Epic: end-to-end AI orchestration for building web apps (the ai-autopilot thesis)

[suleimansh]

The gap

Flue, Pi, and Omnigent are generic agent infra: run any agent, at the harness level. None ship the web-app-specific layer: personas, prompts, and flows that actually know the stack.

The bet

GemStack ai-autopilot: end-to-end for building web apps, opinionated about the stack (vike-* + universal-orm), with personas + prompts baked in, on top of ai-sdk. We sit on harnesses (Pi, Omnigent, WebContainer, Docker), we don't compete with them. Data stays separate (universal-orm) — we skip the DB integration those frameworks bundle.

Ties into the "declare intent, decouple implementation" idea: the app expresses intent ("render this as a toaster"), the implementation is decoupled — a hard guardrail against AI getting UI wrong.

Children

• ai-autopilot: personas + prompts library (stack-aware knowledge) #98 — Personas + prompts library (stack-aware knowledge; the differentiating value; first)
• ai-autopilot: pluggable execution runner (WebContainer + Docker) #99 — Pluggable execution runner (interface + FakeRunner DONE in feat(ai-autopilot): pluggable runner seam (interface + FakeRunner) #103; real adapters FlueRunner/WebContainer/Docker remain, infra-gated)
• ai-autopilot: surfaces (terminal / in-page / background) #100 — Surfaces (terminal / in-page / background; blocked on runner)
• ai-autopilot: Flue orchestration teardown (what to borrow) #101 — Flue orchestration teardown (what to borrow; informs the runner)

Related: #86 (connectors), #1 (umbrella).