vercel-labs/seal

최근 기록된 소스 활동: 2026년 7월 17일 17:25

SkillsMP 카탈로그 업데이트: 2026년 10월 29일 18:30

수집된 skills: 8

GitHub 스타: 8
GitHub 포크: 3

AI 어시스턴트로 설치

이 프롬프트를 복사해 사용 중인 AI 어시스턴트에 입력하세요.

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/vercel-labs/seal.

저장소 README

README 원본 파일: README.md

README 원본 보기

README 전체 펼치기

# seal A personal AI assistant built as a **durable agent**: every agent run is a Vercel workflow, so turns survive restarts, streams can be resumed mid-run, and tool calls can park indefinitely waiting for human approval. Seal is an example app for the [AI SDK for Python](https://ai-python.dev) (the `ai` package) and for [Workflows with Python](https://vercel.com/docs/workflows/python) (`vercel.workflow`). The agent (Claude via the AI Gateway) has three tools: `bash`, `web_fetch`, and `subagent`. Bash runs are gated behind an approval UI when run by the main agent, but not when run by a subagent. (That is silly, but this is a demo app.) ## How it works - **frontend/** — React + Vite chat UI using the AI SDK (`useChat`) and [AI Elements](https://elements.ai-sdk.dev). Reconnecting to a session re-tails the in-flight stream (`useChat({ resume: true })`). - **backend/app/** — FastAPI service. `POST /api/chat` starts (or resumes) a run and streams the AI SDK UI message protocol; other endpoints cover sessions, titles, and private blob attachments. See `app/server.py` for the endpoint list. - **backend/agent/** — the durable agent itself. `driver.py` runs a `run_session` workflow that spawns one child `run_turn` workflow per agent turn and suspends on a hook until it finishes. Tool approvals are workflow hooks too: the turn parks until the user answers, then resumes with the decision. Model calls, stream writes, and session snapshots are all workflow steps, replay-safe via the workflow's deterministic RNG/clock. - **Storage** — durable streams and session snapshots are stored on the workflow SDK's run streams (`agent/stream.py`). Session metadata (`app/sessions.py`) uses Postgres when `DATABASE_URL` is set, local JSON files otherwise. Uses Vercel Blob to store attachments when available. Deployment is two Vercel services (see `vercel.json`): the frontend and the backend, with the workflow worker declared in `backend/pyproject.toml`. ## Development Prereqs: [uv](https://docs.astral.sh/uv/), [pnpm](https://pnpm.io), and the [Vercel CLI](https://vercel.com/docs/cli). ```sh ./dev-setup.sh # sync backend deps (works around a vercel-worker version override) cd frontend && pnpm install vercel dev # serves frontend + backend + worker on :3000 ``` Environment: `AI_GATEWAY_API_KEY` (model access), optional `DATABASE_URL` (Postgres storage), and a blob token for attachments. ### Checks ```sh make ci # everything below make ci-backend # uv sync, ruff, mypy, ty, pytest make ci-frontend # pnpm install, prettier, eslint, tsc, vitest, build ``` ### E2E tests `e2e/` drives a real browser against a running instance: ```sh cd e2e && pnpm install && pnpm run install-browser pnpm test # expects the app at http://localhost:3000 pnpm run test:images # image latency: time to first image, time to all N ``` `test:images` prompts "draw N pictures of things you find interesting" (`N=5` by default) and reports when each image actually painted, measured from the submit click. Timings also land in `/tmp/seal-e2e-images-summary.json`. ## Deployment Deploy as a project to Vercel with `vc deploy`. `DATABASE_URL` must point to a Postgres database, which can most easily be done by configuring a marketplace integration with Neon or similar.

저장소 README를 불러오는 중…