heterogeneous-agent

Debug and implement LobeHub external CLI agent adapters for Claude Code and Codex pipelines.

74|11|Updated Jul 4, 2024
One-click install
npx skills add https://github.com/OpenSourceAGI/qwksearch-research-agent --skill heterogeneous-agent-opensourceagi
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: heterogeneous-agent
Source: https://github.com/OpenSourceAGI/qwksearch-research-agent/tree/main/apps/qwk-in-lobe/.agents/skills/heterogeneous-agent
Command: npx skills add https://github.com/OpenSourceAGI/qwksearch-research-agent --skill heterogeneous-agent-opensourceagi

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Bugs in external CLI agent integrations (Claude Code, Codex) are hard to localize because failures can originate in raw CLI output, event adapters, persistence, or UI hydration. This Skill provides a structured pipeline map, debug order, and invariants to isolate and fix defects in the heterogeneous agent transport layer. ## Core Features & Use Cases - Layered Debug Workflow: Trace bugs from raw CLI stdout through adapters, the executor, persistence, and the gateway UI handler in a fixed order. - Trace Capture Guidance: Read recorded live session traces (meta.json, stdout.jsonl, stderr.log) from .heerogeneous-tracing/ instead of hand-rolling CLI reproductions. - Invariant Checklists: Enforce rules like one tool item mapping to one stable ToolCallPayload.id, message.id-based step boundaries, and subagent thread isolation. - Use Case: When Claude Code duplicates text or Codex collapses multiple tools into one assistant message, follow the debug order to compare raw JSONL against adapted events and add a focused regression test. ## Quick Start Ask the agent to debug why Claude Code tool results never appear in the chat UI using the heterogeneous agent debug workflow.

Frequently Asked Questions about heterogeneous-agent

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I debug Claude Code stream-json adapter bugs?

Start by reading the most recent recorded session via .heerogeneous-tracing/.last-live-trace to inspect raw stdout.jsonl. Then compare raw events against adapted events exposed on window.__HETERO_AGENT_TRACE in dev builds before touching persistence or UI code.

Why does Claude Code duplicate text or thinking in the chat UI?

Duplication usually happens when partial content_block_delta events and the later full assistant block are both emitted. Check the adapter to ensure partial deltas are not re-emitted when the complete assistant event arrives.

Why do multiple Codex tools collapse into one assistant message?

The executor only cuts a new assistant message when it receives a step-boundary signal like stream_start with newStep. Check whether the Codex adapter emits a usable turn boundary from turn.started and turn.completed events.

How are Claude Code tool results mapped to tool messages?

In Claude Code, tool_result blocks arrive in type: 'user' events, not assistant events. The executor resolves each result through the global toolMsgIdByCallId map, so ignoring user events leaves tool rows loading forever.

When should I not start debugging at the UI layer?

Never start with UI rendering or a broad Electron repro when a raw trace or adapter test can prove the fault zone faster. Only move to UI hydration after raw output, adapted events, and persistence invariants all check out.