跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
设计与 UI

demo-5-sandbox-feature-pr

Demo skill for the meta "decode improves decode" flow — launch decode in a sandboxed clone of the course repo, plan and build one small feature inside the Workspace, let Hand-back push the Session Branch, then open a draft PR against the course repo.

decodingai-magazine · 已有 AI 审核记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add decodingai-magazine/building-a-coding-agent-from-scratch-course --skill demo-5-sandbox-feature-pr

技能说明


name: demo-5-sandbox-feature-pr description: Demo skill for the meta "decode improves decode" flow — launch decode in a sandboxed clone of the course repo, plan and build one small feature inside the Workspace, let Hand-back push the Session Branch, then open a draft PR against the course repo.

The meta demo: decode improves decode. Run decode inside an isolated sandbox Workspace that is a clone of this very course repo, plan and implement one small self-contained feature there, and let the harness ship the work back as a branch you turn into a draft PR.

Nothing the model does touches your host checkout — the Workspace is an isolated clone, and the branch only lands on your host through Hand-back (ADR-0012 §8).

1. Launch decode against a sandboxed clone of the course repo (human step)

Sandbox mode is selected by the SANDBOX_MODE environment variable — it is an env var, not a command-line flag; the repo to clone into the Workspace is the --repo flag. Launch the local docker rung (Docker must be running):

SANDBOX_MODE=docker decode --repo git@github.com:decodingai-magazine/building-a-coding-agent-from-scratch-course.git

This clones the course repo into the isolated Workspace (/workspace ≡ host .decode/sandbox) and drops you into the REPL, with bash and the file tools scoped to that clone.

To run the same demo on the remote rung instead, swap SANDBOX_MODE=docker for SANDBOX_MODE=modal — everything else is identical (Modal credentials must be configured).

2. Plan the feature — in plan mode

If you are decode reading this inside the Workspace, start here — step 1 is the human's launch command, already done; do NOT try to run decode yourself. Enter plan mode by calling enter_plan_mode with no arguments (it takes none — you present the plan later, to exit_plan_mode), explore the code read-only, then present a short plan before touching anything. Pick ONE feature — small, self-contained, provable in a single session:

  • A decode --version flag that prints the installed package version, plus its unit test.
  • One small pure helper (e.g. in tools/ or entities/) with a focused unit test.
  • A scoped docstring/README fix that resolves a real inaccuracy you found while exploring.

The plan should name the files to touch, the test that proves the feature, and nothing else. Exit plan mode (exit_plan_mode) once the human approves it.

3. Build it inside the Workspace

Implement the plan: edit, then run the suite inside the Workspace (uv run pytest -q / make unit-tests) until it is green. Keep the final green test output — it goes in the PR body. All of this happens on the sandboxed clone.

4. Get the branch onto the remote

The branch has to be on origin before a PR can point at it. There are two ways it gets there — know which one you are in:

  • Sandbox run (the intended demo, SANDBOX_MODE=docker/modal): just leave the REPL (exit, or /ship while idle). On exit, Hand-back secures the final Workspace onto a deterministic decode/<session-id> Session Branch, auto-commits any uncommitted model work, and git pushes it back to the course repo with your ambient host git credentials — every git command runs host-side, so no credential ever enters the sandbox. The friendly exit line names the branch it pushed. You do not push by hand in this path.

  • No sandbox, or your own feature branch: if decode is running straight on the host (no SANDBOX_MODE), there is no Hand-back — commit your work on a feature branch and push it yourself:

    git push -u origin <your-branch>
    

    Give this push a generous timeout (≥ 180s) and do not kill it. This repo installs a pre-push hook that runs the full unit suite (make unit-tests, ~2 minutes) before the push goes out — a short tool timeout will terminate it mid-run and look like a network hang. It is the gate doing its job; you already ran the suite green in step 3, so just let it finish. Do not reach for git push --no-verify to dodge it — that skips the format/lint/test gate the push is supposed to enforce. Only fall back to --no-verify if the human explicitly tells you to.

5. Open a draft PR against the course repo

Turn the pushed branch into a draft PR with the gh CLI — --head decode/<session-id> for the Hand-back path, or --head <your-branch> for the manual path. Build the body with a heredoc so newlines and backticks land literally (never hand-escape \n into --body):

gh pr create --draft \
  --repo decodingai-magazine/building-a-coding-agent-from-scratch-course \
  --head <branch> --title "<feature>" \
  --body "$(cat <<'EOF'
<what & why, 2-3 lines>

Test evidence:
```
<the green pytest summary line>
```

Built end-to-end inside a decode sandbox Workspace (SANDBOX_MODE=docker) and shipped by Hand-back.
EOF
)"

Report the feature you built, the branch it landed on, and the draft-PR URL.