feat: preflight check for Ollama reachability and configured model tag

[rosspeili]

Problem Statement

Users can finish the full wizard and only discover on the first agent turn that Ollama is not running, the base URL is wrong, or defaults.litellm_model (e.g. ollama/gemma4:e2b) does not match any tag from ollama list.

Settings (#28) expose the model in YAML but nothing validates it before play. Failures show up as generic [Error: ...] lines in the conversation, which feels broken and wastes setup time.

Proposed Solution

• After loading settings (and before or immediately after agent selection), run an optional preflight:
  • Check ollama.base_url / OLLAMA_API_BASE is reachable.
  • Verify the configured LiteLLM model tag exists (e.g. Ollama /api/tags or a small helper).
• Print a Rich panel with actionable fixes: start ollama serve, ollama pull <tag>, edit rooms.settings.yaml, or run python cli.py config init.
• Add --skip-preflight for automation/CI; mock HTTP in tests (no real Ollama in CI).
• Document in docs/EXAMPLES.md or ARCHITECTURE.

Alternatives Considered

• Rely on feat: implement retry_count for agent completions #8 retries only — rejected; retries do not help when the model tag does not exist.
• Fold into Implement ollama.auto_select_first from settings #30 auto_select_first — rejected; auto-select picks a tag; preflight validates the user's configured choice.
• Silent fallback to another model — rejected; surprising vs explicit warning.

Priority

High 🔴

Additional Context

Not a duplicate of #30 (auto-select), #29 (presets in wizard), #32 (config path naming), or #8 (retry loop). Pairs well with a future CLI error UX issue and with #29 once presets land.

[AaineeSinha]

Hey @rosspeili, I'd love to take this one on next.

[rosspeili]

Hey @AaineeSinha, glad you want to take #36, go for it! It's assigned to you.

A few pointers so the PR stays focused and easy to review:

Scope

• Only run preflight when the configured model string starts with ollama/ (skip OpenAI/cloud and custom-function agents for now).
• v1 can check settings.defaults.litellm_model before the wizard starts, no need to validate every agent/orchestrator model in the first PR.

Implementation sketch

• Add a small helper (e.g. rooms/ollama_preflight.py or a function in rooms/settings.py) that:
  1. Hits {settings.ollama.base_url}/api/tags (stdlib urllib is fine, no new deps).
  2. Parses tag names and checks whether the model after ollama/ exists (e.g. gemma4:e2b from ollama/gemma4:e2b).
• Call it from cli.py after load_settings(), before main_menu().
• Show one Rich panel with actionable hints (ollama serve, ollama pull <tag>, edit rooms.settings.yaml).
• Default: warn and let the user continue (or ask once with Confirm.ask), don’t hard-exit unless you think unreachable Ollama should block.

CLI

• Add --skip-preflight next to --config for CI/tests.

Tests

• Mock the HTTP response in tests/, same idea as test_cli_settings_smoke.py. No real Ollama in CI.

Docs

• Short note in docs/EXAMPLES.md is enough for this PR.

Out of scope here: #30 auto-select, #29 presets, #37 inference error UI, keep this PR to preflight only.

Branch off latest main when you start (helps avoid overlap with your other open PRs). Shout if you want a quick “does this hook look right?” before you go deep. Welcome aboard again!