mstro-app 0.1.58 → 0.2.0

package/server/cli/headless/index.ts

@@ -26,14 +26,20 @@ export {
 } from './prompt-utils.js';
 // Main runner class
 export { HeadlessRunner } from './runner.js';
+// Tool watchdog
+export { ToolWatchdog } from './tool-watchdog.js';
 // Types
 export type {
+  ExecutionCheckpoint,
   ExecutionResult,
   HeadlessConfig,
   ImageAttachment,
-  ResolvedHeadlessConfig, 
+  PendingToolMap,
+  ResolvedHeadlessConfig,
   SessionResult,
   SessionState,
+  ToolDurationTracker,
+  ToolTimeoutProfile,
   ToolUseAccumulator,
   ToolUseEvent
 } from './types.js';

package/server/cli/headless/prompt-utils.ts

@@ -7,16 +7,19 @@
  * Utilities for enriching prompts with context from previous conversation.
  */
 
+import { assessApproval } from './stall-assessor.js';
 import type { ImageAttachment, PromptContext } from './types.js';
 
 /**
- * Enrich prompt with context from previous conversation
+ * Enrich prompt with context from previous conversation.
+ * Async because ambiguous short prompts are classified by Haiku.
  */
-export function enrichPromptWithContext(prompt: string, context: PromptContext): string {
+export async function enrichPromptWithContext(prompt: string, context: PromptContext): Promise<string> {
   let enriched = prompt;
 
-  // Detect if this is a continuation/approval prompt
-  const isApprovalOrContinuation = isApprovalPrompt(prompt);
+  // Detect if this is a continuation/approval prompt.
+  // Fast regex path for obvious approvals, Haiku for ambiguous short prompts.
+  const isApprovalOrContinuation = await detectApproval(prompt);
 
   // Add accumulated knowledge from previous prompts
   if (context.accumulatedKnowledge) {
@@ -36,9 +39,38 @@ export function enrichPromptWithContext(prompt: string, context: PromptContext):
 }
 
 /**
- * Detect if a prompt is an approval or continuation
+ * Detect if a prompt is an approval or continuation.
+ * Layer 1: Regex fast path for obvious approvals (free, sync).
+ * Layer 2: Haiku assessment for ambiguous short prompts (<100 chars).
+ */
+async function detectApproval(prompt: string): Promise<boolean> {
+  // Layer 1: fast regex path
+  if (isApprovalPromptFast(prompt)) return true;
+
+  // Layer 2: Haiku for short ambiguous prompts.
+  // Long prompts (>100 chars) are almost certainly new tasks, not approvals.
+  if (prompt.trim().length <= 100) {
+    try {
+      const claudeCmd = process.env.CLAUDE_COMMAND || 'claude';
+      const verdict = await assessApproval(prompt, claudeCmd, false);
+      return verdict.isApproval;
+    } catch {
+      // Haiku failed — fall through to false
+    }
+  }
+
+  return false;
+}
+
+/**
+ * Fast regex-based approval detection (sync, no API call).
+ * Catches obvious affirmatives. Ambiguous cases fall through to Haiku.
  */
 export function isApprovalPrompt(prompt: string): boolean {
+  return isApprovalPromptFast(prompt);
+}
+
+function isApprovalPromptFast(prompt: string): boolean {
   const lower = prompt.toLowerCase().trim();
 
   // Short affirmative responses

package/server/cli/headless/runner.ts

@@ -20,7 +20,7 @@ import type {
 } from './types.js';
 
 // Re-export types for backward compatibility
-export type { HeadlessConfig, ImageAttachment, SessionResult, SessionState, ToolUseEvent } from './types.js';
+export type { ExecutionCheckpoint, HeadlessConfig, ImageAttachment, SessionResult, SessionState, ToolTimeoutProfile, ToolUseEvent } from './types.js';
 
 export class HeadlessRunner {
   private config: ResolvedHeadlessConfig;
@@ -33,9 +33,9 @@ export class HeadlessRunner {
       maxSessions: config.maxSessions || 50,
       maxRetries: config.maxRetries || 3,
       claudeCommand: config.claudeCommand || process.env.CLAUDE_COMMAND || 'claude',
-      verbose: config.verbose || false,
-      noColor: config.noColor || false,
-      improvisationMode: config.improvisationMode || false,
+      verbose: !!config.verbose,
+      noColor: !!config.noColor,
+      improvisationMode: !!config.improvisationMode,
       movementNumber: config.movementNumber ?? 0,
       outputCallback: config.outputCallback,
       thinkingCallback: config.thinkingCallback,
@@ -51,6 +51,11 @@ export class HeadlessRunner {
       stallMaxExtensions: config.stallMaxExtensions ?? 3,
       stallHardCapMs: config.stallHardCapMs ?? 3_600_000,
       model: config.model,
+      toolTimeoutProfiles: config.toolTimeoutProfiles,
+      enableToolWatchdog: config.enableToolWatchdog !== false,
+      maxAutoRetries: config.maxAutoRetries ?? 2,
+      onToolTimeout: config.onToolTimeout,
+      sandboxed: config.sandboxed,
     };
   }
 
@@ -78,22 +83,36 @@ export class HeadlessRunner {
     const sessionId = `direct-${Date.now()}`;
 
     const enrichedPrompt = context
-      ? enrichPromptWithContext(userPrompt, context)
+      ? await enrichPromptWithContext(userPrompt, context)
       : userPrompt;
 
     const result = await this.executePromptCommand(enrichedPrompt, 'main', 1);
 
     if (result.exitCode !== 0) {
+      // Build meaningful error: prefer stderr, fall back to non-JSON stdout lines
+      let errorMessage = result.error;
+      if (!errorMessage && result.output) {
+        const plainLines = result.output.split('\n')
+          .filter(l => l.trim() && !l.trim().startsWith('{'))
+          .join('\n')
+          .trim();
+        if (plainLines) {
+          errorMessage = plainLines.slice(0, 500);
+        }
+      }
       return {
         completed: false,
         needsHandoff: false,
         totalTokens: 0,
         sessionId,
-        error: result.error || 'Execution failed',
+        error: errorMessage || `Claude exited with code ${result.exitCode}`,
         assistantResponse: result.assistantResponse,
         thinkingOutput: result.thinkingOutput,
         toolUseHistory: result.toolUseHistory,
-        claudeSessionId: result.claudeSessionId
+        claudeSessionId: result.claudeSessionId,
+        nativeTimeoutCount: result.nativeTimeoutCount,
+        postTimeoutOutput: result.postTimeoutOutput,
+        resumeBufferedOutput: result.resumeBufferedOutput,
       };
     }
 
@@ -107,7 +126,10 @@ export class HeadlessRunner {
       assistantResponse: result.assistantResponse,
       thinkingOutput: result.thinkingOutput,
       toolUseHistory: result.toolUseHistory,
-      claudeSessionId: result.claudeSessionId
+      claudeSessionId: result.claudeSessionId,
+      nativeTimeoutCount: result.nativeTimeoutCount,
+      postTimeoutOutput: result.postTimeoutOutput,
+      resumeBufferedOutput: result.resumeBufferedOutput,
     };
   }