Problem Statement

After #43, users get early Ollama checks for YAML defaults only. The interactive wizard lets users invite custom agents and set orchestrator models interactively. Those models are not validated before run_session() starts, so users still hit LiteLLM/Ollama errors mid-session after passing startup preflight (especially with --skip-preflight or non-Ollama defaults + Ollama custom agents).

Proposed Solution

1. After wizard builds active_agent_configs (+ optional orchestrator), collect all ollama/… models from the final session config.

2. Reuse shared Ollama helper (issue feat: add timeout to Agent.generate_response() #2) to verify tags once before run_session().

3. On failure: print the same actionable panel as preflight; offer to return to wizard, edit model strings, or abort (do not start session with known-bad tags).

4. Respect --skip-preflight for both startup and post-wizard checks (or document --skip-preflight as global bypass).

5. Smoke test: mock validation hook; assert called with wizard-built configs.

Acceptance criteria

• Custom agent with missing Ollama tag blocked before first turn
• Orchestrator chosen in wizard validated even when YAML default is non-Ollama
• --skip-preflight bypasses both checks
• No duplicate /api/tags calls when tags unchanged from startup check

Alternatives Considered

• Only fix at inference (feat: surface inference failures in CLI with recovery hints (not as agent speech) #37) — better UX but wastes wizard setup time.
• Move all preflight to post-wizard only — rejected; YAML-only users lose early exit.

Priority

Medium 🟡

Additional Context

No response