@namzu/sdk 36.0.0 → 38.0.0

package/dist/utils/json-snapshot.js

@@ -0,0 +1,79 @@
+/** Clone JSON data without invoking getters or silently dropping values. */
+export function cloneJsonValue(value, freeze, path = '$') {
+    const active = new WeakSet();
+    const copies = new WeakMap();
+    const clone = (candidate, at) => {
+        if (candidate === null || typeof candidate === 'string' || typeof candidate === 'boolean') {
+            return candidate;
+        }
+        if (typeof candidate === 'number') {
+            if (!Number.isFinite(candidate) || Object.is(candidate, -0)) {
+                throw new TypeError(`${at} must be a finite JSON number`);
+            }
+            return candidate;
+        }
+        if (typeof candidate !== 'object') {
+            throw new TypeError(`${at} must be a JSON value, not ${typeof candidate}`);
+        }
+        if (active.has(candidate))
+            throw new TypeError(`${at} contains a cycle`);
+        const existing = copies.get(candidate);
+        if (existing)
+            return existing;
+        active.add(candidate);
+        try {
+            if (Array.isArray(candidate)) {
+                const extra = Reflect.ownKeys(candidate).filter((key) => key !== 'length' && (typeof key !== 'string' || !/^(0|[1-9]\d*)$/.test(key)));
+                if (extra.length > 0)
+                    throw new TypeError(`${at} has non-JSON array properties`);
+                const result = new Array(candidate.length);
+                copies.set(candidate, result);
+                for (let index = 0; index < candidate.length; index++) {
+                    const descriptor = Object.getOwnPropertyDescriptor(candidate, String(index));
+                    if (!descriptor)
+                        throw new TypeError(`${at}[${index}] is a sparse array hole`);
+                    if (!descriptor.enumerable || !('value' in descriptor)) {
+                        throw new TypeError(`${at}[${index}] must be an enumerable data property`);
+                    }
+                    result[index] = clone(descriptor.value, `${at}[${index}]`);
+                }
+                if (freeze)
+                    Object.freeze(result);
+                return result;
+            }
+            const prototype = Object.getPrototypeOf(candidate);
+            if (prototype !== Object.prototype && prototype !== null) {
+                throw new TypeError(`${at} must be a plain JSON object`);
+            }
+            // JSON does not preserve prototypes. Canonicalize both admitted plain
+            // shapes to an ordinary object so a disk checkpoint round-trip compares
+            // equal to a fresh preparation of the same semantic value. Properties
+            // are still defined explicitly, so a literal "__proto__" key remains
+            // data rather than changing this object's prototype.
+            const result = {};
+            copies.set(candidate, result);
+            for (const key of Reflect.ownKeys(candidate)) {
+                if (typeof key !== 'string')
+                    throw new TypeError(`${at} has a symbol property`);
+                const descriptor = Object.getOwnPropertyDescriptor(candidate, key);
+                if (!descriptor?.enumerable || !('value' in descriptor)) {
+                    throw new TypeError(`${at}.${key} must be an enumerable data property`);
+                }
+                Object.defineProperty(result, key, {
+                    value: clone(descriptor.value, `${at}.${key}`),
+                    enumerable: true,
+                    writable: !freeze,
+                    configurable: !freeze,
+                });
+            }
+            if (freeze)
+                Object.freeze(result);
+            return result;
+        }
+        finally {
+            active.delete(candidate);
+        }
+    };
+    return clone(value, path);
+}
+//# sourceMappingURL=json-snapshot.js.map

package/dist/utils/json-snapshot.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"json-snapshot.js","sourceRoot":"","sources":["../../src/utils/json-snapshot.ts"],"names":[],"mappings":"AAAA,4EAA4E;AAC5E,MAAM,UAAU,cAAc,CAAC,KAAc,EAAE,MAAe,EAAE,IAAI,GAAG,GAAG;IACzE,MAAM,MAAM,GAAG,IAAI,OAAO,EAAU,CAAA;IACpC,MAAM,MAAM,GAAG,IAAI,OAAO,EAAkB,CAAA;IAE5C,MAAM,KAAK,GAAG,CAAC,SAAkB,EAAE,EAAU,EAAW,EAAE;QACzD,IAAI,SAAS,KAAK,IAAI,IAAI,OAAO,SAAS,KAAK,QAAQ,IAAI,OAAO,SAAS,KAAK,SAAS,EAAE,CAAC;YAC3F,OAAO,SAAS,CAAA;QACjB,CAAC;QACD,IAAI,OAAO,SAAS,KAAK,QAAQ,EAAE,CAAC;YACnC,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,SAAS,CAAC,IAAI,MAAM,CAAC,EAAE,CAAC,SAAS,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;gBAC7D,MAAM,IAAI,SAAS,CAAC,GAAG,EAAE,+BAA+B,CAAC,CAAA;YAC1D,CAAC;YACD,OAAO,SAAS,CAAA;QACjB,CAAC;QACD,IAAI,OAAO,SAAS,KAAK,QAAQ,EAAE,CAAC;YACnC,MAAM,IAAI,SAAS,CAAC,GAAG,EAAE,8BAA8B,OAAO,SAAS,EAAE,CAAC,CAAA;QAC3E,CAAC;QACD,IAAI,MAAM,CAAC,GAAG,CAAC,SAAS,CAAC;YAAE,MAAM,IAAI,SAAS,CAAC,GAAG,EAAE,mBAAmB,CAAC,CAAA;QACxE,MAAM,QAAQ,GAAG,MAAM,CAAC,GAAG,CAAC,SAAS,CAAC,CAAA;QACtC,IAAI,QAAQ;YAAE,OAAO,QAAQ,CAAA;QAE7B,MAAM,CAAC,GAAG,CAAC,SAAS,CAAC,CAAA;QACrB,IAAI,CAAC;YACJ,IAAI,KAAK,CAAC,OAAO,CAAC,SAAS,CAAC,EAAE,CAAC;gBAC9B,MAAM,KAAK,GAAG,OAAO,CAAC,OAAO,CAAC,SAAS,CAAC,CAAC,MAAM,CAC9C,CAAC,GAAG,EAAE,EAAE,CAAC,GAAG,KAAK,QAAQ,IAAI,CAAC,OAAO,GAAG,KAAK,QAAQ,IAAI,CAAC,gBAAgB,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CACrF,CAAA;gBACD,IAAI,KAAK,CAAC,MAAM,GAAG,CAAC;oBAAE,MAAM,IAAI,SAAS,CAAC,GAAG,EAAE,gCAAgC,CAAC,CAAA;gBAChF,MAAM,MAAM,GAAc,IAAI,KAAK,CAAC,SAAS,CAAC,MAAM,CAAC,CAAA;gBACrD,MAAM,CAAC,GAAG,CAAC,SAAS,EAAE,MAAM,CAAC,CAAA;gBAC7B,KAAK,IAAI,KAAK,GAAG,CAAC,EAAE,KAAK,GAAG,SAAS,CAAC,MAAM,EAAE,KAAK,EAAE,EAAE,CAAC;oBACvD,MAAM,UAAU,GAAG,MAAM,CAAC,wBAAwB,CAAC,SAAS,EAAE,MAAM,CAAC,KAAK,CAAC,CAAC,CAAA;oBAC5E,IAAI,CAAC,UAAU;wBAAE,MAAM,IAAI,SAAS,CAAC,GAAG,EAAE,IAAI,KAAK,0BAA0B,CAAC,CAAA;oBAC9E,IAAI,CAAC,UAAU,CAAC,UAAU,IAAI,CAAC,CAAC,OAAO,IAAI,UAAU,CAAC,EAAE,CAAC;wBACxD,MAAM,IAAI,SAAS,CAAC,GAAG,EAAE,IAAI,KAAK,uCAAuC,CAAC,CAAA;oBAC3E,CAAC;oBACD,MAAM,CAAC,KAAK,CAAC,GAAG,KAAK,CAAC,UAAU,CAAC,KAAK,EAAE,GAAG,EAAE,IAAI,KAAK,GAAG,CAAC,CAAA;gBAC3D,CAAC;gBACD,IAAI,MAAM;oBAAE,MAAM,CAAC,MAAM,CAAC,MAAM,CAAC,CAAA;gBACjC,OAAO,MAAM,CAAA;YACd,CAAC;YAED,MAAM,SAAS,GAAG,MAAM,CAAC,cAAc,CAAC,SAAS,CAAC,CAAA;YAClD,IAAI,SAAS,KAAK,MAAM,CAAC,SAAS,IAAI,SAAS,KAAK,IAAI,EAAE,CAAC;gBAC1D,MAAM,IAAI,SAAS,CAAC,GAAG,EAAE,8BAA8B,CAAC,CAAA;YACzD,CAAC;YACD,sEAAsE;YACtE,wEAAwE;YACxE,sEAAsE;YACtE,qEAAqE;YACrE,qDAAqD;YACrD,MAAM,MAAM,GAA4B,EAAE,CAAA;YAC1C,MAAM,CAAC,GAAG,CAAC,SAAS,EAAE,MAAM,CAAC,CAAA;YAC7B,KAAK,MAAM,GAAG,IAAI,OAAO,CAAC,OAAO,CAAC,SAAS,CAAC,EAAE,CAAC;gBAC9C,IAAI,OAAO,GAAG,KAAK,QAAQ;oBAAE,MAAM,IAAI,SAAS,CAAC,GAAG,EAAE,wBAAwB,CAAC,CAAA;gBAC/E,MAAM,UAAU,GAAG,MAAM,CAAC,wBAAwB,CAAC,SAAS,EAAE,GAAG,CAAC,CAAA;gBAClE,IAAI,CAAC,UAAU,EAAE,UAAU,IAAI,CAAC,CAAC,OAAO,IAAI,UAAU,CAAC,EAAE,CAAC;oBACzD,MAAM,IAAI,SAAS,CAAC,GAAG,EAAE,IAAI,GAAG,sCAAsC,CAAC,CAAA;gBACxE,CAAC;gBACD,MAAM,CAAC,cAAc,CAAC,MAAM,EAAE,GAAG,EAAE;oBAClC,KAAK,EAAE,KAAK,CAAC,UAAU,CAAC,KAAK,EAAE,GAAG,EAAE,IAAI,GAAG,EAAE,CAAC;oBAC9C,UAAU,EAAE,IAAI;oBAChB,QAAQ,EAAE,CAAC,MAAM;oBACjB,YAAY,EAAE,CAAC,MAAM;iBACrB,CAAC,CAAA;YACH,CAAC;YACD,IAAI,MAAM;gBAAE,MAAM,CAAC,MAAM,CAAC,MAAM,CAAC,CAAA;YACjC,OAAO,MAAM,CAAA;QACd,CAAC;gBAAS,CAAC;YACV,MAAM,CAAC,MAAM,CAAC,SAAS,CAAC,CAAA;QACzB,CAAC;IACF,CAAC,CAAA;IAED,OAAO,KAAK,CAAC,KAAK,EAAE,IAAI,CAAC,CAAA;AAC1B,CAAC"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@namzu/sdk",
-  "version": "36.0.0",
+  "version": "38.0.0",
   "description": "Open-source AI agent SDK with a built-in runtime. Nothing between you and your agents.",
   "license": "FSL-1.1-MIT",
   "engines": {
@@ -73,8 +73,10 @@
     "zod-to-json-schema": "^3.23.0"
   },
   "dependencies": {
+    "@jitl/quickjs-wasmfile-release-sync": "0.32.0",
     "brace-expansion": "^2.1.4",
-    "minimatch": "^9.0.9"
+    "minimatch": "^9.0.9",
+    "quickjs-emscripten-core": "0.32.0"
   },
   "scripts": {
     "build": "tsc --build",

package/src/bridge/a2a/mapper.ts

@@ -46,6 +46,10 @@ type A2ATransform<K extends RunEvent['type']> =
 const MAPPING: {
 	[K in RunEvent['type']]: A2ATransform<K>
 } = {
+	// Internal cumulative admission ledger, not a public UI event.
+	tool_calls_admitted: null,
+	// Hosted activity is retained in the run log; A2A exposes the resulting answer.
+	hosted_tool: null,
 	run_started: (e, ctx) => statusEvent(e.runId, 'running', false, ctx),
 
 	run_completed: (e, ctx) => {

package/src/bridge/sse/mapper.ts

@@ -34,6 +34,16 @@ type EventTransform<K extends RunEvent['type']> = {
 const MAPPING: {
 	[K in RunEvent['type']]: EventTransform<K>
 } = {
+	// Internal cumulative admission ledger, not a public UI event.
+	tool_calls_admitted: null,
+	hosted_tool: {
+		wire: 'hosted.tool',
+		transform: (e, runId) => ({
+			run_id: runId,
+			iteration: e.iteration,
+			tool: e.tool,
+		}),
+	},
 	run_started: {
 		wire: 'run.started',
 		transform: (e, runId) => ({
@@ -52,7 +62,12 @@ const MAPPING: {
 	// needs in order to see supervision loosen.
 	approval_policy_changed: {
 		wire: 'approval_policy.changed',
-		transform: (e, runId) => ({ run_id: runId, from: e.from, to: e.to, reason: e.reason }),
+		transform: (e, runId) => ({
+			run_id: runId,
+			from: e.from,
+			to: e.to,
+			reason: e.reason,
+		}),
 	},
 
 	iteration_completed: {

package/src/compaction/extractor.ts

@@ -1,6 +1,6 @@
 import type { CompactionConfig } from '../config/runtime.js'
 import { EDIT_TOOLS, READ_TOOLS, SEARCH_TOOLS, SHELL_TOOLS } from '../constants/compaction/index.js'
-import type { WorkingStateManager } from './manager.js'
+import { type WorkingStateManager, truncateStateText } from './manager.js'
 
 function tryParseJson(raw: string): Record<string, unknown> | null {
 	try {
@@ -25,12 +25,17 @@ function splitSentences(text: string): string[] {
 }
 
 function isSubstantive(sentence: string): boolean {
-	const fillerPatterns = [
-		/^(ok|okay|sure|alright|got it|understood|i see|let me|i'll|i will|now i|here)/i,
-		/^(great|perfect|excellent|good|nice|cool|right|yes|no|thanks)/i,
-		/^(let's|we can|we should|i can|i should)/i,
-	]
-	return !fillerPatterns.some((pattern) => pattern.test(sentence))
+	// Discard standalone acknowledgements, not sentences beginning with one:
+	// "No production writes" and even "Node 24" used to match the `no` prefix.
+	return !/^(ok|okay|sure|alright|got it|understood|i see|great|perfect|excellent|good|nice|cool|right|yes|no|thanks)[.!?]*$/i.test(
+		sentence,
+	)
+}
+
+function isExplicitAssistantDecision(sentence: string): boolean {
+	return /^(?:(?:the|my|our) decision is\b|decision\s*:|(?:i|we)(?: have|'ve)? (?:decided|chose|chosen)\b)/i.test(
+		sentence,
+	)
 }
 
 function truncateResult(result: string, maxLen: number): string {
@@ -199,6 +204,14 @@ export function extractFromAssistantMessage(
 	const selected = substantive.slice(0, config.maxSentencesPerTurn)
 
 	for (const sentence of selected) {
-		manager.addAssistantNote(sentence)
+		if (isExplicitAssistantDecision(sentence)) {
+			// A recorded assertion, not verified truth or an instruction from the
+			// user. Keep that provenance when a host later promotes the decision.
+			manager.addDecision(
+				truncateStateText(`Assistant stated: ${sentence}`, config.maxCharsPerNote),
+			)
+		} else {
+			manager.addAssistantNote(sentence)
+		}
 	}
 }

package/src/compaction/manager.ts

@@ -13,6 +13,16 @@ export const MAX_PINS = 40
 /** Characters one pin may carry; a pin is a fact, not a document. */
 export const MAX_PIN_CHARS = 600
 
+/** @internal Keep the omission visible within the existing character budget. */
+export function truncateStateText(text: string, max: number): string {
+	if (max === undefined || text.length <= max) return text
+	const marker = max >= '… [truncated]'.length ? '… [truncated]' : '…'
+	let head = text.slice(0, Math.max(0, max - marker.length))
+	const last = head.charCodeAt(head.length - 1)
+	if (last >= 0xd800 && last <= 0xdbff) head = head.slice(0, -1)
+	return `${head}${marker}`
+}
+
 function createEmptyState(): WorkingState {
 	return {
 		task: '',
@@ -130,7 +140,7 @@ export class WorkingStateManager {
 	}
 
 	addUserRequirement(requirement: string): void {
-		const truncated = requirement.slice(0, this.config.maxCharsPerRequirement)
+		const truncated = truncateStateText(requirement, this.config.maxCharsPerRequirement)
 		this.pushWithEviction(
 			'userRequirements',
 			this.state.userRequirements,

package/src/compaction/verifier.ts

@@ -2,7 +2,7 @@ import type { CompactionConfig } from '../config/runtime.js'
 import { collectChatCompletion } from '../provider/collect-chat-completion.js'
 import { resolveStreamIdleTimeoutMs, withStreamIdleTimeout } from '../provider/idle-timeout.js'
 import type { TokenUsage } from '../types/common/index.js'
-import type { Message } from '../types/message/index.js'
+import type { Message, MessageAttachment, ToolResultBlock } from '../types/message/index.js'
 import type { LLMProvider } from '../types/provider/interface.js'
 import type { WorkingStateManager } from './manager.js'
 import { serializeState } from './serializer.js'
@@ -20,28 +20,90 @@ If the structured state is complete, respond with exactly: COMPLETE
 
 If something important is missing, respond with a brief bullet list of the missing items (no preamble, just the bullets). Each bullet should be a single concise sentence.`
 
-function truncateMessages(messages: Message[], budget: number): string {
-	const lines: string[] = []
-	let charCount = 0
-
-	for (const msg of messages) {
-		if (!msg.content) continue
-		const prefix = `[${msg.role}]: `
-		const content = msg.content
+function* attachmentExcerpt(
+	attachment: MessageAttachment | Exclude<ToolResultBlock, { type: 'text' }>,
+): Generator<string> {
+	const kind =
+		attachment.type === 'stored'
+			? attachment.kind
+			: attachment.type === 'document'
+				? 'document'
+				: 'image'
+	yield `[${kind}`
+	if ('name' in attachment && attachment.name) {
+		yield ' "'
+		yield attachment.name
+		yield '"'
+	}
+	yield ': '
+	yield attachment.mediaType
+	yield ' — not included in this text excerpt]'
+}
 
-		if (charCount + prefix.length + content.length > budget) {
-			const remaining = budget - charCount - prefix.length
-			if (remaining > 0) {
-				lines.push(`${prefix}${content.slice(0, remaining)}...`)
+/** Visible evidence only: never stringify a whole message or its attachment bytes. */
+function* conversationExcerpt(messages: readonly Message[]): Generator<string> {
+	let first = true
+	for (const message of messages) {
+		const hasDetails =
+			message.role === 'tool' ||
+			(message.role === 'assistant' && message.toolCalls?.length) ||
+			(message.role === 'user' && message.attachments?.length)
+		if (!message.content && !hasDetails) continue
+		if (!first) yield '\n\n'
+		first = false
+		yield `[${message.role}]: `
+		if (message.role === 'tool') {
+			yield '[tool result '
+			yield message.toolCallId
+			yield `; isError=${message.isError === true}]\n`
+		}
+		if (typeof message.content === 'string') yield message.content
+		else if (message.content) {
+			for (const [index, block] of message.content.entries()) {
+				if (index > 0) yield '\n'
+				if (block.type === 'text') yield block.text
+				else yield* attachmentExcerpt(block)
+			}
+		}
+		if (message.role === 'assistant') {
+			for (const call of message.toolCalls ?? []) {
+				yield '\n[tool call '
+				yield call.id
+				yield '] '
+				yield call.function.name
+				yield '('
+				yield call.function.arguments
+				yield ')'
+			}
+		}
+		if (message.role === 'user') {
+			for (const attachment of message.attachments ?? []) {
+				yield '\n'
+				yield* attachmentExcerpt(attachment)
 			}
-			break
 		}
-
-		lines.push(`${prefix}${content}`)
-		charCount += prefix.length + content.length
 	}
+}
 
-	return lines.join('\n\n')
+function truncateMessages(messages: readonly Message[], budget: number): string {
+	if (budget <= 0) return ''
+	const parts: string[] = []
+	let remaining = budget
+	for (const part of conversationExcerpt(messages)) {
+		if (part.length > remaining) {
+			parts.push(part.slice(0, remaining))
+			// Role labels, separators and the omission marker all consume the
+			// same budget. Consume fragments lazily so huge arguments and rich
+			// results cannot bypass the cap or require a full serialized copy.
+			let prefix = parts.join('').slice(0, budget - 1)
+			const last = prefix.charCodeAt(prefix.length - 1)
+			if (last >= 0xd800 && last <= 0xdbff) prefix = prefix.slice(0, -1)
+			return `${prefix}…`
+		}
+		parts.push(part)
+		remaining -= part.length
+	}
+	return parts.join('')
 }
 
 /**

package/src/config/runtime.ts

@@ -17,6 +17,8 @@ export const TaskRouterConfigSchema = z
 	.optional()
 
 export const CompactionConfigSchema = z.object({
+	/** Request-only exact duplicate observation masking. Enabled unless false; disabled strategy opts out too. */
+	deduplicateObservations: z.boolean().optional(),
 	/**
 	 * `salience` is the default: every message scored without a model, the
 	 * context held from half the window. `structured` is the previous

package/src/contracts/api.ts

@@ -197,6 +197,7 @@ export type StreamEventType =
 	 * know the answer it is reading came from a provider it did not ask for —
 	 * at a different price, and possibly a different quality.
 	 */
+	| 'hosted.tool'
 	| 'provider.fallback'
 	| 'tool.executing'
 	/** Ephemeral progress from a long-running tool. Not in the transcript. */

package/src/eval/harness-verification.ts

@@ -0,0 +1,270 @@
+import type { CaseResult } from './types.js'
+
+/** A host-owned outcome, with enough identity to check pairing and cite its trace. */
+export interface HarnessTrial {
+	taskId: string
+	trial: number
+	/** Environment/seed/model/limits fingerprint. Equal within a pair, fresh in confirmation. */
+	conditions: string
+	trajectoryId: string
+	result: CaseResult
+}
+
+export interface HarnessAttribution {
+	taskId: string
+	effect: 'improvement' | 'regression' | 'unresolved'
+	reason: string
+	baselineTrajectories: readonly string[]
+	candidateTrajectories: readonly string[]
+}
+
+export interface HarnessVerificationBatch {
+	baselineRevision: string
+	candidateRevision: string
+	baseline: readonly HarnessTrial[]
+	candidate: readonly HarnessTrial[]
+	/** Supplied by an independent trace reviewer, never inferred from scores alone. */
+	attributions: readonly HarnessAttribution[]
+}
+
+export type HarnessBehaviorStatus =
+	| 'recovered'
+	| 'stable-success'
+	| 'regressed'
+	| 'still-failing'
+	| 'mixed'
+	| 'inconclusive'
+
+export interface HarnessTaskComparison {
+	taskId: string
+	status: HarnessBehaviorStatus
+	baselinePasses: number
+	candidatePasses: number
+	trials: number
+}
+
+export interface HarnessComparison {
+	tasks: readonly HarnessTaskComparison[]
+	/** Paired trial pass-rate difference, not held-out pass@1 or causal evidence. */
+	passRateDelta: number | null
+	positiveEvidence: readonly string[]
+	regressions: readonly string[]
+	unresolved: readonly string[]
+	/** Actual recorded consumption from BOTH sides, including unsuccessful runs. */
+	usage: { rollouts: number; tokens: number; costUsd: number; durationMs: number }
+}
+
+function nonempty(value: string): boolean {
+	return typeof value === 'string' && value.trim().length > 0
+}
+
+function key(trial: HarnessTrial): string {
+	return JSON.stringify([trial.taskId, trial.trial])
+}
+
+function indexTrials(trials: readonly HarnessTrial[]): Map<string, HarnessTrial> {
+	const result = new Map<string, HarnessTrial>()
+	const traces = new Set<string>()
+	for (const trial of trials) {
+		if (
+			!nonempty(trial.taskId) ||
+			!nonempty(trial.conditions) ||
+			!nonempty(trial.trajectoryId) ||
+			!Number.isSafeInteger(trial.trial) ||
+			trial.trial < 0
+		)
+			throw new Error(
+				'A harness trial needs a task, nonnegative trial index, conditions and trajectory ID.',
+			)
+		if (result.has(key(trial)) || traces.has(trial.trajectoryId))
+			throw new Error('Duplicate harness trial or trajectory ID.')
+		result.set(key(trial), trial)
+		traces.add(trial.trajectoryId)
+	}
+	return result
+}
+
+function measured(trial: HarnessTrial): boolean {
+	const { result } = trial
+	const scores = Object.values(result.scores)
+	return (
+		!result.run.error &&
+		(result.status === 'passed' || result.status === 'failed') &&
+		scores.length > 0 &&
+		scores.every((s) => !s.unavailable && Number.isFinite(s.score) && s.score >= 0 && s.score <= 1)
+	)
+}
+
+/** HarnessLens Table 6 comparison plus explicit, trace-linked attribution. No model calls. */
+export function compareHarnessTrials(batch: HarnessVerificationBatch): HarnessComparison {
+	if (
+		!nonempty(batch.baselineRevision) ||
+		!nonempty(batch.candidateRevision) ||
+		batch.baselineRevision === batch.candidateRevision
+	)
+		throw new Error('Distinct baseline and candidate revisions are required.')
+	const baseline = indexTrials(batch.baseline)
+	const candidate = indexTrials(batch.candidate)
+	if (!baseline.size || baseline.size !== candidate.size)
+		throw new Error('Complete paired trials are required.')
+	const baselineTraces = new Set(batch.baseline.map((t) => t.trajectoryId))
+	if (batch.candidate.some((t) => baselineTraces.has(t.trajectoryId)))
+		throw new Error('Baseline and candidate must have distinct trajectories.')
+	const groups = new Map<string, HarnessTrial[]>()
+	for (const [id, before] of baseline) {
+		const after = candidate.get(id)
+		if (!after || before.conditions !== after.conditions)
+			throw new Error('Harness trial conditions do not match.')
+		const group = groups.get(before.taskId) ?? []
+		group.push(before)
+		groups.set(before.taskId, group)
+	}
+	const tasks: HarnessTaskComparison[] = []
+	for (const [taskId, before] of groups) {
+		if (new Set(before.map((t) => t.conditions)).size !== before.length)
+			throw new Error('Repeated trials need distinct controlled conditions.')
+		const after = before.map((t) => candidate.get(key(t)) as HarnessTrial)
+		const b = before.filter((t) => t.result.status === 'passed').length
+		const c = after.filter((t) => t.result.status === 'passed').length
+		const status: HarnessBehaviorStatus = ![...before, ...after].every(measured)
+			? 'inconclusive'
+			: b === 0 && c > 0
+				? 'recovered'
+				: b > 0 && c === after.length
+					? 'stable-success'
+					: b > 0 && c === 0
+						? 'regressed'
+						: b === 0 && c === 0
+							? 'still-failing'
+							: 'mixed'
+		tasks.push({ taskId, status, baselinePasses: b, candidatePasses: c, trials: before.length })
+	}
+	const positive = new Set<string>()
+	const regressions = new Set<string>()
+	const unresolved = new Set(
+		tasks.filter((t) => t.status === 'inconclusive' || t.status === 'mixed').map((t) => t.taskId),
+	)
+	// Conservative Namzu policy: even an unattributed observed regression blocks promotion.
+	for (const task of tasks) if (task.status === 'regressed') regressions.add(task.taskId)
+	for (const evidence of batch.attributions) {
+		const task = tasks.find((t) => t.taskId === evidence.taskId)
+		const cites = (ids: readonly string[], trials: readonly HarnessTrial[]) =>
+			ids.length > 0 &&
+			ids.every((id) => trials.some((t) => t.taskId === evidence.taskId && t.trajectoryId === id))
+		if (
+			!task ||
+			!nonempty(evidence.reason) ||
+			!cites(evidence.baselineTrajectories, batch.baseline) ||
+			!cites(evidence.candidateTrajectories, batch.candidate) ||
+			!['improvement', 'regression', 'unresolved'].includes(evidence.effect)
+		)
+			throw new Error('Attribution must explain and cite both sides of an existing task.')
+		if (evidence.effect === 'regression') regressions.add(task.taskId)
+		else if (evidence.effect === 'unresolved') unresolved.add(task.taskId)
+		else if (
+			task.status === 'recovered' ||
+			(task.status === 'stable-success' && task.candidatePasses > task.baselinePasses)
+		)
+			positive.add(task.taskId)
+	}
+	const usage = { rollouts: baseline.size * 2, tokens: 0, costUsd: 0, durationMs: 0 }
+	for (const trial of [...batch.baseline, ...batch.candidate]) {
+		const values = [
+			trial.result.run.totalTokens,
+			trial.result.run.totalCostUsd,
+			trial.result.run.durationMs,
+		]
+		if (values.some((n) => !Number.isFinite(n) || n < 0))
+			throw new Error('Invalid recorded harness usage.')
+		usage.tokens += trial.result.run.totalTokens
+		usage.costUsd += trial.result.run.totalCostUsd
+		usage.durationMs += trial.result.run.durationMs
+	}
+	return {
+		tasks,
+		positiveEvidence: [...positive],
+		regressions: [...regressions],
+		unresolved: [...unresolved],
+		usage,
+		passRateDelta: tasks.some((t) => t.status === 'inconclusive')
+			? null
+			: tasks.reduce((sum, t) => sum + t.candidatePasses - t.baselinePasses, 0) / baseline.size,
+	}
+}
+
+export interface HarnessReview {
+	decision: 'accept' | 'reject' | 'inconclusive'
+	reason: string
+	verification: HarnessComparison
+	confirmation?: HarnessComparison
+}
+
+/**
+ * Evaluate recorded evidence; never applies a patch. Requires at least five tasks,
+ * exactly two trials per task, and a same-size confirmation with at most two reused tasks.
+ * Hosts own isolated execution, task selection, scoring and independent trace review.
+ */
+export function reviewHarnessCandidate(
+	verification: HarnessVerificationBatch,
+	confirmation?: HarnessVerificationBatch,
+): HarnessReview {
+	const first = compareHarnessTrials(verification)
+	const second = confirmation ? compareHarnessTrials(confirmation) : undefined
+	const result = (decision: HarnessReview['decision'], reason: string): HarnessReview => ({
+		decision,
+		reason,
+		verification: first,
+		...(second ? { confirmation: second } : {}),
+	})
+	const sufficient = (report: HarnessComparison) =>
+		report.tasks.length >= 5 && report.tasks.every((t) => t.trials === 2)
+	if (!sufficient(first))
+		return result(
+			'inconclusive',
+			'Verification requires at least five distinct tasks and two paired trials per task.',
+		)
+	if (first.regressions.length || second?.regressions.length)
+		return result('reject', 'Observed or attributed regression blocks promotion.')
+	if (first.unresolved.length || second?.unresolved.length)
+		return result('inconclusive', 'Incomplete or ambiguous behavioral evidence.')
+	if (!first.positiveEvidence.length)
+		return result(
+			'reject',
+			'A metric increase or preservation alone is not attributable improvement.',
+		)
+	if (!confirmation || !second)
+		return result('inconclusive', 'A complete fresh confirmation round is required.')
+	if (
+		verification.baselineRevision !== confirmation.baselineRevision ||
+		verification.candidateRevision !== confirmation.candidateRevision
+	)
+		return result('inconclusive', 'Revisions changed between verification and confirmation.')
+	const firstTasks = new Set(first.tasks.map((t) => t.taskId))
+	if (
+		!sufficient(second) ||
+		second.tasks.length !== first.tasks.length ||
+		second.tasks.filter((t) => firstTasks.has(t.taskId)).length > 2
+	)
+		return result(
+			'inconclusive',
+			'Confirmation must use the same batch size with at most two reused tasks.',
+		)
+	const conditions = new Set(verification.baseline.map((t) => t.conditions))
+	const traces = new Set(
+		[...verification.baseline, ...verification.candidate].map((t) => t.trajectoryId),
+	)
+	if (
+		confirmation.baseline.some((t) => conditions.has(t.conditions)) ||
+		[...confirmation.baseline, ...confirmation.candidate].some((t) => traces.has(t.trajectoryId))
+	)
+		return result('inconclusive', 'Confirmation requires fresh conditions and trajectories.')
+	if (!second.positiveEvidence.length || second.passRateDelta === null || second.passRateDelta <= 0)
+		return result(
+			'reject',
+			'Confirmation needs attributable improvement and a positive primary-metric difference.',
+		)
+	return result(
+		'accept',
+		'Both rounds support improvement without observed regression; confirmation improves the paired pass rate.',
+	)
+}

package/src/eval/index.ts

@@ -1,4 +1,14 @@
 export { evalRunFromQuery, evalRunFromRun } from './from-run.js'
+export { compareHarnessTrials, reviewHarnessCandidate } from './harness-verification.js'
+export type {
+	HarnessTrial,
+	HarnessAttribution,
+	HarnessVerificationBatch,
+	HarnessBehaviorStatus,
+	HarnessTaskComparison,
+	HarnessComparison,
+	HarnessReview,
+} from './harness-verification.js'
 export { formatReport, runExperiment } from './experiment.js'
 export type { ExperimentConfig } from './experiment.js'
 export { judgeScorer } from './judge.js'