elffuss-code / README.md
KikoCis's picture
README: los numeros del gestor de contexto, con referencias
022d4ca verified
|
Raw
History Blame Contribute Delete
4.7 kB
metadata
title: Elffuss Code
emoji: 🧑‍💻
colorFrom: purple
colorTo: pink
sdk: static
app_file: index.html
pinned: false
license: apache-2.0
short_description: A VS Code-style IDE with a local AI agent, in-browser
tags:
  - webgpu
  - transformers.js
  - local-llm
  - agent
  - code
  - gemma

🧑‍💻 Elffuss Code — a VS Code-style IDE with an AI soul, 100% in your browser

Open a local folder and a full web IDE unfolds around it — real Monaco (the VS Code editor), a file tree, tabs — with an AI agent that reads, searches and edits your project for real. The model runs on your GPU via WebGPU. Your code never leaves your machine.

Build the best thing you can with Elffuss — 15-minute challenge, open until Sun 9 Aug 2026. Open one of your own folders and let the agent fix a real bug or add a small feature in 15 minutes, then post a screenshot of the diff (plus your prompt) as a reply on X or a comment on LinkedIn. Every like counts, wherever you post it — X, LinkedIn or the Community tab of this Space. We tally them all on Sunday; the most-liked entry wins and gets featured in this Space and in the repo.

⚠️ Open it full-screen (direct link) — the File System Access API (opening a local folder) is blocked inside the embedded iframe.

demo

In the demo a small model running entirely in the browser reads calc.py, finds a real bug and fixes it with a minimal edit — genuine tool-calling, not autocomplete.

Why it's different

  • Your code stays local. The folder is opened via the File System Access API — no upload, no account.
  • The model runs on your GPU (WebGPU, Gemma-4 E4B via LiteRT-LM; ONNX/transformers.js + our healed LFM2.5-1.2B as fallback).
  • Real tool-calling: code.tree, code.read, code.search, code.edit, code.write — you watch every call.
  • Real editor: Monaco, file tree, tabs, Ctrl+S, and git branch/last commit parsed straight from .git.
  • Zero install, zero build. Vanilla ES modules.
  • Your language: the whole IDE follows your browser's language (English, Spanish and more).
  • External providers (OpenAI, Anthropic, Ollama) are opt-in; keys stay in your browser.
  • Optional local bridge for real node / npm / python execution — downloaded from the main site.

How it's built

Four write-ups on the full stack — model, runtime, agent harness, editor and filesystem access, all in one tab: https://bitacora.utopiaia.com/#elffuss

Links

Gestión de contexto — cómo se mide

El empaquetador de contexto no se elige por intuición. Números medidos a igual presupuesto de tokens:

Sesiones reales de agente (25 sesiones, 174 sondas, recuerdo de hechos, sin juez LLM):

recuerdo
truncar por la cola 7,0 %
versión anterior 15,1 %
esta 65,3 % ± 8,5

Memoria a largo plazo (200 preguntas, F1 del propio repo del banco, modelo real respondiendo, ~8 % del contexto):

F1
el contexto COMPLETO, sin comprimir 22,56
truncar por la cola 6,62
heurísticas escritas a mano 6,07
BM25 sin IDF 20,34
BM25 23,59
embeddings 23,95
fusión por rangos (BM25 + embeddings) 28,09

Recuperar bien bate a tenerlo todo: 28,09 frente a 22,56 usando el 8 % de los tokens. Lo irrelevante no es lastre neutro — distrae.

Otras dos cosas que se miden y no se suponen: activar las heurísticas costaba −3,02 F1, y quitar la IDF cuesta −3,25. Lo semántico no sustituye al léxico, le cubre el punto ciego: sin solape de vocabulario entre pregunta y respuesta los embeddings ganan (24,28 vs 18,60), con solape gana BM25 (29,33 vs 23,58), y la fusión se queda con los dos.

Referencias: BM25 (Robertson et al., TREC-3 1994) · IDF (Spärck Jones, 1972) · RRF (Cormack et al., SIGIR 2009) · MMR (Carbonell & Goldstein, SIGIR 1998) · sumideros de atención (Xiao et al., ICLR 2024) · tiempo de validez (Snodgrass & Ahn, SIGMOD 1985).

La historia completa, con lo que salió mal: https://bitacora.utopiaia.com/posts/16-beaten-by-doing-nothing.html