kevinveenbirkenbachandClaude Opus 5
9aa65a7731
The harness rewrote HOST to localhost, and the provisioner turns localhost into ansible_connection local, so the replayed deploy configured the throwaway container the recipe starts rather than any server. PID 1 there is the entrypoint, not systemd, and sys-svc-container refused with "PID 1 is infinito, not systemd" in all five rows that deploy web-app-docs alongside their own role. The block now runs on the stack host and deploys over ssh into a target container that carries systemd and an ssh server and nothing else. sshd.sh installs that server and authorizes the key pair the run generates; machine.sh reads the target bridge address with container inspect, substitutes it for HOST, and only then runs the placeholder guard, because the address does not exist before that point. The target joins the default bridge so the recipe nested docker run can reach it, and that bridge has no DNS, so the value is an address. PREPARE and engine.sh both fall away. The target is bare on purpose: the replayed deploy installs docker-ce itself, which is the claim under test, so the restart collision engine.sh worked around can no longer occur. Its /var/lib/docker stays a named volume so the installed daemon writes to a real filesystem instead of stacking fuse-overlayfs on fuse-overlayfs. A shape guard refuses a block that starts no container, because this harness would otherwise configure the stack host itself; that guard was exercised in both directions. Teardown now also removes the generated key pair and the inventory the recipe writes, the latter only when the run created the directory. A schema in the test folder documents the flow. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
5.2 KiB
Instructions replay 📖
How CI proves that the ### Production block a role's documentation publishes still deploys that role.
For the workflow catalog see workflows.md.
Where it runs 🏗️
The replay rides along on a deploy row of the sweep instead of holding a job of its own. guide.py marks one row per role:
- the role's smallest variant, the one whose
meta/variants.ymloverlay enables the fewest services, because that is the shape the instructions describe; - only when the role is invokable, because that is what makes the documentation build emit a Production block for it;
- only when the guide's own mode (
composefor a role shipping a container stack,hostotherwise) is enabled inmeta/services.ymland not listed underskipinmeta/tests.yml.
The marked row carries instructions: <mode> in the matrix JSON and a 📖 in its job title, between the variant number and the ⭐ of a priority row.
What the row does 🚀
The marker adds web-app-docs to that row's apps, so the documentation is deployed beside the role under test. Its own CLI test then reads the role's page, takes the Production block out of the rendered HTML, and runs it.
One role per deploy, never a list. guide_role.py names it from APP_ID, the row's own app, which run.py forwards into the infinito container where Ansible runs. A deploy that carries no APP_ID, or one whose app publishes no Production block, draws an invokable peer of the round at random instead.
Reading the block off the page rather than out of the tree is the point: a reader copies from the site, so that is what has to work. Nothing in the repository is consulted at replay time.
The machine 🐳
Placeholders (<your-domain>, <your-clearnet-resolver>) are filled from the deployment's own values, and <your-ssh-public-key> / <your-ssh-private-key-file> from an ed25519 pair generated per run under DOCS_TEST_WORK.
The block runs on the stack host and deploys into a throwaway target container, which is what the published recipe describes: a managed server reached over SSH. machine.sh brings that target up through files/test/compose.yml under the project docs-guide-<role>, and tears it down with down -v --rmi local.
HOST is filled with the target's own bridge address, resolved once the container is up. The target joins the default bridge so the recipe's nested docker run can reach it; that bridge has no DNS, so the value must be an address rather than a name. The placeholder guard therefore runs after this fill, not before it.
The machine starts from the distribution base image, not from an image this project publishes, so the replay also exercises the installation. MACHINE_BASE_IMAGE is read as INFINITO_PARENT_IMAGE out of the .env of the mounted checkout, where the env layer has already composed it from the meta/distros.yml SPOT and honoured any custom.env override. A missing .env fails the test with a pointer to make dotenv. The one build step of files/test/Dockerfile runs init.sh, which returns immediately when the base already carries /sbin/init.
The target receives nothing but an ssh server before the replay. sshd.sh installs one, authorizes the run's public key for root, and machine.sh waits for it to answer. Everything else arrives through the deploy, which is the claim under test.
The target carries no container engine of its own. The replayed deploy installs docker-ce and restarts docker.service; a target that already ran the distribution's engine fails that restart with failed to load listeners: no sockets found via socket activation, because the running docker.socket belongs to the package being replaced. Keeping the target bare removes the collision instead of working around it. Its /var/lib/docker is a named volume so the installed daemon lands on a real filesystem rather than stacking fuse-overlayfs on fuse-overlayfs.
The deploy reaches root over SSH, so a first run needs PermitRootLogin to allow it. SYS_SVC_SSHD_PERMIT_ROOT_LOGIN defaults to no; the published recipe passes prohibit-password through the --vars JSON it already carries.
The block is replayed once systemd reports running or degraded. A block that opens with git clone is a host install: those two lines are dropped and the mounted checkout is used instead, so the replay runs against the tree under test.
Reading a failure 🔍
The CLI test prints the block it fetched and the output of the replay. The role's own deploy snapshots live in rescue-diagnostics-<artifact>.
To reproduce one locally, deploy the role together with the documentation and let the CLI test run:
APP_ID=<role> make compose-deploy mode=reinstall apps=<role>,web-app-docs full_cycle=false