Codewright

English | 简体中文

A powerful and concise coding agent for the command line, written in Python.

Codewright : a streaming LLM loop, a polymorphic tool runtime, multi-agent orchestration, session persistence, MCP tool servers, and an interactive TUI.

Features

• Interactive TUI with streaming output, a status bar, and modal approvals.
• Non-interactive mode for one-shot prompts and scripting.
• Semantic file tools (read_file, list_dir, find_files, search_text) plus a stateful shell tool family (bash-anchored, with background jobs and output paging) and transactional apply_patch edits.
• Task planning: the agent maintains a live, step-by-step plan (update_plan) that streams to the TUI so you can watch progress.
• Permission profiles that gate destructive shell commands and out-of-workspace access.
• Multi-agent: spawn child agents that each run in their own isolated session under a built-in role — explorer (read-only investigation), worker (scoped execution), or default — and report back to the parent. The primary agent's role is set by [agent] default_role / CODEWRIGHT_DEFAULT_ROLE.
• Project instructions: an AGENTS.md (or .agents.md) found in the workspace — searched from the working directory up to the workspace root — is injected into every turn as guidance for the agent.
• Skills: project-scoped instructions and workflows under <workspace>/skills/ (agentskills.io format). Author them by hand, and/or let the agent auto-distill reusable, test-verified knowledge into new skills; it pulls the relevant ones back in on later tasks via the skill tool.
• MCP tool servers over stdio and streamable-http.
• Automatic context compaction: as history approaches the model's context window it is summarized down in place, so long sessions keep running.
• Session persistence + resume via append-only JSONL rollouts.
• Pluggable providers: OpenAI Chat Completions or Responses API, and any OpenAI-compatible endpoint (DeepSeek, Qwen/DashScope, Ollama, local gateways).

Requirements

• Python 3.12 (>=3.12,<3.13)
• uv
• An API key for an LLM provider (see Configuration)

Installation

As a global command (recommended)

Install Codewright as a standalone tool so codewright is available in any directory:

uv tool install --editable .      # run once from the repo root
uv tool update-shell              # if the uv tools bin dir is not on your PATH (then reopen the terminal)

--editable installs from source, so later code changes take effect without reinstalling. Omit it for a pinned snapshot (re-run with --reinstall to update). pipx install --editable . works too.

From source (development)

uv sync                           # install deps incl. the dev group
uv run codewright --help          # run without a global install (inside the repo)

Quick start

# Interactive TUI in the current directory (bare command == `codewright tui`):
codewright

# Interactive TUI in a specific workspace:
codewright --workspace path/to/project

# One non-interactive turn, prints the final message:
codewright run "explain what this project does" --workspace .

# List and resume saved sessions:
codewright list-sessions
codewright resume <session_id>                 # reopen in the TUI
codewright resume <session_id> --message "now add tests"   # one turn, then exit

Configuration

Settings are read from ~/.codewright/config.toml. Effective-value precedence (highest wins): CLI flags > environment variables > config file > built-in defaults.

A model API key is the only thing you must supply — everything else has a default. Provide a key via any one of:

Source	How
OPENAI_API_KEY	export it in your shell (simplest)
CODEWRIGHT_API_KEY	export it in your shell
[llm] api_key_env	name of an env var holding the key (keeps the secret out of the file)
[llm] api_key	the key inline (discouraged; plaintext)

A minimal ~/.codewright/config.toml:

[llm]
model = "gpt-4o"
api_key_env = "OPENAI_API_KEY"
# base_url = "https://api.deepseek.com/v1"   # for OpenAI-compatible providers

See examples/config.example.toml for every option (model/provider, context window, permission profile, shell path, skills test runners, and MCP servers) with its default and explanation.

Useful environment overrides: CODEWRIGHT_MODEL, CODEWRIGHT_API_KEY, CODEWRIGHT_API_KEY_ENV, CODEWRIGHT_BASE_URL (or OPENAI_BASE_URL), CODEWRIGHT_PERMISSION_PROFILE, CODEWRIGHT_DEFAULT_ROLE, CODEWRIGHT_SHELL_PATH, CODEWRIGHT_MAX_CONTEXT_TOKENS, CODEWRIGHT_COMPACT_THRESHOLD.

Commands

Command	Description
codewright	Launch the interactive TUI in the current directory (alias for tui).
codewright tui	Launch the interactive TUI.
codewright run <prompt>	Run a single non-interactive turn and print the final message.
codewright resume <session_id>	Resume a saved session (TUI, or --message for one turn).
codewright list-sessions	List saved sessions in the workspace.

Common flags: --workspace <dir> (default: current dir), --model, --provider-base-url, --api-style {chat_completions,responses}, --permission-profile {read_only,workspace_write,dangerous}. run also accepts --max-context-tokens, --print-session-id, and --no-persist. Run codewright <command> --help for the full list.

TUI keys

Key	Action
Ctrl-C	Interrupt the current turn
Ctrl-D	Quit
PageUp / PageDown	Scroll history
Ctrl-Home / Ctrl-End	Jump to top / bottom
F1 or ?	Toggle status details
y / s / n / a	On an approval prompt: approve / approve for session / deny / abort

Permissions

The active permission profile decides which shell commands run without asking:

• read_only — no shell exec at all; only the read/search file tools.
• workspace_write (default) — exec is auto-allowed inside the workspace; destructive commands (rm -rf, git reset --hard, network access, …) and any cwd outside the workspace still prompt for approval.
• dangerous — nothing is auto-allowed; every exec prompts.

The workspace root is a hard boundary: reads and writes outside it require approval.

Sessions

Each session is recorded as an append-only JSONL rollout under <workspace>/.codewright/sessions/<session_id>.jsonl. run persists by default (disable with --no-persist); the TUI always persists. Use list-sessions and resume to pick up where you left off.

Architecture

The engine sits behind two asyncio.Queues: front ends submit Op instances and consume Event instances, and never touch Session internals. A single submission_loop drives run_turn, which streams the model, dispatches tool calls through the executor, compacts history, and emits events.

Development

uv run pytest                         # full test suite
uv run pytest tests/test_run_turn.py  # a single file
uv run ruff check src tests           # lint
uv run ruff check --fix src tests     # lint + autofix
uv run lint-imports                   # enforce the layering contracts (.importlinter)

License

MIT