cert-manager needs a Google Cloud DNS service-account key to solve
Let's Encrypt DNS-01 challenges for publicly-trusted certs. VSO syncs it
from Vault KV, so the cert-manager namespace needs its own k8s auth role
and a policy granting read on the KV path.
- Add k8s auth role cert_manager_clouddns bound to SA
cert-manager-clouddns in the cert-manager namespace.
- Add policy granting read on
kv/service/kubernetes/au/syd1/cert-manager/clouddns, bound to that role.
Claude-Session: https://claude.ai/code/session_01JUoARVdmhxKQHyyyp1pxeT
## Why
logarchiver encrypts archived logs to an OpenPGP key held in Vault's gpg engine so the private key never leaves Vault (retrieval delegates decryption to `gpg/decrypt/logarchive`, operator-only). This provisions the key and lets the service read only its public key.
## Changes
- Create gpg key `logarchive` (rsa-4096, non-exportable) in the `gpg` mount.
- Add k8s auth role `logging_logarchiver` bound to SA `logarchiver` in the `logging` namespace.
- Add policy granting `read` on `gpg/keys/logarchive` to that role (public key only; no decrypt/export).
Cross-repo: this must apply before the argocd-apps logarchiver Deployment (unkin/argocd-apps) can fetch the key.
https://claude.ai/code/session_015ur3i7D2azsMAWTSVABApv
---------
Co-authored-by: benvin <neotheo@gmail.com>
Reviewed-on: #106
Co-authored-by: Ben Vincent <ben@unkin.net>
Co-committed-by: Ben Vincent <ben@unkin.net>
End-to-end verification of the freshly-applied gitea engine (mint → API call → revoke) surfaced that tokens without read:user get 403 from GET /api/v1/user — the endpoint tea and most Gitea API clients use to validate a login. teabot's personalities would fail their auth check with the current scope sets, while in-scope calls (repo/issue) already work and lease revocation correctly kills tokens (verified 401 after revoke).
- add read:user to the teabot-implementer role scopes
- add read:user to the teabot-reviewer role scopes
Reviewed-on: #105
Co-authored-by: Ben Vincent <ben@unkin.net>
Co-committed-by: Ben Vincent <ben@unkin.net>
## Why
teabot's implementer and reviewer bot users should mint **ephemeral, scoped Gitea tokens** on demand rather than holding standing personal access tokens — Gitea tokens never expire on their own, so a leak lasts until someone notices. This registers and mounts the new `vault-plugin-secrets-gitea` engine (released v0.1.0) and declares its roles, mirroring the rancher engine wiring.
## Change
- Register the plugin in the catalog (`config/plugins/vault-plugin-secrets-gitea.yaml`), pinned to the released v0.1.0 binary `sha256 8f67fbc216effada5fd7399888a710b62fad83be0b31761a439e7dec3d56509b` (sha256 of `/opt/openbao-plugins/vault-plugin-secrets-gitea` from the released `openbao-plugin-secrets-gitea-0.1.0` RPM).
- Add `gitea_secret_backend` + `gitea_secret_backend_role` modules and wire them through `config.hcl`, `environments/au/syd1/terragrunt.hcl`, and `modules/vault_cluster` variables/main, using the `giteavaultsecret` provider from the `terraform-unkin` registry (v0.1.0).
- Mount the engine at `gitea/` against `https://git.unkin.net`; seeded site-admin credentials are read from KV (`service/vault/au/syd1/secret_backend/gitea/config`, keys `admin_username`/`admin_password`) — not stored in git.
- **The seed is consumed create-only**: `lifecycle ignore_changes` on `admin_username`/`admin_password` means the engine reads the KV seed only when first creating `gitea/config`. After `rotate-root` diverges the live password from the seed, a later apply never pushes the stale seed back.
- Add roles with conservative, minimal scopes (write: implies read:):
- `teabot-implementer` — `write:repository`, `write:issue` (clone/push, open PRs, comment).
- `teabot-reviewer` — `read:repository`, `write:issue` (read diffs, post PR reviews/comments).
- TTLs: `ttl` 1h / `max_ttl` 4h on both roles.
## The site-admin bot + KV seed are now provisioned by Terraform (no manual gap)
Per Ben's review, creating the site-admin bot and seeding its credential is no longer a manual step:
- **terraform-git #46** creates the `gitea-vault-admin` site-admin bot and writes its generated password **once** to `kv/service/vault/au/syd1/secret_backend/gitea/config` (create-only KV write; never updated).
- **terraform-vault #102** grants terraform-git write access to that KV path.
## Ordering (merge + apply)
1. **puppet-prod #498** — installs the plugin binary on the vault nodes (Puppet must run).
2. **terraform-vault #100** (`benvin/gitea-deployer-access`) — deployer access to the gitea mount.
3. **terraform-vault #102** (`benvin/gitea-kv-writer`) — terraform-git KV write grant.
4. **terraform-git #46** (`benvin/gitea-vault-admin`) — creates the bot + seeds KV.
5. **This PR** — mounts the engine (reads the seed) and declares roles.
Files here are disjoint from #100 and #102 (no conflict).
**CI note:** the plan for this PR may hard-fail in CI if the plugin isn't yet registered/installed or the KV seed isn't present in the plan's target. If CI plan fails for that ordering reason, that is expected — do not force; apply only once steps 1–4 are live.
## Remaining manual step (one, ordered)
After this PR's first apply, run `vault write -f gitea/config/rotate-root` so the standing seed password is replaced by one only Vault holds. (On future binary upgrades, bump the RPM version in puppet-prod and the catalog `sha256` here together, then `vault write sys/plugins/reload/backend plugin=vault-plugin-secrets-gitea`.)
https://claude.ai/code/session_015ur3i7D2azsMAWTSVABApv
Reviewed-on: #101
Co-authored-by: Ben Vincent <ben@unkin.net>
Co-committed-by: Ben Vincent <ben@unkin.net>
terraform-git's apply (pipeline 108) still fails: pipeline 107 actually wrote the seed but the post-create metadata read 403'd, so terraform tainted the resource — recovery is replace (delete+create), and delete was deliberately not granted. Withholding delete doesn't provide the write-once property anyway (that's lifecycle ignore_changes in terraform-git); it just breaks taint recovery and destroy.
- add delete on the kv data path for the gitea config seed
- add delete on the matching kv metadata path (full destroy support)
After merge+apply, restart the terraform-git apply once more — it will replace the tainted seed and go green, unblocking #101.
Reviewed-on: #104
Co-authored-by: Ben Vincent <ben@unkin.net>
Co-committed-by: Ben Vincent <ben@unkin.net>
terraform-git's main apply still fails after the skip_child_token fix (tfgit #48): the vault_kv_secret_v2 seed resource reads the kv-v2 metadata path during plan/apply, and the grant added in #102 covered kv/data only — Vault returns 403 on GET kv/metadata/.../secret_backend/gitea/config (terraform-git pipeline 107). This is the last blocker before the KV seed lands and terraform-vault #101 can apply.
- add read on kv/metadata/service/vault/au/syd1/secret_backend/gitea/config to the terraform-git seed policy
Reviewed-on: #103
Co-authored-by: Ben Vincent <ben@unkin.net>
Co-committed-by: Ben Vincent <ben@unkin.net>
## Why
terraform-git now provisions the `gitea-vault-admin` site-admin bot and writes its generated password to `kv/service/vault/au/syd1/secret_backend/gitea/config` (as `admin_username` + `admin_password`) so the gitea secrets engine can consume it at creation time. The `woodpecker_terraform_git` / `terraform_git` identity has no write access to that KV path, so its apply would 403 without this grant.
The deployer that *reads* the seed already has read access via `policies/kv/service/vault/secret_backends_read.yaml` (`kv/data/service/vault/+/+/secret_backend/*`), so only the write side is added here.
## Change
- Add `policies/kv/service/vault/au/syd1/secret_backend/gitea/config_write.yaml` granting `create`/`read`/`update` on the gitea config KV path to the `terraform_git` approle and `woodpecker_terraform_git` k8s role.
## Ordering
Merge + apply this before the terraform-git `benvin/gitea-vault-admin` PR applies (which performs the write). Files are disjoint from the other gitea terraform-vault PRs (#100, #101).
https://claude.ai/code/session_015ur3i7D2azsMAWTSVABApv
Reviewed-on: #102
Co-authored-by: Ben Vincent <ben@unkin.net>
Co-committed-by: Ben Vincent <ben@unkin.net>
## Why
The forthcoming `gitea_secret_backend` + role configuration (separate PR, `benvin/gitea-secret-engine`) is applied by terraform-vault under the deployment identity (`tf_vault` approle / `woodpecker_terraform_vault` k8s role). That identity has no access to the `gitea/` mount yet, so writing the engine's config and roles would 403. This mirrors `policies/rancher/admin.yaml`.
## Change
- Add `policies/gitea/admin.yaml` granting the deployer:
- create/read/update/delete on `gitea/config`
- create/update on `gitea/config/rotate-root` (write-only rotation trigger)
- full manage + list on `gitea/roles/*` (and list on `gitea/roles`)
- Deliberately excludes `gitea/creds/*` — minting tokens is for consumers, not the deployer.
- No new catalog or mount grant: plugin registration is already covered by the shared, sudo-protected wildcard in `policies/sys/plugins/catalog/admin.yaml`, and mounting uses the deployer's existing `sys/mounts/*` access — same as the rancher engine.
## Order
Merge and apply this **before** the `benvin/gitea-secret-engine` PR, so the deployer can write the engine config/roles on that apply.
https://claude.ai/code/session_015ur3i7D2azsMAWTSVABApv
Reviewed-on: #100
Co-authored-by: Ben Vincent <ben@unkin.net>
Co-committed-by: Ben Vincent <ben@unkin.net>
## Why
CI installs vault by shelling out to `dnf install vault -y`. That reads
metadata for every enabled repo (appstream/baseos/crb/epel/ha) and downloads
the 169MB vendored vault RPM from the `unkin` repo on **every** plan/apply run
(~39s per job measured in `almalinux9-opentofu:20260606`).
## Change
- Replace `dnf install vault -y` with a pinned `curl` of the upstream vault zip
from the artifactapi `hashicorp-releases` remote proxy, extracted with the
image's `python3` (`python3 -m zipfile`) to `/usr/local/bin/vault`.
- Pin the version via a new `VAULT_VERSION` env var (`1.20.0`); bump the var to
upgrade.
## Speedup
Measured in `git.unkin.net/unkin/almalinux9-opentofu:20260606`:
| approach | time |
|---|---|
| `dnf install vault -y` (current) | ~39s |
| `dnf --disablerepo='*' --enablerepo=unkin` (still pulls 169MB RPM) | ~9s |
| curl zip from artifactapi + python extract (this PR) | ~6.6s |
~32s saved per plan/apply job. The zip is cached by artifactapi after first
fetch (warm ~3s).
## Caveats
- Assumes the `almalinux9-opentofu` image ships `curl` + `python3` (both
present in `:20260606`).
- Relies on the existing artifactapi `hashicorp-releases` generic remote whose
patterns already allow `vault/.*vault_.*_linux_amd64.zip`.
---------
Co-authored-by: benvin <neotheo@gmail.com>
Reviewed-on: #99
Co-authored-by: Ben Vincent <ben@unkin.net>
Co-committed-by: Ben Vincent <ben@unkin.net>
The new **terragrunt-enc** repo manages all encapi ENC data (statuses, roles, node classifications) via Terraform/Terragrunt and needs its own Vault/Consul plumbing, mirroring terraform-git and terraform-incus. This supersedes the dual-write approach in terraform-incus PR #39; the equivalent terraform-incus grant (PR #97) is being closed, so the encapi-token grant is created fresh here for the new approle.
Changes:
- Add approle role `terraform_enc` and k8s auth role `woodpecker_terraform_enc` (bound to the `terraform-enc` ServiceAccount in the `woodpecker` namespace) for CI auth.
- Add consul secret backend role `terraform-enc` plus its ACL rules granting `write` on `infra/terraform/enc/` (its terragrunt state prefix), and a policy letting both auth roles read `consul_root/au/syd1/creds/terraform-enc`.
- Grant both auth roles read on `kv/data/kubernetes/namespace/encapi/default/environment` (the ENCAPI_WRITE_TOKEN) so `make apply` can write to encapi via the encapi provider.
Reviewed-on: #98
Co-authored-by: Ben Vincent <ben@unkin.net>
Co-committed-by: Ben Vincent <ben@unkin.net>
2026-07-24 23:18:56 +10:00
29 changed files with 464 additions and 2 deletions
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.