niceeval 0.12.2-canary.31 → 0.13.1

package/dist/runner/rename-experiment.d.cts

@@ -6,7 +6,7 @@
  */
 import { Effect } from "effect";
 import type { SandboxPlanningServices } from "../sandbox/plan.cjs";
-import { type AdoptionProjectV1, ExplicitAdoptionErrorV1, type ExplicitAdoptionOpenErrorV1, type ExplicitAdoptionReadErrorV1, type ExplicitAdoptionRunPlanV1, type RenameAdoptionPreflightV1, prepareCurrentAdoptionTargetV1, resolveRenameSourceRunV1 } from "./adoption.cjs";
+import { type AdoptionProject, ExplicitAdoptionError, type ExplicitAdoptionOpenError, type ExplicitAdoptionReadError, type ExplicitAdoptionRunPlan, type RenameAdoptionPreflight, prepareCurrentAdoptionTarget, resolveRenameSourceRun } from "./adoption.cjs";
 import type { Config, DiscoveredEval, DiscoveredExperiment } from "./types.cjs";
 /** Kept stable only because the current CLI renders these compatibility keys. */
 export type ExperimentRenameReason = "source-empty" | "target-not-found" | "target-has-results" | "source-unreadable" | "artifact-unavailable" | "nothing-to-migrate";
@@ -56,7 +56,7 @@ export interface ExperimentRenamePlan {
     readonly blocked?: ExperimentRenameBlocked;
 }
 /** A receipt-only replacement for the retired EvalResult. It is never stored. */
-export interface RenamedAttemptSourceReceiptV1 {
+export interface RenamedAttemptSourceReceipt {
     readonly experimentId: string;
     readonly locator: string;
     readonly originRunId: string;
@@ -68,7 +68,7 @@ export interface ExperimentRenameDoneEntry {
     readonly locator: string;
     readonly fingerprint: string;
     readonly verdict: "passed" | "failed";
-    readonly renamedFrom: RenamedAttemptSourceReceiptV1;
+    readonly renamedFrom: RenamedAttemptSourceReceipt;
 }
 export interface RenamedExperiment {
     readonly status: "done";
@@ -88,36 +88,34 @@ export interface ExperimentRenameRejected {
     readonly conflictingEvals?: readonly string[];
     readonly detail?: string;
 }
-export interface RenamePreflightV1 {
-    readonly source: RenameAdoptionPreflightV1;
-    readonly plan: ExplicitAdoptionRunPlanV1;
+export interface RenamePreflight {
+    readonly source: RenameAdoptionPreflight;
+    readonly plan: ExplicitAdoptionRunPlan;
 }
 /** All native Record v1 failures before/while the formal publication runs. */
-export type ExperimentRenameNativeErrorV1 = ExplicitAdoptionOpenErrorV1;
+export type ExperimentRenameNativeError = ExplicitAdoptionOpenError;
 /**
- * Native frozen-view preflight. It selects one old Run, rebuilds the complete
+ * Native Host-reader preflight. It selects one old Run, rebuilds the complete
  * current target for newId, and validates every source Member before any
- * generic writer operation may begin.
+ * reference Run write may begin.
  */
-export declare function preflightExperimentRenameV1(input: {
-    readonly view: Parameters<typeof resolveRenameSourceRunV1>[0]["view"];
-    readonly project: AdoptionProjectV1;
+export declare function preflightExperimentRename(input: {
+    readonly reader: Parameters<typeof resolveRenameSourceRun>[0]["reader"];
+    readonly project: AdoptionProject;
     readonly oldId: string;
     readonly newId: string;
     readonly sourceRunId?: string;
-    readonly startedAt: Parameters<typeof prepareCurrentAdoptionTargetV1>[0]["startedAt"];
+    readonly startedAt: Parameters<typeof prepareCurrentAdoptionTarget>[0]["startedAt"];
     readonly operatorReason: string;
-}): Effect.Effect<RenamePreflightV1, ExplicitAdoptionReadErrorV1>;
+}): Effect.Effect<RenamePreflight, ExplicitAdoptionReadError>;
 /**
  * Reader-only CLI compatibility plan. Expected adoption failures are rendered
  * as stable blocked documents; it never opens a writer or creates a Run.
  */
-export declare function planExperimentRename(input: ExperimentRenameOptions): Effect.Effect<ExperimentRenamePlan, import("../record/index.cjs").RecordResourceLimitInvalid | import("../record/index.cjs").RecordRootInvalid | import("../record/index.cjs").RecordPathInvalid | import("../record/index.cjs").RecordPathTypeInvalid | import("../record/index.cjs").RecordPathAlreadyExists | import("../record/index.cjs").RecordResourceLimitExceeded | import("../record/index.cjs").RecordIoError | import("../record/index.cjs").RecordPermissionError | import("../record/index.cjs").RecordMaintenanceBusy | import("../record/index.cjs").RecordRootConstructionError | import("../record/index.cjs").RecordBootstrapInvalid | import("../record/index.cjs").RecordMigrationRequired | import("../record/index.cjs").RecordFormatUnsupported | import("../record/index.cjs").RecordMigrationInterruptedState | import("../record/index.cjs").RecordReaderClosed | import("../record/index.cjs").RecordHandleInvalid | ExplicitAdoptionErrorV1, import("../record/index.cjs").RecordFileSystem | import("../record/index.cjs").RecordMaintenanceLock>;
+export declare function planExperimentRename(input: ExperimentRenameOptions): Effect.Effect<ExperimentRenamePlan, import("../record/platform/errors.cjs").RecordResourceLimitInvalid | import("../record/platform/errors.cjs").RecordRootInvalid | import("../record/platform/errors.cjs").RecordPathInvalid | import("../record/platform/errors.cjs").RecordPathTypeInvalid | import("../record/platform/errors.cjs").RecordPathAlreadyExists | import("../record/platform/errors.cjs").RecordResourceLimitExceeded | import("../record/platform/errors.cjs").RecordIoError | import("../record/platform/errors.cjs").RecordPermissionError | import("../record/platform/errors.cjs").RecordMaintenanceBusy | import("../record/index.cjs").RecordRootConstructionError | import("../coordination/record-leases.cjs").RecordCoordinationIdentityMismatch | import("../record/reader/errors.cjs").RecordBootstrapInvalid | import("../record/reader/errors.cjs").RecordMigrationRequired | import("../record/reader/errors.cjs").RecordFormatUnsupported | import("../record/reader/errors.cjs").RecordMigrationInterruptedState | import("../record/reader/errors.cjs").RecordReaderClosed | import("../record/reader/errors.cjs").RecordHandleInvalid | ExplicitAdoptionError, import("../coordination/record-leases.cjs").RecordCoordination | import("../record/platform/services.cjs").RecordFileSystem>;
 /**
- * Effect-native formal rename. The frozen reader preflight occurs before the
- * writer lock. The one write-session frozen view then repeats every decision
- * before its sole target Run is created and published.
+ * Effect-native formal rename. A Host reader preflight occurs first. One
+ * later Host read Scope then repeats every decision and publishes the
+ * reference Run through createReferenceRun.
  */
-export declare function renameExperimentV1(input: ExperimentRenameOptions): Effect.Effect<RenamedExperiment, import("../record/index.cjs").RecordRootConstructionError | import("../record/index.cjs").RecordReaderClosed | import("../record/index.cjs").RecordHandleInvalid | import("../record/index.cjs").OpenRecordWriteSessionError | ExplicitAdoptionErrorV1 | ExperimentRenameError, import("../record/index.cjs").RecordFileSystem | import("../record/index.cjs").RecordMaintenanceLock | import("../record/index.cjs").RecordWriterLock | import("../record/index.cjs").RecordEntropy>;
-/** Current CLI compatibility adapter; native callers should use renameExperimentV1. */
-export declare function renameExperiment(input: ExperimentRenameOptions): Effect.Effect<RenamedExperiment, import("../record/index.cjs").RecordRootConstructionError | import("../record/index.cjs").RecordReaderClosed | import("../record/index.cjs").RecordHandleInvalid | import("../record/index.cjs").OpenRecordWriteSessionError | ExplicitAdoptionErrorV1 | ExperimentRenameError, import("../record/index.cjs").RecordFileSystem | import("../record/index.cjs").RecordMaintenanceLock | import("../record/index.cjs").RecordWriterLock | import("../record/index.cjs").RecordEntropy>;
+export declare function renameExperiment(input: ExperimentRenameOptions): Effect.Effect<RenamedExperiment, import("../record/index.cjs").RecordRootConstructionError | import("../record/attachment/errors.cjs").RecordAttachmentClosureInvalid | import("../record/errors/record-errors.cjs").RecordCoreInvalid | import("../record/errors/record-errors.cjs").RecordReferenceInvalid | import("../record/reader/errors.cjs").RecordReaderClosed | import("../record/reader/errors.cjs").RecordHandleInvalid | import("../record/reader/errors.cjs").RecordReaderOpenError | import("../record/writer/errors.cjs").RecordWriterClosed | import("../record/writer/errors.cjs").RecordWriteSessionInvalid | import("../record/writer/errors.cjs").RecordRunAlreadyCreated | import("../record/writer/errors.cjs").RecordDraftHandleInvalid | import("../record/writer/errors.cjs").RecordDraftStateError | import("../record/writer/errors.cjs").RecordAttachmentEncodeError | ExplicitAdoptionError | ExperimentRenameError, import("../coordination/record-leases.cjs").RecordCoordination | import("../record/platform/services.cjs").RecordFileSystem | import("../record/platform/services.cjs").RecordEntropy>;

package/dist/runner/rename-experiment.d.mts

@@ -6,7 +6,7 @@
  */
 import { Effect } from "effect";
 import type { SandboxPlanningServices } from "../sandbox/plan.mjs";
-import { type AdoptionProjectV1, ExplicitAdoptionErrorV1, type ExplicitAdoptionOpenErrorV1, type ExplicitAdoptionReadErrorV1, type ExplicitAdoptionRunPlanV1, type RenameAdoptionPreflightV1, prepareCurrentAdoptionTargetV1, resolveRenameSourceRunV1 } from "./adoption.mjs";
+import { type AdoptionProject, ExplicitAdoptionError, type ExplicitAdoptionOpenError, type ExplicitAdoptionReadError, type ExplicitAdoptionRunPlan, type RenameAdoptionPreflight, prepareCurrentAdoptionTarget, resolveRenameSourceRun } from "./adoption.mjs";
 import type { Config, DiscoveredEval, DiscoveredExperiment } from "./types.mjs";
 /** Kept stable only because the current CLI renders these compatibility keys. */
 export type ExperimentRenameReason = "source-empty" | "target-not-found" | "target-has-results" | "source-unreadable" | "artifact-unavailable" | "nothing-to-migrate";
@@ -56,7 +56,7 @@ export interface ExperimentRenamePlan {
     readonly blocked?: ExperimentRenameBlocked;
 }
 /** A receipt-only replacement for the retired EvalResult. It is never stored. */
-export interface RenamedAttemptSourceReceiptV1 {
+export interface RenamedAttemptSourceReceipt {
     readonly experimentId: string;
     readonly locator: string;
     readonly originRunId: string;
@@ -68,7 +68,7 @@ export interface ExperimentRenameDoneEntry {
     readonly locator: string;
     readonly fingerprint: string;
     readonly verdict: "passed" | "failed";
-    readonly renamedFrom: RenamedAttemptSourceReceiptV1;
+    readonly renamedFrom: RenamedAttemptSourceReceipt;
 }
 export interface RenamedExperiment {
     readonly status: "done";
@@ -88,36 +88,34 @@ export interface ExperimentRenameRejected {
     readonly conflictingEvals?: readonly string[];
     readonly detail?: string;
 }
-export interface RenamePreflightV1 {
-    readonly source: RenameAdoptionPreflightV1;
-    readonly plan: ExplicitAdoptionRunPlanV1;
+export interface RenamePreflight {
+    readonly source: RenameAdoptionPreflight;
+    readonly plan: ExplicitAdoptionRunPlan;
 }
 /** All native Record v1 failures before/while the formal publication runs. */
-export type ExperimentRenameNativeErrorV1 = ExplicitAdoptionOpenErrorV1;
+export type ExperimentRenameNativeError = ExplicitAdoptionOpenError;
 /**
- * Native frozen-view preflight. It selects one old Run, rebuilds the complete
+ * Native Host-reader preflight. It selects one old Run, rebuilds the complete
  * current target for newId, and validates every source Member before any
- * generic writer operation may begin.
+ * reference Run write may begin.
  */
-export declare function preflightExperimentRenameV1(input: {
-    readonly view: Parameters<typeof resolveRenameSourceRunV1>[0]["view"];
-    readonly project: AdoptionProjectV1;
+export declare function preflightExperimentRename(input: {
+    readonly reader: Parameters<typeof resolveRenameSourceRun>[0]["reader"];
+    readonly project: AdoptionProject;
     readonly oldId: string;
     readonly newId: string;
     readonly sourceRunId?: string;
-    readonly startedAt: Parameters<typeof prepareCurrentAdoptionTargetV1>[0]["startedAt"];
+    readonly startedAt: Parameters<typeof prepareCurrentAdoptionTarget>[0]["startedAt"];
     readonly operatorReason: string;
-}): Effect.Effect<RenamePreflightV1, ExplicitAdoptionReadErrorV1>;
+}): Effect.Effect<RenamePreflight, ExplicitAdoptionReadError>;
 /**
  * Reader-only CLI compatibility plan. Expected adoption failures are rendered
  * as stable blocked documents; it never opens a writer or creates a Run.
  */
-export declare function planExperimentRename(input: ExperimentRenameOptions): Effect.Effect<ExperimentRenamePlan, import("../record/index.mjs").RecordResourceLimitInvalid | import("../record/index.mjs").RecordRootInvalid | import("../record/index.mjs").RecordPathInvalid | import("../record/index.mjs").RecordPathTypeInvalid | import("../record/index.mjs").RecordPathAlreadyExists | import("../record/index.mjs").RecordResourceLimitExceeded | import("../record/index.mjs").RecordIoError | import("../record/index.mjs").RecordPermissionError | import("../record/index.mjs").RecordMaintenanceBusy | import("../record/index.mjs").RecordRootConstructionError | import("../record/index.mjs").RecordBootstrapInvalid | import("../record/index.mjs").RecordMigrationRequired | import("../record/index.mjs").RecordFormatUnsupported | import("../record/index.mjs").RecordMigrationInterruptedState | import("../record/index.mjs").RecordReaderClosed | import("../record/index.mjs").RecordHandleInvalid | ExplicitAdoptionErrorV1, import("../record/index.mjs").RecordFileSystem | import("../record/index.mjs").RecordMaintenanceLock>;
+export declare function planExperimentRename(input: ExperimentRenameOptions): Effect.Effect<ExperimentRenamePlan, import("../record/platform/errors.mjs").RecordResourceLimitInvalid | import("../record/platform/errors.mjs").RecordRootInvalid | import("../record/platform/errors.mjs").RecordPathInvalid | import("../record/platform/errors.mjs").RecordPathTypeInvalid | import("../record/platform/errors.mjs").RecordPathAlreadyExists | import("../record/platform/errors.mjs").RecordResourceLimitExceeded | import("../record/platform/errors.mjs").RecordIoError | import("../record/platform/errors.mjs").RecordPermissionError | import("../record/platform/errors.mjs").RecordMaintenanceBusy | import("../record/index.mjs").RecordRootConstructionError | import("../coordination/record-leases.mjs").RecordCoordinationIdentityMismatch | import("../record/reader/errors.mjs").RecordBootstrapInvalid | import("../record/reader/errors.mjs").RecordMigrationRequired | import("../record/reader/errors.mjs").RecordFormatUnsupported | import("../record/reader/errors.mjs").RecordMigrationInterruptedState | import("../record/reader/errors.mjs").RecordReaderClosed | import("../record/reader/errors.mjs").RecordHandleInvalid | ExplicitAdoptionError, import("../coordination/record-leases.mjs").RecordCoordination | import("../record/platform/services.mjs").RecordFileSystem>;
 /**
- * Effect-native formal rename. The frozen reader preflight occurs before the
- * writer lock. The one write-session frozen view then repeats every decision
- * before its sole target Run is created and published.
+ * Effect-native formal rename. A Host reader preflight occurs first. One
+ * later Host read Scope then repeats every decision and publishes the
+ * reference Run through createReferenceRun.
  */
-export declare function renameExperimentV1(input: ExperimentRenameOptions): Effect.Effect<RenamedExperiment, import("../record/index.mjs").RecordRootConstructionError | import("../record/index.mjs").RecordReaderClosed | import("../record/index.mjs").RecordHandleInvalid | import("../record/index.mjs").OpenRecordWriteSessionError | ExplicitAdoptionErrorV1 | ExperimentRenameError, import("../record/index.mjs").RecordFileSystem | import("../record/index.mjs").RecordMaintenanceLock | import("../record/index.mjs").RecordWriterLock | import("../record/index.mjs").RecordEntropy>;
-/** Current CLI compatibility adapter; native callers should use renameExperimentV1. */
-export declare function renameExperiment(input: ExperimentRenameOptions): Effect.Effect<RenamedExperiment, import("../record/index.mjs").RecordRootConstructionError | import("../record/index.mjs").RecordReaderClosed | import("../record/index.mjs").RecordHandleInvalid | import("../record/index.mjs").OpenRecordWriteSessionError | ExplicitAdoptionErrorV1 | ExperimentRenameError, import("../record/index.mjs").RecordFileSystem | import("../record/index.mjs").RecordMaintenanceLock | import("../record/index.mjs").RecordWriterLock | import("../record/index.mjs").RecordEntropy>;
+export declare function renameExperiment(input: ExperimentRenameOptions): Effect.Effect<RenamedExperiment, import("../record/index.mjs").RecordRootConstructionError | import("../record/attachment/errors.mjs").RecordAttachmentClosureInvalid | import("../record/errors/record-errors.mjs").RecordCoreInvalid | import("../record/errors/record-errors.mjs").RecordReferenceInvalid | import("../record/reader/errors.mjs").RecordReaderClosed | import("../record/reader/errors.mjs").RecordHandleInvalid | import("../record/reader/errors.mjs").RecordReaderOpenError | import("../record/writer/errors.mjs").RecordWriterClosed | import("../record/writer/errors.mjs").RecordWriteSessionInvalid | import("../record/writer/errors.mjs").RecordRunAlreadyCreated | import("../record/writer/errors.mjs").RecordDraftHandleInvalid | import("../record/writer/errors.mjs").RecordDraftStateError | import("../record/writer/errors.mjs").RecordAttachmentEncodeError | ExplicitAdoptionError | ExperimentRenameError, import("../coordination/record-leases.mjs").RecordCoordination | import("../record/platform/services.mjs").RecordFileSystem | import("../record/platform/services.mjs").RecordEntropy>;

package/dist/runner/rename-experiment.d.ts

@@ -6,7 +6,7 @@
  */
 import { Effect } from "effect";
 import type { SandboxPlanningServices } from "../sandbox/plan.js";
-import { type AdoptionProjectV1, ExplicitAdoptionErrorV1, type ExplicitAdoptionOpenErrorV1, type ExplicitAdoptionReadErrorV1, type ExplicitAdoptionRunPlanV1, type RenameAdoptionPreflightV1, prepareCurrentAdoptionTargetV1, resolveRenameSourceRunV1 } from "./adoption.js";
+import { type AdoptionProject, ExplicitAdoptionError, type ExplicitAdoptionOpenError, type ExplicitAdoptionReadError, type ExplicitAdoptionRunPlan, type RenameAdoptionPreflight, prepareCurrentAdoptionTarget, resolveRenameSourceRun } from "./adoption.js";
 import type { Config, DiscoveredEval, DiscoveredExperiment } from "./types.js";
 /** Kept stable only because the current CLI renders these compatibility keys. */
 export type ExperimentRenameReason = "source-empty" | "target-not-found" | "target-has-results" | "source-unreadable" | "artifact-unavailable" | "nothing-to-migrate";
@@ -56,7 +56,7 @@ export interface ExperimentRenamePlan {
     readonly blocked?: ExperimentRenameBlocked;
 }
 /** A receipt-only replacement for the retired EvalResult. It is never stored. */
-export interface RenamedAttemptSourceReceiptV1 {
+export interface RenamedAttemptSourceReceipt {
     readonly experimentId: string;
     readonly locator: string;
     readonly originRunId: string;
@@ -68,7 +68,7 @@ export interface ExperimentRenameDoneEntry {
     readonly locator: string;
     readonly fingerprint: string;
     readonly verdict: "passed" | "failed";
-    readonly renamedFrom: RenamedAttemptSourceReceiptV1;
+    readonly renamedFrom: RenamedAttemptSourceReceipt;
 }
 export interface RenamedExperiment {
     readonly status: "done";
@@ -88,36 +88,34 @@ export interface ExperimentRenameRejected {
     readonly conflictingEvals?: readonly string[];
     readonly detail?: string;
 }
-export interface RenamePreflightV1 {
-    readonly source: RenameAdoptionPreflightV1;
-    readonly plan: ExplicitAdoptionRunPlanV1;
+export interface RenamePreflight {
+    readonly source: RenameAdoptionPreflight;
+    readonly plan: ExplicitAdoptionRunPlan;
 }
 /** All native Record v1 failures before/while the formal publication runs. */
-export type ExperimentRenameNativeErrorV1 = ExplicitAdoptionOpenErrorV1;
+export type ExperimentRenameNativeError = ExplicitAdoptionOpenError;
 /**
- * Native frozen-view preflight. It selects one old Run, rebuilds the complete
+ * Native Host-reader preflight. It selects one old Run, rebuilds the complete
  * current target for newId, and validates every source Member before any
- * generic writer operation may begin.
+ * reference Run write may begin.
  */
-export declare function preflightExperimentRenameV1(input: {
-    readonly view: Parameters<typeof resolveRenameSourceRunV1>[0]["view"];
-    readonly project: AdoptionProjectV1;
+export declare function preflightExperimentRename(input: {
+    readonly reader: Parameters<typeof resolveRenameSourceRun>[0]["reader"];
+    readonly project: AdoptionProject;
     readonly oldId: string;
     readonly newId: string;
     readonly sourceRunId?: string;
-    readonly startedAt: Parameters<typeof prepareCurrentAdoptionTargetV1>[0]["startedAt"];
+    readonly startedAt: Parameters<typeof prepareCurrentAdoptionTarget>[0]["startedAt"];
     readonly operatorReason: string;
-}): Effect.Effect<RenamePreflightV1, ExplicitAdoptionReadErrorV1>;
+}): Effect.Effect<RenamePreflight, ExplicitAdoptionReadError>;
 /**
  * Reader-only CLI compatibility plan. Expected adoption failures are rendered
  * as stable blocked documents; it never opens a writer or creates a Run.
  */
-export declare function planExperimentRename(input: ExperimentRenameOptions): Effect.Effect<ExperimentRenamePlan, import("../record/index.js").RecordResourceLimitInvalid | import("../record/index.js").RecordRootInvalid | import("../record/index.js").RecordPathInvalid | import("../record/index.js").RecordPathTypeInvalid | import("../record/index.js").RecordPathAlreadyExists | import("../record/index.js").RecordResourceLimitExceeded | import("../record/index.js").RecordIoError | import("../record/index.js").RecordPermissionError | import("../record/index.js").RecordMaintenanceBusy | import("../record/index.js").RecordRootConstructionError | import("../record/index.js").RecordBootstrapInvalid | import("../record/index.js").RecordMigrationRequired | import("../record/index.js").RecordFormatUnsupported | import("../record/index.js").RecordMigrationInterruptedState | import("../record/index.js").RecordReaderClosed | import("../record/index.js").RecordHandleInvalid | ExplicitAdoptionErrorV1, import("../record/index.js").RecordFileSystem | import("../record/index.js").RecordMaintenanceLock>;
+export declare function planExperimentRename(input: ExperimentRenameOptions): Effect.Effect<ExperimentRenamePlan, import("../record/platform/errors.js").RecordResourceLimitInvalid | import("../record/platform/errors.js").RecordRootInvalid | import("../record/platform/errors.js").RecordPathInvalid | import("../record/platform/errors.js").RecordPathTypeInvalid | import("../record/platform/errors.js").RecordPathAlreadyExists | import("../record/platform/errors.js").RecordResourceLimitExceeded | import("../record/platform/errors.js").RecordIoError | import("../record/platform/errors.js").RecordPermissionError | import("../record/platform/errors.js").RecordMaintenanceBusy | import("../record/index.js").RecordRootConstructionError | import("../coordination/record-leases.js").RecordCoordinationIdentityMismatch | import("../record/reader/errors.js").RecordBootstrapInvalid | import("../record/reader/errors.js").RecordMigrationRequired | import("../record/reader/errors.js").RecordFormatUnsupported | import("../record/reader/errors.js").RecordMigrationInterruptedState | import("../record/reader/errors.js").RecordReaderClosed | import("../record/reader/errors.js").RecordHandleInvalid | ExplicitAdoptionError, import("../coordination/record-leases.js").RecordCoordination | import("../record/platform/services.js").RecordFileSystem>;
 /**
- * Effect-native formal rename. The frozen reader preflight occurs before the
- * writer lock. The one write-session frozen view then repeats every decision
- * before its sole target Run is created and published.
+ * Effect-native formal rename. A Host reader preflight occurs first. One
+ * later Host read Scope then repeats every decision and publishes the
+ * reference Run through createReferenceRun.
  */
-export declare function renameExperimentV1(input: ExperimentRenameOptions): Effect.Effect<RenamedExperiment, import("../record/index.js").RecordRootConstructionError | import("../record/index.js").RecordReaderClosed | import("../record/index.js").RecordHandleInvalid | import("../record/index.js").OpenRecordWriteSessionError | ExplicitAdoptionErrorV1 | ExperimentRenameError, import("../record/index.js").RecordFileSystem | import("../record/index.js").RecordMaintenanceLock | import("../record/index.js").RecordWriterLock | import("../record/index.js").RecordEntropy>;
-/** Current CLI compatibility adapter; native callers should use renameExperimentV1. */
-export declare function renameExperiment(input: ExperimentRenameOptions): Effect.Effect<RenamedExperiment, import("../record/index.js").RecordRootConstructionError | import("../record/index.js").RecordReaderClosed | import("../record/index.js").RecordHandleInvalid | import("../record/index.js").OpenRecordWriteSessionError | ExplicitAdoptionErrorV1 | ExperimentRenameError, import("../record/index.js").RecordFileSystem | import("../record/index.js").RecordMaintenanceLock | import("../record/index.js").RecordWriterLock | import("../record/index.js").RecordEntropy>;
+export declare function renameExperiment(input: ExperimentRenameOptions): Effect.Effect<RenamedExperiment, import("../record/index.js").RecordRootConstructionError | import("../record/attachment/errors.js").RecordAttachmentClosureInvalid | import("../record/errors/record-errors.js").RecordCoreInvalid | import("../record/errors/record-errors.js").RecordReferenceInvalid | import("../record/reader/errors.js").RecordReaderClosed | import("../record/reader/errors.js").RecordHandleInvalid | import("../record/reader/errors.js").RecordReaderOpenError | import("../record/writer/errors.js").RecordWriterClosed | import("../record/writer/errors.js").RecordWriteSessionInvalid | import("../record/writer/errors.js").RecordRunAlreadyCreated | import("../record/writer/errors.js").RecordDraftHandleInvalid | import("../record/writer/errors.js").RecordDraftStateError | import("../record/writer/errors.js").RecordAttachmentEncodeError | ExplicitAdoptionError | ExperimentRenameError, import("../coordination/record-leases.js").RecordCoordination | import("../record/platform/services.js").RecordFileSystem | import("../record/platform/services.js").RecordEntropy>;

package/dist/runner/rename-experiment.js

@@ -5,10 +5,8 @@ const __niceevalExport0 = __niceevalCanonical["ExperimentRenameError"];
 export { __niceevalExport0 as ExperimentRenameError };
 const __niceevalExport1 = __niceevalCanonical["planExperimentRename"];
 export { __niceevalExport1 as planExperimentRename };
-const __niceevalExport2 = __niceevalCanonical["preflightExperimentRenameV1"];
-export { __niceevalExport2 as preflightExperimentRenameV1 };
+const __niceevalExport2 = __niceevalCanonical["preflightExperimentRename"];
+export { __niceevalExport2 as preflightExperimentRename };
 const __niceevalExport3 = __niceevalCanonical["renameExperiment"];
 export { __niceevalExport3 as renameExperiment };
-const __niceevalExport4 = __niceevalCanonical["renameExperimentV1"];
-export { __niceevalExport4 as renameExperimentV1 };
 //# sourceMappingURL=rename-experiment.js.map

package/dist/runner/report.cjs

@@ -96,11 +96,15 @@ function summarize(results, reusedAttempts, startedAt, durationMs, name) {
     let inTok = 0;
     let outTok = 0;
     let cost = 0;
+    let hasEstimatedCost = false;
     for (const r of results) {
         counts[r.verdict] += 1;
         inTok += r.usage?.inputTokens ?? 0;
         outTok += r.usage?.outputTokens ?? 0;
-        cost += r.estimatedCostUSD ?? 0;
+        if (r.estimatedCostUSD !== undefined) {
+            cost += r.estimatedCostUSD;
+            hasEstimatedCost = true;
+        }
     }
     for (const attempt of reusedAttempts) {
         counts[attempt.verdict] += 1;
@@ -115,7 +119,8 @@ function summarize(results, reusedAttempts, startedAt, durationMs, name) {
         errored: counts.errored,
         durationMs,
         usage: { inputTokens: inTok, outputTokens: outTok },
-        estimatedCostUSD: cost || undefined,
+        // 0 是合法估算值:至少一条 fresh 结果带 estimatedCostUSD 就输出累计值(可为 0),全缺才省略。
+        estimatedCostUSD: hasEstimatedCost ? cost : undefined,
         reusedAttempts: Object.freeze([...reusedAttempts]),
         results,
     };

package/dist/runner/report.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"report.js","sourceRoot":"","sources":["report.ts"],"names":[],"mappings":";;AAmBA,kCAYC;AACD,8CAIC;AAED,sCAMC;AAMD,sCAmCC;AAED,8BA4BC;AAnHD,uDAAuD;AACvD,uBAAuB;AACvB,mCAAgC;AAEhC,wCAAqD;AACrD,gDAAyD;AACzD;;;;;;;;;;;;GAYG;AACH,SAAgB,WAAW,CAAC,GAAyB,EAAE,KAAa,EAAE,EAAiB;IACnF,OAAO,eAAM,CAAC,UAAU,CAAC;QACrB,0EAA0E;QAC1E,GAAG,EAAE,GAAG,EAAE,CAAC,OAAO,CAAC,OAAO,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC;QACrC,KAAK,EAAE,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK;KAC1B,CAAC,CAAC,IAAI,CAAC,eAAM,CAAC,QAAQ,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,eAAM,CAAC,IAAI,CAAC,GAAG,EAAE;QAChD,IAAA,6BAAmB,EAAC;YAChB,QAAQ,EAAE,GAAG,CAAC,IAAI;YAClB,QAAQ,EAAE,GAAG,CAAC,QAAQ;YACtB,OAAO,EAAE,GAAG,KAAK,KAAK,IAAA,mBAAS,EAAC,IAAA,sBAAY,EAAC,KAAK,CAAC,CAAC,EAAE;SACzD,CAAC,CAAC;IACP,CAAC,CAAC,CAAC,CAAC,CAAC;AACT,CAAC;AACD,SAAgB,iBAAiB,CAAC,aAA8C,EAAE,KAAoB;IAClG,iFAAiF;IACjF,wEAAwE;IACxE,OAAO,eAAM,CAAC,OAAO,CAAC,aAAa,EAAE,CAAC,GAAG,EAAE,EAAE,CAAC,WAAW,CAAC,GAAG,EAAE,SAAS,KAAK,CAAC,IAAI,EAAE,EAAE,GAAG,EAAE,CAAC,GAAG,CAAC,QAAQ,CAAC,OAAO,EAAE,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,WAAW,EAAE,WAAW,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC;AAC7K,CAAC;AACD,yEAAyE;AACzE,SAAgB,aAAa,CAAC,OAA0B,EAAE,GAAwB;IAC9E,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IAC7D,MAAM,cAAc,GAAG,OAAO,CAAC,cAAc,CAAC,MAAM,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,OAAO,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC;IAClG,MAAM,GAAG,GAAG,SAAS,CAAC,OAAO,EAAE,cAAc,EAAE,OAAO,CAAC,SAAS,EAAE,OAAO,CAAC,UAAU,EAAE,OAAO,CAAC,IAAI,CAAC,CAAC;IACpG,sDAAsD;IACtD,OAAO,EAAE,GAAG,OAAO,EAAE,GAAG,GAAG,EAAE,WAAW,EAAE,OAAO,CAAC,WAAW,EAAE,CAAC;AACpE,CAAC;AACD;;;;GAIG;AACH,SAAgB,aAAa,CAAC,CAAW,EAAE,GAAwB,EAAE,KAAuB;IACxF,MAAM,MAAM,GAAa,EAAE,CAAC;IAC5B,IAAI,CAAC,CAAC,iBAAiB,EAAE,CAAC;QACtB,MAAM,CAAC,iBAAiB,GAAG,CAAC,KAAK,EAAE,SAAS,EAAE,EAAE,CAAC,CAAC,CAAC,iBAAkB,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,KAAK,IAAI,SAAS,CAAC,CAAC;IAClI,CAAC;IACD,IAAI,CAAC,CAAC,cAAc,EAAE,CAAC;QACnB,MAAM,CAAC,cAAc,GAAG,CAAC,MAAM,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,cAAe,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC;IACrG,CAAC;IACD,IAAI,CAAC,CAAC,oBAAoB,EAAE,CAAC;QACzB,MAAM,CAAC,oBAAoB,GAAG,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC,CAAC,oBAAqB,CAAC,aAAa,CAAC,OAAO,EAAE,GAAG,CAAC,CAAC,CAAC;IACpG,CAAC;IACD,IAAI,CAAC,CAAC,OAAO,EAAE,CAAC;QACZ,MAAM,CAAC,OAAO,GAAG,CAAC,KAAK,EAAE,EAAE;YACvB,QAAQ,KAAK,CAAC,IAAI,EAAE,CAAC;gBACjB,KAAK,kBAAkB;oBACnB,OAAO,CAAC,CAAC,OAAQ,CAAC;wBACd,GAAG,KAAK;wBACR,KAAK,EAAE,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;wBAC/C,KAAK,EAAE,KAAK,IAAI,KAAK,CAAC,KAAK;qBAC9B,CAAC,CAAC;gBACP,KAAK,YAAY;oBACb,OAAO,GAAG,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,OAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;gBAClE,KAAK,eAAe;oBAChB,OAAO,GAAG,CAAC,GAAG,CAAC,KAAK,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,OAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;gBACpE,KAAK,sBAAsB;oBACvB,OAAO,GAAG,CAAC,GAAG,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,OAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;gBACjE,KAAK,oBAAoB,CAAC;gBAC1B,KAAK,kBAAkB;oBACnB,OAAO,CAAC,CAAC,OAAQ,CAAC,EAAE,GAAG,KAAK,EAAE,OAAO,EAAE,aAAa,CAAC,KAAK,CAAC,OAAO,EAAE,GAAG,CAAC,EAAE,CAAC,CAAC;gBAChF;oBACI,OAAO,CAAC,CAAC,OAAQ,CAAC,KAAK,CAAC,CAAC;YACjC,CAAC;QACL,CAAC,CAAC;IACN,CAAC;IACD,OAAO,MAAM,CAAC;AAClB,CAAC;AACD,kJAAkJ;AAClJ,SAAgB,SAAS,CAAC,OAAqB,EAAE,cAAqF,EAAE,SAAiB,EAAE,UAAkB,EAAE,IAAoB;IAC/L,MAAM,MAAM,GAAG,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,OAAO,EAAE,CAAC,EAAE,OAAO,EAAE,CAAC,EAAE,CAAC;IAChE,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,IAAI,MAAM,GAAG,CAAC,CAAC;IACf,IAAI,IAAI,GAAG,CAAC,CAAC;IACb,KAAK,MAAM,CAAC,IAAI,OAAO,EAAE,CAAC;QACtB,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;QACvB,KAAK,IAAI,CAAC,CAAC,KAAK,EAAE,WAAW,IAAI,CAAC,CAAC;QACnC,MAAM,IAAI,CAAC,CAAC,KAAK,EAAE,YAAY,IAAI,CAAC,CAAC;QACrC,IAAI,IAAI,CAAC,CAAC,gBAAgB,IAAI,CAAC,CAAC;IACpC,CAAC;IACD,KAAK,MAAM,OAAO,IAAI,cAAc,EAAE,CAAC;QACnC,MAAM,CAAC,OAAO,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;IACjC,CAAC;IACD,OAAO;QACH,IAAI;QACJ,SAAS;QACT,WAAW,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;QACrC,MAAM,EAAE,MAAM,CAAC,MAAM;QACrB,MAAM,EAAE,MAAM,CAAC,MAAM;QACrB,OAAO,EAAE,MAAM,CAAC,OAAO;QACvB,OAAO,EAAE,MAAM,CAAC,OAAO;QACvB,UAAU;QACV,KAAK,EAAE,EAAE,WAAW,EAAE,KAAK,EAAE,YAAY,EAAE,MAAM,EAAE;QACnD,gBAAgB,EAAE,IAAI,IAAI,SAAS;QACnC,cAAc,EAAE,MAAM,CAAC,MAAM,CAAC,CAAC,GAAG,cAAc,CAAC,CAAC;QAClD,OAAO;KACV,CAAC;AACN,CAAC","sourcesContent":["// reporter 编排与运行级汇总。reporter 是「结果消费方」:单个 reporter 抛错只记\n// diagnostic,不能让整次调度崩。\nimport { Effect } from \"effect\";\nimport type { EvalResult, InvocationShape, InvocationSummary, LocalizedText, Reporter, ReporterEvent, ReporterRegistration, } from \"../types.ts\";\nimport { firstLine, formatThrown } from \"../util.ts\";\nimport { reportReporterError } from \"./feedback/sink.ts\";\n/**\n * reporter 调用的统一兜错。返回 void,永不 reject(供 Promise.all 安全聚合,单个 reporter\n * 抛错不终止其它 reporter 的收尾,也不中断在飞的 attempt)。`reg.name`/`reg.required` 决定\n * `reportReporterError()` 收到的诊断身份与判定权重(见 `ReporterRegistration` 的字段注释:\n * 同一个 reporter 反复失败按 `name` 去重折叠,`required` 决定它是否让 completion/CI 退出码\n * 判红)。`stage` 只是这次失败发生在哪个回调阶段的次要上下文,拼进 message,不参与去重身份——\n * 同一 reporter 在 onEvalComplete 与 onInvocationComplete 两个阶段各失败一次,仍然折成同一条诊断、\n * count 累加到 2,而不是拆成两条互不相关的诊断。message 只取 `formatThrown()` 的第一行(与\n * run.ts 的 `describeFailureReason` 同一原则)——reporter 抛出的 Error 完整 `.stack` 含本地\n * 绝对文件路径和调用帧,不适合塞进 agent/ci 那种单行、稳定、机器消费的 envelope;这里没有\n * `EvalResult.error` 那样的落盘字段能保留完整栈,所以「一层可行动摘要」就是这条诊断的全部内容,\n * 不是从更详细的记录里摘出来的简化版——排查真正需要栈时,重跑该 reporter 单独复现更可靠。\n */\nexport function runReporter(reg: ReporterRegistration, stage: string, fn: () => unknown): Effect.Effect<void> {\n    return Effect.tryPromise({\n        // Public Reporter callbacks are the one Promise boundary for this module.\n        try: () => Promise.resolve().then(fn),\n        catch: (cause) => cause,\n    }).pipe(Effect.catchAll((cause) => Effect.sync(() => {\n        reportReporterError({\n            reporter: reg.name,\n            required: reg.required,\n            message: `${stage}: ${firstLine(formatThrown(cause))}`,\n        });\n    })));\n}\nexport function emitReporterEvent(registrations: readonly ReporterRegistration[], event: ReporterEvent): Effect.Effect<void> {\n    // Existing Promise.all starts every reporter immediately and waits for each one;\n    // retain that concurrent, best-effort observable behavior under Effect.\n    return Effect.forEach(registrations, (reg) => runReporter(reg, `event:${event.type}`, () => reg.reporter.onEvent?.(event)), { concurrency: \"unbounded\", discard: true });\n}\n/** 按 eval id 过滤 InvocationSummary 并重新计数 —— eval 级 reporter 只看它观测的那部分。 */\nexport function filterSummary(summary: InvocationSummary, ids: ReadonlySet<string>): InvocationSummary {\n    const results = summary.results.filter((r) => ids.has(r.id));\n    const reusedAttempts = summary.reusedAttempts.filter((attempt) => ids.has(attempt.target.evalId));\n    const sub = summarize(results, reusedAttempts, summary.startedAt, summary.durationMs, summary.name);\n    // completedAt 用原值(summarize 会重新取 now);name 等其余字段原样保留。\n    return { ...summary, ...sub, completedAt: summary.completedAt };\n}\n/**\n * eval 级 reporter 的作用域包装:实例只观测「引用它的那些 eval」。\n * 结果类回调按 eval id 过滤,汇总类回调收到重新计数的子集汇总;\n * shape 由调用方按作用域预先算好(包装器自己看不到 attempts)。\n */\nexport function scopeReporter(r: Reporter, ids: ReadonlySet<string>, shape?: InvocationShape): Reporter {\n    const scoped: Reporter = {};\n    if (r.onInvocationStart) {\n        scoped.onInvocationStart = (evals, fullShape) => r.onInvocationStart!(evals.filter((e) => ids.has(e.id)), shape ?? fullShape);\n    }\n    if (r.onEvalComplete) {\n        scoped.onEvalComplete = (result) => (ids.has(result.id) ? r.onEvalComplete!(result) : undefined);\n    }\n    if (r.onInvocationComplete) {\n        scoped.onInvocationComplete = (summary) => r.onInvocationComplete!(filterSummary(summary, ids));\n    }\n    if (r.onEvent) {\n        scoped.onEvent = (event) => {\n            switch (event.type) {\n                case \"invocation:start\":\n                    return r.onEvent!({\n                        ...event,\n                        evals: event.evals.filter((e) => ids.has(e.id)),\n                        shape: shape ?? event.shape,\n                    });\n                case \"eval:start\":\n                    return ids.has(event.eval.id) ? r.onEvent!(event) : undefined;\n                case \"eval:complete\":\n                    return ids.has(event.result.id) ? r.onEvent!(event) : undefined;\n                case \"invocation:earlyExit\":\n                    return ids.has(event.evalId) ? r.onEvent!(event) : undefined;\n                case \"invocation:summary\":\n                case \"invocation:saved\":\n                    return r.onEvent!({ ...event, summary: filterSummary(event.summary, ids) });\n                default:\n                    return r.onEvent!(event);\n            }\n        };\n    }\n    return scoped;\n}\n/** 全局汇总:verdict 计数 + token / cost 折叠。按 attempt 计(eval 级折叠见 shared/verdict.ts)。不接收 agent/model —— 一次 Invocation 可能横跨多个配置,身份从逐条 `EvalResult` 读取。 */\nexport function summarize(results: EvalResult[], reusedAttempts: readonly import(\"./reuse-readback.ts\").CurrentReusedAttemptReadback[], startedAt: string, durationMs: number, name?: LocalizedText): InvocationSummary {\n    const counts = { passed: 0, failed: 0, skipped: 0, errored: 0 };\n    let inTok = 0;\n    let outTok = 0;\n    let cost = 0;\n    for (const r of results) {\n        counts[r.verdict] += 1;\n        inTok += r.usage?.inputTokens ?? 0;\n        outTok += r.usage?.outputTokens ?? 0;\n        cost += r.estimatedCostUSD ?? 0;\n    }\n    for (const attempt of reusedAttempts) {\n        counts[attempt.verdict] += 1;\n    }\n    return {\n        name,\n        startedAt,\n        completedAt: new Date().toISOString(),\n        passed: counts.passed,\n        failed: counts.failed,\n        skipped: counts.skipped,\n        errored: counts.errored,\n        durationMs,\n        usage: { inputTokens: inTok, outputTokens: outTok },\n        estimatedCostUSD: cost || undefined,\n        reusedAttempts: Object.freeze([...reusedAttempts]),\n        results,\n    };\n}\n"]}
+{"version":3,"file":"report.js","sourceRoot":"","sources":["report.ts"],"names":[],"mappings":";;AAmBA,kCAYC;AACD,8CAIC;AAED,sCAMC;AAMD,sCAmCC;AAED,8BAiCC;AAxHD,uDAAuD;AACvD,uBAAuB;AACvB,mCAAgC;AAEhC,wCAAqD;AACrD,gDAAyD;AACzD;;;;;;;;;;;;GAYG;AACH,SAAgB,WAAW,CAAC,GAAyB,EAAE,KAAa,EAAE,EAAiB;IACnF,OAAO,eAAM,CAAC,UAAU,CAAC;QACrB,0EAA0E;QAC1E,GAAG,EAAE,GAAG,EAAE,CAAC,OAAO,CAAC,OAAO,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC;QACrC,KAAK,EAAE,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK;KAC1B,CAAC,CAAC,IAAI,CAAC,eAAM,CAAC,QAAQ,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,eAAM,CAAC,IAAI,CAAC,GAAG,EAAE;QAChD,IAAA,6BAAmB,EAAC;YAChB,QAAQ,EAAE,GAAG,CAAC,IAAI;YAClB,QAAQ,EAAE,GAAG,CAAC,QAAQ;YACtB,OAAO,EAAE,GAAG,KAAK,KAAK,IAAA,mBAAS,EAAC,IAAA,sBAAY,EAAC,KAAK,CAAC,CAAC,EAAE;SACzD,CAAC,CAAC;IACP,CAAC,CAAC,CAAC,CAAC,CAAC;AACT,CAAC;AACD,SAAgB,iBAAiB,CAAC,aAA8C,EAAE,KAAoB;IAClG,iFAAiF;IACjF,wEAAwE;IACxE,OAAO,eAAM,CAAC,OAAO,CAAC,aAAa,EAAE,CAAC,GAAG,EAAE,EAAE,CAAC,WAAW,CAAC,GAAG,EAAE,SAAS,KAAK,CAAC,IAAI,EAAE,EAAE,GAAG,EAAE,CAAC,GAAG,CAAC,QAAQ,CAAC,OAAO,EAAE,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,WAAW,EAAE,WAAW,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC,CAAC;AAC7K,CAAC;AACD,yEAAyE;AACzE,SAAgB,aAAa,CAAC,OAA0B,EAAE,GAAwB;IAC9E,MAAM,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IAC7D,MAAM,cAAc,GAAG,OAAO,CAAC,cAAc,CAAC,MAAM,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,OAAO,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC;IAClG,MAAM,GAAG,GAAG,SAAS,CAAC,OAAO,EAAE,cAAc,EAAE,OAAO,CAAC,SAAS,EAAE,OAAO,CAAC,UAAU,EAAE,OAAO,CAAC,IAAI,CAAC,CAAC;IACpG,sDAAsD;IACtD,OAAO,EAAE,GAAG,OAAO,EAAE,GAAG,GAAG,EAAE,WAAW,EAAE,OAAO,CAAC,WAAW,EAAE,CAAC;AACpE,CAAC;AACD;;;;GAIG;AACH,SAAgB,aAAa,CAAC,CAAW,EAAE,GAAwB,EAAE,KAAuB;IACxF,MAAM,MAAM,GAAa,EAAE,CAAC;IAC5B,IAAI,CAAC,CAAC,iBAAiB,EAAE,CAAC;QACtB,MAAM,CAAC,iBAAiB,GAAG,CAAC,KAAK,EAAE,SAAS,EAAE,EAAE,CAAC,CAAC,CAAC,iBAAkB,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,KAAK,IAAI,SAAS,CAAC,CAAC;IAClI,CAAC;IACD,IAAI,CAAC,CAAC,cAAc,EAAE,CAAC;QACnB,MAAM,CAAC,cAAc,GAAG,CAAC,MAAM,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,cAAe,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC;IACrG,CAAC;IACD,IAAI,CAAC,CAAC,oBAAoB,EAAE,CAAC;QACzB,MAAM,CAAC,oBAAoB,GAAG,CAAC,OAAO,EAAE,EAAE,CAAC,CAAC,CAAC,oBAAqB,CAAC,aAAa,CAAC,OAAO,EAAE,GAAG,CAAC,CAAC,CAAC;IACpG,CAAC;IACD,IAAI,CAAC,CAAC,OAAO,EAAE,CAAC;QACZ,MAAM,CAAC,OAAO,GAAG,CAAC,KAAK,EAAE,EAAE;YACvB,QAAQ,KAAK,CAAC,IAAI,EAAE,CAAC;gBACjB,KAAK,kBAAkB;oBACnB,OAAO,CAAC,CAAC,OAAQ,CAAC;wBACd,GAAG,KAAK;wBACR,KAAK,EAAE,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;wBAC/C,KAAK,EAAE,KAAK,IAAI,KAAK,CAAC,KAAK;qBAC9B,CAAC,CAAC;gBACP,KAAK,YAAY;oBACb,OAAO,GAAG,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,OAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;gBAClE,KAAK,eAAe;oBAChB,OAAO,GAAG,CAAC,GAAG,CAAC,KAAK,CAAC,MAAM,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,OAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;gBACpE,KAAK,sBAAsB;oBACvB,OAAO,GAAG,CAAC,GAAG,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,OAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;gBACjE,KAAK,oBAAoB,CAAC;gBAC1B,KAAK,kBAAkB;oBACnB,OAAO,CAAC,CAAC,OAAQ,CAAC,EAAE,GAAG,KAAK,EAAE,OAAO,EAAE,aAAa,CAAC,KAAK,CAAC,OAAO,EAAE,GAAG,CAAC,EAAE,CAAC,CAAC;gBAChF;oBACI,OAAO,CAAC,CAAC,OAAQ,CAAC,KAAK,CAAC,CAAC;YACjC,CAAC;QACL,CAAC,CAAC;IACN,CAAC;IACD,OAAO,MAAM,CAAC;AAClB,CAAC;AACD,kJAAkJ;AAClJ,SAAgB,SAAS,CAAC,OAAqB,EAAE,cAAqF,EAAE,SAAiB,EAAE,UAAkB,EAAE,IAAoB;IAC/L,MAAM,MAAM,GAAG,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,OAAO,EAAE,CAAC,EAAE,OAAO,EAAE,CAAC,EAAE,CAAC;IAChE,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,IAAI,MAAM,GAAG,CAAC,CAAC;IACf,IAAI,IAAI,GAAG,CAAC,CAAC;IACb,IAAI,gBAAgB,GAAG,KAAK,CAAC;IAC7B,KAAK,MAAM,CAAC,IAAI,OAAO,EAAE,CAAC;QACtB,MAAM,CAAC,CAAC,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;QACvB,KAAK,IAAI,CAAC,CAAC,KAAK,EAAE,WAAW,IAAI,CAAC,CAAC;QACnC,MAAM,IAAI,CAAC,CAAC,KAAK,EAAE,YAAY,IAAI,CAAC,CAAC;QACrC,IAAI,CAAC,CAAC,gBAAgB,KAAK,SAAS,EAAE,CAAC;YACnC,IAAI,IAAI,CAAC,CAAC,gBAAgB,CAAC;YAC3B,gBAAgB,GAAG,IAAI,CAAC;QAC5B,CAAC;IACL,CAAC;IACD,KAAK,MAAM,OAAO,IAAI,cAAc,EAAE,CAAC;QACnC,MAAM,CAAC,OAAO,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;IACjC,CAAC;IACD,OAAO;QACH,IAAI;QACJ,SAAS;QACT,WAAW,EAAE,IAAI,IAAI,EAAE,CAAC,WAAW,EAAE;QACrC,MAAM,EAAE,MAAM,CAAC,MAAM;QACrB,MAAM,EAAE,MAAM,CAAC,MAAM;QACrB,OAAO,EAAE,MAAM,CAAC,OAAO;QACvB,OAAO,EAAE,MAAM,CAAC,OAAO;QACvB,UAAU;QACV,KAAK,EAAE,EAAE,WAAW,EAAE,KAAK,EAAE,YAAY,EAAE,MAAM,EAAE;QACnD,+DAA+D;QAC/D,gBAAgB,EAAE,gBAAgB,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,SAAS;QACrD,cAAc,EAAE,MAAM,CAAC,MAAM,CAAC,CAAC,GAAG,cAAc,CAAC,CAAC;QAClD,OAAO;KACV,CAAC;AACN,CAAC","sourcesContent":["// reporter 编排与运行级汇总。reporter 是「结果消费方」:单个 reporter 抛错只记\n// diagnostic,不能让整次调度崩。\nimport { Effect } from \"effect\";\nimport type { EvalResult, InvocationShape, InvocationSummary, LocalizedText, Reporter, ReporterEvent, ReporterRegistration, } from \"../types.ts\";\nimport { firstLine, formatThrown } from \"../util.ts\";\nimport { reportReporterError } from \"./feedback/sink.ts\";\n/**\n * reporter 调用的统一兜错。返回 void,永不 reject(供 Promise.all 安全聚合,单个 reporter\n * 抛错不终止其它 reporter 的收尾,也不中断在飞的 attempt)。`reg.name`/`reg.required` 决定\n * `reportReporterError()` 收到的诊断身份与判定权重(见 `ReporterRegistration` 的字段注释:\n * 同一个 reporter 反复失败按 `name` 去重折叠,`required` 决定它是否让 completion/CI 退出码\n * 判红)。`stage` 只是这次失败发生在哪个回调阶段的次要上下文,拼进 message,不参与去重身份——\n * 同一 reporter 在 onEvalComplete 与 onInvocationComplete 两个阶段各失败一次,仍然折成同一条诊断、\n * count 累加到 2,而不是拆成两条互不相关的诊断。message 只取 `formatThrown()` 的第一行(与\n * run.ts 的 `describeFailureReason` 同一原则)——reporter 抛出的 Error 完整 `.stack` 含本地\n * 绝对文件路径和调用帧,不适合塞进 agent/ci 那种单行、稳定、机器消费的 envelope;这里没有\n * `EvalResult.error` 那样的落盘字段能保留完整栈,所以「一层可行动摘要」就是这条诊断的全部内容,\n * 不是从更详细的记录里摘出来的简化版——排查真正需要栈时,重跑该 reporter 单独复现更可靠。\n */\nexport function runReporter(reg: ReporterRegistration, stage: string, fn: () => unknown): Effect.Effect<void> {\n    return Effect.tryPromise({\n        // Public Reporter callbacks are the one Promise boundary for this module.\n        try: () => Promise.resolve().then(fn),\n        catch: (cause) => cause,\n    }).pipe(Effect.catchAll((cause) => Effect.sync(() => {\n        reportReporterError({\n            reporter: reg.name,\n            required: reg.required,\n            message: `${stage}: ${firstLine(formatThrown(cause))}`,\n        });\n    })));\n}\nexport function emitReporterEvent(registrations: readonly ReporterRegistration[], event: ReporterEvent): Effect.Effect<void> {\n    // Existing Promise.all starts every reporter immediately and waits for each one;\n    // retain that concurrent, best-effort observable behavior under Effect.\n    return Effect.forEach(registrations, (reg) => runReporter(reg, `event:${event.type}`, () => reg.reporter.onEvent?.(event)), { concurrency: \"unbounded\", discard: true });\n}\n/** 按 eval id 过滤 InvocationSummary 并重新计数 —— eval 级 reporter 只看它观测的那部分。 */\nexport function filterSummary(summary: InvocationSummary, ids: ReadonlySet<string>): InvocationSummary {\n    const results = summary.results.filter((r) => ids.has(r.id));\n    const reusedAttempts = summary.reusedAttempts.filter((attempt) => ids.has(attempt.target.evalId));\n    const sub = summarize(results, reusedAttempts, summary.startedAt, summary.durationMs, summary.name);\n    // completedAt 用原值(summarize 会重新取 now);name 等其余字段原样保留。\n    return { ...summary, ...sub, completedAt: summary.completedAt };\n}\n/**\n * eval 级 reporter 的作用域包装:实例只观测「引用它的那些 eval」。\n * 结果类回调按 eval id 过滤,汇总类回调收到重新计数的子集汇总;\n * shape 由调用方按作用域预先算好(包装器自己看不到 attempts)。\n */\nexport function scopeReporter(r: Reporter, ids: ReadonlySet<string>, shape?: InvocationShape): Reporter {\n    const scoped: Reporter = {};\n    if (r.onInvocationStart) {\n        scoped.onInvocationStart = (evals, fullShape) => r.onInvocationStart!(evals.filter((e) => ids.has(e.id)), shape ?? fullShape);\n    }\n    if (r.onEvalComplete) {\n        scoped.onEvalComplete = (result) => (ids.has(result.id) ? r.onEvalComplete!(result) : undefined);\n    }\n    if (r.onInvocationComplete) {\n        scoped.onInvocationComplete = (summary) => r.onInvocationComplete!(filterSummary(summary, ids));\n    }\n    if (r.onEvent) {\n        scoped.onEvent = (event) => {\n            switch (event.type) {\n                case \"invocation:start\":\n                    return r.onEvent!({\n                        ...event,\n                        evals: event.evals.filter((e) => ids.has(e.id)),\n                        shape: shape ?? event.shape,\n                    });\n                case \"eval:start\":\n                    return ids.has(event.eval.id) ? r.onEvent!(event) : undefined;\n                case \"eval:complete\":\n                    return ids.has(event.result.id) ? r.onEvent!(event) : undefined;\n                case \"invocation:earlyExit\":\n                    return ids.has(event.evalId) ? r.onEvent!(event) : undefined;\n                case \"invocation:summary\":\n                case \"invocation:saved\":\n                    return r.onEvent!({ ...event, summary: filterSummary(event.summary, ids) });\n                default:\n                    return r.onEvent!(event);\n            }\n        };\n    }\n    return scoped;\n}\n/** 全局汇总:verdict 计数 + token / cost 折叠。按 attempt 计(eval 级折叠见 shared/verdict.ts)。不接收 agent/model —— 一次 Invocation 可能横跨多个配置,身份从逐条 `EvalResult` 读取。 */\nexport function summarize(results: EvalResult[], reusedAttempts: readonly import(\"./reuse-readback.ts\").CurrentReusedAttemptReadback[], startedAt: string, durationMs: number, name?: LocalizedText): InvocationSummary {\n    const counts = { passed: 0, failed: 0, skipped: 0, errored: 0 };\n    let inTok = 0;\n    let outTok = 0;\n    let cost = 0;\n    let hasEstimatedCost = false;\n    for (const r of results) {\n        counts[r.verdict] += 1;\n        inTok += r.usage?.inputTokens ?? 0;\n        outTok += r.usage?.outputTokens ?? 0;\n        if (r.estimatedCostUSD !== undefined) {\n            cost += r.estimatedCostUSD;\n            hasEstimatedCost = true;\n        }\n    }\n    for (const attempt of reusedAttempts) {\n        counts[attempt.verdict] += 1;\n    }\n    return {\n        name,\n        startedAt,\n        completedAt: new Date().toISOString(),\n        passed: counts.passed,\n        failed: counts.failed,\n        skipped: counts.skipped,\n        errored: counts.errored,\n        durationMs,\n        usage: { inputTokens: inTok, outputTokens: outTok },\n        // 0 是合法估算值:至少一条 fresh 结果带 estimatedCostUSD 就输出累计值(可为 0),全缺才省略。\n        estimatedCostUSD: hasEstimatedCost ? cost : undefined,\n        reusedAttempts: Object.freeze([...reusedAttempts]),\n        results,\n    };\n}\n"]}

package/dist/runner/reporters/braintrust.cjs

@@ -95,6 +95,8 @@ function toBraintrustEvent(result, assessment = NOT_PROJECTED_ASSESSMENT) {
         if (requests !== undefined)
             metrics.requests = requests;
     }
+    // estimated_cost_usd 恒为价目表估算(EvalResult.estimatedCostUSD,见 estimateCost);
+    // observed 成本单独留在 result.usage.costUSD,不混入本指标。
     if (result.estimatedCostUSD !== undefined)
         metrics.estimated_cost_usd = result.estimatedCostUSD;
     const metadata = {
@@ -161,25 +163,11 @@ function projectionMetadata(entry, available) {
                 state: "migration-required",
                 from: entry.attachment.from,
                 to: entry.attachment.to,
-                command: entry.attachment.command,
-            };
-        case "migration-unavailable":
-            return {
-                state: "migration-unavailable",
-                from: entry.attachment.from,
-                to: entry.attachment.to,
-                reason: entry.attachment.reason,
+                ...(entry.attachment.command === undefined ? {} : { command: entry.attachment.command }),
             };
         case "unsupported":
-            return { state: "unsupported", schemaId: entry.attachment.schemaId };
         case "invalid":
-            return {
-                state: "invalid",
-                issues: entry.attachment.issues.map((issue) => ({
-                    code: issue.code,
-                    path: [...issue.path],
-                })),
-            };
+            return { state: entry.attachment.state };
     }
 }
 function assertionMetadata(entries) {

package/dist/runner/reporters/braintrust.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"braintrust.js","sourceRoot":"","sources":["braintrust.ts"],"names":[],"mappings":";;AAqFA,gCA8CC;AAOD,8CA4EC;AA7MD,iDAAqD;AAoErD,MAAM,wBAAwB,GAAyB,MAAM,CAAC,MAAM,CAAC;IACjE,KAAK,EAAE,eAAwB;IAC/B,MAAM,EAAE,uDAAgE;CAC3E,CAAC,CAAC;AACH;;;GAGG;AACH,SAAgB,UAAU,CAAC,SAA2B,EAAE;IACpD,IAAI,GAA8B,CAAC;IACnC,IAAI,UAA4C,CAAC;IACjD,OAAO;QACH,KAAK,CAAC,iBAAiB,CAAC,KAAK;YACzB,GAAG,GAAG,MAAM,iBAAiB,EAAE,CAAC;YAChC,UAAU,GAAG,MAAM,GAAG,CAAC,IAAI,CAAC;gBACxB,OAAO,EAAE,MAAM,CAAC,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,OAAO,IAAI,UAAU,CAAC;gBACtE,SAAS,EAAE,MAAM,CAAC,SAAS;gBAC3B,UAAU,EAAE,MAAM,CAAC,UAAU;gBAC7B,cAAc,EAAE,MAAM,CAAC,cAAc;gBACrC,gBAAgB,EAAE,MAAM,CAAC,gBAAgB;gBACzC,MAAM,EAAE,MAAM,CAAC,MAAM;gBACrB,MAAM,EAAE,MAAM,CAAC,MAAM;gBACrB,+DAA+D;gBAC/D,oCAAoC;gBACpC,oDAAoD;gBACpD,qEAAqE;gBACrE,mDAAmD;gBACnD,QAAQ,EAAE;oBACN,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;oBAC7B,GAAG,MAAM,CAAC,QAAQ;iBACrB;gBACD,sDAAsD;gBACtD,UAAU,EAAE,KAAK;aACpB,CAAC,CAAC;QACP,CAAC;QACD,cAAc,CAAC,MAAM;YACjB,UAAU,EAAE,GAAG,CAAC,iBAAiB,CAAC,MAAM,EAAE,wBAAwB,CAAC,CAAC,CAAC;QACzE,CAAC;QACD,KAAK,CAAC,oBAAoB;YACtB,IAAI,CAAC,UAAU;gBACX,OAAO;YACX,IAAI,CAAC;gBACD,MAAM,GAAG,EAAE,KAAK,EAAE,CAAC;gBACnB,MAAM,OAAO,GAAG,MAAM,UAAU,CAAC,SAAS,EAAE,CAAC;gBAC7C,IAAI,OAAO,CAAC,aAAa,EAAE,CAAC;oBACxB,IAAA,wBAAc,EAAC,0BAA0B,OAAO,CAAC,aAAa,EAAE,CAAC,CAAC;gBACtE,CAAC;YACL,CAAC;oBACO,CAAC;gBACL,UAAU,GAAG,SAAS,CAAC;gBACvB,GAAG,GAAG,SAAS,CAAC;YACpB,CAAC;QACL,CAAC;KACJ,CAAC;AACN,CAAC;AACD;;;;;GAKG;AACH,SAAgB,iBAAiB,CAAC,MAAkB,EAAE,aAAmC,wBAAwB;IAC7G,MAAM,gBAAgB,GAAG,UAAU,CAAC,KAAK,KAAK,WAAW;QACrD,CAAC,CAAC,wBAAwB,CAAC,UAAU,CAAC,OAAO,CAAC;QAC9C,CAAC,CAAC,SAAS,CAAC;IAChB,MAAM,OAAO,GAAG,gBAAgB,IAAI,MAAM,CAAC,OAAO,CAAC;IACnD,MAAM,OAAO,GAAsC,EAAE,CAAC;IACtD,IAAI,MAAM,CAAC,SAAS,EAAE,CAAC;QACnB,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,SAAS,CAAC,GAAG,IAAI,CAAC;QAClD,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,EAAE,CAAC;YACzB,OAAO,CAAC,KAAK,GAAG,KAAK,CAAC;YACtB,OAAO,CAAC,GAAG,GAAG,KAAK,GAAG,MAAM,CAAC,UAAU,GAAG,IAAI,CAAC;QACnD,CAAC;IACL,CAAC;IACD,IAAI,MAAM,CAAC,KAAK,EAAE,CAAC;QACf,MAAM,EAAE,WAAW,EAAE,YAAY,EAAE,eAAe,EAAE,mBAAmB,EAAE,eAAe,EAAE,QAAQ,EAAE,GAAG,MAAM,CAAC,KAAK,CAAC;QACpH,IAAI,WAAW,KAAK,SAAS;YACzB,OAAO,CAAC,aAAa,GAAG,WAAW,CAAC;QACxC,IAAI,YAAY,KAAK,SAAS;YAC1B,OAAO,CAAC,iBAAiB,GAAG,YAAY,CAAC;QAC7C,IAAI,WAAW,KAAK,SAAS,IAAI,YAAY,KAAK,SAAS;YACvD,OAAO,CAAC,MAAM,GAAG,WAAW,GAAG,YAAY,CAAC;QAChD,IAAI,eAAe,KAAK,SAAS;YAC7B,OAAO,CAAC,iBAAiB,GAAG,eAAe,CAAC;QAChD,IAAI,mBAAmB,KAAK,SAAS;YACjC,OAAO,CAAC,qBAAqB,GAAG,mBAAmB,CAAC;QACxD,IAAI,eAAe,KAAK,SAAS;YAC7B,OAAO,CAAC,gBAAgB,GAAG,eAAe,CAAC;QAC/C,IAAI,QAAQ,KAAK,SAAS;YACtB,OAAO,CAAC,QAAQ,GAAG,QAAQ,CAAC;IACpC,CAAC;IACD,IAAI,MAAM,CAAC,gBAAgB,KAAK,SAAS;QACrC,OAAO,CAAC,kBAAkB,GAAG,MAAM,CAAC,gBAAgB,CAAC;IACzD,MAAM,QAAQ,GAAyC;QACnD,IAAI,EAAE,MAAM,CAAC,EAAE;QACf,KAAK,EAAE,MAAM,CAAC,KAAK;QACnB,OAAO,EAAE,MAAM,CAAC,OAAO;QACvB,OAAO;QACP,MAAM,EAAE,OAAO;QACf,mBAAmB,EAAE,MAAM,CAAC,mBAAmB;QAC/C,cAAc,EAAE,MAAM,CAAC,cAAc;QACrC,oBAAoB,EAAE,UAAU,CAAC,KAAK;KACzC,CAAC;IACF,MAAM,MAAM,GAAsC,EAAE,CAAC;IACrD,IAAI,UAAU,CAAC,KAAK,KAAK,WAAW,EAAE,CAAC;QACnC,QAAQ,CAAC,UAAU,GAAG,kBAAkB,CAAC,UAAU,CAAC,UAAU,EAAE,CAAC,UAAU,EAAE,EAAE,CAAC,CAAC,EAAE,OAAO,EAAE,iBAAiB,CAAC,UAAU,CAAC,OAAO,CAAC,EAAE,CAAC,CAAC,CAAC;QACtI,QAAQ,CAAC,iBAAiB,GAAG,kBAAkB,CAAC,UAAU,CAAC,OAAO,EAAE,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC,CAAC;QAC5F,IAAI,UAAU,CAAC,KAAK,KAAK,SAAS,EAAE,CAAC;YACjC,QAAQ,CAAC,KAAK,GAAG,kBAAkB,CAAC,UAAU,CAAC,KAAK,EAAE,aAAa,CAAC,CAAC;YACrE,MAAM,KAAK,GAAG,wBAAwB,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC;YACzD,IAAI,KAAK,EAAE,KAAK,KAAK,UAAU;gBAC3B,MAAM,CAAC,gBAAgB,CAAC,GAAG,KAAK,CAAC,MAAM,CAAC;QAChD,CAAC;IACL,CAAC;SACI,CAAC;QACF,QAAQ,CAAC,0BAA0B,GAAG,UAAU,CAAC,MAAM,CAAC;IAC5D,CAAC;IACD,IAAI,MAAM,CAAC,KAAK,KAAK,SAAS;QAC1B,QAAQ,CAAC,KAAK,GAAG,MAAM,CAAC,KAAK,CAAC;IAClC,IAAI,MAAM,CAAC,YAAY,KAAK,SAAS;QACjC,QAAQ,CAAC,UAAU,GAAG,MAAM,CAAC,YAAY,CAAC;IAC9C,IAAI,MAAM,CAAC,UAAU,EAAE,KAAK,IAAI,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC9E,QAAQ,CAAC,KAAK,GAAG,MAAM,CAAC,UAAU,CAAC,KAAK,CAAC;IAC7C,CAAC;IACD,IAAI,MAAM,CAAC,UAAU,KAAK,SAAS;QAC/B,QAAQ,CAAC,UAAU,GAAG,MAAM,CAAC,UAAU,CAAC;IAC5C,4EAA4E;IAC5E,MAAM,EAAE,GAAG,CAAC,MAAM,CAAC,YAAY,IAAI,EAAE,EAAE,MAAM,CAAC,EAAE,EAAE,MAAM,CAAC,KAAK,EAAE,MAAM,CAAC,KAAK,IAAI,EAAE,EAAE,IAAI,MAAM,CAAC,OAAO,EAAE,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;IACpH,OAAO;QACH,EAAE;QACF,KAAK,EAAE,MAAM,CAAC,WAAW,IAAI,MAAM,CAAC,EAAE;QACtC,MAAM,EAAE,iBAAiB,CAAC,MAAM,CAAC,MAAM,CAAC;QACxC,KAAK,EAAE,MAAM,CAAC,KAAK;QACnB,QAAQ;QACR,OAAO;QACP,GAAG,CAAC,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KACxD,CAAC;AACN,CAAC;AACD,SAAS,wBAAwB,CAAQ,KAAuC;IAC5E,OAAO,KAAK,CAAC,KAAK,KAAK,mBAAmB,IAAI,KAAK,CAAC,UAAU,CAAC,KAAK,KAAK,WAAW;QAChF,CAAC,CAAC,KAAK,CAAC,UAAU,CAAC,KAAK;QACxB,CAAC,CAAC,SAAS,CAAC;AACpB,CAAC;AACD,SAAS,kBAAkB,CAAQ,KAAuC,EAAE,SAAiE;IACzI,IAAI,KAAK,CAAC,KAAK,KAAK,mBAAmB,EAAE,CAAC;QACtC,OAAO,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,EAAE,CAAC;IAClC,CAAC;IACD,QAAQ,KAAK,CAAC,UAAU,CAAC,KAAK,EAAE,CAAC;QAC7B,KAAK,WAAW;YACZ,OAAO,EAAE,KAAK,EAAE,WAAW,EAAE,GAAG,SAAS,CAAC,KAAK,CAAC,UAAU,CAAC,KAAK,CAAC,EAAE,CAAC;QACxE,KAAK,aAAa;YACd,OAAO,EAAE,KAAK,EAAE,aAAa,EAAE,CAAC;QACpC,KAAK,oBAAoB;YACrB,OAAO;gBACH,KAAK,EAAE,oBAAoB;gBAC3B,IAAI,EAAE,KAAK,CAAC,UAAU,CAAC,IAAI;gBAC3B,EAAE,EAAE,KAAK,CAAC,UAAU,CAAC,EAAE;gBACvB,OAAO,EAAE,KAAK,CAAC,UAAU,CAAC,OAAO;aACpC,CAAC;QACN,KAAK,uBAAuB;YACxB,OAAO;gBACH,KAAK,EAAE,uBAAuB;gBAC9B,IAAI,EAAE,KAAK,CAAC,UAAU,CAAC,IAAI;gBAC3B,EAAE,EAAE,KAAK,CAAC,UAAU,CAAC,EAAE;gBACvB,MAAM,EAAE,KAAK,CAAC,UAAU,CAAC,MAAM;aAClC,CAAC;QACN,KAAK,aAAa;YACd,OAAO,EAAE,KAAK,EAAE,aAAa,EAAE,QAAQ,EAAE,KAAK,CAAC,UAAU,CAAC,QAAQ,EAAE,CAAC;QACzE,KAAK,SAAS;YACV,OAAO;gBACH,KAAK,EAAE,SAAS;gBAChB,MAAM,EAAE,KAAK,CAAC,UAAU,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC;oBAC5C,IAAI,EAAE,KAAK,CAAC,IAAI;oBAChB,IAAI,EAAE,CAAC,GAAG,KAAK,CAAC,IAAI,CAAC;iBACxB,CAAC,CAAC;aACN,CAAC;IACV,CAAC;AACL,CAAC;AACD,SAAS,iBAAiB,CAAC,OAAiD;IACxE,MAAM,QAAQ,GAAgB,EAAE,CAAC;IACjC,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;QAC1B,MAAM,OAAO,GAAG,KAAK,CAAC,KAAK,CAAC,OAAO,CAAC;QACpC,MAAM,MAAM,GAAG,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC;QAClC,MAAM,IAAI,GAAyC;YAC/C,OAAO,EAAE,KAAK,CAAC,KAAK,CAAC,OAAO;YAC5B,SAAS,EAAE,CAAC,GAAG,OAAO,CAAC,SAAS,CAAC;YACjC,OAAO,EAAE,KAAK,CAAC,KAAK,KAAK,WAAW,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,KAAK;YACjE,QAAQ,EAAE,gBAAgB,CAAC,KAAK,CAAC,KAAK,CAAC,QAAQ,CAAC;YAChD,WAAW,EAAE,KAAK,CAAC,KAAK,CAAC,WAAW,CAAC,GAAG,CAAC,kBAAkB,CAAC;YAC5D,iBAAiB,EAAE,yBAAyB,CAAC,MAAM,CAAC,KAAK,CAAC;SAC7D,CAAC;QACF,IAAI,OAAO,CAAC,KAAK,KAAK,SAAS;YAC3B,IAAI,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC;QAC/B,IAAI,OAAO,CAAC,GAAG,KAAK,SAAS;YACzB,IAAI,CAAC,GAAG,GAAG,OAAO,CAAC,GAAG,CAAC;QAC3B,MAAM,MAAM,GAAG,eAAe,CAAC,KAAK,CAAC,CAAC;QACtC,IAAI,MAAM,KAAK,SAAS;YACpB,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;QACzB,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IACxB,CAAC;IACD,OAAO,QAAQ,CAAC;AACpB,CAAC;AACD,SAAS,eAAe,CAAC,KAAoC;IACzD,IAAI,KAAK,CAAC,KAAK,KAAK,WAAW;QAC3B,OAAO,KAAK,CAAC,MAAM,CAAC;IACxB,QAAQ,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC,KAAK,EAAE,CAAC;QAC/B,KAAK,SAAS;YACV,OAAO,SAAS,CAAC;QACrB,KAAK,YAAY,CAAC;QAClB,KAAK,aAAa,CAAC;QACnB,KAAK,SAAS,CAAC;QACf,KAAK,gBAAgB;YACjB,OAAO,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC,MAAM,CAAC;IACzC,CAAC;AACL,CAAC;AACD,SAAS,gBAAgB,CAAC,QAA6B;IACnD,MAAM,QAAQ,GAAyC,EAAE,KAAK,EAAE,QAAQ,CAAC,KAAK,EAAE,CAAC;IACjF,IAAI,QAAQ,CAAC,KAAK,KAAK,UAAU;QAC7B,QAAQ,CAAC,MAAM,GAAG,QAAQ,CAAC,MAAM,CAAC;IACtC,OAAO,QAAQ,CAAC;AACpB,CAAC;AACD,SAAS,kBAAkB,CAAC,UAAiC;IACzD,QAAQ,UAAU,CAAC,IAAI,EAAE,CAAC;QACtB,KAAK,UAAU;YACX,OAAO,EAAE,IAAI,EAAE,UAAU,CAAC,IAAI,EAAE,UAAU,EAAE,UAAU,CAAC,UAAU,EAAE,CAAC;QACxE,KAAK,SAAS;YACV,OAAO;gBACH,IAAI,EAAE,UAAU,CAAC,IAAI;gBACrB,QAAQ,EAAE,UAAU,CAAC,QAAQ;gBAC7B,GAAG,CAAC,UAAU,CAAC,UAAU,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,UAAU,EAAE,UAAU,CAAC,UAAU,EAAE,CAAC;aACxF,CAAC;QACN,KAAK,WAAW;YACZ,OAAO,EAAE,IAAI,EAAE,UAAU,CAAC,IAAI,EAAE,YAAY,EAAE,UAAU,CAAC,YAAY,EAAE,CAAC;QAC5E,KAAK,kBAAkB;YACnB,OAAO,EAAE,IAAI,EAAE,UAAU,CAAC,IAAI,EAAE,CAAC;IACzC,CAAC;AACL,CAAC;AACD,SAAS,yBAAyB,CAAC,YAAiC;IAChE,QAAQ,YAAY,CAAC,KAAK,EAAE,CAAC;QACzB,KAAK,YAAY;YACb,OAAO,EAAE,KAAK,EAAE,YAAY,CAAC,KAAK,EAAE,CAAC;QACzC,KAAK,QAAQ;YACT,OAAO;gBACH,KAAK,EAAE,YAAY,CAAC,KAAK;gBACzB,MAAM,EAAE,YAAY,CAAC,MAAM;gBAC3B,MAAM,EAAE,YAAY,CAAC,MAAM;aAC9B,CAAC;QACN,KAAK,aAAa;YACd,OAAO;gBACH,KAAK,EAAE,YAAY,CAAC,KAAK;gBACzB,MAAM,EAAE,YAAY,CAAC,MAAM;gBAC3B,MAAM,EAAE,YAAY,CAAC,MAAM;aAC9B,CAAC;IACV,CAAC;AACL,CAAC;AACD,SAAS,aAAa,CAAC,KAAwB;IAC3C,QAAQ,KAAK,CAAC,KAAK,EAAE,CAAC;QAClB,KAAK,UAAU;YACX,OAAO,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,EAAE,MAAM,EAAE,KAAK,CAAC,MAAM,EAAE,UAAU,EAAE,KAAK,CAAC,UAAU,EAAE,CAAC;QACtF,KAAK,SAAS;YACV,OAAO;gBACH,KAAK,EAAE,KAAK,CAAC,KAAK;gBAClB,MAAM,EAAE,KAAK,CAAC,MAAM;gBACpB,OAAO,EAAE,CAAC,GAAG,KAAK,CAAC,OAAO,CAAC;gBAC3B,UAAU,EAAE,KAAK,CAAC,UAAU;aAC/B,CAAC;QACN,KAAK,aAAa;YACd,OAAO;gBACH,KAAK,EAAE,KAAK,CAAC,KAAK;gBAClB,OAAO,EAAE,CAAC,GAAG,KAAK,CAAC,OAAO,CAAC;gBAC3B,UAAU,EAAE,KAAK,CAAC,UAAU;aAC/B,CAAC;IACV,CAAC;AACL,CAAC;AACD,uDAAuD;AACvD,SAAS,iBAAiB,CAAC,MAA4B;IACnD,IAAI,CAAC,MAAM;QACP,OAAO,SAAS,CAAC;IACrB,KAAK,IAAI,CAAC,GAAG,MAAM,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC1C,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC;QACpB,IAAI,CAAC,CAAC,IAAI,KAAK,SAAS,IAAI,CAAC,CAAC,IAAI,KAAK,WAAW;YAC9C,OAAO,CAAC,CAAC,IAAI,CAAC;IACtB,CAAC;IACD,OAAO,SAAS,CAAC;AACrB,CAAC;AACD,uEAAuE;AACvE,MAAM,kBAAkB,GAAG,YAAY,CAAC;AACxC,KAAK,UAAU,iBAAiB;IAC5B,IAAI,CAAC;QACD,OAAO,CAAC,MAAM,uBAAuB,CAAC,kBAAkB,CAAC,CAA6B,CAAC;IAC3F,CAAC;IACD,MAAM,CAAC;QACH,MAAM,IAAI,KAAK,CAAC,2JAA2J,CAAC,CAAC;IACjL,CAAC;AACL,CAAC","sourcesContent":["// Braintrust 报告器:把一次 niceeval 运行作为一个 Braintrust experiment 上报,\n// 每个 attempt 一行(scores = 断言,metrics = 时长/token/成本),跨提交比较与团队共享。\n// `braintrust` 是可选依赖:动态 import,装了才用得上,没装在 onInvocationStart 报错\n// (reporter 错误按框架约定只记 diagnostic,不会让运行崩)。\nimport type { EvalResult, JsonValue, Reporter } from \"../../types.ts\";\nimport type { AssertionCoverageV1, AssertionEntryReadV1, AssertionLimitationV1, AssertionsProjectionV1, ScoreContributionV1, } from \"../../assertions/record/model.ts\";\nimport type { ScoreProjectionV1 } from \"../../eval/record/score.ts\";\nimport type { AttemptSlotProjectedEntry } from \"../../projection/index.ts\";\nimport type { Verdict } from \"../../shared/types.ts\";\nimport { reportActivity } from \"../feedback/sink.ts\";\nexport interface BraintrustConfig {\n    /** Braintrust 项目名;省略时用 \"niceeval\"。 */\n    project?: string;\n    /** Braintrust 项目 id;与 project 给一个即可。 */\n    projectId?: string;\n    /** 实验名;省略时由 Braintrust 自动命名。 */\n    experiment?: string;\n    /** 作为对比基线(diff base)的既有实验名。 */\n    baseExperiment?: string;\n    /** 作为对比基线(diff base)的既有实验 id。 */\n    baseExperimentId?: string;\n    /** true = 更新同名既有实验,而不是新建一个。 */\n    update?: boolean;\n    /** 实验级附加 metadata;与 niceeval 自动写入的字段合并,同名以这里为准。 */\n    metadata?: globalThis.Record<string, JsonValue>;\n    /** API key;省略时 SDK 读 BRAINTRUST_API_KEY 环境变量。 */\n    apiKey?: string;\n}\n/**\n * Braintrust SDK 的最小类型面。本地声明是为了不把 `braintrust` 变成编译期依赖;\n * 动态 import 后按这个形状断言。\n */\ninterface BraintrustSdk {\n    // SDK 里 init 同步返回(登录与建实验都是惰性的);统一 await,同步/异步都兼容。\n    init(options: {\n        project?: string;\n        projectId?: string;\n        experiment?: string;\n        baseExperiment?: string;\n        baseExperimentId?: string;\n        update?: boolean;\n        apiKey?: string;\n        metadata?: globalThis.Record<string, JsonValue>;\n        setCurrent?: boolean;\n    }): BraintrustExperiment | Promise<BraintrustExperiment>;\n    flush(): Promise<void>;\n}\ninterface BraintrustExperiment {\n    log(event: BraintrustLogEvent): void;\n    summarize(): Promise<{\n        experimentUrl?: string;\n    }>;\n}\nexport interface BraintrustLogEvent {\n    id?: string;\n    input?: string;\n    output?: string;\n    error?: EvalResult[\"error\"];\n    scores?: globalThis.Record<string, number>;\n    metadata?: globalThis.Record<string, JsonValue>;\n    metrics?: globalThis.Record<string, number>;\n}\n/**\n * A post-Record Projection calculation supplies this optional assessment. The\n * live Reporter callback currently has no RecordProjection input, so its\n * absence is represented explicitly rather than as a made-up Attachment read.\n */\nexport type BraintrustAssessment = {\n    readonly state: \"projected\";\n    /** Exact Report-style public projections, including every data state. */\n    readonly assertions: AttemptSlotProjectedEntry<AssertionsProjectionV1<unknown>>;\n    readonly verdict: AttemptSlotProjectedEntry<Verdict>;\n    readonly score?: AttemptSlotProjectedEntry<ScoreProjectionV1>;\n} | {\n    readonly state: \"not-projected\";\n    readonly reason: \"reporter-contract-does-not-provide-record-projections\";\n};\nconst NOT_PROJECTED_ASSESSMENT: BraintrustAssessment = Object.freeze({\n    state: \"not-projected\" as const,\n    reason: \"reporter-contract-does-not-provide-record-projections\" as const,\n});\n/**\n * 创建 Braintrust 报告器。挂在 `defineConfig({ reporters })` 上观测整次运行,\n * 或挂在单个 eval 的 `reporters` 上只观测它(同一实例被多个 eval 引用时共享一个实验)。\n */\nexport function Braintrust(config: BraintrustConfig = {}): Reporter {\n    let sdk: BraintrustSdk | undefined;\n    let experiment: BraintrustExperiment | undefined;\n    return {\n        async onInvocationStart(evals) {\n            sdk = await loadBraintrustSdk();\n            experiment = await sdk.init({\n                project: config.projectId ? undefined : (config.project ?? \"niceeval\"),\n                projectId: config.projectId,\n                experiment: config.experiment,\n                baseExperiment: config.baseExperiment,\n                baseExperimentId: config.baseExperimentId,\n                update: config.update,\n                apiKey: config.apiKey,\n                // 不再写顶层单一 agent:一次 Invocation 可能横跨多个 (agent, model, flags) 配置,\n                // 启动时还没有任何结果,写一个必然只代表其中一份配置的值就是撒谎(见\n                // docs/runner.md「Reporter 与运行器事件」)。每行自己的 agent 身份仍在\n                // toBraintrustEvent() 的 metadata.agent 里,逐 attempt 精确;跨行的 agent 集合可以\n                // 从 Braintrust 自己按 metadata.agent 分组得到,不需要实验级再存一份。\n                metadata: {\n                    evals: evals.map((e) => e.id),\n                    ...config.metadata,\n                },\n                // 不设为全局 current experiment,避免污染用户代码里的 braintrust 全局态。\n                setCurrent: false,\n            });\n        },\n        onEvalComplete(result) {\n            experiment?.log(toBraintrustEvent(result, NOT_PROJECTED_ASSESSMENT));\n        },\n        async onInvocationComplete() {\n            if (!experiment)\n                return;\n            try {\n                await sdk?.flush();\n                const summary = await experiment.summarize();\n                if (summary.experimentUrl) {\n                    reportActivity(`Braintrust experiment: ${summary.experimentUrl}`);\n                }\n            }\n            finally {\n                experiment = undefined;\n                sdk = undefined;\n            }\n        },\n    };\n}\n/**\n * EvalResult identity/transport data plus an optional post-Record assessment\n * become one Braintrust row. Assertions, Verdict, and Score only enter through\n * that assessment's public Attachment projections; this function never reads\n * a historical result graph.\n */\nexport function toBraintrustEvent(result: EvalResult, assessment: BraintrustAssessment = NOT_PROJECTED_ASSESSMENT): BraintrustLogEvent {\n    const projectedVerdict = assessment.state === \"projected\"\n        ? availableProjectionValue(assessment.verdict)\n        : undefined;\n    const verdict = projectedVerdict ?? result.verdict;\n    const metrics: globalThis.Record<string, number> = {};\n    if (result.startedAt) {\n        const start = Date.parse(result.startedAt) / 1000;\n        if (Number.isFinite(start)) {\n            metrics.start = start;\n            metrics.end = start + result.durationMs / 1000;\n        }\n    }\n    if (result.usage) {\n        const { inputTokens, outputTokens, cacheReadTokens, cacheCreationTokens, reasoningTokens, requests } = result.usage;\n        if (inputTokens !== undefined)\n            metrics.prompt_tokens = inputTokens;\n        if (outputTokens !== undefined)\n            metrics.completion_tokens = outputTokens;\n        if (inputTokens !== undefined && outputTokens !== undefined)\n            metrics.tokens = inputTokens + outputTokens;\n        if (cacheReadTokens !== undefined)\n            metrics.cache_read_tokens = cacheReadTokens;\n        if (cacheCreationTokens !== undefined)\n            metrics.cache_creation_tokens = cacheCreationTokens;\n        if (reasoningTokens !== undefined)\n            metrics.reasoning_tokens = reasoningTokens;\n        if (requests !== undefined)\n            metrics.requests = requests;\n    }\n    if (result.estimatedCostUSD !== undefined)\n        metrics.estimated_cost_usd = result.estimatedCostUSD;\n    const metadata: globalThis.Record<string, JsonValue> = {\n        eval: result.id,\n        agent: result.agent,\n        attempt: result.attempt,\n        verdict,\n        status: verdict,\n        evaluationAlgorithm: result.evaluationAlgorithm,\n        evaluationKind: result.evaluationKind,\n        assessmentProjection: assessment.state,\n    };\n    const scores: globalThis.Record<string, number> = {};\n    if (assessment.state === \"projected\") {\n        metadata.assertions = projectionMetadata(assessment.assertions, (assertions) => ({ entries: assertionMetadata(assertions.entries) }));\n        metadata.verdictProjection = projectionMetadata(assessment.verdict, (value) => ({ value }));\n        if (assessment.score !== undefined) {\n            metadata.score = projectionMetadata(assessment.score, scoreMetadata);\n            const score = availableProjectionValue(assessment.score);\n            if (score?.state === \"complete\")\n                scores[\"niceeval:score\"] = score.earned;\n        }\n    }\n    else {\n        metadata.assessmentProjectionReason = assessment.reason;\n    }\n    if (result.model !== undefined)\n        metadata.model = result.model;\n    if (result.experimentId !== undefined)\n        metadata.experiment = result.experimentId;\n    if (result.experiment?.flags && Object.keys(result.experiment.flags).length > 0) {\n        metadata.flags = result.experiment.flags;\n    }\n    if (result.skipReason !== undefined)\n        metadata.skipReason = result.skipReason;\n    // 一次运行内 (experiment, eval, agent, model, attempt) 唯一;Braintrust 按 id 合并重复行。\n    const id = [result.experimentId ?? \"\", result.id, result.agent, result.model ?? \"\", `a${result.attempt}`].join(\"|\");\n    return {\n        id,\n        input: result.description ?? result.id,\n        output: lastAssistantText(result.events),\n        error: result.error,\n        metadata,\n        metrics,\n        ...(Object.keys(scores).length > 0 ? { scores } : {}),\n    };\n}\nfunction availableProjectionValue<Value>(entry: AttemptSlotProjectedEntry<Value>): Value | undefined {\n    return entry.state === \"attachment-result\" && entry.attachment.state === \"available\"\n        ? entry.attachment.value\n        : undefined;\n}\nfunction projectionMetadata<Value>(entry: AttemptSlotProjectedEntry<Value>, available: (value: Value) => globalThis.Record<string, JsonValue>): JsonValue {\n    if (entry.state !== \"attachment-result\") {\n        return { state: entry.state };\n    }\n    switch (entry.attachment.state) {\n        case \"available\":\n            return { state: \"available\", ...available(entry.attachment.value) };\n        case \"unavailable\":\n            return { state: \"unavailable\" };\n        case \"migration-required\":\n            return {\n                state: \"migration-required\",\n                from: entry.attachment.from,\n                to: entry.attachment.to,\n                command: entry.attachment.command,\n            };\n        case \"migration-unavailable\":\n            return {\n                state: \"migration-unavailable\",\n                from: entry.attachment.from,\n                to: entry.attachment.to,\n                reason: entry.attachment.reason,\n            };\n        case \"unsupported\":\n            return { state: \"unsupported\", schemaId: entry.attachment.schemaId };\n        case \"invalid\":\n            return {\n                state: \"invalid\",\n                issues: entry.attachment.issues.map((issue) => ({\n                    code: issue.code,\n                    path: [...issue.path],\n                })),\n            };\n    }\n}\nfunction assertionMetadata(entries: readonly AssertionEntryReadV1<unknown>[]): JsonValue {\n    const metadata: JsonValue[] = [];\n    for (const entry of entries) {\n        const display = entry.entry.display;\n        const result = entry.entry.result;\n        const item: globalThis.Record<string, JsonValue> = {\n            entryId: entry.entry.entryId,\n            groupPath: [...display.groupPath],\n            outcome: entry.state === \"available\" ? result.state : entry.state,\n            coverage: coverageMetadata(entry.entry.coverage),\n            limitations: entry.entry.limitations.map(limitationMetadata),\n            scoreContribution: scoreContributionMetadata(result.score),\n        };\n        if (display.label !== undefined)\n            item.label = display.label;\n        if (display.key !== undefined)\n            item.key = display.key;\n        const reason = assertionReason(entry);\n        if (reason !== undefined)\n            item.reason = reason;\n        metadata.push(item);\n    }\n    return metadata;\n}\nfunction assertionReason(entry: AssertionEntryReadV1<unknown>): string | undefined {\n    if (entry.state !== \"available\")\n        return entry.reason;\n    switch (entry.entry.result.state) {\n        case \"matched\":\n            return undefined;\n        case \"mismatched\":\n        case \"unavailable\":\n        case \"errored\":\n        case \"not-applicable\":\n            return entry.entry.result.reason;\n    }\n}\nfunction coverageMetadata(coverage: AssertionCoverageV1): globalThis.Record<string, JsonValue> {\n    const metadata: globalThis.Record<string, JsonValue> = { state: coverage.state };\n    if (coverage.state !== \"complete\")\n        metadata.reason = coverage.reason;\n    return metadata;\n}\nfunction limitationMetadata(limitation: AssertionLimitationV1): globalThis.Record<string, JsonValue> {\n    switch (limitation.kind) {\n        case \"redacted\":\n            return { kind: limitation.kind, fieldCount: limitation.fieldCount };\n        case \"sampled\":\n            return {\n                kind: limitation.kind,\n                captured: limitation.captured,\n                ...(limitation.knownTotal === undefined ? {} : { knownTotal: limitation.knownTotal }),\n            };\n        case \"truncated\":\n            return { kind: limitation.kind, omittedBytes: limitation.omittedBytes };\n        case \"provider-limited\":\n            return { kind: limitation.kind };\n    }\n}\nfunction scoreContributionMetadata(contribution: ScoreContributionV1): globalThis.Record<string, JsonValue> {\n    switch (contribution.state) {\n        case \"not-scored\":\n            return { state: contribution.state };\n        case \"earned\":\n            return {\n                state: contribution.state,\n                points: contribution.points,\n                earned: contribution.earned,\n            };\n        case \"unavailable\":\n            return {\n                state: contribution.state,\n                points: contribution.points,\n                reason: contribution.reason,\n            };\n    }\n}\nfunction scoreMetadata(score: ScoreProjectionV1): globalThis.Record<string, JsonValue> {\n    switch (score.state) {\n        case \"complete\":\n            return { state: score.state, earned: score.earned, comparable: score.comparable };\n        case \"partial\":\n            return {\n                state: score.state,\n                earned: score.earned,\n                reasons: [...score.reasons],\n                comparable: score.comparable,\n            };\n        case \"unavailable\":\n            return {\n                state: score.state,\n                reasons: [...score.reasons],\n                comparable: score.comparable,\n            };\n    }\n}\n/** agent 的最终回复文本(事件流里最后一条 assistant message);没有就不填。 */\nfunction lastAssistantText(events: EvalResult[\"events\"]): string | undefined {\n    if (!events)\n        return undefined;\n    for (let i = events.length - 1; i >= 0; i--) {\n        const e = events[i];\n        if (e.type === \"message\" && e.role === \"assistant\")\n            return e.text;\n    }\n    return undefined;\n}\n/** 动态 import 说明:specifier 放变量里,`braintrust` 不进编译期依赖,没装也能 typecheck。 */\nconst BRAINTRUST_PACKAGE = \"braintrust\";\nasync function loadBraintrustSdk(): Promise<BraintrustSdk> {\n    try {\n        return (await __niceevalDynamicImport(BRAINTRUST_PACKAGE)) as unknown as BraintrustSdk;\n    }\n    catch {\n        throw new Error(\"The 'braintrust' package is required for the Braintrust reporter but was not found. Install it with: npm install braintrust (and set BRAINTRUST_API_KEY).\");\n    }\n}\n"]}
+{"version":3,"file":"braintrust.js","sourceRoot":"","sources":["braintrust.ts"],"names":[],"mappings":";;AAqGA,gCA8CC;AAOD,8CA8EC;AA/MD,iDAAqD;AAoErD,MAAM,wBAAwB,GAAyB,MAAM,CAAC,MAAM,CAAC;IACjE,KAAK,EAAE,eAAwB;IAC/B,MAAM,EAAE,uDAAgE;CAC3E,CAAC,CAAC;AACH;;;GAGG;AACH,SAAgB,UAAU,CAAC,SAA2B,EAAE;IACpD,IAAI,GAA8B,CAAC;IACnC,IAAI,UAA4C,CAAC;IACjD,OAAO;QACH,KAAK,CAAC,iBAAiB,CAAC,KAAK;YACzB,GAAG,GAAG,MAAM,iBAAiB,EAAE,CAAC;YAChC,UAAU,GAAG,MAAM,GAAG,CAAC,IAAI,CAAC;gBACxB,OAAO,EAAE,MAAM,CAAC,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,OAAO,IAAI,UAAU,CAAC;gBACtE,SAAS,EAAE,MAAM,CAAC,SAAS;gBAC3B,UAAU,EAAE,MAAM,CAAC,UAAU;gBAC7B,cAAc,EAAE,MAAM,CAAC,cAAc;gBACrC,gBAAgB,EAAE,MAAM,CAAC,gBAAgB;gBACzC,MAAM,EAAE,MAAM,CAAC,MAAM;gBACrB,MAAM,EAAE,MAAM,CAAC,MAAM;gBACrB,+DAA+D;gBAC/D,oCAAoC;gBACpC,oDAAoD;gBACpD,qEAAqE;gBACrE,mDAAmD;gBACnD,QAAQ,EAAE;oBACN,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;oBAC7B,GAAG,MAAM,CAAC,QAAQ;iBACrB;gBACD,sDAAsD;gBACtD,UAAU,EAAE,KAAK;aACpB,CAAC,CAAC;QACP,CAAC;QACD,cAAc,CAAC,MAAM;YACjB,UAAU,EAAE,GAAG,CAAC,iBAAiB,CAAC,MAAM,EAAE,wBAAwB,CAAC,CAAC,CAAC;QACzE,CAAC;QACD,KAAK,CAAC,oBAAoB;YACtB,IAAI,CAAC,UAAU;gBACX,OAAO;YACX,IAAI,CAAC;gBACD,MAAM,GAAG,EAAE,KAAK,EAAE,CAAC;gBACnB,MAAM,OAAO,GAAG,MAAM,UAAU,CAAC,SAAS,EAAE,CAAC;gBAC7C,IAAI,OAAO,CAAC,aAAa,EAAE,CAAC;oBACxB,IAAA,wBAAc,EAAC,0BAA0B,OAAO,CAAC,aAAa,EAAE,CAAC,CAAC;gBACtE,CAAC;YACL,CAAC;oBACO,CAAC;gBACL,UAAU,GAAG,SAAS,CAAC;gBACvB,GAAG,GAAG,SAAS,CAAC;YACpB,CAAC;QACL,CAAC;KACJ,CAAC;AACN,CAAC;AACD;;;;;GAKG;AACH,SAAgB,iBAAiB,CAAC,MAAkB,EAAE,aAAmC,wBAAwB;IAC7G,MAAM,gBAAgB,GAAG,UAAU,CAAC,KAAK,KAAK,WAAW;QACrD,CAAC,CAAC,wBAAwB,CAAC,UAAU,CAAC,OAAO,CAAC;QAC9C,CAAC,CAAC,SAAS,CAAC;IAChB,MAAM,OAAO,GAAG,gBAAgB,IAAI,MAAM,CAAC,OAAO,CAAC;IACnD,MAAM,OAAO,GAAsC,EAAE,CAAC;IACtD,IAAI,MAAM,CAAC,SAAS,EAAE,CAAC;QACnB,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,MAAM,CAAC,SAAS,CAAC,GAAG,IAAI,CAAC;QAClD,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,EAAE,CAAC;YACzB,OAAO,CAAC,KAAK,GAAG,KAAK,CAAC;YACtB,OAAO,CAAC,GAAG,GAAG,KAAK,GAAG,MAAM,CAAC,UAAU,GAAG,IAAI,CAAC;QACnD,CAAC;IACL,CAAC;IACD,IAAI,MAAM,CAAC,KAAK,EAAE,CAAC;QACf,MAAM,EAAE,WAAW,EAAE,YAAY,EAAE,eAAe,EAAE,mBAAmB,EAAE,eAAe,EAAE,QAAQ,EAAE,GAAG,MAAM,CAAC,KAAK,CAAC;QACpH,IAAI,WAAW,KAAK,SAAS;YACzB,OAAO,CAAC,aAAa,GAAG,WAAW,CAAC;QACxC,IAAI,YAAY,KAAK,SAAS;YAC1B,OAAO,CAAC,iBAAiB,GAAG,YAAY,CAAC;QAC7C,IAAI,WAAW,KAAK,SAAS,IAAI,YAAY,KAAK,SAAS;YACvD,OAAO,CAAC,MAAM,GAAG,WAAW,GAAG,YAAY,CAAC;QAChD,IAAI,eAAe,KAAK,SAAS;YAC7B,OAAO,CAAC,iBAAiB,GAAG,eAAe,CAAC;QAChD,IAAI,mBAAmB,KAAK,SAAS;YACjC,OAAO,CAAC,qBAAqB,GAAG,mBAAmB,CAAC;QACxD,IAAI,eAAe,KAAK,SAAS;YAC7B,OAAO,CAAC,gBAAgB,GAAG,eAAe,CAAC;QAC/C,IAAI,QAAQ,KAAK,SAAS;YACtB,OAAO,CAAC,QAAQ,GAAG,QAAQ,CAAC;IACpC,CAAC;IACD,0EAA0E;IAC1E,+CAA+C;IAC/C,IAAI,MAAM,CAAC,gBAAgB,KAAK,SAAS;QACrC,OAAO,CAAC,kBAAkB,GAAG,MAAM,CAAC,gBAAgB,CAAC;IACzD,MAAM,QAAQ,GAAyC;QACnD,IAAI,EAAE,MAAM,CAAC,EAAE;QACf,KAAK,EAAE,MAAM,CAAC,KAAK;QACnB,OAAO,EAAE,MAAM,CAAC,OAAO;QACvB,OAAO;QACP,MAAM,EAAE,OAAO;QACf,mBAAmB,EAAE,MAAM,CAAC,mBAAmB;QAC/C,cAAc,EAAE,MAAM,CAAC,cAAc;QACrC,oBAAoB,EAAE,UAAU,CAAC,KAAK;KACzC,CAAC;IACF,MAAM,MAAM,GAAsC,EAAE,CAAC;IACrD,IAAI,UAAU,CAAC,KAAK,KAAK,WAAW,EAAE,CAAC;QACnC,QAAQ,CAAC,UAAU,GAAG,kBAAkB,CAAC,UAAU,CAAC,UAAU,EAAE,CAAC,UAAU,EAAE,EAAE,CAAC,CAAC,EAAE,OAAO,EAAE,iBAAiB,CAAC,UAAU,CAAC,OAAO,CAAC,EAAE,CAAC,CAAC,CAAC;QACtI,QAAQ,CAAC,iBAAiB,GAAG,kBAAkB,CAAC,UAAU,CAAC,OAAO,EAAE,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC,CAAC;QAC5F,IAAI,UAAU,CAAC,KAAK,KAAK,SAAS,EAAE,CAAC;YACjC,QAAQ,CAAC,KAAK,GAAG,kBAAkB,CAAC,UAAU,CAAC,KAAK,EAAE,aAAa,CAAC,CAAC;YACrE,MAAM,KAAK,GAAG,wBAAwB,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC;YACzD,IAAI,KAAK,EAAE,KAAK,KAAK,UAAU;gBAC3B,MAAM,CAAC,gBAAgB,CAAC,GAAG,KAAK,CAAC,MAAM,CAAC;QAChD,CAAC;IACL,CAAC;SACI,CAAC;QACF,QAAQ,CAAC,0BAA0B,GAAG,UAAU,CAAC,MAAM,CAAC;IAC5D,CAAC;IACD,IAAI,MAAM,CAAC,KAAK,KAAK,SAAS;QAC1B,QAAQ,CAAC,KAAK,GAAG,MAAM,CAAC,KAAK,CAAC;IAClC,IAAI,MAAM,CAAC,YAAY,KAAK,SAAS;QACjC,QAAQ,CAAC,UAAU,GAAG,MAAM,CAAC,YAAY,CAAC;IAC9C,IAAI,MAAM,CAAC,UAAU,EAAE,KAAK,IAAI,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,UAAU,CAAC,KAAK,CAAC,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC9E,QAAQ,CAAC,KAAK,GAAG,MAAM,CAAC,UAAU,CAAC,KAAK,CAAC;IAC7C,CAAC;IACD,IAAI,MAAM,CAAC,UAAU,KAAK,SAAS;QAC/B,QAAQ,CAAC,UAAU,GAAG,MAAM,CAAC,UAAU,CAAC;IAC5C,4EAA4E;IAC5E,MAAM,EAAE,GAAG,CAAC,MAAM,CAAC,YAAY,IAAI,EAAE,EAAE,MAAM,CAAC,EAAE,EAAE,MAAM,CAAC,KAAK,EAAE,MAAM,CAAC,KAAK,IAAI,EAAE,EAAE,IAAI,MAAM,CAAC,OAAO,EAAE,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;IACpH,OAAO;QACH,EAAE;QACF,KAAK,EAAE,MAAM,CAAC,WAAW,IAAI,MAAM,CAAC,EAAE;QACtC,MAAM,EAAE,iBAAiB,CAAC,MAAM,CAAC,MAAM,CAAC;QACxC,KAAK,EAAE,MAAM,CAAC,KAAK;QACnB,QAAQ;QACR,OAAO;QACP,GAAG,CAAC,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;KACxD,CAAC;AACN,CAAC;AACD,SAAS,wBAAwB,CAAQ,KAAuC;IAC5E,OAAO,KAAK,CAAC,KAAK,KAAK,mBAAmB,IAAI,KAAK,CAAC,UAAU,CAAC,KAAK,KAAK,WAAW;QAChF,CAAC,CAAC,KAAK,CAAC,UAAU,CAAC,KAAK;QACxB,CAAC,CAAC,SAAS,CAAC;AACpB,CAAC;AACD,SAAS,kBAAkB,CAAQ,KAAuC,EAAE,SAAiE;IACzI,IAAI,KAAK,CAAC,KAAK,KAAK,mBAAmB,EAAE,CAAC;QACtC,OAAO,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,EAAE,CAAC;IAClC,CAAC;IACD,QAAQ,KAAK,CAAC,UAAU,CAAC,KAAK,EAAE,CAAC;QAC7B,KAAK,WAAW;YACZ,OAAO,EAAE,KAAK,EAAE,WAAW,EAAE,GAAG,SAAS,CAAC,KAAK,CAAC,UAAU,CAAC,KAAK,CAAC,EAAE,CAAC;QACxE,KAAK,aAAa;YACd,OAAO,EAAE,KAAK,EAAE,aAAa,EAAE,CAAC;QACpC,KAAK,oBAAoB;YACrB,OAAO;gBACH,KAAK,EAAE,oBAAoB;gBAC3B,IAAI,EAAE,KAAK,CAAC,UAAU,CAAC,IAAI;gBAC3B,EAAE,EAAE,KAAK,CAAC,UAAU,CAAC,EAAE;gBACvB,GAAG,CAAC,KAAK,CAAC,UAAU,CAAC,OAAO,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,KAAK,CAAC,UAAU,CAAC,OAAO,EAAE,CAAC;aAC3F,CAAC;QACN,KAAK,aAAa,CAAC;QACnB,KAAK,SAAS;YACV,OAAO,EAAE,KAAK,EAAE,KAAK,CAAC,UAAU,CAAC,KAAK,EAAE,CAAC;IACjD,CAAC;AACL,CAAC;AACD,SAAS,iBAAiB,CAAC,OAA+C;IACtE,MAAM,QAAQ,GAAgB,EAAE,CAAC;IACjC,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;QAC1B,MAAM,OAAO,GAAG,KAAK,CAAC,KAAK,CAAC,OAAO,CAAC;QACpC,MAAM,MAAM,GAAG,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC;QAClC,MAAM,IAAI,GAAyC;YAC/C,OAAO,EAAE,KAAK,CAAC,KAAK,CAAC,OAAO;YAC5B,SAAS,EAAE,CAAC,GAAG,OAAO,CAAC,SAAS,CAAC;YACjC,OAAO,EAAE,KAAK,CAAC,KAAK,KAAK,WAAW,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,KAAK,CAAC,KAAK;YACjE,QAAQ,EAAE,gBAAgB,CAAC,KAAK,CAAC,KAAK,CAAC,QAAQ,CAAC;YAChD,WAAW,EAAE,KAAK,CAAC,KAAK,CAAC,WAAW,CAAC,GAAG,CAAC,kBAAkB,CAAC;YAC5D,iBAAiB,EAAE,yBAAyB,CAAC,MAAM,CAAC,KAAK,CAAC;SAC7D,CAAC;QACF,IAAI,OAAO,CAAC,KAAK,KAAK,SAAS;YAC3B,IAAI,CAAC,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC;QAC/B,IAAI,OAAO,CAAC,GAAG,KAAK,SAAS;YACzB,IAAI,CAAC,GAAG,GAAG,OAAO,CAAC,GAAG,CAAC;QAC3B,MAAM,MAAM,GAAG,eAAe,CAAC,KAAK,CAAC,CAAC;QACtC,IAAI,MAAM,KAAK,SAAS;YACpB,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;QACzB,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IACxB,CAAC;IACD,OAAO,QAAQ,CAAC;AACpB,CAAC;AACD,SAAS,eAAe,CAAC,KAAkC;IACvD,IAAI,KAAK,CAAC,KAAK,KAAK,WAAW;QAC3B,OAAO,KAAK,CAAC,MAAM,CAAC;IACxB,QAAQ,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC,KAAK,EAAE,CAAC;QAC/B,KAAK,SAAS;YACV,OAAO,SAAS,CAAC;QACrB,KAAK,YAAY,CAAC;QAClB,KAAK,aAAa,CAAC;QACnB,KAAK,SAAS,CAAC;QACf,KAAK,gBAAgB;YACjB,OAAO,KAAK,CAAC,KAAK,CAAC,MAAM,CAAC,MAAM,CAAC;IACzC,CAAC;AACL,CAAC;AACD,SAAS,gBAAgB,CAAC,QAA2B;IACjD,MAAM,QAAQ,GAAyC,EAAE,KAAK,EAAE,QAAQ,CAAC,KAAK,EAAE,CAAC;IACjF,IAAI,QAAQ,CAAC,KAAK,KAAK,UAAU;QAC7B,QAAQ,CAAC,MAAM,GAAG,QAAQ,CAAC,MAAM,CAAC;IACtC,OAAO,QAAQ,CAAC;AACpB,CAAC;AACD,SAAS,kBAAkB,CAAC,UAA+B;IACvD,QAAQ,UAAU,CAAC,IAAI,EAAE,CAAC;QACtB,KAAK,UAAU;YACX,OAAO,EAAE,IAAI,EAAE,UAAU,CAAC,IAAI,EAAE,UAAU,EAAE,UAAU,CAAC,UAAU,EAAE,CAAC;QACxE,KAAK,SAAS;YACV,OAAO;gBACH,IAAI,EAAE,UAAU,CAAC,IAAI;gBACrB,QAAQ,EAAE,UAAU,CAAC,QAAQ;gBAC7B,GAAG,CAAC,UAAU,CAAC,UAAU,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,UAAU,EAAE,UAAU,CAAC,UAAU,EAAE,CAAC;aACxF,CAAC;QACN,KAAK,WAAW;YACZ,OAAO,EAAE,IAAI,EAAE,UAAU,CAAC,IAAI,EAAE,YAAY,EAAE,UAAU,CAAC,YAAY,EAAE,CAAC;QAC5E,KAAK,kBAAkB;YACnB,OAAO,EAAE,IAAI,EAAE,UAAU,CAAC,IAAI,EAAE,CAAC;IACzC,CAAC;AACL,CAAC;AACD,SAAS,yBAAyB,CAAC,YAA+B;IAC9D,QAAQ,YAAY,CAAC,KAAK,EAAE,CAAC;QACzB,KAAK,YAAY;YACb,OAAO,EAAE,KAAK,EAAE,YAAY,CAAC,KAAK,EAAE,CAAC;QACzC,KAAK,QAAQ;YACT,OAAO;gBACH,KAAK,EAAE,YAAY,CAAC,KAAK;gBACzB,MAAM,EAAE,YAAY,CAAC,MAAM;gBAC3B,MAAM,EAAE,YAAY,CAAC,MAAM;aAC9B,CAAC;QACN,KAAK,aAAa;YACd,OAAO;gBACH,KAAK,EAAE,YAAY,CAAC,KAAK;gBACzB,MAAM,EAAE,YAAY,CAAC,MAAM;gBAC3B,MAAM,EAAE,YAAY,CAAC,MAAM;aAC9B,CAAC;IACV,CAAC;AACL,CAAC;AACD,SAAS,aAAa,CAAC,KAAsB;IACzC,QAAQ,KAAK,CAAC,KAAK,EAAE,CAAC;QAClB,KAAK,UAAU;YACX,OAAO,EAAE,KAAK,EAAE,KAAK,CAAC,KAAK,EAAE,MAAM,EAAE,KAAK,CAAC,MAAM,EAAE,UAAU,EAAE,KAAK,CAAC,UAAU,EAAE,CAAC;QACtF,KAAK,SAAS;YACV,OAAO;gBACH,KAAK,EAAE,KAAK,CAAC,KAAK;gBAClB,MAAM,EAAE,KAAK,CAAC,MAAM;gBACpB,OAAO,EAAE,CAAC,GAAG,KAAK,CAAC,OAAO,CAAC;gBAC3B,UAAU,EAAE,KAAK,CAAC,UAAU;aAC/B,CAAC;QACN,KAAK,aAAa;YACd,OAAO;gBACH,KAAK,EAAE,KAAK,CAAC,KAAK;gBAClB,OAAO,EAAE,CAAC,GAAG,KAAK,CAAC,OAAO,CAAC;gBAC3B,UAAU,EAAE,KAAK,CAAC,UAAU;aAC/B,CAAC;IACV,CAAC;AACL,CAAC;AACD,uDAAuD;AACvD,SAAS,iBAAiB,CAAC,MAA4B;IACnD,IAAI,CAAC,MAAM;QACP,OAAO,SAAS,CAAC;IACrB,KAAK,IAAI,CAAC,GAAG,MAAM,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC1C,MAAM,CAAC,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC;QACpB,IAAI,CAAC,CAAC,IAAI,KAAK,SAAS,IAAI,CAAC,CAAC,IAAI,KAAK,WAAW;YAC9C,OAAO,CAAC,CAAC,IAAI,CAAC;IACtB,CAAC;IACD,OAAO,SAAS,CAAC;AACrB,CAAC;AACD,uEAAuE;AACvE,MAAM,kBAAkB,GAAG,YAAY,CAAC;AACxC,KAAK,UAAU,iBAAiB;IAC5B,IAAI,CAAC;QACD,OAAO,CAAC,MAAM,uBAAuB,CAAC,kBAAkB,CAAC,CAA6B,CAAC;IAC3F,CAAC;IACD,MAAM,CAAC;QACH,MAAM,IAAI,KAAK,CAAC,2JAA2J,CAAC,CAAC;IACjL,CAAC;AACL,CAAC","sourcesContent":["// Braintrust 报告器:把一次 niceeval 运行作为一个 Braintrust experiment 上报,\n// 每个 attempt 一行(scores = 断言,metrics = 时长/token/成本),跨提交比较与团队共享。\n// `braintrust` 是可选依赖:动态 import,装了才用得上,没装在 onInvocationStart 报错\n// (reporter 错误按框架约定只记 diagnostic,不会让运行崩)。\nimport type { EvalResult, JsonValue, Reporter } from \"../../types.ts\";\nimport type { AssertionCoverage, AssertionEntryRead, AssertionLimitation, AssertionsProjection, ScoreContribution, } from \"../../assertions/record/model.ts\";\nimport type { ScoreProjection } from \"../../eval/record/score.ts\";\n/** Closed assessment rows; this reporter does not reopen Record. */\ntype AttemptSlotProjectedEntry<Value> = {\n    readonly state: \"excluded\" | \"not-recorded\" | \"core-invalid\";\n} | {\n    readonly state: \"attachment-result\";\n    readonly attachment: {\n        readonly state: \"available\";\n        readonly value: Value;\n    } | {\n        readonly state: \"unavailable\" | \"unsupported\" | \"invalid\";\n    } | {\n        readonly state: \"migration-required\";\n        readonly from: string;\n        readonly to: string;\n        readonly command?: string;\n    };\n};\nimport type { Verdict } from \"../../shared/types.ts\";\nimport { reportActivity } from \"../feedback/sink.ts\";\nexport interface BraintrustConfig {\n    /** Braintrust 项目名;省略时用 \"niceeval\"。 */\n    project?: string;\n    /** Braintrust 项目 id;与 project 给一个即可。 */\n    projectId?: string;\n    /** 实验名;省略时由 Braintrust 自动命名。 */\n    experiment?: string;\n    /** 作为对比基线(diff base)的既有实验名。 */\n    baseExperiment?: string;\n    /** 作为对比基线(diff base)的既有实验 id。 */\n    baseExperimentId?: string;\n    /** true = 更新同名既有实验,而不是新建一个。 */\n    update?: boolean;\n    /** 实验级附加 metadata;与 niceeval 自动写入的字段合并,同名以这里为准。 */\n    metadata?: globalThis.Record<string, JsonValue>;\n    /** API key;省略时 SDK 读 BRAINTRUST_API_KEY 环境变量。 */\n    apiKey?: string;\n}\n/**\n * Braintrust SDK 的最小类型面。本地声明是为了不把 `braintrust` 变成编译期依赖;\n * 动态 import 后按这个形状断言。\n */\ninterface BraintrustSdk {\n    // SDK 里 init 同步返回(登录与建实验都是惰性的);统一 await,同步/异步都兼容。\n    init(options: {\n        project?: string;\n        projectId?: string;\n        experiment?: string;\n        baseExperiment?: string;\n        baseExperimentId?: string;\n        update?: boolean;\n        apiKey?: string;\n        metadata?: globalThis.Record<string, JsonValue>;\n        setCurrent?: boolean;\n    }): BraintrustExperiment | Promise<BraintrustExperiment>;\n    flush(): Promise<void>;\n}\ninterface BraintrustExperiment {\n    log(event: BraintrustLogEvent): void;\n    summarize(): Promise<{\n        experimentUrl?: string;\n    }>;\n}\nexport interface BraintrustLogEvent {\n    id?: string;\n    input?: string;\n    output?: string;\n    error?: EvalResult[\"error\"];\n    scores?: globalThis.Record<string, number>;\n    metadata?: globalThis.Record<string, JsonValue>;\n    metrics?: globalThis.Record<string, number>;\n}\n/**\n * A post-Record Projection calculation supplies this optional assessment. The\n * live Reporter callback currently has no RecordProjection input, so its\n * absence is represented explicitly rather than as a made-up Attachment read.\n */\nexport type BraintrustAssessment = {\n    readonly state: \"projected\";\n    /** Exact Report-style public projections, including every data state. */\n    readonly assertions: AttemptSlotProjectedEntry<AssertionsProjection<unknown>>;\n    readonly verdict: AttemptSlotProjectedEntry<Verdict>;\n    readonly score?: AttemptSlotProjectedEntry<ScoreProjection>;\n} | {\n    readonly state: \"not-projected\";\n    readonly reason: \"reporter-contract-does-not-provide-record-projections\";\n};\nconst NOT_PROJECTED_ASSESSMENT: BraintrustAssessment = Object.freeze({\n    state: \"not-projected\" as const,\n    reason: \"reporter-contract-does-not-provide-record-projections\" as const,\n});\n/**\n * 创建 Braintrust 报告器。挂在 `defineConfig({ reporters })` 上观测整次运行,\n * 或挂在单个 eval 的 `reporters` 上只观测它(同一实例被多个 eval 引用时共享一个实验)。\n */\nexport function Braintrust(config: BraintrustConfig = {}): Reporter {\n    let sdk: BraintrustSdk | undefined;\n    let experiment: BraintrustExperiment | undefined;\n    return {\n        async onInvocationStart(evals) {\n            sdk = await loadBraintrustSdk();\n            experiment = await sdk.init({\n                project: config.projectId ? undefined : (config.project ?? \"niceeval\"),\n                projectId: config.projectId,\n                experiment: config.experiment,\n                baseExperiment: config.baseExperiment,\n                baseExperimentId: config.baseExperimentId,\n                update: config.update,\n                apiKey: config.apiKey,\n                // 不再写顶层单一 agent:一次 Invocation 可能横跨多个 (agent, model, flags) 配置,\n                // 启动时还没有任何结果,写一个必然只代表其中一份配置的值就是撒谎(见\n                // docs/runner.md「Reporter 与运行器事件」)。每行自己的 agent 身份仍在\n                // toBraintrustEvent() 的 metadata.agent 里,逐 attempt 精确;跨行的 agent 集合可以\n                // 从 Braintrust 自己按 metadata.agent 分组得到,不需要实验级再存一份。\n                metadata: {\n                    evals: evals.map((e) => e.id),\n                    ...config.metadata,\n                },\n                // 不设为全局 current experiment,避免污染用户代码里的 braintrust 全局态。\n                setCurrent: false,\n            });\n        },\n        onEvalComplete(result) {\n            experiment?.log(toBraintrustEvent(result, NOT_PROJECTED_ASSESSMENT));\n        },\n        async onInvocationComplete() {\n            if (!experiment)\n                return;\n            try {\n                await sdk?.flush();\n                const summary = await experiment.summarize();\n                if (summary.experimentUrl) {\n                    reportActivity(`Braintrust experiment: ${summary.experimentUrl}`);\n                }\n            }\n            finally {\n                experiment = undefined;\n                sdk = undefined;\n            }\n        },\n    };\n}\n/**\n * EvalResult identity/transport data plus an optional post-Record assessment\n * become one Braintrust row. Assertions, Verdict, and Score only enter through\n * that assessment's public Attachment projections; this function never reads\n * a historical result graph.\n */\nexport function toBraintrustEvent(result: EvalResult, assessment: BraintrustAssessment = NOT_PROJECTED_ASSESSMENT): BraintrustLogEvent {\n    const projectedVerdict = assessment.state === \"projected\"\n        ? availableProjectionValue(assessment.verdict)\n        : undefined;\n    const verdict = projectedVerdict ?? result.verdict;\n    const metrics: globalThis.Record<string, number> = {};\n    if (result.startedAt) {\n        const start = Date.parse(result.startedAt) / 1000;\n        if (Number.isFinite(start)) {\n            metrics.start = start;\n            metrics.end = start + result.durationMs / 1000;\n        }\n    }\n    if (result.usage) {\n        const { inputTokens, outputTokens, cacheReadTokens, cacheCreationTokens, reasoningTokens, requests } = result.usage;\n        if (inputTokens !== undefined)\n            metrics.prompt_tokens = inputTokens;\n        if (outputTokens !== undefined)\n            metrics.completion_tokens = outputTokens;\n        if (inputTokens !== undefined && outputTokens !== undefined)\n            metrics.tokens = inputTokens + outputTokens;\n        if (cacheReadTokens !== undefined)\n            metrics.cache_read_tokens = cacheReadTokens;\n        if (cacheCreationTokens !== undefined)\n            metrics.cache_creation_tokens = cacheCreationTokens;\n        if (reasoningTokens !== undefined)\n            metrics.reasoning_tokens = reasoningTokens;\n        if (requests !== undefined)\n            metrics.requests = requests;\n    }\n    // estimated_cost_usd 恒为价目表估算(EvalResult.estimatedCostUSD,见 estimateCost);\n    // observed 成本单独留在 result.usage.costUSD,不混入本指标。\n    if (result.estimatedCostUSD !== undefined)\n        metrics.estimated_cost_usd = result.estimatedCostUSD;\n    const metadata: globalThis.Record<string, JsonValue> = {\n        eval: result.id,\n        agent: result.agent,\n        attempt: result.attempt,\n        verdict,\n        status: verdict,\n        evaluationAlgorithm: result.evaluationAlgorithm,\n        evaluationKind: result.evaluationKind,\n        assessmentProjection: assessment.state,\n    };\n    const scores: globalThis.Record<string, number> = {};\n    if (assessment.state === \"projected\") {\n        metadata.assertions = projectionMetadata(assessment.assertions, (assertions) => ({ entries: assertionMetadata(assertions.entries) }));\n        metadata.verdictProjection = projectionMetadata(assessment.verdict, (value) => ({ value }));\n        if (assessment.score !== undefined) {\n            metadata.score = projectionMetadata(assessment.score, scoreMetadata);\n            const score = availableProjectionValue(assessment.score);\n            if (score?.state === \"complete\")\n                scores[\"niceeval:score\"] = score.earned;\n        }\n    }\n    else {\n        metadata.assessmentProjectionReason = assessment.reason;\n    }\n    if (result.model !== undefined)\n        metadata.model = result.model;\n    if (result.experimentId !== undefined)\n        metadata.experiment = result.experimentId;\n    if (result.experiment?.flags && Object.keys(result.experiment.flags).length > 0) {\n        metadata.flags = result.experiment.flags;\n    }\n    if (result.skipReason !== undefined)\n        metadata.skipReason = result.skipReason;\n    // 一次运行内 (experiment, eval, agent, model, attempt) 唯一;Braintrust 按 id 合并重复行。\n    const id = [result.experimentId ?? \"\", result.id, result.agent, result.model ?? \"\", `a${result.attempt}`].join(\"|\");\n    return {\n        id,\n        input: result.description ?? result.id,\n        output: lastAssistantText(result.events),\n        error: result.error,\n        metadata,\n        metrics,\n        ...(Object.keys(scores).length > 0 ? { scores } : {}),\n    };\n}\nfunction availableProjectionValue<Value>(entry: AttemptSlotProjectedEntry<Value>): Value | undefined {\n    return entry.state === \"attachment-result\" && entry.attachment.state === \"available\"\n        ? entry.attachment.value\n        : undefined;\n}\nfunction projectionMetadata<Value>(entry: AttemptSlotProjectedEntry<Value>, available: (value: Value) => globalThis.Record<string, JsonValue>): JsonValue {\n    if (entry.state !== \"attachment-result\") {\n        return { state: entry.state };\n    }\n    switch (entry.attachment.state) {\n        case \"available\":\n            return { state: \"available\", ...available(entry.attachment.value) };\n        case \"unavailable\":\n            return { state: \"unavailable\" };\n        case \"migration-required\":\n            return {\n                state: \"migration-required\",\n                from: entry.attachment.from,\n                to: entry.attachment.to,\n                ...(entry.attachment.command === undefined ? {} : { command: entry.attachment.command }),\n            };\n        case \"unsupported\":\n        case \"invalid\":\n            return { state: entry.attachment.state };\n    }\n}\nfunction assertionMetadata(entries: readonly AssertionEntryRead<unknown>[]): JsonValue {\n    const metadata: JsonValue[] = [];\n    for (const entry of entries) {\n        const display = entry.entry.display;\n        const result = entry.entry.result;\n        const item: globalThis.Record<string, JsonValue> = {\n            entryId: entry.entry.entryId,\n            groupPath: [...display.groupPath],\n            outcome: entry.state === \"available\" ? result.state : entry.state,\n            coverage: coverageMetadata(entry.entry.coverage),\n            limitations: entry.entry.limitations.map(limitationMetadata),\n            scoreContribution: scoreContributionMetadata(result.score),\n        };\n        if (display.label !== undefined)\n            item.label = display.label;\n        if (display.key !== undefined)\n            item.key = display.key;\n        const reason = assertionReason(entry);\n        if (reason !== undefined)\n            item.reason = reason;\n        metadata.push(item);\n    }\n    return metadata;\n}\nfunction assertionReason(entry: AssertionEntryRead<unknown>): string | undefined {\n    if (entry.state !== \"available\")\n        return entry.reason;\n    switch (entry.entry.result.state) {\n        case \"matched\":\n            return undefined;\n        case \"mismatched\":\n        case \"unavailable\":\n        case \"errored\":\n        case \"not-applicable\":\n            return entry.entry.result.reason;\n    }\n}\nfunction coverageMetadata(coverage: AssertionCoverage): globalThis.Record<string, JsonValue> {\n    const metadata: globalThis.Record<string, JsonValue> = { state: coverage.state };\n    if (coverage.state !== \"complete\")\n        metadata.reason = coverage.reason;\n    return metadata;\n}\nfunction limitationMetadata(limitation: AssertionLimitation): globalThis.Record<string, JsonValue> {\n    switch (limitation.kind) {\n        case \"redacted\":\n            return { kind: limitation.kind, fieldCount: limitation.fieldCount };\n        case \"sampled\":\n            return {\n                kind: limitation.kind,\n                captured: limitation.captured,\n                ...(limitation.knownTotal === undefined ? {} : { knownTotal: limitation.knownTotal }),\n            };\n        case \"truncated\":\n            return { kind: limitation.kind, omittedBytes: limitation.omittedBytes };\n        case \"provider-limited\":\n            return { kind: limitation.kind };\n    }\n}\nfunction scoreContributionMetadata(contribution: ScoreContribution): globalThis.Record<string, JsonValue> {\n    switch (contribution.state) {\n        case \"not-scored\":\n            return { state: contribution.state };\n        case \"earned\":\n            return {\n                state: contribution.state,\n                points: contribution.points,\n                earned: contribution.earned,\n            };\n        case \"unavailable\":\n            return {\n                state: contribution.state,\n                points: contribution.points,\n                reason: contribution.reason,\n            };\n    }\n}\nfunction scoreMetadata(score: ScoreProjection): globalThis.Record<string, JsonValue> {\n    switch (score.state) {\n        case \"complete\":\n            return { state: score.state, earned: score.earned, comparable: score.comparable };\n        case \"partial\":\n            return {\n                state: score.state,\n                earned: score.earned,\n                reasons: [...score.reasons],\n                comparable: score.comparable,\n            };\n        case \"unavailable\":\n            return {\n                state: score.state,\n                reasons: [...score.reasons],\n                comparable: score.comparable,\n            };\n    }\n}\n/** agent 的最终回复文本(事件流里最后一条 assistant message);没有就不填。 */\nfunction lastAssistantText(events: EvalResult[\"events\"]): string | undefined {\n    if (!events)\n        return undefined;\n    for (let i = events.length - 1; i >= 0; i--) {\n        const e = events[i];\n        if (e.type === \"message\" && e.role === \"assistant\")\n            return e.text;\n    }\n    return undefined;\n}\n/** 动态 import 说明:specifier 放变量里,`braintrust` 不进编译期依赖,没装也能 typecheck。 */\nconst BRAINTRUST_PACKAGE = \"braintrust\";\nasync function loadBraintrustSdk(): Promise<BraintrustSdk> {\n    try {\n        return (await __niceevalDynamicImport(BRAINTRUST_PACKAGE)) as unknown as BraintrustSdk;\n    }\n    catch {\n        throw new Error(\"The 'braintrust' package is required for the Braintrust reporter but was not found. Install it with: npm install braintrust (and set BRAINTRUST_API_KEY).\");\n    }\n}\n"]}

package/dist/runner/reporters/braintrust.d.cts

@@ -1,7 +1,23 @@
 import type { EvalResult, JsonValue, Reporter } from "../../types.cjs";
-import type { AssertionsProjectionV1 } from "../../assertions/record/model.cjs";
-import type { ScoreProjectionV1 } from "../../eval/record/score.cjs";
-import type { AttemptSlotProjectedEntry } from "../../projection/index.cjs";
+import type { AssertionsProjection } from "../../assertions/record/model.cjs";
+import type { ScoreProjection } from "../../eval/record/score.cjs";
+/** Closed assessment rows; this reporter does not reopen Record. */
+type AttemptSlotProjectedEntry<Value> = {
+    readonly state: "excluded" | "not-recorded" | "core-invalid";
+} | {
+    readonly state: "attachment-result";
+    readonly attachment: {
+        readonly state: "available";
+        readonly value: Value;
+    } | {
+        readonly state: "unavailable" | "unsupported" | "invalid";
+    } | {
+        readonly state: "migration-required";
+        readonly from: string;
+        readonly to: string;
+        readonly command?: string;
+    };
+};
 import type { Verdict } from "../../shared/types.cjs";
 export interface BraintrustConfig {
     /** Braintrust 项目名;省略时用 "niceeval"。 */
@@ -38,9 +54,9 @@ export interface BraintrustLogEvent {
 export type BraintrustAssessment = {
     readonly state: "projected";
     /** Exact Report-style public projections, including every data state. */
-    readonly assertions: AttemptSlotProjectedEntry<AssertionsProjectionV1<unknown>>;
+    readonly assertions: AttemptSlotProjectedEntry<AssertionsProjection<unknown>>;
     readonly verdict: AttemptSlotProjectedEntry<Verdict>;
-    readonly score?: AttemptSlotProjectedEntry<ScoreProjectionV1>;
+    readonly score?: AttemptSlotProjectedEntry<ScoreProjection>;
 } | {
     readonly state: "not-projected";
     readonly reason: "reporter-contract-does-not-provide-record-projections";
@@ -57,3 +73,4 @@ export declare function Braintrust(config?: BraintrustConfig): Reporter;
  * a historical result graph.
  */
 export declare function toBraintrustEvent(result: EvalResult, assessment?: BraintrustAssessment): BraintrustLogEvent;
+export {};