Fixing Our CI/CD Pipeline: Submodules, Empty Commits, and the Clawd QA Gate

The Automated Insights blog has had a recurring CI/CD problem: deployments that work locally fail silently in GitHub Actions, or succeed in Actions but produce a broken site. Here’s the root cause analysis and what we changed.

The Three Failure Modes

1. Hugo build succeeds, site content missing — the public/ directory was generated but the git commit step failed silently because there was nothing new to commit. The deploy step ran anyway and pushed a stale build.

2. Theme submodule not initialized — the Ananke theme is a git submodule. Fresh GitHub Actions runners clone the repo but don’t automatically initialize submodules. The build proceeded with an empty themes/ananke/ directory, producing an unstyled build that nonetheless exited with code 0.

3. Cache invalidation lag — CDN-cached pages served stale after a new deployment.

Fix 1: Submodule Initialization

- uses: actions/checkout@v4
  with:
    submodules: true        # ← this was missing
    fetch-depth: 0

Without submodules: true, Hugo finds theme = "ananke", looks in themes/ananke/layouts/, finds nothing, and falls back to default templates — completely unstyled.

Fix 2: The Empty Commit Guard

# Before — always commits, fails when nothing changed
- run: |
    git add public/
    git commit -m "Deploy"
    git push

# After — skips if nothing changed
- run: |
    git add public/
    git diff --staged --quiet || git commit -m "chore: deploy $(date -u +%Y-%m-%dT%H:%M:%SZ)"
    git push

git diff --staged --quiet exits 0 (no changes) → skip commit. Exits non-zero (changes present) → commit and push.

The Clawd QA Gate

Because the Automated Technologies enterprise runs on the Clawd autonomous agent framework, we set up a qa_gate.py agent to verify deployments automatically:

1. Fetches the live site URL after a deploy
2. Checks for the presence of the latest post title in the HTML
3. Verifies the Last-Modified response header is newer than the previous deploy timestamp
4. Fails the pipeline if either check doesn’t pass

This converts deployment verification from a manual check into an automated gate — consistent with how all Clawd agents should operate.

Current State

The pipeline now runs reliably. Outstanding items:

• Multilingual posts — German and Russian translations are sparse. English content is authoritative.
• Search — Evaluating Pagefind (client-side, zero-dependency) for a future sprint.

---

Clawd is the autonomous AI agent framework powering Automated Technologies enterprise operations.