CRITICAL: Task Tool agents don't use MCP communication protocol - integration disconnect #10
Description
Task Tool MCP Integration Failure
Issue Summary
Severity: CRITICAL
Impact: Complete isolation of Task tool agents from MCP communication protocol
Status: Architectural integration failure - agents operate independently
When using Claude Code's Task tool to launch specialized agents, those agents do NOT inherit or use the MCP communication protocol, creating a complete disconnect between the intended coordination system and actual agent behavior.
Related Issues & Partial Solutions
🔗 Issue #24: Implement strict agent ownership validation - Provides mitigation by preventing Task tool agents from cross-contaminating tasks they don't own
🔗 Issue #26: Implement MCP Prompts for agent protocol guidance - Could provide Task tool agents with real-time protocol instructions even in isolated environments
Note: These related issues provide partial mitigation but don't solve the core Task tool isolation problem, which requires external enhancement to Claude Code's Task tool itself.
Expected vs Actual Behavior
Expected Workflow
// User launches agent via Task tool
Task(subagent_type="senior-frontend-engineer", prompt="Fix tests")
// Agent should automatically:
1. mcp__agent_comm__check_tasks(agent="senior-frontend-engineer")
2. mcp__agent_comm__submit_plan(agent="senior-frontend-engineer", content="...")
3. mcp__agent_comm__report_progress(agent="senior-frontend-engineer", updates=[...])
4. mcp__agent_comm__mark_complete(agent="senior-frontend-engineer", status="DONE")Actual Workflow
// User launches agent via Task tool
Task(subagent_type="senior-frontend-engineer", prompt="Fix tests") ✅
// Agent operates in complete isolation:
1. NO MCP tool access ❌
2. NO task context retrieval ❌
3. NO progress reporting ❌
4. NO plan submission ❌
5. NO completion tracking ❌
6. Returns fabricated detailed results ❌
Root Cause Analysis
Primary Cause: Process Isolation
Task tool launches agents in isolated environments/processes that:
- Don't inherit MCP tool access from main Claude Code session
- Can't communicate with MCP server running for main session
- Have no knowledge of task context or coordination protocol
- Generate responses without any validation mechanism
Current Mitigation Strategy
Until the core Task tool integration is fixed:
Immediate Workaround (RECOMMENDED)
Stop using Task tool for MCP-dependent operations. Use direct MCP approach in main session.
Partial Mitigations from Related Issues
- Issue feat: Implement strict agent ownership validation #24: Ownership validation prevents accidental cross-contamination
- Issue feat: Implement MCP Prompts for agent protocol guidance #26: MCP Prompts could educate Task tool agents about proper coordination requirements
Long-term Solution Required
This issue requires external enhancement to Claude Code's Task tool to support MCP integration. Our server-side improvements (#24, #26) provide mitigation but don't solve the fundamental isolation problem.
Priority: CRITICAL - Blocks fundamental agent coordination architecture
Status: Requires external Task tool enhancement