OpenCode (DeepSeek V4 Pro via OpenRouter) in a Vercel Sandbox. Starter only (AGENTS.md, serve.py, player/, helpers/dom/view_kit). No hardcoded demo code. Cursor did not edit workspace/.
2026-10-15 · task -18pE0jzLZc1 · opencode · openrouter/deepseek/deepseek-v4-pro-0813
Build a real Linux desktop I can use for short jobs. Each round is one assignment, like creating a file with exact text, or looking something up and writing the answer down. I see the actual screen. I can click, type, open programs, and switch windows. When I say I am done, score the result: the files and the state I left, not the particular clicks I used. Keep the screens so I can replay the episode without doing it live again. Restarting a job should be a clean machine, not last round's files. Several different jobs on the same desktop.
Running the prompt-created workspace on this machine — not a mock. Save an episode, then open Replay below to scrub the timeline.
opening · 0 ms
Exact files the agent wrote — not rewritten by hand.
# AI : shared Action contract — the page posts {"move": "..."}; keep this shape.
from pydantic import BaseModel
class Action(BaseModel):
move: str