Feature Models Round Table

Take a feature request from a multi-model consensus on what to build, straight through to an implemented, reviewed, tested change. This skill is intentionally thin: it sequences two existing skills and passes their artifacts. It does not re-implement either — to change how consensus or building works, change models-roundtable or implement-feature, not this skill.

Pipeline: models-roundtable (answer → consensus) → implement-feature (decompose + build).

Steps

1. Consensus understanding/plan. Run models-roundtable on the feature request — frame the task for it as scoping/approach, e.g. "Determine what this feature requires and the best approach: <request>." Do not pre-analyze the request yourself; the bias-free fan-out is the point of models-roundtable. It returns a Consensus answer + Agreements + Open caveats, written to .ai-workflow/roundtable/<id>/report.md.

2. Resolve caveats (gate). If the report has Open caveats (low-confidence or orchestrator-decided points), settle them with the user before any code — unless --auto, then record an assumption per caveat. This satisfies implement-feature's intake gate.

3. Build. Invoke implement-feature with the settled consensus as the plan (--from-roundtable .ai-workflow/roundtable/<id>/report.md). It takes the Consensus answer as the plan, breaks it into to-tasks vertical-slice tasks, runs implement-and-review per task (independent tasks in parallel, FE/BE cross-reviewed TDD), integrates in dependency order, then runs the feature-wide full-review, leaving tests/build green.

4. Report. Relay implement-feature's final report, prefaced with the consensus summary and how any caveats were resolved. Return both report paths.

Rules