HHostingByteSoftware
Ctrl K
ComunidadGuías…

AGENTS.md

Rama main · réplica de lectura del repositorio original.

Archivo original ↗Preguntar a KiBot
# Agent Guide

This file defines repository-wide rules for agents working on
`self-hosted-services`. Follow it for every task in this repository.

## 1. How to Use This Guide

The requirement words in this document have consistent meanings:

- **MUST** and **MUST NOT** are mandatory.
- **SHOULD** and **SHOULD NOT** are the default. Deviate only when the task
  requires it, and explain why.
- **MAY** is optional.

When rules appear to conflict, apply them in this order:

1. Protect production systems, credentials, and user data.
2. Follow the user's explicit scope and approvals.
3. Preserve tenant isolation and repository architecture.
4. Follow the implementation and validation standards below.

If a higher-priority rule prevents a lower-priority requirement, stop at the
safe boundary and report what remains.

### 1.1 Scoped Guides

Read the nearest scoped guide before changing files in these areas:

- [`charts/services/AGENTS.md`](charts/services/AGENTS.md): Helm, Kubernetes,
  application integration, secrets, ingress, and observability.
- [`roles/AGENTS.md`](roles/AGENTS.md): Ansible role design and validation.
- [`scripts/AGENTS.md`](scripts/AGENTS.md): Python and CI-oriented scripting.

Scoped guides supplement this file. Repository-wide safety and workflow rules
continue to apply; this root guide takes precedence if instructions conflict.

## 2. Safety Boundaries

### 2.1 Production Commands

- You MUST NOT run `kubectl`, `ansible-playbook`, or any `helm` command that
  connects to or mutates a cluster without explicit user permission. These
  commands can affect live infrastructure.
- You MAY run local, non-mutating Helm validation commands without explicit
  permission, including `helm lint`, `helm template`, and `helm unittest`.
- Permission to inspect or edit the repository does not imply permission to
  deploy, reconcile, or mutate a cluster.

### 2.2 Sensitive Files and Secrets

- You MUST NOT read or modify any file whose name contains `inventory`, such
  as `inventory_home.yml`. These files contain sensitive production details
  and credentials.
- You MUST NOT commit real secrets.
- You MAY commit clearly identified dummy or template secrets used only for
  tests, such as `molecule/default/sample_secrets.yml`.

## 3. Working in the Repository

### 3.1 Scope and Git Workflow

- You MUST use a dedicated git worktree for each new task by default. If the
  user explicitly asks you to make the change directly on a branch, do not
  create or use a worktree. The user may override the worktree default at any
  time.
- You MUST keep each change focused on one feature or fix. Do not include
  opportunistic cleanup or unrelated edits.
- If a refactor is necessary, perform it as a separate step and commit before
  implementing the feature that depends on it.
- Preserve existing user changes and do not stage or commit them as part of
  your work.

### 3.2 Commit Messages

- All commits MUST follow the Conventional Commits format:
  `<type>[optional scope]: <description>`.
- The type MUST be one of `build`, `chore`, `ci`, `docs`, `feat`, `fix`,
  `perf`, `refactor`, `revert`, `style`, or `test`.
- Every commit MUST include a body that explains why the change is needed.
- Breaking changes MUST use `!` after the type or scope and include a
  `BREAKING CHANGE:` footer that describes the migration required.

### 3.3 Continuous Improvement

When a task reveals a durable repository pattern or pitfall, you SHOULD update
the applicable guide. Unless guide maintenance is part of the requested task,
leave that change unstaged and uncommitted for user review.

## 4. Project Architecture

### 4.1 Opinionated Configuration

- You SHOULD prefer a robust, opinionated setup over a broadly configurable
  framework.
- You MUST keep configuration knobs to a minimum, especially in
  `charts/services/values.yaml`.
- All applications MUST be part of the `charts/services` chart.

### 4.2 Multitenancy

- Only the primary tenant (`.Values.primaryTenant: true`) MAY deploy system
  charts and core infrastructure.
- You MUST deploy application charts and isolated service instances, including
  `PostgresCluster` and MinIO `tenant` resources, once per tenant.

### 4.3 Namespace Placement

Use the following rules when choosing a namespace for a resource or Helm
release:

- **`kube-system`:** Use only for Kubernetes components, node-level
  daemons, and cluster-wide add-ons that are part of the cluster itself. This
  includes cert-manager, trust-manager, External Secrets, device plugins, node
  feature detection, node problem detection, descheduler, and similar system
  components. Do not place shared application services here.
- **`platform`:** Use for shared services and operators owned by the primary
  tenant. This includes ingress, GitOps, observability, backup, security,
  shared storage, and operators that manage shared or tenant resources, such
  as the MinIO Operator and PGO. The `platform` namespace
  is shared by multiple tenants, but MUST be created and managed only by the
  primary tenant.
- **Application namespace:** Use `.Values.applicationsNamespace` for
  tenant-owned workloads and isolated service instances. This includes
  application charts, Redis, MinIO `tenant` resources, `PostgresCluster`
  resources, and their Secrets, ConfigMaps, Services, and monitoring objects.
  These resources MUST remain isolated per tenant.

When a HelmChart deploys a platform service, its `metadata.namespace` and
`spec.targetNamespace` MUST both be `platform`. When it deploys a system
component, both MUST normally be `kube-system`. Keep the HelmChart custom
resource in the same namespace as the release unless the controller requires
another placement and the reason is documented. A controller's namespace and
the namespace of the resources it manages are separate decisions.

Before moving a resource between namespaces, update all namespaced support
objects and references, including Secrets, ConfigMaps, ServiceAccounts,
monitoring objects, service DNS names, and Traefik middleware names.
Kubernetes objects cannot be renamed between namespaces.

### 4.4 Shared Services

You MUST use the following platform services whenever the application supports
them, except where a scoped guide gives more specific implementation details:

- **Database:** PostgreSQL, not an embedded database such as SQLite.
- **Cache and sessions:** Redis.
- **Object storage:** MinIO through its S3-compatible API.
- **Multimedia PVCs:** MUST NEVER be backed up by Velero.
- **SSO:** Authentik, preferably through OAuth/OIDC.
- **Outbound email:** the shared `smtp` secret, backed by AWS SES by default.
- **Metrics:** you SHOULD enable metrics when supported.
- **VPN routing:** route all pod egress through the central `gluetun` service.

## 5. Validation

Use `.woodpecker/lint.yaml` as the source of truth for linting and validation.
Do not duplicate its complete tool list in agent guides. Follow any additional
validation requirements in the applicable scoped guide.

## 6. Definition of Done

A task is complete only when:

- The requested implementation is complete and follows every applicable agent
  guide.
- Targeted validation for the changed behavior passes.
- All additional validation required by the applicable scoped guide passes, or
  each skipped check is reported with its reason.
- The final diff has been reviewed for unrelated changes.
- The diff contains no real secrets, sensitive inventory data, or unexpected
  generated files.
- No production-affecting command was run without explicit user approval.