qwenpaw-tui 0.1.0

A terminal chat UI for QwenPaw — drive an existing QwenPaw installation over ACP.

QwenPaw-TUI

A terminal chat UI for QwenPaw.

QwenPaw-TUI is a small, fast Textual front-end for an existing QwenPaw installation. It spawns qwenpaw acp and drives it over ACP (Agent Client Protocol) — streaming replies and thinking, rendering tool calls as inline panels, handling permission prompts, and forwarding slash commands (/clear, /compact, …) straight to the agent. Its command-line tool is paw.

It only speaks ACP and never imports QwenPaw, so it stays light and ships independently. Everything else — provider keys, model selection, memory, tools — lives in QwenPaw; paw just drives the agent you've already configured.

Install

Install QwenPaw first (see the QwenPaw docs), then install the UI:

pip install qwenpaw-tui   # finds `qwenpaw` on your PATH

paw does not bundle or install QwenPaw, so there's nothing to conflict with your existing setup — it just drives whatever qwenpaw is on your PATH (or an explicit --agent-cmd).

Usage

paw                              # interactive chat with your QwenPaw
paw --agent writer               # pick a specific agent
paw --agent-cmd "qwenpaw acp"    # drive an explicit ACP command

For non-interactive / one-shot use, run QwenPaw directly: qwenpaw chat.

Inside the chat: enter sends, shift+enter inserts a newline, esc interrupts the current turn, and ctrl+c quits.

Type / to open an overlay suggestion list. It includes paw's own commands, QwenPaw commands advertised over ACP, and argument completions such as /theme cyberpunk. Use up/down to pick, tab or enter to insert a highlighted suggestion, and esc to dismiss. Once an exact full command is typed, enter submits it.

Useful local commands:

• /theme opens the theme gallery; /theme <prompt> generates a persistent vibe from your words.
• /inspect toggles between friendly chat and deeper tool/thought inspection.

Model and provider commands (e.g. /model) are QwenPaw's — paw forwards them to the agent. Configure providers and models with QwenPaw's own tools (qwenpaw models config-key, qwenpaw models set-llm).

Pasting a file path or data URL attaches it to the prompt. Pasting long text stores it as a temporary attachment and inserts a reference, keeping the input usable.

How it finds QwenPaw

paw resolves the agent to drive in this order:

1. --agent-cmd "<command>" — used verbatim.
2. PATH — runs qwenpaw acp.

If QwenPaw isn't on your PATH, install it first or pass --agent-cmd.

How it works

paw is an ACP client. It spawns the QwenPaw agent as a subprocess and exchanges JSON-RPC over stdio. Because QwenPaw already ships a full ACP agent (qwenpaw acp), paw reuses the backend for tools, memory, advertised slash commands, permissions, and model switching.

The agent's stderr is drained to a log file under paw's state dir (PAW_STATE_DIR, or an OS default) so chatty tools (e.g. a headless browser) can't deadlock the stdio stream.

Develop

pip install -e ".[dev]"
pytest -m "not integration"  # unit + transport + UI + CLI tests

Run the live development-checkout integration tests with provider keys:

DASHSCOPE_API_KEY=... \
OPENAI_API_KEY=... \
ANTHROPIC_API_KEY=... \
GEMINI_API_KEY=... \
DEEPSEEK_API_KEY=... \
  pytest tests/test_qwenpaw_dev_integration.py -m integration -vv

The integration suite defaults to /Users/erkang.zhu/code/ContinueLearningBench/submodules/QwenPaw, or set QWENPAW_DEV_CHECKOUT=/path/to/QwenPaw. It starts QwenPaw through uv run --project <checkout> python -m qwenpaw acp, verifies the imported module comes from that checkout, and sets temporary QWENPAW_WORKING_DIR, QWENPAW_SECRET_DIR, and PAW_STATE_DIR so it does not write to ~/.qwenpaw or ~/.copaw. It covers multi-turn conversation, a live read_file tool call through ACP, and two-turn provider smoke tests for DashScope, OpenAI, Anthropic, Gemini, and DeepSeek. Provider cases skip cleanly if the corresponding API key is absent.

Default live-test models can be overridden with:

• PAW_E2E_DASHSCOPE_MODEL
• PAW_E2E_OPENAI_MODEL
• PAW_E2E_ANTHROPIC_MODEL
• PAW_E2E_GEMINI_MODEL
• PAW_E2E_DEEPSEEK_MODEL

Against a local QwenPaw checkout

To test paw against an in-development QwenPaw (e.g. a sibling ../QwenPaw editable install) without touching your normal QwenPaw setup, point paw at that checkout's interpreter with --agent-cmd, and isolate its data with QWENPAW_WORKING_DIR:

QWENPAW_WORKING_DIR="$PWD/.devdata" \
  paw --agent-cmd "/path/to/QwenPaw/.venv/bin/python -m qwenpaw acp"

paw forwards its environment to the spawned agent, so any vars you set (QWENPAW_WORKING_DIR, provider keys, etc.) reach QwenPaw. The agent uses .devdata/.devdata.secret for its config, sessions, and secrets, leaving ~/.qwenpaw untouched.

First time, seed a provider key and pick a model into .devdata using QwenPaw's own config (run the dev interpreter with the same working dir):

DEV="/path/to/QwenPaw/.venv/bin/python"
# store an API key for a provider (id: `dashscope` or `openai`) — paste the
# key at the prompt, e.g. $DASHSCOPE_API_KEY / $OPENAI_API_KEY
QWENPAW_WORKING_DIR="$PWD/.devdata" "$DEV" -m qwenpaw models config-key dashscope
# then choose the active model
QWENPAW_WORKING_DIR="$PWD/.devdata" "$DEV" -m qwenpaw models set-llm

The key is encrypted into .devdata.secret, so it stays out of your normal install (and out of git).

License

MIT