Baize is a workspace-native supervisor for local coding agents. It is designed to route work across agents such as Codex, Gemini CLI, GitHub Copilot CLI, and OpenCode while preserving workspace state and handoff context.

This repository currently contains the first Rust MVP skeleton:

• Rust workspace crates;
• core domain types;
• TOML config loader;
• SQLite append-only event log and MVP query tables;
• git workspace status inspection;
• ACP JSON-RPC message primitives;
• provider registry and health probing stubs;
• local daemon with HTTP + SSE workspace/session/handoff/permission APIs;
• Gemini stream-json and Codex exec --json prompt execution paths;
• ratatui TUI shell;
• baize CLI entrypoint.

Once Rust is installed:

cargo run -p baize-cli -- status .
cargo run -p baize-cli -- providers
cargo run -p baize-cli -- doctor
cargo run -p baize-cli -- validate
cargo run -p baize-cli -- validate gemini
cargo run -p baize-cli -- daemon
cargo run -p baize-cli -- tui

For local sandboxed development, keep Baize data inside the repository:

BAIZE_DATA_DIR=.baize/data cargo run -p baize-cli -- daemon

For TUI usage, provider setup, keyboard shortcuts, local API examples and test commands, see doc/quickstart.md.

Useful daemon endpoints:

GET  /health
GET  /providers
GET  /providers/:id/validate
POST /providers/check
POST /providers/validate
GET  /workspaces
POST /workspaces
GET  /workspaces/:id/status
GET  /sessions
POST /sessions
GET  /sessions/:id
POST /sessions/:id/prompt
POST /sessions/:id/cancel
POST /sessions/:id/handoff
GET  /sessions/:id/events
GET  /sessions/:id/diff
GET  /permissions
POST /permissions
GET  /permissions/:id
POST /permissions/:id/approve
POST /permissions/:id/deny
GET  /events

POST /sessions/:id/prompt accepts an optional timeout_seconds field. Use a short timeout for smoke tests when a provider might be waiting for authentication.

The first adapter execution paths are Gemini --prompt --output-format stream-json and Codex exec --json. Tests use fake executors and parser fixtures so they do not spend model quota.