normogen/docs/adr/zero-knowledge-encryption.md
goose 38bf0ae8b4
Some checks failed
Lint and Build / format (push) Successful in 33s
Lint and Build / clippy (push) Successful in 1m36s
Lint and Build / build (push) Successful in 3m37s
Lint and Build / test (push) Failing after 2m33s
docs: update ADR for ZK Phase 2 (wrapped-DEK recovery) completion
Records the implemented wrapped-DEK recovery model, new endpoints, and the
password-change re-wrapping. Strikes the 'future' Phase 2 items as done.
2026-06-29 08:51:03 -03:00

90 lines
3.8 KiB
Markdown

# ADR: Zero-Knowledge Encryption (Phase 1)
**Status**: Implemented (Phase 1 + Phase 2)
**Date**: 2026-06-28 (Phase 1), 2026-06-29 (Phase 2)
## Context
The original design (`docs/product/encryption.md`, `docs/adr/mongodb-schema-decision.md`)
specified client-side zero-knowledge encryption: the server should never hold
plaintext user data or encryption keys. Until this change, the `EncryptedField`
type and recovery-phrase fields were aspirational scaffolding — data was always
stored and processed as plaintext JSON.
## Decision
Implement **client-side zero-knowledge encryption** using AES-256-GCM via the
browser's Web Crypto API. The server becomes a blind store: it holds opaque
ciphertext and can never decrypt user data.
### Key derivation: double PBKDF2
From the user's password, the client derives two independent values:
- **Auth secret**: `PBKDF2(password, "normogen-auth-v1", 150k iters, SHA-256)`
base64, sent to the server as the "password". The server PBKDF2-hashes it
(as before).
- **Encryption key**: `PBKDF2(password, "normogen-enc-v1", 150k iters, SHA-256)`
AES-GCM `CryptoKey`, kept in memory only, never transmitted.
The server cannot derive the encryption key because it never sees the raw
password or the enc-domain PBKDF2 output.
### What's encrypted
- Medication data blobs (name, dosage, frequency, route, etc.)
- Appointment data blobs (title, provider, date/time, etc.)
- Profile display name
### What stays plaintext (queryable)
- IDs (`medicationId`, `appointmentId`, `userId`, `profileId`)
- Medication `active` flag (top-level, filterable)
- Appointment `status` (top-level, filterable)
- Timestamps
### In-memory key lifecycle
The encryption key lives only in browser memory for the authenticated session.
It is NOT persisted. A page reload requires re-entering the password to
re-derive it. This is the core ZK trade-off.
## Consequences
### Phase 1 limitations (documented)
- **Forgotten password = data loss.** No recovery key-wrapping yet (Phase 2).
- **Page reload requires re-authentication** to re-derive the enc key.
- No data migration was needed (no real user data existed).
### Phase 2 (future)
- ~~Recovery key-wrapping (wrap the enc key under a recovery-derived key so a
forgotten password doesn't lose data).~~ **Done (Phase 2).**
- ~~Password-change re-wrapping (re-wrap the enc key under the new password).~~ **Done (Phase 2).**
- ~~Wrapped-DEK model (separate data-encryption key + key-encryption key).~~ **Done (Phase 2).**
## Phase 2: Wrapped-DEK Recovery (Implemented)
Phase 2 introduced the wrapped-DEK model so a forgotten password doesn't lose
all encrypted data:
- A **random DEK** (data encryption key) encrypts all user data (not a key
derived directly from the password).
- The DEK is **wrapped** (AES-GCM encrypted) under a password-derived KEK and
a recovery-phrase-derived KEK. Both wrapped forms are stored on the server.
- At login, the client derives the password KEK and **unwraps** the DEK.
- At recovery, the client derives the recovery KEK, unwraps the DEK from the
recovery-wrapped form, **re-wraps** it under the new password, and sends the
new wrapped form to the server.
- Password change also re-wraps the DEK under the new password.
- New endpoints: `GET /api/auth/recovery-info` (returns the recovery-wrapped
DEK), updated `POST /api/auth/recover-password` (accepts new wrapped DEK).
- New frontend: `RecoveryPage` (email → recovery phrase → new password flow),
recovery phrase field on `RegisterPage`.
The server never holds the DEK or any KEK — only the wrapped forms.
## Relationship to the original design doc
`docs/product/encryption.md` describes a fuller zero-knowledge architecture
(per-user keys, deterministic encryption for search, key rotation). This Phase 1
implements the core property (server can't read data) with the simplest viable
design. The fuller features remain future work.