@iowarp/clio-coder 0.3.7 → 0.3.9

package/src/domains/dispatch/write-boundary.ts

@@ -21,18 +21,27 @@
  * Everything is pinned to the baseline commit recorded at snapshot time, so a
  * commit step that moves HEAD without touching the working tree reads as "no
  * change" instead of as a wholesale rewrite.
+ *
+ * A window's diff answers "what changed", never "who changed it". An operator
+ * editing the checkout while a fleet runs produces changes inside the window
+ * that no step made, and rolling those back to the baseline destroys work
+ * nobody asked Clio to touch. So a change is blamed on the window only when it
+ * intersects what the window's own runs recorded writing; everything else is
+ * reported as an unattributed concurrent change and left exactly as it is.
  */
 
 import { execFileSync } from "node:child_process";
 import { createHash } from "node:crypto";
 import { existsSync, lstatSync, mkdirSync, readlinkSync, realpathSync, rmSync, writeFileSync } from "node:fs";
-import { dirname, join, relative, resolve, sep } from "node:path";
+import { dirname, isAbsolute, join, relative, resolve, sep } from "node:path";
+import { asDirectoryPathBoundary, pathBoundaryEntryCovers } from "../../core/path-boundary.js";
 import { clioStateDir } from "../../core/xdg.js";
 import { atomicWrite } from "../../engine/session.js";
 import { isGitRepository } from "../../tools/compete-worktrees.js";
 import {
 	describeWriteBoundary,
 	normalizeWriteBoundary,
+	normalizeWriteBoundaryEntry,
 	type WriteBoundary,
 	writeBoundaryCovers,
 } from "../agents/write-boundary.js";
@@ -71,10 +80,15 @@ function gitBytes(root: string, args: ReadonlyArray<string>): Buffer {
  *
  * Enforcement sees exactly what git sees, so an ignored path is outside it: a
  * repository decides what counts as its own content, and this is not the place
- * to overrule that. The one thing subtracted is Clio's own state directory when
- * an operator has placed it inside the workspace. Receipts, code-step logs, and
- * boundary verdicts are the orchestrator writing its journal while the step
- * runs, and blaming the step for them would make every window a violation.
+ * to overrule that. The consequence is that a boundary declared over an ignored
+ * path could never be verified, which is why `assertWriteBoundaryVisibleToGit`
+ * refuses that declaration before anything runs rather than letting this
+ * function certify an unobserved window as clean.
+ *
+ * The one thing subtracted is Clio's own state directory when an operator has
+ * placed it inside the workspace. Receipts, code-step logs, and boundary
+ * verdicts are the orchestrator writing its journal while the step runs, and
+ * blaming the step for them would make every window a violation.
  */
 function dirtyPaths(root: string): string[] {
 	const raw = git(root, ["status", "--porcelain=v1", "-z", "-uall", "--no-renames"]);
@@ -222,14 +236,98 @@ export interface WriteBoundaryUnrecoverable {
 }
 
 /**
- * `clean` means nothing outside the allowlist changed. `rolled-back` means
- * something did and the repository now looks as it did before. Anything else is
- * `rollback-incomplete`: the working tree is left exactly as the step made it
- * and handed to the operator with the list, because a rollback that guesses at
- * content it never recorded destroys work.
+ * `clean` means no change outside the allowlist is attributable to this
+ * window's steps. `rolled-back` means one was and the repository now looks as
+ * it did before. Anything else is `rollback-incomplete`: the working tree is
+ * left exactly as the step made it and handed to the operator with the list,
+ * because a rollback that guesses at content it never recorded destroys work.
+ *
+ * A `clean` verdict can still carry `unattributed` paths. Those changed inside
+ * the window and no run in it recorded writing them, so they are the
+ * operator's, another process's, or a build artifact's, and nothing here may
+ * touch them.
  */
 export type WriteBoundaryStatus = "clean" | "rolled-back" | "rollback-incomplete";
 
+/**
+ * What the window's own runs recorded aiming a successful mutation at.
+ *
+ * This is the write-side witness the diff cannot be: the checkout says a path
+ * changed, and only the run's own tool-call record says whether the run is what
+ * changed it. The record comes from the same fold that grounds a sealed
+ * mutation report, so the boundary and the result contract judge one set of
+ * facts.
+ *
+ * A run can write through a channel no tool event enumerates, and then its path
+ * set is a lower bound rather than the whole list. That is not silently
+ * tolerated: the recorder knows which tools can do it (bash, verify, dispatch,
+ * steer, and any tool this process has no schema for) and reports the record
+ * open, which arrives here as `complete: false`. So the set is only ever read
+ * as a closed list for a run that could not have written outside it, and
+ * enforcement over a run that could stays exactly as strong as it was before
+ * attribution existed.
+ */
+export interface WriteBoundaryAttribution {
+	/**
+	 * Paths the window's runs aimed a successful mutation at. Absolute, or
+	 * relative to the workspace root; anything resolving outside the workspace is
+	 * dropped, because git could not have reported it as a change either.
+	 */
+	recorded: ReadonlyArray<string>;
+	/**
+	 * False when at least one step in the window has no closed write record: a
+	 * code step running a registered command, a runtime that publishes no tool
+	 * telemetry, a run whose telemetry came back with holes, or a run that
+	 * called a tool able to write a path its arguments do not name. An open
+	 * record cannot clear anybody, so the window falls back to blaming every
+	 * change outside the allowlist, and the verdict says it did.
+	 */
+	complete: boolean;
+	/** Durable causes for an incomplete record. Empty only when complete is true. */
+	downgrades?: ReadonlyArray<WriteBoundaryAttributionDowngrade>;
+}
+
+export type WriteBoundaryAttributionDowngradeReason =
+	| "opaque_tool_succeeded"
+	| "incomplete_tool_telemetry"
+	| "code_step_untracked"
+	| "write_record_unavailable";
+
+/** Why one step could not offer a closed write record. */
+export interface WriteBoundaryAttributionDowngrade {
+	reason: WriteBoundaryAttributionDowngradeReason;
+	/** Tool whose successful call opened the record, or null for a non-tool cause. */
+	tool: string | null;
+	/** Exact worker call when telemetry supplied one, or null for a non-tool cause. */
+	toolCallId: string | null;
+	/** Run that observed the cause, when one reached dispatch. */
+	runId: string | null;
+	/** Fleet step whose record was downgraded. */
+	stepId: string | null;
+}
+
+/** One operator-facing explanation shared by live and durable surfaces. */
+export function describeWriteBoundaryAttributionDowngrade(downgrade: WriteBoundaryAttributionDowngrade): string {
+	if (downgrade.reason === "opaque_tool_succeeded") {
+		const tool = downgrade.tool ?? "unknown";
+		const call = downgrade.toolCallId === null ? "" : ` (call ${downgrade.toolCallId})`;
+		return `a successful '${tool}' call${call} can write paths its arguments do not name`;
+	}
+	if (downgrade.reason === "incomplete_tool_telemetry") {
+		return downgrade.runId === null
+			? "tool telemetry was incomplete"
+			: `run '${downgrade.runId}' reported incomplete tool telemetry`;
+	}
+	if (downgrade.reason === "code_step_untracked") {
+		return downgrade.stepId === null
+			? "a code step ran through the untracked command runner"
+			: `code step '${downgrade.stepId}' ran through the untracked command runner`;
+	}
+	return downgrade.stepId === null
+		? "no usable write record was available"
+		: `step '${downgrade.stepId}' supplied no usable write record`;
+}
+
 export interface WriteBoundaryVerdict {
 	version: 1;
 	/** Which scheduling window this covers, e.g. `wave-2` or `revalidate-suite.check.1`. */
@@ -243,12 +341,27 @@ export interface WriteBoundaryVerdict {
 	checkedAt: string;
 	changedPaths: ReadonlyArray<string>;
 	violations: ReadonlyArray<string>;
+	/**
+	 * Paths outside the allowlist that changed inside the window and that no run
+	 * in it recorded writing. Recorded so the operator can see them, never rolled
+	 * back: the window did not make them, so the baseline commit is not their
+	 * prior content.
+	 */
+	unattributed: ReadonlyArray<string>;
+	/** Whether every step in the window offered an enumerable write record. */
+	attributionComplete: boolean;
+	/** Why attribution was incomplete, including the causal tool call when known. */
+	attributionDowngrades: ReadonlyArray<WriteBoundaryAttributionDowngrade>;
 	rolledBack: ReadonlyArray<WriteBoundaryRollback>;
 	unrecoverable: ReadonlyArray<WriteBoundaryUnrecoverable>;
 	status: WriteBoundaryStatus;
 	/** Typed failure reason, or null when the window stayed inside its boundary. */
 	reason: typeof WRITE_BOUNDARY_VIOLATION_REASON | null;
-	/** Operator-facing message naming the offending paths and the declaration. */
+	/**
+	 * Operator-facing message naming the offending paths and the declaration, or
+	 * naming the concurrent changes the window saw but did not cause. Null only
+	 * when neither happened.
+	 */
 	detail: string | null;
 	digest: string;
 }
@@ -258,6 +371,12 @@ export interface EnforceWriteBoundaryInput {
 	window: string;
 	stepIds: ReadonlyArray<string>;
 	allow: WriteBoundary;
+	/**
+	 * What the window's runs recorded writing. Absent means no caller offered a
+	 * record, which reads the same as an incomplete one: every change outside the
+	 * allowlist is blamed on the window.
+	 */
+	attribution?: WriteBoundaryAttribution;
 }
 
 function canonicalVerdict(verdict: Omit<WriteBoundaryVerdict, "digest">): string {
@@ -317,6 +436,22 @@ function rollbackPath(
 	return { rolledBack: { path: change.path, action: "restored", restoredFrom: `${snapshot.head}:${change.path}` } };
 }
 
+/**
+ * Whether one changed path falls inside what a run recorded writing.
+ *
+ * A recorded target is usually the exact file a write or edit named, but a
+ * command's delete target can be a directory, so a change under a recorded
+ * prefix counts as the same write. Matching is on segment boundaries: `work`
+ * does not cover `workspace/a.txt`.
+ */
+function withinRecordedWrite(recorded: ReadonlySet<string>, path: string): boolean {
+	if (recorded.has(path)) return true;
+	for (const target of recorded) {
+		if (path.startsWith(`${target}/`)) return true;
+	}
+	return false;
+}
+
 /**
  * Compare, roll back, and return the verdict. Never throws for a violation: a
  * violation is evidence the caller records and acts on, not an exception that
@@ -325,7 +460,34 @@ function rollbackPath(
 export function enforceWriteBoundary(input: EnforceWriteBoundaryInput): WriteBoundaryVerdict {
 	const allow = normalizeWriteBoundary(input.allow);
 	const changes = diffWorkspace(input.snapshot);
-	const violations = changes.filter((change) => !writeBoundaryCovers(allow, change.path));
+	const outside = changes.filter((change) => !writeBoundaryCovers(allow, change.path));
+	const attributionComplete = input.attribution?.complete === true;
+	const attributionDowngrades: ReadonlyArray<WriteBoundaryAttributionDowngrade> = attributionComplete
+		? []
+		: input.attribution?.downgrades !== undefined && input.attribution.downgrades.length > 0
+			? input.attribution.downgrades.map((entry) => ({ ...entry }))
+			: [
+					{
+						reason: "write_record_unavailable",
+						tool: null,
+						toolCallId: null,
+						runId: null,
+						stepId: input.stepIds.length === 1 ? (input.stepIds[0] ?? null) : null,
+					},
+				];
+	const recorded = new Set<string>();
+	for (const target of input.attribution?.recorded ?? []) {
+		const relativeTarget = repoRelative(input.snapshot.root, target);
+		if (relativeTarget !== null) recorded.add(relativeTarget);
+	}
+	// Without a complete record nothing can be cleared, so the window keeps the
+	// whole diff. With one, only what the runs actually wrote is theirs.
+	const violations = attributionComplete
+		? outside.filter((change) => withinRecordedWrite(recorded, change.path))
+		: outside;
+	const unattributed = attributionComplete
+		? outside.filter((change) => !withinRecordedWrite(recorded, change.path))
+		: [];
 	const rolledBack: WriteBoundaryRollback[] = [];
 	const unrecoverable: WriteBoundaryUnrecoverable[] = [];
 	for (const change of violations) {
@@ -335,15 +497,15 @@ export function enforceWriteBoundary(input: EnforceWriteBoundaryInput): WriteBou
 	}
 	const status: WriteBoundaryStatus =
 		violations.length === 0 ? "clean" : unrecoverable.length === 0 ? "rolled-back" : "rollback-incomplete";
-	const attribution =
+	const who =
 		input.stepIds.length === 1
 			? `step '${input.stepIds[0]}'`
 			: `steps ${input.stepIds.map((id) => `'${id}'`).join(", ")} ran concurrently in one checkout, so the write cannot be attributed to one of them`;
-	const detail =
+	const violationDetail =
 		violations.length === 0
-			? null
+			? []
 			: [
-					`${attribution} wrote outside its declared boundary.`,
+					`${who} wrote outside its declared boundary.`,
 					`Unauthorized paths: ${violations.map((change) => change.path).join(", ")}.`,
 					`Declared writes: ${describeWriteBoundary(allow)}.`,
 					status === "rolled-back"
@@ -351,8 +513,23 @@ export function enforceWriteBoundary(input: EnforceWriteBoundaryInput): WriteBou
 						: `Rollback is incomplete; the working tree is left as the step made it. Unrestorable: ${unrecoverable
 								.map((entry) => `${entry.path} (${entry.reason})`)
 								.join("; ")}.`,
+					attributionComplete
+						? "Each was blamed on this window because a run in it recorded writing that path."
+						: `Attribution was incomplete because ${attributionDowngrades
+								.map(describeWriteBoundaryAttributionDowngrade)
+								.join("; ")}, so every change outside the declaration was blamed on this window.`,
 					"If the change was legitimate, widen the step's `writes:` declaration to cover it.",
-				].join(" ");
+				];
+	const concurrentDetail =
+		unattributed.length === 0
+			? []
+			: [
+					`Concurrent changes were seen in this window that no run in it recorded writing: ${unattributed
+						.map((change) => change.path)
+						.join(", ")}.`,
+					"They were left untouched, because a path this window did not write is not this window's to restore.",
+				];
+	const detail = [...violationDetail, ...concurrentDetail];
 	const body: Omit<WriteBoundaryVerdict, "digest"> = {
 		version: 1,
 		window: input.window,
@@ -363,11 +540,14 @@ export function enforceWriteBoundary(input: EnforceWriteBoundaryInput): WriteBou
 		checkedAt: new Date().toISOString(),
 		changedPaths: changes.map((change) => change.path),
 		violations: violations.map((change) => change.path),
+		unattributed: unattributed.map((change) => change.path),
+		attributionComplete,
+		attributionDowngrades,
 		rolledBack,
 		unrecoverable,
 		status,
 		reason: violations.length === 0 ? null : WRITE_BOUNDARY_VIOLATION_REASON,
-		detail,
+		detail: detail.length === 0 ? null : detail.join(" "),
 	};
 	return { ...body, digest: createHash("sha256").update(canonicalVerdict(body), "utf8").digest("hex") };
 }
@@ -387,16 +567,76 @@ export function assertWriteBoundaryInsideRoot(root: string, allow: WriteBoundary
 		// path a step is about to create.
 		while (probe !== realRoot && !existsSync(probe)) probe = dirname(probe);
 		const resolved = realpathSync(probe);
-		if (resolved !== realRoot && !resolved.startsWith(realRoot + sep)) {
+		// The root is inside itself. The prefix walk above stops at the root when
+		// a declared directory does not exist yet, and `pathBoundaryEntryCovers`
+		// answers a different question: it asks whether a changed path falls under
+		// a declaration, where the directory entry itself is not a change git ever
+		// reports. Asking it about the root would refuse every boundary naming a
+		// path the step is about to create.
+		if (resolved !== realRoot && !pathBoundaryEntryCovers(asDirectoryPathBoundary(realRoot), resolved)) {
 			throw new Error(`write boundary: entry '${entry}' resolves to ${resolved}, outside the workspace at ${realRoot}`);
 		}
 	}
 }
 
-/** Repo-relative form of an absolute path, or null when it is outside. */
+/**
+ * Refuse a declaration git will never report on.
+ *
+ * `git status` does not list an ignored path, so a boundary entry the
+ * repository ignores names a region enforcement is blind to: every window over
+ * it is certified clean without a single observation, which is a stronger claim
+ * than "nothing was checked" and reads as a passing one. That is the same
+ * defect `assertWriteBoundaryInsideRoot` refuses for an entry that leaves the
+ * workspace through a symlink, so it is refused in the same place and at the
+ * same time, before anything runs.
+ *
+ * A tracked path is reportable whatever the ignore rules say, and git's own
+ * index-aware check is what decides that, so this never refuses a path the
+ * repository is actually following.
+ */
+export function assertWriteBoundaryVisibleToGit(root: string, allow: WriteBoundary): void {
+	// Entry by entry rather than through `normalizeWriteBoundary`, whose 32-entry
+	// cap is a per-step declaration limit. A caller may legitimately ask this
+	// question about every entry a whole plan declared.
+	const entries = [...new Set(allow.map(normalizeWriteBoundaryEntry))].sort();
+	if (entries.length === 0) return;
+	const absolute = resolve(root);
+	let raw: string;
+	try {
+		raw = git(absolute, ["check-ignore", "-v", "-z", "--stdin"], `${entries.join("\0")}\0`);
+	} catch (error) {
+		// `check-ignore` exits 1 when nothing matched, which is the common answer
+		// and not a failure. Any other exit is a real inability to decide, and
+		// enforcement fails closed rather than assuming visibility.
+		const status = (error as { status?: number }).status;
+		if (status === 1) return;
+		throw new Error(
+			`write boundary: could not determine whether ${describeWriteBoundary(entries)} is ignored by git in ${absolute}: ${(error as Error).message}`,
+		);
+	}
+	const fields = raw.split("\0");
+	const refusals: string[] = [];
+	// `-v -z` emits source, line number, pattern, and pathname per match.
+	for (let index = 0; index + 3 < fields.length; index += 4) {
+		const [source, line, pattern, path] = fields.slice(index, index + 4);
+		if (path === undefined || path.length === 0) continue;
+		refusals.push(`'${path}' is ignored by ${source}:${line}:${pattern}`);
+	}
+	if (refusals.length === 0) return;
+	throw new Error(
+		`write boundary: ${refusals.join(", ")}. git status never reports an ignored path, so a boundary declared over one is verified against nothing and every window would be certified clean without an observation. Declare a path the repository tracks, or stop ignoring this one.`,
+	);
+}
+
+/**
+ * Workspace-relative form of a path, or null when it resolves outside. A
+ * relative input is taken as already workspace-relative, which is how a
+ * recorded write target and an absolute state directory both normalize here.
+ */
 function repoRelative(root: string, path: string): string | null {
-	const rel = relative(resolve(root), resolve(path));
-	if (rel.length === 0 || rel.startsWith("..") || rel.startsWith(`..${sep}`)) return null;
+	const base = resolve(root);
+	const rel = relative(base, resolve(base, path));
+	if (rel.length === 0 || rel === ".." || rel.startsWith(`..${sep}`) || isAbsolute(rel)) return null;
 	return rel.split(sep).join("/");
 }
 

package/src/domains/eval/artifacts/store.ts

@@ -3,6 +3,11 @@ import { join } from "node:path";
 import { assertSafeId } from "../../../core/safe-id.js";
 import { safeResourceWrite } from "../../../core/safe-resource-write.js";
 import type { EvalArtifactV4, EvalTokenAccountingV4 } from "../schema/artifact.js";
+import { assertEvalBehaviorReferencesVerdictV1, parseEvalBehaviorVerdictV1 } from "../schema/behavioral.js";
+import { parseEvalBehaviorMetricsV1 } from "../schema/behavioral-metrics.js";
+import { parseEvalExecutionEnvelopeV1, parseEvalExecutionMatrixDimensionsV1 } from "../schema/execution-envelope.js";
+import { parseEvalServingConfigurationV1 } from "../schema/serving.js";
+import { parseEvalVerdictEnvelopeV1 } from "../schema/verdict.js";
 import { evalRoot } from "../store.js";
 import { redactArtifactForStorage } from "./redact.js";
 
@@ -45,6 +50,14 @@ export function parseEvalArtifactV4(value: unknown, source: string): EvalArtifac
 	const summary = asRecord(value.summary, `${source}.summary`);
 	const matrix = asRecord(value.matrix, `${source}.matrix`);
 	const suite = asRecord(value.suite, `${source}.suite`);
+	const servingConfiguration =
+		value.servingConfiguration === undefined
+			? undefined
+			: parseEvalServingConfigurationV1(value.servingConfiguration, `${source}.servingConfiguration`);
+	const aggregates =
+		value.aggregates === undefined
+			? undefined
+			: (readArray(value, source, "aggregates") as NonNullable<EvalArtifactV4["aggregates"]>);
 	return {
 		version: 4,
 		evalId: readString(value, source, "evalId"),
@@ -62,7 +75,11 @@ export function parseEvalArtifactV4(value: unknown, source: string): EvalArtifac
 			target: readString(matrix, `${source}.matrix`, "target"),
 			model: readNullableString(matrix, `${source}.matrix`, "model"),
 			thinking: readNullableString(matrix, `${source}.matrix`, "thinking"),
+			...(matrix.dimensions === undefined
+				? {}
+				: { dimensions: parseEvalExecutionMatrixDimensionsV1(matrix.dimensions, `${source}.matrix.dimensions`) }),
 		},
+		...(servingConfiguration === undefined ? {} : { servingConfiguration }),
 		summary: {
 			runs: readNumber(summary, `${source}.summary`, "runs"),
 			passed: readNumber(summary, `${source}.summary`, "passed"),
@@ -71,6 +88,7 @@ export function parseEvalArtifactV4(value: unknown, source: string): EvalArtifac
 			tokens: parseTokenAccounting(summary.tokens, `${source}.summary.tokens`),
 			wallTimeMs: readNumber(summary, `${source}.summary`, "wallTimeMs"),
 		},
+		...(aggregates === undefined ? {} : { aggregates }),
 		results: readArray(value, source, "results").map((entry, index) => parseResult(entry, `${source}.results[${index}]`)),
 	};
 }
@@ -108,6 +126,46 @@ function parseTokenAccounting(value: unknown, source: string): EvalTokenAccounti
 function parseResult(value: unknown, source: string): EvalArtifactV4["results"][number] {
 	const record = asRecord(value, source);
 	const target = asRecord(record.target, `${source}.target`);
+	const verdict =
+		record.verdict === undefined ? undefined : parseEvalVerdictEnvelopeV1(record.verdict, `${source}.verdict`);
+	const behavioral =
+		record.behavioral === undefined ? undefined : parseEvalBehaviorVerdictV1(record.behavioral, `${source}.behavioral`);
+	const behavioralMetrics =
+		record.behavioralMetrics === undefined
+			? undefined
+			: parseEvalBehaviorMetricsV1(record.behavioralMetrics, `${source}.behavioralMetrics`);
+	const executionEnvelope =
+		record.executionEnvelope === undefined
+			? undefined
+			: parseEvalExecutionEnvelopeV1(record.executionEnvelope, `${source}.executionEnvelope`);
+	if (behavioral !== undefined && verdict === undefined) {
+		throw new Error(`${source}.behavioral: sibling document requires a verdict`);
+	}
+	if (behavioral !== undefined && verdict !== undefined) {
+		assertEvalBehaviorReferencesVerdictV1(behavioral, verdict, `${source}.behavioral`);
+	}
+	if (behavioralMetrics !== undefined) {
+		if (behavioral === undefined) throw new Error(`${source}.behavioralMetrics: requires a behavioral verdict`);
+		if (behavioralMetrics.scenarioId !== readString(record, source, "taskId")) {
+			throw new Error(`${source}.behavioralMetrics.scenarioId: conflicts with result taskId`);
+		}
+		if (
+			behavioralMetrics.target.id !== readString(target, `${source}.target`, "id") ||
+			behavioralMetrics.target.model !== readNullableString(target, `${source}.target`, "model")
+		) {
+			throw new Error(`${source}.behavioralMetrics.target: conflicts with result target`);
+		}
+	}
+	if (executionEnvelope !== undefined) {
+		if (behavioral === undefined) throw new Error(`${source}.executionEnvelope: requires a behavioral verdict`);
+		if (
+			executionEnvelope.target !== readString(target, `${source}.target`, "id") ||
+			executionEnvelope.corpus.id !== behavioral.corpus.id ||
+			executionEnvelope.corpus.version !== behavioral.corpus.version
+		) {
+			throw new Error(`${source}.executionEnvelope: conflicts with result target or behavioral corpus`);
+		}
+	}
 	return {
 		assignmentId: readNullableString(record, source, "assignmentId"),
 		terminalReceiptDigest: readNullableDigest(record, source, "terminalReceiptDigest"),
@@ -122,6 +180,10 @@ function parseResult(value: unknown, source: string): EvalArtifactV4["results"][
 		failureClass: readNullableString(record, source, "failureClass"),
 		metrics: asRecord(record.metrics, `${source}.metrics`) as Record<string, number | string | boolean | null>,
 		artifacts: asRecord(record.artifacts, `${source}.artifacts`) as Record<string, string | string[] | null>,
+		...(verdict === undefined ? {} : { verdict }),
+		...(behavioral === undefined ? {} : { behavioral }),
+		...(behavioralMetrics === undefined ? {} : { behavioralMetrics }),
+		...(executionEnvelope === undefined ? {} : { executionEnvelope }),
 	};
 }