@mastra/code-sdk 1.0.3-alpha.0 → 1.1.0-alpha.2

package/dist/evals/context-builder.js

@@ -1,120 +1,161 @@
 import { extractTrajectoryFromTrace } from "@mastra/core/evals";
+//#region src/evals/context-builder.ts
+/**
+* Build an evaluation context from a AgentController session.
+*
+* This extracts messages from storage, builds trajectory from trace spans,
+* and packages everything into the format scorers expect.
+*/
 async function buildEvalContext(options) {
-  const { controller, session, lastNTurns } = options;
-  const threadId = options.threadId ?? session.thread.getId();
-  if (!threadId) {
-    throw new Error("No thread ID available. Start a session before building eval context.");
-  }
-  const mastra = controller.getMastra();
-  const storage = mastra?.getStorage();
-  const rawMessages = await getRawMessages(storage, threadId, lastNTurns);
-  const { inputMessages, systemMessages, outputMessages } = categorizeMessages(rawMessages);
-  const { trajectory, traceId } = await extractSessionTrajectory(storage, threadId);
-  const requestContext = buildRequestContext(session, threadId);
-  return {
-    agentInput: {
-      inputMessages,
-      rememberedMessages: [],
-      // AgentController doesn't separate these; memory recall is transparent
-      systemMessages,
-      taggedSystemMessages: {}
-    },
-    agentOutput: outputMessages,
-    trajectory,
-    requestContext,
-    threadId,
-    traceId
-  };
+	const { controller, session, lastNTurns } = options;
+	const threadId = options.threadId ?? session.thread.getId();
+	if (!threadId) throw new Error("No thread ID available. Start a session before building eval context.");
+	const storage = controller.getMastra()?.getStorage();
+	const { inputMessages, systemMessages, outputMessages } = categorizeMessages(await getRawMessages(storage, threadId, lastNTurns));
+	const { trajectory, traceId } = await extractSessionTrajectory(storage, threadId);
+	const requestContext = buildRequestContext(session, threadId);
+	return {
+		agentInput: {
+			inputMessages,
+			rememberedMessages: [],
+			systemMessages,
+			taggedSystemMessages: {}
+		},
+		agentOutput: outputMessages,
+		trajectory,
+		requestContext,
+		threadId,
+		traceId
+	};
 }
+/**
+* Fetch raw MastraDBMessage[] from memory storage for a thread.
+*/
 async function getRawMessages(storage, threadId, lastNTurns) {
-  if (!storage) return [];
-  const memoryStore = await storage.getStore("memory");
-  if (!memoryStore) return [];
-  const result = await memoryStore.listMessages({
-    threadId,
-    perPage: false
-  });
-  let messages = result.messages;
-  if (lastNTurns !== void 0 && lastNTurns > 0) {
-    messages = trimToLastNTurns(messages, lastNTurns);
-  }
-  return messages;
+	if (!storage) return [];
+	const memoryStore = await storage.getStore("memory");
+	if (!memoryStore) return [];
+	let messages = (await memoryStore.listMessages({
+		threadId,
+		perPage: false
+	})).messages;
+	if (lastNTurns !== void 0 && lastNTurns > 0) messages = trimToLastNTurns(messages, lastNTurns);
+	return messages;
 }
+/**
+* Trim messages to the last N user-assistant turn pairs.
+*/
 function trimToLastNTurns(messages, turns) {
-  let turnCount = 0;
-  let cutoffIndex = 0;
-  for (let i = messages.length - 1; i >= 0; i--) {
-    if (messages[i].role === "user") {
-      turnCount++;
-      if (turnCount === turns) {
-        cutoffIndex = i;
-        break;
-      }
-    }
-  }
-  return messages.slice(cutoffIndex);
+	let turnCount = 0;
+	let cutoffIndex = 0;
+	for (let i = messages.length - 1; i >= 0; i--) if (messages[i].role === "user") {
+		turnCount++;
+		if (turnCount === turns) {
+			cutoffIndex = i;
+			break;
+		}
+	}
+	return messages.slice(cutoffIndex);
 }
+/**
+* Categorize messages into input (user), system, and output (assistant) groups.
+*/
 function categorizeMessages(messages) {
-  const inputMessages = [];
-  const systemMessages = [];
-  const outputMessages = [];
-  for (const msg of messages) {
-    switch (msg.role) {
-      case "user":
-        inputMessages.push(msg);
-        break;
-      case "system":
-        systemMessages.push({
-          role: "system",
-          content: extractTextContent(msg)
-        });
-        break;
-      case "assistant":
-        outputMessages.push(msg);
-        break;
-    }
-  }
-  return { inputMessages, systemMessages, outputMessages };
+	const inputMessages = [];
+	const systemMessages = [];
+	const outputMessages = [];
+	for (const msg of messages) switch (msg.role) {
+		case "user":
+			inputMessages.push(msg);
+			break;
+		case "system":
+			systemMessages.push({
+				role: "system",
+				content: extractTextContent(msg)
+			});
+			break;
+		case "assistant":
+			outputMessages.push(msg);
+			break;
+	}
+	return {
+		inputMessages,
+		systemMessages,
+		outputMessages
+	};
 }
+/**
+* Extract plain text content from a MastraDBMessage.
+*/
 function extractTextContent(msg) {
-  if (!msg.content?.parts) return "";
-  return msg.content.parts.filter((p) => p.type === "text" && "text" in p).map((p) => p.text).join("\n");
+	if (!msg.content?.parts) return "";
+	return msg.content.parts.filter((p) => p.type === "text" && "text" in p).map((p) => p.text).join("\n");
 }
+/**
+* Extract trajectory from trace spans in observability storage.
+* Looks up traces by threadId metadata.
+*/
 async function extractSessionTrajectory(storage, threadId) {
-  if (!storage) return { trajectory: void 0, traceId: void 0 };
-  try {
-    const observabilityStore = await storage.getStore("observability");
-    if (!observabilityStore) return { trajectory: void 0, traceId: void 0 };
-    const result = await observabilityStore.listTraces({
-      filters: { threadId },
-      pagination: { page: 0, perPage: 1 },
-      orderBy: { field: "startedAt", direction: "DESC" }
-    });
-    if (!result?.spans?.length) return { trajectory: void 0, traceId: void 0 };
-    const traceId = result.spans[0].traceId;
-    const trace = await observabilityStore.getTrace({ traceId });
-    if (!trace?.spans?.length) return { trajectory: void 0, traceId };
-    const trajectory = extractTrajectoryFromTrace(trace.spans);
-    return { trajectory, traceId };
-  } catch {
-    return { trajectory: void 0, traceId: void 0 };
-  }
+	if (!storage) return {
+		trajectory: void 0,
+		traceId: void 0
+	};
+	try {
+		const observabilityStore = await storage.getStore("observability");
+		if (!observabilityStore) return {
+			trajectory: void 0,
+			traceId: void 0
+		};
+		const result = await observabilityStore.listTraces({
+			filters: { threadId },
+			pagination: {
+				page: 0,
+				perPage: 1
+			},
+			orderBy: {
+				field: "startedAt",
+				direction: "DESC"
+			}
+		});
+		if (!result?.spans?.length) return {
+			trajectory: void 0,
+			traceId: void 0
+		};
+		const traceId = result.spans[0].traceId;
+		const trace = await observabilityStore.getTrace({ traceId });
+		if (!trace?.spans?.length) return {
+			trajectory: void 0,
+			traceId
+		};
+		return {
+			trajectory: extractTrajectoryFromTrace(trace.spans),
+			traceId
+		};
+	} catch {
+		return {
+			trajectory: void 0,
+			traceId: void 0
+		};
+	}
 }
+/**
+* Build request context from AgentController state for scorer evaluation.
+*/
 function buildRequestContext(session, threadId) {
-  const state = session.state.get();
-  return {
-    threadId,
-    mode: state.currentMode ?? state.mode,
-    modelId: session.model.get(),
-    projectPath: state.projectPath,
-    projectName: state.projectName,
-    gitBranch: state.gitBranch,
-    thinkingLevel: state.thinkingLevel,
-    yolo: state.yolo,
-    smartEditing: state.smartEditing
-  };
+	const state = session.state.get();
+	return {
+		threadId,
+		mode: state.currentMode ?? state.mode,
+		modelId: session.model.get(),
+		projectPath: state.projectPath,
+		projectName: state.projectName,
+		gitBranch: state.gitBranch,
+		thinkingLevel: state.thinkingLevel,
+		yolo: state.yolo,
+		smartEditing: state.smartEditing
+	};
 }
-export {
-  buildEvalContext
-};
+//#endregion
+export { buildEvalContext };
+
 //# sourceMappingURL=context-builder.js.map

package/dist/evals/context-builder.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../src/evals/context-builder.ts"],"sourcesContent":["/**\n * Eval Context Builder for MastraCode\n *\n * Converts AgentController session data into scorer-compatible format.\n * Supports both agent-type scorers (ScorerRunInputForAgent) and\n * trajectory scorers (Trajectory).\n */\n\nimport type { MastraDBMessage } from '@mastra/core/agent';\nimport type { AgentController, Session } from '@mastra/core/agent-controller';\nimport { extractTrajectoryFromTrace } from '@mastra/core/evals';\nimport type { ScorerRunInputForAgent, ScorerRunOutputForAgent, Trajectory } from '@mastra/core/evals';\nimport type { CoreMessage, CoreSystemMessage } from '@mastra/core/llm';\nimport type { MastraCompositeStore } from '@mastra/core/storage';\n\nexport type MastraCodeEvalContext = {\n  /** Scorer-compatible agent input (messages, system prompt, etc.) */\n  agentInput: ScorerRunInputForAgent;\n  /** Scorer-compatible agent output (response messages) */\n  agentOutput: ScorerRunOutputForAgent;\n  /** Execution trajectory extracted from trace spans */\n  trajectory: Trajectory | undefined;\n  /** Request context with session metadata */\n  requestContext: Record<string, unknown>;\n  /** The thread ID this context was built from */\n  threadId: string;\n  /** The trace ID for observability linkage */\n  traceId: string | undefined;\n};\n\nexport type BuildContextOptions = {\n  /** AgentController instance to extract data from */\n  controller: AgentController<any>;\n  /** Session whose state/model/thread is being evaluated */\n  session: Session<any>;\n  /** Thread ID to build context for (defaults to current thread) */\n  threadId?: string;\n  /** Limit messages to the last N turns (user+assistant pairs). Undefined = all messages */\n  lastNTurns?: number;\n};\n\n/**\n * Build an evaluation context from a AgentController session.\n *\n * This extracts messages from storage, builds trajectory from trace spans,\n * and packages everything into the format scorers expect.\n */\nexport async function buildEvalContext(options: BuildContextOptions): Promise<MastraCodeEvalContext> {\n  const { controller, session, lastNTurns } = options;\n  const threadId = options.threadId ?? session.thread.getId();\n\n  if (!threadId) {\n    throw new Error('No thread ID available. Start a session before building eval context.');\n  }\n\n  const mastra = controller.getMastra();\n  const storage = mastra?.getStorage();\n\n  // 1. Get raw MastraDB messages from memory storage\n  const rawMessages = await getRawMessages(storage, threadId, lastNTurns);\n\n  // 2. Split messages into input/output categories\n  const { inputMessages, systemMessages, outputMessages } = categorizeMessages(rawMessages);\n\n  // 3. Extract trajectory from observability traces\n  const { trajectory, traceId } = await extractSessionTrajectory(storage, threadId);\n\n  // 4. Build request context from AgentController state\n  const requestContext = buildRequestContext(session, threadId);\n\n  return {\n    agentInput: {\n      inputMessages,\n      rememberedMessages: [], // AgentController doesn't separate these; memory recall is transparent\n      systemMessages,\n      taggedSystemMessages: {},\n    },\n    agentOutput: outputMessages,\n    trajectory,\n    requestContext,\n    threadId,\n    traceId,\n  };\n}\n\n/**\n * Fetch raw MastraDBMessage[] from memory storage for a thread.\n */\nasync function getRawMessages(\n  storage: MastraCompositeStore | undefined,\n  threadId: string,\n  lastNTurns?: number,\n): Promise<MastraDBMessage[]> {\n  if (!storage) return [];\n\n  const memoryStore = await storage.getStore('memory');\n  if (!memoryStore) return [];\n\n  const result = await memoryStore.listMessages({\n    threadId,\n    perPage: false,\n  });\n\n  let messages = result.messages;\n\n  if (lastNTurns !== undefined && lastNTurns > 0) {\n    messages = trimToLastNTurns(messages, lastNTurns);\n  }\n\n  return messages;\n}\n\n/**\n * Trim messages to the last N user-assistant turn pairs.\n */\nfunction trimToLastNTurns(messages: MastraDBMessage[], turns: number): MastraDBMessage[] {\n  // Walk backward, counting user messages as turn boundaries.\n  // When we find the Nth user message, slice from there to include\n  // that turn and everything after it.\n  let turnCount = 0;\n  let cutoffIndex = 0;\n\n  for (let i = messages.length - 1; i >= 0; i--) {\n    if (messages[i]!.role === 'user') {\n      turnCount++;\n      if (turnCount === turns) {\n        cutoffIndex = i;\n        break;\n      }\n    }\n  }\n\n  return messages.slice(cutoffIndex);\n}\n\n/**\n * Categorize messages into input (user), system, and output (assistant) groups.\n */\nfunction categorizeMessages(messages: MastraDBMessage[]): {\n  inputMessages: MastraDBMessage[];\n  systemMessages: CoreMessage[];\n  outputMessages: MastraDBMessage[];\n} {\n  const inputMessages: MastraDBMessage[] = [];\n  const systemMessages: CoreSystemMessage[] = [];\n  const outputMessages: MastraDBMessage[] = [];\n\n  for (const msg of messages) {\n    switch (msg.role) {\n      case 'user':\n        inputMessages.push(msg);\n        break;\n      case 'system':\n        systemMessages.push({\n          role: 'system',\n          content: extractTextContent(msg),\n        });\n        break;\n      case 'assistant':\n        outputMessages.push(msg);\n        break;\n    }\n  }\n\n  return { inputMessages, systemMessages, outputMessages };\n}\n\n/**\n * Extract plain text content from a MastraDBMessage.\n */\nfunction extractTextContent(msg: MastraDBMessage): string {\n  if (!msg.content?.parts) return '';\n\n  return msg.content.parts\n    .filter((p): p is { type: 'text'; text: string } => p.type === 'text' && 'text' in p)\n    .map(p => p.text)\n    .join('\\n');\n}\n\n/**\n * Extract trajectory from trace spans in observability storage.\n * Looks up traces by threadId metadata.\n */\nasync function extractSessionTrajectory(\n  storage: MastraCompositeStore | undefined,\n  threadId: string,\n): Promise<{ trajectory: Trajectory | undefined; traceId: string | undefined }> {\n  if (!storage) return { trajectory: undefined, traceId: undefined };\n\n  try {\n    const observabilityStore = await storage.getStore('observability');\n    if (!observabilityStore) return { trajectory: undefined, traceId: undefined };\n\n    // Find traces associated with this thread\n    const result = await observabilityStore.listTraces({\n      filters: { threadId },\n      pagination: { page: 0, perPage: 1 },\n      orderBy: { field: 'startedAt', direction: 'DESC' },\n    });\n\n    if (!result?.spans?.length) return { trajectory: undefined, traceId: undefined };\n\n    const traceId = result.spans[0]!.traceId;\n    const trace = await observabilityStore.getTrace({ traceId });\n\n    if (!trace?.spans?.length) return { trajectory: undefined, traceId };\n\n    const trajectory = extractTrajectoryFromTrace(trace.spans);\n    return { trajectory, traceId };\n  } catch {\n    // Trace extraction is best-effort\n    return { trajectory: undefined, traceId: undefined };\n  }\n}\n\n/**\n * Build request context from AgentController state for scorer evaluation.\n */\nfunction buildRequestContext(session: Session<any>, threadId: string): Record<string, unknown> {\n  const state = session.state.get() as Record<string, unknown>;\n\n  return {\n    threadId,\n    mode: state.currentMode ?? state.mode,\n    modelId: session.model.get(),\n    projectPath: state.projectPath,\n    projectName: state.projectName,\n    gitBranch: state.gitBranch,\n    thinkingLevel: state.thinkingLevel,\n    yolo: state.yolo,\n    smartEditing: state.smartEditing,\n  };\n}\n"],"mappings":"AAUA,SAAS,kCAAkC;AAqC3C,eAAsB,iBAAiB,SAA8D;AACnG,QAAM,EAAE,YAAY,SAAS,WAAW,IAAI;AAC5C,QAAM,WAAW,QAAQ,YAAY,QAAQ,OAAO,MAAM;AAE1D,MAAI,CAAC,UAAU;AACb,UAAM,IAAI,MAAM,uEAAuE;AAAA,EACzF;AAEA,QAAM,SAAS,WAAW,UAAU;AACpC,QAAM,UAAU,QAAQ,WAAW;AAGnC,QAAM,cAAc,MAAM,eAAe,SAAS,UAAU,UAAU;AAGtE,QAAM,EAAE,eAAe,gBAAgB,eAAe,IAAI,mBAAmB,WAAW;AAGxF,QAAM,EAAE,YAAY,QAAQ,IAAI,MAAM,yBAAyB,SAAS,QAAQ;AAGhF,QAAM,iBAAiB,oBAAoB,SAAS,QAAQ;AAE5D,SAAO;AAAA,IACL,YAAY;AAAA,MACV;AAAA,MACA,oBAAoB,CAAC;AAAA;AAAA,MACrB;AAAA,MACA,sBAAsB,CAAC;AAAA,IACzB;AAAA,IACA,aAAa;AAAA,IACb;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACF;AAKA,eAAe,eACb,SACA,UACA,YAC4B;AAC5B,MAAI,CAAC,QAAS,QAAO,CAAC;AAEtB,QAAM,cAAc,MAAM,QAAQ,SAAS,QAAQ;AACnD,MAAI,CAAC,YAAa,QAAO,CAAC;AAE1B,QAAM,SAAS,MAAM,YAAY,aAAa;AAAA,IAC5C;AAAA,IACA,SAAS;AAAA,EACX,CAAC;AAED,MAAI,WAAW,OAAO;AAEtB,MAAI,eAAe,UAAa,aAAa,GAAG;AAC9C,eAAW,iBAAiB,UAAU,UAAU;AAAA,EAClD;AAEA,SAAO;AACT;AAKA,SAAS,iBAAiB,UAA6B,OAAkC;AAIvF,MAAI,YAAY;AAChB,MAAI,cAAc;AAElB,WAAS,IAAI,SAAS,SAAS,GAAG,KAAK,GAAG,KAAK;AAC7C,QAAI,SAAS,CAAC,EAAG,SAAS,QAAQ;AAChC;AACA,UAAI,cAAc,OAAO;AACvB,sBAAc;AACd;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,SAAO,SAAS,MAAM,WAAW;AACnC;AAKA,SAAS,mBAAmB,UAI1B;AACA,QAAM,gBAAmC,CAAC;AAC1C,QAAM,iBAAsC,CAAC;AAC7C,QAAM,iBAAoC,CAAC;AAE3C,aAAW,OAAO,UAAU;AAC1B,YAAQ,IAAI,MAAM;AAAA,MAChB,KAAK;AACH,sBAAc,KAAK,GAAG;AACtB;AAAA,MACF,KAAK;AACH,uBAAe,KAAK;AAAA,UAClB,MAAM;AAAA,UACN,SAAS,mBAAmB,GAAG;AAAA,QACjC,CAAC;AACD;AAAA,MACF,KAAK;AACH,uBAAe,KAAK,GAAG;AACvB;AAAA,IACJ;AAAA,EACF;AAEA,SAAO,EAAE,eAAe,gBAAgB,eAAe;AACzD;AAKA,SAAS,mBAAmB,KAA8B;AACxD,MAAI,CAAC,IAAI,SAAS,MAAO,QAAO;AAEhC,SAAO,IAAI,QAAQ,MAChB,OAAO,CAAC,MAA2C,EAAE,SAAS,UAAU,UAAU,CAAC,EACnF,IAAI,OAAK,EAAE,IAAI,EACf,KAAK,IAAI;AACd;AAMA,eAAe,yBACb,SACA,UAC8E;AAC9E,MAAI,CAAC,QAAS,QAAO,EAAE,YAAY,QAAW,SAAS,OAAU;AAEjE,MAAI;AACF,UAAM,qBAAqB,MAAM,QAAQ,SAAS,eAAe;AACjE,QAAI,CAAC,mBAAoB,QAAO,EAAE,YAAY,QAAW,SAAS,OAAU;AAG5E,UAAM,SAAS,MAAM,mBAAmB,WAAW;AAAA,MACjD,SAAS,EAAE,SAAS;AAAA,MACpB,YAAY,EAAE,MAAM,GAAG,SAAS,EAAE;AAAA,MAClC,SAAS,EAAE,OAAO,aAAa,WAAW,OAAO;AAAA,IACnD,CAAC;AAED,QAAI,CAAC,QAAQ,OAAO,OAAQ,QAAO,EAAE,YAAY,QAAW,SAAS,OAAU;AAE/E,UAAM,UAAU,OAAO,MAAM,CAAC,EAAG;AACjC,UAAM,QAAQ,MAAM,mBAAmB,SAAS,EAAE,QAAQ,CAAC;AAE3D,QAAI,CAAC,OAAO,OAAO,OAAQ,QAAO,EAAE,YAAY,QAAW,QAAQ;AAEnE,UAAM,aAAa,2BAA2B,MAAM,KAAK;AACzD,WAAO,EAAE,YAAY,QAAQ;AAAA,EAC/B,QAAQ;AAEN,WAAO,EAAE,YAAY,QAAW,SAAS,OAAU;AAAA,EACrD;AACF;AAKA,SAAS,oBAAoB,SAAuB,UAA2C;AAC7F,QAAM,QAAQ,QAAQ,MAAM,IAAI;AAEhC,SAAO;AAAA,IACL;AAAA,IACA,MAAM,MAAM,eAAe,MAAM;AAAA,IACjC,SAAS,QAAQ,MAAM,IAAI;AAAA,IAC3B,aAAa,MAAM;AAAA,IACnB,aAAa,MAAM;AAAA,IACnB,WAAW,MAAM;AAAA,IACjB,eAAe,MAAM;AAAA,IACrB,MAAM,MAAM;AAAA,IACZ,cAAc,MAAM;AAAA,EACtB;AACF;","names":[]}
+{"version":3,"file":"context-builder.js","names":[],"sources":["../../src/evals/context-builder.ts"],"sourcesContent":["/**\n * Eval Context Builder for MastraCode\n *\n * Converts AgentController session data into scorer-compatible format.\n * Supports both agent-type scorers (ScorerRunInputForAgent) and\n * trajectory scorers (Trajectory).\n */\n\nimport type { MastraDBMessage } from '@mastra/core/agent';\nimport type { AgentController, Session } from '@mastra/core/agent-controller';\nimport { extractTrajectoryFromTrace } from '@mastra/core/evals';\nimport type { ScorerRunInputForAgent, ScorerRunOutputForAgent, Trajectory } from '@mastra/core/evals';\nimport type { CoreMessage, CoreSystemMessage } from '@mastra/core/llm';\nimport type { MastraCompositeStore } from '@mastra/core/storage';\n\nexport type MastraCodeEvalContext = {\n  /** Scorer-compatible agent input (messages, system prompt, etc.) */\n  agentInput: ScorerRunInputForAgent;\n  /** Scorer-compatible agent output (response messages) */\n  agentOutput: ScorerRunOutputForAgent;\n  /** Execution trajectory extracted from trace spans */\n  trajectory: Trajectory | undefined;\n  /** Request context with session metadata */\n  requestContext: Record<string, unknown>;\n  /** The thread ID this context was built from */\n  threadId: string;\n  /** The trace ID for observability linkage */\n  traceId: string | undefined;\n};\n\nexport type BuildContextOptions = {\n  /** AgentController instance to extract data from */\n  controller: AgentController<any>;\n  /** Session whose state/model/thread is being evaluated */\n  session: Session<any>;\n  /** Thread ID to build context for (defaults to current thread) */\n  threadId?: string;\n  /** Limit messages to the last N turns (user+assistant pairs). Undefined = all messages */\n  lastNTurns?: number;\n};\n\n/**\n * Build an evaluation context from a AgentController session.\n *\n * This extracts messages from storage, builds trajectory from trace spans,\n * and packages everything into the format scorers expect.\n */\nexport async function buildEvalContext(options: BuildContextOptions): Promise<MastraCodeEvalContext> {\n  const { controller, session, lastNTurns } = options;\n  const threadId = options.threadId ?? session.thread.getId();\n\n  if (!threadId) {\n    throw new Error('No thread ID available. Start a session before building eval context.');\n  }\n\n  const mastra = controller.getMastra();\n  const storage = mastra?.getStorage();\n\n  // 1. Get raw MastraDB messages from memory storage\n  const rawMessages = await getRawMessages(storage, threadId, lastNTurns);\n\n  // 2. Split messages into input/output categories\n  const { inputMessages, systemMessages, outputMessages } = categorizeMessages(rawMessages);\n\n  // 3. Extract trajectory from observability traces\n  const { trajectory, traceId } = await extractSessionTrajectory(storage, threadId);\n\n  // 4. Build request context from AgentController state\n  const requestContext = buildRequestContext(session, threadId);\n\n  return {\n    agentInput: {\n      inputMessages,\n      rememberedMessages: [], // AgentController doesn't separate these; memory recall is transparent\n      systemMessages,\n      taggedSystemMessages: {},\n    },\n    agentOutput: outputMessages,\n    trajectory,\n    requestContext,\n    threadId,\n    traceId,\n  };\n}\n\n/**\n * Fetch raw MastraDBMessage[] from memory storage for a thread.\n */\nasync function getRawMessages(\n  storage: MastraCompositeStore | undefined,\n  threadId: string,\n  lastNTurns?: number,\n): Promise<MastraDBMessage[]> {\n  if (!storage) return [];\n\n  const memoryStore = await storage.getStore('memory');\n  if (!memoryStore) return [];\n\n  const result = await memoryStore.listMessages({\n    threadId,\n    perPage: false,\n  });\n\n  let messages = result.messages;\n\n  if (lastNTurns !== undefined && lastNTurns > 0) {\n    messages = trimToLastNTurns(messages, lastNTurns);\n  }\n\n  return messages;\n}\n\n/**\n * Trim messages to the last N user-assistant turn pairs.\n */\nfunction trimToLastNTurns(messages: MastraDBMessage[], turns: number): MastraDBMessage[] {\n  // Walk backward, counting user messages as turn boundaries.\n  // When we find the Nth user message, slice from there to include\n  // that turn and everything after it.\n  let turnCount = 0;\n  let cutoffIndex = 0;\n\n  for (let i = messages.length - 1; i >= 0; i--) {\n    if (messages[i]!.role === 'user') {\n      turnCount++;\n      if (turnCount === turns) {\n        cutoffIndex = i;\n        break;\n      }\n    }\n  }\n\n  return messages.slice(cutoffIndex);\n}\n\n/**\n * Categorize messages into input (user), system, and output (assistant) groups.\n */\nfunction categorizeMessages(messages: MastraDBMessage[]): {\n  inputMessages: MastraDBMessage[];\n  systemMessages: CoreMessage[];\n  outputMessages: MastraDBMessage[];\n} {\n  const inputMessages: MastraDBMessage[] = [];\n  const systemMessages: CoreSystemMessage[] = [];\n  const outputMessages: MastraDBMessage[] = [];\n\n  for (const msg of messages) {\n    switch (msg.role) {\n      case 'user':\n        inputMessages.push(msg);\n        break;\n      case 'system':\n        systemMessages.push({\n          role: 'system',\n          content: extractTextContent(msg),\n        });\n        break;\n      case 'assistant':\n        outputMessages.push(msg);\n        break;\n    }\n  }\n\n  return { inputMessages, systemMessages, outputMessages };\n}\n\n/**\n * Extract plain text content from a MastraDBMessage.\n */\nfunction extractTextContent(msg: MastraDBMessage): string {\n  if (!msg.content?.parts) return '';\n\n  return msg.content.parts\n    .filter((p): p is { type: 'text'; text: string } => p.type === 'text' && 'text' in p)\n    .map(p => p.text)\n    .join('\\n');\n}\n\n/**\n * Extract trajectory from trace spans in observability storage.\n * Looks up traces by threadId metadata.\n */\nasync function extractSessionTrajectory(\n  storage: MastraCompositeStore | undefined,\n  threadId: string,\n): Promise<{ trajectory: Trajectory | undefined; traceId: string | undefined }> {\n  if (!storage) return { trajectory: undefined, traceId: undefined };\n\n  try {\n    const observabilityStore = await storage.getStore('observability');\n    if (!observabilityStore) return { trajectory: undefined, traceId: undefined };\n\n    // Find traces associated with this thread\n    const result = await observabilityStore.listTraces({\n      filters: { threadId },\n      pagination: { page: 0, perPage: 1 },\n      orderBy: { field: 'startedAt', direction: 'DESC' },\n    });\n\n    if (!result?.spans?.length) return { trajectory: undefined, traceId: undefined };\n\n    const traceId = result.spans[0]!.traceId;\n    const trace = await observabilityStore.getTrace({ traceId });\n\n    if (!trace?.spans?.length) return { trajectory: undefined, traceId };\n\n    const trajectory = extractTrajectoryFromTrace(trace.spans);\n    return { trajectory, traceId };\n  } catch {\n    // Trace extraction is best-effort\n    return { trajectory: undefined, traceId: undefined };\n  }\n}\n\n/**\n * Build request context from AgentController state for scorer evaluation.\n */\nfunction buildRequestContext(session: Session<any>, threadId: string): Record<string, unknown> {\n  const state = session.state.get() as Record<string, unknown>;\n\n  return {\n    threadId,\n    mode: state.currentMode ?? state.mode,\n    modelId: session.model.get(),\n    projectPath: state.projectPath,\n    projectName: state.projectName,\n    gitBranch: state.gitBranch,\n    thinkingLevel: state.thinkingLevel,\n    yolo: state.yolo,\n    smartEditing: state.smartEditing,\n  };\n}\n"],"mappings":";;;;;;;;AA+CA,eAAsB,iBAAiB,SAA8D;CACnG,MAAM,EAAE,YAAY,SAAS,eAAe;CAC5C,MAAM,WAAW,QAAQ,YAAY,QAAQ,OAAO,MAAM;CAE1D,IAAI,CAAC,UACH,MAAM,IAAI,MAAM,uEAAuE;CAIzF,MAAM,UADS,WAAW,UACL,CAAC,EAAE,WAAW;CAMnC,MAAM,EAAE,eAAe,gBAAgB,mBAAmB,mBAAmB,MAHnD,eAAe,SAAS,UAAU,UAAU,CAGkB;CAGxF,MAAM,EAAE,YAAY,YAAY,MAAM,yBAAyB,SAAS,QAAQ;CAGhF,MAAM,iBAAiB,oBAAoB,SAAS,QAAQ;CAE5D,OAAO;EACL,YAAY;GACV;GACA,oBAAoB,CAAC;GACrB;GACA,sBAAsB,CAAC;EACzB;EACA,aAAa;EACb;EACA;EACA;EACA;CACF;AACF;;;;AAKA,eAAe,eACb,SACA,UACA,YAC4B;CAC5B,IAAI,CAAC,SAAS,OAAO,CAAC;CAEtB,MAAM,cAAc,MAAM,QAAQ,SAAS,QAAQ;CACnD,IAAI,CAAC,aAAa,OAAO,CAAC;CAO1B,IAAI,YAAW,MALM,YAAY,aAAa;EAC5C;EACA,SAAS;CACX,CAAC,EAAA,CAEqB;CAEtB,IAAI,eAAe,KAAA,KAAa,aAAa,GAC3C,WAAW,iBAAiB,UAAU,UAAU;CAGlD,OAAO;AACT;;;;AAKA,SAAS,iBAAiB,UAA6B,OAAkC;CAIvF,IAAI,YAAY;CAChB,IAAI,cAAc;CAElB,KAAK,IAAI,IAAI,SAAS,SAAS,GAAG,KAAK,GAAG,KACxC,IAAI,SAAS,EAAE,CAAE,SAAS,QAAQ;EAChC;EACA,IAAI,cAAc,OAAO;GACvB,cAAc;GACd;EACF;CACF;CAGF,OAAO,SAAS,MAAM,WAAW;AACnC;;;;AAKA,SAAS,mBAAmB,UAI1B;CACA,MAAM,gBAAmC,CAAC;CAC1C,MAAM,iBAAsC,CAAC;CAC7C,MAAM,iBAAoC,CAAC;CAE3C,KAAK,MAAM,OAAO,UAChB,QAAQ,IAAI,MAAZ;EACE,KAAK;GACH,cAAc,KAAK,GAAG;GACtB;EACF,KAAK;GACH,eAAe,KAAK;IAClB,MAAM;IACN,SAAS,mBAAmB,GAAG;GACjC,CAAC;GACD;EACF,KAAK;GACH,eAAe,KAAK,GAAG;GACvB;CACJ;CAGF,OAAO;EAAE;EAAe;EAAgB;CAAe;AACzD;;;;AAKA,SAAS,mBAAmB,KAA8B;CACxD,IAAI,CAAC,IAAI,SAAS,OAAO,OAAO;CAEhC,OAAO,IAAI,QAAQ,MAChB,QAAQ,MAA2C,EAAE,SAAS,UAAU,UAAU,CAAC,CAAC,CACpF,KAAI,MAAK,EAAE,IAAI,CAAC,CAChB,KAAK,IAAI;AACd;;;;;AAMA,eAAe,yBACb,SACA,UAC8E;CAC9E,IAAI,CAAC,SAAS,OAAO;EAAE,YAAY,KAAA;EAAW,SAAS,KAAA;CAAU;CAEjE,IAAI;EACF,MAAM,qBAAqB,MAAM,QAAQ,SAAS,eAAe;EACjE,IAAI,CAAC,oBAAoB,OAAO;GAAE,YAAY,KAAA;GAAW,SAAS,KAAA;EAAU;EAG5E,MAAM,SAAS,MAAM,mBAAmB,WAAW;GACjD,SAAS,EAAE,SAAS;GACpB,YAAY;IAAE,MAAM;IAAG,SAAS;GAAE;GAClC,SAAS;IAAE,OAAO;IAAa,WAAW;GAAO;EACnD,CAAC;EAED,IAAI,CAAC,QAAQ,OAAO,QAAQ,OAAO;GAAE,YAAY,KAAA;GAAW,SAAS,KAAA;EAAU;EAE/E,MAAM,UAAU,OAAO,MAAM,EAAE,CAAE;EACjC,MAAM,QAAQ,MAAM,mBAAmB,SAAS,EAAE,QAAQ,CAAC;EAE3D,IAAI,CAAC,OAAO,OAAO,QAAQ,OAAO;GAAE,YAAY,KAAA;GAAW;EAAQ;EAGnE,OAAO;GAAE,YADU,2BAA2B,MAAM,KAClC;GAAG;EAAQ;CAC/B,QAAQ;EAEN,OAAO;GAAE,YAAY,KAAA;GAAW,SAAS,KAAA;EAAU;CACrD;AACF;;;;AAKA,SAAS,oBAAoB,SAAuB,UAA2C;CAC7F,MAAM,QAAQ,QAAQ,MAAM,IAAI;CAEhC,OAAO;EACL;EACA,MAAM,MAAM,eAAe,MAAM;EACjC,SAAS,QAAQ,MAAM,IAAI;EAC3B,aAAa,MAAM;EACnB,aAAa,MAAM;EACnB,WAAW,MAAM;EACjB,eAAe,MAAM;EACrB,MAAM,MAAM;EACZ,cAAc,MAAM;CACtB;AACF"}

package/dist/evals/index.js

@@ -1,8 +1,5 @@
+import { createOutcomeScorer } from "./scorers/outcome.js";
+import { createEfficiencyScorer } from "./scorers/efficiency.js";
+import "./scorers/index.js";
 import { buildEvalContext } from "./context-builder.js";
-import { createOutcomeScorer, createEfficiencyScorer } from "./scorers/index.js";
-export {
-  buildEvalContext,
-  createEfficiencyScorer,
-  createOutcomeScorer
-};
-//# sourceMappingURL=index.js.map
+export { buildEvalContext, createEfficiencyScorer, createOutcomeScorer };

package/dist/evals/scorers/classify-command.js

@@ -1,52 +1,73 @@
+//#region src/evals/scorers/classify-command.ts
+/**
+* Shared command classification utilities for scoring.
+*
+* Used by both live scorers (outcome.ts) and offline scorers (outcome-match.ts)
+* to consistently identify build, test, and other command types.
+*/
+/**
+* Check if a command string is a build/typecheck command.
+* Handles compound commands (e.g. `pnpm build && pnpm test`) by checking each segment.
+*/
 function isBuildCommand(command) {
-  const segments = command.split(/\s*(?:&&|\|\||;)\s*/);
-  return segments.some((seg) => {
-    const trimmed = seg.trim();
-    if (!/\b(tsc|typecheck|type-check|build)\b/.test(trimmed)) return false;
-    const primaryVerb = trimmed.replace(/^(?:pnpm|npm|yarn|npx|bunx?|turbo)\s+(?:run\s+)?/, "").split(/\s/)[0] ?? "";
-    return !/^(test|vitest|jest|mocha|pytest)$/.test(primaryVerb);
-  });
+	return command.split(/\s*(?:&&|\|\||;)\s*/).some((seg) => {
+		const trimmed = seg.trim();
+		if (!/\b(tsc|typecheck|type-check|build)\b/.test(trimmed)) return false;
+		const primaryVerb = trimmed.replace(/^(?:pnpm|npm|yarn|npx|bunx?|turbo)\s+(?:run\s+)?/, "").split(/\s/)[0] ?? "";
+		return !/^(test|vitest|jest|mocha|pytest)$/.test(primaryVerb);
+	});
 }
+/**
+* Check if a command string is a test command.
+*/
 function isTestCommand(command) {
-  return /\b(test|vitest|jest|mocha|pytest|spec)\b/.test(command);
+	return /\b(test|vitest|jest|mocha|pytest|spec)\b/.test(command);
 }
+/**
+* Extract exit code from an `execute_command` result.
+* Only trusts explicit exitCode/code fields — no heuristic fallbacks.
+*/
 function getExitCode(result) {
-  if (result == null) return null;
-  if (typeof result === "string") {
-    const exitMatch = result.match(/exit(?:ed)?\s+(?:with\s+)?(?:code\s+)?(\d+)/i);
-    if (exitMatch) return parseInt(exitMatch[1], 10);
-    return null;
-  }
-  if (typeof result === "object") {
-    const obj = result;
-    if ("exitCode" in obj && typeof obj.exitCode === "number") return obj.exitCode;
-    if ("code" in obj && typeof obj.code === "number") return obj.code;
-  }
-  return null;
+	if (result == null) return null;
+	if (typeof result === "string") {
+		const exitMatch = result.match(/exit(?:ed)?\s+(?:with\s+)?(?:code\s+)?(\d+)/i);
+		if (exitMatch) return parseInt(exitMatch[1], 10);
+		return null;
+	}
+	if (typeof result === "object") {
+		const obj = result;
+		if ("exitCode" in obj && typeof obj.exitCode === "number") return obj.exitCode;
+		if ("code" in obj && typeof obj.code === "number") return obj.code;
+	}
+	return null;
 }
+/**
+* Determine if a tool call result indicates success (exit code 0, no explicit error).
+* Works with both structured result objects and string results.
+*/
 function isSuccessResult(result, error) {
-  if (error) return false;
-  if (!result) return false;
-  if (typeof result === "object") {
-    const obj = result;
-    const exitCode = obj.exitCode ?? obj.code;
-    if (typeof exitCode === "number") return exitCode === 0;
-  }
-  if (typeof result === "string") {
-    const exitMatch = result.match(/exit(?:ed)?\s+(?:with\s+)?(?:code\s+)?(\d+)/i);
-    if (exitMatch) return exitMatch[1] === "0";
-  }
-  return false;
+	if (error) return false;
+	if (!result) return false;
+	if (typeof result === "object") {
+		const obj = result;
+		const exitCode = obj.exitCode ?? obj.code;
+		if (typeof exitCode === "number") return exitCode === 0;
+	}
+	if (typeof result === "string") {
+		const exitMatch = result.match(/exit(?:ed)?\s+(?:with\s+)?(?:code\s+)?(\d+)/i);
+		if (exitMatch) return exitMatch[1] === "0";
+	}
+	return false;
 }
+/**
+* Match a file path from tool output against an expected path.
+* Supports exact match and suffix match (e.g. "src/foo.ts" matches "/abs/path/src/foo.ts").
+*/
 function matchFilePath(actual, expected) {
-  if (actual === expected) return true;
-  return actual.endsWith(`/${expected}`) || expected.endsWith(`/${actual}`);
+	if (actual === expected) return true;
+	return actual.endsWith(`/${expected}`) || expected.endsWith(`/${actual}`);
 }
-export {
-  getExitCode,
-  isBuildCommand,
-  isSuccessResult,
-  isTestCommand,
-  matchFilePath
-};
+//#endregion
+export { getExitCode, isBuildCommand, isSuccessResult, isTestCommand, matchFilePath };
+
 //# sourceMappingURL=classify-command.js.map

package/dist/evals/scorers/classify-command.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../../../src/evals/scorers/classify-command.ts"],"sourcesContent":["/**\n * Shared command classification utilities for scoring.\n *\n * Used by both live scorers (outcome.ts) and offline scorers (outcome-match.ts)\n * to consistently identify build, test, and other command types.\n */\n\n/**\n * Check if a command string is a build/typecheck command.\n * Handles compound commands (e.g. `pnpm build && pnpm test`) by checking each segment.\n */\nexport function isBuildCommand(command: string): boolean {\n  const segments = command.split(/\\s*(?:&&|\\|\\||;)\\s*/);\n  return segments.some(seg => {\n    const trimmed = seg.trim();\n    if (!/\\b(tsc|typecheck|type-check|build)\\b/.test(trimmed)) return false;\n    // Don't reject if \"test\" only appears as a substring (e.g. \"build-test-output\").\n    // Only reject if a test runner is the primary verb of this segment.\n    const primaryVerb = trimmed.replace(/^(?:pnpm|npm|yarn|npx|bunx?|turbo)\\s+(?:run\\s+)?/, '').split(/\\s/)[0] ?? '';\n    return !/^(test|vitest|jest|mocha|pytest)$/.test(primaryVerb);\n  });\n}\n\n/**\n * Check if a command string is a test command.\n */\nexport function isTestCommand(command: string): boolean {\n  return /\\b(test|vitest|jest|mocha|pytest|spec)\\b/.test(command);\n}\n\n/**\n * Extract exit code from an `execute_command` result.\n * Only trusts explicit exitCode/code fields — no heuristic fallbacks.\n */\nexport function getExitCode(result: unknown): number | null {\n  if (result == null) return null;\n\n  if (typeof result === 'string') {\n    const exitMatch = result.match(/exit(?:ed)?\\s+(?:with\\s+)?(?:code\\s+)?(\\d+)/i);\n    if (exitMatch) return parseInt(exitMatch[1]!, 10);\n    return null;\n  }\n\n  if (typeof result === 'object') {\n    const obj = result as Record<string, unknown>;\n    if ('exitCode' in obj && typeof obj.exitCode === 'number') return obj.exitCode;\n    if ('code' in obj && typeof obj.code === 'number') return obj.code;\n  }\n\n  return null;\n}\n\n/**\n * Determine if a tool call result indicates success (exit code 0, no explicit error).\n * Works with both structured result objects and string results.\n */\nexport function isSuccessResult(result: unknown, error?: unknown): boolean {\n  if (error) return false;\n  if (!result) return false;\n\n  if (typeof result === 'object') {\n    const obj = result as Record<string, unknown>;\n    const exitCode = (obj.exitCode as number | undefined) ?? (obj.code as number | undefined);\n    if (typeof exitCode === 'number') return exitCode === 0;\n  }\n\n  if (typeof result === 'string') {\n    const exitMatch = result.match(/exit(?:ed)?\\s+(?:with\\s+)?(?:code\\s+)?(\\d+)/i);\n    if (exitMatch) return exitMatch[1] === '0';\n  }\n\n  // No explicit exit code — can't determine\n  return false;\n}\n\n/**\n * Match a file path from tool output against an expected path.\n * Supports exact match and suffix match (e.g. \"src/foo.ts\" matches \"/abs/path/src/foo.ts\").\n */\nexport function matchFilePath(actual: string, expected: string): boolean {\n  if (actual === expected) return true;\n  return actual.endsWith(`/${expected}`) || expected.endsWith(`/${actual}`);\n}\n"],"mappings":"AAWO,SAAS,eAAe,SAA0B;AACvD,QAAM,WAAW,QAAQ,MAAM,qBAAqB;AACpD,SAAO,SAAS,KAAK,SAAO;AAC1B,UAAM,UAAU,IAAI,KAAK;AACzB,QAAI,CAAC,uCAAuC,KAAK,OAAO,EAAG,QAAO;AAGlE,UAAM,cAAc,QAAQ,QAAQ,oDAAoD,EAAE,EAAE,MAAM,IAAI,EAAE,CAAC,KAAK;AAC9G,WAAO,CAAC,oCAAoC,KAAK,WAAW;AAAA,EAC9D,CAAC;AACH;AAKO,SAAS,cAAc,SAA0B;AACtD,SAAO,2CAA2C,KAAK,OAAO;AAChE;AAMO,SAAS,YAAY,QAAgC;AAC1D,MAAI,UAAU,KAAM,QAAO;AAE3B,MAAI,OAAO,WAAW,UAAU;AAC9B,UAAM,YAAY,OAAO,MAAM,8CAA8C;AAC7E,QAAI,UAAW,QAAO,SAAS,UAAU,CAAC,GAAI,EAAE;AAChD,WAAO;AAAA,EACT;AAEA,MAAI,OAAO,WAAW,UAAU;AAC9B,UAAM,MAAM;AACZ,QAAI,cAAc,OAAO,OAAO,IAAI,aAAa,SAAU,QAAO,IAAI;AACtE,QAAI,UAAU,OAAO,OAAO,IAAI,SAAS,SAAU,QAAO,IAAI;AAAA,EAChE;AAEA,SAAO;AACT;AAMO,SAAS,gBAAgB,QAAiB,OAA0B;AACzE,MAAI,MAAO,QAAO;AAClB,MAAI,CAAC,OAAQ,QAAO;AAEpB,MAAI,OAAO,WAAW,UAAU;AAC9B,UAAM,MAAM;AACZ,UAAM,WAAY,IAAI,YAAoC,IAAI;AAC9D,QAAI,OAAO,aAAa,SAAU,QAAO,aAAa;AAAA,EACxD;AAEA,MAAI,OAAO,WAAW,UAAU;AAC9B,UAAM,YAAY,OAAO,MAAM,8CAA8C;AAC7E,QAAI,UAAW,QAAO,UAAU,CAAC,MAAM;AAAA,EACzC;AAGA,SAAO;AACT;AAMO,SAAS,cAAc,QAAgB,UAA2B;AACvE,MAAI,WAAW,SAAU,QAAO;AAChC,SAAO,OAAO,SAAS,IAAI,QAAQ,EAAE,KAAK,SAAS,SAAS,IAAI,MAAM,EAAE;AAC1E;","names":[]}
+{"version":3,"file":"classify-command.js","names":[],"sources":["../../../src/evals/scorers/classify-command.ts"],"sourcesContent":["/**\n * Shared command classification utilities for scoring.\n *\n * Used by both live scorers (outcome.ts) and offline scorers (outcome-match.ts)\n * to consistently identify build, test, and other command types.\n */\n\n/**\n * Check if a command string is a build/typecheck command.\n * Handles compound commands (e.g. `pnpm build && pnpm test`) by checking each segment.\n */\nexport function isBuildCommand(command: string): boolean {\n  const segments = command.split(/\\s*(?:&&|\\|\\||;)\\s*/);\n  return segments.some(seg => {\n    const trimmed = seg.trim();\n    if (!/\\b(tsc|typecheck|type-check|build)\\b/.test(trimmed)) return false;\n    // Don't reject if \"test\" only appears as a substring (e.g. \"build-test-output\").\n    // Only reject if a test runner is the primary verb of this segment.\n    const primaryVerb = trimmed.replace(/^(?:pnpm|npm|yarn|npx|bunx?|turbo)\\s+(?:run\\s+)?/, '').split(/\\s/)[0] ?? '';\n    return !/^(test|vitest|jest|mocha|pytest)$/.test(primaryVerb);\n  });\n}\n\n/**\n * Check if a command string is a test command.\n */\nexport function isTestCommand(command: string): boolean {\n  return /\\b(test|vitest|jest|mocha|pytest|spec)\\b/.test(command);\n}\n\n/**\n * Extract exit code from an `execute_command` result.\n * Only trusts explicit exitCode/code fields — no heuristic fallbacks.\n */\nexport function getExitCode(result: unknown): number | null {\n  if (result == null) return null;\n\n  if (typeof result === 'string') {\n    const exitMatch = result.match(/exit(?:ed)?\\s+(?:with\\s+)?(?:code\\s+)?(\\d+)/i);\n    if (exitMatch) return parseInt(exitMatch[1]!, 10);\n    return null;\n  }\n\n  if (typeof result === 'object') {\n    const obj = result as Record<string, unknown>;\n    if ('exitCode' in obj && typeof obj.exitCode === 'number') return obj.exitCode;\n    if ('code' in obj && typeof obj.code === 'number') return obj.code;\n  }\n\n  return null;\n}\n\n/**\n * Determine if a tool call result indicates success (exit code 0, no explicit error).\n * Works with both structured result objects and string results.\n */\nexport function isSuccessResult(result: unknown, error?: unknown): boolean {\n  if (error) return false;\n  if (!result) return false;\n\n  if (typeof result === 'object') {\n    const obj = result as Record<string, unknown>;\n    const exitCode = (obj.exitCode as number | undefined) ?? (obj.code as number | undefined);\n    if (typeof exitCode === 'number') return exitCode === 0;\n  }\n\n  if (typeof result === 'string') {\n    const exitMatch = result.match(/exit(?:ed)?\\s+(?:with\\s+)?(?:code\\s+)?(\\d+)/i);\n    if (exitMatch) return exitMatch[1] === '0';\n  }\n\n  // No explicit exit code — can't determine\n  return false;\n}\n\n/**\n * Match a file path from tool output against an expected path.\n * Supports exact match and suffix match (e.g. \"src/foo.ts\" matches \"/abs/path/src/foo.ts\").\n */\nexport function matchFilePath(actual: string, expected: string): boolean {\n  if (actual === expected) return true;\n  return actual.endsWith(`/${expected}`) || expected.endsWith(`/${actual}`);\n}\n"],"mappings":";;;;;;;;;;;AAWA,SAAgB,eAAe,SAA0B;CAEvD,OADiB,QAAQ,MAAM,qBACjB,CAAC,CAAC,MAAK,QAAO;EAC1B,MAAM,UAAU,IAAI,KAAK;EACzB,IAAI,CAAC,uCAAuC,KAAK,OAAO,GAAG,OAAO;EAGlE,MAAM,cAAc,QAAQ,QAAQ,oDAAoD,EAAE,CAAC,CAAC,MAAM,IAAI,CAAC,CAAC,MAAM;EAC9G,OAAO,CAAC,oCAAoC,KAAK,WAAW;CAC9D,CAAC;AACH;;;;AAKA,SAAgB,cAAc,SAA0B;CACtD,OAAO,2CAA2C,KAAK,OAAO;AAChE;;;;;AAMA,SAAgB,YAAY,QAAgC;CAC1D,IAAI,UAAU,MAAM,OAAO;CAE3B,IAAI,OAAO,WAAW,UAAU;EAC9B,MAAM,YAAY,OAAO,MAAM,8CAA8C;EAC7E,IAAI,WAAW,OAAO,SAAS,UAAU,IAAK,EAAE;EAChD,OAAO;CACT;CAEA,IAAI,OAAO,WAAW,UAAU;EAC9B,MAAM,MAAM;EACZ,IAAI,cAAc,OAAO,OAAO,IAAI,aAAa,UAAU,OAAO,IAAI;EACtE,IAAI,UAAU,OAAO,OAAO,IAAI,SAAS,UAAU,OAAO,IAAI;CAChE;CAEA,OAAO;AACT;;;;;AAMA,SAAgB,gBAAgB,QAAiB,OAA0B;CACzE,IAAI,OAAO,OAAO;CAClB,IAAI,CAAC,QAAQ,OAAO;CAEpB,IAAI,OAAO,WAAW,UAAU;EAC9B,MAAM,MAAM;EACZ,MAAM,WAAY,IAAI,YAAoC,IAAI;EAC9D,IAAI,OAAO,aAAa,UAAU,OAAO,aAAa;CACxD;CAEA,IAAI,OAAO,WAAW,UAAU;EAC9B,MAAM,YAAY,OAAO,MAAM,8CAA8C;EAC7E,IAAI,WAAW,OAAO,UAAU,OAAO;CACzC;CAGA,OAAO;AACT;;;;;AAMA,SAAgB,cAAc,QAAgB,UAA2B;CACvE,IAAI,WAAW,UAAU,OAAO;CAChC,OAAO,OAAO,SAAS,IAAI,UAAU,KAAK,SAAS,SAAS,IAAI,QAAQ;AAC1E"}