Bug: Streaming chat crashes with 'Unsupported renderer' and cancel scope error

[lbedner]

Bug Description

Streaming chat crashes with two linked errors:

1. Marko markdown rendering fails with "Unsupported renderer"
2. Async cleanup crashes with "Attempted to exit cancel scope in a different task"

Error Messages

Streaming error: Unsupported renderer <class 'marko._Renderer'>
Error: Unsupported renderer <class 'marko._Renderer'>

[error] Streaming failed: Attempted to exit cancel scope in a different task than it was entered in

Followed by full crash with asyncio.exceptions.CancelledError.

Steps to Reproduce

1. Start interactive chat: my-app ai chat
2. Switch to a model like gpt-5-mini: /model gpt-5-mini
3. Send a message that generates a response with certain markdown elements
4. Chat crashes with the above errors

Root Cause Analysis

Issue 1: Marko Renderer

The custom TerminalRenderer in app/cli/marko_terminal_renderer.py may not have handlers for all markdown element types. When marko encounters an element without a matching render_* method, it falls back to the default _Renderer which throws the error.

Likely missing renderers for:

• Certain GFM elements (task lists, strikethrough, etc.)
• Edge case elements the AI generated

Issue 2: Cancel Scope Crossing

When the marko error occurs mid-stream:

• The streaming async generator is interrupted
• pydantic_ai's streaming cleanup runs via cancel_and_wait_for_background_tasks()
• The cancel scope was entered in one task but exits in another
• This is a known anyio/pydantic_ai interaction issue

Files Involved

• app/cli/marko_terminal_renderer.py - Custom renderer (needs more element handlers)
• app/cli/ai_rendering.py - Streaming renderer
• app/services/ai/service.py - stream_chat() method
• app/cli/ai.py - Interactive chat loop

Suggested Fix

1. Add fallback handling in TerminalRenderer - Catch unknown elements and render as plain text instead of crashing
2. Wrap streaming in try/except - Gracefully handle rendering errors without crashing the whole session
3. Consider anyio compatibility - May need to ensure cancel scopes don't cross task boundaries

Context

Error occurred after switching models mid-session. The AI response contained markdown that triggered the renderer issue.