Gum Agent Mesh

Multi-AI-agent orchestrator inside VS Code — Codex + Claude + DeepSeek working as a coordinated team, not in parallel chaos.

A single prompt is routed through a 4-phase pipeline: one agent analyzes the repo, another implements, a third reviews the result, and a fourth summarizes for you. No race conditions, no token waste from duplicate work.

⬡ Why this exists

Most "use multiple AIs" setups send the same request to every model in parallel. They each guess differently, edit the same files, and the last write wins. This extension fixes that with sequential pipeline orchestration and event-driven coordination — no polling, no timer-based wake-up.

The 3 musketeers

Roles are configurable — bring your own agent (Anthropic API, OpenAI-compatible, any CLI).

Pipeline

USER PROMPT
   ↓
1. ANALYZE     → analyzer reads repo, proposes plan (no edits)
   ↓
2. IMPLEMENT   → implementer executes, runs build/check
   ↓
3. REVIEW      → reviewer audits the result (read-only)
   ↓
4. SUMMARIZE   → summarizer produces a concise markdown report
   ↓
USER REPLY

Each phase is gated on the previous output — so reviewers see real diffs and summarizers see real results, not guesses.

Key features

• Sequential pipeline — no race condition, no concurrent writers
• Heartbeat-aware timeout — extends automatically when a worker is making progress
• Late-reply rescue — replies that arrive after timeout still surface as supplemental notes
• In-process broker — agents communicate via Node EventEmitter, latency in ms
• Auto-stdin for CLI agents — multi-line prompts pass cleanly through Windows cmd
• File explorer in panel — click a file → inserts @path reference into chat
• Drag-drop & paste image — screenshots are saved to .agent-mesh/attachments/ and referenced
• Markdown rendering — code blocks, lists, links all render in chat
• Dashboard welcome — agents online, tasks in progress, tokens spent, recent activity
• Append-only JSON Lines persistence — thread history survives reload, zero native dependencies
• Bus bridge — file-watcher bridge to existing agent-hub style file buses
• CLI Tools menu — install / login / diagnose Claude Code & Codex CLI from inside the panel

Setup

1. Install the extension from the Marketplace.
2. Open Settings (Ctrl+,) → search "Gum Agent Mesh".
3. Provide credentials for any provider you want:
   • gumAgentMesh.anthropicApiKey — Anthropic API key (orchestrator + workers)
   • gumAgentMesh.deepseekApiKey — DeepSeek API key (summarizer)
4. Optionally configure gumAgentMesh.customAgents to wire in CLI agents (Claude Code, Codex, …).
5. Open the Agent Mesh activity bar icon → start typing.

Example customAgents config

"gumAgentMesh.customAgents": [
  {
    "id": "claude-cli",
    "name": "Claude Code",
    "provider": "cli",
    "meshRole": "implementer",
    "command": "claude",
    "args": ["--print", "{{prompt}}"]
  },
  {
    "id": "codex-cli",
    "name": "Codex",
    "provider": "cli",
    "meshRole": "reviewer",
    "command": "codex",
    "args": ["exec", "{{prompt}}"]
  },
  {
    "id": "deepseek",
    "name": "DeepSeek",
    "provider": "deepseek",
    "meshRole": "summarizer",
    "model": "deepseek-chat"
  }
]

The orchestrator infers roles by id automatically (claude* → implementer, codex* → reviewer, deepseek* → summarizer) — meshRole only needed if you want to override.

Commands

All available from Ctrl+Shift+P:

• Gum Agent Mesh: Open Panel — show the chat
• Gum Agent Mesh: CLI Tools — quickpick menu for install / login / diagnose
• Gum Agent Mesh: Install Claude Code CLI
• Gum Agent Mesh: Login Claude Code CLI
• Gum Agent Mesh: Install Codex CLI
• Gum Agent Mesh: Login Codex CLI
• Gum Agent Mesh: Diagnose Agents — checks every configured CLI/API, reports to Output channel

Architecture notes

• No polling anywhere — file watcher uses OS push events (vscode.workspace.createFileSystemWatcher); in-process broker uses EventEmitter.
• Each agent is a class implementing one handle(message) method. Adapters exist for: Anthropic SDK, OpenAI-compatible HTTP (DeepSeek), child-process CLI, mock.
• Prompt caching is enabled on every Anthropic system prompt block.
• Diagnostics output channel logs CLI invocations with command, args, cwd, duration, exit code, and stderr tail on failure.

Requirements

• VS Code 1.85+
• Node 18+ (extension host runs in VS Code's Electron node)
• For CLI agents: install claude (@anthropic-ai/claude-code) and codex (@openai/codex) globally via npm
• API keys for any cloud agent you enable

Status & roadmap

Status: 0.1.0 — early access. Workflow stable on Windows/macOS; field-tested on real refactors of 5–10 minute scope.

Planned:

• Streaming agent responses
• File-claim locking (prevent multi-writer conflict at the broker layer)
• More provider adapters (Gemini, local Ollama)
• MCP server export so external tools can drive the mesh

License

MIT