feat: wire max_rounds safety boundary into model.act() (#97)

[VforVitorio]

Summary

First of two PRs for #97 (robust auto mode). This one is the safety core — stops runaway tool loops before they burn the context window. No visual changes; follow-up PR 2 (feat/auto-mode-ux) will add the spinner color per mode, round counter, and first-time auto warning.

What changed

• AgentSettings.max_rounds default 50→10, docstring explains the trade-off. The max_rounds field already existed but was never wired to model.act() — effectively dead code.
• _run_turn now passes max_prediction_rounds=get_settings().agent.max_rounds to model.act() in the non-strict branch. Strict mode is untouched (uses model.respond() which has no round concept).
• Limit-reached detection covers both SDK paths:
  • Well-behaved model: SDK forces a tool-free final round, model produces a text answer, ActResult.rounds == cap → flag set.
  • Stubborn model: model still emits tool_call on the final round, SDK raises LMStudioPredictionError("Model requested tool use on final prediction round."). _run_turn catches this specifically (matching on the "final prediction round" substring constant _MAX_ROUNDS_ERR_MARKER) and sets the flag. Any other LMStudioPredictionError re-raises. Critical: LMStudioPredictionError is a subclass of LMStudioServerError, which run() catches at the top level as a "LM Studio disconnected" signal. Without the specific catch, hitting the round limit would show the disconnect screen.
• run() prints an inline warning when the flag is set, matching the ctx window warning style (no panel, just coloured text above the Rule separator):

  ⚠ agent stopped after N rounds — raise the limit with LMCODE_AGENT__MAX_ROUNDS=N, set agent.max_rounds in config, or pass --max-rounds N on the CLI. Type to continue the conversation.

• CLI --max-rounds flag is now actually wired: mutates settings.agent.max_rounds for the session. Previously accepted by Typer but silently ignored. Default changed from 50 to None (fall back to config).

Tests

5 new tests in tests/test_agent/test_core.py:

• test_run_turn_passes_max_rounds_to_act — the kwarg flows from config into the SDK call.
• test_run_turn_max_rounds_none_when_config_zero — a non-positive max_rounds passes None to the SDK (which rejects values < 1).
• test_run_turn_flags_limit_reached_when_rounds_equal_cap — well-behaved case sets the flag.
• test_run_turn_flags_limit_reached_on_final_round_error — stubborn case catches the SDK error without re-raising.
• test_run_turn_reraises_unrelated_prediction_error — pins the catch to the marker string, not to LMStudioPredictionError broadly.

Also fixed _make_mock_model to return a mock with concrete rounds: int and total_time_seconds: float — the post-act comparison getattr(result, "rounds", 0) >= cap crashes on the default MagicMock.rounds with TypeError.

Test plan

• uv run lmcode chat --max-rounds 3, Tab to auto, ask for a task requiring ≥ 4 tool calls (e.g. "read calculator.py, greet.py and data.json then add a comment to each"). Expect the ⚠ warning after the model stops.
• uv run lmcode chat (default 10) with normal single-step tasks — should behave identically to before.
• LMCODE_AGENT__MAX_ROUNDS=1 uv run lmcode chat — env var override works.
• CI green.

Follow-up — PR 2 (feat/auto-mode-ux)

Depends on this PR being merged first (uses max_rounds as the denominator for the round counter). Will add:

• Spinner colour per permission mode (reuses existing _MODE_COLORS)
• Round counter in the spinner label (working… · round 3/10), auto mode only
• First-time auto warning on Tab into auto
• /status line showing current max_rounds
• Regression test: _always_allowed_tools survives ask → auto → strict → ask cycling

🤖 Generated with Claude Code