pi-go

Pi-go is a terminal‑based coding assistant written in Go that runs from a single binary. It lets you converse with large language models without leaving your command‑line workflow, and it bundles a set of development‑oriented tools that can be invoked directly from the chat interface.

When you start the program with `pi` you are presented with a text‑based user interface. You choose a model either with a flag such as `--model gpt-5.6-sol` or later in the session using the `/model` slash command, which swaps the provider without restarting the tool. After you type a prompt, the agent forwards the request to the selected LLM, receives the response and can automatically call one of the sandboxed utilities. Typical inputs include a request to create a file, edit code, run a shell command, or query a git repository. The output appears in the same pane, showing the LLM's answer together with any tool output, such as file contents, grep matches, or a git diff. All tool actions are restricted to the directory defined as the project root, enforced by the Go `os.Root` sandbox, so the assistant cannot modify files outside that scope.

The “Memory Palace” stores conversational context across four layers. A SQLite database holds structured data, an FTS5 index enables full‑text search, semantic embeddings allow similarity queries, and a temporal knowledge graph records the order of events. You can retrieve stored snippets with `/memory` commands, which return matching text or embeddings that the LLM can reference in later replies. This persistent memory is useful for large projects where the assistant needs to recall earlier design decisions or code snippets.

For language‑specific assistance, pi‑go includes an LSP client that connects to JSON‑RPC servers for Go, TypeScript, Python and Rust. When you ask for formatting, diagnostics or type information, the agent forwards the request to the appropriate language server and returns the formatted code or error list in the terminal. The subagent system runs separate processes for distinct roles, explore, plan, designer, reviewer and generic task runners. You can launch a planning session with `/plan`, start a design review with `/agents`, or execute a specification using `/run`. Each subagent works concurrently, and you can monitor their status with the `/agents` command.

Sessions are recorded in an append‑only JSONL log.