diff --git a/CHANGELOG.md b/CHANGELOG.md
index b61b318fa..daf954ff0 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -9,20 +9,20 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Added
 
-- `POST /sessions/{sessionId}/chats/read` accepts an optional `messageIds` array (up to 100) naming exactly which messages to acknowledge. Baileys acknowledges individual messages, so without it only the newest message still held in memory got a receipt: a burst left its earlier messages unread, and a session restarted since the message arrived had nothing to acknowledge and answered `false` under a `200`. Each id is resolved through the message store, so a group receipt carries the `participant` that attributes it. Omitting the field keeps the previous behaviour, an empty array is refused with a `400` rather than silently acknowledging the newest message, and the whatsapp-web.js engine ignores the field because its own receipt is chat-level. The `SessionMarkChatRead` agent tool takes the same list, and all five clients expose the field through a dedicated read-request type. Thanks @m7fz7.
+- `POST /sessions/{sessionId}/chats/read` takes an optional `messageIds` array (up to 100) naming which messages to acknowledge. Baileys acknowledges individual messages, so without it a burst left its earlier messages unread. Ids resolve through the message store, so a group receipt carries its `participant`. Available on the agent tool and all five clients; ignored by whatsapp-web.js. Thanks @m7fz7.
 
 ### Changed
 
-- `POST /sessions/{sessionId}/chats/unread` publishes its own `MarkChatUnreadDto` instead of sharing `MarkChatReadDto` with the read route. The request body is unchanged (`chatId` alone), but a client generated from the contract sees the schema under a new name. Sharing the class would have advertised `messageIds` on a route that discards it.
+- `POST /sessions/{sessionId}/chats/unread` publishes its own `MarkChatUnreadDto` rather than sharing `MarkChatReadDto`. The body is unchanged (`chatId` alone), but a generated client sees the schema under a new name.
 
 ### Fixed
 
 - The dashboard CSP nonce is substituted at every occurrence in the served document, not only the first. One placeholder exists today, so a second would have been left reading the literal text and its script refused by the browser.
-- Outbound webhook deliveries survive a hard crash. Fan-out is fire-and-forget, so a crash between persisting a message and completing its POST lost the delivery with no record in either mode, while the documented contract promises at-least-once. Each delivery is now recorded before it is attempted, retired once the queue or the inline send owns it, and a bounded sweep replays whatever is left stranded using the stored idempotency key so the retry stays deduplicable at the receiver.
+- Outbound webhook deliveries survive a hard crash. Fan-out was fire-and-forget, so a crash between persisting a message and completing its POST lost the delivery, against a documented at-least-once contract. Deliveries are now recorded before they are attempted, and a bounded sweep replays whatever is stranded under its stored idempotency key.
 - Settled outbound delivery records are pruned after `WEBHOOK_OUTBOX_RETENTION_DAYS` (default 7). A record that can still be replayed is never pruned on age, and a non-positive window falls back to the default rather than letting the table grow without bound.
-- `PLUGIN_STATE_DIR` moves the plugin registry and per-plugin storage off the default `./data`. It was the one piece of state with no path knob, so a test run rewrote the developer's own registry and every suite in that run inherited whatever the previous one left in it.
+- `PLUGIN_STATE_DIR` moves the plugin registry and per-plugin storage off the default `./data`. It was the one piece of state with no path knob, so a test run rewrote the developer's own registry.
 - The e2e lane sweeps the throwaway state roots it creates. Each suite gets its own, nothing removed them, and the temp directory accumulated hundreds of entries over a few days of runs.
-- e2e assertions are no longer answered by unrelated processes on the host. supertest starts a listener per request on the wildcard address and then dials 127.0.0.1, and macOS both permits that bind over a port another process holds on 127.0.0.1 specifically and routes the connection to the more specific holder, so a run could assert against a status the app has no route for. Each suite's server now listens on loopback while it initialises, which supertest reuses instead of opening one, taking the lane from 230 listeners a run to 27.
+- e2e assertions are no longer answered by unrelated processes on the host. supertest binds its per-request listener to the wildcard address and then dials 127.0.0.1, which on macOS lets a process holding that port on 127.0.0.1 answer instead. Each suite's server now listens on loopback during init, which supertest reuses.
 
 ### Tests
 
diff --git a/src/core/agent-tools/tools/session.tools.spec.ts b/src/core/agent-tools/tools/session.tools.spec.ts
index 84cbf657b..076881410 100644
--- a/src/core/agent-tools/tools/session.tools.spec.ts
+++ b/src/core/agent-tools/tools/session.tools.spec.ts
@@ -1,3 +1,4 @@
+import { BadRequestException } from '@nestjs/common';
 import { invokeTool } from '../tool-invoker';
 import { sessionTools } from './session.tools';
 import type { AnyToolDescriptor } from '../tool-descriptor';
@@ -104,6 +105,27 @@ describe('sessionTools', () => {
     expect(sendSeen).toHaveBeenCalledWith('s1', '628111@c.us', ['M1', 'M2']);
   });
 
+  // The REST body rejects a whitespace-only id, and this path never sees that DTO: invokeTool parses
+  // the tool's own schema and calls the service directly, so an id the schema accepts is sent as a
+  // receipt key. Asserting the service was not reached is the point; a 400 that still delegated
+  // would leave the defect in place.
+  it('SessionMarkChatRead refuses a whitespace-only message id instead of sending it as a key', async () => {
+    const sendSeen = jest.fn().mockResolvedValue(true);
+    const tool = makeTools({ sendSeen } as unknown as SessionService).get('SessionMarkChatRead')!;
+
+    const failure = await run(tool, { sessionId: 's1', chatId: '628111@c.us', messageIds: ['  '] }).catch(
+      (e: unknown) => e,
+    );
+    expect(failure).toBeInstanceOf(BadRequestException);
+    expect(JSON.stringify((failure as BadRequestException).getResponse())).toContain('no whitespace');
+    expect(sendSeen).not.toHaveBeenCalled();
+
+    // Control: a well-formed id still reaches the service, so the refusal above is this rule and not
+    // a schema that turns everything away.
+    await run(tool, { sessionId: 's1', chatId: '628111@c.us', messageIds: ['3EB0C767D26B8A3F1A2B'] });
+    expect(sendSeen).toHaveBeenCalledWith('s1', '628111@c.us', ['3EB0C767D26B8A3F1A2B']);
+  });
+
   it('SessionMarkChatUnread maps the markUnread result to a success field', async () => {
     const markUnread = jest.fn().mockResolvedValue(true);
     const out = await run(makeTools({ markUnread } as unknown as SessionService).get('SessionMarkChatUnread')!, {
diff --git a/src/core/agent-tools/tools/session.tools.ts b/src/core/agent-tools/tools/session.tools.ts
index 916d203e8..0fb75dd3a 100644
--- a/src/core/agent-tools/tools/session.tools.ts
+++ b/src/core/agent-tools/tools/session.tools.ts
@@ -2,7 +2,11 @@ import { z } from 'zod';
 import { ApiKeyRole } from '../../../modules/auth/entities/api-key.entity';
 import type { SessionService } from '../../../modules/session/session.service';
 import { SessionResponseDto } from '../../../modules/session/dto/session-response.dto';
-import { MARK_READ_MESSAGE_IDS_MAX } from '../../../modules/session/dto/mark-chat-read.dto';
+import {
+  MARK_READ_MESSAGE_IDS_MAX,
+  MARK_READ_MESSAGE_ID_MESSAGE,
+  MARK_READ_MESSAGE_ID_PATTERN,
+} from '../../../modules/session/dto/mark-chat-read.dto';
 import { defineTool, type AnyToolDescriptor } from '../tool-descriptor';
 
 const sessionId = z.string().min(1).describe('Session UUID (the session id, not the name)');
@@ -91,7 +95,9 @@ export function sessionTools(session: SessionService): AnyToolDescriptor[] {
         sessionId,
         chatId: z.string().describe('Chat JID (e.g. 1234567890@c.us)'),
         messageIds: z
-          .array(z.string())
+          // The element rule comes from the DTO rather than being restated here: the REST body
+          // rejects a whitespace-only id, and this path reaches the engine without the DTO at all.
+          .array(z.string().regex(MARK_READ_MESSAGE_ID_PATTERN, MARK_READ_MESSAGE_ID_MESSAGE))
           .nonempty()
           .max(MARK_READ_MESSAGE_IDS_MAX)
           .optional()
diff --git a/src/modules/session/dto/mark-chat-read.dto.ts b/src/modules/session/dto/mark-chat-read.dto.ts
index d47648721..e993de6fa 100644
--- a/src/modules/session/dto/mark-chat-read.dto.ts
+++ b/src/modules/session/dto/mark-chat-read.dto.ts
@@ -8,6 +8,16 @@ import { ArrayMaxSize, ArrayNotEmpty, IsArray, IsNotEmpty, IsOptional, IsString,
  */
 export const MARK_READ_MESSAGE_IDS_MAX = 100;
 
+/**
+ * The shape of one message id: a single non-whitespace token. Exported for the same reason as the
+ * cap above, so the agent tool holds the rule rather than restating it. Restating it is how the two
+ * surfaces drifted: the tool copied the count and left `['  ']` reaching the engine as a receipt key.
+ */
+export const MARK_READ_MESSAGE_ID_PATTERN = /^\S{1,128}$/;
+
+/** Rejection message for {@link MARK_READ_MESSAGE_ID_PATTERN}, shared so both surfaces answer alike. */
+export const MARK_READ_MESSAGE_ID_MESSAGE = 'each messageIds entry must be a non-empty id with no whitespace';
+
 export class MarkChatReadDto {
   @ApiProperty({
     description: "Chat ID in the active engine's native format (e.g. 1234567890@c.us on whatsapp-web.js)",
@@ -46,8 +56,6 @@ export class MarkChatReadDto {
   @ArrayMaxSize(MARK_READ_MESSAGE_IDS_MAX)
   @IsString({ each: true })
   @IsNotEmpty({ each: true })
-  // Engine-neutral structural check: a message id is one non-whitespace token. @IsNotEmpty alone
-  // accepts '   ', which would reach sendNode as a receipt key.
-  @Matches(/^\S{1,128}$/, { each: true, message: 'each messageIds entry must be a non-empty id with no whitespace' })
+  @Matches(MARK_READ_MESSAGE_ID_PATTERN, { each: true, message: MARK_READ_MESSAGE_ID_MESSAGE })
   messageIds?: string[];
 }