Implement interactive CLI with prompt-toolkit

[jx-qiu]

This pull request introduces an interactive mode to the RPent CLI, allowing users to steer the agent in real-time by typing messages during execution. The implementation adds a new --interactive flag, a terminal UI for user input, and integrates interactive input handling into the agent loop for both the api and claude/codex backends.

[wilburx813]

Curious about what “/default” is for, seems the default prompt is already pre-filled and can be submitted by pressing Enter.

[wilburx813]

maybe there are some bugs? it seems doesn't react to my messages.

my command:

  /mnt/public/ljy/.venv-opi-libero/bin/python \
    rpent/cli/main.py \
    --interactive \
    --suite libero_spatial_task \
    --task 0 \
    --seed 0 \
    --libero-type pro \
    --cerebrum claude_code \
    --model claude-opus-4-7 \
    --claude-code-max-budget-usd 30 \
    --max-turns 100 \
    --cerebrum-timeout-s 2400 \
    --cuda-device 1 \
    --vla-endpoint http://127.0.0.1:45665

[Copilot]

Pull request overview

This PR adds an interactive mode to the RPent CLI so users can type messages during agent execution to steer subsequent turns, and wires that interactive input into the api, claude_code, and codex planners via a shared TUI helper.

Changes:

• Add --interactive/-i CLI flag and a prompt-toolkit-based terminal UI that streams user lines into a queue.
• Integrate queued user steering into planner loops (pydantic-ai loop injection, Claude session continuation, Codex steer/interrupt thread).
• Document interactive mode in both English and Chinese READMEs and add prompt-toolkit as a dependency.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
rpent/planner/codex.py	Adds optional interactive input queue handling for Codex via a steering thread and renders userMessage events.
rpent/planner/claude_code.py	Adds optional interactive queue loop to continue a stateful Claude SDK client session one user line per turn.
rpent/planner/base.py	Extends the Planner protocol with an optional input_queue for interactive steering.
rpent/planner/api_loop.py	Adds interactive steering support by draining queued user messages into an active pydantic-ai run and looping between runs.
rpent/cli/tui.py	New prompt-toolkit TUI module for interactive input, help/quit tokens, and queue helpers.
rpent/cli/main.py	Adds --interactive/-i, starts the interactive reader, and passes the queue to the selected planner.
README.md	Documents interactive mode usage and provides an example command.
README.zh-CN.md	Documents interactive mode usage in Chinese and provides an example command.
pyproject.toml	Adds prompt-toolkit dependency needed for the interactive TUI.

Comments suppressed due to low confidence (1)

rpent/planner/api_loop.py:264

• The outer interactive loop currently continues after a run ends even if turns has reached max_turns, allowing another run to start and exceed the intended budget. Add an explicit turns >= max_turns check after the run closes to stop the session once the limit is hit.

                # finish => end the whole session immediately (chosen behavior).
                if finish_result is not None or quit_requested or not interactive:
                    break

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[jx-qiu]

Curious about what “/default” is for, seems the default prompt is already pre-filled and can be submitted by pressing Enter.

Removed the /default command.