Epic: The AI framework — turnkey end-to-end AI orchestration for building software (web-app first)

[suleimansh]

The bet

A turnkey AI framework for building web apps. npm install and it works: prompts, personas, and flows already wired, opinionated about the stack (Vike/Next + universal-orm). Claude Code for the web, supercharged, not low-level agent infra you assemble yourself.

This extends the ai-autopilot thesis (#97, closed) from "the seed" (personas + runner + surfaces, all shipped) to the full product. Flue/Pi/Omnigent are generic harnesses we sit on; none ship the web-app-specific layer. That layer is the moat.

Scope (agreed 2026-07-02)

Agnostic core + domain presets. The engine (the loop + the state layer: decisions ledger, CODE-OVERVIEW, review/quality/security triggers) is language- and framework-agnostic and runs on any codebase; most prompts are generic. The web-app-specific value (Vike/Next personas, UI-flow => QA+UX) sits on top as a preset. Web-app is the flagship preset and the launch wedge, we lead the story there, not with "works on any code" (the crowded/generic pitch). Non-web presets (e.g. Python) are possible later.

Execution model (updated 2026-07-03, see #165): the shipped product wraps the Claude Code CLI and drives it as a black-box outer loop — we send prompts, read the code it produced, and gate on outcomes (builds / serves / review-passes / PR-exists); we never reach inside its loop. This lets users spend a Claude Max subscription and keeps us provider-swappable (Codex/opencode behind the same driver adapter). @gemstack/ai-sdk remains the lower-level, provider-agnostic engine (pay-per-token) that the orchestration primitives and any non-wrapped paths run on — it is the engine, not the shell. Guardrails (swappable driver, own the event stream) live in #165. (Verify Claude Code ToS before relying on the subscription path.)

Positioning

• Turnkey: zero wiring. npm install and go.
• Opinionated: knows Vike/Next + universal-orm. Data stays separate (universal-orm), as agreed.
• Open source, collaborative: the community grows the prompt/flow library via PRs, and it stays turnkey for everyone.
• Name (decided 2026-07-03): "The Framework", tagline "The (AI) Framework", domain the-framework.ai, npm @gemstack/framework (npm i -g). Positioning shorthand: "Vite for AI" (turnkey/zero-config) vs competitors' "webpack for AI" (low-level).

Children (buildable, in priority order)

Status (2026-07-02): the whole state layer + loop + bootstrap are built and merged. Only #115 (deliberately scope-open) and the infra-gated pieces remain.

• Built-in prompts library (review, quality, security, refactor, UX, QA, knowledge base) #111 — Built-in prompts library (P0, first): review (TLDR + thorough), code quality, security audit, refactoring, UX, QA, knowledge base / business context. The v1 and the main OSS contribution surface.
• Decisions bookkeeping — stop the AI re-pitching rejected ideas #112 — Decisions bookkeeping (P1): persist rejected ideas so the AI stops re-pitching them. Novel; nobody ships it.
• "The loop" — event-triggered prompt chains (major change → review/quality/security; UI flow → QA/UX) #113 — "The loop" (P1): event to prompt-chain policy. A major change fires review + code quality + security; a new UI flow fires QA + UX. Runs on the autopilot Supervisor + surfaces.
• Scale mode — maintain CODE-OVERVIEW.md for large-codebase navigation #114 — Scale mode (P2): maintain a CODE-OVERVIEW.md so the agent navigates a large codebase fast.
• Web-app preset: Next.js + Vike targeting (persona/prompt/loop layer) #115 — Web-app preset (Next.js + Vike): the flagship domain preset on top of the agnostic core; framework-specific personas/prompts at the persona/prompt/loop layer only. (Scope deliberately open pending v1 traction.)
• Bootstrap mode — zero to a production-grade app (scope prompt + full-fledged loop + narration) #116 — Bootstrap mode (P-high): zero to a production-grade app, scope prompt + full-fledged loop + dev-style narration. Likely our best first-run / acquisition hook. (Framework side done Generalize the surface stream to be event-type generic #120-123; only the live capstone Bootstrap capstone example + live verification (infra-gated) #124 is infra-gated.)
• AI framework: presentation layer — headless server + CLI/web chat UIs #165 — Presentation layer: headless server + two thin UIs (CLI + localhost web chat) on one backend, over the existing EventStream. Open decision: own shell vs. sit on opencode.

Parked (not children yet)

• Paid-hosting business model (decision, not a buildable issue). (Product name decided 2026-07-03 — see Positioning.)

Foundation: #97.

Landscape + positioning (validated 2026-07-02, competitive scan)

The "prompt collection + workflow + OSS PRs" positioning is already crowded: BMAD, GitHub Spec Kit, OpenSpec, GSD, Superpowers (170k+ combined stars) all sit there, betting on spec-driven dev. Cloudflare published their internal review-orchestration system (runs "the loop" at scale in CI; found AGENTS.md-style instruction files rot fast, so they built a reviewer that detects material changes and forces updates). Claude Squad / Baton orchestrate parallelism, not quality loops. LangGraph/CrewAI = infra layer, not competitors.

So: prompts are table stakes / the on-ramp, not the moat. Lead the product with the auto-maintained state layer + the loop, which is the whitespace:

• Decisions bookkeeping (Decisions bookkeeping — stop the AI re-pitching rejected ideas #112) — nobody has a rejected-ideas ledger. Most original.
• Semantic event-triggered loops ("The loop" — event-triggered prompt chains (major change → review/quality/security; UI flow → QA/UX) #113) — others are command-driven or CI-every-PR, not change-type-aware.
• Browser-agent QA on UI changes — nobody ships it inside a workflow framework (near-standalone).
• Auto-maintained CODE-OVERVIEW.md (Scale mode — maintain CODE-OVERVIEW.md for large-codebase navigation #114) — validated by Cloudflare's "instructions rot" finding.

Our bet (review/QA loops + persistent state) differs from the spec-driven pack. Lead with it so we don't read as "the sixth spec framework."

Which agent do we orchestrate? (updated 2026-07-03) The product wraps the Claude Code CLI as an outer loop; @gemstack/ai-sdk is the underlying provider-agnostic engine, not the shell. Wrapping a vendor CLI does raise a real "could Anthropic eat this layer" risk — the #165 guardrails are how we mitigate it: (1) the wrapped CLI is a swappable driver adapter (Claude Code first, Codex/opencode behind the same seam), so no single vendor is load-bearing; the only hard dependency is the code/outcome seam we own; (2) we own our own event stream (the Supervisor already emits it) rather than betting the UX on any one vendor's UI. The defensible value stays in the loop + state layer + team/multiplayer, not the shell.

Business model: keep the OSS local version from being so complete that hosting is pointless. The paid wedge is team/multiplayer (shared decision log, review history, org-wide business context), not "we run it for you."

[brillout]

AI framework for building web apps

I think we can generalize it for any source code. Python, JS, any Vite framework, etc.

The way to think about it is that it's like Claude Code (for the web) but supercharged with agent loops and built-in prompts/skills.

We run on @gemstack/ai-sdk (provider-agnostic runtime, calls model APIs directly). We do NOT wrap Claude Code / a CLI,

Can a Claude subscriptions (e.g. I have Max 20x) be used then? I was told that's only possible by wrapping Claude Code.

[brillout]

AI framework for building web apps

I think we can generalize it for any source code. Python, JS, any Vite framework, etc.

How about we do both: some prompts are tailored to web app building, while most prompts are generic (prog lang agnostic).

[suleimansh]

Scope: agreed, do both. The loop + state layer (decisions ledger, CODE-OVERVIEW, review/quality/security) is already language-agnostic, so the engine runs on any codebase; web-app bits (Vike/Next personas, UI-flow => QA+UX) sit on top as presets. I'd still lead the launch story with web-app building though, that's the wedge.

Claude sub: your info's right. ai-sdk calls the API (pay-per-token), doesn't spend a Max sub, that's tied to Claude Code. Fits the runner seam (#99): ai-sdk stays the default, add an optional "Claude Code runner" for people who want to use their sub. One optional backend, not the core.