Skip to main content

DELETE /api/principals/{id}

Remove a principal -- DISABLING it, CHECKING ITS KEYS IN, and REASSIGNING ITS KEYS to a recipient.

WARNING: this does considerably more than the path suggests, and the extra work is not optional. Removing a principal REASSIGNS ITS KEYS to another principal, and the handler performs a three-step sequence:

  1. if the principal is enabled, it is DISABLED and that disable is written;
  2. an emergency key check-in (EMERGENCY_CHECKIN) is sent for the principal, so keys it holds checked out at a remote engine are checked in while it still exists;
  3. it is then removed, with its keys reassigned to the recipient.

A refused check-in does not stop the removal. A principal holding nothing checked out is the ordinary case, so the refusal is classified rather than fatal: the removal goes ahead and the 200 body reports checkedIn: false with the key server's answer in checkInRefusal.

If step 3 fails, the principal is left DISABLED, its keys checked in, and still present. The 409 says so and reports what the re-read before the removal saw. This is not a rollback-safe operation and there is no undo.

The recipient defaults to the calling session's own user. Name a different one with ?recipient=<name>.

WARNING: the recipient lookup resolves PASSWORD principals ONLY. A group, a certificate principal and an XAuth id are all unresolvable as recipients even when they exist. In particular, on a DELEGATED session the session user IS an XAuth id, so the default can never resolve there and ?recipient= must be given. When the lookup fails nothing is written -- the principal is untouched and still enabled, and no check-in is sent.

On success the answer is 200 with a body reporting whether the check-in ran. (It was 204 with no body until TE81-457 / ADR 0134: a 204 cannot say that the check-in was refused.)

Authentication​

Send the session id as Authorization: Bearer <session id>, and an RFC 9449 DPoP proof in the DPoP header if your session is bound to a key. The two are required together, not as alternatives — see Authenticating and Proving you hold your key.

This surface sets no cookie

There is no session cookie on any response and none is read on any request. Do not set credentials: "include" on a browser fetch.

Parameters​

NameInRequiredTypeDescription
idpathyesstringA principal id as 16 hex digits. This is the id field of a principal projection -- NOT the principalId decimal string beside it. The two are the same eight bytes in two spellings, and only the hex one is a path segment.
recipientquerynostringThe name of the PASSWORD principal that inherits this principal's keys. Omit to use the calling session's own user. Must not be empty if given.

Request​

No request body.

Responses​

200​

The principal is gone and its keys were reassigned. Verified by a fresh read.

checkedIn is true when the emergency check-in sent before the removal succeeded. When it is false, checkInRefusal carries the key server's refusal verbatim and the removal went ahead anyway -- a principal holding nothing checked out is the ordinary cause.

FieldTypeRequiredDescription
idstringyes
removedtrue (constant)yes
recipientstringyesThe name of the principal that inherited the keys.
recipientDefaultedbooleanyestrue when no ?recipient= was named and the session's own user was used.
disabledFirstbooleanyestrue when the principal was enabled and this request disabled it.
checkedInbooleanyes
checkInRefusalstring or nullyes
detailstringyes

Example

{
"id": "...",
"removed": true,
"recipient": "...",
"recipientDefaulted": true,
"disabledFirst": true,
"checkedIn": true,
"checkInRefusal": "...",
"detail": "..."
}

400​

BAD_REQUEST -- the id is not 16 hex digits, or recipient was given as an empty string.

401​

No live session, a DPoP proof that did not verify, or a certificate session whose certificate the key server now refuses.

SESSION_NOT_FOUND -- the credential names no session this Gateway holds. SESSION_EXPIRED -- it aged out. Both carry reauth: true; obtain a new credential rather than retrying.

PROOF_* -- the sender constraint refused. See the DPoP section of this document for the full table; PROOF_MALFORMED is a 400 and PROOF_REPLAY_CACHE_FULL a 503, and the rest are here.

SCOPED_CREDENTIAL_EXPIRED -- a scoped browser credential passed its fixed lifetime. Use does not extend it: an idle timeout would be refreshed by exactly the traffic a stolen handle produces.

CERTIFICATE_REVOKED -- on a session established by certificate login, the key server re-checks the certificate that session was established with on its operations, and it is now named by revocation evidence (ADR 0115). detail is the key server's sentence. Every later operation of the session is refused the same way; log in with a different certificate.

403​

Authenticated, and refused. The codes differ in what they say about where the refusal came from: ACCESS_DENIED, PRINCIPAL_DISABLED and DELEGATOR_REFUSED are the key server's -- the last one on a DELEGATED session whose delegator has since been disabled, lost its DELEGATION flag, been unenrolled or had its certificate revoked (ADR-0123); its detail is the key server's sentence naming the delegator, and no new login helps until an administrator restores it. REVOCATION_STATUS_UNAVAILABLE and REVOCATION_EVIDENCE_STALE are the key server's too, on a session established by certificate login: the same re-check found no current revocation evidence for that certificate's issuer under a policy that requires it (ADR 0115) -- the credential is not what was refused, and an operator importing a current CRL for the issuer is what helps -- while NOT_WRITABLE, BUILT_IN_PRINCIPAL, OWN_ACCOUNT and OBJECT_NOT_IN_CREDENTIAL_SCOPE are decided by the Gateway before anything is sent.

404​

NOT_FOUND -- no such principal. Nothing was written.

409​

NO_RECIPIENT -- none was named and the session carries no user name. Nothing was written.

NO_SUCH_RECIPIENT -- the named (or defaulted) recipient could not be resolved. Nothing was written: the principal is untouched and still enabled. recipientWasDefaulted says whether you named it.

RECIPIENT_IS_TARGET -- the named recipient IS the principal being removed. Nothing was written: no disable, no check-in, no removal.

NOT_SAVED -- the principal was still present when read back after removal. It may now be disabled and its keys checked in: the earlier steps had already run. disabledOnReRead reports what the re-read before the removal saw; checkedIn and checkInRefusal report the check-in.

500​

A fault rather than a refusal.

OP_FAILED is the catch-all: the key server refused in words that match no classified pattern, so the Gateway cannot say more than that the operation did not happen. Some key server refusals that are conceptually a caller error arrive here rather than as a 4xx, because the wire carries no distinguishing code -- where a route can recognise one from its message it maps it, and each such mapping is documented on the operation.

503​

A capacity ceiling or a missing deployment prerequisite. retryable says which.

A scope of session or process names the ceiling that was hit, and the response carries Retry-After.

BUILD_NOT_ADMITTED (scope: process, retryable: true, Retry-After): the key server does not admit this Gateway's client build. Every key-server connection the Gateway opens attests the signed release it runs (TE_RELEASE_MANIFEST) right after it authenticates; while that attestation is refused -- the release is not registered at the key server's code-signing level, the level was raised under it, or the installed files no longer match the signed manifest -- every route but GET /api/health and GET /api/ready answers this, and detail names the cause. It is never a 401 or a 403: no credential or role of the caller's is at fault. The Gateway re-attempts on a timer and recovers without a restart once the key server admits it; GET /api/ready says when.

Error handling​

Every failure answers JSON carrying at least code and detail. Match on code — detail is written for a human debugging the call and its wording is not part of the contract. See the error model.

StatusMeaning
400BAD_REQUEST -- the id is not 16 hex digits, or recipient was given as an empty string.
401No live session, a DPoP proof that did not verify, or a certificate session whose certificate the key server now refuses. SESSION_NOT_FOUND -- the credential names no session this Gateway holds. SESSION_EXPIRED -- it aged out. Both carry reauth: true; obtain a new credential rather than retrying. PROOF_* -- the sender constraint refused. See the DPoP section of this document for the full table; PROOF_MALFORMED is a 400 and PROOF_REPLAY_CACHE_FULL a 503, and the rest are here. SCOPED_CREDENTIAL_EXPIRED -- a scoped browser credential passed its fixed lifetime. Use does not extend it: an idle timeout would be refreshed by exactly the traffic a stolen handle produces. CERTIFICATE_REVOKED -- on a session established by certificate login, the key server re-checks the certificate that session was established with on its operations, and it is now named by revocation evidence (ADR 0115). detail is the key server's sentence. Every later operation of the session is refused the same way; log in with a different certificate.
403Authenticated, and refused. The codes differ in what they say about where the refusal came from: ACCESS_DENIED, PRINCIPAL_DISABLED and DELEGATOR_REFUSED are the key server's -- the last one on a DELEGATED session whose delegator has since been disabled, lost its DELEGATION flag, been unenrolled or had its certificate revoked (ADR-0123); its detail is the key server's sentence naming the delegator, and no new login helps until an administrator restores it. REVOCATION_STATUS_UNAVAILABLE and REVOCATION_EVIDENCE_STALE are the key server's too, on a session established by certificate login: the same re-check found no current revocation evidence for that certificate's issuer under a policy that requires it (ADR 0115) -- the credential is not what was refused, and an operator importing a current CRL for the issuer is what helps -- while NOT_WRITABLE, BUILT_IN_PRINCIPAL, OWN_ACCOUNT and OBJECT_NOT_IN_CREDENTIAL_SCOPE are decided by the Gateway before anything is sent.
404NOT_FOUND -- no such principal. Nothing was written.
409NO_RECIPIENT -- none was named and the session carries no user name. Nothing was written. NO_SUCH_RECIPIENT -- the named (or defaulted) recipient could not be resolved. Nothing was written: the principal is untouched and still enabled. recipientWasDefaulted says whether you named it. RECIPIENT_IS_TARGET -- the named recipient IS the principal being removed. Nothing was written: no disable, no check-in, no removal. NOT_SAVED -- the principal was still present when read back after removal. It may now be disabled and its keys checked in: the earlier steps had already run. disabledOnReRead reports what the re-read before the removal saw; checkedIn and checkInRefusal report the check-in.
500A fault rather than a refusal. OP_FAILED is the catch-all: the key server refused in words that match no classified pattern, so the Gateway cannot say more than that the operation did not happen. Some key server refusals that are conceptually a caller error arrive here rather than as a 4xx, because the wire carries no distinguishing code -- where a route can recognise one from its message it maps it, and each such mapping is documented on the operation.
503A capacity ceiling or a missing deployment prerequisite. retryable says which. A scope of session or process names the ceiling that was hit, and the response carries Retry-After. BUILD_NOT_ADMITTED (scope: process, retryable: true, Retry-After): the key server does not admit this Gateway's client build. Every key-server connection the Gateway opens attests the signed release it runs (TE_RELEASE_MANIFEST) right after it authenticates; while that attestation is refused -- the release is not registered at the key server's code-signing level, the level was raised under it, or the installed files no longer match the signed manifest -- every route but GET /api/health and GET /api/ready answers this, and detail names the cause. It is never a 401 or a 403: no credential or role of the caller's is at fault. The Gateway re-attempts on a timer and recovers without a restart once the key server admits it; GET /api/ready says when.