Skip to content

Expert Panel System

The basic unit of LingXiao is not “an assistant” but a Leader + Worker expert panel. You provide the goal; the Leader handles judgment, decomposition, planning, DAG construction, team assembly, and task dispatch. Worker experts execute research, frontend, backend, testing, review, documentation, and other work in parallel.

  • Understand user goals and decompose them into a dependency-aware task graph (DAG)
  • Select appropriate Worker roles based on task type
  • Handle user confirmation and interaction
  • Monitor execution progress, handle blockers and failures
  • Verify task results, extract PASS/FAIL/BLOCKED verdicts
  • Coordinate cross-task dependencies, wrap up and summarize
  • Receive tasks dispatched by the Leader, execute in independent processes
  • Have independent context, tool calls, runtime state, and logs
  • Submit results and verification evidence via attempt_completion
  • Report progress or request help via send_message
RoleIdentifierResponsibilities
ArchitectarchitectArchitecture design, interface boundaries, module decomposition, risk control
BackendbackendBackend implementation, state machines, APIs, databases, task scheduling
FrontendfrontendWebUI/TUI interaction, state projection, visualization workbench
FullstackfullstackCross-stack implementation, contract-first, end-to-end verification
CodingcodingGeneral coding implementation
ResearchresearchResearch, solution comparison, external verification
ExploreexploreCode exploration, architecture understanding, entry point localization
VerifyverifyVerification, regression testing, build checks
ReviewreviewCode review, quality assessment
QAqaTest case writing, quality assurance
PlannerplannerPlanning, task decomposition
EvaluatorevaluatorEvaluation, acceptance
UX Designerux_designerUsability review, interaction design

Each Worker runs in an independent process, communicating with the Leader via IPC:

  • Heartbeat mechanism: 30-second heartbeat timeout for automatic Worker liveness detection
  • State channel: Worker state, logs, conversations, and results flow through agent communication channels
  • Supervision and intervention: The Leader uses orchestration tools rather than process flags for supervision

LingXiao supports extending expert capabilities through multiple approaches:

  1. Role registration: Define new roles in builtinRoles.ts or via the WebUI role management interface
  2. Skill system: Bind domain knowledge and execution flows to roles via skill_names
  3. Tool permissions: Configure allowed/requested tool sets per role
  4. Dynamic roles: The Leader can dynamically create roles at runtime via define_agent_role
// Role definition example
{
name: "security-auditor",
description: "Security audit expert",
systemPrompt: "You are a security audit expert...",
allowedTools: ["code_search", "file_read", "shell"],
skillNames: ["security-review"],
suggestedModel: "claude-sonnet-4"
}

The WebUI Agent panel shows each Worker’s real-time:

  • Role identity and current task
  • Tool call records (parameters, permissions, output, duration)
  • Runtime status and live logs
  • Result receipts and verification evidence