@wix/pathgrade 1.0.30 → 1.0.32

package/dist/runners/adapter-loader.d.ts

@@ -2,9 +2,6 @@ import type { ResolvedPathgradeConfig } from '../config/pathgrade.js';
 import { type SpawnVitest } from './vitest-invocation.js';
 import type { RunnerAdapter } from './adapter.js';
 import type { RunnerInvocationAdapter } from './invocation.js';
-export interface ExternalAdapterFactoryInput {
-    config: ResolvedPathgradeConfig;
-}
 export declare function loadRunnerAdapter(input: {
     adapterName?: string;
     cwd?: string;

package/dist/runners/model-builders.d.ts

@@ -1,4 +1,6 @@
 import type { ReportGroupInput } from '../reporting/types.js';
 import type { AdapterRunHandle } from './adapter.js';
 import type { NormalizedRunSnapshot } from './model.js';
-export declare function buildNormalizedRunSnapshotFromReportGroups(run: AdapterRunHandle, groups: ReportGroupInput[]): NormalizedRunSnapshot;
+export declare function buildNormalizedRunSnapshotFromReportGroups(run: AdapterRunHandle, groups: ReportGroupInput[], options?: {
+    cwd?: string;
+}): NormalizedRunSnapshot;

package/dist/runners/model-builders.js

@@ -1,4 +1,6 @@
-export function buildNormalizedRunSnapshotFromReportGroups(run, groups) {
+import { resolveSourceMetadata } from '../reporting/source-metadata.js';
+export function buildNormalizedRunSnapshotFromReportGroups(run, groups, options = {}) {
+    const cwd = options.cwd ?? process.cwd();
     return {
         version: 1,
         completeness: 'final',
@@ -12,12 +14,18 @@ export function buildNormalizedRunSnapshotFromReportGroups(run, groups) {
                         message: diagnostic.message,
                     })) } : {}),
             },
-            units: groups.map((group, index) => ({
-                id: unitId(group.groupName, index),
-                runId: `${run.adapterName}:run`,
-                displayName: group.groupName,
-                groupingHints: [{ kind: 'suite', key: group.groupName, label: group.groupName, order: index }],
-            })),
+            units: groups.map((group, index) => {
+                const source = group.sourceFile && group.sourceRevision
+                    ? { sourceFile: group.sourceFile, sourceRevision: group.sourceRevision }
+                    : resolveSourceMetadata(group.sourceFile ?? group.cases.map(testCase => testCase.filePath ?? testCase.sourceRef).find(Boolean), cwd);
+                return {
+                    id: unitId(group.groupName, index),
+                    runId: `${run.adapterName}:run`,
+                    displayName: group.groupName,
+                    ...source,
+                    groupingHints: [{ kind: 'suite', key: group.groupName, label: group.groupName, order: index }],
+                };
+            }),
             cases: groups.flatMap((group, groupIndex) => group.cases.map((testCase, caseIndex) => {
                 const caseId = testCase.caseId ?? `${unitId(group.groupName, groupIndex)}:case-${caseIndex + 1}`;
                 const attemptId = `${caseId}:attempt-1`;
@@ -33,6 +41,10 @@ export function buildNormalizedRunSnapshotFromReportGroups(run, groups) {
                         : {}),
                     ...(evaluation.recordedAt ? { recordedAt: evaluation.recordedAt } : {}),
                     ...(evaluation.agent ? { agent: evaluation.agent } : {}),
+                    ...(evaluation.evaluationDefinitionKey
+                        ? { evaluationDefinitionKey: evaluation.evaluationDefinitionKey }
+                        : {}),
+                    ...(evaluation.scorerRevision ? { scorerRevision: evaluation.scorerRevision } : {}),
                 }));
                 return {
                     id: caseId,

package/dist/runners/model-validation.d.ts

@@ -3,10 +3,11 @@ export interface NormalizedRunValidationResult {
     ok: boolean;
     errors: NormalizedRunValidationError[];
 }
-export interface NormalizedRunValidationError {
+interface NormalizedRunValidationError {
     path: string;
     message: string;
 }
 export declare function validateNormalizedRunSnapshot(snapshot: NormalizedRunSnapshot, options: {
     completeness: SnapshotCompleteness;
 }): NormalizedRunValidationResult;
+export {};

package/dist/runners/model.d.ts

@@ -25,6 +25,8 @@ export interface EvalUnitRecord {
     id: string;
     runId: string;
     displayName: string;
+    sourceFile?: string;
+    sourceRevision?: string;
     diagnostics?: Diagnostic[];
     nativeReferences?: NativeReference[];
     groupingHints?: GroupingHint[];
@@ -92,6 +94,8 @@ export interface EvaluationRecord {
     scoringDurationMs?: number;
     recordedAt?: string;
     agent?: import('../sdk/types.js').AgentExecutionMetadata;
+    evaluationDefinitionKey?: string;
+    scorerRevision?: string;
 }
 export interface AssertionRecord {
     id: string;

package/dist/runners/repeated-attempts.d.ts

@@ -1,7 +1,2 @@
 import type { NormalizedRunSnapshot } from './model.js';
-export type RepeatAggregationFailure = 'incomplete_attempts' | 'case_identity_mismatch';
-export declare class RepeatAggregationError extends Error {
-    readonly reason: RepeatAggregationFailure;
-    constructor(reason: RepeatAggregationFailure, message: string);
-}
 export declare function mergeRepeatedRunSnapshots(snapshots: readonly NormalizedRunSnapshot[]): NormalizedRunSnapshot;

package/dist/runners/repeated-attempts.js

@@ -1,6 +1,6 @@
 import { createHash, randomUUID } from 'node:crypto';
 import { validateNormalizedRunSnapshot } from './model-validation.js';
-export class RepeatAggregationError extends Error {
+class RepeatAggregationError extends Error {
     reason;
     constructor(reason, message) {
         super(message);

package/dist/runners/report-projection.js

@@ -16,9 +16,15 @@ export function projectNormalizedRunSnapshotToReportInput(snapshot, options = {}
             ?? preferredGroupingLabel(unit?.groupingHints)
             ?? unit?.displayName;
         const key = groupName ?? snapshot.model.run.adapterName;
-        const group = groups.get(key) ?? { groupName: key, cases: [] };
+        const group = groups.get(key) ?? {
+            groupName: key,
+            ...(unit?.sourceFile ? { sourceFile: unit.sourceFile } : {}),
+            ...(unit?.sourceRevision ? { sourceRevision: unit.sourceRevision } : {}),
+            cases: [],
+        };
         group.cases.push({
             caseId: runCase.id,
+            ...(runCase.repeatKey ? { repeatKey: runCase.repeatKey } : {}),
             name: runCase.name,
             state: runCase.state,
             ...(runCase.state === 'skipped' || runCase.state === 'pending' ? { reportable: repeated || runCase.scoringPolicy.kind !== 'non-scoring' } : {}),
@@ -53,6 +59,10 @@ function projectedEvaluations(runCase, attemptEvaluations) {
             ...(evaluation.scoringDurationMs !== undefined ? { scoringDurationMs: evaluation.scoringDurationMs } : {}),
             ...(evaluation.recordedAt ? { recordedAt: evaluation.recordedAt } : {}),
             ...(evaluation.agent ? { agent: evaluation.agent } : {}),
+            ...(evaluation.evaluationDefinitionKey
+                ? { evaluationDefinitionKey: evaluation.evaluationDefinitionKey }
+                : {}),
+            ...(evaluation.scorerRevision ? { scorerRevision: evaluation.scorerRevision } : {}),
         }));
     }
     if (runCase.scoringPolicy.kind === 'score') {
@@ -70,6 +80,10 @@ function projectedEvaluations(runCase, attemptEvaluations) {
             ...(evaluation.scoringDurationMs !== undefined ? { scoringDurationMs: evaluation.scoringDurationMs } : {}),
             ...(evaluation.recordedAt ? { recordedAt: evaluation.recordedAt } : {}),
             ...(evaluation.agent ? { agent: evaluation.agent } : {}),
+            ...(evaluation.evaluationDefinitionKey
+                ? { evaluationDefinitionKey: evaluation.evaluationDefinitionKey }
+                : {}),
+            ...(evaluation.scorerRevision ? { scorerRevision: evaluation.scorerRevision } : {}),
         }];
 }
 function preferredGroupingLabel(hints) {

package/dist/runners/vitest-adapter.js

@@ -2,9 +2,11 @@ import { discoverPathgradeEvalFiles } from '../evals/discovery.js';
 import { buildNormalizedRunSnapshotFromReportGroups } from './model-builders.js';
 import { buildDiagnosticsReport } from '../sdk/diagnostics.js';
 export function createVitestAdapter(options = {}) {
+    let discoveryCwd = process.cwd();
     return {
         name: 'vitest',
         async discover(input) {
+            discoveryCwd = input.cwd;
             const include = input.include ?? ['**/*.eval.ts'];
             const exclude = input.exclude ?? [];
             const files = discoverPathgradeEvalFiles({
@@ -28,11 +30,12 @@ export function createVitestAdapter(options = {}) {
                 exitCode: input.signal?.aborted ? 1 : 0,
                 native: {
                     testModules: options.testModules ?? [],
+                    cwd: discoveryCwd,
                 },
             };
         },
         async collectNormalizedRunSnapshot(run) {
-            return buildNormalizedRunSnapshotFromReportGroups(run, collectVitestReportGroups(readVitestRunNative(run).testModules));
+            return buildNormalizedRunSnapshotFromReportGroups(run, collectVitestReportGroups(readVitestRunNative(run).testModules), { cwd: readVitestRunNative(run).cwd });
         },
     };
 }
@@ -56,12 +59,13 @@ export function collectVitestReportGroups(testModules) {
 function readVitestRunNative(run) {
     if (isVitestRunNative(run.native))
         return run.native;
-    return { testModules: [] };
+    return { testModules: [], cwd: process.cwd() };
 }
 function isVitestRunNative(native) {
     return typeof native === 'object'
         && native !== null
-        && Array.isArray(native.testModules);
+        && Array.isArray(native.testModules)
+        && typeof native.cwd === 'string';
 }
 function getGroupName(testCase) {
     const modulePath = testCase.module.relativeModuleId;
@@ -94,6 +98,8 @@ function toReportCaseInput(testCase, groupName) {
             scoringDurationMs: entry.scoringDurationMs,
             recordedAt: entry.recordedAt,
             agent: entry.agent,
+            evaluationDefinitionKey: entry.evaluationDefinitionKey,
+            scorerRevision: entry.scorerRevision,
         })),
         diagnostics: normalized.diagnostics,
     };

package/dist/runners/vitest-invocation.d.ts

@@ -1,5 +1,5 @@
 import type { RunnerInvocationAdapter } from './invocation.js';
-export interface SpawnVitestRequest {
+interface SpawnVitestRequest {
     argv: string[];
     env: NodeJS.ProcessEnv;
     cwd: string;
@@ -8,4 +8,4 @@ export type SpawnVitest = (req: SpawnVitestRequest) => Promise<number> | number;
 export declare function createVitestInvocationAdapter(input?: {
     spawnVitest?: SpawnVitest;
 }): RunnerInvocationAdapter;
-export declare function resolveVitestCli(cwd: string): string;
+export {};

package/dist/runners/vitest-invocation.js

@@ -47,7 +47,7 @@ async function defaultSpawnVitest(req) {
         child.on('close', code => resolve(code ?? 0));
     });
 }
-export function resolveVitestCli(cwd) {
+function resolveVitestCli(cwd) {
     const requireFromCwd = createRequire(path.join(cwd, 'package.json'));
     const packageJsonPath = requireFromCwd.resolve('vitest/package.json');
     const packageJson = JSON.parse(fs.readFileSync(packageJsonPath, 'utf8'));

package/dist/runners/vitest-lifecycle.d.ts

@@ -1,7 +1,6 @@
 import type { Agent, PathgradeTestMeta } from '../sdk/types.js';
 import { type ResultObserverHandle } from '../sdk/result-capture.js';
 import { type CaseContextProviderHandle } from '../sdk/case-context.js';
-import { type LifecycleAgentOwner } from '../sdk/lifecycle.js';
 import type { AdapterLifecycleHooks, EvalResultEvent } from './adapter.js';
 export type VitestAfterEachFn = (fn: (ctx: {
     task: {
@@ -31,7 +30,6 @@ export interface VitestLifecycleInstallHandle {
 }
 export declare function createVitestLifecycleHooks(): AdapterLifecycleHooks;
 export declare function installVitestLifecycle(input: VitestLifecycleInstallInput): VitestLifecycleInstallHandle;
-export declare function currentVitestAgentOwner(resultCaptureActive: boolean): LifecycleAgentOwner | null;
 export declare function trackVitestAgent(agent: Agent, resultCaptureActive: boolean): void;
 export declare function untrackVitestAgent(agent: Agent): void;
 export declare function releaseVitestAgent(agent: Agent): void;

package/dist/runners/vitest-lifecycle.js

@@ -43,7 +43,7 @@ export function installVitestLifecycle(input) {
         fileContextHandle.restore();
     }
 }
-export function currentVitestAgentOwner(resultCaptureActive) {
+function currentVitestAgentOwner(resultCaptureActive) {
     const current = getCurrentCaseContext();
     if (current.status === 'active') {
         return current.context.scope === 'runner-case'

package/dist/sdk/ask-bus/handler.d.ts

@@ -1,12 +1,12 @@
 import type { AskHandler } from './types.js';
 import type { AskUserReactionPreviewEntry, Reaction } from '../types.js';
-export type UnmatchedAskUserDisposition = 'error' | 'first-option' | 'decline';
+type UnmatchedAskUserDisposition = 'error' | 'first-option' | 'decline';
 export interface AskUserHandlerContext {
     reactions: readonly Reaction[];
     onUnmatchedAskUser: UnmatchedAskUserDisposition;
     firedOnce: Set<number>;
 }
-export interface AskUserUnmatchedSignal {
+interface AskUserUnmatchedSignal {
     batchId: string;
     turnNumber: number;
     /**
@@ -29,3 +29,4 @@ export interface AskUserHandlerApi {
     getUnmatchedError(): AskUserUnmatchedSignal | null;
 }
 export declare function createAskUserHandler(ctx: AskUserHandlerContext): AskUserHandlerApi;
+export {};

package/dist/sdk/diagnostics.d.ts

@@ -1,7 +1,7 @@
 import type { LogEntry } from '../types.js';
 import type { ReactionFiredEntry, TurnDetail } from './types.js';
 import { type RuntimePolicyAuditEntry } from './runtime-policy.js';
-export interface DiagnosticsScorer {
+interface DiagnosticsScorer {
     name: string;
     type: string;
     score: number;
@@ -9,10 +9,10 @@ export interface DiagnosticsScorer {
     details?: string;
     status?: 'ok' | 'error' | 'skipped';
 }
-export interface DiagnosticsTurnDetail extends TurnDetail {
+interface DiagnosticsTurnDetail extends TurnDetail {
     apiRetries: number;
 }
-export type DiagnosticsRuntimePolicyEntry = RuntimePolicyAuditEntry;
+type DiagnosticsRuntimePolicyEntry = RuntimePolicyAuditEntry;
 export interface DiagnosticsReport {
     turns: number;
     totalDurationMs: number;
@@ -39,3 +39,4 @@ export interface BuildDiagnosticsReportInput {
     log: LogEntry[];
 }
 export declare function buildDiagnosticsReport(input: BuildDiagnosticsReportInput): DiagnosticsReport;
+export {};

package/dist/sdk/evaluate.js

@@ -1,4 +1,5 @@
 import { countShellCommandsFromLog, extractSkillsFromLog, extractToolEventsFromLog } from '../tool-events.js';
+import { createScorerRevision } from '../reporting/comparison-contract.js';
 import { getRuntime } from './eval-runtime.js';
 import { emitEvalResult } from './result-capture.js';
 import { runJudgePipeline } from './judge-pipeline.js';
@@ -73,6 +74,7 @@ function makeEvaluateAgent() {
             : undefined;
         if (isFirstEval)
             conversationAttributed.add(agent);
+        const scorerRevision = createScorerRevision(scorers);
         const recordedResult = {
             ...evalResult,
             tokenUsage: deltaTokenUsage,
@@ -82,6 +84,7 @@ function makeEvaluateAgent() {
             ...(opts?.evaluationDefinitionKey
                 ? { evaluationDefinitionKey: opts.evaluationDefinitionKey }
                 : {}),
+            ...(scorerRevision ? { scorerRevision } : {}),
             trial: buildTrialResult(agent.log, { ...evalResult, tokenUsage: deltaTokenUsage }, agent.scenarioEvidence, conversationTokens, conversationCost),
         };
         emitEvalResult({ result: recordedResult, agent });
@@ -120,6 +123,7 @@ async function fromSnapshot(snapshotPath, scorers, opts) {
     };
     const scoringStartedAt = performance.now();
     const evalResult = await evaluateWithContext(ctx, scorers, { ...opts, llm: trackedLLM });
+    const scorerRevision = createScorerRevision(scorers);
     const recordedResult = {
         ...evalResult,
         resultKind: 'snapshot',
@@ -128,6 +132,7 @@ async function fromSnapshot(snapshotPath, scorers, opts) {
         ...(opts?.evaluationDefinitionKey
             ? { evaluationDefinitionKey: opts.evaluationDefinitionKey }
             : {}),
+        ...(scorerRevision ? { scorerRevision } : {}),
         trial: buildTrialResult(snapshot.log, evalResult, snapshot.scenarioEvidence),
     };
     maybeThrowOnScorerErrors(recordedResult, opts?.onScorerError ?? 'skip');

package/dist/sdk/index.d.ts

@@ -21,6 +21,7 @@ export { DEFAULT_COPY_IGNORE } from '../providers/copy-filter.js';
 export { extractToolEventsFromLog } from '../tool-events.js';
 export { collectSensitiveEnvValues, sanitizePersistenceValue, } from '../tool-event-results.js';
 export { parseEnvFile } from '../utils/env.js';
+export { parsePathgradeReport } from '../reporting/report-parser.js';
 export { createAskBus, requireAskBusForLiveBatches, AskBusTimeoutError } from './ask-bus/bus.js';
 export { toAskUserToolEvent } from './ask-bus/projection.js';
 export type { AskUserToolEvent, AskUserToolEventArguments, AskUserToolEventQuestionArgument, } from './ask-bus/projection.js';
@@ -40,6 +41,7 @@ export type { DiagnosticsReport } from './diagnostics.js';
 export type { ExpectedMcpStartupStatus, ExpectedMcpToolCall, McpStartupStatusEvidence, McpToolCallEvidence, McpApprovalEvidence, ExpectedMcpApproval, McpInvocationResult, } from './mcp-evidence.js';
 export type { McpPolicyDenialReason, McpToolCallRequest, McpToolPolicyDecision, } from './mcp-safety.js';
 export type { ToolEvent, McpToolCallClassification } from '../tool-events.js';
+export type { ComparisonContract, ComparisonUnavailableReason, PathgradeGroupReport, PathgradeReport, StrippedTrialResult, } from '../types.js';
 export type { LLMPort, EvalRuntime } from './eval-runtime.js';
 export { createAgentLLM, createLLMClient, ProviderNotSupportedError } from '../utils/llm.js';
 export type { CreateLLMClientOptions, LLMProviderAdapter, TokenUsage as LLMTokenUsage } from '../utils/llm.js';

package/dist/sdk/index.js

@@ -19,6 +19,7 @@ export { DEFAULT_COPY_IGNORE } from '../providers/copy-filter.js';
 export { extractToolEventsFromLog } from '../tool-events.js';
 export { collectSensitiveEnvValues, sanitizePersistenceValue, } from '../tool-event-results.js';
 export { parseEnvFile } from '../utils/env.js';
+export { parsePathgradeReport } from '../reporting/report-parser.js';
 export { createAskBus, requireAskBusForLiveBatches, AskBusTimeoutError } from './ask-bus/bus.js';
 export { toAskUserToolEvent } from './ask-bus/projection.js';
 export { buildAskBatchLogEntries } from './agent-result-log.js';

package/dist/sdk/judge-tool-runner.d.ts

@@ -1,7 +1,6 @@
 import type { ToolCapableLLMPort } from '../utils/llm-types.js';
 import type { JudgeScorer, ScorerContext, ScorerResultEntry, TokenUsage } from './types.js';
-import { type JudgeErrorCode, type JudgeToolCallRecord, type ToolExecutor } from './judge-tool-session.js';
-export type { JudgeErrorCode, JudgeToolCallRecord };
+import { type JudgeToolCallRecord, type ToolExecutor } from './judge-tool-session.js';
 export interface RunJudgeWithToolsResult {
     entry: ScorerResultEntry;
     rounds: number;

package/dist/sdk/lifecycle.js

@@ -81,6 +81,7 @@ function recordResult(result, agent, attribution) {
             ...(result.evaluationDefinitionKey
                 ? { evaluationDefinitionKey: result.evaluationDefinitionKey }
                 : {}),
+            ...(result.scorerRevision ? { scorerRevision: result.scorerRevision } : {}),
             trial: result.trial,
             resultKind: result.resultKind ?? 'evaluated',
             ...(result.scoringDurationMs !== undefined

package/dist/sdk/mcp-mock-approvals.d.ts

@@ -9,7 +9,7 @@ export interface McpMockApprovalRule {
     decision: 'approve' | 'deny';
     label?: string;
 }
-export interface CompiledMcpTool {
+interface CompiledMcpTool {
     name: string;
     description?: string;
     inputSchema: Record<string, unknown>;
@@ -19,7 +19,7 @@ export interface CompiledMcpTool {
         response: McpMockJsonValue;
     }>[];
 }
-export interface CompiledMcpServer {
+interface CompiledMcpServer {
     name: string;
     instructions?: string;
     tools: ReadonlyMap<string, CompiledMcpTool>;
@@ -38,3 +38,4 @@ export declare function matchesMcpApprovalArguments(actual: Record<string, McpMo
 export declare function canonicalizeMcpArguments(value: McpMockJsonValue): string;
 export declare function createMcpArgumentDigest(key: Buffer, value: McpMockJsonValue): string;
 export declare function createMcpReceiptKey(): Buffer;
+export {};

package/dist/sdk/reaction-loader.d.ts

@@ -1,4 +1,3 @@
-import type { AskUserOption, AskUserQuestion, AskUserReaction, Message, Reaction, TextReaction } from './types.js';
+import type { Message, Reaction } from './types.js';
 export declare function loadReactionSnapshotMessages(snapshotPath: string): Promise<Message[]>;
 export declare function loadReactionsFromFile(filePath: string): Promise<Reaction[]>;
-export type { AskUserOption, AskUserQuestion, AskUserReaction, TextReaction };

package/dist/sdk/reaction-preview.d.ts

@@ -1,6 +1,3 @@
-import type { AskUserReaction, Message, Reaction, ReactionPreviewEntry, ReactionPreviewResult, TextReaction } from './types.js';
-export declare function isAskUserReaction(reaction: Reaction): reaction is AskUserReaction;
-export declare function isTextReaction(reaction: Reaction): reaction is TextReaction;
+import type { Message, Reaction, ReactionPreviewEntry, ReactionPreviewResult } from './types.js';
 export declare function inspectReactions(agentMessage: string, reactions: Reaction[], firedOnce: Set<number>): ReactionPreviewEntry[];
-export declare function pickReactionReply(agentMessage: string, reactions: Reaction[], firedOnce: Set<number>): string | null;
 export declare function previewReactions(messages: Message[], reactions: Reaction[]): ReactionPreviewResult;

package/dist/sdk/reaction-preview.js

@@ -20,10 +20,10 @@ function testUnless(pattern, input) {
         return false;
     }
 }
-export function isAskUserReaction(reaction) {
+function isAskUserReaction(reaction) {
     return 'whenAsked' in reaction;
 }
-export function isTextReaction(reaction) {
+function isTextReaction(reaction) {
     return 'when' in reaction;
 }
 export function inspectReactions(agentMessage, reactions, firedOnce) {
@@ -71,7 +71,7 @@ export function inspectReactions(agentMessage, reactions, firedOnce) {
         };
     });
 }
-export function pickReactionReply(agentMessage, reactions, firedOnce) {
+function pickReactionReply(agentMessage, reactions, firedOnce) {
     const evaluations = inspectReactions(agentMessage, reactions, firedOnce);
     const match = evaluations.find((evaluation) => evaluation.fired && evaluation.kind === 'text');
     return match && match.kind === 'text' ? match.reply ?? null : null;

package/dist/sdk/run-scorer.d.ts

@@ -1,10 +1,7 @@
 import type { LLMPort } from '../utils/llm-types.js';
-import type { CheckScorer, Scorer, ScoreScorer, ScorerContext, ScorerResultEntry, ToolUsageScorer } from './types.js';
+import type { Scorer, ScorerContext, ScorerResultEntry } from './types.js';
 export interface RunScorerOptions {
     /** Required for `judge` scorers; ignored for other types. */
     llm?: LLMPort;
 }
 export declare function runScorer(scorer: Scorer, ctx: ScorerContext, opts?: RunScorerOptions): Promise<ScorerResultEntry>;
-export declare function runCheckScorer(scorer: CheckScorer, ctx: ScorerContext): Promise<ScorerResultEntry>;
-export declare function runScoreScorer(scorer: ScoreScorer, ctx: ScorerContext): Promise<ScorerResultEntry>;
-export declare function runToolUsageScorer(scorer: ToolUsageScorer, ctx: ScorerContext): Promise<ScorerResultEntry>;

package/dist/sdk/run-scorer.js

@@ -43,7 +43,7 @@ function buildEntryFromOutcome(scorer, outcome) {
         errorCode: outcome.code,
     };
 }
-export async function runCheckScorer(scorer, ctx) {
+async function runCheckScorer(scorer, ctx) {
     try {
         const passed = await scorer.fn(ctx);
         return {
@@ -59,7 +59,7 @@ export async function runCheckScorer(scorer, ctx) {
         return makeErroredResult('check', scorer.name, scorer.weight, error);
     }
 }
-export async function runScoreScorer(scorer, ctx) {
+async function runScoreScorer(scorer, ctx) {
     try {
         const raw = await scorer.fn(ctx);
         if (typeof raw === 'number') {
@@ -84,7 +84,7 @@ export async function runScoreScorer(scorer, ctx) {
         return makeErroredResult('score', scorer.name, scorer.weight, error);
     }
 }
-export async function runToolUsageScorer(scorer, ctx) {
+async function runToolUsageScorer(scorer, ctx) {
     try {
         const toolEvents = ctx.toolEvents;
         if (toolEvents.length === 0) {

package/dist/sdk/runtime-policy.d.ts

@@ -15,7 +15,6 @@ export interface RuntimePolicyAuditEntry {
 export declare const NONINTERACTIVE_RUNTIME_POLICY: RuntimePolicyDescriptor;
 export declare function planRuntimePolicies(agent: AgentName, transport?: AgentTransport): RuntimePolicyDescriptor[];
 export declare function renderRuntimePolicy(policy: RuntimePolicyDescriptor, context: RuntimePolicyRenderContext): string;
-export declare function renderRuntimePolicies(policies: RuntimePolicyDescriptor[], context: RuntimePolicyRenderContext): string;
 export declare function prependRuntimePolicies(instruction: string, policies: RuntimePolicyDescriptor[], context: RuntimePolicyRenderContext): string;
 export declare function getRuntimePolicyLogMetadata(turnResult: AgentTurnResult): Record<string, unknown>;
 export declare function extractRuntimePolicyAuditEntries(log: LogEntry[]): RuntimePolicyAuditEntry[];

package/dist/sdk/runtime-policy.js

@@ -32,7 +32,7 @@ export function renderRuntimePolicy(policy, context) {
         'Ask only the earliest outstanding approval checkpoint.',
     ].join('\n');
 }
-export function renderRuntimePolicies(policies, context) {
+function renderRuntimePolicies(policies, context) {
     return policies
         .map((policy) => renderRuntimePolicy(policy, context))
         .filter((policy) => policy.length > 0)

package/dist/sdk/types.d.ts

@@ -372,6 +372,8 @@ export interface RecordedEvalResult extends Omit<EvalResult, 'score'> {
     scoringDurationMs?: number;
     recordedAt?: string;
     evaluationDefinitionKey?: string;
+    /** Privacy-safe revision of the declared scorers used by this evaluation. */
+    scorerRevision?: string;
 }
 export type EvaluationResultKind = 'evaluated' | 'synthetic_no_evaluation' | 'snapshot';
 export interface ScorerResultEntry {
@@ -393,6 +395,7 @@ export interface PathgradeTestMeta {
     scorers: ScorerResultEntry[];
     /** Stable identity for this evaluation definition across separate runs. */
     evaluationDefinitionKey?: string;
+    scorerRevision?: string;
     trial?: TrialResult;
     diagnostics?: DiagnosticsReport;
     resultKind?: EvaluationResultKind;

package/dist/tool-events.d.ts

@@ -44,7 +44,6 @@ export interface ToolEventResult {
     exitCode?: number;
     truncated?: boolean;
 }
-export declare function summarizeToolEvents(events: ToolEvent[]): string;
 /**
  * Map from provider-specific tool names to normalized Pathgrade actions.
  * Conservative: only map names we're confident about.

package/dist/tool-events.js

@@ -1,4 +1,4 @@
-export function summarizeToolEvents(events) {
+function summarizeToolEvents(events) {
     return events.map((event) => `${event.action}:${event.providerToolName}`).join(', ');
 }
 /**