Delegated token lifecycle
The three deliberate constraints
1. TTL ≤ 300 s (hard cap 900 in code)
A delegated token is a fast-path credential, not a source of truth. The shorter it lives, the
shorter the maximum staleness after a revocation. The 900-second cap is enforced in
TokenExchangeGrant — configuration cannot raise it, because raising it silently converts “revoked
at the next check” into “revoked at expiry”.
2. No refresh token — the re-exchange IS the freshness check
A refresh token would be a long-lived credential in the agent’s hands — exactly what delegation
exists to avoid. Instead, when the token expires the orchestrator exchanges again, and every
exchange re-verifies: agent still active, user’s session still alive in the SessionRegistry,
grant still Active, scopes still inside the intersection. Revocation, suspension, logout — all
land within one TTL, automatically.
3. Introspection-mandatory verification
A resource server that trusts local signature parsing keeps honouring a delegated token for its
whole TTL even after revocation — and a legacy verifier that ignores act would treat it as the
user’s full authority. So delegated tokens are verified by asking the server
(client guide): the answer
includes session liveness and the authoritative claims. typ: delegated+jwt and a dedicated aud
are hygiene layers on top — never the defence.
Revocation, end to end
| Who revokes | How | Lands |
|---|---|---|
| The user | DELETE /iam/me/delegations/{grantId} — one click, no step-up (revoking must always be easier than granting) |
Next introspection/checkDelegated citing pds_dgr denies; next exchange refuses |
| An admin | POST …/delegation-grants/{id}/revoke (console kill-switch) |
Same |
| Agent suspension | POST …/agents/{id}/suspend |
Every grant of that agent stops working (agent layer denies) |
| User logout / session revoked | Server session registry | Next exchange refuses (sid liveness re-checked); introspection reports inactive |
| Webhook push | Sealed audit events (iam.delegation.grant.revoked, …) are pushed to subscribers |
PEPs/agents learn without polling |
Maximum staleness in every row: one delegated TTL (≤ 5 minutes by default).
