feat: surface inference failures in CLI with recovery hints (not as agent speech)

[rosspeili]

Problem Statement

When LiteLLM or a custom function fails, Agent.generate_response() returns strings such as [Error: ...] or [Timeout Error: ...]. cli.py prints these as normal agent dialogue (colored name + content).

Users cannot distinguish a failed inference from in-character speech, and there is no in-session guidance (retry turn, skip agent, check model/API key, increase timeout).

Proposed Solution

• Detect error/timeout prefixes in the session loop (small helper, e.g. is_inference_failure(content)).
• Render a distinct Rich panel (not agent voice): model id, failure type, and hints (Ollama down, wrong tag, API key, raise defaults.timeout in settings).
• Optional prompts: retry this turn / skip agent / end session (CLI only; does not replace agent-level retries in feat: implement retry_count for agent completions #8).
• Add unit tests with mocked Session.generate_next_turn or agent responses.

Alternatives Considered

• Only implement feat: implement retry_count for agent completions #8 retry_count in agent.py — rejected; users still see opaque error text as dialogue if all retries fail.
• Raise exceptions and crash the session — rejected; too brittle for long multi-agent runs.
• Log errors only to stderr — rejected; misses the primary UX surface (terminal).

Priority

High 🔴

Additional Context

Works with timeouts already on main and with future #8 retries (retry in agent, clear UI in CLI). Not a duplicate of #23 (Skillware/tools). Consider doing after or in parallel with preflight (#2 above).