Bug: Task tool fails with "tools: Tool names must be unique" error

Environment

• Claude Code Version: Latest (as of 2025-10-31)
• OS: macOS 25.0.0 (Darwin)
• Node: 22+
• Project Type: Next.js 16 application with MCP servers

Description

The Task tool consistently fails to launch subprocess agents with a 400 API error: "tools: Tool names must be unique". This occurs for all subagent types, including general-purpose, and happens during agent initialization before any execution begins.

Steps to Reproduce

1. Invoke Task tool with subagent_type="general-purpose"
2. Provide any valid prompt (e.g., "Create test files for API routes")
3. Observe API Error 400 response

Actual Behavior

{
  "type": "error",
  "error": {
    "type": "invalid_request_error",
    "message": "tools: Tool names must be unique."
  }
}

Error appears immediately during agent initialization, before any task execution.

Expected Behavior

Task agent should initialize successfully and execute the provided prompt autonomously.

Test Cases Attempted

• ❌ 4 parallel agents (single message, multiple Task invocations) - all failed
• ❌ 1 sequential agent - failed
• ❌ Agent with explicit model: "sonnet" parameter - failed
• ❌ Different subagent types - all failed

Request IDs

• req_011CUezFfa6UVuTcb13YUK1f
• req_011CUezFeoxPTtpm1Ui1wbj7
• req_011CUezFeapTWu9numGWTFmM
• req_011CUezFemyUWAUnTVHhmoGG
• req_011CUezHAwDVuXkegeUVzKmC

Investigation Results

• ✅ Verified no duplicate tools in parent environment (49 unique tools)
• ✅ Error is consistent across all invocation patterns
• ✅ Not specific to parallel execution (single agent also fails)
• ✅ Not model-specific (tried default and sonnet)
• ✅ Occurs at API validation level (400 error from Anthropic)

Root Cause Hypothesis

The Task tool's implementation likely has a bug where it double-adds tools when creating subprocess agents:

1. Inherits all parent tools from the main session
2. Re-adds them from the agent type definition
3. Results in duplicate tool names passed to the Anthropic API
4. API validation rejects the request

Impact

• Severity: Medium
• Workaround Available: Yes (direct implementation without agents)
• Blocked Feature: Parallel autonomous agent execution for large-scale tasks
• User Impact: Reduced efficiency - tasks that could be parallelized across 4+ agents must be done sequentially

Active MCP Servers

• GitHub MCP (26 tools: create_issue, create_pull_request, etc.)
• Microsoft Learn MCP (3 tools: docs_search, code_sample_search, docs_fetch)
• IDE MCP (2 tools: getDiagnostics, executeCode)

Use Case That Failed

Attempted to launch 4 parallel agents to create test coverage for 94 API routes:

• Agent 1: Entitlements routes (4 routes)
• Agent 2: Patient cycle management (21 routes)
• Agent 3: Clinic admin routes (33 routes)
• Agent 4: Mobile app & AI routes (33 routes)

All agents failed immediately with the duplicate tools error.

Suggested Fix

Add validation in the Task tool implementation to:

1. Deduplicate tools before passing to Anthropic API
2. Provide clear error messages if duplicates are detected
3. Consider tool namespacing for subagents to prevent conflicts