Feedback: multi-model orchestration session (Claude Code + Workflow tool)

Not a bug — sharing an outstanding session pattern.

Used Claude Code (Fable 5) as orchestrator to run a design "home-run derby" across 6 models (Opus 5, Sonnet 5, Fable, GPT-5.6 Sol/Terra via CLI, Grok 4.5): two competitive rounds, anonymized cross-model judging panels with a published scoreboard, plus a 13-seat brainstorm with adversarial critic/adjudicator meta-seats — all banked to git between stages.

The Workflow tool's per-agent model/effort control and background task notifications made a genuinely novel multi-vendor deliberation workflow possible, and the quality delta over single-model brainstorming was obvious (the judges caught real factual errors in entries, unanimous winner on an axis no other seat found). More first-class support for this pattern would be 🔥.