diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index e176a2e..470151b 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -71,3 +71,44 @@ jobs:
           files: coverage.xml
           fail_ci_if_error: false
           token: ${{ secrets.CODECOV_TOKEN }}
+
+  deploy-smoke:
+    # Catches breakages that the Python-only matrix can't see — e.g.
+    # PR #6 round 1 shipped a Dockerfile whose runtime stage left the
+    # venv installed in editable mode pointing at a non-existent /src,
+    # so `import pypi_winnow_downloads` failed at startup. Lint, type
+    # checks, and pytest were all green; the bug only surfaced when QA
+    # ran the container by hand. This job exercises the deploy/
+    # examples that the docs in deploy/README.md tell users to copy.
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+
+      - name: Build Docker image
+        run: docker build -f deploy/docker/Dockerfile -t pypi-winnow-downloads:smoke .
+
+      - name: Verify entrypoint resolves (winnow-collect --help)
+        # Override the Dockerfile's ENTRYPOINT so we can pass --help
+        # without also passing the hardcoded --config flag (which would
+        # require a config file the container doesn't ship). Exit 0 from
+        # --help is the smoke signal that the venv links resolve and
+        # `import pypi_winnow_downloads` works.
+        run: docker run --rm --entrypoint /opt/venv/bin/winnow-collect pypi-winnow-downloads:smoke --help
+
+      - name: Validate Compose example syntax
+        env:
+          # compose.yml.example uses ${BADGE_HOST:?…} so `compose config`
+          # errors without it. Any non-empty placeholder works.
+          BADGE_HOST: badges.example.com
+        run: docker compose -f deploy/docker/compose.yml.example config
+
+      - name: Validate Caddyfile example syntax
+        # Mount the deploy/caddy/ directory into the official caddy:2
+        # image and run `caddy validate` against the example. Catches
+        # syntax-level breakage and confirms the Caddyfile parses with
+        # the version users will most likely run.
+        run: |
+          docker run --rm \
+            -v "${GITHUB_WORKSPACE}/deploy/caddy:/etc/caddy:ro" \
+            caddy:2 \
+            caddy validate --config /etc/caddy/Caddyfile.example --adapter caddyfile
diff --git a/CHANGELOG.md b/CHANGELOG.md
index f52b1aa..498d622 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -9,6 +9,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Added
 
+- **`deploy-smoke` CI job** in `.github/workflows/ci.yml` exercises the four `deploy/` example artifacts that the Python-only matrix can't reach: builds the multi-stage Dockerfile, runs the container with overridden entrypoint to assert `winnow-collect --help` exits 0 (the bug class that took down PR #6 round 1 — venv installed in editable mode pointing at a non-existent `/src` so `import pypi_winnow_downloads` failed at startup), validates `deploy/docker/compose.yml.example` with `docker compose config` (`BADGE_HOST` substitution), and validates `deploy/caddy/Caddyfile.example` via `caddy validate` inside the official `caddy:2` image. Catches Dockerfile / compose / Caddyfile breakage before users hit it. Skips `systemd-analyze` on the timer because the referenced `.service` declares a binary at `/usr/local/bin/winnow-collect` that doesn't exist on a fresh CI runner. Closes #7.
 - **`.github/PULL_REQUEST_TEMPLATE.md`** auto-fills new human-authored PRs with Summary, Test plan, and CHANGELOG sections matching this repo's CI commands (`uv run pytest --cov`, `uv run ruff check src/ tests/`, `uv run ruff format --check src/ tests/`, `uv run mypy src/pypi_winnow_downloads/`). Dependabot bypasses PR templates by design and is handled separately by the new auto-CHANGELOG workflow.
 - **`.github/workflows/dependabot-changelog.yml`** auto-appends a `## [Unreleased]` → `### Changed` entry to Dependabot-authored PRs so they satisfy the per-PR CHANGELOG rule without manual intervention. Runs on `pull_request_target`, filters to `dependabot[bot]`, mints a GitHub App installation token via `actions/create-github-app-token`, fetches metadata via `dependabot/fetch-metadata@v3.1.0` (the v3 line fixed empty `prevVersion`/`newVersion` on grouped PRs), and pushes the CHANGELOG commit under the `cmeans-claude-dev[bot]` identity. The App-token push is the load-bearing piece: `secrets.GITHUB_TOKEN`-authored pushes do NOT re-trigger required `pull_request` checks (lint, typecheck, test) under GitHub's anti-loop policy, which would leave Dependabot PRs un-mergeable under the `main-protection` ruleset's required-status-checks rule. Loop guard skips when the last commit is already by the bot; idempotency guard skips when the PR number is already referenced in `CHANGELOG.md`. Operator must configure two repo secrets (`BOT_APP_ID`, `BOT_APP_PRIVATE_KEY`) before the workflow can run. Validated end-to-end on `cmeans/mcp-synology` (PRs #57 + #58 + #60 + #61, the latter being live verification on a real grouped Dependabot bump). Cross-repo playbook lives in awareness `dependabot-pr-hygiene-playbook`.