@citeark/agent 0.3.20

package/runtime/bootstrap-autodl-runtime.sh

@@ -0,0 +1,314 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+CLAUDE_CODE_VERSION="${CLAUDE_CODE_VERSION:-2.1.227}"
+CODEX_VERSION="${CODEX_VERSION:-0.159.1}"
+OPENCODE_VERSION="${OPENCODE_VERSION:-1.18.14}"
+RELAY_SOURCE="${RELAY_SOURCE:-/root/codex-pro-relay.mjs}"
+BASELINE_REQUIREMENTS="${BASELINE_REQUIREMENTS:-/root/citeark-requirements-baseline.txt}"
+RUNTIME_VENV_SOURCE="${RUNTIME_VENV_SOURCE:-/root/citeark-create-runtime-venv.sh}"
+PYTHON_PACKAGE_INDEX="${PYTHON_PACKAGE_INDEX:-https://pypi.tuna.tsinghua.edu.cn/simple}"
+PYTHON_PACKAGE_FALLBACK_INDEX="${PYTHON_PACKAGE_FALLBACK_INDEX:-https://mirrors.aliyun.com/pypi/simple}"
+HUGGINGFACE_ENDPOINT="${HUGGINGFACE_ENDPOINT:-https://hf-mirror.com}"
+PYTHON_RUNTIME="${PYTHON_RUNTIME:-/opt/citeark/miniconda3/bin/python}"
+RUST_TOOLCHAIN_VERSION="${RUST_TOOLCHAIN_VERSION:-1.85.1}"
+RUSTUP_DIST_SERVER="${RUSTUP_DIST_SERVER:-https://rsproxy.cn}"
+RUSTUP_UPDATE_ROOT="${RUSTUP_UPDATE_ROOT:-https://rsproxy.cn/rustup}"
+LEGACY_TOKENIZERS_SDIST_URL="${LEGACY_TOKENIZERS_SDIST_URL:-https://files.pythonhosted.org/packages/23/6e/89019979dc7a3a86f36da5de707cf2eec72046bcb8656305a26df5c85ca7/tokenizers-0.9.4.tar.gz}"
+LEGACY_TOKENIZERS_SDIST_SHA256="${LEGACY_TOKENIZERS_SDIST_SHA256:-3ea3038008f1f74c8a1e1e2e73728690eed2d7fa4db0a51bcea391e644672426}"
+
+export DEBIAN_FRONTEND=noninteractive
+if [ -r /etc/network_turbo ]; then
+  # AutoDL 官方学术加速仅作用于本次镜像构建进程，不写入持久配置。
+  set +u
+  # shellcheck disable=SC1091
+  source /etc/network_turbo
+  set -u
+fi
+apt-get update
+apt-get install -y --no-install-recommends \
+  build-essential \
+  ca-certificates \
+  cmake \
+  curl \
+  ffmpeg \
+  git \
+  git-lfs \
+  jq \
+  libgl1 \
+  libglib2.0-0 \
+  libgomp1 \
+  libsndfile1 \
+  libssl-dev \
+  ninja-build \
+  openssh-client \
+  pkg-config \
+  poppler-utils \
+  python3 \
+  python3-pip \
+  python3-venv \
+  rsync \
+  tar \
+  unzip \
+  util-linux \
+  wget
+
+# AutoDL 学术代理适合 GitHub/HuggingFace，但会显著拖慢 npm/PyPI；
+# 后续包管理恢复直连，任务运行时再由 CiteArk 的统一路由代理接管失败回退。
+unset http_proxy https_proxy HTTP_PROXY HTTPS_PROXY all_proxy ALL_PROXY
+
+if ! command -v node >/dev/null \
+  || [ "$(node -p 'process.versions.node.split(`.`)[0]')" != "22" ] \
+  || ! node --help 2>&1 | grep -q -- '--use-env-proxy'; then
+  curl -fsSL https://deb.nodesource.com/setup_22.x | bash -
+  apt-get install -y --no-install-recommends nodejs
+fi
+if ! node --help 2>&1 | grep -q -- '--use-env-proxy'; then
+  echo "Node.js 运行时缺少 --use-env-proxy" >&2
+  exit 1
+fi
+
+# 旧 Python 包可能需要源码编译，但仓库内的 rust-toolchain 常固定到
+# 已无法解析当代依赖元数据的 Cargo。镜像构建时安装固定工具链并直接
+# 暴露实际二进制，避免运行时 rustup 按项目覆盖成旧工具链。
+RUSTUP_TEMP_HOME=/opt/citeark/rustup-build
+RUSTUP_TEMP_CARGO_HOME=/opt/citeark/rustup-build-cargo
+rm -rf /opt/citeark/rust "$RUSTUP_TEMP_HOME" "$RUSTUP_TEMP_CARGO_HOME"
+curl --fail --silent --show-error --location --retry 5 --retry-all-errors \
+  --connect-timeout 15 --max-time 120 \
+  https://rsproxy.cn/rustup-init.sh -o /tmp/citeark-rustup-init.sh
+RUSTUP_HOME="$RUSTUP_TEMP_HOME" \
+CARGO_HOME="$RUSTUP_TEMP_CARGO_HOME" \
+RUSTUP_DIST_SERVER="$RUSTUP_DIST_SERVER" \
+RUSTUP_UPDATE_ROOT="$RUSTUP_UPDATE_ROOT" \
+  sh /tmp/citeark-rustup-init.sh -y --profile minimal \
+    --default-toolchain "$RUST_TOOLCHAIN_VERSION" --no-modify-path
+RUST_TOOLCHAIN_SOURCE="$(find "$RUSTUP_TEMP_HOME/toolchains" -mindepth 1 -maxdepth 1 -type d -name "$RUST_TOOLCHAIN_VERSION-*" -print -quit)"
+if [ -z "$RUST_TOOLCHAIN_SOURCE" ]; then
+  echo "未找到已安装的 Rust 工具链：$RUST_TOOLCHAIN_VERSION" >&2
+  exit 1
+fi
+mv "$RUST_TOOLCHAIN_SOURCE" /opt/citeark/rust
+rm -rf "$RUSTUP_TEMP_HOME" "$RUSTUP_TEMP_CARGO_HOME" /tmp/citeark-rustup-init.sh
+ln -sfn /opt/citeark/rust/bin/rustc /usr/local/bin/rustc
+ln -sfn /opt/citeark/rust/bin/cargo /usr/local/bin/cargo
+ln -sfn /opt/citeark/rust/bin/rustdoc /usr/local/bin/rustdoc
+
+if ! id node >/dev/null 2>&1; then
+  useradd --create-home --uid 10001 --shell /bin/bash node
+fi
+
+npm install --global \
+  "@anthropic-ai/claude-code@${CLAUDE_CODE_VERSION}" \
+  "@openai/codex@${CODEX_VERSION}" \
+  "opencode-ai@${OPENCODE_VERSION}"
+
+install -d -m 0755 /opt/citeark
+if [ ! -e /opt/citeark/miniconda3 ] \
+  && [ -d /root/miniconda3 ] \
+  && [ ! -L /root/miniconda3 ]; then
+  mv /root/miniconda3 /opt/citeark/miniconda3
+fi
+if [ -d /opt/citeark/miniconda3 ]; then
+  ln -sfn /opt/citeark/miniconda3 /root/miniconda3
+  chmod 0755 /opt/citeark/miniconda3
+fi
+
+if [ ! -r "$BASELINE_REQUIREMENTS" ]; then
+  echo "缺少 AutoDL 基线 Python 依赖清单：$BASELINE_REQUIREMENTS" >&2
+  exit 1
+fi
+if [ ! -r "$RUNTIME_VENV_SOURCE" ]; then
+  echo "缺少 AutoDL 运行虚拟环境脚本：$RUNTIME_VENV_SOURCE" >&2
+  exit 1
+fi
+if [ ! -x "$PYTHON_RUNTIME" ]; then
+  PYTHON_RUNTIME="$(command -v python3)"
+fi
+
+# 不把前一次镜像构建期间未完成或已经失去约束的包下载固化进镜像；
+# 最终基线安装产生的 pip wheel 缓存仍会保留给运行任务使用。
+rm -rf /opt/citeark/cache/pip /opt/citeark/cache/uv
+install -d -m 1777 \
+  /opt/citeark/cache \
+  /opt/citeark/cache/cargo \
+  /opt/citeark/cache/huggingface \
+  /opt/citeark/cache/pip \
+  /opt/citeark/cache/torch \
+  /opt/citeark/cache/uv
+chown -R root:root /opt/citeark/cache
+cat > /opt/citeark/cache/cargo/config.toml <<'EOF'
+[source.crates-io]
+replace-with = "rsproxy-sparse"
+
+[source.rsproxy-sparse]
+registry = "sparse+https://rsproxy.cn/index/"
+
+[net]
+git-fetch-with-cli = true
+EOF
+rm -rf /opt/citeark/venv
+"$PYTHON_RUNTIME" -m venv --system-site-packages /opt/citeark/venv
+pip_install_with_domestic_fallback() {
+  local index
+  for index in "$PYTHON_PACKAGE_INDEX" "$PYTHON_PACKAGE_FALLBACK_INDEX"; do
+    if PIP_CACHE_DIR=/opt/citeark/cache/pip \
+      PIP_INDEX_URL="$index" \
+      PIP_DEFAULT_TIMEOUT=60 \
+      PIP_RETRIES=8 \
+      /opt/citeark/venv/bin/python -m pip install "$@"; then
+      return 0
+    fi
+    printf '%s\n' "Python 镜像源失败，切换国内备用源：$index" >&2
+  done
+  return 1
+}
+pip_install_with_domestic_fallback --upgrade pip setuptools uv wheel
+pip_install_with_domestic_fallback \
+  --upgrade-strategy only-if-needed \
+  --requirement "$BASELINE_REQUIREMENTS"
+if ! /opt/citeark/venv/bin/python -c 'import torch' >/dev/null 2>&1; then
+  pip_install_with_domestic_fallback torch torchvision
+fi
+TORCH_PUBLIC_VERSION="$(/opt/citeark/venv/bin/python -c 'import torch; print(torch.__version__.split("+")[0])')"
+if ! /opt/citeark/venv/bin/python -c 'import torchaudio' >/dev/null 2>&1; then
+  pip_install_with_domestic_fallback \
+    --upgrade-strategy only-if-needed \
+    "torchaudio==$TORCH_PUBLIC_VERSION"
+fi
+
+# tokenizers 0.9.4 没有 CPython 3.10 wheel，且其 Rust 清单没有锁定
+# serde。今天直接构建会拉到已经移除 serde::export 的版本。只固定
+# 构建期 serde 到该包发布年代的兼容版本，包版本和运行时代码不变。
+LEGACY_WHEELHOUSE=/opt/citeark/wheels
+LEGACY_BUILD_ROOT=/tmp/citeark-tokenizers-0.9.4-build
+rm -rf "$LEGACY_WHEELHOUSE" "$LEGACY_BUILD_ROOT"
+install -d -m 0755 "$LEGACY_WHEELHOUSE" "$LEGACY_BUILD_ROOT"
+curl --fail --silent --show-error --location --retry 5 --retry-all-errors \
+  --connect-timeout 15 --max-time 120 \
+  "$LEGACY_TOKENIZERS_SDIST_URL" -o "$LEGACY_BUILD_ROOT/tokenizers-0.9.4.tar.gz"
+printf '%s  %s\n' "$LEGACY_TOKENIZERS_SDIST_SHA256" "$LEGACY_BUILD_ROOT/tokenizers-0.9.4.tar.gz" \
+  | sha256sum --check --status
+tar -xzf "$LEGACY_BUILD_ROOT/tokenizers-0.9.4.tar.gz" -C "$LEGACY_BUILD_ROOT"
+for manifest in \
+  "$LEGACY_BUILD_ROOT/tokenizers-0.9.4/Cargo.toml" \
+  "$LEGACY_BUILD_ROOT/tokenizers-0.9.4/tokenizers-lib/Cargo.toml"; do
+  sed -i 's/serde = { version = "1.0",/serde = { version = "=1.0.118",/' "$manifest"
+  grep -q 'serde = { version = "=1.0.118",' "$manifest"
+done
+CARGO_HOME=/opt/citeark/cache/cargo \
+CARGO_NET_GIT_FETCH_WITH_CLI=true \
+RUSTFLAGS='-A invalid_reference_casting' \
+UV_CONFIG_FILE=/dev/null \
+UV_INDEX="$PYTHON_PACKAGE_INDEX" \
+UV_DEFAULT_INDEX="$PYTHON_PACKAGE_FALLBACK_INDEX" \
+UV_INDEX_STRATEGY=first-index \
+  /opt/citeark/venv/bin/uv build \
+    --python /usr/bin/python3.10 --wheel \
+    "$LEGACY_BUILD_ROOT/tokenizers-0.9.4" \
+    --out-dir "$LEGACY_WHEELHOUSE"
+LEGACY_TOKENIZERS_WHEEL="$(find "$LEGACY_WHEELHOUSE" -maxdepth 1 -type f -name 'tokenizers-0.9.4-cp310-*.whl' -print -quit)"
+if [ -z "$LEGACY_TOKENIZERS_WHEEL" ]; then
+  echo "未生成 tokenizers 0.9.4 的 CPython 3.10 wheel" >&2
+  exit 1
+fi
+LEGACY_TOKENIZERS_WHEEL_SHA256="$(sha256sum "$LEGACY_TOKENIZERS_WHEEL" | cut -d' ' -f1)"
+jq -n \
+  --arg package "tokenizers==0.9.4" \
+  --arg sourceSha256 "$LEGACY_TOKENIZERS_SDIST_SHA256" \
+  --argjson buildConstraints '["serde==1.0.118","RUSTFLAGS=-A invalid_reference_casting"]' \
+  --arg wheel "$(basename "$LEGACY_TOKENIZERS_WHEEL")" \
+  --arg wheelSha256 "$LEGACY_TOKENIZERS_WHEEL_SHA256" \
+  '{schemaVersion:"0.1",package:$package,sourceSha256:$sourceSha256,buildConstraints:$buildConstraints,wheel:$wheel,wheelSha256:$wheelSha256}' \
+  > "$LEGACY_WHEELHOUSE/tokenizers-0.9.4-cp310-provenance.json"
+rm -rf "$LEGACY_BUILD_ROOT"
+
+cat > /etc/pip.conf <<EOF
+[global]
+index-url = $PYTHON_PACKAGE_INDEX
+find-links = $LEGACY_WHEELHOUSE
+timeout = 60
+retries = 8
+disable-pip-version-check = true
+EOF
+cat > /etc/profile.d/citeark-python-index.sh <<EOF
+export UV_CONFIG_FILE="/dev/null"
+export PIP_INDEX_URL="$PYTHON_PACKAGE_INDEX"
+export PIP_DEFAULT_TIMEOUT="60"
+export PIP_RETRIES="8"
+export UV_INDEX="$PYTHON_PACKAGE_INDEX"
+export UV_DEFAULT_INDEX="$PYTHON_PACKAGE_FALLBACK_INDEX"
+export UV_INDEX_STRATEGY="first-index"
+export UV_HTTP_TIMEOUT="60"
+export UV_HTTP_RETRIES="8"
+export PIP_FIND_LINKS="$LEGACY_WHEELHOUSE"
+export UV_FIND_LINKS="$LEGACY_WHEELHOUSE"
+export CARGO_HOME="/opt/citeark/cache/cargo"
+export CARGO_NET_GIT_FETCH_WITH_CLI="true"
+export HF_ENDPOINT="$HUGGINGFACE_ENDPOINT"
+export HF_HUB_ETAG_TIMEOUT="30"
+export HF_HUB_DOWNLOAD_TIMEOUT="120"
+EOF
+chmod 0644 /etc/pip.conf /etc/profile.d/citeark-python-index.sh
+
+install -m 0644 "$RELAY_SOURCE" /opt/citeark/codex-pro-relay.mjs
+install -m 0644 "$BASELINE_REQUIREMENTS" /opt/citeark/requirements-baseline.txt
+install -m 0555 "$RUNTIME_VENV_SOURCE" /opt/citeark/create-runtime-venv.sh
+ln -sfn /opt/citeark/venv/bin/uv /usr/local/bin/uv
+chown -R node:node /opt/citeark/cache /opt/citeark/venv
+install -d -o node -g node -m 0700 /root/citeark
+
+NODE_VERSION="$(node --version)"
+PYTHON_VERSION="$(/opt/citeark/venv/bin/python --version 2>&1)"
+TORCH_VERSION="$(/opt/citeark/venv/bin/python -c 'import torch; print(torch.__version__)')"
+TORCH_CUDA="$(/opt/citeark/venv/bin/python -c 'import torch; print(torch.version.cuda or "none")')"
+RUST_VERSION="$(rustc --version)"
+CARGO_VERSION="$(cargo --version)"
+jq -n \
+  --arg builtAt "$(date -u +%Y-%m-%dT%H:%M:%SZ)" \
+  --arg node "$NODE_VERSION" \
+  --arg python "$PYTHON_VERSION" \
+  --arg torch "$TORCH_VERSION" \
+  --arg torchCuda "$TORCH_CUDA" \
+  --arg rust "$RUST_VERSION" \
+  --arg cargo "$CARGO_VERSION" \
+  --arg legacyTokenizersWheelSha256 "$LEGACY_TOKENIZERS_WHEEL_SHA256" \
+  --arg pythonPackageIndex "$PYTHON_PACKAGE_INDEX" \
+  --arg pythonPackageFallbackIndex "$PYTHON_PACKAGE_FALLBACK_INDEX" \
+  '{schemaVersion:"0.1",profile:"citeark-autodl-research-baseline",builtAt:$builtAt,node:$node,python:$python,torch:$torch,torchCuda:$torchCuda,rust:$rust,cargo:$cargo,legacyTokenizersWheelSha256:$legacyTokenizersWheelSha256,pythonPackageIndex:$pythonPackageIndex,pythonPackageFallbackIndex:$pythonPackageFallbackIndex}' \
+  > /opt/citeark/runtime-manifest.json
+/opt/citeark/venv/bin/python -m pip freeze \
+  > /opt/citeark/python-packages.lock
+
+claude --version
+codex --version
+opencode --version
+node --version
+python3 --version
+jq --version
+rustc --version
+cargo --version
+/opt/citeark/venv/bin/python -c 'import numpy, pandas, pydantic, scipy, sklearn, soundfile, torch, torchaudio, transformers, yaml; print("baseline scientific stack: ready"); print("torch cuda available:", torch.cuda.is_available())'
+runuser -u node -- env HOME=/home/node \
+  /opt/citeark/venv/bin/python -c 'import numpy, pandas, pydantic, scipy, sklearn, soundfile, torch, torchaudio, transformers, yaml; print("node user scientific stack: ready"); print("node user torch cuda available:", torch.cuda.is_available())'
+RUNTIME_SMOKE_ENV=/tmp/citeark-runtime-venv-smoke
+rm -rf "$RUNTIME_SMOKE_ENV"
+/opt/citeark/create-runtime-venv.sh "$RUNTIME_SMOKE_ENV"
+runuser -u node -- env \
+  HOME=/home/node \
+  "$RUNTIME_SMOKE_ENV/bin/python" -c 'import numpy, scipy, soundfile, torch, torchaudio, yaml; print("isolated runtime scientific stack: ready"); print("runtime torch cuda available:", torch.cuda.is_available())'
+rm -rf "$RUNTIME_SMOKE_ENV"
+LEGACY_SMOKE_ENV=/tmp/citeark-tokenizers-smoke
+rm -rf "$LEGACY_SMOKE_ENV"
+runuser -u node -- env HOME=/home/node UV_CONFIG_FILE=/dev/null \
+  /opt/citeark/venv/bin/uv venv --python /usr/bin/python3.10 "$LEGACY_SMOKE_ENV"
+runuser -u node -- env HOME=/home/node UV_CONFIG_FILE=/dev/null \
+  UV_FIND_LINKS="$LEGACY_WHEELHOUSE" \
+  /opt/citeark/venv/bin/uv pip install --python "$LEGACY_SMOKE_ENV/bin/python" \
+    --no-index --no-deps tokenizers==0.9.4
+runuser -u node -- env HOME=/home/node \
+  "$LEGACY_SMOKE_ENV/bin/python" -c 'import tokenizers; assert tokenizers.__version__ == "0.9.4"; print("legacy tokenizers wheel: ready")'
+rm -rf "$LEGACY_SMOKE_ENV"
+cat /opt/citeark/runtime-manifest.json
+
+echo "CiteArk AutoDL runtime bootstrap completed. Save this instance as a private image before production use."

package/runtime/create-runtime-venv.sh

@@ -0,0 +1,41 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+TARGET_VIRTUAL_ENV="${1:-/job/runtime-venv}"
+BASE_VIRTUAL_ENV="${CITEARK_BASE_VIRTUAL_ENV:-/opt/citeark/venv}"
+BASE_PYTHON="$BASE_VIRTUAL_ENV/bin/python"
+
+if [ ! -x "$BASE_PYTHON" ]; then
+  printf '%s\n' "CiteArk Python 基线不存在：$BASE_PYTHON" >&2
+  exit 70
+fi
+
+if [ ! -x "$TARGET_VIRTUAL_ENV/bin/python" ]; then
+  "$BASE_PYTHON" -m venv --system-site-packages "$TARGET_VIRTUAL_ENV"
+fi
+
+BASE_SITE_PACKAGES="$($BASE_PYTHON - <<'PY'
+import os
+import sys
+
+seen = set()
+for value in sys.path:
+    path = os.path.realpath(value or "")
+    if not path or path in seen or not os.path.isdir(path):
+        continue
+    if os.path.basename(path) not in {"site-packages", "dist-packages"}:
+        continue
+    seen.add(path)
+    print(path)
+PY
+)"
+if [ -z "$BASE_SITE_PACKAGES" ]; then
+  printf '%s\n' "CiteArk Python 基线没有可继承的 site-packages" >&2
+  exit 70
+fi
+
+TARGET_SITE_PACKAGES="$($TARGET_VIRTUAL_ENV/bin/python -c 'import site; print(site.getsitepackages()[0])')"
+PTH_PATH="$TARGET_SITE_PACKAGES/citeark-base-runtime.pth"
+printf '%s\n' "$BASE_SITE_PACKAGES" > "$PTH_PATH.tmp"
+mv "$PTH_PATH.tmp" "$PTH_PATH"
+chown -R node:node "$TARGET_VIRTUAL_ENV"

package/runtime/install-local-cpu-runtime.sh

@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+python3.12 -m venv /opt/citeark/venv
+/opt/citeark/venv/bin/python -m pip install --no-cache-dir --upgrade pip setuptools wheel
+/opt/citeark/venv/bin/python -m pip install --no-cache-dir \
+  --index-url https://download.pytorch.org/whl/cpu \
+  torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0
+/opt/citeark/venv/bin/python -m pip install --no-cache-dir \
+  numpy==1.26.4 scipy==1.17.1 pandas==2.2.3 matplotlib==3.10.8 \
+  scikit-learn==1.7.2 pyyaml==6.0.3 jsonschema==4.26.0 tqdm==4.67.3
+/opt/citeark/venv/bin/python - <<'PY'
+import json, pathlib, platform, numpy, scipy, torch, torchvision, torchaudio
+assert torch.version.cuda is None
+assert torchvision.ops.nms(torch.tensor([[0.,0.,2.,2.]]), torch.tensor([1.]), 0.5).tolist() == [0]
+pathlib.Path('/opt/citeark/runtime-manifest.json').write_text(json.dumps({
+    'schemaVersion':'0.2', 'profile':'local-cpu', 'architecture':platform.machine(),
+    'python':platform.python_version(), 'torch':torch.__version__, 'torchCuda':'none',
+    'numpy':numpy.__version__, 'scipy':scipy.__version__
+})+'\n')
+print('CPU scientific runtime imports and native operators: ready')
+PY
+/opt/citeark/venv/bin/python -m pip freeze > /opt/citeark/python-packages.lock

package/runtime/install-scientific-runtime.sh

@@ -0,0 +1,153 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+BASE_VIRTUAL_ENV="${CITEARK_BASE_VIRTUAL_ENV:-/opt/citeark/venv}"
+BASELINE_REQUIREMENTS="${CITEARK_BASELINE_REQUIREMENTS:-/opt/citeark/requirements-baseline.txt}"
+RUNTIME_VENV_SOURCE="${CITEARK_RUNTIME_VENV_SOURCE:-/opt/citeark/create-runtime-venv.sh}"
+PYTHON_RUNTIME="${CITEARK_PYTHON_RUNTIME:-python3.12}"
+PYTHON_PACKAGE_INDEX="${CITEARK_PYTHON_PACKAGE_INDEX:-https://pypi.org/simple}"
+TORCH_INDEX="${CITEARK_TORCH_INDEX:-https://download.pytorch.org/whl/cu128}"
+TORCH_VERSION="${CITEARK_TORCH_VERSION:-2.8.0}"
+TORCHVISION_VERSION="${CITEARK_TORCHVISION_VERSION:-0.23.0}"
+TORCHAUDIO_VERSION="${CITEARK_TORCHAUDIO_VERSION:-2.8.0}"
+CAUSAL_CONV1D_WHEEL_URL="${CITEARK_CAUSAL_CONV1D_WHEEL_URL:-https://github.com/Dao-AILab/causal-conv1d/releases/download/v1.7.0/causal_conv1d-1.7.0%2Bcu12torch2.8cxx11abiTRUE-cp312-cp312-linux_x86_64.whl}"
+CAUSAL_CONV1D_WHEEL_SHA256="${CITEARK_CAUSAL_CONV1D_WHEEL_SHA256:-a1fcd5d022b67ddfc9e1c291f890004c1c081a03d66b532f6eebe31219168e0e}"
+HUGGINGFACE_ENDPOINT="${CITEARK_HUGGINGFACE_ENDPOINT:-https://huggingface.co}"
+RUNTIME_PROFILE="${CITEARK_RUNTIME_PROFILE:-gcp-scientific-runtime}"
+PIP_CACHE_ROOT="${CITEARK_PIP_CACHE_ROOT:-/tmp/citeark-runtime-pip-cache}"
+
+if [ ! -r "$BASELINE_REQUIREMENTS" ]; then
+  printf '%s\n' "缺少科学运行基线依赖：$BASELINE_REQUIREMENTS" >&2
+  exit 70
+fi
+if [ ! -r "$RUNTIME_VENV_SOURCE" ]; then
+  printf '%s\n' "缺少运行虚拟环境脚本：$RUNTIME_VENV_SOURCE" >&2
+  exit 70
+fi
+if ! command -v "$PYTHON_RUNTIME" >/dev/null 2>&1; then
+  printf '%s\n' "缺少科学运行 Python：$PYTHON_RUNTIME" >&2
+  exit 70
+fi
+
+rm -rf "$BASE_VIRTUAL_ENV" "$PIP_CACHE_ROOT"
+install -d -m 0755 /opt/citeark
+install -d -m 1777 /opt/citeark/cache
+install -d -m 0755 "$PIP_CACHE_ROOT"
+
+"$PYTHON_RUNTIME" -m venv --system-site-packages "$BASE_VIRTUAL_ENV"
+
+pip_install() {
+  PIP_CACHE_DIR="$PIP_CACHE_ROOT" \
+  PIP_INDEX_URL="$PYTHON_PACKAGE_INDEX" \
+  PIP_DEFAULT_TIMEOUT=120 \
+  PIP_RETRIES=8 \
+    "$BASE_VIRTUAL_ENV/bin/python" -m pip install "$@"
+}
+
+pip_install --upgrade pip setuptools uv wheel
+
+PIP_CACHE_DIR="$PIP_CACHE_ROOT" \
+PIP_DEFAULT_TIMEOUT=180 \
+PIP_RETRIES=8 \
+  "$BASE_VIRTUAL_ENV/bin/python" -m pip install \
+    --index-url "$TORCH_INDEX" \
+    "torch==$TORCH_VERSION" \
+    "torchvision==$TORCHVISION_VERSION" \
+    "torchaudio==$TORCHAUDIO_VERSION"
+
+# Mamba-2 的官方快速路径依赖与 PyTorch/CUDA/Python ABI 精确匹配的
+# causal-conv1d 扩展。PyPI 默认只返回源码包，在运行容器没有 nvcc 时会
+# 让每个复现 Agent 重复搜索 wheel 或退回慢速实现。镜像仅缓存官方发布
+# 且固定摘要的匹配 wheel；任务虚拟环境仍可按证据覆盖它。
+pip_install --no-deps \
+  "${CAUSAL_CONV1D_WHEEL_URL}#sha256=${CAUSAL_CONV1D_WHEEL_SHA256}"
+
+# 先固定 GPU 版 PyTorch，再安装依赖它的科学包，避免 accelerate 等包
+# 从 PyPI 另行解析并下载另一套 PyTorch/CUDA 运行库。
+pip_install --upgrade-strategy only-if-needed --requirement "$BASELINE_REQUIREMENTS"
+
+install -m 0555 "$RUNTIME_VENV_SOURCE" /opt/citeark/create-runtime-venv.sh
+if [ "$(realpath "$BASELINE_REQUIREMENTS")" != "$(realpath -m /opt/citeark/requirements-baseline.txt)" ]; then
+  install -m 0644 "$BASELINE_REQUIREMENTS" /opt/citeark/requirements-baseline.txt
+else
+  chmod 0644 /opt/citeark/requirements-baseline.txt
+fi
+ln -sfn "$BASE_VIRTUAL_ENV/bin/uv" /usr/local/bin/uv
+
+cat > /etc/pip.conf <<EOF
+[global]
+index-url = $PYTHON_PACKAGE_INDEX
+timeout = 120
+retries = 8
+disable-pip-version-check = true
+EOF
+
+cat > /etc/profile.d/citeark-scientific-runtime.sh <<EOF
+export UV_CONFIG_FILE="/dev/null"
+export PIP_INDEX_URL="$PYTHON_PACKAGE_INDEX"
+export PIP_DEFAULT_TIMEOUT="120"
+export PIP_RETRIES="8"
+export UV_INDEX="$PYTHON_PACKAGE_INDEX"
+export UV_INDEX_STRATEGY="first-index"
+export UV_HTTP_TIMEOUT="120"
+export UV_HTTP_RETRIES="8"
+export HF_ENDPOINT="$HUGGINGFACE_ENDPOINT"
+export HF_HUB_ETAG_TIMEOUT="30"
+export HF_HUB_DOWNLOAD_TIMEOUT="120"
+EOF
+chmod 0644 /etc/pip.conf /etc/profile.d/citeark-scientific-runtime.sh
+
+PYTHON_VERSION="$($BASE_VIRTUAL_ENV/bin/python --version 2>&1)"
+TORCH_INSTALLED_VERSION="$($BASE_VIRTUAL_ENV/bin/python -c 'import torch; print(torch.__version__)')"
+TORCH_CUDA_VERSION="$($BASE_VIRTUAL_ENV/bin/python -c 'import torch; print(torch.version.cuda or "none")')"
+NUMPY_VERSION="$($BASE_VIRTUAL_ENV/bin/python -c 'import numpy; print(numpy.__version__)')"
+SCIPY_VERSION="$($BASE_VIRTUAL_ENV/bin/python -c 'import scipy; print(scipy.__version__)')"
+CAUSAL_CONV1D_VERSION="$($BASE_VIRTUAL_ENV/bin/python -c 'from importlib.metadata import version; print(version("causal-conv1d"))')"
+
+jq -n \
+  --arg builtAt "$(date -u +%Y-%m-%dT%H:%M:%SZ)" \
+  --arg profile "$RUNTIME_PROFILE" \
+  --arg python "$PYTHON_VERSION" \
+  --arg torch "$TORCH_INSTALLED_VERSION" \
+  --arg torchCuda "$TORCH_CUDA_VERSION" \
+  --arg numpy "$NUMPY_VERSION" \
+  --arg scipy "$SCIPY_VERSION" \
+  --arg causalConv1d "$CAUSAL_CONV1D_VERSION" \
+  --arg pythonPackageIndex "$PYTHON_PACKAGE_INDEX" \
+  --arg huggingFaceEndpoint "$HUGGINGFACE_ENDPOINT" \
+  '{schemaVersion:"0.2",$profile,$builtAt,$python,$torch,$torchCuda,$numpy,$scipy,$causalConv1d,$pythonPackageIndex,$huggingFaceEndpoint}' \
+  > /opt/citeark/runtime-manifest.json
+
+"$BASE_VIRTUAL_ENV/bin/python" -m pip freeze > /opt/citeark/python-packages.lock
+"$BASE_VIRTUAL_ENV/bin/python" - <<'PY'
+import accelerate
+import causal_conv1d
+import datasets
+import fastapi
+import fastapi_cli
+import jsonschema
+import lm_eval
+import numpy
+import pandas
+import pydantic
+import scipy
+import sklearn
+import soundfile
+import torch
+import torchaudio
+import torchvision
+import transformers
+import yaml
+
+assert torch.version.cuda == "12.8", torch.version.cuda
+boxes = torch.tensor([[0.0, 0.0, 10.0, 10.0], [1.0, 1.0, 9.0, 9.0]])
+scores = torch.tensor([0.9, 0.8])
+assert torchvision.ops.nms(boxes, scores, 0.5).tolist() == [0]
+print("scientific runtime imports: ready")
+print("torch:", torch.__version__)
+print("torch cuda runtime:", torch.version.cuda)
+print("torch cuda available during image build:", torch.cuda.is_available())
+PY
+
+rm -rf "$PIP_CACHE_ROOT"
+cat /opt/citeark/runtime-manifest.json

package/runtime/mineru/parse.py

@@ -0,0 +1,62 @@
+"""Shared, offline CPU parser. Cloud hosts only transport its inputs/outputs."""
+import argparse
+import hashlib
+import importlib.metadata
+import json
+import os
+from pathlib import Path
+import time
+
+
+def main():
+    args = argparse.ArgumentParser()
+    args.add_argument('--input', required=True)
+    args.add_argument('--output', required=True)
+    args.add_argument('--source-digest', required=True)
+    args.add_argument('--threads', type=int, default=4)
+    options = args.parse_args()
+    if not 1 <= options.threads <= 64:
+        raise ValueError('CPU threads must be between 1 and 64')
+    source = Path(options.input)
+    if hashlib.sha256(source.read_bytes()).hexdigest() != options.source_digest:
+        raise ValueError('Paper source digest mismatch')
+    for package, version in [('mineru', '4.0.10'), ('mineru-llama-cpp', '0.1.2'),
+                             ('docvortex', '0.5.8'), ('onnxruntime', '1.30.0')]:
+        if importlib.metadata.version(package) != version:
+            raise ValueError(f'Expected {package}=={version}')
+    # Ignore user-level MinerU remote/LLM configuration. The host supplies only
+    # a model directory; no external parsing or inference endpoint is accepted.
+    model_home = os.environ.get('CITEARK_MINERU_HOME', str(Path.home() / '.citeark' / 'mineru-4.0.10'))
+    for key in list(os.environ):
+        if key.startswith('MINERU_'):
+            del os.environ[key]
+    config_path = Path(options.output).resolve().with_suffix('.config.json')
+    config_path.parent.mkdir(parents=True, exist_ok=True)
+    config_path.write_text('{}')
+    os.environ.update(MINERU_HOME=model_home, MINERU_CONFIG=str(config_path),
+                      MINERU_MODEL_SOURCE='local', MINERU_MODEL_SMALL_BACKEND='onnx',
+                      MINERU_MODEL_VLM_ENGINE='llama-cpp', MINERU_TABLE_DEVICE='cpu',
+                      MINERU_INTRA_OP_NUM_THREADS=str(options.threads), MINERU_INTER_OP_NUM_THREADS='1',
+                      OMP_NUM_THREADS=str(options.threads), OPENBLAS_NUM_THREADS=str(options.threads),
+                      MKL_NUM_THREADS=str(options.threads), CUDA_VISIBLE_DEVICES='',
+                      HF_HUB_OFFLINE='1', HF_HUB_DISABLE_TELEMETRY='1', DO_NOT_TRACK='1')
+    import mineru_llama_cpp
+    engine = mineru_llama_cpp.Engine
+
+    class CpuEngine(engine):
+        def __init__(self, *args, **kwargs):
+            kwargs.update(n_gpu_layers=0, n_threads=options.threads, n_parallel=1)
+            super().__init__(*args, **kwargs)
+
+    mineru_llama_cpp.Engine = CpuEngine
+    from mineru.parser import parse
+    from mineru.kit.common import save_parse_result
+    start = time.monotonic()
+    result = parse(source, tier='standard', ocr_mode='auto', image_analysis=True)
+    save_parse_result(result, Path(options.output), 'zip')
+    print(json.dumps({'event': 'paper_parse_completed', 'sourceDigest': options.source_digest,
+                      'seconds': time.monotonic() - start, 'mineruVersion': '4.0.10', 'gpu': 0}), flush=True)
+
+
+if __name__ == '__main__':
+    main()

package/runtime/mineru/requirements.txt

@@ -0,0 +1,4 @@
+mineru==4.0.10
+mineru-llama-cpp==0.1.2
+docvortex==0.5.8
+onnxruntime==1.30.0

package/runtime/requirements-baseline.txt

@@ -0,0 +1,38 @@
+accelerate>=1,<2
+beautifulsoup4>=4.12,<5
+datasets>=2.21,<4
+einops>=0.8,<1
+evaluate>=0.4,<1
+fastapi>=0.115,<1
+fastapi-cli>=0.0.10,<1
+huggingface-hub>=0.28,<1
+jsonschema>=4,<5
+lxml>=5,<7
+lm-eval==0.4.2
+matplotlib>=3.8,<4
+numpy>=1.26,<2
+opencv-python-headless>=4.10,<4.12
+pandas>=2.2,<3
+pillow>=10,<13
+polars>=1,<2
+pydantic>=2.10,<3
+pyarrow>=18,<24
+pymupdf>=1.25,<2
+pytest>=8,<10
+pyyaml>=6,<7
+requests>=2.32,<3
+safetensors>=0.5,<1
+scikit-learn>=1.4,<2
+scipy>=1.12,<1.18
+seaborn>=0.13,<1
+sentencepiece>=0.2,<1
+soundfile>=0.12,<1
+statsmodels>=0.14,<1
+tensorboard>=2.18,<3
+tiktoken>=0.8,<1
+torch==2.8.0
+torchaudio==2.8.0
+torchvision==0.23.0
+tqdm>=4.66,<5
+transformers>=4.48,<5
+xgboost>=2,<3.1