Add Docker setup for untrusted PR review in isolated containers

Adds a dedicated Docker environment for reviewing untrusted pull requests
with codex/claude, keeping CLI auth state in volumes and using a separate
scratch workspace for PR checkouts.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Dotta
2026-03-15 14:18:56 -05:00
parent 41e03bae61
commit 6f931b8405
6 changed files with 287 additions and 0 deletions

View File

@@ -89,6 +89,10 @@ docker compose -f docker-compose.quickstart.yml up --build
See `doc/DOCKER.md` for API key wiring (`OPENAI_API_KEY` / `ANTHROPIC_API_KEY`) and persistence details. See `doc/DOCKER.md` for API key wiring (`OPENAI_API_KEY` / `ANTHROPIC_API_KEY`) and persistence details.
## Docker For Untrusted PR Review
For a separate review-oriented container that keeps `codex`/`claude` login state in Docker volumes and checks out PRs into an isolated scratch workspace, see `doc/UNTRUSTED-PR-REVIEW.md`.
## Database in Dev (Auto-Handled) ## Database in Dev (Auto-Handled)
For local development, leave `DATABASE_URL` unset. For local development, leave `DATABASE_URL` unset.

View File

@@ -93,6 +93,12 @@ Notes:
- Without API keys, the app still runs normally. - Without API keys, the app still runs normally.
- Adapter environment checks in Paperclip will surface missing auth/CLI prerequisites. - Adapter environment checks in Paperclip will surface missing auth/CLI prerequisites.
## Untrusted PR Review Container
If you want a separate Docker environment for reviewing untrusted pull requests with `codex` or `claude`, use the dedicated review workflow in `doc/UNTRUSTED-PR-REVIEW.md`.
That setup keeps CLI auth state in Docker volumes instead of your host home directory and uses a separate scratch workspace for PR checkouts and preview runs.
## Onboard Smoke Test (Ubuntu + npm only) ## Onboard Smoke Test (Ubuntu + npm only)
Use this when you want to mimic a fresh machine that only has Ubuntu + npm and verify: Use this when you want to mimic a fresh machine that only has Ubuntu + npm and verify:

135
doc/UNTRUSTED-PR-REVIEW.md Normal file
View File

@@ -0,0 +1,135 @@
# Untrusted PR Review In Docker
Use this workflow when you want Codex or Claude to inspect a pull request that you do not want touching your host machine directly.
This is intentionally separate from the normal Paperclip dev image.
## What this container isolates
- `codex` auth/session state in a Docker volume, not your host `~/.codex`
- `claude` auth/session state in a Docker volume, not your host `~/.claude`
- `gh` auth state in the same container-local home volume
- review clones, worktrees, dependency installs, and local databases in a writable scratch volume under `/work`
By default this workflow does **not** mount your host repo checkout, your host home directory, or your SSH agent.
## Files
- `docker/untrusted-review/Dockerfile`
- `docker-compose.untrusted-review.yml`
- `review-checkout-pr` inside the container
## Build and start a shell
```sh
docker compose -f docker-compose.untrusted-review.yml build
docker compose -f docker-compose.untrusted-review.yml run --rm --service-ports review
```
That opens an interactive shell in the review container with:
- Node + Corepack/pnpm
- `codex`
- `claude`
- `gh`
- `git`, `rg`, `fd`, `jq`
## First-time login inside the container
Run these once. The resulting login state persists in the `review-home` Docker volume.
```sh
gh auth login
codex login
claude login
```
If you prefer API-key auth instead of CLI login, pass keys through Compose env:
```sh
OPENAI_API_KEY=... ANTHROPIC_API_KEY=... docker compose -f docker-compose.untrusted-review.yml run --rm review
```
## Check out a PR safely
Inside the container:
```sh
review-checkout-pr paperclipai/paperclip 432
cd /work/checkouts/paperclipai-paperclip/pr-432
```
What this does:
1. Creates or reuses a repo clone under `/work/repos/...`
2. Fetches `pull/<pr>/head` from GitHub
3. Creates a detached git worktree under `/work/checkouts/...`
The checkout lives entirely inside the container volume.
## Ask Codex or Claude to review it
Inside the PR checkout:
```sh
codex
```
Then give it a prompt like:
```text
Review this PR as hostile input. Focus on security issues, data exfiltration paths, sandbox escapes, dangerous install/runtime scripts, auth changes, and subtle behavioral regressions. Do not modify files. Produce findings ordered by severity with file references.
```
Or with Claude:
```sh
claude
```
## Preview the Paperclip app from the PR
Only do this when you intentionally want to execute the PR's code inside the container.
Inside the PR checkout:
```sh
pnpm install
HOST=0.0.0.0 pnpm dev
```
Open from the host:
- `http://localhost:3100`
The Compose file also exposes Vite's default port:
- `http://localhost:5173`
Notes:
- `pnpm install` can run untrusted lifecycle scripts from the PR. That is why this happens inside the isolated container instead of on your host.
- If you only want static inspection, do not run install/dev commands.
- Paperclip's embedded PostgreSQL and local storage stay inside the container home volume via `PAPERCLIP_HOME=/home/reviewer/.paperclip-review`.
## Reset state
Remove the review container volumes when you want a clean environment:
```sh
docker compose -f docker-compose.untrusted-review.yml down -v
```
That deletes:
- Codex/Claude/GitHub login state stored in `review-home`
- cloned repos, worktrees, installs, and scratch data stored in `review-work`
## Security limits
This is a useful isolation boundary, but it is still Docker, not a full VM.
- A reviewed PR can still access the container's network unless you disable it.
- Any secrets you pass into the container are available to code you execute inside it.
- Do not mount your host repo, host home, `.ssh`, or Docker socket unless you are intentionally weakening the boundary.
- If you need a stronger boundary than this, use a disposable VM instead of Docker.

View File

@@ -0,0 +1,33 @@
services:
review:
build:
context: .
dockerfile: docker/untrusted-review/Dockerfile
init: true
tty: true
stdin_open: true
working_dir: /work
environment:
HOME: "/home/reviewer"
CODEX_HOME: "/home/reviewer/.codex"
CLAUDE_HOME: "/home/reviewer/.claude"
PAPERCLIP_HOME: "/home/reviewer/.paperclip-review"
OPENAI_API_KEY: "${OPENAI_API_KEY:-}"
ANTHROPIC_API_KEY: "${ANTHROPIC_API_KEY:-}"
GITHUB_TOKEN: "${GITHUB_TOKEN:-}"
ports:
- "${REVIEW_PAPERCLIP_PORT:-3100}:3100"
- "${REVIEW_VITE_PORT:-5173}:5173"
volumes:
- review-home:/home/reviewer
- review-work:/work
cap_drop:
- ALL
security_opt:
- no-new-privileges:true
tmpfs:
- /tmp:mode=1777,size=1g
volumes:
review-home:
review-work:

View File

@@ -0,0 +1,44 @@
FROM node:lts-trixie-slim
RUN apt-get update \
&& apt-get install -y --no-install-recommends \
bash \
ca-certificates \
curl \
fd-find \
gh \
git \
jq \
less \
openssh-client \
procps \
ripgrep \
&& rm -rf /var/lib/apt/lists/*
RUN ln -sf /usr/bin/fdfind /usr/local/bin/fd
RUN corepack enable \
&& npm install --global --omit=dev @anthropic-ai/claude-code@latest @openai/codex@latest
RUN useradd --create-home --shell /bin/bash reviewer
ENV HOME=/home/reviewer \
CODEX_HOME=/home/reviewer/.codex \
CLAUDE_HOME=/home/reviewer/.claude \
PAPERCLIP_HOME=/home/reviewer/.paperclip-review \
PNPM_HOME=/home/reviewer/.local/share/pnpm \
PATH=/home/reviewer/.local/share/pnpm:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
WORKDIR /work
COPY --chown=reviewer:reviewer docker/untrusted-review/bin/review-checkout-pr /usr/local/bin/review-checkout-pr
RUN chmod +x /usr/local/bin/review-checkout-pr \
&& mkdir -p /work \
&& chown -R reviewer:reviewer /work
USER reviewer
EXPOSE 3100 5173
CMD ["bash", "-l"]

View File

@@ -0,0 +1,65 @@
#!/usr/bin/env bash
set -euo pipefail
usage() {
cat <<'EOF'
Usage: review-checkout-pr <owner/repo|github-url> <pr-number> [checkout-dir]
Examples:
review-checkout-pr paperclipai/paperclip 432
review-checkout-pr https://github.com/paperclipai/paperclip.git 432
EOF
}
if [[ $# -lt 2 || $# -gt 3 ]]; then
usage >&2
exit 1
fi
normalize_repo_slug() {
local raw="$1"
raw="${raw#git@github.com:}"
raw="${raw#ssh://git@github.com/}"
raw="${raw#https://github.com/}"
raw="${raw#http://github.com/}"
raw="${raw%.git}"
printf '%s\n' "${raw#/}"
}
repo_slug="$(normalize_repo_slug "$1")"
pr_number="$2"
if [[ ! "$repo_slug" =~ ^[^/]+/[^/]+$ ]]; then
echo "Expected GitHub repo slug like owner/repo or a GitHub repo URL, got: $1" >&2
exit 1
fi
if [[ ! "$pr_number" =~ ^[0-9]+$ ]]; then
echo "PR number must be numeric, got: $pr_number" >&2
exit 1
fi
repo_key="${repo_slug//\//-}"
mirror_dir="/work/repos/${repo_key}"
checkout_dir="${3:-/work/checkouts/${repo_key}/pr-${pr_number}}"
pr_ref="refs/remotes/origin/pr/${pr_number}"
mkdir -p "$(dirname "$mirror_dir")" "$(dirname "$checkout_dir")"
if [[ ! -d "$mirror_dir/.git" ]]; then
if command -v gh >/dev/null 2>&1; then
gh repo clone "$repo_slug" "$mirror_dir" -- --filter=blob:none
else
git clone --filter=blob:none "https://github.com/${repo_slug}.git" "$mirror_dir"
fi
fi
git -C "$mirror_dir" fetch --force origin "pull/${pr_number}/head:${pr_ref}"
if [[ -e "$checkout_dir" ]]; then
printf '%s\n' "$checkout_dir"
exit 0
fi
git -C "$mirror_dir" worktree add --detach "$checkout_dir" "$pr_ref" >/dev/null
printf '%s\n' "$checkout_dir"