# ============================================

# DeepTutor Sandbox Runner Sidecar Image

# ============================================

# A deliberately small, least-privileged image whose *only* job is to execute

# untrusted shell commands on behalf of the main app, isolated in its own

# container. The main app talks to it over HTTP via RunnerSidecarBackend

# (see deeptutor/services/sandbox/backends.py), pointed here through

# DEEPTUTOR_SANDBOX_RUNNER_URL.

#

# Build/run (normally orchestrated by docker-compose, not by hand):

# docker build -f Dockerfile.runner -t deeptutor-sandbox-runner:local .

# docker run --rm -p 8900:8900 deeptutor-sandbox-runner:local

#

# Why no app code beyond server.py: the runner ships ONLY the stdlib HTTP server

# plus a set of common CLI tools. It must not depend on the DeepTutor package or

# any heavy framework — keeping the attack surface and image size minimal.

# ============================================

FROM python:3.11-slim

# ----- Common CLI + data tooling -------------------------------------------

# A pragmatic toolbelt for the kinds of shell tasks the model runs (clone a

# repo, fetch a URL, grep code, slice JSON). numpy/pandas are installed via pip

# so simple data crunching works out of the box. Trim this list if image size

# matters more than coverage for your deployment.

RUN apt-get update && apt-get install -y --no-install-recommends \

git \

curl \

ca-certificates \

ripgrep \

jq \

build-essential \

fonts-wqy-zenhei \

&& rm -rf /var/lib/apt/lists/*

# fonts-wqy-zenhei: a CJK font so reportlab-generated PDFs render Chinese/JP/KR

# instead of tofu boxes (the pdf SKILL.md registers it from

# /usr/share/fonts/truetype/wqy/wqy-zenhei.ttc). It MUST be a TrueType font:

# reportlab cannot embed CFF/OpenType outlines, so fonts-noto-cjk (CFF .otf)

# would fail to register. python:3.11-slim ships no CJK fonts on its own.

# build-essential ships gcc / g++ / make + libc headers so the `code_execution`

# tool can compile and run C (`cc`) and C++ (`c++ -std=c++17`) snippets, not

# just Python. Drop it if your deployment only needs Python execution.

# Python data + office-document stack. Kept separate so it is easy to drop.

# --no-cache-dir keeps the layer lean.

#

# numpy/pandas cover simple data crunching. The rest back the built-in office

# skills (deeptutor/skills/builtin/{docx,pptx,xlsx,pdf}/SKILL.md): the model

# writes short Python against these libs and runs it via the `exec` tool, so

# they MUST be present here in the sidecar — the runner image carries no

# deeptutor deps of its own. All ship as wheels (no extra apt needed). Keep this

# list in sync with the libraries those SKILL.md playbooks promise are available.

RUN pip install --no-cache-dir \

numpy \

pandas \

python-docx \

python-pptx \

openpyxl \

pypdf \

pdfplumber \

PyMuPDF \

reportlab \

lxml \

defusedxml \

Pillow

# Optional, NOT installed by default: LibreOffice (`soffice`) for format

# conversion (.doc→.docx, →PDF) and Excel formula recalculation. It is large

# (~400MB+), so the office skills gate every use on `command -v soffice` and

# degrade gracefully when absent. To enable it for your deployment, uncomment:

# RUN apt-get update && apt-get install -y --no-install-recommends \

# libreoffice-writer libreoffice-calc libreoffice-impress \

# && rm -rf /var/lib/apt/lists/*

# ----- Non-root user --------------------------------------------------------

# Commands run as an unprivileged user (uid 1000) so a sandbox escape cannot act

# as root inside the container. uid 1000 matches the host user that owns the

# shared task-workspace volume in the common single-user deployment, so files

# written into ./data/user stay readable by the main app.

RUN useradd --create-home --uid 1000 --shell /bin/bash runner

WORKDIR /app

# Ship just the server module. We copy it to a flat path and run it directly

# (python /app/server.py) rather than `python -m deeptutor...`: the runner image

# intentionally does NOT contain the deeptutor package, so the module path would

# not resolve. Direct-file execution is the simple, dependency-free choice.

COPY deeptutor/services/sandbox/runner/server.py /app/server.py

# The workspace shared with the main app is mounted here at runtime by

# docker-compose (./data/user:/app/data/user), at the *same* path in both

# containers so the mount contract (host_path == sandbox_path) holds.

ENV RUNNER_PORT=8900 \

PYTHONUNBUFFERED=1

EXPOSE 8900

USER runner

CMD ["python", "/app/server.py"]