@runtypelabs/sdk 6.2.0 → 6.3.0

package/dist/index.d.ts

@@ -3086,7 +3086,7 @@ interface paths {
         };
         /**
          * Stream execution events
-         * @description Reconnect to a durable Claude Managed agent turn: replay-and-tail its Server-Sent Events. Resolves the per-conversation session owner by (agentId, conversationId), replays buffered events strictly past the `after` cursor, then live-tails if the turn is still running. Each event carries an SSE `id:` line — a durable row seq, or (unified vocabulary) a composite `<seq>.<subIndex>` when one durable row translates to multiple events. Reconnect after a disconnect (tab reload, sleep, stream timeout) by passing the last seen id verbatim as `after` to resume without missing or duplicating events. The vocabulary follows the same negotiation as the execution stream (`?events=`, `X-Persona-Version`, then the default-sse-event-format flag).
+         * @description Reconnect to a durable Claude Managed agent turn: replay-and-tail its Server-Sent Events. Resolves the per-conversation session owner by (agentId, conversationId), replays buffered events strictly past the `after` cursor, then live-tails if the turn is still running. Each event carries an SSE `id:` line — a durable row seq, or (unified vocabulary) a composite `<seq>.<subIndex>` when one durable row translates to multiple events. Reconnect after a disconnect (tab reload, sleep, stream timeout) by passing the last seen id verbatim as `after` to resume without missing or duplicating events. Replayed and live-tailed frames use the unified execution vocabulary.
          */
         get: {
             parameters: {
@@ -11072,6 +11072,16 @@ interface paths {
                                         content: string;
                                         role: string;
                                     }[];
+                                    toolMocks?: {
+                                        input?: unknown;
+                                        isError?: boolean;
+                                        output: null | boolean | number | string | unknown[] | {
+                                            [key: string]: unknown;
+                                        };
+                                        sourceToolExecutionId?: string;
+                                        toolName: string;
+                                        truncated?: boolean;
+                                    }[];
                                     variables?: {
                                         [key: string]: unknown;
                                     };
@@ -11080,7 +11090,7 @@ interface paths {
                             }[];
                             name: string;
                             /**
-                             * @description How captured (saved_from_run) cases replay: "next_step" (default) or "continue" (serve recorded toolMocks and run past the first action).
+                             * @description How cases carrying toolMocks replay — captured (saved_from_run) or hand-authored alike: "next_step" (default; captured cases only) captures the first tool intent and stops, or "continue" serves each case's toolMocks by tool name and runs past the first action.
                              * @enum {string}
                              */
                             recordedToolMode?: "next_step" | "continue";
@@ -11529,6 +11539,16 @@ interface paths {
                                         content: string;
                                         role: string;
                                     }[];
+                                    toolMocks?: {
+                                        input?: unknown;
+                                        isError?: boolean;
+                                        output: null | boolean | number | string | unknown[] | {
+                                            [key: string]: unknown;
+                                        };
+                                        sourceToolExecutionId?: string;
+                                        toolName: string;
+                                        truncated?: boolean;
+                                    }[];
                                     variables?: {
                                         [key: string]: unknown;
                                     };
@@ -11537,7 +11557,7 @@ interface paths {
                             }[];
                             name: string;
                             /**
-                             * @description How captured (saved_from_run) cases replay: "next_step" (default) or "continue" (serve recorded toolMocks and run past the first action).
+                             * @description How cases carrying toolMocks replay — captured (saved_from_run) or hand-authored alike: "next_step" (default; captured cases only) captures the first tool intent and stops, or "continue" serves each case's toolMocks by tool name and runs past the first action.
                              * @enum {string}
                              */
                             recordedToolMode?: "next_step" | "continue";
@@ -12097,6 +12117,12 @@ interface paths {
                                 json?: unknown;
                                 text?: string;
                             };
+                            /** @description Per-case graders. When set, they REPLACE the suite-level graders for this case (not appended). Omit to score the case with the suite graders. */
+                            graders?: ({
+                                kind: string;
+                            } & {
+                                [key: string]: unknown;
+                            })[];
                             /** @default {} */
                             input?: {
                                 checkpoint?: {
@@ -12121,7 +12147,9 @@ interface paths {
                                 toolMocks?: {
                                     input?: unknown;
                                     isError?: boolean;
-                                    output?: unknown;
+                                    output: null | boolean | number | string | unknown[] | {
+                                        [key: string]: unknown;
+                                    };
                                     sourceToolExecutionId?: string;
                                     toolName: string;
                                     truncated?: boolean;
@@ -12495,6 +12523,12 @@ interface paths {
                                 json?: unknown;
                                 text?: string;
                             };
+                            /** @description Per-case graders. When set, they REPLACE the suite-level graders for this case (not appended). Omit to score the case with the suite graders. */
+                            graders?: ({
+                                kind: string;
+                            } & {
+                                [key: string]: unknown;
+                            })[];
                             /** @default {} */
                             input?: {
                                 checkpoint?: {
@@ -12519,7 +12553,9 @@ interface paths {
                                 toolMocks?: {
                                     input?: unknown;
                                     isError?: boolean;
-                                    output?: unknown;
+                                    output: null | boolean | number | string | unknown[] | {
+                                        [key: string]: unknown;
+                                    };
                                     sourceToolExecutionId?: string;
                                     toolName: string;
                                     truncated?: boolean;
@@ -12806,6 +12842,12 @@ interface paths {
                             json?: unknown;
                             text?: string;
                         } | null;
+                        /** @description Per-case graders that REPLACE the suite graders for this case. null clears the override so the case uses the suite graders again. */
+                        graders?: ({
+                            kind: string;
+                        } & {
+                            [key: string]: unknown;
+                        })[] | null;
                         input?: {
                             checkpoint?: {
                                 forkMessageIndex: number;
@@ -12829,7 +12871,9 @@ interface paths {
                             toolMocks?: {
                                 input?: unknown;
                                 isError?: boolean;
-                                output?: unknown;
+                                output: null | boolean | number | string | unknown[] | {
+                                    [key: string]: unknown;
+                                };
                                 sourceToolExecutionId?: string;
                                 toolName: string;
                                 truncated?: boolean;
@@ -13256,7 +13300,9 @@ interface paths {
                                 toolMocks?: {
                                     input?: unknown;
                                     isError?: boolean;
-                                    output?: unknown;
+                                    output: null | boolean | number | string | unknown[] | {
+                                        [key: string]: unknown;
+                                    };
                                     sourceToolExecutionId?: string;
                                     toolName: string;
                                     truncated?: boolean;
@@ -26005,6 +26051,8 @@ interface paths {
                             data: {
                                 behavior?: unknown;
                                 createdAt: string;
+                                /** @description Canonical public URL of this surface (MCP, API, A2A, AG-UI, webhook, chat). Null for surface types reached through their own channels (Slack, email, schedule, ...). Clients should use this instead of constructing the URL themselves. */
+                                endpoint: string | null;
                                 environment: string;
                                 id: string;
                                 inbound?: unknown;
@@ -26114,6 +26162,8 @@ interface paths {
                         "application/json": {
                             behavior?: unknown;
                             createdAt: string;
+                            /** @description Canonical public URL of this surface (MCP, API, A2A, AG-UI, webhook, chat). Null for surface types reached through their own channels (Slack, email, schedule, ...). Clients should use this instead of constructing the URL themselves. */
+                            endpoint: string | null;
                             environment: string;
                             id: string;
                             inbound?: unknown;
@@ -26443,6 +26493,8 @@ interface paths {
                         "application/json": {
                             behavior?: unknown;
                             createdAt: string;
+                            /** @description Canonical public URL of this surface (MCP, API, A2A, AG-UI, webhook, chat). Null for surface types reached through their own channels (Slack, email, schedule, ...). Clients should use this instead of constructing the URL themselves. */
+                            endpoint: string | null;
                             environment: string;
                             id: string;
                             inbound?: unknown;
@@ -26580,6 +26632,8 @@ interface paths {
                         "application/json": {
                             behavior?: unknown;
                             createdAt: string;
+                            /** @description Canonical public URL of this surface (MCP, API, A2A, AG-UI, webhook, chat). Null for surface types reached through their own channels (Slack, email, schedule, ...). Clients should use this instead of constructing the URL themselves. */
+                            endpoint: string | null;
                             environment: string;
                             id: string;
                             inbound?: unknown;
@@ -39006,6 +39060,8 @@ interface paths {
                         "application/json": {
                             data: {
                                 createdAt: string;
+                                /** @description Canonical public URL of this surface (MCP, API, A2A, AG-UI, webhook, chat). Null for surface types reached through their own channels (Slack, email, schedule, ...). Clients should use this instead of constructing the URL themselves. */
+                                endpoint: string | null;
                                 environment: string;
                                 id: string;
                                 inbound: {
@@ -40672,7 +40728,7 @@ interface paths {
         put?: never;
         /**
          * Execute tool
-         * @description Execute a tool by ID with the provided parameters. Non-flow tools return the documented JSON result. Flow tools return a text/event-stream response whose frames use the ExecutionStreamEvent vocabulary by default; use a raw/streaming client helper for flow tools.
+         * @description Execute a tool by ID with the provided parameters. Non-flow tools return the documented JSON result. Flow tools return a text/event-stream response whose frames use the ExecutionStreamEvent vocabulary; use a raw/streaming client helper for flow tools.
          */
         post: {
             parameters: {
@@ -41165,6 +41221,7 @@ interface paths {
                             };
                             id: string;
                             orgId?: string | null;
+                            orgName: string | null;
                             preferences: {
                                 notifications: boolean;
                                 theme: string;
@@ -42039,6 +42096,8 @@ interface components {
             createdAt: string;
             enabled: boolean;
             expected?: unknown;
+            /** @description Per-case graders that REPLACE the suite graders for this case, or null when the case uses the suite graders. */
+            graders: unknown[] | null;
             id: string;
             /** @description { variables?, messages? } — what the case feeds the target. */
             input?: unknown;
@@ -42152,6 +42211,16 @@ interface components {
                             content: string;
                             role: string;
                         }[];
+                        toolMocks?: {
+                            input?: unknown;
+                            isError?: boolean;
+                            output: null | boolean | number | string | unknown[] | {
+                                [key: string]: unknown;
+                            };
+                            sourceToolExecutionId?: string;
+                            toolName: string;
+                            truncated?: boolean;
+                        }[];
                         variables?: {
                             [key: string]: unknown;
                         };
@@ -42160,7 +42229,7 @@ interface components {
                 }[];
                 name: string;
                 /**
-                 * @description How captured (saved_from_run) cases replay: "next_step" (default) or "continue" (serve recorded toolMocks and run past the first action).
+                 * @description How cases carrying toolMocks replay — captured (saved_from_run) or hand-authored alike: "next_step" (default; captured cases only) captures the first tool intent and stops, or "continue" serves each case's toolMocks by tool name and runs past the first action.
                  * @enum {string}
                  */
                 recordedToolMode?: "next_step" | "continue";
@@ -42368,6 +42437,10 @@ interface components {
             completedAt?: string;
             durationMs?: number;
             executionId: string;
+            externalAgent?: {
+                contextId?: string;
+                taskId?: string;
+            };
             failedSteps?: number;
             finalOutput?: string;
             iterations?: number;
@@ -42698,6 +42771,10 @@ interface components {
             approvalId: string;
             description?: string;
             executionId: string;
+            externalAgent?: {
+                contextId?: string;
+                taskId?: string;
+            };
             iteration?: number;
             parameters?: {
                 [key: string]: unknown;
@@ -42841,382 +42918,6 @@ interface components {
             /** @description Present only when a raw tool_… reference could not be emitted as a portable tool:<name> (the tool was deleted, or its name is shadowed by an older same-named tool). Those references are left as raw ids, which the ensure surface rejects; each warning explains how to fix it. */
             warnings?: string[];
         };
-        FlowSSEEvent: {
-            executionContext?: {
-                [key: string]: unknown;
-            };
-            executionId?: string;
-            flowId: string;
-            flowName?: string;
-            input?: unknown;
-            seq?: number;
-            source?: string;
-            startedAt: string;
-            toolContext?: {
-                executionId: string;
-                stepId: string;
-                toolId: string;
-            };
-            totalSteps?: number;
-            /** @enum {string} */
-            type: "flow_start";
-        } | {
-            claudeManagedAgentId?: string;
-            completedAt?: string;
-            completedSteps?: number;
-            duration?: number;
-            executionContext?: {
-                [key: string]: unknown;
-            };
-            executionId?: string;
-            executionTime?: number;
-            failedSteps?: number;
-            finalOutput?: string;
-            flowId?: string;
-            flowName?: string;
-            output?: unknown;
-            seq?: number;
-            source?: string;
-            success?: boolean;
-            successfulSteps?: number;
-            toolContext?: {
-                executionId: string;
-                stepId: string;
-                toolId: string;
-            };
-            totalSteps?: number;
-            totalTokensUsed?: number;
-            /** @enum {string} */
-            type: "flow_complete";
-        } | {
-            blockReason?: string;
-            code?: string;
-            error: string | {
-                code: string;
-                message: string;
-                stepId?: string;
-                stepType?: string;
-            };
-            executionId?: string;
-            executionTime?: number;
-            flowId?: string;
-            seq?: number;
-            timestamp?: string;
-            toolContext?: {
-                executionId: string;
-                stepId: string;
-                toolId: string;
-            };
-            /** @enum {string} */
-            type: "flow_error";
-            upgradeUrl?: string;
-        } | {
-            approvalId?: string;
-            awaitReason?: string;
-            awaitedAt: string;
-            crawlId?: string;
-            elicitation?: {
-                message: string;
-                /** @enum {string} */
-                mode: "form" | "url";
-                pauseCount?: number;
-                requestedSchema?: {
-                    [key: string]: unknown;
-                };
-                serverName?: string;
-                url?: string;
-            };
-            executionId?: string;
-            flowId: string;
-            /** @enum {string} */
-            origin?: "webmcp" | "sdk";
-            pageOrigin?: string;
-            parameters?: {
-                [key: string]: unknown;
-            };
-            seq?: number;
-            stepId?: string;
-            timeout?: number;
-            toolCallId?: string;
-            toolId?: string;
-            toolName?: string;
-            /** @enum {string} */
-            type: "flow_await";
-        } | {
-            estimatedTokens?: number;
-            executionId?: string;
-            id?: string;
-            index?: number;
-            name?: string;
-            outputVariable?: string;
-            seq?: number;
-            startedAt: string;
-            stepId?: string;
-            stepName?: string;
-            stepType?: string;
-            toolContext?: {
-                executionId: string;
-                stepId: string;
-                toolId: string;
-            };
-            totalSteps?: number;
-            /** @enum {string} */
-            type: "step_start";
-        } | {
-            delta?: string;
-            executionId?: string;
-            id?: string;
-            messageId?: string;
-            partId?: string;
-            seq?: number;
-            text?: string;
-            toolContext?: {
-                executionId: string;
-                stepId: string;
-                toolId: string;
-            };
-            toolId?: string;
-            /** @enum {string} */
-            type: "step_delta";
-        } | {
-            completedAt?: string;
-            duration?: number;
-            durationMs?: number;
-            error?: string | null;
-            executionId?: string;
-            executionTime?: number;
-            id?: string;
-            index?: number;
-            name?: string;
-            output?: unknown;
-            result?: unknown;
-            seq?: number;
-            stepId?: string;
-            stepName?: string;
-            stepType?: string;
-            /** @enum {string} */
-            stopReason?: "end_turn" | "max_tool_calls" | "length" | "content_filter" | "error" | "unknown";
-            success?: boolean;
-            tokensUsed?: number;
-            toolContext?: {
-                executionId: string;
-                stepId: string;
-                toolId: string;
-            };
-            /** @enum {string} */
-            type: "step_complete";
-            unresolvedVariables?: string[];
-        } | {
-            error: string;
-            executionId?: string;
-            executionTime?: number;
-            id?: string;
-            index?: number;
-            name?: string;
-            seq?: number;
-            stepType?: string;
-            /** @enum {string} */
-            type: "step_error";
-        } | {
-            error?: string;
-            executionId?: string;
-            id: string;
-            index?: number;
-            name?: string;
-            seq?: number;
-            skippedAt: string;
-            stepType: string;
-            totalSteps: number;
-            /** @enum {string} */
-            type: "step_skip";
-            when: string;
-        } | ({
-            elicitation?: {
-                message: string;
-                /** @enum {string} */
-                mode: "form" | "url";
-                pauseCount?: number;
-                requestedSchema?: {
-                    [key: string]: unknown;
-                };
-                serverName?: string;
-                url?: string;
-            };
-            executionId?: string;
-            reason?: string;
-            seq?: number;
-            /** @enum {string} */
-            type: "step_await";
-        } & {
-            [key: string]: unknown;
-        }) | ({
-            agentContext?: {
-                executionId: string;
-                iteration: number;
-                seq: number;
-            };
-            executionId?: string;
-            hiddenParameterNames?: string[];
-            name?: string;
-            parameters?: {
-                [key: string]: unknown;
-            };
-            providerOptions?: {
-                [key: string]: unknown;
-            };
-            seq?: number;
-            startedAt?: string;
-            stepId?: string;
-            toolCallId?: string;
-            toolId?: string;
-            toolName?: string;
-            /** @enum {string} */
-            toolType: "flow" | "mcp" | "builtin" | "custom" | "external" | "advisor" | "subagent" | "local";
-            /** @enum {string} */
-            type: "tool_start";
-        } & {
-            [key: string]: unknown;
-        }) | ({
-            delta?: string;
-            executionId?: string;
-            seq?: number;
-            toolId?: string;
-            /** @enum {string} */
-            type: "tool_delta";
-        } & {
-            [key: string]: unknown;
-        }) | {
-            delta: string;
-            executionId?: string;
-            seq?: number;
-            stepId?: string;
-            toolCallId?: string;
-            toolId?: string;
-            /** @enum {string} */
-            type: "tool_input_delta";
-        } | {
-            executionId?: string;
-            hiddenParameterNames?: string[];
-            parameters: {
-                [key: string]: unknown;
-            };
-            providerOptions?: {
-                [key: string]: unknown;
-            };
-            seq?: number;
-            stepId?: string;
-            toolCallId?: string;
-            toolId?: string;
-            toolName?: string;
-            /** @enum {string} */
-            type: "tool_input_complete";
-        } | {
-            agentContext?: {
-                executionId: string;
-                iteration: number;
-                seq: number;
-            };
-            completedAt?: string;
-            error?: string;
-            executionId?: string;
-            executionTime?: number;
-            name?: string;
-            result?: unknown;
-            seq?: number;
-            stepId?: string;
-            success: boolean;
-            toolCallId?: string;
-            toolCost?: number;
-            toolId?: string;
-            toolName?: string;
-            /** @enum {string} */
-            type: "tool_complete";
-        } | {
-            agentContext?: {
-                executionId: string;
-                iteration: number;
-                seq: number;
-            };
-            error: string;
-            executionId?: string;
-            executionTime?: number;
-            failedAt?: string;
-            name: string;
-            seq?: number;
-            toolId: string;
-            /** @enum {string} */
-            type: "tool_error";
-        } | {
-            executionId?: string;
-            id: string;
-            seq?: number;
-            text: string;
-            /** @enum {string} */
-            type: "chunk";
-        } | ({
-            executionId?: string;
-            seq?: number;
-            /** @enum {string} */
-            type: "text_start";
-        } & {
-            [key: string]: unknown;
-        }) | ({
-            executionId?: string;
-            seq?: number;
-            /** @enum {string} */
-            type: "text_end";
-        } & {
-            [key: string]: unknown;
-        }) | ({
-            executionId?: string;
-            seq?: number;
-            /** @enum {string} */
-            type: "reason_start";
-        } & {
-            [key: string]: unknown;
-        }) | ({
-            executionId?: string;
-            seq?: number;
-            /** @enum {string} */
-            type: "reason_delta";
-        } & {
-            [key: string]: unknown;
-        }) | ({
-            executionId?: string;
-            seq?: number;
-            /** @enum {string} */
-            type: "reason_complete";
-        } & {
-            [key: string]: unknown;
-        }) | ({
-            executionId?: string;
-            seq?: number;
-            /** @enum {string} */
-            type: "source";
-        } & {
-            [key: string]: unknown;
-        }) | ({
-            executionId?: string;
-            seq?: number;
-            /** @enum {string} */
-            type: "fallback_start";
-        } & {
-            [key: string]: unknown;
-        }) | ({
-            executionId?: string;
-            seq?: number;
-            /** @enum {string} */
-            type: "fallback_complete";
-        } & {
-            [key: string]: unknown;
-        }) | ({
-            executionId?: string;
-            seq?: number;
-            /** @enum {string} */
-            type: "fallback_exhausted";
-        } & {
-            [key: string]: unknown;
-        });
         FlowValidationIssue: {
             code: string;
             details?: unknown;
@@ -43744,10 +43445,12 @@ interface components {
  * the rest of the SDK (and `types.ts`) re-exports, so the published surface
  * never exposes `paths['/v1/...'][...]` indexing.
  *
- * Two flavors live here:
- *  1. Streaming event unions — sourced from named `components.schemas` (the
- *     OAS 3.2 SSE itemSchema unions). Clean `$ref` → named type.
- *  2. Entity types — sourced by indexing into `paths`, because entities are
+ * Three flavors live here:
+ *  1. The unified streaming wire union — sourced from the named OAS 3.2
+ *     `ExecutionStreamEvent` itemSchema component.
+ *  2. The compatibility-only FlowBuilder callback union — hand-maintained
+ *     because it is a client presentation shape, not a server wire schema.
+ *  3. Entity types — sourced by indexing into `paths`, because entities are
  *     inline in the spec (no named components, no operationIds). The `Flow`
  *     alias below is the REFERENCE PATTERN the entity-codegen workflow clones.
  */
@@ -43762,12 +43465,121 @@ type ExecutionStreamEvent = components['schemas']['ExecutionStreamEvent'];
  */
 type DispatchEvent = components['schemas']['ExecutionStreamEvent'];
 /**
- * Legacy flow-execution event union. Kept for the SDK's FlowBuilder convenience
- * types (`StreamEventOf<FlowStreamEvent, 'flow_start' | 'step_delta' | …>`),
- * which use the historical event names. The wire is the unified stream;
- * {@link ExecutionStreamEvent} is the current shape.
- */
-type FlowStreamEvent = components['schemas']['FlowSSEEvent'];
+ * SDK presentation shape for FlowBuilder callbacks.
+ *
+ * This is deliberately hand-written instead of sourced from OpenAPI: the public
+ * wire has one vocabulary ({@link ExecutionStreamEvent}), while the SDK
+ * reverse-translates those frames to its long-standing callback names for
+ * compatibility. These types describe that client-side presentation layer, not
+ * an alternate server wire contract.
+ */
+type FlowStreamEvent = (LegacyFlowEventBase & {
+    type: 'flow_start';
+    flowId: string;
+    flowName?: string;
+    totalSteps?: number;
+    startedAt?: string;
+    input?: unknown;
+}) | (LegacyFlowEventBase & {
+    type: 'flow_complete';
+    flowId?: string;
+    flowName?: string;
+    totalSteps?: number;
+    successfulSteps?: number;
+    failedSteps?: number;
+    skippedSteps?: number;
+    executionTime?: number;
+    duration?: number;
+    durationMs?: number;
+    success?: boolean;
+    finalOutput?: unknown;
+    completedAt?: string;
+}) | (LegacyFlowEventBase & {
+    type: 'flow_error';
+    flowId?: string;
+    stepId?: string;
+    error?: string | Record<string, unknown> | null;
+    code?: string;
+    timestamp?: string;
+}) | (LegacyFlowEventBase & {
+    type: 'flow_await';
+    flowId?: string;
+    stepId?: string;
+    toolId?: string;
+    toolCallId?: string;
+    toolName?: string;
+    awaitedAt?: string;
+    awaitReason?: string;
+    status?: string;
+}) | (LegacyFlowEventBase & {
+    type: 'step_start';
+    id?: string;
+    stepId?: string;
+    name?: string;
+    stepName?: string;
+    index?: number;
+    stepType?: string;
+    input?: unknown;
+    startedAt?: string;
+}) | (LegacyFlowEventBase & {
+    type: 'step_delta';
+    id?: string;
+    stepId?: string;
+    index?: number;
+    text?: string;
+    delta?: string;
+    partId?: string;
+}) | (LegacyFlowEventBase & {
+    type: 'step_complete';
+    id?: string;
+    stepId?: string;
+    name?: string;
+    stepName?: string;
+    index?: number;
+    stepType?: string;
+    result?: unknown;
+    executionTime?: number;
+    duration?: number;
+    durationMs?: number;
+    success?: boolean;
+    stopReason?: string;
+    completedAt?: string;
+}) | (LegacyFlowEventBase & {
+    type: 'step_error';
+    id?: string;
+    stepId?: string;
+    name?: string;
+    stepName?: string;
+    error?: string | Record<string, unknown> | null;
+    success?: false;
+}) | (LegacyFlowEventBase & {
+    type: 'step_await';
+    id?: string;
+    stepId?: string;
+    name?: string;
+    stepName?: string;
+    index?: number;
+    stepType?: string;
+    toolId?: string;
+    toolCallId?: string;
+    toolName?: string;
+    parameters?: unknown;
+}) | LegacyFlowNamedEvent<'step_skip' | 'tool_start' | 'tool_delta' | 'tool_input_delta' | 'tool_input_complete' | 'tool_complete' | 'tool_error' | 'chunk' | 'text_start' | 'text_end' | 'reason_start' | 'reason_delta' | 'reason_complete' | 'source' | 'fallback_start' | 'fallback_complete' | 'fallback_exhausted'>;
+type LegacyFlowNamedEvent<T extends string> = T extends unknown ? LegacyFlowEventBase & {
+    type: T;
+} : never;
+interface LegacyFlowEventBase {
+    executionContext?: Record<string, unknown>;
+    executionId?: string;
+    seq?: number;
+    source?: string;
+    toolContext?: {
+        executionId: string;
+        stepId: string;
+        toolId: string;
+    };
+    [key: string]: unknown;
+}
 /**
  * @deprecated Alias of {@link ExecutionStreamEvent}. Prefer `ExecutionStreamEvent`.
  */
@@ -43790,20 +43602,11 @@ type StreamEventOf<U, T extends string> = Extract<U, {
 /**
  * Options for the flow stream consumers.
  *
- * `unified` controls translation of the SSE branch from the unified 33-event
- * wire vocabulary back to the legacy flow event shapes:
- *   - `undefined` (default): AUTO-DETECT — translate iff the stream is detected
- *     as unified (an execution stream always leads with `execution_start`).
- *     This makes the public `streamEvents(response)` / `processStream(response,
- *     cb)` pattern correct on a `dispatch()` response without the caller knowing
- *     the wire format, and keeps a still-legacy stream (the eval stream)
- *     flowing through untouched.
- *   - `true`: force translation (internal callers that KNOW the route is unified).
- *   - `false`: force legacy passthrough (never translate).
- * The non-streaming JSON branch always stays legacy — the api only translates
- * the SSE writer, never the JSON `events` array.
+ * Kept for source compatibility with older SDK consumers. Unified SSE
+ * translation is unconditional; `unified` is ignored.
  */
 interface StreamConsumeOptions {
+    /** @deprecated The API wire is always unified; this option is ignored. */
     unified?: boolean;
 }
 /**
@@ -43840,7 +43643,7 @@ declare function parseFinalBuffer(buffer: string): string | null;
  * })
  * ```
  */
-declare function processStream(response: Response, callbacks?: StreamCallbacks, options?: StreamConsumeOptions): Promise<FlowSummary>;
+declare function processStream(response: Response, callbacks?: StreamCallbacks, _options?: StreamConsumeOptions): Promise<FlowSummary>;
 /**
  * Create an async iterator over SSE events
  *
@@ -43855,7 +43658,7 @@ declare function processStream(response: Response, callbacks?: StreamCallbacks,
  * }
  * ```
  */
-declare function streamEvents(response: Response, options?: StreamConsumeOptions): AsyncGenerator<StreamEvent>;
+declare function streamEvents(response: Response, _options?: StreamConsumeOptions): AsyncGenerator<StreamEvent>;
 
 /**
  * FlowResult - Wrapper for streaming flow execution responses
@@ -43893,13 +43696,11 @@ declare class FlowResult {
     private response;
     private consumed;
     private cachedSummary;
-    private streamOptions;
     /**
-     * @param options.unified - The response is a unified-vocabulary SSE stream
-     *   (`/dispatch` since the unified-SSE cutover). Set by dispatch call sites;
-     *   left unset by the still-legacy eval stream producer.
+     * @param _options Deprecated stream-format options retained for source
+     *   compatibility. The API wire is always unified.
      */
-    constructor(response: Response, summary?: FlowSummary, options?: StreamConsumeOptions);
+    constructor(response: Response, summary?: FlowSummary, _options?: StreamConsumeOptions);
     /**
      * Get the raw Response object for manual handling
      *
@@ -44456,15 +44257,10 @@ interface Message$1 {
 /**
  * Union type of all possible flow-execution SSE events.
  *
- * Sourced from the GENERATED OpenAPI spec (`components.schemas.FlowSSEEvent`,
- * a 26-member discriminated union on `type`) rather than hand-mirrored
- * interfaces. The per-event member types below are back-compat aliases that
- * narrow this union via its `type` discriminant. The generated union is a
- * superset of the old hand-written shapes: some fields are optional or
- * field-name-aliased (`id`/`stepId`, `name`/`stepName`, `text`/`delta`,
- * `duration`/`durationMs`/`executionTime`), and `flow_error.error` may be a
- * string OR a structured `{ code, message, ... }` object — consumers must
- * read these fields defensively.
+ * This compatibility-only presentation union is produced by the SDK's unified
+ * wire adapter. It preserves the long-standing FlowBuilder callback names and
+ * field aliases without exposing a second server vocabulary. The per-event
+ * aliases below narrow it via the `type` discriminant.
  */
 type StreamEvent = FlowStreamEvent;
 type FlowStartEvent = StreamEventOf<FlowStreamEvent, 'flow_start'>;
@@ -46446,16 +46242,19 @@ interface EvalCaseCheckpoint {
     forkMessageIndex: number;
 }
 /**
- * One recorded tool result attached to a captured case (mirror of `ToolMock` in
- * `@runtypelabs/shared`). A future next-step replay serves these by tool name;
- * editing `output` turns the case into a counterfactual.
+ * One tool result a `recordedToolMode: 'continue'` suite serves by tool name
+ * instead of dispatching the real tool (mirror of `ToolMock` in
+ * `@runtypelabs/shared`). Either captured from a real run or hand-authored in
+ * `defineEval` directly — a `toolType: 'local'` tool otherwise pauses the run
+ * on its first call, so a mock is how an eval case gets an agent past it.
  */
 interface EvalToolMock {
     toolName: string;
     input?: unknown;
     output: unknown;
-    /** True when the recorded call failed: `output` is its error payload and a replay serves it as a failure. */
+    /** True when the call should replay as a FAILURE: `output` is its error payload. */
     isError?: boolean;
+    /** Present when captured: the `tool_executions.id` this mock came from. */
     sourceToolExecutionId?: string;
     /** True when `output` was capped at capture (the case is then non-fully-replayable). */
     truncated?: boolean;
@@ -46464,9 +46263,19 @@ interface EvalToolMock {
 interface EvalCaseInput {
     variables?: Record<string, unknown>;
     messages?: EvalMessage[];
-    /** Present on `saved_from_run` cases: where the run was forked. */
+    /**
+     * Present on `saved_from_run` cases: where the run was forked. Execution
+     * state, not something `defineEval` accepts as authored input — it round-
+     * trips through `client.evals.pull()` but is dropped if you feed a pulled
+     * case straight back into `defineEval` (the server rejects it too).
+     */
     checkpoint?: EvalCaseCheckpoint;
-    /** Present on `saved_from_run` cases: recorded tool results, for replay/counterfactuals. */
+    /**
+     * Tool results served instead of dispatching the real tool, when the
+     * suite's `recordedToolMode` is `'continue'`. Present on `saved_from_run`
+     * cases (recorded from the fork) — and, as of round-2 item 7, also
+     * writable directly on a hand-authored case with no `checkpoint` at all.
+     */
     toolMocks?: EvalToolMock[];
 }
 /** The target a suite evaluates — a saved flow or agent, by portable name. */
@@ -46580,7 +46389,8 @@ interface EvalPullResult {
 }
 /**
  * One grader's verdict for one case (mirror of `@runtypelabs/shared`'s
- * `GraderOutcome`). `graderIndex` is the position in the suite's grader list.
+ * `GraderOutcome`). `graderIndex` is the index into the case's effective grader
+ * array (`case.graders ?? suite.graders`).
  */
 interface GraderOutcome {
     graderIndex: number;
@@ -47491,6 +47301,11 @@ interface EvalSuiteCase {
     name: string;
     input: EvalCaseInput;
     expected: CaseExpected | null;
+    /**
+     * Per-case graders that REPLACE the suite graders for this case, or null when
+     * the case is scored with the suite's `graders`.
+     */
+    graders: GraderConfig[] | null;
     notes: string | null;
     origin: string;
     enabled: boolean;
@@ -47518,6 +47333,11 @@ interface EvalSuiteCaseInput {
     name: string;
     input?: EvalCaseInput;
     expected?: CaseExpected;
+    /**
+     * Per-case graders. When set, they REPLACE the suite graders for this case
+     * (not appended). Omit to score the case with the suite graders.
+     */
+    graders?: GraderConfig[];
     notes?: string;
     enabled?: boolean;
     /** Case provenance. Defaults to 'manual' server-side when omitted. */
@@ -47606,6 +47426,12 @@ interface UpdateEvalCaseInput {
     name?: string;
     input?: EvalCaseInput;
     expected?: CaseExpected | null;
+    /**
+     * Per-case graders that REPLACE the suite graders for this case. `null` clears
+     * the override (the case scores with the suite graders again); `undefined`
+     * leaves it unchanged.
+     */
+    graders?: GraderConfig[] | null;
     notes?: string | null;
     enabled?: boolean;
 }
@@ -51773,10 +51599,10 @@ interface ParsedSSEEvent {
 interface AgentStreamCallbacks {
     onAgentStart?: (event: AgentStartEvent) => void;
     /**
-     * @deprecated No longer emitted on the unified wire vocabulary (the default
-     * since the unified-SSE cutover): iteration boundaries are folded into the
+     * @deprecated Not represented on the public unified wire: iteration
+     * boundaries are folded into the
      * `iteration` field on `onTurnStart` / `onTurnComplete` / tool events. Read
-     * that field instead. Still fires on a legacy stream during the rollout.
+     * that field instead. Retained only for source compatibility.
      */
     onIterationStart?: (event: AgentIterationStartEvent) => void;
     onTurnStart?: (event: AgentTurnStartEvent) => void;
@@ -51791,10 +51617,10 @@ interface AgentStreamCallbacks {
     onApprovalStart?: (event: AgentApprovalStartEvent) => void;
     onApprovalComplete?: (event: AgentApprovalCompleteEvent) => void;
     /**
-     * @deprecated No longer emitted on the unified wire vocabulary (the default
-     * since the unified-SSE cutover): iteration boundaries are folded into the
+     * @deprecated Not represented on the public unified wire: iteration
+     * boundaries are folded into the
      * `iteration` field on turn/tool events. Use `onTurnComplete` for per-turn
-     * cost/token data. Still fires on a legacy stream during the rollout.
+     * cost/token data. Retained only for source compatibility.
      */
     onIterationComplete?: (event: AgentIterationCompleteEvent) => void;
     onReflection?: (event: AgentReflectionEvent) => void;
@@ -53339,17 +53165,15 @@ declare function parseOffloadedOutputId(value: string): string | undefined;
 declare function parseLedgerArtifactRelativePath(value: string): string | undefined;
 
 /**
- * Unified → legacy SSE event adapter (the client-side mirror of the api's
+ * Unified wire → stable SDK callback adapter (the client-side mirror of the api's
  * `apps/api/src/lib/unified-event-stream.ts`).
  *
  * As of the unified-SSE cutover (runtypelabs/core unified-sse-default) the
  * execution streams the SDK consumes — `/dispatch`, `/dispatch/resume`,
- * `/agents/{id}/execute`, `/agents/{id}/resume` — default to the 33-event
- * unified vocabulary (`unifiedSSEEventSchema` in `@runtypelabs/shared`). The
- * SDK opts INTO that vocabulary explicitly with `?events=unified` on those
- * requests (see the call sites in `endpoints.ts` / `runtype.ts` /
- * `flows-namespace.ts`), then translates the unified frames back into the
- * SDK's stable, hand-written legacy event shapes here.
+ * `/agents/{id}/execute`, `/agents/{id}/resume` — use the 33-event unified
+ * vocabulary (`unifiedSSEEventSchema` in `@runtypelabs/shared`). The SDK
+ * translates those frames back into its stable, hand-written callback shapes
+ * here.
  *
  * Why translate instead of rewriting every consumer switch: the SDK's public
  * callback contracts (`StreamCallbacks`, `AgentStreamCallbacks`) and their event
@@ -53361,52 +53185,49 @@ declare function parseLedgerArtifactRelativePath(value: string): string | undefi
  *
  * Scope: only the events the SDK's consumers actually read are reconstructed;
  * everything else (artifact channel, `source`, `custom`, fallback live-beat,
- * `step_skip`) maps to zero legacy events. Two unified-contract folds are NOT
+ * `step_skip`) maps to zero callback events. Two unified-contract folds are NOT
  * reconstructed because the unified vocabulary deliberately discards them:
  *  - `agent_iteration_start` / `agent_iteration_complete` are folded into the
  *    `iteration` field on turn/tool frames, so `onIterationStart` /
  *    `onIterationComplete` no longer fire.
  *  - the additive `fallback` summary on `step_complete` / `turn_complete` is
- *    dropped (the SDK never consumed the legacy `fallback_*` events either).
+ *    dropped (the SDK callback surface never consumed `fallback_*` events).
  *
  * Each translator is STATEFUL (channel block ids, current step/turn/iteration,
  * media accumulation) and therefore must be constructed once PER STREAM — never
  * shared across executions.
  */
 type Json = Record<string, unknown>;
-/** A reconstructed legacy event (shape validated by the consumer switch/cast). */
-type LegacyEvent = Json & {
-    type: string;
-};
 /**
- * Query string that opts a streaming request into the unified SSE vocabulary
- * (`?events=unified`, resolved by the api's `resolveSseEventFormat`). The SDK
- * requests it explicitly on the cut-over routes so it controls its own wire
- * format regardless of the server-side `default-sse-event-format` rollout flag.
+ * @deprecated Unified events are now the only public execution-stream format.
+ * Retained so published SDK consumers can compile while removing old opt-in
+ * plumbing at their convenience.
  */
 declare const UNIFIED_EVENTS_QUERY = "events=unified";
-/** Append the unified-events opt-in to a request path, preserving any existing query. */
+/**
+ * @deprecated Unified events are now unconditional; the request path is
+ * returned unchanged and no negotiation query is appended.
+ */
 declare function withUnifiedEvents(path: string): string;
-/** True when `type` is an unambiguous unified-vocabulary event discriminator. */
+/** True when `type` belongs to the canonical unified vocabulary. */
 declare function isUnifiedEventType(type: unknown): boolean;
+/** A reconstructed compatibility event (validated by the consumer switch/cast). */
+type CompatibilityEvent = Json & {
+    type: string;
+};
 /**
- * Create a stateful translator from the unified vocabulary to the legacy flow
- * event shapes consumed by `stream-utils.ts` / `flows-namespace.ts` /
- * `client.ts`. Returns 0+ legacy events per unified event. Construct one per
+ * Create a stateful translator from the unified vocabulary to the stable flow
+ * callback shapes consumed by `stream-utils.ts` / `flows-namespace.ts` /
+ * `client.ts`. Returns 0+ callback events per unified event. Construct one per
  * stream.
- *
- * Tolerant by design: unified frames are translated; legacy flow frames pass
- * through unchanged (so a legacy stream — e.g. during the rollout window —
- * still works). The overlapping `step_start` / `step_complete` read both
- * vocabularies' field names.
  */
-declare function createFlowEventTranslator(): (event: unknown) => LegacyEvent[];
+declare function createFlowEventTranslator(): (event: unknown) => CompatibilityEvent[];
 /**
- * Create a stateful translator from the unified vocabulary to the legacy
- * `agent_*` event shapes dispatched by `processAgentStream` in `endpoints.ts`.
- * Returns 0+ legacy events per unified event. Construct one per stream.
+ * Create a stateful translator from the unified vocabulary to the stable
+ * `agent_*` callback shapes dispatched by `processAgentStream` in `endpoints.ts`.
+ * Returns 0+ callback events per unified event. Construct one per stream.
  */
-declare function createAgentEventTranslator(): (event: unknown) => LegacyEvent[];
+declare function createAgentEventTranslator(): (event: unknown) => CompatibilityEvent[];
 
 /**
  * BatchBuilder - Fluent builder for batch operations