docs(brencher): add AGENTS.md checklist for new stacks
Python package / build (3.12.3) (pull_request) Successful in 2m1s
Python package / build (3.13) (pull_request) Successful in 2m17s
Python package / build (3.14) (pull_request) Successful in 2m13s

This commit is contained in:
2026-09-15 13:58:24 +02:00
parent 055b7e2015
commit 5eef68a1e0
+99
View File
@@ -0,0 +1,99 @@
# Adding a new stack to brencher
A "stack" is one self-contained Docker Swarm service that brencher manages
end-to-end: detect HEAD changes, build, deploy, and expose service status to
the UI. This file documents how to wire one up so it picks up commits
without an operator clicking "Update" in the UI every time.
## Repository layout
Brencher looks at exactly one commit on one ref per environment. For a new
stack `my_stack`:
1. Push branch `feat/my_stack` (or whatever the team convention is) to a
Git repo listed in `remotes` of the config file.
2. Drop the repo in `gitea` or `github` so it's reachable through the same
remotes that other stacks use.
3. Make sure `my_stack/stack-compose.yml` and a Dockerfile (typically
`my_stack/Dockerfile`) exist at the path the config will point at.
## Config file: `backend/configs/<stack>.py`
The shape is fixed by `enironment.Environment` and the steps it pulls in.
Look at `backend/configs/hermes.py` for the canonical pattern. A new config
must:
- Build a `GitClone` with at least one remote whose `url` points to the repo
above.
- Build a `DockerSwarmCheck(stack_name="<stack>")` to introspect the
deployed service.
- Build **`GitUnmerge(clone, dockerSwarmCheck)`** and wrap it into
`SharedStateHolderInMemory(unmerge=unmerge)`. This is what makes
autodeploy work — `GitUnmerge` reads the `org.brencher.version` label
from the running swarm task and converts it into
`desired_branches`, so each polling cycle asks "is HEAD ≠ what's
deployed?" without any UI input.
- Wire a `CheckoutMerged(clone, desired_branches=state, push=False, …)`.
Push stays `False` for first-party stacks; set `True` only when you want
brencher to push auto-named branches back to origin.
- Build a `DockerComposeBuild` (or `DockerSwarmBuild` if a Swarmfile is
used) pointing at the compose path in the repo.
- Build a `DockerSwarmDeploy(wd=clone, buildDocker=…, stackChecker=…, …)`
with the same `stack_name`. Image tag is `auto-<short-sha>` driven by
`checkoutMerged.progress().version`.
- Optionally append a `SimpleLog(message={"userLinks": {...}})` for the
link to surface in the UI.
- Set `__all__` and export the `Environment` named after the stack id.
- Add it to `Environment.pipeline` in the canonical order:
`clone → state → checkoutMerged → dockerSwarmCheck → build → deploy →
optionals`. Build and deploy are explicit `DockerComposeBuild` and
`DockerSwarmDeploy` siblings; do not collapse them into one step.
If you skip the `GitUnmerge` line, `state.branches` will stay empty and
the stack will only deploy when an operator pushes a `branches` update
through the UI. Don't skip it; it's the difference between a self-driving
stack and a manual one. See `connectivity_tester.py` history for the bug
this caused.
## Verification checklist
- `python3 -c "from backend.configs.my_stack import my_stack"` works in
the brencher container.
- `PROFILES="my_stack" python3 backend/app.py` boots and the UI renders a
row for `my_stack` with non-empty branches (from `GitUnmerge`,
pointing at whatever image is currently deployed).
- Push to the configured branch, wait one full polling cycle
(~`RESET_INTERVAL` = 180s, plus the next pipeline tick), then confirm
the service was redeployed with a new `org.brencher.version` label.
- No UI `environment_update` was issued in between.
## Common foot-guns
- `dockerfile:` path in `stack-compose.yml` is interpreted relative to the
compose file's directory. The compose file and the Dockerfile must
live side-by-side; use `dockerfile: Dockerfile`, not
`dockerfile: <relative-path>`.
- Removing the `name:` key from the top of `stack-compose.yml` is fine
for `docker stack deploy` — the stack name comes from the CLI option
brencher sets.
- `EXTRA_ENDPOINTS` and friends are env vars; if you reference them in
compose, define them in the `envs=lambda: {...}` of the
`DockerComposeBuild` and `DockerSwarmDeploy`. `substitute` uses
`env.get(...)` for deploys but `env[...]` (raises) for builds, so an
unlisted variable kills the build but quietly produces an empty string
in the deploy.
- Compose files must not assume Docker networks that don't exist. Use
either an `external: true` net that already exists in Swarm, drop
the `networks:` block entirely, or `driver: overlay` for a
stack-private net.
## Tracking
When you add a stack, also:
- Open a tracking issue in `admin/brencher` titled
`[ready-to-test] <stack>: …` with `state=` from your
`GitUnmerge`-derived branches and what URL/port the service exposes.
- Add the stack id to the `PROFILES` env var in the brencher stack
compose (`docker_stack` / Swarm declarative) so it gets picked up on
next rollout. The admin overlay is where this lives.