Skip to content

SkillPilot Deployment Process

This document describes the current automated deployment workflow for Linux servers. Operators use the stable repository-root entrypoint ./deploy_skillpilot.sh; it delegates to the generic deployment engine scripts/deploy.sh.

Stable production entrypoint

The normal production command is:

./deploy_skillpilot.sh

The production host must provide Poppler (pdfinfo, pdftohtml) and Fontconfig (fc-list, fc-match) and Liberation Sans/Mono in all four regular/bold/italic/bold-italic faces for the learning-goal-book build and its fail-closed publication checks. On Rocky Linux, RHEL, or Fedora:

sudo dnf install poppler-utils fontconfig liberation-sans-fonts liberation-mono-fonts

The deployment checks these commands before updating the Git checkout or copying assets. A missing production dependency therefore cannot leave a new index.html next to stale hashed frontend assets.

Learning-goal-book PDFs and their public indexes are generated, not committed. The deployment uses npm ci, installs the lockfile-matched Playwright Chromium, checks that it can launch, and generates/verifies all four books before Vite. Browser OS libraries must already be available; there is no silent fallback to a system browser. See the book build and history-migration runbook, especially before using an old production checkout after the one-time Git history cleanup.

Generated shell files such as backend/src/main/resources/static/index.html, sw.js, and version.json are deliberately not tracked. Vite creates them and their content-hashed CSS/JavaScript files as one artifact during deployment; Git must never update only the HTML half of that bundle.

The root entrypoint pins VITE_SKILLPILOT_COACH_VARIANT=openai-mcp, which is the current production architecture for the German coach, and then executes scripts/deploy.sh. The engine continues to validate the variant before the build and again in the local and publicly served frontend artifacts. This keeps one operational command without removing the deployment guardrail.

An intentional rollback uses the same entrypoint with an explicit command-line option:

./deploy_skillpilot.sh --coach-variant visible-session

On the production server, the historical launcher in the operator home directory can remain the everyday entrypoint as a symlink to the versioned script:

ln -s /home/enpasos/skillpilot/deploy_skillpilot.sh \
  /home/enpasos/deploy_skillpilot.sh

The versioned entrypoint resolves symlinks before locating the repository, so running /home/enpasos/deploy_skillpilot.sh still executes the checked-in deployment engine from /home/enpasos/skillpilot. There is no second copy of the deployment logic to maintain.

Reviewed curriculum asset changes and transparency inventory

After approved Layer A curriculum or image changes, prepare the runtime assets and check the inventory before committing:

npm --prefix app run prepare:runtime-assets
node scripts/check_ai_transparency_inventory.mjs

If measured curriculum/image counts differ, inspect an explicit patch proposal:

node scripts/check_ai_transparency_inventory.mjs --emit-layer-a-patch

This command only prints a patch; it does not write files or approve content. Review every proposed change, apply the accepted patch and rerun the normal check. The proposal is restricted to measured Layer A inventory fields; policy and review statements remain unchanged. Missing or unequal asset copies, changed fixed media hashes, and unsupported schema versions still block the proposal. A normal CI/build/deploy never silently refreshes the inventory. C2PA container-marker detection is not verification of a credential or its authenticity.

Overview

The deployment process currently does all of the following: 1. Require an explicit, valid frontend coach variant for this artifact. 2. Check that the target systemd service is reachable, that the exact restart command has a passwordless sudo grant, and that the required PDF validation tools are available. 3. Stash local working-tree changes. 4. Pull the latest code from Git and, if HEAD changed, restart the freshly checked-out deployment engine. 5. Validate the OpenAI V1 release contract and any explicit public-URL overrides before copying assets, building, or restarting the service. 6. Deploy curriculum decks from curricula/.../json/ into both frontend and backend static data folders. 7. Deploy whitepaper assets into app/public/whitepaper and the comic folders. 8. Deploy quickstart/story assets into app/public/. 9. Install frontend dependencies and verify the committed AI-transparency inventory against the exact assets to be deployed. 10. Rebuild the React app. 11. Verify the requested coach variant and the referenced CSS/JavaScript shell assets in the generated backend static artifact. 12. Build the backend in an isolated deployment directory, separate from the default Gradle build directory used by the currently running production bootRun process. 13. Verify that the processed backend resources contain the exact deployed Git commit and a complete frontend shell before restarting the service. 14. For the openai-mcp variant, run the focused backend security and contract tests. 15. Restart the skillpilot system service. 16. Wait until the public readiness endpoint returns HTTP 200 (shared process and database readiness, not provider acceptance). 17. Verify the deployed CSS/JavaScript shell assets, coach variant, and AI-transparency copy against the public host. 18. For the openai-mcp variant, require the public path-based OpenAI V1 smoke; then run the source-rationale deployment smoke against the public host.

The Deployment Engine (scripts/deploy.sh)

This is the current automation flow:

#!/bin/bash
set -e

# Deploy from the repository that contains this script.
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
PROJECT_ROOT="$(cd "${SCRIPT_DIR}/.." && pwd)"
cd "${PROJECT_ROOT}"

SERVICE_NAME="${SKILLPILOT_SERVICE_NAME:-skillpilot}"

echo "Pruefe explizite Coach-Variante..."
# The script accepts only visible-session, openai-mcp, or legacy and aborts
# before Git/build/restart when VITE_SKILLPILOT_COACH_VARIANT is absent.

echo "Pruefe Restart-Voraussetzungen..."
# The script validates systemctl access and the command-specific NOPASSWD grant
# before doing expensive build work.

echo "Pruefe Lernzielbuch-PDF-Werkzeuge..."
# pdfinfo and pdftohtml are required before the checkout can be updated.

if [ "${SKILLPILOT_SKIP_GIT_UPDATE:-0}" = "1" ]; then
  echo "Ueberspringe Git-Update (SKILLPILOT_SKIP_GIT_UPDATE=1)."
else
  echo "Stash local changes..."
  git stash

  echo "Hole Updates..."
  git pull
fi

echo "Pruefe konsistente OpenAI-Plugin-V1-Versionierung..."
node scripts/check_openai_plugin_versioning.mjs

if [ "${VITE_SKILLPILOT_COACH_VARIANT}" = "openai-mcp" ]; then
  echo "Pruefe exakte OpenAI-Plugin-V1-Runtime-Konfiguration..."
  node scripts/validate_openai_v1_runtime_config.mjs
fi

echo "Pruefe unveraenderten OpenAI-Plugin-V1-Snapshot..."
node scripts/openai_plugin_release.mjs verify

echo "Deploying Vocabulary Decks..."
python3 scripts/deploy_decks.py

echo "Deploying Whitepaper assets..."
python3 scripts/deploy_whitepaper.py

echo "Deploying Story assets..."
python3 scripts/deploy_story.py

cd app
echo "Installiere Abhaengigkeiten..."
npm ci

echo "Installiere den zum Lockfile passenden Chromium..."
if [ -z "${GOAL_BOOK_CHROMIUM_EXECUTABLE_PATH:-}" ]; then
  ./node_modules/.bin/playwright install chromium
fi
# The executable deploy.sh additionally checks Poppler/Fontconfig and launches
# Chromium before the build; use that script as the authoritative procedure.

echo "Baue Anwendung..."
npm run build

echo "Pruefe KI-Transparenznachweis..."
npm run check:ai-transparency-inventory

echo "Pruefe Coach-Variante im Frontend-Artefakt..."
node ../scripts/verify_frontend_coach_variant.mjs \
  ../backend/src/main/resources/static \
  "${VITE_SKILLPILOT_COACH_VARIANT}"

echo "Pruefe KI-Transparenz im Frontend-Artefakt..."
node ../scripts/verify_ai_transparency_artifact.mjs \
  ../backend/src/main/resources/static

echo "Pruefe Frontend-Shell-Assets im Build-Artefakt..."
node ../scripts/verify_frontend_shell_assets.mjs \
  ../backend/src/main/resources/static

cd ../backend
chmod +x gradlew
export SKILLPILOT_BACKEND_BUILD_DIR="${PROJECT_ROOT}/tmp/deploy/backend-build"
./gradlew clean build -x test

if [ "${VITE_SKILLPILOT_COACH_VARIANT}" = "openai-mcp" ]; then
  echo "Pruefe eingebettete Backend-Buildkennung..."
  node ../scripts/validate_openai_v1_runtime_config.mjs \
    --built-application "${SKILLPILOT_BACKEND_BUILD_DIR}/resources/main/application.yml"
fi

node ../scripts/verify_frontend_shell_assets.mjs \
  "${SKILLPILOT_BACKEND_BUILD_DIR}/resources/main/static"
cd ..

echo "Starte Service neu..."
sudo -n -- "$(command -v systemctl)" restart "${SERVICE_NAME}"

SMOKE_BASE_URL="${SKILLPILOT_BASE_URL:-https://skillpilot.com}"
echo "Warte auf oeffentliche Readiness..."
# Polls /actuator/health/readiness until HTTP 200 or the configured timeout.

echo "Pruefe ausgelieferte Frontend-Shell-Assets..."
node scripts/verify_frontend_shell_assets.mjs \
  "${SMOKE_BASE_URL}"

echo "Pruefe ausgelieferte Coach-Variante..."
node scripts/verify_frontend_coach_variant.mjs \
  "${SMOKE_BASE_URL}" \
  "${VITE_SKILLPILOT_COACH_VARIANT}"

echo "Pruefe ausgelieferte KI-Transparenz..."
node scripts/verify_ai_transparency_artifact.mjs \
  "${SMOKE_BASE_URL}"

echo "Pruefe Quellenbegruendungs-Smoke-Test..."
cd app
npm run smoke:goal-source-rationales:deployment -- --base-url="${SMOKE_BASE_URL}"

Why this order?

  1. Coach-variant preflight first: every deploy must state the intended frontend contract; there is no production default that could silently choose Visible Session or MCP.
  2. Restart and build-tool preflight: the script fails before stashing, copying assets, or building if the current environment cannot reach the systemd service, lacks a command-specific passwordless restart grant, or lacks pdfinfo/pdftohtml. It never opens a general sudo password prompt.
  3. git stash + git pull: the current script assumes deployment happens from a possibly dirty working tree and protects the pull by stashing first. If the pull changes HEAD, it restarts the newly checked-out deployment engine before continuing.
  4. OpenAI V1 preflight: the checked-in release contract is validated before asset copying, build and restart. The three canonical V1 public URL variables may be absent, in which case the versioned application defaults are used. An explicitly supplied value, including an empty value, must equal the canonical V1 value exactly; a stale alias, typo, whitespace or other origin fails closed.
  5. Deck/story/whitepaper deployment must happen before the frontend build so it receives the exact public asset set.
  6. Frontend build before inventory verification synchronizes the generated runtime assets into both frontend and the non-versioned backend build tree. This prevents the inventory check from comparing a current frontend asset with a stale backend copy.
  7. AI-transparency inventory check binds current visualization providers and C2PA container markers, illustration collections, canonical goal/card counts, and podcast hashes to the reviewed inventory under docs/legal/. Asset drift therefore stops deployment before backend build or restart.
  8. Frontend artifact verification must finish before backend build or restart. The shell verifier reads index.html, rejects cross-origin stylesheet/module references, and checks that every referenced local file is present and nonempty.
  9. Isolated backend build and build-identity verification produce the updated server artifact outside the default backend/build directory. This prevents Gradle clean/processResources during a failed deployment from changing the classpath of the still-running production bootRun process. Gradle embeds the full lowercase HEAD commit into both skillpilot.openai.coach.v1.server-build and the MCP server-version; the deployment engine verifies the processed resource before restart. SKILLPILOT_SERVER_BUILD is not a runtime setting and cannot replace this artifact identity.
  10. Focused OpenAI security and contract tests run before restart for the openai-mcp artifact.
  11. systemctl restart activates the freshly built frontend/backend bundle.
  12. Public readiness wait absorbs the normal Spring Boot and reverse-proxy startup window after systemctl restart. A temporary 502 therefore does not produce a false failed deployment. This group checks only shared process and database readiness. OpenAI has a separate openaiReadiness health group; the profile-specific host/release acceptance remains required before activating a profile. A shared readiness success or the public-contract smoke below does not replace that acceptance.
  13. Public shell verification after readiness fetches index.html and the exact referenced CSS/module assets with cache bypass headers. It requires successful, nonempty same-origin responses with the expected content types, so missing hashed assets or an HTML error page served as CSS stop deployment.
  14. Mandatory OpenAI V1 public-contract smoke runs after readiness for every openai-mcp deployment. It verifies the dedicated mcp-coach-v1.skillpilot.com TLS certificate, direct responses without redirects, HTTP 200 plus the exact resource in path-specific protected-resource metadata, and HTTP 401 plus the exact WWW-Authenticate metadata reference at https://mcp-coach-v1.skillpilot.com/mcp. The discarded main-origin routes and the internal transport route must return HTTP 404; all five reserved sibling hosts must remain fail-closed with HTTP 404.
  15. Further deployment smoke tests check that the public host serves the intended coach variant in both version metadata and HTML and contains the reviewed DE/EN audio, coach, and legal transparency copy. The source-rationale smoke then detects the active curriculum mode: repository deployments must serve the two exact compatibility indexes, while package deployments must expose Catalog API 1.2 and working generation-bound source-evidence routes.

Asset deployment details

  • scripts/deploy_decks.py
  • scans curricula/**/json/ for files matching _deck*.json
  • copies them to:
    • app/public/data/
    • backend/src/main/resources/static/data/
  • scripts/deploy_whitepaper.py
  • copies docs/whitepaper/ into app/public/whitepaper/
  • copies comic assets for comic1, comic2, and comic3
  • scripts/deploy_story.py
  • copies docs/quickstart/* into app/public/

Operational notes

  • ./deploy_skillpilot.sh is the normal production entrypoint and always pins openai-mcp. Stale ambient VITE_SKILLPILOT_COACH_VARIANT values are intentionally ignored.
  • VITE_SKILLPILOT_COACH_VARIANT remains mandatory for a direct scripts/deploy.sh engine call and must be exactly visible-session, openai-mcp, or legacy.
  • Multilingual MCP deployment: ./deploy_skillpilot.sh
  • Visible Session rollback: ./deploy_skillpilot.sh --coach-variant visible-session
  • The openai-mcp build uses the same V1 App for every backend-supported communication locale.
  • The canonical OpenAI V1 public values are safe, versioned application defaults:
  • MCP endpoint: https://mcp-coach-v1.skillpilot.com/mcp
  • OAuth resource: https://mcp-coach-v1.skillpilot.com/mcp
  • protected-resource metadata: https://mcp-coach-v1.skillpilot.com/.well-known/oauth-protected-resource/mcp V1 binds exactly two active content-addressed MCP Apps UI resources on the fixed widget domain https://mcp-coach-v1.skillpilot.com: the image-only goal renderer and interactive card learning. Previously advertised image hash URIs remain passive and byte-identically readable. The card-review tool is app-only and unbound; ordinary tools remain UI-free. Permanent-ID and Level-2 setup remain exclusively in the SkillPilot WebGUI. Bare MCP ImageContent is not the visibility contract, and the runtime applies no openai/userAgent or client-surface gate. These URLs are immutable contract values rather than environment settings. Obsolete SKILLPILOT_OPENAI_DE_* URL names and newly invented SKILLPILOT_OPENAI_COACH_V1_* URL overrides fail closed. Remove stale SKILLPILOT_OPENAI_DE_UI_ORIGIN, obsolete V1-origin and old locale-bound mTLS/smoke variables before the first subdomain deployment. The current neutral edge uses only SKILLPILOT_OPENAI_COACH_V1_MTLS_EDGE_MODE=disabled|observe|enforce and a root-owned Nginx mode file; old SKILLPILOT_OPENAI_DE_* names stay forbidden.
  • Treat MCP tool descriptions, input/output schemas, annotations, server instructions, and skills as versioned model-facing metadata. A server deploy updates compatible live result behavior, but it does not rewrite an existing ChatGPT conversation or its earlier tool results. For a developer-mode connection, select Refresh on the connection after deployment, verify the discovered metadata, and test in a new conversation. For a published plugin, scan, review, and publish a new metadata snapshot. A browser-page reload is not a substitute. See OpenAI's metadata refresh procedure and published MCP metadata versioning.
  • One skillpilot-server Spring Boot artifact hosts every coach line. Values belonging to V1 use SKILLPILOT_OPENAI_COACH_V1_*; genuine shared process policies use SKILLPILOT_OPENAI_* without locale/version segments.
  • The additive Nginx templates are deploy/nginx/skillpilot-mcp-coaches.conf for inclusion inside http {} and deploy/nginx/skillpilot-main-vhost-openai-deny-locations.conf for inclusion only inside the existing skillpilot.com HTTPS server {} block before its general location /. The first file activates only neutral V1 and keeps neutral V2 through V9 at 404; the second prevents a main-origin or internal-path alias. The dedicated coaches template also applies the OpenAI client-mTLS boundary only to V1 /mcp; it does not replace existing vHosts or protect OAuth/discovery paths with a client certificate.
  • Production uses exactly one systemd EnvironmentFile, normally /etc/skillpilot/skillpilot.env. Before copying assets or building, ./deploy_skillpilot.sh verifies that this is the file configured for the service and rejects removed OpenAI names as well as attempted public-URL overrides. It inspects variable names only; OAuth, database, and other secret values are not logged or printed. The same forbidden names must not be supplied by unit-level Environment= or PassEnvironment= settings. A stale global systemd manager environment is rejected as well. A nonstandard file path must be selected explicitly with SKILLPILOT_SERVICE_ENV_FILE.
  • If systemd marks that one file as optional (ignore_errors=yes) and it is absent, the preflight accepts the canonical application defaults after the other environment channels have been checked. A missing required file (ignore_errors=no) remains a deployment error.
  • Keep an environment file containing OAuth or database secrets root-owned and mode 0600. Do not weaken it to make the deployment preflight read it. When the deploy user cannot traverse or read the root-protected file, the allowlisted content check is reported as SKIP; Spring's exact V1 startup validation remains the final fail-closed boundary. Unit-level and global systemd sources are still checked before the build.
  • Do not maintain SKILLPILOT_SERVER_BUILD in /etc/skillpilot/skillpilot.env. The backend build embeds the full Git commit into the jar and the deploy verifies it before restart. Rebuilding a commit therefore carries the correct build identity without an operator editing the service environment.
  • git stash is part of the current script behavior.
  • Operators should be aware that locally modified files will be stashed, not merged or deployed.
  • The initial backend build runs with -x test; an openai-mcp deployment then runs the focused OpenAI security, OAuth, contract and end-to-end tests before restart.
  • CI remains responsible for the complete backend regression suite.
  • The service name defaults to skillpilot.
  • Override with SKILLPILOT_SERVICE_NAME=<service-name> when deploying an environment with a different unit name.
  • The script normally stashes local changes and pulls from Git before building.
  • Set SKILLPILOT_SKIP_GIT_UPDATE=1 only when the exact desired tree is already present on the server, for example after applying a patch manually in an SSH recovery deployment.
  • The post-restart smoke test defaults to https://skillpilot.com.
  • Override with SKILLPILOT_BASE_URL=https://staging.example.org for another host.
  • The public readiness wait defaults to 180 seconds with a 5-second interval.
  • Override with SKILLPILOT_DEPLOY_READINESS_TIMEOUT_SECONDS=<seconds> and SKILLPILOT_DEPLOY_READINESS_INTERVAL_SECONDS=<seconds> when a target environment has a different startup profile.
  • Deployments as enpasos need a command-specific, passwordless sudo grant for the restart. The script deliberately uses sudo -n and never asks for a Linux password:
enpasos ALL=(root) NOPASSWD: /usr/bin/systemctl restart skillpilot

Confirm the actual binary path on the server with command -v systemctl before creating the rule. Do not grant unrestricted passwordless sudo.

Prerequisites on Server

  • Python 3: Required to run scripts/deploy_decks.py, scripts/deploy_whitepaper.py, and scripts/deploy_story.py.
  • Node.js & npm: Required for installing dependencies and building the frontend.
  • Java: Required for the backend Gradle build.
  • Git: Required for pulling updates.
  • Sudo Access: A command-specific NOPASSWD grant is required for restarting the skillpilot system service when deploying as a non-root user.