[li-yazhou]

Problem

In CLI mode, streaming reasoning tokens from models like DeepSeek-V4 are
printed one per line with a ✻ prefix, producing output like:

🐈 nanobot
✻ The
✻  user
✻  is
✻  asking
✻  me
✻  to
✻  confirm
✻  whether
✻  the
✻  nan
✻ obot
✻  agent
✻ .

This is because _print_cli_reasoning calls console.print() for every
individual delta token, and Rich's Console.print() adds a newline
each time.

Solution

Add a module-level buffer (_reasoning_buf) that accumulates incoming
reasoning deltas and only flushes to the display when:

• A newline is encountered
• Sentence-ending punctuation (. / ! / ? / 。 / ！ / ？) is reached
• The buffer exceeds ~60 characters

A new _flush_reasoning() helper handles the actual print, and both
the direct hook path (_cli_progress) and the bus path
(_maybe_print_interactive_progress) flush remaining content when
reasoning_end is signaled.

After

✻ The user is asking me to confirm whether the nanobot agent

Changes

• nanobot/cli/commands.py — 3 modifications:
  1. Module-level _reasoning_buf variable
  2. _print_cli_reasoning → buffer + flush; new _flush_reasoning
  3. Handle reasoning_end in both _cli_progress and
     _maybe_print_interactive_progress

[li-yazhou]

The effect before adding the buffer is shown in the image below.

[Re-bin]

This is ready to merge. The original fix addresses the CLI reasoning-token display issue, and the follow-up commit keeps the buffering state local to each CLI progress path instead of using module-level mutable state. I also added focused regression coverage for buffering until a sentence boundary and flushing on reasoning_end.

Verified locally:

• CLI/reasoning/channel focused tests: 43 passed
• Full pytest after installing the missing local test dependency lxml_html_clean: 3201 passed, 4 skipped