@probelabs/probe 0.6.0-rc230 → 0.6.0-rc231

package/build/agent/ProbeAgent.js

@@ -2496,10 +2496,9 @@ ${extractGuidance}
       toolDefinitions += `${taskToolDefinition}\n`;
     }
 
-    // Always include attempt_completion (unless explicitly disabled in raw AI mode)
-    if (isToolAllowed('attempt_completion')) {
-      toolDefinitions += `${attemptCompletionToolDefinition}\n`;
-    }
+    // Always include attempt_completion unconditionally - it's a completion signal, not a tool
+    // This ensures agents can always complete their work, regardless of tool restrictions
+    toolDefinitions += `${attemptCompletionToolDefinition}\n`;
 
     // Delegate tool (require both enableDelegate flag AND allowedTools permission)
     // Place after attempt_completion as it's an optional tool
@@ -3304,8 +3303,9 @@ Follow these instructions carefully:
           if (this.enableSkills && this.allowedTools.isEnabled('listSkills')) validTools.push('listSkills');
           if (this.enableSkills && this.allowedTools.isEnabled('useSkill')) validTools.push('useSkill');
           if (this.allowedTools.isEnabled('readImage')) validTools.push('readImage');
-          // Always allow attempt_completion - it's a completion signal, not a tool
+          // Always allow attempt_completion in validTools - it's a completion signal, not a tool
           // This ensures agents can complete even when disableTools: true is set (fixes #333)
+          // The tool DEFINITION may be hidden in raw AI mode, but we still need to recognize it
           validTools.push('attempt_completion');
 
           // Edit tools (require both allowEdit flag AND allowedTools permission)

package/build/agent/index.js

@@ -9199,7 +9199,15 @@ function createTaskTool(options = {}) {
           });
           return `Error: Invalid task parameters - ${validation.error.message}`;
         }
-        const { action, tasks, id, title, description, status, priority, dependencies, after } = validation.data;
+        const { action, tasks: rawTasks, id, title, description, status, priority, dependencies, after } = validation.data;
+        let tasks = rawTasks;
+        if (typeof rawTasks === "string") {
+          try {
+            tasks = JSON.parse(rawTasks);
+          } catch (e) {
+            return `Error: Invalid tasks JSON - ${e.message}`;
+          }
+        }
         switch (action) {
           case "create": {
             if (tasks && Array.isArray(tasks)) {
@@ -9374,7 +9382,8 @@ var init_taskTool = __esm({
     });
     taskSchema = external_exports.object({
       action: external_exports.enum(["create", "update", "complete", "delete", "list"]),
-      tasks: external_exports.array(external_exports.union([external_exports.string(), taskItemSchema])).optional(),
+      // Accept both array and JSON string (AI models sometimes serialize as string)
+      tasks: external_exports.union([external_exports.array(external_exports.union([external_exports.string(), taskItemSchema])), external_exports.string()]).optional(),
       id: external_exports.string().optional(),
       title: external_exports.string().optional(),
       description: external_exports.string().optional(),
@@ -9485,6 +9494,25 @@ SKIP TASKS for single-goal requests, even if they require multiple searches:
 **Key insight**: Multiple *internal steps* (search, read, analyze) are NOT the same as multiple *goals*.
 A single investigation with many steps is still ONE task, not many.
 
+## Task Granularity
+
+Tasks represent LOGICAL UNITS OF WORK, not individual files or steps:
+- "Fix 8 similar test files" \u2192 ONE task (same type of fix across files)
+- "Update API + tests + docs" \u2192 THREE tasks (different types of work)
+- "Implement feature in 5 files" \u2192 ONE task (single feature)
+
+**Rule of thumb**: If you're creating more than 3-4 tasks, you're probably too granular.
+
+**Anti-patterns to avoid**:
+- One task per file \u274C
+- One task per function \u274C
+- One task per repository (when same type of work) \u274C
+
+**Good patterns**:
+- One task per distinct deliverable \u2713
+- One task per phase (implement, test, document) \u2713
+- One task per different type of work \u2713
+
 MODIFY TASKS when (during execution):
 - You discover the problem is more complex than expected \u2192 Add new tasks
 - A single task covers too much scope \u2192 Split into smaller tasks
@@ -55823,6 +55851,7 @@ var require_pattern = __commonJS({
     "use strict";
     Object.defineProperty(exports2, "__esModule", { value: true });
     var code_1 = require_code2();
+    var util_1 = require_util3();
     var codegen_1 = require_codegen();
     var error = {
       message: ({ schemaCode }) => (0, codegen_1.str)`must match pattern "${schemaCode}"`,
@@ -55835,10 +55864,18 @@ var require_pattern = __commonJS({
       $data: true,
       error,
       code(cxt) {
-        const { data, $data, schema, schemaCode, it } = cxt;
+        const { gen, data, $data, schema, schemaCode, it } = cxt;
         const u = it.opts.unicodeRegExp ? "u" : "";
-        const regExp = $data ? (0, codegen_1._)`(new RegExp(${schemaCode}, ${u}))` : (0, code_1.usePattern)(cxt, schema);
-        cxt.fail$data((0, codegen_1._)`!${regExp}.test(${data})`);
+        if ($data) {
+          const { regExp } = it.opts.code;
+          const regExpCode = regExp.code === "new RegExp" ? (0, codegen_1._)`new RegExp` : (0, util_1.useFunc)(gen, regExp);
+          const valid = gen.let("valid");
+          gen.try(() => gen.assign(valid, (0, codegen_1._)`${regExpCode}(${schemaCode}, ${u}).test(${data})`), () => gen.assign(valid, false));
+          cxt.fail$data((0, codegen_1._)`!${valid}`);
+        } else {
+          const regExp = (0, code_1.usePattern)(cxt, schema);
+          cxt.fail$data((0, codegen_1._)`!${regExp}.test(${data})`);
+        }
       }
     };
     exports2.default = def;
@@ -72250,10 +72287,8 @@ Workspace: ${this.allowedFolders.join(", ")}`;
           toolDefinitions += `${taskToolDefinition}
 `;
         }
-        if (isToolAllowed("attempt_completion")) {
-          toolDefinitions += `${attemptCompletionToolDefinition}
+        toolDefinitions += `${attemptCompletionToolDefinition}
 `;
-        }
         if (this.enableDelegate && isToolAllowed("delegate")) {
           toolDefinitions += `${delegateToolDefinition}
 `;

package/build/agent/tasks/taskTool.js

@@ -23,7 +23,8 @@ export const taskItemSchema = z.object({
  */
 export const taskSchema = z.object({
   action: z.enum(['create', 'update', 'complete', 'delete', 'list']),
-  tasks: z.array(z.union([z.string(), taskItemSchema])).optional(),
+  // Accept both array and JSON string (AI models sometimes serialize as string)
+  tasks: z.union([z.array(z.union([z.string(), taskItemSchema])), z.string()]).optional(),
   id: z.string().optional(),
   title: z.string().optional(),
   description: z.string().optional(),
@@ -142,6 +143,25 @@ SKIP TASKS for single-goal requests, even if they require multiple searches:
 **Key insight**: Multiple *internal steps* (search, read, analyze) are NOT the same as multiple *goals*.
 A single investigation with many steps is still ONE task, not many.
 
+## Task Granularity
+
+Tasks represent LOGICAL UNITS OF WORK, not individual files or steps:
+- "Fix 8 similar test files" → ONE task (same type of fix across files)
+- "Update API + tests + docs" → THREE tasks (different types of work)
+- "Implement feature in 5 files" → ONE task (single feature)
+
+**Rule of thumb**: If you're creating more than 3-4 tasks, you're probably too granular.
+
+**Anti-patterns to avoid**:
+- One task per file ❌
+- One task per function ❌
+- One task per repository (when same type of work) ❌
+
+**Good patterns**:
+- One task per distinct deliverable ✓
+- One task per phase (implement, test, document) ✓
+- One task per different type of work ✓
+
 MODIFY TASKS when (during execution):
 - You discover the problem is more complex than expected → Add new tasks
 - A single task covers too much scope → Split into smaller tasks
@@ -314,7 +334,17 @@ export function createTaskTool(options = {}) {
           return `Error: Invalid task parameters - ${validation.error.message}`;
         }
 
-        const { action, tasks, id, title, description, status, priority, dependencies, after } = validation.data;
+        const { action, tasks: rawTasks, id, title, description, status, priority, dependencies, after } = validation.data;
+
+        // Parse tasks if passed as JSON string (common AI model behavior)
+        let tasks = rawTasks;
+        if (typeof rawTasks === 'string') {
+          try {
+            tasks = JSON.parse(rawTasks);
+          } catch (e) {
+            return `Error: Invalid tasks JSON - ${e.message}`;
+          }
+        }
 
         switch (action) {
           case 'create': {

package/cjs/agent/ProbeAgent.cjs

@@ -36310,7 +36310,15 @@ function createTaskTool(options = {}) {
           });
           return `Error: Invalid task parameters - ${validation.error.message}`;
         }
-        const { action, tasks, id, title, description, status, priority, dependencies, after } = validation.data;
+        const { action, tasks: rawTasks, id, title, description, status, priority, dependencies, after } = validation.data;
+        let tasks = rawTasks;
+        if (typeof rawTasks === "string") {
+          try {
+            tasks = JSON.parse(rawTasks);
+          } catch (e4) {
+            return `Error: Invalid tasks JSON - ${e4.message}`;
+          }
+        }
         switch (action) {
           case "create": {
             if (tasks && Array.isArray(tasks)) {
@@ -36485,7 +36493,8 @@ var init_taskTool = __esm({
     });
     taskSchema = external_exports.object({
       action: external_exports.enum(["create", "update", "complete", "delete", "list"]),
-      tasks: external_exports.array(external_exports.union([external_exports.string(), taskItemSchema])).optional(),
+      // Accept both array and JSON string (AI models sometimes serialize as string)
+      tasks: external_exports.union([external_exports.array(external_exports.union([external_exports.string(), taskItemSchema])), external_exports.string()]).optional(),
       id: external_exports.string().optional(),
       title: external_exports.string().optional(),
       description: external_exports.string().optional(),
@@ -36596,6 +36605,25 @@ SKIP TASKS for single-goal requests, even if they require multiple searches:
 **Key insight**: Multiple *internal steps* (search, read, analyze) are NOT the same as multiple *goals*.
 A single investigation with many steps is still ONE task, not many.
 
+## Task Granularity
+
+Tasks represent LOGICAL UNITS OF WORK, not individual files or steps:
+- "Fix 8 similar test files" \u2192 ONE task (same type of fix across files)
+- "Update API + tests + docs" \u2192 THREE tasks (different types of work)
+- "Implement feature in 5 files" \u2192 ONE task (single feature)
+
+**Rule of thumb**: If you're creating more than 3-4 tasks, you're probably too granular.
+
+**Anti-patterns to avoid**:
+- One task per file \u274C
+- One task per function \u274C
+- One task per repository (when same type of work) \u274C
+
+**Good patterns**:
+- One task per distinct deliverable \u2713
+- One task per phase (implement, test, document) \u2713
+- One task per different type of work \u2713
+
 MODIFY TASKS when (during execution):
 - You discover the problem is more complex than expected \u2192 Add new tasks
 - A single task covers too much scope \u2192 Split into smaller tasks
@@ -82501,6 +82529,7 @@ var require_pattern = __commonJS({
     "use strict";
     Object.defineProperty(exports2, "__esModule", { value: true });
     var code_1 = require_code2();
+    var util_1 = require_util3();
     var codegen_1 = require_codegen();
     var error2 = {
       message: ({ schemaCode }) => (0, codegen_1.str)`must match pattern "${schemaCode}"`,
@@ -82513,10 +82542,18 @@ var require_pattern = __commonJS({
       $data: true,
       error: error2,
       code(cxt) {
-        const { data: data2, $data, schema, schemaCode, it } = cxt;
+        const { gen, data: data2, $data, schema, schemaCode, it } = cxt;
         const u4 = it.opts.unicodeRegExp ? "u" : "";
-        const regExp = $data ? (0, codegen_1._)`(new RegExp(${schemaCode}, ${u4}))` : (0, code_1.usePattern)(cxt, schema);
-        cxt.fail$data((0, codegen_1._)`!${regExp}.test(${data2})`);
+        if ($data) {
+          const { regExp } = it.opts.code;
+          const regExpCode = regExp.code === "new RegExp" ? (0, codegen_1._)`new RegExp` : (0, util_1.useFunc)(gen, regExp);
+          const valid = gen.let("valid");
+          gen.try(() => gen.assign(valid, (0, codegen_1._)`${regExpCode}(${schemaCode}, ${u4}).test(${data2})`), () => gen.assign(valid, false));
+          cxt.fail$data((0, codegen_1._)`!${valid}`);
+        } else {
+          const regExp = (0, code_1.usePattern)(cxt, schema);
+          cxt.fail$data((0, codegen_1._)`!${regExp}.test(${data2})`);
+        }
       }
     };
     exports2.default = def;
@@ -98927,10 +98964,8 @@ Workspace: ${this.allowedFolders.join(", ")}`;
           toolDefinitions += `${taskToolDefinition}
 `;
         }
-        if (isToolAllowed("attempt_completion")) {
-          toolDefinitions += `${attemptCompletionToolDefinition}
+        toolDefinitions += `${attemptCompletionToolDefinition}
 `;
-        }
         if (this.enableDelegate && isToolAllowed("delegate")) {
           toolDefinitions += `${delegateToolDefinition}
 `;

package/cjs/index.cjs

@@ -35480,7 +35480,15 @@ function createTaskTool(options = {}) {
           });
           return `Error: Invalid task parameters - ${validation.error.message}`;
         }
-        const { action, tasks, id, title, description, status, priority, dependencies, after } = validation.data;
+        const { action, tasks: rawTasks, id, title, description, status, priority, dependencies, after } = validation.data;
+        let tasks = rawTasks;
+        if (typeof rawTasks === "string") {
+          try {
+            tasks = JSON.parse(rawTasks);
+          } catch (e4) {
+            return `Error: Invalid tasks JSON - ${e4.message}`;
+          }
+        }
         switch (action) {
           case "create": {
             if (tasks && Array.isArray(tasks)) {
@@ -35655,7 +35663,8 @@ var init_taskTool = __esm({
     });
     taskSchema = external_exports.object({
       action: external_exports.enum(["create", "update", "complete", "delete", "list"]),
-      tasks: external_exports.array(external_exports.union([external_exports.string(), taskItemSchema])).optional(),
+      // Accept both array and JSON string (AI models sometimes serialize as string)
+      tasks: external_exports.union([external_exports.array(external_exports.union([external_exports.string(), taskItemSchema])), external_exports.string()]).optional(),
       id: external_exports.string().optional(),
       title: external_exports.string().optional(),
       description: external_exports.string().optional(),
@@ -35766,6 +35775,25 @@ SKIP TASKS for single-goal requests, even if they require multiple searches:
 **Key insight**: Multiple *internal steps* (search, read, analyze) are NOT the same as multiple *goals*.
 A single investigation with many steps is still ONE task, not many.
 
+## Task Granularity
+
+Tasks represent LOGICAL UNITS OF WORK, not individual files or steps:
+- "Fix 8 similar test files" \u2192 ONE task (same type of fix across files)
+- "Update API + tests + docs" \u2192 THREE tasks (different types of work)
+- "Implement feature in 5 files" \u2192 ONE task (single feature)
+
+**Rule of thumb**: If you're creating more than 3-4 tasks, you're probably too granular.
+
+**Anti-patterns to avoid**:
+- One task per file \u274C
+- One task per function \u274C
+- One task per repository (when same type of work) \u274C
+
+**Good patterns**:
+- One task per distinct deliverable \u2713
+- One task per phase (implement, test, document) \u2713
+- One task per different type of work \u2713
+
 MODIFY TASKS when (during execution):
 - You discover the problem is more complex than expected \u2192 Add new tasks
 - A single task covers too much scope \u2192 Split into smaller tasks
@@ -79234,6 +79262,7 @@ var require_pattern = __commonJS({
     "use strict";
     Object.defineProperty(exports2, "__esModule", { value: true });
     var code_1 = require_code2();
+    var util_1 = require_util3();
     var codegen_1 = require_codegen();
     var error2 = {
       message: ({ schemaCode }) => (0, codegen_1.str)`must match pattern "${schemaCode}"`,
@@ -79246,10 +79275,18 @@ var require_pattern = __commonJS({
       $data: true,
       error: error2,
       code(cxt) {
-        const { data: data2, $data, schema, schemaCode, it } = cxt;
+        const { gen, data: data2, $data, schema, schemaCode, it } = cxt;
         const u4 = it.opts.unicodeRegExp ? "u" : "";
-        const regExp = $data ? (0, codegen_1._)`(new RegExp(${schemaCode}, ${u4}))` : (0, code_1.usePattern)(cxt, schema);
-        cxt.fail$data((0, codegen_1._)`!${regExp}.test(${data2})`);
+        if ($data) {
+          const { regExp } = it.opts.code;
+          const regExpCode = regExp.code === "new RegExp" ? (0, codegen_1._)`new RegExp` : (0, util_1.useFunc)(gen, regExp);
+          const valid = gen.let("valid");
+          gen.try(() => gen.assign(valid, (0, codegen_1._)`${regExpCode}(${schemaCode}, ${u4}).test(${data2})`), () => gen.assign(valid, false));
+          cxt.fail$data((0, codegen_1._)`!${valid}`);
+        } else {
+          const regExp = (0, code_1.usePattern)(cxt, schema);
+          cxt.fail$data((0, codegen_1._)`!${regExp}.test(${data2})`);
+        }
       }
     };
     exports2.default = def;
@@ -95660,10 +95697,8 @@ Workspace: ${this.allowedFolders.join(", ")}`;
           toolDefinitions += `${taskToolDefinition}
 `;
         }
-        if (isToolAllowed("attempt_completion")) {
-          toolDefinitions += `${attemptCompletionToolDefinition}
+        toolDefinitions += `${attemptCompletionToolDefinition}
 `;
-        }
         if (this.enableDelegate && isToolAllowed("delegate")) {
           toolDefinitions += `${delegateToolDefinition}
 `;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@probelabs/probe",
-  "version": "0.6.0-rc230",
+  "version": "0.6.0-rc231",
   "description": "Node.js wrapper for the probe code search tool",
   "main": "src/index.js",
   "module": "src/index.js",
@@ -79,7 +79,7 @@
     "@ai-sdk/openai": "^2.0.10",
     "@anthropic-ai/claude-agent-sdk": "^0.1.46",
     "@modelcontextprotocol/sdk": "^1.0.0",
-    "@probelabs/maid": "^0.0.23",
+    "@probelabs/maid": "^0.0.24",
     "adm-zip": "^0.5.16",
     "ai": "^5.0.0",
     "ajv": "^8.17.1",

package/src/agent/ProbeAgent.js

@@ -2496,10 +2496,9 @@ ${extractGuidance}
       toolDefinitions += `${taskToolDefinition}\n`;
     }
 
-    // Always include attempt_completion (unless explicitly disabled in raw AI mode)
-    if (isToolAllowed('attempt_completion')) {
-      toolDefinitions += `${attemptCompletionToolDefinition}\n`;
-    }
+    // Always include attempt_completion unconditionally - it's a completion signal, not a tool
+    // This ensures agents can always complete their work, regardless of tool restrictions
+    toolDefinitions += `${attemptCompletionToolDefinition}\n`;
 
     // Delegate tool (require both enableDelegate flag AND allowedTools permission)
     // Place after attempt_completion as it's an optional tool
@@ -3304,8 +3303,9 @@ Follow these instructions carefully:
           if (this.enableSkills && this.allowedTools.isEnabled('listSkills')) validTools.push('listSkills');
           if (this.enableSkills && this.allowedTools.isEnabled('useSkill')) validTools.push('useSkill');
           if (this.allowedTools.isEnabled('readImage')) validTools.push('readImage');
-          // Always allow attempt_completion - it's a completion signal, not a tool
+          // Always allow attempt_completion in validTools - it's a completion signal, not a tool
           // This ensures agents can complete even when disableTools: true is set (fixes #333)
+          // The tool DEFINITION may be hidden in raw AI mode, but we still need to recognize it
           validTools.push('attempt_completion');
 
           // Edit tools (require both allowEdit flag AND allowedTools permission)

package/src/agent/tasks/taskTool.js

@@ -23,7 +23,8 @@ export const taskItemSchema = z.object({
  */
 export const taskSchema = z.object({
   action: z.enum(['create', 'update', 'complete', 'delete', 'list']),
-  tasks: z.array(z.union([z.string(), taskItemSchema])).optional(),
+  // Accept both array and JSON string (AI models sometimes serialize as string)
+  tasks: z.union([z.array(z.union([z.string(), taskItemSchema])), z.string()]).optional(),
   id: z.string().optional(),
   title: z.string().optional(),
   description: z.string().optional(),
@@ -142,6 +143,25 @@ SKIP TASKS for single-goal requests, even if they require multiple searches:
 **Key insight**: Multiple *internal steps* (search, read, analyze) are NOT the same as multiple *goals*.
 A single investigation with many steps is still ONE task, not many.
 
+## Task Granularity
+
+Tasks represent LOGICAL UNITS OF WORK, not individual files or steps:
+- "Fix 8 similar test files" → ONE task (same type of fix across files)
+- "Update API + tests + docs" → THREE tasks (different types of work)
+- "Implement feature in 5 files" → ONE task (single feature)
+
+**Rule of thumb**: If you're creating more than 3-4 tasks, you're probably too granular.
+
+**Anti-patterns to avoid**:
+- One task per file ❌
+- One task per function ❌
+- One task per repository (when same type of work) ❌
+
+**Good patterns**:
+- One task per distinct deliverable ✓
+- One task per phase (implement, test, document) ✓
+- One task per different type of work ✓
+
 MODIFY TASKS when (during execution):
 - You discover the problem is more complex than expected → Add new tasks
 - A single task covers too much scope → Split into smaller tasks
@@ -314,7 +334,17 @@ export function createTaskTool(options = {}) {
           return `Error: Invalid task parameters - ${validation.error.message}`;
         }
 
-        const { action, tasks, id, title, description, status, priority, dependencies, after } = validation.data;
+        const { action, tasks: rawTasks, id, title, description, status, priority, dependencies, after } = validation.data;
+
+        // Parse tasks if passed as JSON string (common AI model behavior)
+        let tasks = rawTasks;
+        if (typeof rawTasks === 'string') {
+          try {
+            tasks = JSON.parse(rawTasks);
+          } catch (e) {
+            return `Error: Invalid tasks JSON - ${e.message}`;
+          }
+        }
 
         switch (action) {
           case 'create': {