# ADR 0009 — Team access: membership authorises, a grant unlocks

- Status: accepted
- Date: 2026-07-31
- Builds on: [ADR 0001](0001-e2ee-trust-model.md)

## Context

M3 makes vaults shareable. The obvious way to model that is one concept — "access" — with a role
attached, and to let the server hand it out. Every hosted competitor works that way, and it is what
the imported design drew: a members table with a role column, and a share button beside each item.

This architecture cannot implement that concept, and the interesting part of M3 was working out
what it can implement instead.

The server holds ciphertext and no keys. A vault key is 32 random bytes sealed to each member's
X25519 public key (`docs/crypto.md` §3), and only a client holding the plaintext key can produce a
seal for somebody else. So "give Bob access" decomposes into two operations that live on different
machines and cannot be performed by the same actor:

- deciding that the server will **serve** Bob this vault's rows, which is a database write; and
- **wrapping** the vault key to Bob's public key, which needs a client that already holds it.

The schema anticipated this — `team`, `team_membership`, `vault.team_id` and `vault_key_grant` have
existed since the first migration — but nothing had had to name the split.

## Decision

**Membership is authorisation. A grant is access. The product says so out loud.**

`VaultAccessService` resolves a team vault through `team_membership`, mapping the role to
`PermissionFlags` by a union with no Deny rules. That decides what the server serves and nothing
else. Whether the caller can read what it serves is decided by whether they hold a grant, which the
server records, cannot produce and cannot verify.

Five consequences, each of which is a place where a more reassuring design was rejected:

- **A member with no grant is a normal state, not an error.** `VaultSummary.WrappedVaultKey` is null
  and the vault appears in their list saying it is waiting for a key. Hiding it until a grant existed
  would have been tidier and would have implied the server was the thing granting access.
- **The roles are only the ones that are enforceable.** There is no `ConnectOnly`, despite the design
  asking for one and `TeamRole` having room. SSH terminates on the client, so opening a session needs
  the credential's plaintext on that machine; "may connect but may not read the key" cannot be
  enforced here, and shipping it as a role would have been a lie in a dropdown. `Connect` rides along
  with `Read` and is documented as an interface hint.
- **Sharing verifies the recipient's key against the append-only key log, or refuses.** A directory
  lookup is a claim by the server about a third party's public key; wrapping to an unverified claim
  hands the vault to whoever made it. `KeyLogAudit` reads the whole log, checks its hash chain from
  genesis, and refuses unless the offered key appears in it unchanged. There is no override flag,
  because a flag that exists gets used on the day the log is briefly unreachable.
- **Removal is named for what it does.** It revokes grants and flags the vault for rekey. It does not
  claim to reach anything already downloaded, and the interface says the remediation is rotating the
  credential — the same non-retroactive limit ADR 0001 records.
- **Ownership is sole, so handing it over is one write and not a role change.** If membership
  authorises, the owner's membership is the last authority in the team, and a transfer that stopped
  halfway would leave nobody with the standing to finish it — owned twice if the promotion went first,
  owned by nobody if the demotion did, and in either case recoverable only by an operator editing the
  database. So `POST /teams/{id}/owner` promotes the recipient and demotes the outgoing owner to
  **admin** in one transaction, `ChangeRoleAsync` refuses `Owner` outright, and the recipient must
  already be an active member — handing a team to an id supplied once is the same mistake as adding
  somebody straight to the owner role. Demoting rather than removing is the deliberate half: removing
  them would revoke their vault key grants and flag every team vault for rekey, which is a far larger
  act than the one being asked for, and somebody handing over a team is usually staying in it.

One thing is deliberately **not** built, and it is a refusal rather than an omission:

- **The rekey itself.** Only a client holding the current vault key can re-wrap every item's data key
  under a new one. The server records that a rotation is owed and the interface reports it. M5.

Two smaller choices, recorded because the alternative was written down first and rejected:

- **No `v_user_vault_permission` view.** ADR-adjacent notes and the old `VaultAccessService` remark
  both anticipated one. The rules turned out to be about sixteen lines of C# shared by the two
  methods that need them; a view would have moved the authorisation model into migrations, where a
  test cannot reach it without a container.
- **Host key trust stays vault-scoped to the personal vault.** Pins in a team vault are listed but
  not consulted at connect time. Consulting them would let any member with Write pre-approve a
  fingerprint that another member's client then trusts silently for a host in their *own* vault,
  which is a cross-boundary trust escalation. Scoping trust properly needs a scope on the SSH connect
  path (`IKnownHostStore.FindAsync` takes host, port and algorithm and knows nothing about vaults);
  until that exists, the safe direction is the narrow one, and the cost — approving a team host's key
  once per member per machine — is stated in the README rather than hidden.

### An invitation is membership decided before there is an account to hold it

A membership names an account: `team_membership.user_id` is not nullable and carries a foreign key, so
somebody who has never signed in here has nothing for that row to point at. `MembershipStatus.Invited`
has existed since the first migration and is still never written — not as an oversight, but because a
membership waiting for a person is the one shape this model cannot store. An invitation is therefore its
own record, `team_invitation`, held against an **address**, and it becomes an ordinary active membership
the moment an account with that address signs in.

That extends the model rather than bending it. An invitation grants nothing readable and cannot be a
step towards sharing, because there is no account and so no public key to wrap a vault to. It moves the
first half of the split earlier and leaves the second half exactly where it was.

Three decisions inside it belong here, because each had a more convenient alternative:

- **The claim requires `email_verified` on the access token, and nothing relaxes it.** This is the whole
  of the security boundary. Membership is authorisation, so an invitation that could be taken by anybody
  able to obtain a token asserting somebody else's address is a way into a team — the same attack
  `OidcOptions.AllowEmailLinking` exists to refuse, arriving by another door and deserving the same bar.
  An unverified or absent claim claims nothing and logs a warning, which is the only signal an operator
  gets that their provider is not sending it. There is deliberately no setting to trust an unverified
  address: a flag that exists is a flag somebody turns on for the afternoon their provider is
  misconfigured, and this is the one it must not be possible to turn on.
- **Nothing is sent, and the product says so rather than implying a mail path it has not got.** There is
  no token and no link — the row is a standing instruction, and telling the invitee to go and sign in
  happens over a channel this server does not carry. A link nobody can deliver would be worse than none.
  The compensation, such as it is, is real: an invitation that is not a bearer credential is one that
  cannot be forwarded, intercepted or replayed.
- **An address that already has an account here is accepted rather than refused.** Refusing and pointing
  at the directory would have been tidier, and would have turned the endpoint into an oracle for which
  addresses have accounts on this deployment, answerable by anybody willing to create a team first. Only
  an address already belonging to a member of *this* team is refused, and that is a fact the caller can
  already read off the members table, so naming it leaks nothing.

## Consequences

The sharing graph is visible to the operator: who is in which team, which vaults exist, and who holds
a grant are all plaintext rows. That was already true of metadata generally (`docs/crypto.md` §10)
and is not made worse here, but it is now a graph rather than a list. Invitations widen it by one
edge — an address that has been invited is on the graph before its owner has ever been here — which is
the same class of fact and worth naming rather than leaving to be noticed.

A malicious granter can seal garbage. The recipient detects it as a tag failure and the grant's
Ed25519 signature names who issued it — detectable and attributable, which is the most that is
achievable without the server holding a key.

The two-step model costs a step in the interface and buys the property the whole product is for. It
also makes a class of bug impossible: there is no code path on the server that could accidentally
grant read access to plaintext, because there is no plaintext on the server to grant.