Multi-skill walkthroughs.

Scopes the app — who it's for, what it retrieves over, the one job it has to nail.

Picks the platform and a kind: vector-db pack — pgvector, Pinecone, Turbopuffer, or Supabase Vectors — against your scale and budget.

Runs the Designer → Builder → Evaluator loop: chunking strategy, embedding model, retrieval pipeline, and a retrieval eval that scores recall before you wire generation.

Adds the generation model on top of retrieved context, with prompt caching on the system + retrieved chunks to cut cost and latency.

Runs the "Phase: AI app?" sweep — prompt-injection surfaces and tool-use boundaries, not just generic lint.

Gates on the audit, ships to staging, then production with a health check.

Scopes what the agent is for and where its PRs land before any topology is chosen.

Planner picks the subagent topology, tool budget, and loop shape; the Evaluator gates the build on a task fixture suite so regressions surface before merge.

Sets model routing — a stronger model for the orchestrator, cheaper workers for the fan-out — with caching on shared context.

The agent's PRs go through the same commit gate as a human's — branch, audit, commit, push, open the PR.

Produces a model-version migration diff PR — the SDK call sites, params, and any retired-model swaps in one reviewable change.

Re-runs the eval goldset on the new model and blocks if any score drops more than regression_epsilon.

Cuts the release once the regression suite is green — bumps the version, drafts the changelog, hands off to ship.