Seven crates, one binary.

The crates.

From the binary you run down to the schema every other crate speaks.

• alforria

  The binary. It parses your command and routes it to the terminal UI, the local server, the web client, or a run mode that streams JSON events and exits with the right code.

• alforria-tui

  The terminal UI: a ratatui chat view with markdown and diff rendering, 30+ dialogs and 33 bundled themes, built as an HTTP and SSE client of the server. It never touches the engine directly.

• alforria-server

  The HTTP surface: v1 and v2 routers, the SSE stream, PTY websockets, auth and origin checks, OpenAPI export, and the embedded web client, all from one process.

• alforria-core

  The session engine. It owns the agent loop and the built-in tools, plus permissions, snapshots, config discovery, the model catalog, MCP, SQLite storage and the event bus the rest of the workspace speaks through.

• alforria-ui

  The web client, built from the SolidJS source in web/, packed into a deterministic zstd tarball and embedded in the binary. Decompressed once, on first use.

• alforria-llm

  The model layer: six protocol adapters that stream each response as events, report token usage, and handle retries and the prompt-cache policy.

• alforria-schema

  The wire contract: DTOs for sessions, events and the server API. It depends on no other crate, which is what makes the contract enforceable everywhere above it.

Dependencies point one way.

Read across: a dot means the row's crate depends on the column's. Every dot sits above the diagonal, so nothing depends on a crate listed above it. Select a crate to trace it.

What the selected crate depends onWhat depends on it

Behavior you can check.

Recorded behavior is frozen first and the code is written to replay it. Where the checks stop, this page says so.

• Golden fixtures first

  Recorded API traffic, event manifests and LLM protocol streams are frozen into fixtures, and the code is checked against them.

• Wire compatibility

  Schema golden tests compare parsed JSON values, including field names, union tagging, and omitted versus null fields. Float formatting and key order may differ. Separate tests check byte-identical OpenAPI documents.

• A deterministic agent loop

  A scripted model drives the full agent loop end to end through the real server and wire protocol, so sessions, retries and tool calls are covered by tests that run the same way every time. The home page replay uses a scripted model in the same way.

• Rendered frames and real browsers

  Headless ratatui buffer assertions check content in real rendered frames, including dialogs at 80 and 140 columns. The web client has an end-to-end suite that drives a real server in Chromium, plus a reducer scale test over 20 sessions of 2,000 parts.