@ls-stack/agent-eval 0.16.1 → 0.17.0

package/dist/index.d.mts

@@ -666,6 +666,46 @@ declare const assertionFailureSchema: z$1.ZodObject<{
 }, z$1.core.$strip>;
 /** Assertion failure metadata captured for one case run. */
 type AssertionFailure = z$1.infer<typeof assertionFailureSchema>;
+/** Severity level for one log captured during a case run. */
+declare const runLogLevelSchema: z$1.ZodEnum<{
+  error: "error";
+  log: "log";
+  info: "info";
+  warn: "warn";
+}>;
+/** Severity level for one log captured during a case run. */
+type RunLogLevel = z$1.infer<typeof runLogLevelSchema>;
+/** Eval runner phase that emitted a captured case log. */
+declare const runLogPhaseSchema: z$1.ZodEnum<{
+  eval: "eval";
+  derive: "derive";
+  outputsSchema: "outputsSchema";
+  scorer: "scorer";
+}>;
+/** Eval runner phase that emitted a captured case log. */
+type RunLogPhase = z$1.infer<typeof runLogPhaseSchema>;
+/** Schema for one persisted log entry captured during a case run. */
+declare const runLogEntrySchema: z$1.ZodObject<{
+  timestamp: z$1.ZodString;
+  level: z$1.ZodEnum<{
+    error: "error";
+    log: "log";
+    info: "info";
+    warn: "warn";
+  }>;
+  phase: z$1.ZodEnum<{
+    eval: "eval";
+    derive: "derive";
+    outputsSchema: "outputsSchema";
+    scorer: "scorer";
+  }>;
+  message: z$1.ZodString;
+  args: z$1.ZodDefault<z$1.ZodArray<z$1.ZodUnknown>>;
+  truncated: z$1.ZodDefault<z$1.ZodBoolean>;
+  source: z$1.ZodOptional<z$1.ZodString>;
+}, z$1.core.$strip>;
+/** Persisted log entry captured during a case run. */
+type RunLogEntry = z$1.infer<typeof runLogEntrySchema>;
 /** Trace payload captured while computing one score for a case. */
 declare const scoreTraceSchema: z$1.ZodObject<{
   trace: z$1.ZodArray<z$1.ZodObject<{
@@ -907,6 +947,25 @@ declare const caseDetailSchema: z$1.ZodObject<{
     message: string;
     stack?: string | undefined;
   }, string>>]>>;
+  logs: z$1.ZodDefault<z$1.ZodArray<z$1.ZodObject<{
+    timestamp: z$1.ZodString;
+    level: z$1.ZodEnum<{
+      error: "error";
+      log: "log";
+      info: "info";
+      warn: "warn";
+    }>;
+    phase: z$1.ZodEnum<{
+      eval: "eval";
+      derive: "derive";
+      outputsSchema: "outputsSchema";
+      scorer: "scorer";
+    }>;
+    message: z$1.ZodString;
+    args: z$1.ZodDefault<z$1.ZodArray<z$1.ZodUnknown>>;
+    truncated: z$1.ZodDefault<z$1.ZodBoolean>;
+    source: z$1.ZodOptional<z$1.ZodString>;
+  }, z$1.core.$strip>>>;
   error: z$1.ZodNullable<z$1.ZodObject<{
     name: z$1.ZodOptional<z$1.ZodString>;
     message: z$1.ZodString;
@@ -1564,6 +1623,12 @@ declare const apiCallsConfigSchema: z$1.ZodObject<{
 }, z$1.core.$strip>;
 /** Authored API calls config accepted from `agent-evals.config.ts`. */
 type ApiCallsConfigInput = z$1.infer<typeof apiCallsConfigSchema>;
+/** Schema for workspace-level run log capture options. */
+declare const runLogsConfigSchema: z$1.ZodObject<{
+  captureConsole: z$1.ZodOptional<z$1.ZodBoolean>;
+}, z$1.core.$strip>;
+/** Workspace-level run log capture options. */
+type RunLogsConfigInput = z$1.infer<typeof runLogsConfigSchema>;
 /** Resolved LLM-calls config sent to the UI with all defaults applied. */
 type ResolvedLlmCallsConfig = {
   kinds: string[];
@@ -1737,6 +1802,16 @@ type AgentEvalsConfig = {
    * ```
    */
   apiCalls?: ApiCallsConfigInput;
+  /**
+   * Configuration for case run logs.
+   *
+   * Console capture is enabled by default and stores `console.log`,
+   * `console.info`, `console.warn`, and `console.error` calls made during
+   * active case-owned phases. Set `captureConsole: false` to keep console
+   * output visible in the terminal without persisting it to case details.
+   * Manual `evalLog(...)` calls are still persisted.
+   */
+  runLogs?: RunLogsConfigInput;
   /**
    * Optional controls for the operation cache. When omitted, the cache is
    * enabled and stored under `<workspaceRoot>/.agent-evals/cache`.
@@ -1872,6 +1947,9 @@ declare const agentEvalsConfigSchema: z$1.ZodObject<{
       }>>>;
     }, z$1.core.$strip>>>;
   }, z$1.core.$strip>>;
+  runLogs: z$1.ZodOptional<z$1.ZodObject<{
+    captureConsole: z$1.ZodOptional<z$1.ZodBoolean>;
+  }, z$1.core.$strip>>;
   cache: z$1.ZodOptional<z$1.ZodObject<{
     enabled: z$1.ZodOptional<z$1.ZodBoolean>;
     dir: z$1.ZodOptional<z$1.ZodString>;
@@ -2247,6 +2325,110 @@ declare const cacheEntrySchema: z$1.ZodObject<{
 }, z$1.core.$strip>;
 /** Persisted cache file contents. */
 type CacheEntry = z$1.infer<typeof cacheEntrySchema>;
+/** Debug-only raw key metadata stored outside the reusable cache entry. */
+declare const cacheDebugKeyEntrySchema: z$1.ZodObject<{
+  version: z$1.ZodLiteral<1>;
+  key: z$1.ZodString;
+  namespace: z$1.ZodString;
+  operationType: z$1.ZodEnum<{
+    span: "span";
+    value: "value";
+  }>;
+  operationName: z$1.ZodString;
+  storedAt: z$1.ZodString;
+  codeFingerprint: z$1.ZodString;
+  rawKey: z$1.ZodUnknown;
+}, z$1.core.$strip>;
+/** Debug-only raw cache key entry. May contain sensitive prompt/input data. */
+type CacheDebugKeyEntry = z$1.infer<typeof cacheDebugKeyEntrySchema>;
+/** Cache lookup response with optional debug-only raw key data. */
+declare const cacheEntryWithDebugKeySchema: z$1.ZodObject<{
+  version: z$1.ZodLiteral<1>;
+  key: z$1.ZodString;
+  namespace: z$1.ZodString;
+  operationType: z$1.ZodOptional<z$1.ZodEnum<{
+    span: "span";
+    value: "value";
+  }>>;
+  operationName: z$1.ZodOptional<z$1.ZodString>;
+  spanName: z$1.ZodOptional<z$1.ZodString>;
+  spanKind: z$1.ZodOptional<z$1.ZodString>;
+  storedAt: z$1.ZodString;
+  codeFingerprint: z$1.ZodString;
+  recording: z$1.ZodObject<{
+    returnValue: z$1.ZodUnknown;
+    finalAttributes: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
+    finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
+      error: "error";
+      running: "running";
+      ok: "ok";
+      cancelled: "cancelled";
+    }>>;
+    finalError: z$1.ZodOptional<z$1.ZodObject<{
+      name: z$1.ZodOptional<z$1.ZodString>;
+      message: z$1.ZodString;
+      stack: z$1.ZodOptional<z$1.ZodString>;
+      capturedAt: z$1.ZodOptional<z$1.ZodString>;
+    }, z$1.core.$catchall<z$1.ZodUnknown>>>;
+    finalErrors: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
+      name: z$1.ZodOptional<z$1.ZodString>;
+      message: z$1.ZodString;
+      stack: z$1.ZodOptional<z$1.ZodString>;
+      capturedAt: z$1.ZodOptional<z$1.ZodString>;
+    }, z$1.core.$catchall<z$1.ZodUnknown>>>>;
+    finalWarning: z$1.ZodOptional<z$1.ZodObject<{
+      name: z$1.ZodOptional<z$1.ZodString>;
+      message: z$1.ZodString;
+      stack: z$1.ZodOptional<z$1.ZodString>;
+      capturedAt: z$1.ZodOptional<z$1.ZodString>;
+    }, z$1.core.$catchall<z$1.ZodUnknown>>>;
+    finalWarnings: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
+      name: z$1.ZodOptional<z$1.ZodString>;
+      message: z$1.ZodString;
+      stack: z$1.ZodOptional<z$1.ZodString>;
+      capturedAt: z$1.ZodOptional<z$1.ZodString>;
+    }, z$1.core.$catchall<z$1.ZodUnknown>>>>;
+    ops: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
+      kind: z$1.ZodLiteral<"setOutput">;
+      key: z$1.ZodString;
+      value: z$1.ZodUnknown;
+    }, z$1.core.$strip>, z$1.ZodObject<{
+      kind: z$1.ZodLiteral<"appendOutput">;
+      key: z$1.ZodString;
+      value: z$1.ZodUnknown;
+    }, z$1.core.$strip>, z$1.ZodObject<{
+      kind: z$1.ZodLiteral<"mergeOutput">;
+      key: z$1.ZodString;
+      patch: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
+    }, z$1.core.$strip>, z$1.ZodObject<{
+      kind: z$1.ZodLiteral<"incrementOutput">;
+      key: z$1.ZodString;
+      delta: z$1.ZodNumber;
+    }, z$1.core.$strip>, z$1.ZodObject<{
+      kind: z$1.ZodLiteral<"checkpoint">;
+      name: z$1.ZodString;
+      data: z$1.ZodUnknown;
+    }, z$1.core.$strip>, z$1.ZodObject<{
+      kind: z$1.ZodLiteral<"subSpan">;
+      span: z$1.ZodType<SerializedCacheSpan, unknown, z$1.core.$ZodTypeInternals<SerializedCacheSpan, unknown>>;
+    }, z$1.core.$strip>], "kind">>;
+  }, z$1.core.$strip>;
+  debugKey: z$1.ZodOptional<z$1.ZodObject<{
+    version: z$1.ZodLiteral<1>;
+    key: z$1.ZodString;
+    namespace: z$1.ZodString;
+    operationType: z$1.ZodEnum<{
+      span: "span";
+      value: "value";
+    }>;
+    operationName: z$1.ZodString;
+    storedAt: z$1.ZodString;
+    codeFingerprint: z$1.ZodString;
+    rawKey: z$1.ZodUnknown;
+  }, z$1.core.$strip>>;
+}, z$1.core.$strip>;
+/** Cache lookup response returned by cache APIs when raw-key debug data exists. */
+type CacheEntryWithDebugKey = z$1.infer<typeof cacheEntryWithDebugKeySchema>;
 /** Persisted per-owner cache file containing multiple cache entries. */
 declare const cacheFileSchema: z$1.ZodObject<{
   version: z$1.ZodLiteral<1>;
@@ -2326,6 +2508,26 @@ declare const cacheFileSchema: z$1.ZodObject<{
 }, z$1.core.$strip>;
 /** Persisted per-owner cache file contents. */
 type CacheFile = z$1.infer<typeof cacheFileSchema>;
+/** Persisted per-owner debug file containing raw cache key metadata. */
+declare const cacheDebugKeyFileSchema: z$1.ZodObject<{
+  version: z$1.ZodLiteral<1>;
+  owner: z$1.ZodString;
+  entries: z$1.ZodRecord<z$1.ZodString, z$1.ZodObject<{
+    version: z$1.ZodLiteral<1>;
+    key: z$1.ZodString;
+    namespace: z$1.ZodString;
+    operationType: z$1.ZodEnum<{
+      span: "span";
+      value: "value";
+    }>;
+    operationName: z$1.ZodString;
+    storedAt: z$1.ZodString;
+    codeFingerprint: z$1.ZodString;
+    rawKey: z$1.ZodUnknown;
+  }, z$1.core.$strip>>;
+}, z$1.core.$strip>;
+/** Persisted per-owner raw cache key debug file contents. */
+type CacheDebugKeyFile = z$1.infer<typeof cacheDebugKeyFileSchema>;
 //#endregion
 //#region ../shared/src/utils/extractCacheHits.d.ts
 /**
@@ -2673,6 +2875,19 @@ declare function defineEval<TInput = unknown, TOutputs extends EvalOutputs = Eva
 declare function repoFile(path: string, mimeType?: string): RepoFileRef;
 //#endregion
 //#region ../sdk/src/runtime.d.ts
+/**
+ * Raw-key debug payload passed alongside cache writes.
+ *
+ * `rawKey` may include prompt text, user input, or other sensitive material.
+ * Runners store it outside the reusable cache so projects can gitignore the
+ * debug folder while keeping hash-only cache entries shareable.
+ */
+type CacheDebugKeyWrite = {
+  rawKey: unknown;
+  operationType: CacheOperationType;
+  operationName: string;
+  codeFingerprint: string;
+};
 /**
  * Adapter used by the SDK to read and write cache entries.
  *
@@ -2680,8 +2895,15 @@ declare function repoFile(path: string, mimeType?: string): RepoFileRef;
  * starts executing.
  */
 type CacheAdapter = {
-  /** Return the stored entry for `keyHash` under `namespace`, or `null`. */lookup(namespace: string, keyHash: string): Promise<CacheEntry | null>; /** Persist a cache entry. Must be safe under concurrent calls. */
-  write(entry: CacheEntry): Promise<void>;
+  /** Return the stored entry for `keyHash` under `namespace`, or `null`. */lookup(namespace: string, keyHash: string): Promise<CacheEntry | null>;
+  /**
+   * Persist a cache entry. Must be safe under concurrent calls.
+   *
+   * `debugKey` is optional and contains the authored raw key value for
+   * debugging. It may contain sensitive prompt/input data and should be stored
+   * separately from reusable cache files.
+   */
+  write(entry: CacheEntry, debugKey?: CacheDebugKeyWrite): Promise<void>;
 };
 /** Runner-supplied cache context attached to an eval case scope. */
 type CacheScopeContext = {
@@ -2703,7 +2925,8 @@ type EvalCaseScope = {
   nextEvalIdCounter: number; /** Authored input for the current case, when provided by the runner. */
   input?: unknown;
   outputs: Record<string, unknown>; /** Structured assertion failures recorded for the current case. */
-  assertionFailures: AssertionFailure[];
+  assertionFailures: AssertionFailure[]; /** Logs captured from manual `evalLog(...)` calls and enabled console calls. */
+  logs: RunLogEntry[];
   spans: EvalTraceSpan[];
   checkpoints: Map<string, unknown>;
   spanStack: string[];
@@ -2735,6 +2958,7 @@ type EvalCaseScope = {
  * modules imported while a run is being prepared.
  */
 type EvalRuntimeScope = 'env' | 'cases' | 'eval' | 'derive' | 'outputsSchema' | 'scorer';
+type EvalLogLevelInput = RunLogLevel | 'warning';
 /** Error thrown when an eval assertion fails during case execution. */
 declare class EvalAssertionError extends Error {
   constructor(message: string);
@@ -2750,6 +2974,14 @@ declare function getCurrentScope(): EvalCaseScope | undefined;
  * while validating outputs, and `scorer` while computing scores.
  */
 declare function isInEvalScope(): EvalRuntimeScope | null;
+/**
+ * Record a manual log entry on the active eval case.
+ *
+ * Values are formatted with Node-style console formatting and capped before
+ * persistence so a single log cannot make run artifacts unbounded. Calls made
+ * outside active case-owned eval phases are ignored.
+ */
+declare function evalLog(level: EvalLogLevelInput, ...args: unknown[]): void;
 /**
  * Register background work that should settle before eval finalization.
  *
@@ -2845,6 +3077,352 @@ declare function incrementEvalOutput(key: string, delta: number): void;
  */
 declare function evalAssert(condition: boolean, message: string): void;
 //#endregion
+//#region ../../node_modules/.pnpm/seroval@1.5.2/node_modules/seroval/dist/types/core/constants.d.ts
+declare const enum SerovalConstant {
+  Null = 0,
+  Undefined = 1,
+  True = 2,
+  False = 3,
+  NegZero = 4,
+  Inf = 5,
+  NegInf = 6,
+  Nan = 7
+}
+declare const enum SerovalNodeType {
+  Number = 0,
+  String = 1,
+  Constant = 2,
+  BigInt = 3,
+  IndexedValue = 4,
+  Date = 5,
+  RegExp = 6,
+  Set = 7,
+  Map = 8,
+  Array = 9,
+  Object = 10,
+  NullConstructor = 11,
+  Promise = 12,
+  Error = 13,
+  AggregateError = 14,
+  TypedArray = 15,
+  BigIntTypedArray = 16,
+  WKSymbol = 17,
+  Reference = 18,
+  ArrayBuffer = 19,
+  DataView = 20,
+  Boxed = 21,
+  PromiseConstructor = 22,
+  PromiseSuccess = 23,
+  PromiseFailure = 24,
+  Plugin = 25,
+  SpecialReference = 26,
+  IteratorFactory = 27,
+  IteratorFactoryInstance = 28,
+  AsyncIteratorFactory = 29,
+  AsyncIteratorFactoryInstance = 30,
+  StreamConstructor = 31,
+  StreamNext = 32,
+  StreamThrow = 33,
+  StreamReturn = 34,
+  Sequence = 35
+}
+declare const enum SerovalObjectFlags {
+  None = 0,
+  NonExtensible = 1,
+  Sealed = 2,
+  Frozen = 3
+}
+declare const enum Symbols {
+  AsyncIterator = 0,
+  HasInstance = 1,
+  IsConcatSpreadable = 2,
+  Iterator = 3,
+  Match = 4,
+  MatchAll = 5,
+  Replace = 6,
+  Search = 7,
+  Species = 8,
+  Split = 9,
+  ToPrimitive = 10,
+  ToStringTag = 11,
+  Unscopables = 12
+}
+declare const enum ErrorConstructorTag {
+  Error = 0,
+  EvalError = 1,
+  RangeError = 2,
+  ReferenceError = 3,
+  SyntaxError = 4,
+  TypeError = 5,
+  URIError = 6
+}
+//#endregion
+//#region ../../node_modules/.pnpm/seroval@1.5.2/node_modules/seroval/dist/types/core/special-reference.d.ts
+declare const enum SpecialReference {
+  MapSentinel = 0,
+  PromiseConstructor = 1,
+  PromiseSuccess = 2,
+  PromiseFailure = 3,
+  StreamConstructor = 4,
+  ArrayBufferConstructor = 5
+}
+//#endregion
+//#region ../../node_modules/.pnpm/seroval@1.5.2/node_modules/seroval/dist/types/core/types.d.ts
+interface SerovalBaseNode {
+  t: SerovalNodeType;
+  i: number | undefined;
+  s: unknown;
+  c: string | undefined;
+  m: string | undefined;
+  p: SerovalObjectRecordNode | undefined;
+  e: SerovalMapRecordNode | undefined;
+  a: (SerovalNode | 0)[] | undefined;
+  f: SerovalNode | undefined;
+  b: number | undefined;
+  o: SerovalObjectFlags | undefined;
+  l: number | undefined;
+}
+type SerovalObjectRecordKey = string | SerovalNode;
+interface SerovalObjectRecordNode {
+  k: SerovalObjectRecordKey[];
+  v: SerovalNode[];
+}
+interface SerovalMapRecordNode {
+  k: SerovalNode[];
+  v: SerovalNode[];
+}
+interface SerovalNumberNode extends SerovalBaseNode {
+  t: SerovalNodeType.Number;
+  s: number;
+}
+interface SerovalStringNode extends SerovalBaseNode {
+  t: SerovalNodeType.String;
+  s: string;
+}
+interface SerovalConstantNode extends SerovalBaseNode {
+  t: SerovalNodeType.Constant;
+  s: SerovalConstant;
+}
+type SerovalPrimitiveNode = SerovalNumberNode | SerovalStringNode | SerovalConstantNode;
+interface SerovalIndexedValueNode extends SerovalBaseNode {
+  t: SerovalNodeType.IndexedValue;
+  i: number;
+}
+interface SerovalBigIntNode extends SerovalBaseNode {
+  t: SerovalNodeType.BigInt;
+  s: string;
+}
+interface SerovalDateNode extends SerovalBaseNode {
+  t: SerovalNodeType.Date;
+  i: number;
+  s: string;
+}
+interface SerovalRegExpNode extends SerovalBaseNode {
+  t: SerovalNodeType.RegExp;
+  i: number;
+  c: string;
+  m: string;
+}
+interface SerovalArrayBufferNode extends SerovalBaseNode {
+  t: SerovalNodeType.ArrayBuffer;
+  i: number;
+  s: string;
+  f: SerovalNodeWithID;
+}
+interface SerovalTypedArrayNode extends SerovalBaseNode {
+  t: SerovalNodeType.TypedArray;
+  i: number;
+  c: string;
+  f: SerovalNode;
+  b: number;
+  l: number;
+}
+interface SerovalBigIntTypedArrayNode extends SerovalBaseNode {
+  t: SerovalNodeType.BigIntTypedArray;
+  i: number;
+  c: string;
+  f: SerovalNode;
+  b: number;
+  l: number;
+}
+type SerovalSemiPrimitiveNode = SerovalBigIntNode | SerovalDateNode | SerovalRegExpNode | SerovalTypedArrayNode | SerovalBigIntTypedArrayNode;
+interface SerovalSetNode extends SerovalBaseNode {
+  t: SerovalNodeType.Set;
+  i: number;
+  a: SerovalNode[];
+}
+interface SerovalMapNode extends SerovalBaseNode {
+  t: SerovalNodeType.Map;
+  i: number;
+  e: SerovalMapRecordNode;
+  f: SerovalNodeWithID;
+}
+interface SerovalArrayNode extends SerovalBaseNode {
+  t: SerovalNodeType.Array;
+  a: (SerovalNode | 0)[];
+  i: number;
+  o: SerovalObjectFlags;
+}
+interface SerovalObjectNode extends SerovalBaseNode {
+  t: SerovalNodeType.Object;
+  p: SerovalObjectRecordNode;
+  i: number;
+  o: SerovalObjectFlags;
+}
+interface SerovalNullConstructorNode extends SerovalBaseNode {
+  t: SerovalNodeType.NullConstructor;
+  p: SerovalObjectRecordNode;
+  i: number;
+  o: SerovalObjectFlags;
+}
+interface SerovalPromiseNode extends SerovalBaseNode {
+  t: SerovalNodeType.Promise;
+  s: 0 | 1;
+  f: SerovalNode;
+  i: number;
+}
+interface SerovalErrorNode extends SerovalBaseNode {
+  t: SerovalNodeType.Error;
+  s: ErrorConstructorTag;
+  m: string;
+  p: SerovalObjectRecordNode | undefined;
+  i: number;
+}
+interface SerovalAggregateErrorNode extends SerovalBaseNode {
+  t: SerovalNodeType.AggregateError;
+  i: number;
+  m: string;
+  p: SerovalObjectRecordNode | undefined;
+}
+interface SerovalWKSymbolNode extends SerovalBaseNode {
+  t: SerovalNodeType.WKSymbol;
+  i: number;
+  s: Symbols;
+}
+interface SerovalReferenceNode extends SerovalBaseNode {
+  t: SerovalNodeType.Reference;
+  i: number;
+  s: string;
+}
+interface SerovalDataViewNode extends SerovalBaseNode {
+  t: SerovalNodeType.DataView;
+  i: number;
+  f: SerovalNode;
+  b: number;
+  l: number;
+}
+interface SerovalBoxedNode extends SerovalBaseNode {
+  t: SerovalNodeType.Boxed;
+  i: number;
+  f: SerovalNode;
+}
+interface SerovalPromiseConstructorNode extends SerovalBaseNode {
+  t: SerovalNodeType.PromiseConstructor;
+  i: number;
+  s: number;
+  f: SerovalNodeWithID;
+}
+interface SerovalPromiseResolveNode extends SerovalBaseNode {
+  t: SerovalNodeType.PromiseSuccess;
+  i: number;
+  a: [resolver: SerovalNodeWithID, resolved: SerovalNode];
+}
+interface SerovalPromiseRejectNode extends SerovalBaseNode {
+  t: SerovalNodeType.PromiseFailure;
+  i: number;
+  a: [resolver: SerovalNodeWithID, resolved: SerovalNode];
+}
+interface SerovalPluginNode extends SerovalBaseNode {
+  t: SerovalNodeType.Plugin;
+  i: number;
+  s: Record<string, SerovalNode>;
+  c: string;
+}
+/**
+ * Represents special values as placeholders
+ */
+interface SerovalSpecialReferenceNode extends SerovalBaseNode {
+  t: SerovalNodeType.SpecialReference;
+  i: number;
+  s: SpecialReference;
+}
+interface SerovalIteratorFactoryNode extends SerovalBaseNode {
+  t: SerovalNodeType.IteratorFactory;
+  i: number;
+  f: SerovalNodeWithID;
+}
+interface SerovalIteratorFactoryInstanceNode extends SerovalBaseNode {
+  t: SerovalNodeType.IteratorFactoryInstance;
+  a: [instance: SerovalNodeWithID, sequence: SerovalNodeWithID];
+}
+interface SerovalAsyncIteratorFactoryNode extends SerovalBaseNode {
+  t: SerovalNodeType.AsyncIteratorFactory;
+  i: number;
+  a: [promise: SerovalNodeWithID, symbol: SerovalNodeWithID];
+}
+interface SerovalAsyncIteratorFactoryInstanceNode extends SerovalBaseNode {
+  t: SerovalNodeType.AsyncIteratorFactoryInstance;
+  a: [instance: SerovalNodeWithID, sequence: SerovalNodeWithID];
+}
+interface SerovalStreamConstructorNode extends SerovalBaseNode {
+  t: SerovalNodeType.StreamConstructor;
+  i: number;
+  a: SerovalNode[];
+  f: SerovalNodeWithID;
+}
+interface SerovalStreamNextNode extends SerovalBaseNode {
+  t: SerovalNodeType.StreamNext;
+  i: number;
+  f: SerovalNode;
+}
+interface SerovalStreamThrowNode extends SerovalBaseNode {
+  t: SerovalNodeType.StreamThrow;
+  i: number;
+  f: SerovalNode;
+}
+interface SerovalStreamReturnNode extends SerovalBaseNode {
+  t: SerovalNodeType.StreamReturn;
+  i: number;
+  f: SerovalNode;
+}
+interface SerovalSequenceNode extends SerovalBaseNode {
+  t: SerovalNodeType.Sequence;
+  i: number;
+  s: number;
+  a: SerovalNode[];
+  l: number;
+}
+type SerovalSyncNode = SerovalPrimitiveNode | SerovalIndexedValueNode | SerovalSemiPrimitiveNode | SerovalSetNode | SerovalMapNode | SerovalArrayNode | SerovalObjectNode | SerovalNullConstructorNode | SerovalPromiseNode | SerovalErrorNode | SerovalAggregateErrorNode | SerovalWKSymbolNode | SerovalReferenceNode | SerovalArrayBufferNode | SerovalDataViewNode | SerovalBoxedNode | SerovalPluginNode | SerovalSpecialReferenceNode | SerovalIteratorFactoryNode | SerovalIteratorFactoryInstanceNode | SerovalAsyncIteratorFactoryNode | SerovalAsyncIteratorFactoryInstanceNode | SerovalSequenceNode;
+type SerovalAsyncNode = SerovalPromiseNode | SerovalPromiseConstructorNode | SerovalPromiseResolveNode | SerovalPromiseRejectNode | SerovalStreamConstructorNode | SerovalStreamNextNode | SerovalStreamThrowNode | SerovalStreamReturnNode;
+type SerovalNode = SerovalSyncNode | SerovalAsyncNode;
+type SerovalNodeWithID = Extract<SerovalNode, {
+  i: number;
+}>;
+//#endregion
+//#region ../../node_modules/.pnpm/seroval@1.5.2/node_modules/seroval/dist/types/core/tree/index.d.ts
+interface SerovalJSON {
+  t: SerovalNode;
+  f: number;
+  m: number[];
+}
+//#endregion
+//#region ../sdk/src/cacheSerialization.d.ts
+declare const serializedCacheValueMarker = "__agentEvalsCacheSerialization";
+declare const serializedCacheValueVersion = "seroval-web-v1";
+/** Seroval-backed persisted representation for one cached value. */
+type SerializedCacheValue = {
+  [serializedCacheValueMarker]: typeof serializedCacheValueVersion;
+  value: SerovalJSON;
+};
+/** Serialize one cached value with Seroval plus the Web API plugin set. */
+declare function serializeCacheValue(value: unknown): Promise<unknown>;
+/** Revive one cached value, while preserving legacy JSON-round-tripped data. */
+declare function deserializeCacheValue(value: unknown): unknown;
+/** Serialize all rich values captured in a cache recording before persistence. */
+declare function serializeCacheRecording(recording: CacheRecording): Promise<CacheRecording>;
+/** Revive all rich values captured in a cache recording after lookup. */
+declare function deserializeCacheRecording(recording: CacheRecording): CacheRecording;
+//#endregion
 //#region ../sdk/src/traceDiagnostics.d.ts
 /** Severity used when attaching a recoverable diagnostic to an active span. */
 type CaptureEvalSpanErrorLevel = 'error' | 'warning';
@@ -3110,11 +3688,11 @@ type EvalRunner = {
   listCache(): Promise<CacheListItem[]>;
   /**
    * Return the full persisted cache entry for `namespace` + `key`, including
-   * its recording. Returns `null` when no entry matches. Used by the case
-   * drawer's Cache hits tab to lazily fetch the cached return value when a
-   * row is expanded.
+   * its recording and optional raw-key debug metadata. Returns `null` when no
+   * entry matches. Used by the case drawer's Cache hits tab to lazily fetch
+   * the cached return value when a row is expanded.
    */
-  getCacheEntry(namespace: string, key: string): Promise<CacheEntry | null>;
+  getCacheEntry(namespace: string, key: string): Promise<CacheEntryWithDebugKey | null>;
   /**
    * Remove cache entries matching `filter`, or all entries when no filter is
    * supplied.
@@ -3169,4 +3747,4 @@ declare function createRunner({
  */
 declare function runCli(argv: string[]): Promise<void>;
 //#endregion
-export { type AgentEvalsConfig, type ApiCallEntry, type ApiCallMetric, type ApiCallMetricFormat, type ApiCallMetricPlacement, type ApiCallMetricValue, type ApiCallsConfigInput, type AssertionFailure, type CacheAdapter, type CacheEntry, type CacheFile, type CacheHitEntry, type CacheKeyHashInput, type CacheKeyHashOptions, type CacheListItem, type CacheMode, type CacheOperationType, type CacheRecording, type CacheRecordingFrame, type CacheRecordingOp, type CacheScopeContext, type CacheStatus, type CaptureEvalSpanErrorLevel, type CaptureEvalSpanErrorOptions, type CaseDetail, type CaseRow, type CellValue, type ColumnDef, type ColumnFormat, type ColumnKind, type CreateRunRequest, DEFAULT_API_CALLS_CONFIG, DEFAULT_LLM_CALLS_CONFIG, type DerivedStatus, EvalAssertionError, type EvalCase, type EvalCaseScope, type EvalChartAggregate, type EvalChartAxis, type EvalChartBuiltinMetric, type EvalChartColor, type EvalChartConfig, type EvalChartMetric, type EvalChartTooltipExtra, type EvalChartType, type EvalChartsConfig, type EvalColumnOverride, type EvalColumns, type EvalDefinition, type EvalDeriveContext, type EvalDisplayStatus, type EvalExecuteContext, type EvalFreshnessStatus, type EvalManualScoreDef, type EvalOutputs, type EvalOutputsSchema, type EvalRunner, type EvalRuntimeScope, type EvalScoreContext, type EvalScoreDef, type EvalScoreFn, type EvalSetOutput, type EvalStatAggregate, type EvalStatItem, type EvalStatsConfig, type EvalSummary, type EvalTraceSpan, type EvalTraceSpanError, type EvalTraceSpanWarning, type EvalTraceTree, type FileRef, type JsonCell, type LlmCallEntry, type LlmCallMetric, type LlmCallMetricFormat, type LlmCallMetricPlacement, type LlmCallMetricValue, type LlmCallsConfigInput, type NumberDisplayOptions, type RepoFileRef, type ResolvedApiCallMetric, type ResolvedApiCallsConfig, type ResolvedLlmCallMetric, type ResolvedLlmCallsConfig, type RunArtifactRef, type RunInEvalScopeOptions, type RunManifest, type RunSummary, type ScalarCell, type ScopedCaseSummary, type ScoreTrace, type SerializedCacheSpan, type SpanCacheOptions, type SseEnvelope, type SseEventType, type TraceActiveSpan, type TraceAttributeDisplay, type TraceAttributeDisplayFormat, type TraceAttributeDisplayInput, type TraceAttributeDisplayPlacement, type TraceAttributeTransform, type TraceAttributeTransformContext, type TraceCacheInfo, type TraceCacheRef, type TraceDisplayConfig, type TraceDisplayInputConfig, type TraceSpanInfo, type TrialSelectionMode, type UpdateManualScoreRequest, agentEvalsConfigSchema, apiCallMetricFormatSchema, apiCallMetricPlacementSchema, apiCallMetricSchema, apiCallsConfigSchema, appendToEvalOutput, assertionFailureSchema, buildTraceTree, cacheEntrySchema, cacheFileSchema, cacheListItemSchema, cacheModeSchema, cacheOperationTypeSchema, cacheRecordingOpSchema, cacheRecordingSchema, cacheStatusSchema, captureEvalSpanError, caseDetailSchema, caseRowSchema, cellValueSchema, columnDefSchema, columnFormatSchema, columnKindSchema, createRunRequestSchema, createRunner, defineEval, deriveScopedSummaryFromCases, deriveStatusFromCaseRows, deriveStatusFromChildStatuses, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalFreshnessStatusSchema, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, extractApiCalls, extractCacheHits, extractLlmCalls, fileRefSchema, getCurrentScope, getEvalCaseInput, getEvalDisplayStatus, getEvalRegistry, getEvalTitle, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, jsonCellSchema, llmCallMetricFormatSchema, llmCallMetricPlacementSchema, llmCallMetricSchema, llmCallsConfigSchema, mergeEvalOutput, nextEvalId, numberDisplayOptionsSchema, repoFile, repoFileRefSchema, resolveApiCallsConfig, resolveLlmCallsConfig, runArtifactRefSchema, runCli, runInEvalRuntimeScope, runInEvalScope, runInExistingEvalScope, runManifestSchema, runSummarySchema, scoreTraceSchema, serializedCacheSpanSchema, setEvalOutput, setScopeCacheContext, spanCacheOptionsSchema, sseEnvelopeSchema, startEvalBackgroundJob, traceAttributeDisplayFormatSchema, traceAttributeDisplayInputSchema, traceAttributeDisplayPlacementSchema, traceAttributeDisplaySchema, traceCacheRefSchema, traceDisplayConfigSchema, traceDisplayInputConfigSchema, traceSpanErrorSchema, traceSpanKindSchema, traceSpanSchema, traceSpanWarningSchema, trialSelectionModeSchema, updateManualScoreRequestSchema, z };
+export { type AgentEvalsConfig, type ApiCallEntry, type ApiCallMetric, type ApiCallMetricFormat, type ApiCallMetricPlacement, type ApiCallMetricValue, type ApiCallsConfigInput, type AssertionFailure, type CacheAdapter, type CacheDebugKeyEntry, type CacheDebugKeyFile, type CacheDebugKeyWrite, type CacheEntry, type CacheEntryWithDebugKey, type CacheFile, type CacheHitEntry, type CacheKeyHashInput, type CacheKeyHashOptions, type CacheListItem, type CacheMode, type CacheOperationType, type CacheRecording, type CacheRecordingFrame, type CacheRecordingOp, type CacheScopeContext, type CacheStatus, type CaptureEvalSpanErrorLevel, type CaptureEvalSpanErrorOptions, type CaseDetail, type CaseRow, type CellValue, type ColumnDef, type ColumnFormat, type ColumnKind, type CreateRunRequest, DEFAULT_API_CALLS_CONFIG, DEFAULT_LLM_CALLS_CONFIG, type DerivedStatus, EvalAssertionError, type EvalCase, type EvalCaseScope, type EvalChartAggregate, type EvalChartAxis, type EvalChartBuiltinMetric, type EvalChartColor, type EvalChartConfig, type EvalChartMetric, type EvalChartTooltipExtra, type EvalChartType, type EvalChartsConfig, type EvalColumnOverride, type EvalColumns, type EvalDefinition, type EvalDeriveContext, type EvalDisplayStatus, type EvalExecuteContext, type EvalFreshnessStatus, type EvalManualScoreDef, type EvalOutputs, type EvalOutputsSchema, type EvalRunner, type EvalRuntimeScope, type EvalScoreContext, type EvalScoreDef, type EvalScoreFn, type EvalSetOutput, type EvalStatAggregate, type EvalStatItem, type EvalStatsConfig, type EvalSummary, type EvalTraceSpan, type EvalTraceSpanError, type EvalTraceSpanWarning, type EvalTraceTree, type FileRef, type JsonCell, type LlmCallEntry, type LlmCallMetric, type LlmCallMetricFormat, type LlmCallMetricPlacement, type LlmCallMetricValue, type LlmCallsConfigInput, type NumberDisplayOptions, type RepoFileRef, type ResolvedApiCallMetric, type ResolvedApiCallsConfig, type ResolvedLlmCallMetric, type ResolvedLlmCallsConfig, type RunArtifactRef, type RunInEvalScopeOptions, type RunLogEntry, type RunLogLevel, type RunLogPhase, type RunLogsConfigInput, type RunManifest, type RunSummary, type ScalarCell, type ScopedCaseSummary, type ScoreTrace, type SerializedCacheSpan, type SerializedCacheValue, type SpanCacheOptions, type SseEnvelope, type SseEventType, type TraceActiveSpan, type TraceAttributeDisplay, type TraceAttributeDisplayFormat, type TraceAttributeDisplayInput, type TraceAttributeDisplayPlacement, type TraceAttributeTransform, type TraceAttributeTransformContext, type TraceCacheInfo, type TraceCacheRef, type TraceDisplayConfig, type TraceDisplayInputConfig, type TraceSpanInfo, type TrialSelectionMode, type UpdateManualScoreRequest, agentEvalsConfigSchema, apiCallMetricFormatSchema, apiCallMetricPlacementSchema, apiCallMetricSchema, apiCallsConfigSchema, appendToEvalOutput, assertionFailureSchema, buildTraceTree, cacheDebugKeyEntrySchema, cacheDebugKeyFileSchema, cacheEntrySchema, cacheEntryWithDebugKeySchema, cacheFileSchema, cacheListItemSchema, cacheModeSchema, cacheOperationTypeSchema, cacheRecordingOpSchema, cacheRecordingSchema, cacheStatusSchema, captureEvalSpanError, caseDetailSchema, caseRowSchema, cellValueSchema, columnDefSchema, columnFormatSchema, columnKindSchema, createRunRequestSchema, createRunner, defineEval, deriveScopedSummaryFromCases, deriveStatusFromCaseRows, deriveStatusFromChildStatuses, deserializeCacheRecording, deserializeCacheValue, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalFreshnessStatusSchema, evalLog, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, extractApiCalls, extractCacheHits, extractLlmCalls, fileRefSchema, getCurrentScope, getEvalCaseInput, getEvalDisplayStatus, getEvalRegistry, getEvalTitle, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, jsonCellSchema, llmCallMetricFormatSchema, llmCallMetricPlacementSchema, llmCallMetricSchema, llmCallsConfigSchema, mergeEvalOutput, nextEvalId, numberDisplayOptionsSchema, repoFile, repoFileRefSchema, resolveApiCallsConfig, resolveLlmCallsConfig, runArtifactRefSchema, runCli, runInEvalRuntimeScope, runInEvalScope, runInExistingEvalScope, runLogEntrySchema, runLogLevelSchema, runLogPhaseSchema, runLogsConfigSchema, runManifestSchema, runSummarySchema, scoreTraceSchema, serializeCacheRecording, serializeCacheValue, serializedCacheSpanSchema, setEvalOutput, setScopeCacheContext, spanCacheOptionsSchema, sseEnvelopeSchema, startEvalBackgroundJob, traceAttributeDisplayFormatSchema, traceAttributeDisplayInputSchema, traceAttributeDisplayPlacementSchema, traceAttributeDisplaySchema, traceCacheRefSchema, traceDisplayConfigSchema, traceDisplayInputConfigSchema, traceSpanErrorSchema, traceSpanKindSchema, traceSpanSchema, traceSpanWarningSchema, trialSelectionModeSchema, updateManualScoreRequestSchema, z };