Fix blank responses on reasoning models

[matteo-brandolino]

Problem

When using regolo chat with reasoning models, the assistant appeared completely frozen — no output was printed,
neither during the thinking phase nor for the final response.

Two independent bugs caused this:

1. max_tokens too low

run_chat had a hardcoded default of max_tokens=200. Reasoning models spend a large portion of the token budget on internal thinking
(reasoning_content) before producing the visible answer. With only 200 tokens, the budget was exhausted during reasoning, leaving 0
tokens for the final response.

2. reasoning_content silently ignored

handle_search_text_chat_completions only read delta.get("content") and discarded reasoning_content entirely. During the whole thinking
phase, nothing was printed — the assistant looked stuck.

---

Solution

src/regolo/client/regolo_client.py

• Added an inner resolve() function that distinguishes content from reasoning_content and returns a "thinking" role tag for reasoning
  tokens.
• Added _state["in_reasoning"] to track the transition between thinking and answer phases.

src/regolo/cli.py

• Added --max-tokens flag (default: 2048, user-configurable) forwarded to run_chat.
• Added differentiated rendering for the thinking phase:
  • Thinking... header (dim + italic) on first reasoning token
  • Reasoning text rendered in dim style
  • Visual separator ───────────────────── on transition to the final answer

---

Test Plan

• Run regolo chat with a reasoning model ) and verify thinking tokens stream in real-time with dim styling
• Verify the ───────────────────── separator appears before the final answer, which is printed at full brightness
• Test --max-tokens 512 and --max-tokens 4096 to confirm the flag is forwarded correctly
• Verify non-reasoning models (no reasoning_content in deltas) continue to work as before