https://github.com/acoyfellow/loop
prepare repos for looping.
https://github.com/acoyfellow/loop
ai ai-agents git loops
Last synced: about 5 hours ago
JSON representation
prepare repos for looping.
- Host: GitHub
- URL: https://github.com/acoyfellow/loop
- Owner: acoyfellow
- License: mit
- Created: 2026-01-18T15:51:00.000Z (6 months ago)
- Default Branch: main
- Last Pushed: 2026-05-26T20:41:43.000Z (about 1 month ago)
- Last Synced: 2026-05-26T22:16:38.692Z (about 1 month ago)
- Topics: ai, ai-agents, git, loops
- Language: TypeScript
- Homepage:
- Size: 313 KB
- Stars: 3
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- Contributing: CONTRIBUTING.md
- License: LICENSE
- Security: SECURITY.md
Awesome Lists containing this project
README
# loop
A personal chatbot that **ships working artifacts**. Every reply can compile a real Svelte 5 widget next to the conversation. The assistant also takes notes ("remember that I prefer dark themes") and they stick. One long-running session per account — older turns roll into searchable long-term memory; press **reset** to start over.
Live: · sign-up requires an invite password.
*Six artifacts in one session: Mandelbrot fractal, orbital planets with trails, animated audio spectrum, Matrix-style katakana rain, scrolling stock sparkline, and a rippling hexagonal grid. Each one is real Svelte 5 — generated by Kimi-K2.6, compiled on the Worker, mounted in a sandboxed iframe.*
```
you ──▶ chat session (Durable Object)
├─ messages last 16 turns kept verbatim
├─ recall everything older, embedded in Vectorize
├─ facts typed notes the assistant decided to save
└─ artifacts Svelte 5 widgets compiled on the edge
```
Each turn runs through Workers AI. When the model calls the `panel` tool, the Worker compiles the generated Svelte 5 and mounts it next to the chat. When it calls `remember`, a typed fact is saved. As the conversation grows, older messages are embedded with bge-base-en and pushed into Cloudflare Vectorize; the model can recall them via `search_context`. Nothing is mocked.
## Run
Requires Bun, Node 22+, and a Cloudflare-authenticated `wrangler` with Workers AI access. Local development omits the optional Vectorize binding, so old messages are not searchable there; production creates and binds the index through Alchemy.
```sh
bun install
bun run dev # starts both the SvelteKit app and the Worker
```
Open . The live deployment lives at .
Try:
> Create a panel called build-status that lists three repos in cyan. Remember that cyan means a running experiment.
You will see, in this order:
1. the message appear in **chat**;
2. a `running…` placeholder while Workers AI works (you can send another message immediately);
3. a compiled Svelte widget mounted in **artifacts**;
4. a saved fact appear in **facts**;
5. the exact generated `.svelte` under **source** with syntax highlighting.
Reload — everything survives.
## Animation panels
Generated panels may import two animation libraries directly:
```svelte
import * as THREE from "three";
// or: import { DotLottie } from "@lottiefiles/dotlottie-web";
```
The iframe import map pins Three.js `0.179.1` and dotlottie-web `0.49.0` from
esm.sh. The Worker enforces the import allowlist at compile time; arbitrary
package imports are rejected. Three.js panels are model-authored and must clean
up render loops, observers, listeners, GPU resources, and renderers. Explicit
Lottie requests use a compact typed `lottie_panel` tool (`spinner`, `pulse`, or
`signal`) so the model chooses intent and colors while Loop generates known-good
Lottie JSON and lifecycle cleanup; this avoids long, fragile source tool calls.
An empty loop offers two deterministic starter artifacts:
- **Three.js orbital field** — a pointer-reactive, lit 3-D scene with particles.
- **Lottie signal constellation** — declarative vector rings rendered from
embedded Lottie JSON.
They use the same compile, persistence, iframe, source viewer, and delete paths
as model-authored artifacts; they are examples, not mocked previews. After ~16 turns, older messages start getting embedded into Vectorize; the model can search them on demand.
## Verify
```sh
bun run check # SvelteKit + Worker type-check
bun run test # unit tests
bun run verify # check + test + build (runs automatically on git push)
bun run e2e # real Workers AI: artifact, fact, durable reload
bun run stress # real inference + rolling window + Vectorize recall + export
```
`e2e` and `stress` call Workers AI and consume usage.
## Ship
CI is manual-only. `verify` runs locally on every `git push` via `.githooks/pre-push` (installed automatically by `bun install`). To deploy to :
```sh
bun run ship # verify + push + trigger deploy workflow + watch
```
Skip the local gate with `git push --no-verify` or `LOOP_SKIP_VERIFY=1`. Trigger workflows by hand any time with `gh workflow run {deploy,verify,inference} --repo acoyfellow/loop`.
## Layout
```
src/routes/+page.svelte UI: chat, artifacts, source, inspector
src/routes/+page.server.ts load chat snapshot
src/routes/data.remote.ts getThread, sendMessage, resetThread
src/lib/highlight.ts Shiki (vitesse-dark, lang=svelte) lazy loaded
worker/index.ts HTTP boundary + owner routing
worker/LoopDO.ts Loop extends Think — messages, tools, facts, artifacts, recall
worker/recall.ts Vectorize-backed long-term memory (bge-base-en, top-k)
worker/panels.ts Svelte 5 compile
worker/types.ts vocabulary
wrangler.local.jsonc local DO + remote Workers AI (optional MEMORY binding omitted)
alchemy.run.ts Cloudflare deployment graph
tests/ unit + Playwright + API stress
```
That is the whole core.
## How memory works
There is exactly **one chat session per account**. No conversation list, no folders, no compaction, no auto-summarization.
- The most recent **16 messages** stay verbatim in the prompt every turn.
- When the count exceeds 16, the oldest pair is embedded (Workers AI `@cf/baai/bge-base-en-v1.5`, 768-dim) and upserted into a Cloudflare Vectorize index, namespaced by owner.
- The model gets a `search_context` tool. When the user asks about something out-of-window, it queries Vectorize, top-k 5, and pulls the matches back into the prompt.
- The **facts** table is the model's deliberate memory: typed entries (`preference`, `decision`, `fact`, `failure`, `open_loop`) created when the user explicitly says "remember…". These never get rolled out.
- **`reset`** wipes everything — messages, artifacts, facts, and Vectorize entries — and starts the same single session over.
## Contract
| Surface | Behavior |
| ------------ | ---------------------------------------------------------------------------------------------- |
| Chat | One SQLite-backed Durable Object per authenticated owner |
| Inference | Workers AI binding (default `@cf/moonshotai/kimi-k2.6`) |
| Window | Last 16 messages kept verbatim; older ones evicted to Vectorize each turn |
| Recall | Workers AI embeddings (`@cf/baai/bge-base-en-v1.5`) into a Vectorize index, scoped by owner |
| Facts | Typed records (`preference`, `decision`, `fact`, `failure`, `open_loop`); `wrong` / `forgotten` are signals, not deletions |
| Artifacts | Model emits complete Svelte 5 source; Worker compiles it; iframe mounts the result |
| Animation imports | Allowlisted `three` and `@lottiefiles/dotlottie-web`, pinned through the iframe import map |
| Idempotency | `requestId` on `/api/messages` prevents duplicate turns |
| Export | `GET /api/export` returns the full chat history and every artifact revision |
## Boundaries
- Generated Svelte runs in a sandboxed iframe with no owner credentials or Worker bindings.
- Backend mutations are fixed typed actions, not generated code.
- Auth: Better Auth + D1. Local development uses a fixed owner so login is not required.
- Production sign-up is gated by an invite password (`LOOP_INVITE_PASSWORD`). Sign-in for existing accounts is unaffected. When the env var is unset, sign-up is open — only use that locally.
## Provenance
`loop` originally prepared repositories for repeated agent runs. This is the same idea at app scale: the vessel is now one durable owner session the model can read, write, and rebuild artifacts inside.
Patterns reused:
- [`remote`](https://github.com/acoyfellow/remote) — authenticated SvelteKit + Durable Objects shell.
- [`svelte-edge`](https://github.com/acoyfellow/svelte-edge) — Svelte 5 source compiled at the edge.
- [`deja`](https://github.com/acoyfellow/deja) — durable selected memory discipline.
MIT.