1188,7 +1299,7 @@ async function readRuntimeSupervisorRun(runDir, opts = {}) {
 	assertResultMatchesJournal(normalized.root, result, resultPath);
 	const childSpawns = normalized.events.filter((event) => event.kind === "spawned" && nonEmptyString(event.id) !== null).filter((event) => event.id !== normalized.root);
 	const workerIds = new Set(childSpawns.map((event) => nonEmptyString(event.id)).filter((id) => id !== null));
-	const transcripts = await workerTranscripts(runDir, normalized.events, workerIds);
+	const transcripts = await nodeTranscripts(runDir, normalized.events, /* @__PURE__ */ new Set([normalized.root, ...workerIds]));
 	const workers = childSpawns.map((event) => {
 		const workerId = nonEmptyString(event.id);
 		const transcript = transcripts.get(workerId);
@@ -1204,6 +1315,7 @@ async function readRuntimeSupervisorRun(runDir, opts = {}) {
 				retained: 0
 			},
 			nativeSession: transcript?.nativeSession ?? null,
+			workspaceCaptures: transcript?.workspaceCaptures ?? null,
 			patchPath: null
 		};
 	});
@@ -1230,6 +1342,7 @@ async function readRuntimeSupervisorRun(runDir, opts = {}) {
 		harnessMissingReason: "Runtime FileRunContext has no external worker-token join",
 		limits: sourceLimits(normalized.root, normalized.events, workerIds),
 		rootTranscriptRef: await isFile(rootStream) ? rootStream : null,
+		rootWorkspaceCaptures: transcripts.get(normalized.root)?.workspaceCaptures ?? null,
 		traceCommand: "unavailable — Runtime FileRunContext records no provider-session trace identity"
 	};
 }
@@ -1484,4 +1597,4 @@ async function runSupervisorRunCommand(argv, io = PROCESS_IO) {
 //#endregion
 export { writeSupervisorRunReport as a, readRuntimeSupervisorRun as c, renderSupervisorRunHeadline as d, renderSupervisorRunMarkdown as f, rollupSupervisorRuns as h, reportSupervisorRound as i, runtimeSupervisorRunReader as l, parsePatch as m, analyzeSupervisorRun as n, writeSupervisorRunReportSafe as o, analyzeSupervisorRunSources as p, findSupervisorRunDirs as r, isRuntimeSupervisorRunDir as s, runSupervisorRunCommand as t, renderSupervisorRollupMarkdown as u };
 
-//# sourceMappingURL=report-command-DzPpFgjB.js.map
+//# sourceMappingURL=report-command-D0XL2Cvi.js.map