diff --git a/CHANGELOG.md b/CHANGELOG.md
index 7f9c610e..5e7e9aad 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -9,9 +9,42 @@ Versioning: [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
 
+## [2.26.0] - 2026-10-04
+
+Minor release: the Verbara.Sdk `2.7.0` / Verbara.Sdk.Pro `2.17.1-pro` cascade, with the queue-call
+ownership behaviour it brings; idle `Available` agents keep receiving queue calls; and message history
+is readable again on PostgreSQL. There is no database migration.
+
+### Upgrading from 2.25.x
+
+- **AMI read class `agent`.** Queued calls now get an owner only when the queue reports that a member
+  took the call, which arrives on the AMI `agent` class. The bundled
+  `docker/asterisk-config/manager.conf.template` already grants it. If Platform connects to your own
+  Asterisk with its own AMI user, add `agent` to that user's `read =` line in `manager.conf` and
+  reload the manager (`manager reload`), or queue calls stay without an owner.
+
+### Known issues
+
+- **Compose start-up order leaves Asterisk's `queue_members` column cache empty** when Asterisk starts
+  before the API has run its migrations (#333). Running `module reload app_queue.so` in the Asterisk
+  container after the API is up refreshes the cache.
+
 ### Fixed
 
-- **Message history can be read again on PostgreSQL.** Every read of a stored message failed with a server error (500). Postgres stores message content as `jsonb`, which reorders the keys of each block, and the reader required the block's type to come first. The bug was latent since message storage first shipped (1.2.0). It broke `GET /api/v1/conversations/{id}/messages`, `GET /api/v1/supervisor/conversations/{id}/messages`, typification suggestions, the GDPR export, the conversation summary report, delivery-status webhooks, de-duplication of redelivered inbound messages and email thread matching. The in-memory store was not affected. No migration is needed: rows already stored are read correctly after the upgrade.
+- **Message history can be read again on PostgreSQL.** Every read of a stored message failed with a server error (500). Postgres stores message content as `jsonb`, which reorders the keys of each block, and the reader required the block's type to come first. The bug was latent since message storage first shipped (1.2.0). It broke `GET /api/v1/conversations/{id}/messages`, `GET /api/v1/supervisor/conversations/{id}/messages`, typification suggestions, the GDPR export, the conversation summary report, delivery-status webhooks, de-duplication of redelivered inbound messages and email thread matching. The in-memory store was not affected. No migration is needed: rows already stored are read correctly after the upgrade. (#337)
+- **An idle `Available` agent no longer stops receiving queue calls after a minute.** Every 60 s the
+  realtime reconciliation (and every queue-membership save) re-wrote the agent's Asterisk Realtime
+  `queue_members` rows with `paused = 1`, so `app_queue` skipped an agent the console still showed as
+  `Available` until the agent changed state by hand, and callers waited while a free agent sat idle.
+  Verbara.Sdk.Pro `2.17.1-pro` now keeps an existing row's `paused` value on re-upsert, and Platform
+  re-asserts the agent's real pause state after every member upsert it issues (reconcile pass and
+  membership save): it re-reads the agent under the same per-agent lock the state bridge uses and
+  pauses only a non-routable agent (`Available` and `Busy` are routable) or one with a pending pause.
+  This also unpauses a member created, or re-created after orphan cleanup, for an agent who is
+  already `Available`. Rows already stuck at `paused = 1` are corrected by the first reconcile after
+  the upgrade. **Known limitation:** the per-agent lock is per process. With several API replicas,
+  one pod's reconcile can still briefly race another pod's state change; the next state change or
+  reconcile re-asserts the correct value. A cross-pod lock is a separate change. (#338)
 
 ### Dependencies
 
@@ -42,29 +75,14 @@ Versioning: [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
   overload and the `StateChanged` event — to the primary connection. The interface gives them default
   bodies, so the wrapper compiled clean while answering `false`, dropping the outcome and raising
   nothing; four tests now lock the forwarding. Platform does not call any of them yet.
+- **`Scalar.AspNetCore` `2.17.10` → `2.17.11`** (Dependabot, #322).
 - **Deliberately out of scope** (`Verbara.Sdk/ADR-0040` D4): `Microsoft.Extensions.TimeProvider.Testing`
   stays at `10.10.0`, and neither `OpenTelemetry` itself nor `NATS.Client.Core` gains a direct pin.
   The second hop, Sdk `2.8.0` (H142), is a separate change. decision_ref `Verbara.Sdk/ADR-0040`.
 - **Verbara.Sdk.Pro `2.17.0-pro` → `2.17.1-pro`** — all 22 `Verbara.Sdk.Pro.*` pins move together
   (Pro packages depend on their siblings at the same version, so a partial bump risks `NU1605`).
   A patch with no API change: the `queue_members` upsert stops overwriting `paused` on conflict (see
-  *Fixed*). (#N)
-
-### Fixed
-
-- **An idle `Available` agent no longer stops receiving queue calls after a minute.** Every 60 s the
-  realtime reconciliation (and every queue-membership save) re-wrote the agent's Asterisk Realtime
-  `queue_members` rows with `paused = 1`, so `app_queue` skipped an agent the console still showed as
-  `Available` until the agent changed state by hand, and callers waited while a free agent sat idle.
-  Verbara.Sdk.Pro `2.17.1-pro` now keeps an existing row's `paused` value on re-upsert, and Platform
-  re-asserts the agent's real pause state after every member upsert it issues (reconcile pass and
-  membership save): it re-reads the agent under the same per-agent lock the state bridge uses and
-  pauses only a non-routable agent (`Available` and `Busy` are routable) or one with a pending pause.
-  This also unpauses a member created, or re-created after orphan cleanup, for an agent who is
-  already `Available`. Rows already stuck at `paused = 1` are corrected by the first reconcile after
-  the upgrade. **Known limitation:** the per-agent lock is per process. With several API replicas,
-  one pod's reconcile can still briefly race another pod's state change; the next state change or
-  reconcile re-asserts the correct value. A cross-pod lock is a separate change. (#N)
+  *Fixed*). (#338)
 
 ---
 
diff --git a/Directory.Build.props b/Directory.Build.props
index dec39c38..2ab69c80 100644
--- a/Directory.Build.props
+++ b/Directory.Build.props
@@ -26,7 +26,7 @@
     <RepositoryType>git</RepositoryType>
     <RepositoryUrl>https://github.com/verbara/verbara-platform</RepositoryUrl>
     <PackageTags>verbara;platform;contact-center;omnichannel;native-aot</PackageTags>
-    <PackageVersion>2.25.0</PackageVersion>
+    <PackageVersion>2.26.0</PackageVersion>
   </PropertyGroup>
 
   <PropertyGroup Condition="$(MSBuildProjectDirectory.Contains('src'))">