wsshow/feikong-teams

Purpose and Entry Points

fkteams is a locally runnable Go multi-agent AI assistant for development, research, analysis, remote operations, and automation. The binary starts at cmd/fkteams/main.go; Cobra-style CLI commands expose interactive CLI/TUI, Web, API server, login, model management, sessions, skills, tools, and updates under internal/adapters/transport/cli/commands/. fkteams web/serve host the HTTP API and embedded React frontend.

Layering and Composition

The repository follows ports-and-adapters boundaries: domain types/events in internal/domain/, application orchestration in internal/app/, contracts in internal/ports/, and implementations in internal/adapters/. Bootstrap assembles configured runtimes, tools, skills, channels, schedulers, and environment dependencies through internal/bootstrap/runtimes/runtimes.go, internal/bootstrap/tools/tools.go, and internal/bootstrap/services/.

Agent Catalog and Task Orchestration

Agent definitions include coordinator, coder, researcher, analyst, tasker, moderator, and custom agents. The catalog/registry selects agent behavior and prompts from internal/app/agent/catalog/registry.go and internal/app/agent/catalog/; internal/app/agent/runner.go executes selected agents. Chat lifecycle and long-running task ownership—including interruption and queued follow-ups—are coordinated by internal/app/chat/service.go and internal/app/chat/taskstream/manager.go.

Runtime Execution Pipeline

The Eino runtime adapter is the primary agent execution engine: internal/adapters/runtime/eino/engine/engine.go, internal/adapters/runtime/eino/runner.go, and internal/adapters/runtime/eino/factory.go. It routes model messages and tool calls through middleware for dispatch, skills/context injection, summaries, steering, filesystem access, result trimming, patching, and destructive-operation checks. Runtime-facing abstractions preserve the boundary in internal/ports/runtime/runtime.go; turn execution, approvals, checkpoints, and event emission are handled in internal/runtime/turn/, internal/runtime/approval/, and internal/runtime/events/.

Models, Tools, and Extensions

Model providers are registered behind runtime provider adapters, including OpenAI-compatible services, Claude, Gemini, DeepSeek, Ollama, Qwen, Ark, OpenRouter, and GitHub Copilot (internal/adapters/runtime/eino/providers/register/register.go, internal/adapters/model/providers/copilot/copilot.go). Built-in tools cover filesystem, command execution, Git, SSH, search, documents, spreadsheets, scripts, and scheduling under internal/adapters/tools/builtin/. MCP clients (internal/adapters/tools/mcp/client.go) and SkillHub-backed skills (internal/adapters/skill/skillhub/provider.go) extend the tool/knowledge surface.

Persistence, Memory, and Scheduled Work

Local file storage persists sessions, transcripts, attachments, metadata, and memory via internal/adapters/storage/file/history/. Application memory extracts, ranks, injects, and manages long-term context in internal/app/memory/; the memory model adapter is internal/adapters/model/memory/adapter.go. Scheduled tasks are collected/executed in internal/app/schedule/ and persist through the file-backed cron adapter at internal/adapters/scheduler/filecron/scheduler.go.

Transports and External Interfaces

The HTTP server routes chat, streaming/WebSocket events, OpenAI-compatible endpoints, session/file/share/skill/schedule APIs, authentication, and preview endpoints through internal/adapters/transport/http/server.go, internal/adapters/transport/http/router/router.go, and internal/adapters/transport/http/handler/. The embedded Bun/Vite React UI begins at web/src/main.tsx and drives task-stream chat from web/src/features/chat/TaskStreamController.tsx. CLI/TUI rendering consumes the same event model through internal/adapters/transport/cli/runtime/. Channel bridges connect Discord, QQ, and Weixin in internal/adapters/transport/channel/.

Runtime and Deployment Boundaries

The backend is Go; the embedded frontend is TypeScript/React built with Bun. Docker deployment is supplied by Dockerfile and docker-compose.yml. External dependencies include configured LLM providers, GitHub Copilot OAuth, optional MCP servers, SkillHub, Discord, QQ, Weixin, DuckDuckGo search, and local executables such as Git, SSH, Bun, and uv. Local workspace and tool policies are central security boundaries, enforced by command security, path guarding, approvals, and runtime tool policy.