lightcode: Instructions file for GitHub Copilot

Copilot Instructions

Build and test commands

• Build the main binary: go build ./cmd/lightcode or make build
• Run the app: go run ./cmd/lightcode/main.go
• Run the full test suite: go test ./...
• Run the existing focused test file: go test ./internal/server/tools
• Run a single test: go test ./internal/server/tools -run TestSkill_BasicSkill

There is no repo-defined lint target or linter config in this repository. Do not invent one in a Copilot session.

go test ./... is currently not green: internal/server/tools/skill_test.go expects SKILL_PATH-based behavior, but internal/server/tools/skill.go resolves the skills directory through config.SkillsPath() from ~/.lightcode/config.json.

High-level architecture

cmd/lightcode/main.go is the only real entrypoint. By default it starts the local HTTP server and then launches the Bubble Tea TUI; -server runs only the server and -tui runs only the TUI.

The app is split into two halves that communicate over localhost HTTP instead of direct package calls:

• internal/server exposes endpoints for session CRUD, message history, streaming chat completions, and the current todo list.
• internal/tui is the Bubble Tea frontend. It talks to the server through internal/tui/client, including streaming assistant responses from /chat-completion.

Persistent app state lives under ~/.lightcode/:

• config.json stores theme, port, providers, skills path, and the currently selected model.
• lightcode.db is a SQLite database managed through GORM.
• skills/ holds user skills that the skill tool loads.

The agent loop is in internal/server/agent/agent.go. Each iteration reads the session's stored messages from SQLite, converts them into llm.Chat messages, appends the session todo list as an extra user message, calls the model, persists the assistant reply, executes any tool calls, persists tool outputs, and repeats until the model returns plain assistant text or the iteration limit is hit.