Rich Claude Code TUI Clone

Build a Python terminal interface that recreates the interaction grammar of a modern coding agent: conversation-first layout, streaming responses, compact tool summaries, permission gates, rich diffs, transcript navigation, and restrained motion.

Do not treat Rich as a drop-in replacement for Ink. Rich is the rendering layer. Input editing, keyboard dispatch, viewport state, search, and modal ownership must be implemented by the application or delegated to an optional widget framework.

When to Activate

Activate when:

• Rebuilding a Claude Code-like terminal experience in Python
• Converting a TypeScript/Ink TUI design into Rich renderables
• Adding full-screen mode, transcript export, tool panels, diffs, or themes to a Python agent CLI
• Deciding whether a Rich-only CLI is sufficient or a Textual application is justified
• Reviewing a Python coding-agent TUI for rendering, input, accessibility, or test gaps

For a generic TypeScript/OpenRouter scaffold, use create-agent-tui instead. This skill is specifically for the Python/Rich implementation path.

Source Boundary

Use:

• Official Rich documentation and repository examples for API behavior
• Official Textual materials for the optional widget-framework upgrade path
• Toad only as a high-level architecture reference for a current Rich/Textual agentic TUI
• The user-supplied visual report as a design target

Do not copy proprietary product code, trademarks, or mascot art. Recreate interaction patterns and visual hierarchy. Use project-owned names and artwork unless the user has rights to branded assets. Toad is AGPL-licensed: study its architecture, but do not copy implementation code into an incompatibly licensed project.

First Decision: Choose the Fidelity Level

Start at the lowest level that meets the acceptance criteria. A full Claude Code visual clone normally ends at textual-app; a convincing first milestone can remain rich-screen.

Read references/rich-api-map.md before selecting a level.

Non-Negotiable Architecture

Use one-way state flow:

agent events + terminal input + timers
              |
              v
         reduce(state, event)
              |
              v
         immutable AppState
              |
              v
         render_app(state)
              |
              v
        Rich renderable tree

Keep transport, state transitions, and rendering separate. Never let an agent callback write directly to stdout while a Live display owns the terminal.

Recommended project shape:

src/
  app.py
  state.py
  theme.py
  runtime/
    agent_events.py
    input.py
    terminal.py
    transcript.py
  renderables/
    logo.py
    messages.py
    tools.py
    diff.py
    prompt.py
    footer.py
    dialogs.py
tests/
  test_render_snapshots.py
  test_reducer.py
  test_transcript.py

For the component map and state model, read references/claude-code-parity-roadmap.md.

Build Workflow

Step 1: Write a Parity Contract

Record the requested fidelity before coding:

## TUI Parity Contract
- Fidelity level: scrollback | rich-screen | textual-app
- Required modes: prompt | transcript | fullscreen
- Required message types:
- Required tool renderers:
- Prompt features: multiline | history | slash completion | vim | paste
- Permission flows:
- Diff requirements: unified | split | word-level
- Theme targets: dark | light | daltonized | ANSI
- Motion policy: standard | reduced
- Terminal targets: Linux | macOS | Windows Terminal | WSL
- Brand assets authorized: yes | no

Step 2: Scaffold the Rendering Layer

Use a custom Theme, then implement renderables in this order:

1. Logo header and status notices
2. User, assistant, thinking, and tool message blocks
3. Spinner row with verb, elapsed time, and state color
4. Prompt shell and footer
5. Permission dialog
6. Structured diff
7. Transcript layout and export

Use Text, Panel, Group, Table.grid, Rule, Markdown, and Syntax before creating custom renderables. Implement __rich_console__ only when composition is no longer enough.

Step 3: Add the Runtime Loop

For rich-screen, keep one Live owner:

with Live(
    render_app(state),
    console=console,
    screen=True,
    refresh_per_second=20,
    vertical_overflow="crop",
) as live:
    while state.running:
        event = await event_queue.get()
        state = reduce(state, event)
        live.update(render_app(state))

Use auto_refresh=False and live.update(..., refresh=True) when deterministic repaint timing matters. Use a transient non-screen Live display for the scrollback MVP.

Step 4: Implement Streaming Deliberately

Do not re-render the entire transcript for every token:

1. Freeze completed message blocks.
2. Cache renderables by message id, width, theme, and content version.
3. Rebuild only the active streaming suffix.
4. Parse Markdown at stable block boundaries when possible.
5. Throttle repaints independently from token arrival.

Rich Markdown is sufficient for complete blocks. A production streaming renderer may need a stable prefix plus an active suffix, especially for open code fences and long responses.

Step 5: Treat Input as a Separate Subsystem

Rich does not provide a multiline prompt editor, fuzzy completion, Vim mode, cursor editing, mouse selection, or modal focus management.

Choose one:

• Basic MVP: asyncio plus a small raw-terminal input adapter
• Intermediate: a dedicated prompt library for multiline editing and history
• Full parity: Textual widgets, bindings, screens, and reactive state

Do not fake widget behavior with scattered ANSI writes inside render functions.

Step 6: Add Transcript and Fullscreen Behavior

For transcript export, use a recording console:

recording = Console(record=True, width=width)
recording.print(render_transcript(state))
plain_text = recording.export_text(clear=True, styles=False)

For alternate-screen rendering, use Live(screen=True) or Console.screen(). Keep transcript search, viewport slicing, sticky prompts, and unseen-message counters in application state.

Step 7: Verify at Multiple Widths and Capabilities

Snapshot renderables at 60, 100, and 140 columns. Test:

• Dark, light, daltonized, ANSI, and no-color modes
• Reduced-motion mode
• Long paths, wide characters, wrapped code, and partial Markdown fences
• Permission allow, deny, and cancel paths
• Streaming updates, resize events, transcript export, and search
• Windows Terminal or WSL behavior if Windows is a target

Read references/input-fullscreen-testing.md for the detailed test matrix.

Definition of Done

A finished implementation:

• Has a declared fidelity level and terminal support matrix
• Keeps agent events separate from terminal writes
• Has one owner for each live terminal region
• Renders complete and streaming messages without transcript-wide repaint churn
• Uses explicit fallbacks for Unicode, color, and reduced motion
• Keeps dangerous tool approvals keyboard-operable and auditable
• Exports a plain-text transcript
• Has snapshot tests at narrow and wide widths
• Documents which Claude Code-like features are implemented, deferred, or delegated to Textual

Gotchas

1. Rich is not Ink. It renders terminal content; it does not provide a React component lifecycle or

a full input widget system.

1. Alternate screen changes scrollback expectations. Use normal printing for shell-like history and

screen=True only when the application should own the viewport.

1. Concurrent writes corrupt live displays. Route logs, subprocess output, and agent deltas through

events.

1. Virtualization is application behavior. Rich Layout does not virtualize a long transcript.
2. Streaming Markdown is incremental parsing work. Cache completed blocks and isolate the unstable

suffix.

1. Unicode width varies by terminal. Test mascot art, box borders, braille spinners, and wide glyphs;

ship ASCII fallbacks.

1. Color is not the only signal. Pair colors with labels, prefixes, or icons.
2. Reference implementations have licenses. Learn from Toad's structure; do not paste AGPL code into

a project with incompatible distribution terms.

References

• Rich API map
• Claude Code parity roadmap
• Input, fullscreen, and testing
• Research ledger

Skill Metadata

Created: 2026-06-02 Version: 1.0.0