# Authority DPoP & mTLS Implementation Plan (2025-10-19)

## Purpose
- Provide the implementation blueprint for AUTH-DPOP-11-001 and AUTH-MTLS-11-002.
- Unify sender-constraint validation across Authority, downstream services, and clients.
- Capture deterministic, testable steps that unblock UI/Signer guilds depending on DPoP/mTLS hardening.

## Scope
- Token endpoint validation, issuance, and storage changes inside `StellaOps.Authority`.
- Shared security primitives consumed by Authority, Scanner, Signer, CLI, and UI.
- Operator-facing configuration, auditing, and observability.
- Out of scope: PoE enforcement (Signer) and CLI/UI client UX; those teams consume the new capabilities.

> **Status update (2025-10-19):** `ValidateDpopProofHandler`, `AuthorityClientCertificateValidator`, and the supporting storage/audit plumbing now live in `src/Authority/StellaOps.Authority`. DPoP proofs populate `cnf.jkt`, mTLS bindings enforce certificate thumbprints via `cnf.x5t#S256`, and token documents persist the sender constraint metadata. In-memory nonce issuance is wired (Redis implementation to follow). Documentation and configuration references were updated (`docs/11_AUTHORITY.md`). Targeted unit/integration tests were added; running the broader test suite is currently blocked by pre-existing `StellaOps.Concelier.Storage.Mongo` build errors.
>
> **Status update (2025-10-20):** Redis-backed nonce configuration is exposed through `security.senderConstraints.dpop.nonce` with sample YAML (`etc/authority.yaml.sample`) and architecture docs refreshed. Operator guide now includes concrete Redis/required audiences snippet; nonce challenge regression remains covered by `ValidateDpopProof_IssuesNonceChallenge_WhenNonceMissing`.
>
> **Status update (2025-10-23):** mTLS enforcement now honours `security.senderConstraints.mtls.enforceForAudiences`, automatically rejecting non-mTLS clients targeting audiences such as `signer`. Certificate bindings validate thumbprint, issuer, subject, serial number, and SAN values, producing deterministic error codes for operators. Introspection responses include `cnf.x5t#S256`, and new unit tests cover audience enforcement, binding mismatches, and bootstrap storage. Docs/sample config updated accordingly.

## Design Summary
- Extract the existing Scanner `DpopProofValidator` stack into a shared `StellaOps.Auth.Security` library used by Authority and resource servers.
- Extend Authority configuration (`authority.yaml`) with strongly-typed `senderConstraints.dpop` and `senderConstraints.mtls` sections (map to sample already shown in architecture doc).
- Require DPoP proofs on `/token` when the registered client policy is `senderConstraint=dpop`; bind issued access tokens via `cnf.jkt`.
- Introduce Authority-managed nonce issuance for “high value” audiences (default: `signer`, `attestor`) with Redis-backed persistence and deterministic auditing.
- Enable OAuth 2.0 mTLS (RFC 8705) by storing certificate bindings per client, requesting client certificates at TLS termination, and stamping `cnf.x5t#S256` into issued tokens plus introspection output.
- Surface structured logs and counters for both DPoP and mTLS flows; provide integration tests that cover success, replay, invalid proof, and certificate mismatch cases.

## AUTH-DPOP-11-001 — Proof Validation & Nonce Handling

**Shared validator**
- Move `DpopProofValidator`, option types, and replay cache interfaces from `StellaOps.Scanner.Core` into a new assembly `StellaOps.Auth.Security`.
- Provide pluggable caches: `InMemoryDpopReplayCache` (existing) and new `RedisDpopReplayCache` (leveraging the Authority Redis connection).
- Ensure the validator exposes the validated `SecurityKey`, `jti`, and `iat` so Authority can construct the `cnf` claim and compute nonce expiry.

**Configuration model**
- Extend `StellaOpsAuthorityOptions.Security` with a `SenderConstraints` property containing:
  - `Dpop` (`enabled`, `allowedAlgorithms`, `maxAgeSeconds`, `clockSkewSeconds`, `replayWindowSeconds`, `nonce` settings with `enabled`, `ttlSeconds`, `requiredAudiences`, `maxIssuancePerMinute`).
  - `Mtls` (`enabled`, `requireChainValidation`, `clientCaBundle`, `allowedSubjectPatterns`, `allowedSanTypes`).
- Bind from YAML (`authority.security.senderConstraints.*`) while preserving backwards compatibility (defaults keep both disabled).

**Token endpoint pipeline**
- Introduce a scoped OpenIddict handler `ValidateDpopProofHandler` inserted before `ValidateClientCredentialsHandler`.
- Determine the required sender constraint from client metadata:
  - Add `AuthorityClientMetadataKeys.SenderConstraint` storing `dpop` or `mtls`.
  - Optionally allow per-client overrides for nonce requirement.
- When `dpop` is required:
  - Read the `DPoP` header from the ASP.NET request, reject with `invalid_token` + `WWW-Authenticate: DPoP error="invalid_dpop_proof"` if absent.
  - Call the shared validator with method/URI. Enforce algorithm allowlist and `iat` window from options.
  - Persist the `jkt` thumbprint plus replay cache state in the OpenIddict transaction (`AuthorityOpenIddictConstants.DpopKeyThumbprintProperty`, `DpopIssuedAtProperty`).
  - When the requested audience intersects `SenderConstraints.Dpop.Nonce.RequiredAudiences`, require `nonce` in the proof; on first failure respond with HTTP 401, `error="use_dpop_nonce"`, and include `DPoP-Nonce` header (see nonce note below). Cache the rejection reason for audit logging.

**Nonce service**
- Add `IDpopNonceStore` with methods `IssueAsync(audience, clientId, jkt)` and `TryConsumeAsync(nonce, audience, clientId, jkt)`.
- Default implementation `RedisDpopNonceStore` storing SHA-256 hashes of nonces keyed by `audience:clientId:jkt`. TTL comes from `SenderConstraints.Dpop.Nonce.Ttl`.
- Create helper `DpopNonceIssuer` used by `ValidateDpopProofHandler` to issue nonces when missing/expired, enforcing issuance rate limits (per options) and tagging audit/log records.
- On successful validation (nonce supplied and consumed) stamp metadata into the transaction for auditing.
- Update `ClientCredentialsHandlers` to observe nonce enforcement: when a nonce challenge was sent, emit structured audit with `nonce_issued`, `audiences`, and `retry`.

**Token issuance**
- In `HandleClientCredentialsHandler`, if the transaction contains a validated DPoP key:
  - Build `cnf.jkt` using thumbprint from validator.
  - Include `auth_time`/`dpop_jti` as needed for diagnostics.
  - Persist the thumbprint alongside token metadata in Mongo (extend `AuthorityTokenDocument` with `SenderConstraint`, `KeyThumbprint`, `Nonce` fields).

**Auditing & observability**
- Emit new audit events:
  - `authority.dpop.proof.validated` (success/failure, clientId, audience, thumbprint, nonce status, jti).
  - `authority.dpop.nonce.issued` and `authority.dpop.nonce.consumed`.
- Metrics (Prometheus style):
  - `authority_dpop_validations_total{result,reason}`.
  - `authority_dpop_nonce_issued_total{audience}` and `authority_dpop_nonce_fails_total{reason}`.
- Structured logs include `authority.sender_constraint=dpop`, `authority.dpop_thumbprint`, `authority.dpop_nonce`.

**Testing**
- Unit tests for the handler pipeline using fake OpenIddict transactions.
- Replay/nonce tests with in-memory and Redis stores.
- Integration tests in `StellaOps.Authority.Tests` covering:
  - Valid DPoP proof issuing `cnf.jkt`.
  - Missing header → challenge with nonce.
  - Replayed `jti` rejected.
  - Invalid nonce rejected even after issuance.
- Contract tests to ensure `/.well-known/openid-configuration` advertises `dpop_signing_alg_values_supported` and `dpop_nonce_supported` when enabled.

## AUTH-MTLS-11-002 — Certificate-Bound Tokens

**Configuration model**
- Reuse `SenderConstraints.Mtls` described above; include:
  - `enforceForAudiences` list (defaults `signer`, `attestor`, `scheduler`).
  - `certificateRotationGraceSeconds` for overlap.
  - `allowedClientCertificateAuthorities` absolute paths.

**Kestrel/TLS pipeline**
- Configure Kestrel with `ClientCertificateMode.AllowCertificate` globally and implement middleware that enforces certificate presence only when the resolved client requires mTLS.
- Add `IAuthorityClientCertificateValidator` that validates presented certificate chain, SANs (`dns`, `uri`, optional SPIFFE), and thumbprint matches one of the stored bindings.
- Cache validation results per connection id to avoid rehashing on every request.

**Client registration & storage**
- Extend `AuthorityClientDocument` with `List<AuthorityClientCertificateBinding>` containing:
  - `Thumbprint`, `SerialNumber`, `Subject`, `NotBefore`, `NotAfter`, `Sans`, `CreatedAt`, `UpdatedAt`, `Label`.
- Provide admin API mutations (`/admin/clients/{id}/certificates`) for ops tooling (deferred implementation but schema ready).
- Update plugin provisioning store (`StandardClientProvisioningStore`) to map descriptors with certificate bindings and `senderConstraint`.
- Persist binding state in Mongo migrations (index on `{clientId, thumbprint}`).

**Token issuance & introspection**
- Add a transaction property capturing the validated client certificate thumbprint.
- `HandleClientCredentialsHandler`:
  - When mTLS required, ensure certificate info present; reject otherwise.
  - Stamp `cnf` claim: `principal.SetClaim("cnf", JsonSerializer.Serialize(new { x5t#S256 = thumbprint }))`.
  - Store binding metadata in issued token document for audit.
- Update `ValidateAccessTokenHandler` and introspection responses to surface `cnf.x5t#S256`.
- Ensure refresh tokens (if ever enabled) copy the binding data.

**Auditing & observability**
- Audit events:
  - `authority.mtls.handshake` (success/failure, clientId, thumbprint, issuer, subject).
  - `authority.mtls.binding.missing` when a required client posts without a cert.
- Metrics:
  - `authority_mtls_handshakes_total{result}`.
  - `authority_mtls_certificate_rotations_total`.
- Logs include `authority.sender_constraint=mtls`, `authority.mtls_thumbprint`, `authority.mtls_subject`.

**Testing**
- Unit tests for certificate validation rules (SAN mismatches, expiry, CA trust).
- Integration tests running Kestrel with test certificates:
  - Successful token issuance with bound certificate.
  - Request without certificate → `invalid_client`.
  - Token introspection reveals `cnf.x5t#S256`.
  - Rotation scenario (old + new cert allowed during grace window).

## Implementation Checklist

**DPoP work-stream**
1. Extract shared validator into `StellaOps.Auth.Security`; update Scanner references.
2. Introduce configuration classes and bind from YAML/environment.
3. Implement nonce store (Redis + in-memory), handler integration, and OpenIddict transaction plumbing.
4. Stamp `cnf.jkt`, audit events, and metrics; update Mongo documents and migrations.
5. Extend docs: `docs/modules/authority/architecture.md`, `docs/security/audit-events.md`, `docs/security/rate-limits.md`, CLI/UI references.

**mTLS work-stream**
1. Extend client document/schema and provisioning stores with certificate bindings + sender constraint flag.
2. Configure Kestrel/middleware for optional client certificates and validation service.
3. Update token issuance/introspection to honour certificate bindings and emit `cnf.x5t#S256`.
4. Add auditing/metrics and integration tests (happy path + failure).
5. Refresh operator documentation (`docs/modules/authority/operations/backup-restore.md`, `docs/modules/authority/operations/monitoring.md`, sample `authority.yaml`) to cover certificate lifecycle.

Both streams should conclude with `dotnet test src/Authority/StellaOps.Authority/StellaOps.Authority.sln` and documentation cross-links so dependent guilds can unblock UI/Signer work.