# gstack CI eval runner — pre-baked toolchain + deps # Built and pushed by the eval workflows' build-image job whenever its content-hash # tag (Dockerfile.ci, bun.lock, patches/**) is missing from the registry FROM ubuntu:24.04 ENV DEBIAN_FRONTEND=noninteractive # Switch apt sources to Hetzner's public mirror. # Ubicloud runners (Hetzner FSN1-DC21) hit reliable connection timeouts to # archive.ubuntu.com:80 — observed 90+ second outages on multiple builds. # Hetzner's mirror is publicly accessible from any cloud and route-local for # Ubicloud, so this fixes both reliability and latency. Ubuntu 24.04 uses # the deb822 sources format at /etc/apt/sources.list.d/ubuntu.sources. # # Using HTTP (not HTTPS) intentionally: the base ubuntu:24.04 image ships # without ca-certificates, so HTTPS apt fails with "No system certificates # available." Apt's security model verifies via GPG-signed Release files, # not TLS, so HTTP here is no weaker than the upstream defaults. RUN sed -i \ -e 's|http://archive.ubuntu.com/ubuntu|http://mirror.hetzner.com/ubuntu/packages|g' \ -e 's|http://security.ubuntu.com/ubuntu|http://mirror.hetzner.com/ubuntu/packages|g' \ /etc/apt/sources.list.d/ubuntu.sources # Also make apt itself resilient — per-package retries + generous timeouts. # Hetzner's mirror is reliable but individual packages can still blip; the # retry config means a single failed fetch doesn't nuke the whole build. RUN printf 'Acquire::Retries "5";\nAcquire::http::Timeout "30";\nAcquire::https::Timeout "30";\n' \ > /etc/apt/apt.conf.d/80-retries # System deps (retry apt-get update + install as a unit — even Hetzner can blip). # Includes xz-utils so the Node.js .tar.xz download below can decompress. # python3: bin/gstack-jsonl-merge, gstack-brain-sync, gstack-detach, and other # bash bins shell out to it (macOS ships python3; the base image doesn't). # file: skill-validation's no-compiled-binaries-in-git check runs `file --mime-type`. # poppler-utils: make-pdf's e2e gates hard-require pdftotext/pdffonts/pdfinfo in CI. # zsh: the sourced-helper guard sources every skill helper under bash and zsh # (macOS's default shell) and fails under CI=true when zsh is missing. RUN for i in 1 2 3; do \ apt-get update && apt-get install -y --no-install-recommends \ git curl unzip xz-utils ca-certificates jq bc gpg python3 python3-venv file poppler-utils gcc libc6-dev zsh && break || \ (echo "apt retry $i/3 after failure"; sleep 10); \ done \ && rm -rf /var/lib/apt/lists/* # Git >= 2.44 for --no-lazy-fetch. bin/gstack-safe-git puts it on every # /deslop-shared-libs Git read, and noble's git 2.43 exits 129 on it, so every # paid shared-libs case ran the unsupported-Git fallback (local history # unavailable) instead of the path it was written for. The Ubuntu Git # Maintainers PPA ships current Git for noble; the build fails if the flag is # missing so the image can never silently fall back again. RUN curl --retry 5 --retry-delay 5 --retry-connrefused -fsSL \ "https://keyserver.ubuntu.com/pks/lookup?op=get&search=0xE1DD270288B4E6030699E45FA1715D88E1DF1F24" \ | gpg --dearmor -o /usr/share/keyrings/git-core-ppa.gpg \ && echo "deb [signed-by=/usr/share/keyrings/git-core-ppa.gpg] https://ppa.launchpadcontent.net/git-core/ppa/ubuntu noble main" \ > /etc/apt/sources.list.d/git-core-ppa.list \ && for i in 1 2 3; do \ apt-get update && apt-get install -y --no-install-recommends git && break || \ (echo "apt retry $i/3 after failure"; sleep 10); \ done \ && rm -rf /var/lib/apt/lists/* \ && git --no-lazy-fetch --version RUN python3 -m venv /tmp/gstack-ci-venv \ && /tmp/gstack-ci-venv/bin/python -m pip --version \ && rm -rf /tmp/gstack-ci-venv # Direct builds produce the trusted CSO launcher and watchdog. Check the exact # static-C capability here so the cached eval image cannot reach a slice without it. RUN printf 'int main(void) { return 0; }\n' > /tmp/gstack-cso-cc-probe.c \ && cc -std=c11 -static /tmp/gstack-cso-cc-probe.c -o /tmp/gstack-cso-cc-probe \ && /tmp/gstack-cso-cc-probe \ && rm -f /tmp/gstack-cso-cc-probe.c /tmp/gstack-cso-cc-probe # GitHub CLI RUN curl --retry 5 --retry-delay 5 --retry-connrefused -fsSL https://cli.github.com/packages/githubcli-archive-keyring.gpg \ | gpg --dearmor -o /usr/share/keyrings/githubcli-archive-keyring.gpg \ && echo "deb [arch=$(dpkg --print-architecture) signed-by=/usr/share/keyrings/githubcli-archive-keyring.gpg] https://cli.github.com/packages stable main" \ | tee /etc/apt/sources.list.d/github-cli.list > /dev/null \ && for i in 1 2 3; do \ apt-get update && apt-get install -y --no-install-recommends gh && break || \ (echo "gh install retry $i/3"; sleep 10); \ done \ && rm -rf /var/lib/apt/lists/* # Node.js 22 LTS (needed for claude CLI). # Install from the official nodejs.org tarball instead of NodeSource's apt setup. # NodeSource's setup_22.x script runs its own `apt-get update` + `apt-get install gnupg`, # both of which depend on archive.ubuntu.com / security.ubuntu.com being reachable. # Ubicloud CI runners frequently can't reach those mirrors (connection timeouts), # and "gnupg" was renamed to "gpg" on Ubuntu 24.04 anyway, so NodeSource's script # fails before it can add its own repo. Direct tarball download is network-simpler # (one host: nodejs.org) and doesn't touch apt at all. ENV NODE_VERSION=22.20.0 RUN curl --retry 5 --retry-delay 5 --retry-connrefused -fsSL "https://nodejs.org/dist/v${NODE_VERSION}/node-v${NODE_VERSION}-linux-x64.tar.xz" -o /tmp/node.tar.xz \ && tar -xJ -C /usr/local --strip-components=1 --no-same-owner -f /tmp/node.tar.xz \ && rm -f /tmp/node.tar.xz \ && node --version \ && npm --version # Bun (install to /usr/local so non-root users can access it). # The exact release archive, verified by SHA-256 before it is unpacked # (#1706). The old `curl https://bun.sh/install | bash -s bun-vX.Y.Z` ran # whatever script the network served at build time, and before that the # `| BUN_VERSION=x.y.z bash` form silently installed latest on every rebuild. # Hashes come from the release's SHASUMS256.txt. The image builds linux/amd64 # only (the Node.js tarball above is linux-x64); an architecture without a # verified archive fails here. test/ci-image-cli-pin.test.ts pins this shape. ARG BUN_VERSION=1.4.2 ARG TARGETARCH ENV BUN_INSTALL="/usr/local" RUN case "${TARGETARCH:-amd64}" in \ amd64) bun_target=linux-x64; bun_sha256=36368faef7527875d5ffa52e53cd48021741f2a83eb6208a8dd64068d422a913 ;; \ *) echo "Dockerfile.ci: no verified Bun ${BUN_VERSION} archive for ${TARGETARCH}" >&2; exit 1 ;; \ esac \ && curl --retry 5 --retry-delay 5 --retry-connrefused -fsSL \ "https://github.com/oven-sh/bun/releases/download/bun-v${BUN_VERSION}/bun-${bun_target}.zip" -o /tmp/bun.zip \ && echo "${bun_sha256} /tmp/bun.zip" | sha256sum -c - \ && unzip -q /tmp/bun.zip -d /tmp/bun \ && install -D -m 0755 "/tmp/bun/bun-${bun_target}/bun" "${BUN_INSTALL}/bin/bun" \ && rm -rf /tmp/bun /tmp/bun.zip \ && test "$(bun --version)" = "${BUN_VERSION}" # Claude CLI — pinned to an EXACT version, same discipline as the bun pin # above. The PTY harness (test/helpers/claude-pty-runner.ts) screen-scrapes # this CLI's TUI (trust dialog, input prompt, spinner glyphs); an unpinned # install rebuilt weekly rode the TUI wherever it drifted, and that drift # broke the harness three separate times (welcome-screen wedge on 2.1.233, # skillify HOME discovery on 2.1.237, guard/freeze hooks on 2.1.162). # Bump deliberately, via a PR that runs the PTY gate against the new TUI. # test/ci-image-cli-pin.test.ts fails the free suite if this pin is removed. # 2.1.292 is the npm latest users run (October 2026). On it the fork-subagent # gate is on by default in interactive sessions: the Agent schema drops # run_in_background and reviewers run in the background, which broke the # /autoplan publication guard while CI still ran 2.1.284 and never saw it. # Pinning the version users run is the point; the daily # autoplan-schema-canary case watches the latest release ahead of the pin. # (2.1.284 was the first pin to recognize the eval model claude-fable-5-1; # 2.1.292 keeps that profile.) RUN npm i -g @anthropic-ai/claude-code@2.1.292 # Codex CLI — exact pin, same discipline as the claude pin above, so the # Codex eval files execute in CI instead of self-skipping. Installed OFF PATH: # a skill's outside-voice probe (`command -v codex`) must keep reporting # not_installed in every other paid case. The paid shard runner puts this # directory on PATH only for the Codex files (scripts/lib/paid-cases.ts scopeCodexAccess). # The image carries no credential: the eval workflows run `codex login # --with-api-key` from the OPENAI_API_KEY secret in same-repo runs only # (codex 0.160.0 does not read OPENAI_API_KEY from the environment). ENV GSTACK_CI_CODEX_BIN_DIR=/opt/codex/bin RUN npm i -g --prefix=/opt/codex @openai/codex@0.160.0 \ && test -x /opt/codex/bin/codex \ && ! command -v codex # Playwright system deps (Chromium) — needed for browse E2E tests RUN npx playwright install-deps chromium # Linux has neither Helvetica nor Arial. make-pdf's print CSS stacks fall back # to Liberation Sans (metric-compatible Arial clone, SIL OFL 1.1) so PDFs don't # render in DejaVu Sans. playwright install-deps happens to pull this in today, # but the dep is implicit and could change — install explicitly so upgrades # can't silently regress rendering. # # Xvfb is also installed here so the browse --headed integration tests # (headed-xvfb, headed-orphan-cleanup) can exercise the Linux container # auto-spawn path on every CI run. Without Xvfb in the image, the most # common production --headed path goes untested. # fonts-noto-color-emoji: the make-pdf emoji render gate needs a color-emoji # fallback font (mirrors make-pdf-gate.yml's Ubuntu setup step). RUN for i in 1 2 3; do \ apt-get update && apt-get install -y --no-install-recommends fonts-liberation fonts-noto-color-emoji fontconfig xvfb x11-utils && break || \ (echo "fonts-liberation install retry $i/3"; sleep 10); \ done \ && fc-cache -f \ && rm -rf /var/lib/apt/lists/* # Pre-install dependencies (cached layer — only rebuilds when package.json or # bun.lock changes). Copy BOTH so install is deterministic and matches local # resolution. Without bun.lock here, bun install resolved transitive deps # differently in CI vs local (observed on v1.28.0.0: socks landed but # smart-buffer + ip-address didn't make it into the cached node_modules). # patches/ rides along: bun.lock's patchedDependencies (playwright-core # windowsHide, v1.67) makes install fail without the patch files present — # and the workflows' image-tag hash includes patches/** so editing a patch # rebuilds this layer. COPY package.json bun.lock /workspace/ COPY patches /workspace/patches WORKDIR /workspace RUN bun install --frozen-lockfile && rm -rf /tmp/* # Install Playwright Chromium to a shared location accessible by all users ENV PLAYWRIGHT_BROWSERS_PATH=/opt/playwright-browsers RUN npx playwright install chromium \ && chmod -R a+rX /opt/playwright-browsers # Verify everything works RUN bun --version && node --version && claude --version && /opt/codex/bin/codex --version && jq --version && gh --version \ && python3 --version && command -v file && command -v pdftotext && command -v pdffonts && command -v pdfinfo \ && npx playwright --version \ && fc-match "Liberation Sans" | grep -qi "Liberation" \ || (echo "ERROR: fonts-liberation not installed — make-pdf PDFs will render in DejaVu Sans" && exit 1) # At runtime: checkout overwrites /workspace, but node_modules persists # if we move it out of the way and symlink back # Save node_modules + package.json snapshot for cache validation at runtime RUN mv /workspace/node_modules /opt/node_modules_cache \ && cp /workspace/package.json /opt/node_modules_cache/.package.json \ && cp /workspace/bun.lock /opt/node_modules_cache/.bun.lock # Claude CLI refuses --dangerously-skip-permissions as root. # Create a non-root user for eval runs (GH Actions overrides USER, so # the workflow must set options.user or use gosu/su-exec at runtime). RUN useradd -m -s /bin/bash runner \ && chmod -R a+rX /opt/node_modules_cache \ && mkdir -p /home/runner/.gstack && chown -R runner:runner /home/runner/.gstack \ && chmod 1777 /tmp \ && mkdir -p /home/runner/.bun && chown -R runner:runner /home/runner/.bun