feat: robust auto mode implementation

[VforVitorio]

What problem does this solve?

Currently, �uto mode is present but might feel a bit undercooked. It needs to seamlessly run tools without user intervention, but we also need to ensure it doesn't get into infinite loops and handles autonomous failures gracefully.

Proposed solution

1. Guarantee that �uto mode bypasses all confirmation prompts correctly.
2. Implement safety boundaries (e.g., maximum consecutive tool calls or a token limit) to prevent runaway LLM loops.
3. Establish a clear UI indicator (like a distinct spinner or logging format) so the user knows the agent is taking autonomous multi-step actions.
4. Integrate with the new Yes and always allow option from the extended interactive prompt.

Alternatives considered

Sticking to the current minimal implementation, but that risks a poor UX if the model gets stuck in an execution loop without user boundaries.

[VforVitorio]

Resolved in two stacked PRs, both merged to dev:

• feat: wire max_rounds safety boundary into model.act() (#97) #104 — max_rounds safety boundary: wires max_prediction_rounds into model.act(), catches the LMStudioPredictionError on stubborn final rounds, and flags well-behaved turns that hit the cap.
• feat: auto-mode UX — per-mode spinner, round counter, first-time warning (#97) #105 — auto-mode UX layer: per-mode spinner colour (orange/blue/red), live round N/M counter in auto mode, first-time auto-mode amber warning on Tab, and max rounds row in /status.

Default cap is now 10 (down from the previously unused 50), overridable via LMCODE_AGENT__MAX_ROUNDS=N, agent.max_rounds in config, or --max-rounds N on the CLI.