@runtypelabs/sdk 1.9.1 → 1.10.0

package/dist/index.d.cts

@@ -3142,11 +3142,16 @@ interface RunTaskStateSlice {
     bestCandidateVerified?: boolean;
     verificationRequired?: boolean;
     lastVerificationPassed?: boolean;
+    consecutiveBlockedVerificationSessions?: number;
     isCreationTask?: boolean;
+    /** For creation tasks: allowed write root (e.g. "public/"). Writes must stay under this path. */
+    outputRoot?: string;
     sessions: Array<{
         actionKeys?: string[];
         hadTextOutput?: boolean;
         wroteFiles?: boolean;
+        verificationAttempted?: boolean;
+        verificationBlocked?: boolean;
     }>;
 }
 interface RunTaskToolTraceSlice {
@@ -3165,6 +3170,7 @@ interface RunTaskToolTraceSlice {
     bestCandidateVerified: boolean;
     verificationAttempted: boolean;
     verificationPassed: boolean;
+    verificationBlocked: boolean;
     localToolLoopGuardTriggered: boolean;
     forcedTurnEndReason?: string;
     bestCandidatePath?: string;
@@ -4197,6 +4203,8 @@ interface AgentExecuteRequest {
     debugMode?: boolean;
     /** Model ID to use for this session (overrides agent config) */
     model?: string;
+    /** Enable reasoning/thinking for models that support it (e.g. Gemini 3, o-series) */
+    reasoning?: boolean;
     /** Runtime tools to make available during execution */
     tools?: {
         runtimeTools?: AgentRuntimeToolDefinition[];
@@ -4298,6 +4306,8 @@ interface RunTaskSessionSummary {
     verificationAttempted?: boolean;
     /** Whether the latest verification command passed during the session */
     verificationPassed?: boolean;
+    /** Whether the verification command was blocked (e.g. unsafe command rejected) */
+    verificationBlocked?: boolean;
     /** Best candidate file identified during the session */
     bestCandidatePath?: string;
     /** Recent action keys used during the session */
@@ -4373,12 +4383,16 @@ interface RunTaskState {
     lastVerificationPassed?: boolean;
     /** Whether this task is creating something new rather than modifying existing files */
     isCreationTask?: boolean;
+    /** For creation tasks: allowed write root (e.g. "public/"). Writes must stay under this path. */
+    outputRoot?: string;
     /** Workflow variant string (e.g. 'create', 'modify', or custom) */
     workflowVariant?: string;
+    /** Number of consecutive sessions where verification was blocked */
+    consecutiveBlockedVerificationSessions?: number;
     /** Arbitrary bag for workflow-specific data */
     workflowState?: Record<string, unknown>;
 }
-type RunTaskResumeState = Pick<RunTaskState, 'originalMessage' | 'bootstrapContext' | 'workflowPhase' | 'planPath' | 'planWritten' | 'bestCandidatePath' | 'bestCandidateReason' | 'candidatePaths' | 'recentReadPaths' | 'recentActionKeys' | 'bestCandidateNeedsVerification' | 'bestCandidateVerified' | 'verificationRequired' | 'lastVerificationPassed' | 'isCreationTask' | 'workflowVariant' | 'workflowState'>;
+type RunTaskResumeState = Pick<RunTaskState, 'originalMessage' | 'bootstrapContext' | 'workflowPhase' | 'planPath' | 'planWritten' | 'bestCandidatePath' | 'bestCandidateReason' | 'candidatePaths' | 'recentReadPaths' | 'recentActionKeys' | 'bestCandidateNeedsVerification' | 'bestCandidateVerified' | 'verificationRequired' | 'lastVerificationPassed' | 'consecutiveBlockedVerificationSessions' | 'isCreationTask' | 'outputRoot' | 'workflowVariant' | 'workflowState'>;
 /**
  * Callback invoked after each session completes.
  * Return `false` to stop the loop early.
@@ -4455,6 +4469,8 @@ interface RunTaskOptions {
     localTools?: Record<string, LocalToolDefinition>;
     /** Model ID to use (overrides agent's configured model) */
     model?: string;
+    /** Enable reasoning/thinking for models that support it (e.g. Gemini 3, o-series) */
+    reasoning?: boolean;
     /** Previous messages from a prior run (for continuation/resume) */
     previousMessages?: AgentMessage[];
     /** New user message for continuation (appended after previous context) */

package/dist/index.d.ts

@@ -3142,11 +3142,16 @@ interface RunTaskStateSlice {
     bestCandidateVerified?: boolean;
     verificationRequired?: boolean;
     lastVerificationPassed?: boolean;
+    consecutiveBlockedVerificationSessions?: number;
     isCreationTask?: boolean;
+    /** For creation tasks: allowed write root (e.g. "public/"). Writes must stay under this path. */
+    outputRoot?: string;
     sessions: Array<{
         actionKeys?: string[];
         hadTextOutput?: boolean;
         wroteFiles?: boolean;
+        verificationAttempted?: boolean;
+        verificationBlocked?: boolean;
     }>;
 }
 interface RunTaskToolTraceSlice {
@@ -3165,6 +3170,7 @@ interface RunTaskToolTraceSlice {
     bestCandidateVerified: boolean;
     verificationAttempted: boolean;
     verificationPassed: boolean;
+    verificationBlocked: boolean;
     localToolLoopGuardTriggered: boolean;
     forcedTurnEndReason?: string;
     bestCandidatePath?: string;
@@ -4197,6 +4203,8 @@ interface AgentExecuteRequest {
     debugMode?: boolean;
     /** Model ID to use for this session (overrides agent config) */
     model?: string;
+    /** Enable reasoning/thinking for models that support it (e.g. Gemini 3, o-series) */
+    reasoning?: boolean;
     /** Runtime tools to make available during execution */
     tools?: {
         runtimeTools?: AgentRuntimeToolDefinition[];
@@ -4298,6 +4306,8 @@ interface RunTaskSessionSummary {
     verificationAttempted?: boolean;
     /** Whether the latest verification command passed during the session */
     verificationPassed?: boolean;
+    /** Whether the verification command was blocked (e.g. unsafe command rejected) */
+    verificationBlocked?: boolean;
     /** Best candidate file identified during the session */
     bestCandidatePath?: string;
     /** Recent action keys used during the session */
@@ -4373,12 +4383,16 @@ interface RunTaskState {
     lastVerificationPassed?: boolean;
     /** Whether this task is creating something new rather than modifying existing files */
     isCreationTask?: boolean;
+    /** For creation tasks: allowed write root (e.g. "public/"). Writes must stay under this path. */
+    outputRoot?: string;
     /** Workflow variant string (e.g. 'create', 'modify', or custom) */
     workflowVariant?: string;
+    /** Number of consecutive sessions where verification was blocked */
+    consecutiveBlockedVerificationSessions?: number;
     /** Arbitrary bag for workflow-specific data */
     workflowState?: Record<string, unknown>;
 }
-type RunTaskResumeState = Pick<RunTaskState, 'originalMessage' | 'bootstrapContext' | 'workflowPhase' | 'planPath' | 'planWritten' | 'bestCandidatePath' | 'bestCandidateReason' | 'candidatePaths' | 'recentReadPaths' | 'recentActionKeys' | 'bestCandidateNeedsVerification' | 'bestCandidateVerified' | 'verificationRequired' | 'lastVerificationPassed' | 'isCreationTask' | 'workflowVariant' | 'workflowState'>;
+type RunTaskResumeState = Pick<RunTaskState, 'originalMessage' | 'bootstrapContext' | 'workflowPhase' | 'planPath' | 'planWritten' | 'bestCandidatePath' | 'bestCandidateReason' | 'candidatePaths' | 'recentReadPaths' | 'recentActionKeys' | 'bestCandidateNeedsVerification' | 'bestCandidateVerified' | 'verificationRequired' | 'lastVerificationPassed' | 'consecutiveBlockedVerificationSessions' | 'isCreationTask' | 'outputRoot' | 'workflowVariant' | 'workflowState'>;
 /**
  * Callback invoked after each session completes.
  * Return `false` to stop the loop early.
@@ -4455,6 +4469,8 @@ interface RunTaskOptions {
     localTools?: Record<string, LocalToolDefinition>;
     /** Model ID to use (overrides agent's configured model) */
     model?: string;
+    /** Enable reasoning/thinking for models that support it (e.g. Gemini 3, o-series) */
+    reasoning?: boolean;
     /** Previous messages from a prior run (for continuation/resume) */
     previousMessages?: AgentMessage[];
     /** New user message for continuation (appended after previous context) */

package/dist/index.js

@@ -1896,7 +1896,7 @@ var TOOL_NAME_PATTERN = /^[A-Za-z][A-Za-z0-9_]{1,63}$/;
 var DEFAULT_MAX_CODE_LENGTH = 12e3;
 var DEFAULT_MAX_TIMEOUT_MS = 3e4;
 var DEFAULT_BLOCKED_CODE_PATTERNS = [
-  /\b(?:child_process|fs|net|tls|http|https|os)\b/i,
+  /\b(?:child_process|fs|net|tls|os)\b/i,
   /\b(?:process|Deno|Bun)\b/i,
   /\b(?:require|import)\s*\(/i,
   /\beval\s*\(/i,
@@ -2715,7 +2715,7 @@ var executionPhase = {
           `After that, you may update "${normalizedPlanPath}" with progress.`
         ].join(" ");
       }
-      if (normalizedPathArg && normalizedPathArg !== normalizedPlanPath) {
+      if (!ctx.state.isCreationTask && normalizedPathArg && normalizedPathArg !== normalizedPlanPath) {
         const allowedWriteTargets = new Set(
           [
             normalizedPlanPath,
@@ -2732,6 +2732,23 @@ var executionPhase = {
           ].join(" ");
         }
       }
+      if (ctx.state.isCreationTask && normalizedPathArg && normalizedPathArg !== normalizedPlanPath) {
+        const outputRoot = ctx.state.outputRoot ? ctx.state.outputRoot.trim().replace(/\\/g, "/").replace(/\/+/g, "/").replace(/\/$/, "") || void 0 : void 0;
+        if (!outputRoot) {
+          return [
+            `Blocked by marathon execution guard: creation tasks require outputRoot. Writes outside the plan are not allowed.`,
+            `Plan path: "${normalizedPlanPath}". Create files only under the configured output root.`
+          ].join(" ");
+        }
+        const rootPrefix = outputRoot + "/";
+        const isUnderRoot = normalizedPathArg === outputRoot || normalizedPathArg.startsWith(rootPrefix);
+        if (!isUnderRoot) {
+          return [
+            `Blocked by marathon execution guard: ${toolName} must target the plan or paths under outputRoot "${outputRoot}/".`,
+            `"${normalizedPathArg}" is outside the allowed output root.`
+          ].join(" ");
+        }
+      }
     }
     return void 0;
   },
@@ -2744,6 +2761,17 @@ var executionPhase = {
   },
   buildRecoveryMessage(state) {
     const recent = state.sessions.slice(-2);
+    if (recent.length >= 2 && recent.every(
+      (session) => session.verificationAttempted === true && session.wroteFiles !== true
+    )) {
+      return [
+        "Recovery instruction:",
+        "You have attempted verification in multiple sessions but none passed, and no files were written.",
+        "If the project lacks test/lint/build tooling, verification cannot succeed.",
+        "Focus on completing any remaining implementation. Signal TASK_COMPLETE when done.",
+        "Do not retry run_check unless you have a specific command likely to succeed."
+      ].join("\n");
+    }
     const normalizedPlanPath = typeof state.planPath === "string" && state.planPath.trim() ? normalizeCandidatePath(state.planPath) : void 0;
     const recentPlanOnlyLoop = Boolean(normalizedPlanPath) && recent.length === 2 && recent.every((session) => {
       const specificActionKeys = (session.actionKeys || []).map((actionKey) => actionKey.replace(/\\/g, "/")).filter((actionKey) => !actionKey.startsWith("server:"));
@@ -2786,6 +2814,57 @@ var executionPhase = {
 };
 function classifyVariant(message) {
   const lower = message.toLowerCase();
+  const modificationVerbs = [
+    "fix",
+    "update",
+    "change",
+    "modify",
+    "edit",
+    "refactor",
+    "improve",
+    "add to",
+    "remove",
+    "delete",
+    "rename",
+    "migrate"
+  ];
+  const hasModificationVerb = modificationVerbs.some(
+    (v) => lower.startsWith(v) || new RegExp(`\\b${v}\\b`).test(lower)
+  );
+  const creationPatterns = [
+    /^create\b/,
+    /^build\b/,
+    /^make\b/,
+    /^generate\b/,
+    /^scaffold\b/,
+    /^set up\b/,
+    /^setup\b/,
+    /^bootstrap\b/,
+    /^initialize\b/,
+    /^init\b/,
+    /^write a\b/,
+    /^write an\b/,
+    /^implement a\b/,
+    /^implement an\b/,
+    /^start a\b/,
+    /^start an\b/,
+    /^new\b/
+  ];
+  const hasCreationStart = creationPatterns.some((p) => p.test(lower));
+  const creationVerbs = [
+    "build",
+    "create",
+    "make",
+    "generate",
+    "scaffold",
+    "implement"
+  ];
+  const hasCreationVerb = creationVerbs.some(
+    (v) => new RegExp(`\\b${v}\\b`).test(lower)
+  );
+  if ((hasCreationStart || hasCreationVerb) && !hasModificationVerb) {
+    return "create";
+  }
   const externalVerbs = [
     "fetch",
     "browse",
@@ -2825,49 +2904,9 @@ function classifyVariant(message) {
   ];
   const hasExternalVerb = externalVerbs.some((v) => lower.includes(v));
   const hasExternalTarget = externalTargets.some((t) => lower.includes(t));
-  const modificationVerbs = [
-    "fix",
-    "update",
-    "change",
-    "modify",
-    "edit",
-    "refactor",
-    "improve",
-    "add to",
-    "remove",
-    "delete",
-    "rename",
-    "migrate"
-  ];
-  const hasModificationVerb = modificationVerbs.some(
-    (v) => lower.startsWith(v) || new RegExp(`\\b${v}\\b`).test(lower)
-  );
   if (hasExternalVerb && hasExternalTarget && !hasModificationVerb) {
     return "external";
   }
-  const creationPatterns = [
-    /^create\b/,
-    /^build\b/,
-    /^make\b/,
-    /^generate\b/,
-    /^scaffold\b/,
-    /^set up\b/,
-    /^setup\b/,
-    /^bootstrap\b/,
-    /^initialize\b/,
-    /^init\b/,
-    /^write a\b/,
-    /^write an\b/,
-    /^implement a\b/,
-    /^implement an\b/,
-    /^start a\b/,
-    /^start an\b/,
-    /^new\b/
-  ];
-  const hasCreationStart = creationPatterns.some((p) => p.test(lower));
-  if (hasCreationStart && !hasModificationVerb) {
-    return "create";
-  }
   return "modify";
 }
 async function generateBootstrapContext(message, localTools, variant) {
@@ -4705,6 +4744,7 @@ var _AgentsEndpoint = class _AgentsEndpoint {
       bestCandidateVerified: false,
       verificationAttempted: false,
       verificationPassed: false,
+      verificationBlocked: false,
       localToolLoopGuardTriggered: false
     };
   }
@@ -4926,6 +4966,7 @@ var _AgentsEndpoint = class _AgentsEndpoint {
       if (typeof parsed.success !== "boolean") return void 0;
       return {
         success: parsed.success,
+        ...typeof parsed.blocked === "boolean" ? { blocked: parsed.blocked } : {},
         ...typeof parsed.command === "string" ? { command: parsed.command } : {},
         ...typeof parsed.output === "string" ? { output: parsed.output } : {},
         ...typeof parsed.error === "string" ? { error: parsed.error } : {}
@@ -5153,9 +5194,11 @@ var _AgentsEndpoint = class _AgentsEndpoint {
       bestCandidateVerified: Boolean(resumeState.bestCandidateVerified),
       ...resumeState.verificationRequired !== void 0 ? { verificationRequired: resumeState.verificationRequired } : {},
       lastVerificationPassed: Boolean(resumeState.lastVerificationPassed),
+      ...resumeState.consecutiveBlockedVerificationSessions !== void 0 ? { consecutiveBlockedVerificationSessions: resumeState.consecutiveBlockedVerificationSessions } : {},
       ...resumeState.isCreationTask !== void 0 ? { isCreationTask: resumeState.isCreationTask } : {},
       ...resumeState.workflowVariant !== void 0 ? { workflowVariant: resumeState.workflowVariant } : {},
-      ...resumeState.workflowState !== void 0 ? { workflowState: resumeState.workflowState } : {}
+      ...resumeState.workflowState !== void 0 ? { workflowState: resumeState.workflowState } : {},
+      ...typeof resumeState.outputRoot === "string" && resumeState.outputRoot.trim() ? { outputRoot: resumeState.outputRoot.trim().replace(/\\/g, "/").replace(/\/+/g, "/") } : {}
     };
   }
   buildPhaseInstructions(state, workflow) {
@@ -5282,6 +5325,10 @@ var _AgentsEndpoint = class _AgentsEndpoint {
             } else if (state.workflowPhase === "execution") {
               trace.executionFileWritten = true;
               trace.verificationPassed = false;
+              if (!this.isMarathonArtifactPath(normalizedPathArg)) {
+                trace.bestCandidatePath = normalizedPathArg;
+                trace.bestCandidateReason = "written by agent during execution";
+              }
               if (normalizedBestCandidatePath && normalizedPathArg === normalizedBestCandidatePath) {
                 trace.bestCandidateWritten = true;
               }
@@ -5291,6 +5338,9 @@ var _AgentsEndpoint = class _AgentsEndpoint {
           if (verificationResult) {
             trace.verificationAttempted = true;
             trace.verificationPassed = verificationResult.success;
+            if (verificationResult.blocked) {
+              trace.verificationBlocked = true;
+            }
           }
           const summarizedResult = verificationResult ? [
             verificationResult.command || "verification",
@@ -5613,6 +5663,7 @@ var _AgentsEndpoint = class _AgentsEndpoint {
     };
     state.workflowVariant = classifiedVariant;
     state.isCreationTask = seededResumeState?.isCreationTask ?? state.workflowVariant === "create";
+    state.outputRoot = seededResumeState?.outputRoot ?? (state.isCreationTask ? "public/" : void 0);
     this.updateWorkflowPhase(state, this.createEmptyToolTrace(), workflow);
     let recordId;
     const localToolNames = options.localTools ? Object.keys(options.localTools) : void 0;
@@ -5685,6 +5736,7 @@ var _AgentsEndpoint = class _AgentsEndpoint {
         messages,
         debugMode: options.debugMode,
         model: options.model,
+        ...options.reasoning !== void 0 ? { reasoning: options.reasoning } : {},
         ...options.toolIds?.length ? { tools: { toolIds: options.toolIds } } : {},
         ...requestContextManagement ? { contextManagement: requestContextManagement } : {}
       };
@@ -5783,6 +5835,7 @@ var _AgentsEndpoint = class _AgentsEndpoint {
         hadTextOutput: Boolean(sessionResult.result.trim()),
         verificationAttempted: sessionTrace.verificationAttempted,
         verificationPassed: sessionTrace.verificationPassed,
+        verificationBlocked: sessionTrace.verificationBlocked || void 0,
         bestCandidatePath: sessionTrace.bestCandidatePath || void 0,
         actionKeys: sessionTrace.actionKeys.slice(-5),
         completedAt: (/* @__PURE__ */ new Date()).toISOString()
@@ -5823,6 +5876,21 @@ var _AgentsEndpoint = class _AgentsEndpoint {
       }
       if (sessionTrace.verificationAttempted) {
         state.lastVerificationPassed = sessionTrace.verificationPassed;
+        if (sessionTrace.verificationBlocked && !sessionTrace.verificationPassed) {
+          state.consecutiveBlockedVerificationSessions = (state.consecutiveBlockedVerificationSessions || 0) + 1;
+        } else {
+          state.consecutiveBlockedVerificationSessions = 0;
+        }
+      }
+      if ((state.consecutiveBlockedVerificationSessions || 0) >= 2 && state.verificationRequired) {
+        state.verificationRequired = false;
+        state.lastVerificationPassed = true;
+        if (!state.planWritten) {
+          state.planWritten = true;
+        }
+        if (!state.bestCandidateVerified) {
+          state.bestCandidateVerified = true;
+        }
       }
       const modelKey = options.model || "default";
       if (!state.costByModel) state.costByModel = {};
@@ -6401,9 +6469,12 @@ Do NOT redo any of the above work.`
       const replayHistoryMessages = this.sanitizeReplayHistoryMessages(
         continuationContext.previousMessages
       );
+      const continuationGuardrail = "IMPORTANT: You are continuing a previously completed task. The conversation above shows your prior work. Do NOT redo any of it. Build on what was already accomplished. If there is nothing new to do, respond with TASK_COMPLETE.";
       const defaultContinueMessage = "Continue the task. Review your prior work above and proceed with any remaining work. If everything is already complete, respond with TASK_COMPLETE.";
       const userMessage = continuationContext.newUserMessage || defaultContinueMessage;
       const userContent = [
+        continuationGuardrail,
+        "",
         userMessage,
         phaseBlock,
         toolsBlock,
@@ -6414,10 +6485,6 @@ Do NOT redo any of the above work.`
       ].join("\n");
       const fullHistoryMessages = [
         ...replayHistoryMessages,
-        {
-          role: "system",
-          content: "IMPORTANT: You are continuing a previously completed task. The conversation above shows your prior work. Do NOT redo any of it. Build on what was already accomplished. If there is nothing new to do, respond with TASK_COMPLETE."
-        },
         {
           role: "user",
           content: userContent