buzz/docs/operator-nip-fi-bindings.md

# NIP-FI Community Issuer Bindings

A deployment that provisions communities under one host suffix, for example
`<name>.communities.buzz.xyz`, stores the NIP-FI issuers allowed to vouch for
each community in `community_nip_fi_issuers`. The relay writes these rows when
it creates a community; `buzz-admin nip-fi` backfills and edits them.

A host may hold a binding only when it is exactly one DNS label under the
suffix, has no port, and is not *reserved*. A host is reserved when
`https://<host>` is a `command_audiences` entry of any issuer in
`BUZZ_NIP_FI_ISSUERS`, so a command audience can never become a community
assertion audience. The relay refuses to start while a reserved host holds a
binding, and the operator lookup route reports no issuers for a host that is not
bindable or for an issuer that is not configured.

A community with durable `requires_nip_fi=true` always retains its last issuer.
The internal `allow_last` exception for unbindable hosts cannot override this
policy. Bind a replacement first, or archive/delete the community for containment.
Bulk backfill skips required communities whose bindings are missing; its closing
report still counts them as unbound. Repair those bindings with an explicit bind.

## Configuration

`BUZZ_NIP_FI_PROVISION_BINDING` is read in every NIP-FI mode. Unset or empty
binds nothing.

```json
{
  "host_suffix": "communities.buzz.xyz",
  "authorized_issuers": ["https://app.builderlab.xyz/api/goose"]
}
```

- `host_suffix` is a lowercase domain name without a port or trailing dot.
- `authorized_issuers` is non-empty, and every entry must be an `issuer` in
  `BUZZ_NIP_FI_ISSUERS`. Each issuer is 1 to 2048 bytes.
- The suffix must not cover any `BUZZ_NIP_FI_COMMUNITIES` host, and every
  `canonical_uri` there must be readable. The relay and `buzz-admin` both
  refuse to start otherwise.

`buzz-admin` parses the same variables with the relay's parser. Run it with the
relay's environment: `BUZZ_NIP_FI_PROVISION_BINDING`, `BUZZ_NIP_FI_ISSUERS`,
`BUZZ_NIP_FI_COMMUNITIES`, and `DATABASE_URL`.

### Database community source

`BUZZ_NIP_FI_COMMUNITY_SOURCE=database` resolves communities from these
bindings instead of `BUZZ_NIP_FI_COMMUNITIES`. The relay refuses to start if
`BUZZ_NIP_FI_PROVISION_BINDING` is unset or `BUZZ_NIP_FI_COMMUNITIES` is set.
In `off` mode no snapshot is loaded.

| Variable | Default | Bound |
|----------|---------|-------|
| `BUZZ_NIP_FI_COMMUNITY_SOURCE` | `env` | `env` or `database` |
| `BUZZ_NIP_FI_SNAPSHOT_MAX_ROWS` | 1,000,000 | 1 to 10,000,000 community-issuer rows; a larger load fails |
| `BUZZ_NIP_FI_SNAPSHOT_REFRESH_TIMEOUT_MS` | 10,000 | under the 30s refresh interval |
| `BUZZ_NIP_FI_RESOLVER_NEGATIVE_CACHE_MAX` | 100,000 | 1 to 1,000,000 hosts |
| `BUZZ_NIP_FI_RESOLVER_MAX_INFLIGHT` | 8 | at least 1, clamped below the database pool size |

Each pod loads a snapshot before it binds any listener, and refreshes it
every 30s. A removed binding stops admitting new connections within 30s plus
the configured refresh timeout (40s by default) while refreshes succeed, and within two minutes of the last good load while they
fail. Open WebSocket sessions are not rechecked when a binding is removed;
they end at their own session deadline. Alert on
`time() - buzz_nip_fi_binding_snapshot_loaded_timestamp_seconds > 120` and on
`absent(buzz_nip_fi_binding_snapshot_loaded_timestamp_seconds)`.

## Initial policy on operator creation

A trusted NIP-98 operator may send `initial_requires_nip_fi: true` with
`create_only: true` to `POST /operator/communities`. The community, owner,
issuer bindings and required policy (generation 1) commit in one transaction.
Omitted, null or false creates an optional community at generation 0. A non-null
policy field without `create_only: true` returns 400
`initial_policy_requires_create_only`.

Required insertion needs an effective provision binding for the host; otherwise
400 `nip_fi_binding_required` leaves no row. Reserved hosts retain the existing
`reserved_host` refusal. The successful response includes the stored
`requires_nip_fi` and `status: "created"` for insertion or `"existed"` for a
same-owner replay. A replay changes neither policy nor bindings, including after
an operator disable or a change in issuer configuration. A different owner
retains the existing conflict. Legacy ensure and deployment bootstrap remain
optional and do not accept an explicit initial policy.

This policy is independent of `BUZZ_NIP_FI_UPGRADES_ENABLED`. Mode `off` stores
it but bypasses enforcement. The production creation paths are operator
create-only, legacy operator ensure, and deployment bootstrap; only create-only
accepts this choice.

Upgrade the **whole serving relay fleet** before callers request true. Older
writers may ignore the unknown field and insert an optional row. Checking the
created response for stored true detects that incompatible writer, but cannot
undo its insertion; a later existing-row lookup remains read-only. Turn caller
targeting off before rolling back the relay. Turning it off affects future
insertions only; it does not clear existing policy.

## Commands

```sh
buzz-admin nip-fi backfill
buzz-admin nip-fi bind --host acme.communities.buzz.xyz --issuer https://issuer.example
buzz-admin nip-fi unbind --host acme.communities.buzz.xyz --issuer https://issuer.example
```

`backfill` binds the authorized issuers to every active, unbound, bindable
community, a page of 1000 hosts at a time. Archived communities are bound on
purpose, so unarchiving leaves a community that can log in. A community whose
deletion has started is skipped. It always prints its counts and exits 1 while
any bindable community remains unbound:

```json
{"rows_bound": 12, "unbound": 0, "unbound_hosts": [], "truncated": false}
```

`unbound_hosts` lists at most 100 hosts; `truncated` is true when there are
more.

`bind` refuses a reserved host, a host that is not one label under the suffix,
and an issuer not in `BUZZ_NIP_FI_ISSUERS`.

`unbind` works without `BUZZ_NIP_FI_PROVISION_BINDING`. It refuses to remove a
community's last issuer, because an empty community looks never-provisioned and
the next `backfill` would bind it again. To swap issuers, bind the new one and
then unbind the old one. To stop a community logging in, archive or delete it.
When the provisioning config is present and shows the host is not bindable (a
reserved host or one outside the suffix), `unbind` removes the last issuer too;
this is how to clear a reserved-host binding that blocks relay startup.

`backfill` never adds the default issuers beside an operator `bind`: both lock
the community row, and the backfill only inserts while the community still has
no binding.

## Lookup route

`GET /operator/communities/nip-fi-binding?host=<host>` is operator-signed and
returns `{"host": ..., "active": ..., "issuers": [...]}`. kgoose relies on this
contract:

- `host` is the host stored on the community row.
- Every lifecycle state answers 200. `active` is `false` once the community is
  archived or its deletion has started quiescing, including a tombstoned
  community; it stays `true` while a deletion is only requested.
- `issuers` is `[]` when `BUZZ_NIP_FI_PROVISION_BINDING` is unset, the host is
  not bindable, or nothing is bound. Only issuers in `BUZZ_NIP_FI_ISSUERS` are
  reported, in byte order.
- 404 means no community has this host; 400 means the host is malformed.

## Rollout

1. Deploy the relay with `BUZZ_NIP_FI_PROVISION_BINDING` set. New communities
   are bound at creation from then on.
2. Run `buzz-admin nip-fi backfill` with the relay's environment. Repeat it until
   it exits 0.
3. Do not turn on the database binding source in enforce mode while any
   bindable community is unbound: an unbound community fails closed.

## Database source runbook

`BUZZ_NIP_FI_COMMUNITY_SOURCE=database` reads communities from these bindings
instead of `BUZZ_NIP_FI_COMMUNITIES`. Move one mode at a time, and only move
forward when the step's go signals hold.

### Off

Set `BUZZ_NIP_FI_COMMUNITY_SOURCE=database` with `BUZZ_NIP_FI_MODE=off`. No
snapshot loads in `off`, but every `BUZZ_NIP_FI_RESOLVER_*` and
`BUZZ_NIP_FI_SNAPSHOT_*` value is validated, so a bad one fails this deploy.

- Go: the rollout completes, and `buzz-admin nip-fi backfill` exits 0.
- No-go: a pod fails config validation. Fix the named variable.

### Shadow

Set `BUZZ_NIP_FI_MODE=shadow`. Each pod now loads the snapshot before any
listener binds, refreshes it every 30 s, and records what enforce would decide
without refusing anything. Shadow labels database-source records
`community="bound"` or `"unmapped"`, never a host.

- Go, held for at least a day of normal traffic:
  - `time() - buzz_nip_fi_binding_snapshot_loaded_timestamp_seconds` stays
    under 60 on every pod.
  - `buzz_nip_fi_binding_snapshot_load_failures_total` is flat.
  - `buzz_nip_fi_binding_snapshot_rows / buzz_nip_fi_binding_snapshot_max_rows`
    is under 0.5.
  - `buzz_nip_fi_resolver_miss_rejections_total` is flat.
  - The non-`admit` share of `buzz_nip_fi_shadow_total` is only what you expect
    to deny; `community="unmapped"` comes only from hosts you do not serve.
  - `stage="owner"` counts as would-deny: each is an agent enforce will refuse
    for its token owner.
- No-go: any of these fails. Return to `off` and investigate.

### Enforce

Set `BUZZ_NIP_FI_MODE=enforce`. Watch the same signals, plus the rate of 503
responses on login.

- Go: the shadow signals keep holding, and 503s stay at their shadow baseline.
- No-go: logins return 503 or are refused unexpectedly. Set `mode=off` (below).

### Alerts

- Stale snapshot: `time() - buzz_nip_fi_binding_snapshot_loaded_timestamp_seconds > 120`,
  together with `absent(buzz_nip_fi_binding_snapshot_loaded_timestamp_seconds)`
  for a database-source pod.
- Near the row cap: `buzz_nip_fi_binding_snapshot_rows / buzz_nip_fi_binding_snapshot_max_rows > 0.8`.
  `buzz_nip_fi_binding_snapshot_rows` is the size of the last good load.
- Over the cap or too slow: any increase of
  `buzz_nip_fi_binding_snapshot_load_failures_total{reason=~"over_cap|timeout"}`.
- Live lookups shed: any increase of `buzz_nip_fi_resolver_miss_rejections_total`.

### When the snapshot is too big or too slow

A load fails when the snapshot has more rows than
`BUZZ_NIP_FI_SNAPSHOT_MAX_ROWS` or takes longer than
`BUZZ_NIP_FI_SNAPSHOT_REFRESH_TIMEOUT_MS`. Both get worse as communities are
added, so they usually arrive during normal growth, not at a deploy.

- New pods: the startup load fails and the process exits, so new pods
  crash-loop and a rollout stalls. Boot already fails when the database is
  unreachable; what this source adds is that snapshot size and load time can
  also fail it.
- Running pods: their refreshes fail but the last good snapshot keeps
  answering. Once it is 120 s old it is stale, and every login's host goes to
  a live lookup. Live lookups share `BUZZ_NIP_FI_RESOLVER_MAX_INFLIGHT` slots
  (8 by default). Concurrent logins for the same host share one lookup, and a
  login waits up to 50 ms for a free slot; how many are shed depends on login
  load across distinct hosts and on lookup latency. In enforce, a shed login
  gets 503 (`authorization_unavailable`); in shadow it
  is only recorded as a would-deny.

Ways out, in order:

1. Raise `BUZZ_NIP_FI_SNAPSHOT_MAX_ROWS` (up to 10,000,000) or
   `BUZZ_NIP_FI_SNAPSHOT_REFRESH_TIMEOUT_MS` (under 30,000), whichever the
   failure `reason` names, and roll out.
2. If that cannot land quickly, set `BUZZ_NIP_FI_MODE=off`. `off` loads no
   snapshot, so pods boot and admit as they did before NIP-FI.

## Assertion policy ID change

This release changes every verified assertion's policy ID once, on purpose.
The base ID (`derive_assertion_policy_id` in `crates/buzz-auth/src/nip_fi/config.rs`)
no longer hashes the configured community audiences. The ID carried on a
verified assertion is now `IssuerPolicy::id_for_community`: a hash of that base
and the one community audience the request resolved to. The issuer is still
committed through the base.

Policy IDs from earlier releases therefore do not match. Inside the relay the ID
is only carried on the verified assertion and never compared with a stored ID, so
admission is unaffected. Anything outside the relay that recorded old policy IDs,
such as dashboards, log queries or allow-lists, should expect new values after
the upgrade.