fix(comments): suggestion с вечным 409 (#496)

expectedText брался из debounced REST-снапшота (getAnchoredText по
pages/info), а метка ставилась в live-доке — при расхождении (док ушёл
вперёд за окно дебаунса) apply строго сравнивал текст под меткой с
устаревшим stored selection и давал 409 на каждый вызов.

MCP-клиент теперь в transform-фазе (та же версия live-дока, где ставится
метка) перечитывает фактическую подстроку под меткой и, если она
отличается от сохранённого selection, синкает её через новый эндпоинт
POST /comments/resync-suggestion-anchor. Best-effort: сбой синка не
откатывает уже заякоренный комментарий, а лишь выдаёт мягкое
предупреждение. Совпадающий снапшот не делает лишнего round-trip.

Сервер: resyncSuggestionAnchor правит только stored selection
незаселённой suggestion своего автора (guards: top-level, есть
suggestedText, не applied/resolved, отличается от suggestedText),
идемпотентно, без ws-бродкаста.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

apps/server/src/app.module.ts

@@ -25,7 +25,7 @@ import { CacheModule } from '@nestjs/cache-manager';
 import KeyvRedis from '@keyv/redis';
 import { LoggerModule } from './common/logger/logger.module';
 import { ClsModule } from 'nestjs-cls';
-import { NoopAuditModule } from './integrations/audit/audit.module';
+import { AuditModule } from './integrations/audit/audit.module';
 import { ThrottleModule } from './integrations/throttle/throttle.module';
 import { McpModule } from './integrations/mcp/mcp.module';
 import { SandboxModule } from './integrations/sandbox/sandbox.module';
@@ -55,7 +55,7 @@ try {
       middleware: { mount: true },
     }),
     LoggerModule,
-    NoopAuditModule,
+    AuditModule,
     CoreModule,
     DatabaseModule,
     EnvironmentModule,

apps/server/src/collaboration/yjs.util.spec.ts

@@ -529,4 +529,79 @@ describe('replaceYjsMarkedText', () => {
     expect(result).toEqual({ applied: false, currentText: 'abcdef' });
     expect(text.toDelta()).toEqual(before);
   });
+
+  // #496: apply must NOT silently strip the replaced run's inline formatting.
+  // Build a paragraph and format the marked range with extra marks, then assert
+  // the replacement carries them.
+  function buildFormatted(
+    runs: Array<{ text: string; attrs?: Record<string, any> }>,
+  ): { fragment: Y.XmlFragment; text: Y.XmlText } {
+    const ydoc = new Y.Doc();
+    const fragment = ydoc.getXmlFragment('default');
+    const para = new Y.XmlElement('paragraph');
+    fragment.insert(0, [para]);
+    const text = new Y.XmlText();
+    para.insert(0, [text]);
+    text.insert(0, runs.map((r) => r.text).join(''));
+    let offset = 0;
+    for (const run of runs) {
+      if (run.attrs) text.format(offset, run.text.length, run.attrs);
+      offset += run.text.length;
+    }
+    return { fragment, text };
+  }
+
+  it('preserves the original run formatting (bold + link) on the replacement', () => {
+    const { fragment, text } = buildFormatted([
+      { text: 'see ' },
+      {
+        text: 'old',
+        attrs: {
+          comment: { commentId: 'c1', resolved: false },
+          bold: true,
+          link: { href: 'https://x.test' },
+        },
+      },
+      { text: ' end' },
+    ]);
+
+    const result = replaceYjsMarkedText(fragment, 'c1', 'old', 'new');
+
+    expect(result).toEqual({ applied: true, currentText: 'new' });
+    // The comment anchor AND the bold/link marks survive the delete+insert.
+    expect(text.toDelta()).toEqual([
+      { insert: 'see ' },
+      {
+        insert: 'new',
+        attributes: {
+          comment: { commentId: 'c1', resolved: false },
+          bold: true,
+          link: { href: 'https://x.test' },
+        },
+      },
+      { insert: ' end' },
+    ]);
+  });
+
+  it('mixed formatting under the mark: replacement takes the DOMINANT run marks', () => {
+    // Marked run = "bold" (bold, 4 chars) + "x" (plain, 1 char), same commentId.
+    // The dominant (longer) segment is bold, so the flat replacement is bold.
+    const { fragment, text } = buildFormatted([
+      {
+        text: 'bold',
+        attrs: { comment: { commentId: 'c1', resolved: false }, bold: true },
+      },
+      { text: 'x', attrs: { comment: { commentId: 'c1', resolved: false } } },
+    ]);
+
+    const result = replaceYjsMarkedText(fragment, 'c1', 'boldx', 'Z');
+
+    expect(result).toEqual({ applied: true, currentText: 'Z' });
+    expect(text.toDelta()).toEqual([
+      {
+        insert: 'Z',
+        attributes: { comment: { commentId: 'c1', resolved: false }, bold: true },
+      },
+    ]);
+  });
 });

apps/server/src/collaboration/yjs.util.ts

@@ -145,6 +145,10 @@ type MarkedSegment = {
   length: number;
   text: string;
   markAttrs: Record<string, any>;
+  // The FULL attribute set of this delta run — the `comment` mark plus any
+  // inline formatting (bold/italic/code/link/…). Captured so apply can carry the
+  // original run's formatting onto the replacement instead of dropping it.
+  attributes: Record<string, any>;
 };
 
 /**
@@ -202,6 +206,7 @@ export function replaceYjsMarkedText(
             length,
             text: insert,
             markAttrs: markAttr,
+            attributes,
           });
         }
         offset += length;
@@ -251,15 +256,26 @@ export function replaceYjsMarkedText(
     return { applied: false, currentText: joinedText };
   }
 
-  // 3. All guards passed: delete the marked run and re-insert newText with the
-  // same comment attributes at the same offset. Atomic within the caller's
-  // transaction.
+  // 3. All guards passed: delete the marked run and re-insert newText at the
+  // same offset. Atomic within the caller's transaction.
   const start = segments[0].offset;
   const len = segments.reduce((sum, s) => sum + s.length, 0);
-  const markAttrs = segments[0].markAttrs;
+
+  // Carry the ORIGINAL run's formatting onto the replacement (#496): inserting
+  // with only the `comment` mark silently dropped bold/italic/code/link of the
+  // replaced text. Yjs applies one flat attribute set to the whole insert, so
+  // when the marked run mixes formatting we pick the DOMINANT segment (the one
+  // covering the most characters) and apply its attributes — a v1 that preserves
+  // the common single-format case exactly and, for a mixed run, keeps the
+  // prevailing style rather than losing all of it. The dominant segment already
+  // carries the `comment` mark (every collected segment does), so the anchor is
+  // preserved; we defensively re-assert it in case a future attribute shape
+  // omits it.
+  const dominant = segments.reduce((a, b) => (b.length > a.length ? b : a));
+  const insertAttrs = { ...dominant.attributes, comment: dominant.markAttrs };
 
   node.delete(start, len);
-  node.insert(start, newText, { comment: markAttrs });
+  node.insert(start, newText, insertAttrs);
 
   return { applied: true, currentText: newText };
 }

apps/server/src/core/comment/comment.controller.ts

@@ -16,6 +16,7 @@ import { UpdateCommentDto } from './dto/update-comment.dto';
 import { ResolveCommentDto } from './dto/resolve-comment.dto';
 import { ApplySuggestionDto } from './dto/apply-suggestion.dto';
 import { DismissSuggestionDto } from './dto/dismiss-suggestion.dto';
+import { ResyncSuggestionAnchorDto } from './dto/resync-suggestion-anchor.dto';
 import { PageIdDto, CommentIdDto } from './dto/comments.input';
 import { AuthUser } from '../../common/decorators/auth-user.decorator';
 import { AuthWorkspace } from '../../common/decorators/auth-workspace.decorator';
@@ -235,6 +236,39 @@ export class CommentController {
     return this.commentService.applySuggestion(comment, user, provenance);
   }
 
+  @HttpCode(HttpStatus.OK)
+  @Post('resync-suggestion-anchor')
+  async resyncSuggestionAnchor(
+    @Body() dto: ResyncSuggestionAnchorDto,
+    @AuthUser() user: User,
+    @AuthWorkspace() workspace: Workspace,
+  ) {
+    const comment = await this.commentRepo.findById(dto.commentId, {
+      includeCreator: true,
+      includeResolvedBy: true,
+    });
+    if (!comment) {
+      throw new NotFoundException('Comment not found');
+    }
+
+    const page = await this.pageRepo.findById(comment.pageId);
+    if (!page || page.deletedAt) {
+      throw new NotFoundException('Page not found');
+    }
+
+    // Authorize BEFORE revealing structural detail (mirrors apply/dismiss).
+    // Re-anchoring does NOT change the page text — it only corrects the stored
+    // selection metadata — so the page-level gate is comment access. The service
+    // further restricts it to the suggestion's own author.
+    await this.pageAccessService.validateCanComment(page, user, workspace.id);
+
+    return this.commentService.resyncSuggestionAnchor(
+      comment,
+      dto.selection,
+      user,
+    );
+  }
+
   @HttpCode(HttpStatus.OK)
   @Post('dismiss-suggestion')
   async dismissSuggestion(

apps/server/src/core/comment/comment.service.apply-suggestion.spec.ts

@@ -146,11 +146,19 @@ describe('CommentService — applySuggestion', () => {
       'page-1',
       expect.objectContaining({ operation: 'commentDeleted', commentId: 'c-1' }),
     );
+    // #496: hard-deleted row → the audit payload is the only surviving record.
     expect(auditService.log).toHaveBeenCalledWith(
       expect.objectContaining({
         event: AuditEvent.COMMENT_SUGGESTION_APPLIED,
         resourceType: AuditResource.COMMENT,
         resourceId: 'c-1',
+        metadata: expect.objectContaining({
+          pageId: 'page-1',
+          suggestedText: 'new text',
+          selection: 'old text',
+          commentAuthor: 'user-1',
+          decidedBy: 'user-1',
+        }),
       }),
     );
     expect(result.outcome).toBe('deleted');
@@ -189,17 +197,25 @@ describe('CommentService — applySuggestion', () => {
     expect(resolvePatch.resolvedAt).toBeInstanceOf(Date);
     expect(resolvePatch.resolvedById).toBe('user-1');
 
-    // NOT deleted; broadcast an update, not a deletion.
+    // NOT deleted.
     expect(commentRepo.deleteComment).not.toHaveBeenCalled();
     expect(collaborationGateway.handleYjsEvent).not.toHaveBeenCalledWith(
       'deleteCommentMark',
       expect.anything(),
       expect.anything(),
     );
+    // #496 dedup: resolveComment broadcasts `commentResolved` with the enriched
+    // row; finalize must NOT ALSO emit a redundant `commentUpdated`. So the
+    // thread receives exactly ONE resolve broadcast and no update broadcast.
     expect(wsService.emitCommentEvent).toHaveBeenCalledWith(
       'space-1',
       'page-1',
-      expect.objectContaining({ operation: 'commentUpdated', comment: UPDATED }),
+      expect.objectContaining({ operation: 'commentResolved', comment: UPDATED }),
+    );
+    expect(wsService.emitCommentEvent).not.toHaveBeenCalledWith(
+      'space-1',
+      'page-1',
+      expect.objectContaining({ operation: 'commentUpdated' }),
     );
 
     expect(auditService.log).toHaveBeenCalledWith(
@@ -211,6 +227,36 @@ describe('CommentService — applySuggestion', () => {
     expect(result.outcome).toBe('resolved');
   });
 
+  it('re-entry: already applied+resolved WITH replies → emits commentUpdated (dedup does not over-suppress)', async () => {
+    // suggestionAppliedAt set → idempotent finalize; resolvedAt set → resolveComment
+    // is skipped, so there is NO commentResolved broadcast. The applied-stamp state
+    // must still reach clients via a single commentUpdated.
+    const { service, wsService } = makeService(
+      { applied: false, currentText: 'new text' },
+      true,
+    );
+
+    await service.applySuggestion(
+      suggestionComment({
+        suggestionAppliedAt: new Date(),
+        resolvedAt: new Date(),
+      }),
+      user(),
+    );
+
+    expect(wsService.emitCommentEvent).toHaveBeenCalledWith(
+      'space-1',
+      'page-1',
+      expect.objectContaining({ operation: 'commentUpdated', comment: UPDATED }),
+    );
+    // Nothing resolved this time (already resolved) → no resolve broadcast.
+    expect(wsService.emitCommentEvent).not.toHaveBeenCalledWith(
+      'space-1',
+      'page-1',
+      expect.objectContaining({ operation: 'commentResolved' }),
+    );
+  });
+
   // --- error / rejection branches -----------------------------------------
 
   it('applied=false and currentText differs → ConflictException with currentText in payload', async () => {

apps/server/src/core/comment/comment.service.dismiss-suggestion.spec.ts

@@ -107,11 +107,21 @@ describe('CommentService — dismissSuggestion', () => {
       'page-1',
       expect.objectContaining({ operation: 'commentDeleted', commentId: 'c-1' }),
     );
+    // #496: the row is hard-deleted, so the audit payload must carry the
+    // decision's substance (what was suggested, the anchored text, who authored
+    // it, who decided) — it is the only surviving record.
     expect(auditService.log).toHaveBeenCalledWith(
       expect.objectContaining({
         event: AuditEvent.COMMENT_SUGGESTION_DISMISSED,
         resourceType: AuditResource.COMMENT,
         resourceId: 'c-1',
+        metadata: expect.objectContaining({
+          pageId: 'page-1',
+          suggestedText: 'new text',
+          selection: 'old text',
+          commentAuthor: 'user-1',
+          decidedBy: 'user-1',
+        }),
       }),
     );
     expect(result.outcome).toBe('deleted');

apps/server/src/core/comment/comment.service.resync-anchor.spec.ts

@@ -0,0 +1,126 @@
+import { BadRequestException, ForbiddenException } from '@nestjs/common';
+import { CommentService } from './comment.service';
+
+/**
+ * Coverage for CommentService.resyncSuggestionAnchor (#496): re-anchoring a
+ * suggestion's stored selection (== apply-time expectedText) to the live-doc
+ * substring. The service is built directly with jest-mocked deps (the
+ * @InjectQueue tokens can't be resolved by Test.createTestingModule — see the
+ * sibling specs).
+ */
+describe('CommentService — resyncSuggestionAnchor', () => {
+  const UPDATED = { id: 'c-1', selection: 'new anchor', __updated: true } as any;
+
+  function makeService() {
+    const commentRepo: any = {
+      updateComment: jest.fn(async () => undefined),
+      findById: jest.fn(async () => UPDATED),
+    };
+    const service = new CommentService(
+      commentRepo,
+      {} as any,
+      { emitCommentEvent: jest.fn() } as any,
+      {} as any,
+      { add: jest.fn() } as any,
+      { add: jest.fn() } as any,
+      { log: jest.fn() } as any,
+    );
+    return { service, commentRepo };
+  }
+
+  const suggestion = (over?: Partial<any>): any => ({
+    id: 'c-1',
+    creatorId: 'user-1',
+    parentCommentId: null,
+    selection: 'old anchor',
+    suggestedText: 'new text',
+    suggestionAppliedAt: null,
+    resolvedAt: null,
+    ...over,
+  });
+  const user = (over?: Partial<any>): any => ({ id: 'user-1', ...over });
+
+  it('persists the new selection and returns the enriched comment', async () => {
+    const { service, commentRepo } = makeService();
+
+    const out = await service.resyncSuggestionAnchor(
+      suggestion(),
+      'new anchor',
+      user(),
+    );
+
+    expect(commentRepo.updateComment).toHaveBeenCalledWith(
+      { selection: 'new anchor' },
+      'c-1',
+    );
+    expect(out).toBe(UPDATED);
+  });
+
+  it('is idempotent: no write when the anchor already matches', async () => {
+    const { service, commentRepo } = makeService();
+
+    const out = await service.resyncSuggestionAnchor(
+      suggestion({ selection: 'same' }),
+      'same',
+      user(),
+    );
+
+    expect(commentRepo.updateComment).not.toHaveBeenCalled();
+    expect(out).toEqual(suggestion({ selection: 'same' }));
+  });
+
+  it('rejects a non-author (only the suggestion owner may re-anchor)', async () => {
+    const { service, commentRepo } = makeService();
+    await expect(
+      service.resyncSuggestionAnchor(suggestion(), 'new anchor', user({ id: 'other' })),
+    ).rejects.toBeInstanceOf(ForbiddenException);
+    expect(commentRepo.updateComment).not.toHaveBeenCalled();
+  });
+
+  it('rejects a reply / a comment with no suggestion', async () => {
+    const { service } = makeService();
+    await expect(
+      service.resyncSuggestionAnchor(
+        suggestion({ parentCommentId: 'p-1' }),
+        'x',
+        user(),
+      ),
+    ).rejects.toBeInstanceOf(BadRequestException);
+    await expect(
+      service.resyncSuggestionAnchor(
+        suggestion({ suggestedText: null }),
+        'x',
+        user(),
+      ),
+    ).rejects.toBeInstanceOf(BadRequestException);
+  });
+
+  it('rejects re-anchoring an already applied or resolved suggestion', async () => {
+    const { service } = makeService();
+    await expect(
+      service.resyncSuggestionAnchor(
+        suggestion({ suggestionAppliedAt: new Date() }),
+        'x',
+        user(),
+      ),
+    ).rejects.toBeInstanceOf(BadRequestException);
+    await expect(
+      service.resyncSuggestionAnchor(
+        suggestion({ resolvedAt: new Date() }),
+        'x',
+        user(),
+      ),
+    ).rejects.toBeInstanceOf(BadRequestException);
+  });
+
+  it('rejects a no-op selection equal to the suggested text', async () => {
+    const { service } = makeService();
+    await expect(
+      service.resyncSuggestionAnchor(
+        suggestion({ suggestedText: 'new text' }),
+        'new text',
+        user(),
+      ),
+    ).rejects.toBeInstanceOf(BadRequestException);
+  });
+});

apps/server/src/core/comment/comment.service.ts

@@ -370,6 +370,76 @@ export class CommentService {
     return updatedComment;
   }
 
+  /**
+   * Re-sync a suggestion's stored `selection` (== apply-time expectedText) to the
+   * RAW substring the inline mark actually covers in the LIVE document (#496).
+   *
+   * The MCP client creates the comment from a DEBOUNCED REST snapshot, then
+   * anchors the mark in the live collab doc. When the two disagree (the doc moved
+   * on in the debounce window) the stored selection no longer equals the marked
+   * text, so EVERY apply 409s ("the commented text changed"). After anchoring the
+   * client re-reads the exact marked substring and calls this to store it, making
+   * apply's strict equality hold.
+   *
+   * Only meaningful for an un-settled top-level suggestion authored by the
+   * caller: applying/resolving freezes the anchor, and a reply-carrying thread is
+   * preserved rather than mutated. The new text must still differ from the
+   * suggestion (else "apply" would be a no-op), preserving create()'s invariant.
+   */
+  async resyncSuggestionAnchor(
+    comment: Comment,
+    selection: string,
+    user: User,
+  ): Promise<Comment> {
+    if (comment.creatorId !== user.id) {
+      throw new ForbiddenException(
+        'You can only re-anchor your own suggestion',
+      );
+    }
+    if (comment.parentCommentId) {
+      throw new BadRequestException(
+        'Only a top-level comment can carry a suggested edit',
+      );
+    }
+    if (!comment.suggestedText) {
+      throw new BadRequestException('This comment has no suggested edit');
+    }
+    // A settled suggestion's anchor is frozen: re-anchoring an applied/resolved
+    // thread is meaningless and could resurrect a stale expectedText.
+    if (comment.suggestionAppliedAt || comment.resolvedAt) {
+      throw new BadRequestException(
+        'Cannot re-anchor a suggestion that was already applied or resolved',
+      );
+    }
+    const trimmed = selection.trim();
+    if (trimmed.length === 0) {
+      throw new BadRequestException('The re-anchored selection cannot be empty');
+    }
+    // Same no-op guard as create(): the suggestion must differ from the text it
+    // replaces, or apply becomes indistinguishable from already-applied.
+    if (trimmed === comment.suggestedText.trim()) {
+      throw new BadRequestException(
+        'A suggested edit must differ from the selected text',
+      );
+    }
+
+    // Idempotent: nothing to persist when the anchor already matches.
+    if (comment.selection === selection) {
+      return comment;
+    }
+
+    await this.commentRepo.updateComment({ selection }, comment.id);
+
+    const updatedComment = await this.commentRepo.findById(comment.id, {
+      includeCreator: true,
+      includeResolvedBy: true,
+    });
+
+    // Re-anchoring only corrects stored metadata; it does not change the page
+    // text or the comment body, so no ws broadcast / notification is warranted.
+    return updatedComment;
+  }
+
   /**
    * Apply the suggested edit carried by a top-level inline comment: atomically
    * replace the text under the comment mark in the collaborative document with
@@ -524,7 +594,7 @@ export class CommentService {
         resourceType: AuditResource.COMMENT,
         resourceId: comment.id,
         spaceId: comment.spaceId,
-        metadata: { pageId: comment.pageId },
+        metadata: this.suggestionAuditMetadata(comment, user),
       });
       return { ...updatedComment, outcome: 'resolved' };
     }
@@ -538,7 +608,7 @@ export class CommentService {
       resourceType: AuditResource.COMMENT,
       resourceId: comment.id,
       spaceId: comment.spaceId,
-      metadata: { pageId: comment.pageId },
+      metadata: this.suggestionAuditMetadata(comment, user),
     });
     return settled;
   }
@@ -577,8 +647,10 @@ export class CommentService {
 
       // Auto-resolve the thread. resolveComment handles the resolve mark, its ws
       // broadcast and the resolve notification. Stay defensive on re-entry.
+      let didResolveBroadcast = false;
       if (!comment.resolvedAt) {
         await this.resolveComment(comment, true, user, provenance);
+        didResolveBroadcast = true;
       }
 
       const updatedComment = await this.commentRepo.findById(comment.id, {
@@ -586,18 +658,27 @@ export class CommentService {
         includeResolvedBy: true,
       });
 
-      this.wsService.emitCommentEvent(comment.spaceId, comment.pageId, {
-        operation: 'commentUpdated',
-        pageId: comment.pageId,
-        comment: updatedComment,
-      });
+      // #496 dedup: resolveComment already broadcast `commentResolved` carrying
+      // the fully-enriched row (the applied stamps were persisted above, before
+      // that call, so its re-read reflects them). Emitting `commentUpdated` here
+      // too made the client receive TWO events for one apply. Broadcast the
+      // update ONLY when we did NOT resolve — i.e. the rare re-entry on an
+      // already-resolved thread, where the applied-stamp change still needs a
+      // broadcast and resolveComment did not run.
+      if (!didResolveBroadcast) {
+        this.wsService.emitCommentEvent(comment.spaceId, comment.pageId, {
+          operation: 'commentUpdated',
+          pageId: comment.pageId,
+          comment: updatedComment,
+        });
+      }
 
       this.auditService.log({
         event: AuditEvent.COMMENT_SUGGESTION_APPLIED,
         resourceType: AuditResource.COMMENT,
         resourceId: comment.id,
         spaceId: comment.spaceId,
-        metadata: { pageId: comment.pageId },
+        metadata: this.suggestionAuditMetadata(comment, user),
       });
 
       return { ...updatedComment, outcome: 'resolved' };
@@ -616,7 +697,7 @@ export class CommentService {
       resourceType: AuditResource.COMMENT,
       resourceId: comment.id,
       spaceId: comment.spaceId,
-      metadata: { pageId: comment.pageId },
+      metadata: this.suggestionAuditMetadata(comment, user),
     });
 
     return settled;
@@ -627,14 +708,17 @@ export class CommentService {
    * inline `comment` anchor mark, then ATOMICALLY hard-delete the row only if it
    * is still childless. Shared by the apply/dismiss no-replies branches (#329).
    *
-   * ORDER MATTERS: the anchor mark is removed FIRST and FATALLY (mirrors
-   * applySuggestion, which mutates the doc before writing the DB). The row
-   * delete is irreversible, so if the mark removal fails — including the
-   * COLLAB_DISABLE_REDIS "no live instance" hard-error — we must NOT delete the
-   * row and report success, or the document is left with a permanent orphan
-   * anchor pointing at a comment that no longer exists (the exact data-integrity
-   * bug #329 targets). Let the exception propagate (→ 5xx); the operation is
-   * then repeatable with row + mark still consistent.
+   * ORDER MATTERS (updated #399 → #496): what runs FIRST and FATALLY here is the
+   * mark-removal ENQUEUE (a fast, durable Redis add), NOT the mark op itself.
+   * deleteCommentMark awaits only the enqueue, so a failed add throws BEFORE the
+   * irreversible row delete — the row + mark stay consistent and the operation is
+   * repeatable. The actual anchor strip then runs off the HTTP path in the worker
+   * (idempotent, 3 retries). Only an EXHAUSTED-retries job could leave the doc
+   * with an orphan anchor pointing at a hard-deleted comment (the data-integrity
+   * bug #329 targets); that residual divergence is now self-healed by the
+   * resolve/unresolve mark worker, which strips an orphan mark whenever its
+   * comment row is gone (#496), and it is meanwhile VISIBLE via BullMQ failed-job
+   * metrics rather than a silently-swallowed warn.
    *
    * RACE (#338 F4): the caller read `hasChildren` BEFORE the (slow) mark
    * removal, so a reply can land in that window. `comments.parent_comment_id` is
@@ -732,6 +816,27 @@ export class CommentService {
     return this.generalQueue.add(QueueJob.COMMENT_MARK_UPDATE, jobData);
   }
 
+  /**
+   * Build the audit metadata for a suggestion apply/dismiss decision (#496).
+   * The subject comment is HARD-DELETED on the childless path, so the audit row
+   * is the only surviving record — capture the decision's substance (what was
+   * suggested, the anchored text it replaced, who authored it, who decided)
+   * before the row can vanish. `decidedBy` is the acting user; `commentAuthor`
+   * is the suggestion's creator.
+   */
+  private suggestionAuditMetadata(
+    comment: Comment,
+    user: User,
+  ): Record<string, any> {
+    return {
+      pageId: comment.pageId,
+      suggestedText: comment.suggestedText ?? null,
+      selection: comment.selection ?? null,
+      commentAuthor: comment.creatorId ?? null,
+      decidedBy: user.id,
+    };
+  }
+
   private async queueCommentNotification(
     content: any,
     oldMentionIds: string[],

apps/server/src/core/comment/dto/resync-suggestion-anchor.dto.ts

@@ -0,0 +1,20 @@
+import { IsString, IsUUID, MaxLength, MinLength } from 'class-validator';
+
+/**
+ * #496: after the MCP client anchors a suggestion in the LIVE collab doc, it
+ * re-reads the exact substring under the new mark and syncs it here as the
+ * comment's stored `selection` (== apply-time expectedText). Fixes the perpetual
+ * 409 where expectedText came from a debounced REST snapshot while the mark sat
+ * in the live doc.
+ */
+export class ResyncSuggestionAnchorDto {
+  @IsUUID()
+  commentId: string;
+
+  // The raw substring the mark now covers in the live document. Bounded like the
+  // create-time selection (2000) so a legitimate anchored span is never cut.
+  @IsString()
+  @MinLength(1)
+  @MaxLength(2000)
+  selection: string;
+}

apps/server/src/integrations/audit/audit.module.ts

@@ -1,14 +1,19 @@
 import { Global, Module } from '@nestjs/common';
-import { AUDIT_SERVICE, NoopAuditService } from './audit.service';
+import { AUDIT_SERVICE } from './audit.service';
+import { DatabaseAuditService } from './database-audit.service';
 
+// #496: bind the audit token to a real DB-backed trail (was NoopAuditService,
+// which silently dropped every event). Kysely (@Global DatabaseModule) and
+// ClsService (@Global ClsModule) are both globally available, so this module
+// needs no extra imports.
 @Global()
 @Module({
   providers: [
     {
       provide: AUDIT_SERVICE,
-      useClass: NoopAuditService,
+      useClass: DatabaseAuditService,
     },
   ],
   exports: [AUDIT_SERVICE],
 })
-export class NoopAuditModule {}
+export class AuditModule {}

apps/server/src/integrations/audit/database-audit.service.spec.ts

@@ -0,0 +1,171 @@
+import { DatabaseAuditService } from './database-audit.service';
+import { AUDIT_CONTEXT_KEY } from '../../common/middlewares/audit-context.middleware';
+import { AuditEvent, AuditResource } from '../../common/events/audit-events';
+
+/**
+ * Observable-property coverage for the DB-backed audit trail (#496): every
+ * assertion pins what actually reaches the `audit` table (or that nothing does),
+ * driven through a chainable Kysely mock that captures the inserted rows.
+ */
+describe('DatabaseAuditService', () => {
+  function makeService(clsContext: any) {
+    const inserted: any[] = [];
+    const updated: any[] = [];
+    let failNextInsert = false;
+
+    const db: any = {
+      insertInto: jest.fn(() => ({
+        values: jest.fn((rows: any) => ({
+          execute: jest.fn(async () => {
+            if (failNextInsert) {
+              failNextInsert = false;
+              throw new Error('boom');
+            }
+            inserted.push(rows);
+          }),
+        })),
+      })),
+      updateTable: jest.fn(() => ({
+        set: jest.fn((patch: any) => ({
+          where: jest.fn(() => ({
+            execute: jest.fn(async () => {
+              updated.push(patch);
+            }),
+          })),
+        })),
+      })),
+    };
+
+    const store: Record<string, any> = { [AUDIT_CONTEXT_KEY]: clsContext };
+    const cls: any = {
+      get: jest.fn((key: string) => store[key]),
+      set: jest.fn((key: string, val: any) => {
+        store[key] = val;
+      }),
+    };
+
+    const service = new DatabaseAuditService(db, cls);
+    return {
+      service,
+      inserted,
+      updated,
+      cls,
+      store,
+      failInsert: () => {
+        failNextInsert = true;
+      },
+    };
+  }
+
+  const applyPayload = () => ({
+    event: AuditEvent.COMMENT_SUGGESTION_APPLIED,
+    resourceType: AuditResource.COMMENT,
+    resourceId: 'c-1',
+    spaceId: 'space-1',
+    metadata: { pageId: 'p-1', suggestedText: 'new', decidedBy: 'u-2' },
+  });
+
+  const ctx = (over?: any) => ({
+    workspaceId: 'ws-1',
+    actorId: 'u-2',
+    actorType: 'user',
+    ipAddress: '10.0.0.1',
+    userAgent: 'jest',
+    ...over,
+  });
+
+  it('log() persists a row with the CLS context merged onto the payload', async () => {
+    const { service, inserted } = makeService(ctx());
+    service.log(applyPayload());
+    // log() is fire-and-forget; flush the microtask queue.
+    await Promise.resolve();
+    await Promise.resolve();
+
+    expect(inserted).toHaveLength(1);
+    expect(inserted[0]).toMatchObject({
+      workspaceId: 'ws-1',
+      actorId: 'u-2',
+      actorType: 'user',
+      event: AuditEvent.COMMENT_SUGGESTION_APPLIED,
+      resourceType: AuditResource.COMMENT,
+      resourceId: 'c-1',
+      spaceId: 'space-1',
+      ipAddress: '10.0.0.1',
+      metadata: { pageId: 'p-1', suggestedText: 'new', decidedBy: 'u-2' },
+    });
+  });
+
+  it('log() is a no-op when there is no workspace in scope', async () => {
+    const { service, inserted } = makeService(ctx({ workspaceId: null }));
+    service.log(applyPayload());
+    await Promise.resolve();
+    expect(inserted).toHaveLength(0);
+  });
+
+  it('log() drops EXCLUDED_AUDIT_EVENTS (e.g. comment.created)', async () => {
+    const { service, inserted } = makeService(ctx());
+    service.log({
+      event: AuditEvent.COMMENT_CREATED,
+      resourceType: AuditResource.COMMENT,
+      resourceId: 'c-1',
+      spaceId: 'space-1',
+    });
+    await Promise.resolve();
+    await Promise.resolve();
+    expect(inserted).toHaveLength(0);
+  });
+
+  it('a failed insert never throws to the caller (audit is a side-record)', async () => {
+    const { service, failInsert } = makeService(ctx());
+    failInsert();
+    // Must not reject / throw.
+    expect(() => service.log(applyPayload())).not.toThrow();
+    await Promise.resolve();
+    await Promise.resolve();
+  });
+
+  it('logWithContext() persists with an explicit (non-request) context', async () => {
+    const { service, inserted } = makeService(undefined);
+    service.logWithContext(applyPayload(), ctx({ actorType: 'system' }) as any);
+    await Promise.resolve();
+    await Promise.resolve();
+    expect(inserted).toHaveLength(1);
+    expect(inserted[0].actorType).toBe('system');
+    expect(inserted[0].workspaceId).toBe('ws-1');
+  });
+
+  it('logBatchWithContext() inserts only non-excluded events', async () => {
+    const { service, inserted } = makeService(undefined);
+    service.logBatchWithContext(
+      [
+        applyPayload(),
+        {
+          event: AuditEvent.COMMENT_CREATED,
+          resourceType: AuditResource.COMMENT,
+          resourceId: 'c-2',
+        },
+      ],
+      ctx() as any,
+    );
+    await Promise.resolve();
+    await Promise.resolve();
+    // Batch is a single insert call carrying only the applied event.
+    expect(inserted).toHaveLength(1);
+    expect(inserted[0]).toHaveLength(1);
+    expect(inserted[0][0].event).toBe(AuditEvent.COMMENT_SUGGESTION_APPLIED);
+  });
+
+  it('setActorId / setActorType mutate the ambient CLS context', () => {
+    const { service, store } = makeService(ctx({ actorId: null }));
+    service.setActorId('u-9');
+    service.setActorType('api_key');
+    expect(store[AUDIT_CONTEXT_KEY].actorId).toBe('u-9');
+    expect(store[AUDIT_CONTEXT_KEY].actorType).toBe('api_key');
+  });
+
+  it('updateRetention() writes the workspace retention window', async () => {
+    const { service, updated } = makeService(ctx());
+    await service.updateRetention('ws-1', 30);
+    expect(updated).toEqual([{ auditRetentionDays: 30 }]);
+  });
+});

apps/server/src/integrations/audit/database-audit.service.ts

@@ -0,0 +1,160 @@
+import { Injectable, Logger } from '@nestjs/common';
+import { InjectKysely } from 'nestjs-kysely';
+import { ClsService } from 'nestjs-cls';
+import { KyselyDB } from '@docmost/db/types/kysely.types';
+import {
+  AuditContext,
+  AUDIT_CONTEXT_KEY,
+} from '../../common/middlewares/audit-context.middleware';
+import {
+  AuditLogPayload,
+  ActorType,
+  EXCLUDED_AUDIT_EVENTS,
+} from '../../common/events/audit-events';
+import { AuditLogContext, IAuditService } from './audit.service';
+
+/**
+ * Minimal DB-backed audit trail (#496). Replaces NoopAuditService so that
+ * decision-bearing events — notably comment.suggestion_applied /
+ * comment.suggestion_dismissed, whose subject comment is HARD-DELETED on the
+ * childless path — leave a durable record of who decided what. Without this the
+ * events were emitted (comment.service / *.controller) but swallowed, so an
+ * applied/dismissed suggestion was unrecoverable once the row was gone.
+ *
+ * Rows land in the pre-existing `audit` table (migration 20260228T223532). The
+ * per-request actor/workspace/ip come from the CLS AuditContext populated by
+ * AuditContextMiddleware + AuditActorInterceptor; callers that run OUTSIDE a
+ * request (queue workers, imports) pass an explicit context via
+ * logWithContext / logBatchWithContext.
+ *
+ * Audit is a side-record: a write failure MUST NOT break the originating
+ * request, so every persistence path swallows its error with a warn. Events in
+ * EXCLUDED_AUDIT_EVENTS (high-volume/low-signal) are dropped.
+ */
+@Injectable()
+export class DatabaseAuditService implements IAuditService {
+  private readonly logger = new Logger(DatabaseAuditService.name);
+
+  constructor(
+    @InjectKysely() private readonly db: KyselyDB,
+    private readonly cls: ClsService,
+  ) {}
+
+  /**
+   * Persist a single event using the ambient request-scoped AuditContext. A
+   * no-op when there is no workspace in scope (the table's workspace_id is NOT
+   * NULL) or the event is excluded. Fire-and-forget: the returned promise is not
+   * awaited by hot callers, and its rejection is swallowed here.
+   */
+  log(payload: AuditLogPayload): void {
+    const context = this.cls?.get<AuditContext>(AUDIT_CONTEXT_KEY);
+    if (!context?.workspaceId) {
+      // No workspace in scope — nothing we can attribute the row to. This is
+      // expected for events emitted outside an HTTP request; those callers must
+      // use logWithContext instead.
+      return;
+    }
+    void this.persist(payload, {
+      workspaceId: context.workspaceId,
+      actorId: context.actorId ?? undefined,
+      actorType: context.actorType,
+      ipAddress: context.ipAddress ?? undefined,
+      userAgent: context.userAgent ?? undefined,
+    });
+  }
+
+  /** Persist a single event with an explicit (non-request) context. */
+  logWithContext(payload: AuditLogPayload, context: AuditLogContext): void {
+    if (!context?.workspaceId) return;
+    void this.persist(payload, context);
+  }
+
+  /** Persist a batch of events sharing one explicit context (imports). */
+  logBatchWithContext(
+    payloads: AuditLogPayload[],
+    context: AuditLogContext,
+  ): void {
+    if (!context?.workspaceId || payloads.length === 0) return;
+    const rows = payloads
+      .filter((p) => !EXCLUDED_AUDIT_EVENTS.has(p.event))
+      .map((p) => this.toRow(p, context));
+    if (rows.length === 0) return;
+    this.db
+      .insertInto('audit')
+      .values(rows)
+      .execute()
+      .catch((err: any) =>
+        this.logger.warn(`Failed to persist ${rows.length} audit events: ${err?.message}`),
+      );
+  }
+
+  /** Update the ambient request actor (e.g. after login resolves the user). */
+  setActorId(actorId: string): void {
+    const context = this.cls?.get<AuditContext>(AUDIT_CONTEXT_KEY);
+    if (context) {
+      context.actorId = actorId;
+      this.cls.set(AUDIT_CONTEXT_KEY, context);
+    }
+  }
+
+  /** Update the ambient request actor type (user | system | api_key). */
+  setActorType(actorType: ActorType): void {
+    const context = this.cls?.get<AuditContext>(AUDIT_CONTEXT_KEY);
+    if (context) {
+      context.actorType = actorType;
+      this.cls.set(AUDIT_CONTEXT_KEY, context);
+    }
+  }
+
+  /** Persist a workspace's audit-log retention window (days). */
+  async updateRetention(
+    workspaceId: string,
+    retentionDays: number,
+  ): Promise<void> {
+    try {
+      await this.db
+        .updateTable('workspaces')
+        .set({ auditRetentionDays: retentionDays })
+        .where('id', '=', workspaceId)
+        .execute();
+    } catch (err: any) {
+      this.logger.warn(
+        `Failed to update audit retention for workspace ${workspaceId}: ${err?.message}`,
+      );
+    }
+  }
+
+  private async persist(
+    payload: AuditLogPayload,
+    context: AuditLogContext,
+  ): Promise<void> {
+    if (EXCLUDED_AUDIT_EVENTS.has(payload.event)) return;
+    try {
+      await this.db
+        .insertInto('audit')
+        .values(this.toRow(payload, context))
+        .execute();
+    } catch (err: any) {
+      // Audit is a side-record; never let a failed write surface to the caller.
+      this.logger.warn(
+        `Failed to persist audit event ${payload.event}: ${err?.message}`,
+      );
+    }
+  }
+
+  private toRow(payload: AuditLogPayload, context: AuditLogContext) {
+    return {
+      workspaceId: context.workspaceId,
+      actorId: context.actorId ?? null,
+      actorType: context.actorType ?? 'user',
+      event: payload.event,
+      resourceType: payload.resourceType,
+      resourceId: payload.resourceId ?? null,
+      spaceId: payload.spaceId ?? null,
+      // jsonb columns: node-postgres serializes plain objects to JSON.
+      changes: payload.changes ? (payload.changes as any) : null,
+      metadata: payload.metadata ? (payload.metadata as any) : null,
+      ipAddress: context.ipAddress ?? null,
+    };
+  }
+}

apps/server/src/integrations/mcp/mcp.service.ts

@@ -34,6 +34,8 @@ import {
   isMetricsEnabled,
   observeMcpTool,
   incConnectTimeout,
+  incGetPageCacheHit,
+  incGetPageCacheMiss,
 } from '../metrics/metrics.registry';
 
 // Minimal shape of the embedded MCP HTTP handler exported by @docmost/mcp/http.
@@ -357,6 +359,10 @@ export class McpService implements OnModuleDestroy {
                       observeMcpTool(labels?.tool ?? 'other', value);
                     } else if (name === 'collab_connect_timeouts_total') {
                       incConnectTimeout();
+                    } else if (name === 'mcp_getpage_cache_hits_total') {
+                      incGetPageCacheHit();
+                    } else if (name === 'mcp_getpage_cache_misses_total') {
+                      incGetPageCacheMiss();
                     }
                   }
                 : undefined,

apps/server/src/integrations/metrics/metrics.constants.ts

@@ -25,6 +25,15 @@ export const METRIC_COLLAB_CONNECT_TIMEOUTS_TOTAL =
 export const METRIC_COLLAB_AUTH_DURATION = 'collab_auth_duration_seconds';
 export const METRIC_MCP_TOOL_DURATION = 'mcp_tool_duration_seconds';
 
+// #479 — getPage PM→Markdown conversion cache hit/miss counters. Emitted by the
+// MCP package via its dependency-neutral onMetric sink and routed onto these two
+// prom counters by the mcp.service onMetric callback; a >50% hit-rate is the
+// success signal for the getPage perf work. Same "do not rename" contract.
+export const METRIC_MCP_GETPAGE_CACHE_HITS_TOTAL =
+  'mcp_getpage_cache_hits_total';
+export const METRIC_MCP_GETPAGE_CACHE_MISSES_TOTAL =
+  'mcp_getpage_cache_misses_total';
+
 // Histogram buckets (seconds). Chosen to give useful p50/p95/p99 resolution
 // for typical web/DB latencies without exploding series cardinality.
 export const HTTP_BUCKETS = [

apps/server/src/integrations/metrics/metrics.registry.ts

@@ -24,6 +24,8 @@ import {
   METRIC_DB_QUERY_DURATION,
   METRIC_HTTP_REQUEST_DURATION,
   METRIC_MCP_TOOL_DURATION,
+  METRIC_MCP_GETPAGE_CACHE_HITS_TOTAL,
+  METRIC_MCP_GETPAGE_CACHE_MISSES_TOTAL,
   sizeBucket,
 } from './metrics.constants';
 
@@ -61,6 +63,9 @@ let connectTimeoutsCounter: Counter | null = null;
 let collabConnectHist: Histogram | null = null;
 let collabAuthHist: Histogram | null = null;
 let mcpToolHist: Histogram<'tool'> | null = null;
+// #479 — getPage conversion-cache hit/miss counters.
+let getPageCacheHitsCounter: Counter | null = null;
+let getPageCacheMissesCounter: Counter | null = null;
 
 // #402 — read-on-scrape source for collab_docs_open. The gauge is NEVER
 // inc/dec'd (that drifts under crashes/handoffs); instead its collect() callback
@@ -175,6 +180,18 @@ function init(): void {
     buckets: MCP_TOOL_BUCKETS,
     registers: [registry],
   });
+
+  getPageCacheHitsCounter = new Counter({
+    name: METRIC_MCP_GETPAGE_CACHE_HITS_TOTAL,
+    help: 'Total getPage PM→Markdown conversions served from the cache (skipped)',
+    registers: [registry],
+  });
+
+  getPageCacheMissesCounter = new Counter({
+    name: METRIC_MCP_GETPAGE_CACHE_MISSES_TOTAL,
+    help: 'Total getPage PM→Markdown conversions computed (cache misses)',
+    registers: [registry],
+  });
 }
 
 // Runs once when this module is first imported. Safe to call again (idempotent).
@@ -247,6 +264,14 @@ export function observeCollabAuth(seconds: number): void {
   collabAuthHist?.observe(seconds);
 }
 
+export function incGetPageCacheHit(): void {
+  getPageCacheHitsCounter?.inc();
+}
+
+export function incGetPageCacheMiss(): void {
+  getPageCacheMissesCounter?.inc();
+}
+
 export function observeMcpTool(tool: string, seconds: number): void {
   // `tool` MUST be a bounded, registration-derived MCP tool name (the caller
   // guarantees it comes from the registered-tool set) — never free-form input —

apps/server/src/integrations/metrics/metrics.spec.ts

@@ -10,6 +10,8 @@ import {
   incConnectTimeout,
   incDocLoad,
   incDocUnload,
+  incGetPageCacheHit,
+  incGetPageCacheMiss,
   isMetricsEnabled,
   observeCollabAuth,
   observeCollabConnect,
@@ -197,6 +199,8 @@ describe('metrics helpers are safe no-ops when METRICS_PORT is unset', () => {
       incDocLoad();
       incDocUnload();
       incConnectTimeout();
+      incGetPageCacheHit();
+      incGetPageCacheMiss();
       // Registering a source must not create the gauge or invoke the fn.
       registerDocsOpenSource(() => {
         throw new Error('docsOpenSource must NOT be called when disabled');

apps/server/src/integrations/queue/processors/general-queue.processor.comment-mark.spec.ts

@@ -139,13 +139,19 @@ describe('GeneralQueueProcessor — COMMENT_MARK_UPDATE (#399)', () => {
     );
   });
 
-  it('skips (no throw) when the comment row has vanished', async () => {
+  it('reconcile (#496): comment row vanished → strips the orphan anchor mark', async () => {
     const { proc, collaborationGateway, commentRepo } = makeProc();
     commentRepo.findById.mockResolvedValue(undefined);
 
     await expect(
       proc.process(job({ ...base, action: 'resolve', ts: 1000 })),
     ).resolves.toBeUndefined();
-    expect(collaborationGateway.handleYjsEvent).not.toHaveBeenCalled();
+    // A resolve/unresolve mark job whose comment row is gone leaves a silent
+    // orphan; the worker self-heals by stripping the anchor instead of returning.
+    expect(collaborationGateway.handleYjsEvent).toHaveBeenCalledWith(
+      'deleteCommentMark',
+      'page.page-1',
+      { commentId: 'c-1', user: { id: 'user-1' } },
+    );
   });
 });

apps/server/src/integrations/queue/processors/general-queue.processor.ts

@@ -95,7 +95,8 @@ export class GeneralQueueProcessor
    * #399: apply a comment's inline-mark mirror in the collab Y.Doc, off the HTTP
    * critical path. Runs the SAME gateway path the synchronous comment.service
    * code used (byte-identical mark op):
-   *   - resolve / unresolve → resolveCommentMark (flip the `resolved` attribute);
+   *   - resolve / unresolve → resolveCommentMark (flip the `resolved` attribute),
+   *     OR strip an orphan anchor when the comment row has vanished (#496);
    *   - delete → deleteCommentMark (strip the ephemeral-suggestion anchor #329).
    * The op is idempotent, so a BullMQ retry is safe. Throwing propagates to
    * WorkerHost → the job is retried and, on exhaustion, surfaces in failed-job
@@ -133,7 +134,18 @@ export class GeneralQueueProcessor
     // of this resolve), skip it rather than flip the mark to a stale state.
     const comment = await this.commentRepo.findById(commentId);
     if (!comment) {
-      // The comment vanished (e.g. hard-deleted) → nothing left to mirror.
+      // #496 reconcile: the comment row is GONE (e.g. an ephemeral apply/dismiss
+      // hard-deleted it while this resolve/unresolve mark job sat in the queue),
+      // but its inline anchor may still live in the doc — a silent orphan mark
+      // pointing at a comment that no longer exists. Self-heal by stripping it
+      // instead of just returning: this closes the divergence the fire-and-forget
+      // resolve/unresolve enqueue (comment.service resolveComment) could leave.
+      // Idempotent — deleteCommentMark on an already-absent mark is a no-op.
+      await this.getCollaborationGateway().handleYjsEvent(
+        'deleteCommentMark',
+        documentName,
+        { commentId, user },
+      );
       return;
     }
     const wantResolved = action === 'resolve';

packages/mcp/src/client/comments.ts

@@ -450,6 +450,12 @@ export function CommentsMixin<TBase extends GConstructor<DocmostClientContext>>(
     // can surface the closest-block / spans-multiple-blocks hint built from the
     // LIVE document (the pre-check page is not in scope there).
     let liveNotFoundError: Error | null = null;
+    // #496: the RAW substring the mark actually covers in the LIVE doc. The
+    // stored selection (payload.selection) came from a DEBOUNCED REST snapshot,
+    // which can differ from the live doc — and apply compares the marked live
+    // text to the stored selection strictly, so a stale snapshot 409s on EVERY
+    // apply. Captured here (same doc version the mark is set in) and synced below.
+    let liveAnchoredSelection: string | null = null;
     try {
       const collabToken = await this.getCollabTokenWithReauth();
       // Open the collab doc by the canonical UUID, never the slugId (#260). The
@@ -489,6 +495,12 @@ export function CommentsMixin<TBase extends GConstructor<DocmostClientContext>>(
           }
           if (applyAnchorInDoc(doc, selection as string, newCommentId)) {
             anchored = true;
+            // For a suggestion, re-read the exact substring now under the mark
+            // (the mark is an attribute, so it does not change the raw text) to
+            // sync as the stored expectedText after the mutation resolves.
+            if (hasSuggestion) {
+              liveAnchoredSelection = getAnchoredText(doc, selection as string);
+            }
             return doc;
           }
           // Selection text not found in the LIVE document: abort the write. The
@@ -527,6 +539,36 @@ export function CommentsMixin<TBase extends GConstructor<DocmostClientContext>>(
       );
     }
 
+    // #496: sync the stored selection (== apply-time expectedText) to the RAW
+    // substring the mark actually covers in the LIVE doc when it diverged from
+    // the debounced REST snapshot we stored at create time. Without this, apply
+    // strictly compares the marked live text to a stale stored selection and
+    // 409s every time. Best-effort: the comment is already correctly anchored, so
+    // a resync failure must NOT roll it back — it only risks a later apply 409,
+    // which we surface as a soft warning.
+    if (
+      hasSuggestion &&
+      liveAnchoredSelection != null &&
+      liveAnchoredSelection !== payload.selection
+    ) {
+      try {
+        await this.client.post("/comments/resync-suggestion-anchor", {
+          commentId: newCommentId,
+          selection: liveAnchoredSelection,
+        });
+        // Reflect the corrected anchor in the returned comment.
+        if (result.data) result.data.selection = liveAnchoredSelection;
+      } catch (e) {
+        if (process.env.DEBUG) {
+          console.error("Failed to resync suggestion anchor:", e);
+        }
+        result.warning =
+          "The suggestion was anchored, but its stored selection could not be " +
+          "synced to the live document; applying it may report a conflict if the " +
+          "text changed. Re-create the suggestion if Apply fails.";
+      }
+    }
+
     // Soft warning (like editPageText): the selection only matched after
     // stripping markdown, so the caller likely quoted a styled fragment.
     if (anchorNormalized) {

packages/mcp/src/client/context.ts

@@ -23,6 +23,7 @@ import { acquireCollabSession } from "../lib/collab-session.js";
 import { withPageLock, isUuid } from "../lib/page-lock.js";
 import { getCollabToken, performLogin } from "../lib/auth-utils.js";
 import { formatDocmostAxiosError } from "./errors.js";
+import { GetPageConversionCache } from "./getpage-cache.js";
 
 // A generic mixin base constructor (issue #450). Each domain mixin is a factory
 // `<T extends GConstructor<DocmostClientContext>>(Base: T) => class extends Base`
@@ -159,6 +160,13 @@ export abstract class DocmostClientContext {
   // bypassed on a forced refresh (the 401/403 reauth path). null = no token yet.
   protected collabTokenCache: { token: string; mintedAt: number } | null = null;
 
+  // Content-addressed conversion cache for getPage (issue #479). Keyed on
+  // (canonical pageId, updatedAt, optionsHash) -> the converted Markdown, so a
+  // re-read of an UNCHANGED page skips the expensive convertProseMirrorToMarkdown
+  // tree walk. Per-instance (a DocmostClient is built per user / per chat), so a
+  // cached conversion can never leak across identities. See getpage-cache.ts.
+  protected getPageCache = new GetPageConversionCache();
+
   // Two construction forms:
   //  - new DocmostClient(config)                  // discriminated union (current)
   //  - new DocmostClient(baseURL, email, password) // legacy positional creds

packages/mcp/src/client/getpage-cache.ts

@@ -0,0 +1,147 @@
+// Content-addressed LRU cache for the PM->Markdown conversion in getPage
+// (issue #479). getPage is the dominant agent op (812 calls / 2h, p95 840ms);
+// the bulk of its cost is convertProseMirrorToMarkdown — a full ProseMirror-tree
+// walk over the page content (hundreds of KB of JSON on large pages) run on
+// EVERY read. Since agents re-read far more than they write (812 reads vs 28
+// writes in the sample), most conversions re-produce the SAME markdown from
+// UNCHANGED content. This cache skips the recomputation on a hit.
+//
+// KEY = (pageId, updatedAt, optionsHash):
+//   - pageId: the page's CANONICAL UUID (resultData.id), not the agent-supplied
+//     slugId — so a slugId read and a UUID read of the same page share one entry.
+//   - updatedAt: comes from the SAME /pages/info response as `content`, so the
+//     two are mutually consistent; a changed page yields a new updatedAt -> a new
+//     key -> automatic, precise invalidation (no stale markdown is ever served).
+//   - optionsHash: a stable hash of the conversion options. getPage passes
+//     `{dropResolvedCommentAnchors:true}` while exportPageMarkdown passes the
+//     defaults (#328) — DIFFERENT output for the same content, so the options
+//     MUST be part of the key or a hit would serve the wrong variant.
+//
+// BOUNDS: evict the LEAST-recently-used entry when EITHER the entry count OR the
+// total stored bytes would exceed its cap. Large pages are hundreds of KB, so a
+// byte cap (not just a count cap) is what actually bounds memory. A Map iterates
+// in insertion order, so the first key is the LRU entry; a hit re-inserts its key
+// to move it to the most-recently-used end.
+//
+// This module is dependency-neutral (no axios/client/prom-client): a plain class
+// the shared client context owns one instance of, so the cache persists across
+// getPage calls on a single DocmostClient instance (built per user / per chat).
+
+/** A stable, order-insensitive hash of the conversion options object. */
+export function hashConvertOptions(options: unknown): string {
+  // JSON.stringify with SORTED keys makes the hash independent of key order, so
+  // {a:1,b:2} and {b:2,a:1} collapse to one entry. undefined/null options -> a
+  // fixed empty-object key, matching a caller that passes no options at all.
+  if (options === undefined || options === null) return "{}";
+  return stableStringify(options);
+}
+
+function stableStringify(value: any): string {
+  if (value === null || typeof value !== "object") return JSON.stringify(value);
+  if (Array.isArray(value)) return `[${value.map(stableStringify).join(",")}]`;
+  const keys = Object.keys(value).sort();
+  return `{${keys
+    .map((k) => `${JSON.stringify(k)}:${stableStringify(value[k])}`)
+    .join(",")}}`;
+}
+
+interface CacheEntry {
+  markdown: string;
+  bytes: number;
+}
+
+export interface GetPageCacheOptions {
+  /** Max number of entries before LRU eviction. Default 50. */
+  maxEntries?: number;
+  /** Max total stored bytes before LRU eviction. Default 10 MB. */
+  maxBytes?: number;
+}
+
+export class GetPageConversionCache {
+  private readonly maxEntries: number;
+  private readonly maxBytes: number;
+  // Insertion-ordered: the FIRST key is the least-recently-used entry.
+  private readonly map = new Map<string, CacheEntry>();
+  private totalBytes = 0;
+
+  constructor(opts: GetPageCacheOptions = {}) {
+    // A non-positive/NaN cap is treated as "use the default", never as an
+    // unbounded (or always-empty) cache — a silently unbounded cache would leak
+    // memory, and an always-empty one would defeat the whole optimization.
+    this.maxEntries =
+      Number.isFinite(opts.maxEntries) && (opts.maxEntries as number) > 0
+        ? Math.floor(opts.maxEntries as number)
+        : 50;
+    this.maxBytes =
+      Number.isFinite(opts.maxBytes) && (opts.maxBytes as number) > 0
+        ? Math.floor(opts.maxBytes as number)
+        : 10 * 1024 * 1024;
+  }
+
+  /** Compose the content-addressed key from its three parts. */
+  static key(pageId: string, updatedAt: string, optionsHash: string): string {
+    // A space separates the parts so no combination of values can collide by
+    // concatenation: a canonical UUID and an ISO updatedAt never contain a
+    // space, so the boundaries between the three parts are unambiguous.
+    return `${pageId} ${updatedAt} ${optionsHash}`;
+  }
+
+  /**
+   * Return the cached markdown for `key`, or undefined on a miss. A hit moves
+   * the entry to the most-recently-used end (delete + re-set) so the LRU order
+   * reflects real access, not just insertion.
+   */
+  get(key: string): string | undefined {
+    const entry = this.map.get(key);
+    if (entry === undefined) return undefined;
+    this.map.delete(key);
+    this.map.set(key, entry);
+    return entry.markdown;
+  }
+
+  /**
+   * Store `markdown` under `key`, then evict LRU entries until BOTH caps hold.
+   * Re-storing an existing key refreshes its value and recency (its old bytes
+   * are subtracted first, so totalBytes stays exact).
+   */
+  set(key: string, markdown: string): void {
+    // Byte size of the stored string (UTF-8). A single entry larger than the
+    // whole byte cap is still stored (so getPage always gets a hit next time),
+    // then the eviction loop below simply cannot shrink below it — accepted:
+    // one oversized page is bounded by the page itself, not a cache leak.
+    const bytes = Buffer.byteLength(markdown, "utf8");
+    const existing = this.map.get(key);
+    if (existing !== undefined) {
+      this.totalBytes -= existing.bytes;
+      this.map.delete(key);
+    }
+    this.map.set(key, { markdown, bytes });
+    this.totalBytes += bytes;
+    this.evict();
+  }
+
+  /** Evict the LRU entry until both the count and byte caps are satisfied. */
+  private evict(): void {
+    while (
+      this.map.size > this.maxEntries ||
+      (this.totalBytes > this.maxBytes && this.map.size > 1)
+    ) {
+      // The first key in insertion order is the least-recently-used.
+      const oldest = this.map.keys().next().value as string | undefined;
+      if (oldest === undefined) break;
+      const entry = this.map.get(oldest);
+      this.map.delete(oldest);
+      if (entry) this.totalBytes -= entry.bytes;
+    }
+  }
+
+  /** Current entry count (test/introspection). */
+  get size(): number {
+    return this.map.size;
+  }
+
+  /** Current total stored bytes (test/introspection). */
+  get bytes(): number {
+    return this.totalBytes;
+  }
+}

packages/mcp/src/client/read.ts

@@ -10,7 +10,14 @@ import {
   filterComment,
   filterSearchResult,
 } from "../lib/filters.js";
-import { convertProseMirrorToMarkdown } from "../lib/markdown-converter.js";
+import {
+  convertProseMirrorToMarkdown,
+  type ConvertProseMirrorToMarkdownOptions,
+} from "../lib/markdown-converter.js";
+import {
+  GetPageConversionCache,
+  hashConvertOptions,
+} from "./getpage-cache.js";
 import {
   collectInternalFileNodes,
   normalizeFileUrl,
@@ -395,6 +402,19 @@ export function ReadMixin<TBase extends GConstructor<DocmostClientContext>>(Base
 
   /** Raw page info including the ProseMirror JSON content and slugId. */
 
+  /**
+   * Overridable seam over convertProseMirrorToMarkdown (issue #479). Production
+   * just delegates; it exists as a method so a unit test can spy on it and
+   * assert the conversion is genuinely SKIPPED on a getPage cache HIT (the whole
+   * point of the cache) — an ESM named import cannot be intercepted otherwise.
+   */
+  protected convertPageMarkdown(
+    content: any,
+    options: ConvertProseMirrorToMarkdownOptions,
+  ): string {
+    return convertProseMirrorToMarkdown(content, options);
+  }
+
   async getPage(pageId: string) {
     await this.ensureAuthenticated();
     const resultData = await this.getPageRaw(pageId);
@@ -403,13 +423,55 @@ export function ReadMixin<TBase extends GConstructor<DocmostClientContext>>(Base
     // discussions. Active anchors are kept. (The lossless exportPageMarkdown
     // round-trip deliberately does NOT pass this flag — resolved anchors there
     // must be preserved.)
-    let content = resultData.content
-      ? convertProseMirrorToMarkdown(resultData.content, {
-          dropResolvedCommentAnchors: true,
-        })
-      : "";
+    //
+    // Content-addressed conversion cache (issue #479): the PM->Markdown walk is
+    // the dominant cost of this hot read op. Key on the page's canonical UUID +
+    // updatedAt (both from THIS /pages/info response, so mutually consistent) +
+    // a hash of the conversion options. A hit returns the cached markdown and
+    // skips the walk; a miss converts and stores. The cached value is the
+    // conversion output BEFORE the {{SUBPAGES}} substitution below, which uses
+    // live subpage data and stays outside the cache — so the final result is
+    // byte-identical to the uncached path.
+    const convertOptions = { dropResolvedCommentAnchors: true };
+    let content = "";
+    if (resultData.content) {
+      // Only cache when we have a stable identity+version for the key. Both come
+      // from the same response; if either is missing (unexpected server shape),
+      // fall back to converting uncached rather than keying on a partial tuple.
+      const cacheable =
+        typeof resultData.id === "string" &&
+        typeof resultData.updatedAt === "string";
+      const cacheKey = cacheable
+        ? GetPageConversionCache.key(
+            resultData.id,
+            resultData.updatedAt,
+            hashConvertOptions(convertOptions),
+          )
+        : null;
 
-    // Always fetch subpages to provide context to the agent
+      const cached = cacheKey ? this.getPageCache.get(cacheKey) : undefined;
+      if (cached !== undefined) {
+        content = cached;
+        this.onMetricFn?.("mcp_getpage_cache_hits_total", 1);
+      } else {
+        // Goes through the convertPageMarkdown seam (not the raw import) so a
+        // test can assert the conversion is SKIPPED on a hit (issue #479 F2).
+        content = this.convertPageMarkdown(resultData.content, convertOptions);
+        if (cacheKey) this.getPageCache.set(cacheKey, content);
+        // A non-cacheable page (missing id/updatedAt) is still a genuine
+        // conversion, so it counts as a miss for an honest hit-rate.
+        this.onMetricFn?.("mcp_getpage_cache_misses_total", 1);
+      }
+    }
+
+    // Always fetch subpages to provide context to the agent.
+    //
+    // NOT parallelizable with the page fetch (issue #479 asked to check): the
+    // sidebar-pages endpoint REQUIRES spaceId in its POST body, and spaceId is
+    // only known FROM this page fetch's response (resolvePageId yields the UUID
+    // but never the spaceId). So `Promise.all([pageFetch, subpagesFetch])` would
+    // have to invent a spaceId it does not have — the two calls are inherently
+    // sequential. Correctness wins; the conversion cache above is the real speedup.
     let subpages: any[] = [];
     try {
       // `pageId` may be a slugId, but the sidebar-pages endpoint requires the

packages/mcp/test/mock/create-comment.test.mjs

@@ -553,6 +553,189 @@ test("suggestedText: the stored selection is the doc's RAW typographic substring
   assert.equal(createPayload.suggestedText, "goodbye");
 });
 
+// -----------------------------------------------------------------------------
+// 8b) #496: the DEBOUNCED REST snapshot (pages/info) DIFFERS from the LIVE collab
+//     doc (mutatePage) — the doc moved on in the debounce window. The stored
+//     selection is captured from the snapshot at create time, but the mark is set
+//     in the live doc, so apply would 409 forever. After anchoring, the client
+//     re-reads the RAW substring under the mark from the LIVE doc and POSTs it to
+//     /comments/resync-suggestion-anchor so the stored expectedText matches.
+// -----------------------------------------------------------------------------
+test("suggestion: re-syncs the stored selection to the LIVE doc substring when the REST snapshot lagged", async () => {
+  let createPayload = null;
+  let resyncPayload = null;
+
+  const { baseURL } = await spawn(async (req, res) => {
+    const raw = await readBody(req);
+    if (req.url === "/api/auth/login") {
+      sendJson(res, 200, { success: true }, {
+        "Set-Cookie": "authToken=t; Path=/; HttpOnly",
+      });
+      return;
+    }
+    if (req.url === "/api/pages/info") {
+      // DEBOUNCED snapshot: ASCII quotes (stale — the live doc has since been
+      // typographically corrected).
+      sendJson(res, 200, {
+        data: {
+          id: "33333333-3333-3333-3333-333333333333",
+          content: {
+            type: "doc",
+            content: [
+              {
+                type: "paragraph",
+                content: [{ type: "text", text: 'he said "hello" loudly' }],
+              },
+            ],
+          },
+        },
+      });
+      return;
+    }
+    if (req.url === "/api/comments/create") {
+      createPayload = JSON.parse(raw);
+      sendJson(res, 200, {
+        data: {
+          id: "cmt-resync-1",
+          content: createPayload.content,
+          selection: createPayload.selection,
+          suggestedText: createPayload.suggestedText,
+          type: createPayload.type,
+        },
+      });
+      return;
+    }
+    if (req.url === "/api/comments/resync-suggestion-anchor") {
+      resyncPayload = JSON.parse(raw);
+      sendJson(res, 200, { data: { id: "cmt-resync-1" } });
+      return;
+    }
+    sendJson(res, 404, { message: "not found" });
+  });
+
+  class TestClient extends DocmostClient {
+    async getCollabTokenWithReauth() {
+      return "collab-token";
+    }
+    async resolvePageId() {
+      return "33333333-3333-3333-3333-333333333333";
+    }
+    async mutatePage(pageId, collabToken, apiUrl, transform) {
+      // LIVE doc: SMART quotes (what the mark is actually set over).
+      const doc = {
+        type: "doc",
+        content: [
+          {
+            type: "paragraph",
+            content: [{ type: "text", text: "he said “hello” loudly" }],
+          },
+        ],
+      };
+      const out = transform(doc);
+      return { doc: out, verify: { ok: true } };
+    }
+  }
+
+  const client = new TestClient(baseURL, "user@example.com", "pw");
+
+  const result = await client.createComment(
+    "33333333-3333-3333-3333-333333333333",
+    "please change",
+    "inline",
+    '"hello"', // ASCII quotes
+    undefined,
+    "goodbye",
+  );
+
+  assert.equal(result.success, true);
+  assert.equal(result.anchored, true);
+  // Create stored the STALE snapshot substring (ASCII quotes).
+  assert.equal(createPayload.selection, '"hello"');
+  // …then the client re-synced to the LIVE marked substring (smart quotes).
+  assert.ok(resyncPayload, "/comments/resync-suggestion-anchor must be called");
+  assert.equal(resyncPayload.commentId, "cmt-resync-1");
+  assert.equal(resyncPayload.selection, "“hello”");
+  // The returned comment reflects the corrected anchor.
+  assert.equal(result.data.selection, "“hello”");
+});
+
+// -----------------------------------------------------------------------------
+// 8c) #496: when the REST snapshot and the LIVE doc AGREE, no resync round-trip
+//     is made (the common case must stay a single write).
+// -----------------------------------------------------------------------------
+test("suggestion: no resync call when the snapshot already matches the live doc", async () => {
+  let resyncCalls = 0;
+
+  const { baseURL } = await spawn(async (req, res) => {
+    const raw = await readBody(req);
+    if (req.url === "/api/auth/login") {
+      sendJson(res, 200, { success: true }, {
+        "Set-Cookie": "authToken=t; Path=/; HttpOnly",
+      });
+      return;
+    }
+    if (req.url === "/api/pages/info") {
+      sendJson(res, 200, {
+        data: {
+          id: "44444444-4444-4444-4444-444444444444",
+          content: {
+            type: "doc",
+            content: [
+              { type: "paragraph", content: [{ type: "text", text: "Hello brave world" }] },
+            ],
+          },
+        },
+      });
+      return;
+    }
+    if (req.url === "/api/comments/create") {
+      const p = JSON.parse(raw);
+      sendJson(res, 200, {
+        data: { id: "cmt-nosync-1", content: p.content, selection: p.selection, suggestedText: p.suggestedText, type: p.type },
+      });
+      return;
+    }
+    if (req.url === "/api/comments/resync-suggestion-anchor") {
+      resyncCalls++;
+      sendJson(res, 200, { data: {} });
+      return;
+    }
+    sendJson(res, 404, { message: "not found" });
+  });
+
+  class TestClient extends DocmostClient {
+    async getCollabTokenWithReauth() {
+      return "collab-token";
+    }
+    async resolvePageId() {
+      return "44444444-4444-4444-4444-444444444444";
+    }
+    async mutatePage(pageId, collabToken, apiUrl, transform) {
+      const doc = {
+        type: "doc",
+        content: [
+          { type: "paragraph", content: [{ type: "text", text: "Hello brave world" }] },
+        ],
+      };
+      const out = transform(doc);
+      return { doc: out, verify: { ok: true } };
+    }
+  }
+
+  const client = new TestClient(baseURL, "user@example.com", "pw");
+  const result = await client.createComment(
+    "44444444-4444-4444-4444-444444444444",
+    "rename",
+    "inline",
+    "brave",
+    undefined,
+    "bold",
+  );
+
+  assert.equal(result.anchored, true);
+  assert.equal(resyncCalls, 0, "matching snapshot must NOT trigger a resync round-trip");
+});
+
 // -----------------------------------------------------------------------------
 // 8) #408: a not-found selection error QUOTES the closest block text so the
 //    model can self-correct instead of blind-retrying.

packages/mcp/test/mock/getpage-conversion-cache.test.mjs

@@ -0,0 +1,251 @@
+// Mock-HTTP integration tests for the getPage conversion cache (issue #479).
+// A local http.createServer stands in for Docmost (same harness style as
+// get-page-context.test.mjs) so everything is deterministic and offline.
+//
+// Verifies end-to-end through the real client that:
+//   - the FIRST getPage of a page is a MISS (mcp_getpage_cache_misses_total)
+//     and converts the content (the server's convert-representative counter);
+//   - a SECOND getPage of the same (pageId, updatedAt) is a HIT
+//     (mcp_getpage_cache_hits_total) and returns BYTE-IDENTICAL output while
+//     skipping the conversion;
+//   - a changed updatedAt is a fresh key -> MISS again;
+//   - the returned shape still resolves page + subpages.
+import { test, after, mock } from "node:test";
+import assert from "node:assert/strict";
+import http from "node:http";
+import { DocmostClient } from "../../build/client.js";
+
+function readBody(req) {
+  return new Promise((resolve) => {
+    let raw = "";
+    req.on("data", (chunk) => (raw += chunk));
+    req.on("end", () => resolve(raw));
+  });
+}
+
+function sendJson(res, status, obj, extraHeaders = {}) {
+  res.writeHead(status, { "Content-Type": "application/json", ...extraHeaders });
+  res.end(JSON.stringify(obj));
+}
+
+const openServers = [];
+after(async () => {
+  await Promise.all(openServers.map((s) => new Promise((r) => s.close(r))));
+});
+
+const PAGE_UUID = "00000000-0000-4000-8000-000000000010";
+const SPACE_UUID = "00000000-0000-4000-8000-0000000000aa";
+const CHILD_UUID = "00000000-0000-4000-8000-0000000000bb";
+
+// A small ProseMirror doc so the converter produces non-trivial markdown.
+function makeDoc(text) {
+  return {
+    type: "doc",
+    content: [
+      {
+        type: "paragraph",
+        content: [{ type: "text", text }],
+      },
+    ],
+  };
+}
+
+// state.info counts /pages/info hits; state.updatedAt / state.text drive the
+// content+version returned; state.sidebar counts sidebar-pages hits;
+// state.subpages (when set) drives the child list the sidebar endpoint returns,
+// so a test can vary the live subpages across two reads of the same page.
+function spawn(state) {
+  return new Promise((resolve) => {
+    const server = http.createServer(async (req, res) => {
+      await readBody(req);
+      if (req.url === "/api/auth/login") {
+        return sendJson(res, 200, { success: true }, {
+          "Set-Cookie": "authToken=t; Path=/; HttpOnly",
+        });
+      }
+      if (req.url === "/api/pages/info") {
+        state.info++;
+        return sendJson(res, 200, {
+          success: true,
+          data: {
+            id: PAGE_UUID,
+            slugId: "slug123456",
+            title: "Cached Page",
+            parentPageId: null,
+            spaceId: SPACE_UUID,
+            updatedAt: state.updatedAt,
+            content: makeDoc(state.text),
+          },
+        });
+      }
+      if (req.url === "/api/pages/sidebar-pages") {
+        state.sidebar++;
+        const items = state.subpages ?? [
+          { id: CHILD_UUID, title: "Child", hasChildren: false },
+        ];
+        return sendJson(res, 200, {
+          success: true,
+          data: {
+            items,
+            meta: { hasNextPage: false, nextCursor: null },
+          },
+        });
+      }
+      return sendJson(res, 404, { message: "not found" });
+    });
+    server.listen(0, "127.0.0.1", () => {
+      openServers.push(server);
+      resolve(`http://127.0.0.1:${server.address().port}/api`);
+    });
+  });
+}
+
+function makeClient(baseURL, metrics) {
+  return new DocmostClient({
+    apiUrl: baseURL,
+    getToken: async () => "access",
+    onMetric: (name, value) => {
+      metrics[name] = (metrics[name] ?? 0) + value;
+    },
+  });
+}
+
+test("first read MISS, second read HIT with byte-identical output; convert runs once", async () => {
+  const state = { info: 0, sidebar: 0, updatedAt: "2026-01-01T00:00:00Z", text: "Hello world" };
+  const baseURL = await spawn(state);
+  const metrics = {};
+  const client = makeClient(baseURL, metrics);
+
+  const first = await client.getPage(PAGE_UUID);
+  assert.equal(metrics["mcp_getpage_cache_misses_total"], 1, "first read is a miss");
+  assert.equal(metrics["mcp_getpage_cache_hits_total"] ?? 0, 0, "no hit yet");
+
+  const second = await client.getPage(PAGE_UUID);
+  assert.equal(metrics["mcp_getpage_cache_hits_total"], 1, "second read is a hit");
+  assert.equal(metrics["mcp_getpage_cache_misses_total"], 1, "still one miss");
+
+  // BYTE-IDENTICAL: the cache only skips recomputation, never changes output.
+  assert.deepEqual(second, first, "cached result is identical to the uncached one");
+  assert.equal(
+    JSON.stringify(second),
+    JSON.stringify(first),
+    "serialized output is byte-identical",
+  );
+
+  // The page fetch + subpages fetch still happen every call (only the CPU
+  // conversion is cached); both reads hit /pages/info and sidebar-pages.
+  assert.equal(state.info, 2, "both reads still fetch /pages/info");
+  assert.equal(state.sidebar, 2, "both reads still fetch subpages");
+
+  // Shape sanity: content present, subpages resolved.
+  assert.equal(typeof second.data.content, "string");
+  assert.ok(second.data.content.includes("Hello world"));
+  assert.deepEqual(second.data.subpages, [{ id: CHILD_UUID, title: "Child" }]);
+});
+
+test("a changed updatedAt is a fresh key -> MISS again, with the NEW content", async () => {
+  const state = { info: 0, sidebar: 0, updatedAt: "2026-01-01T00:00:00Z", text: "Version one" };
+  const baseURL = await spawn(state);
+  const metrics = {};
+  const client = makeClient(baseURL, metrics);
+
+  const a = await client.getPage(PAGE_UUID); // miss
+  const b = await client.getPage(PAGE_UUID); // hit
+  assert.equal(metrics["mcp_getpage_cache_misses_total"], 1);
+  assert.equal(metrics["mcp_getpage_cache_hits_total"], 1);
+  assert.ok(a.data.content.includes("Version one"));
+
+  // The page changes: new updatedAt AND new content.
+  state.updatedAt = "2026-02-02T00:00:00Z";
+  state.text = "Version two";
+
+  const c = await client.getPage(PAGE_UUID); // miss on the new key
+  assert.equal(metrics["mcp_getpage_cache_misses_total"], 2, "changed version -> miss");
+  assert.equal(metrics["mcp_getpage_cache_hits_total"], 1, "no stale hit");
+  assert.ok(c.data.content.includes("Version two"), "the NEW content is served");
+  assert.ok(!c.data.content.includes("Version one"), "no stale markdown");
+
+  const d = await client.getPage(PAGE_UUID); // hit on the new key
+  assert.equal(metrics["mcp_getpage_cache_hits_total"], 2, "the new snapshot caches too");
+});
+
+test("a slugId read and a UUID read of the same page share one cache entry", async () => {
+  // resolvePageId maps the slugId -> UUID via /pages/info; the cache keys on the
+  // canonical UUID (resultData.id), so both inputs land on the same entry.
+  const state = { info: 0, sidebar: 0, updatedAt: "2026-01-01T00:00:00Z", text: "Shared" };
+  const baseURL = await spawn(state);
+  const metrics = {};
+  const client = makeClient(baseURL, metrics);
+
+  await client.getPage(PAGE_UUID);          // miss (keyed on UUID)
+  await client.getPage("slug123456");        // the server returns the same id -> HIT
+  assert.equal(metrics["mcp_getpage_cache_misses_total"], 1, "one conversion total");
+  assert.equal(metrics["mcp_getpage_cache_hits_total"], 1, "slugId read hits the UUID entry");
+});
+
+test("on a conversion HIT, the {{SUBPAGES}} block reflects the LIVE subpages, not the cached ones", async () => {
+  // The whole byte-identity guarantee: the cache stores the conversion output
+  // BEFORE the {{SUBPAGES}} substitution, so a re-read of an UNCHANGED page still
+  // splices the FRESH subpage list. The page body itself contains {{SUBPAGES}}
+  // (converts to a literal placeholder); getPage replaces it with the live list.
+  const CHILD_A = "00000000-0000-4000-8000-0000000000a1";
+  const CHILD_B = "00000000-0000-4000-8000-0000000000b2";
+  const state = {
+    info: 0,
+    sidebar: 0,
+    updatedAt: "2026-01-01T00:00:00Z", // FIXED across both reads -> conversion cache HIT
+    text: "Body before {{SUBPAGES}} body after",
+    subpages: [{ id: CHILD_A, title: "Alpha", hasChildren: false }],
+  };
+  const baseURL = await spawn(state);
+  const metrics = {};
+  const client = makeClient(baseURL, metrics);
+
+  // Read 1: MISS (converts). The substitution runs with list A.
+  const first = await client.getPage(PAGE_UUID);
+  assert.equal(metrics["mcp_getpage_cache_misses_total"], 1, "first read converts (miss)");
+  assert.ok(first.data.content.includes("[Alpha](page:" + CHILD_A + ")"), "list A spliced in");
+  assert.ok(!first.data.content.includes("{{SUBPAGES}}"), "placeholder consumed");
+
+  // The subpages change while the PAGE CONTENT/updatedAt do NOT: same conversion
+  // cache key -> a HIT that skips the CPU walk, but the live substitution must
+  // still run on the NEW list B.
+  state.subpages = [{ id: CHILD_B, title: "Beta", hasChildren: false }];
+
+  const second = await client.getPage(PAGE_UUID);
+  assert.equal(metrics["mcp_getpage_cache_hits_total"], 1, "second read is a conversion HIT");
+  assert.equal(metrics["mcp_getpage_cache_misses_total"], 1, "no second conversion");
+
+  // The cache did NOT freeze the subpages block: list B is present, list A gone.
+  assert.ok(second.data.content.includes("[Beta](page:" + CHILD_B + ")"), "live list B spliced in on a HIT");
+  assert.ok(!second.data.content.includes("Alpha"), "stale list A is NOT frozen into the output");
+  assert.deepEqual(second.data.subpages, [{ id: CHILD_B, title: "Beta" }], "subpages field reflects list B");
+});
+
+test("a cache HIT SKIPS the convertProseMirrorToMarkdown CPU walk (called once across MISS+HIT)", async () => {
+  // The single reason the cache exists: on a hit the expensive PM-tree walk must
+  // NOT run. The miss counter alone can't prove this — a broken hit branch that
+  // re-converted (same output, misses=1) would leave every other assert green.
+  // So spy directly on the conversion seam and assert the call COUNT.
+  const state = { info: 0, sidebar: 0, updatedAt: "2026-01-01T00:00:00Z", text: "Body text" };
+  const baseURL = await spawn(state);
+  const metrics = {};
+  const client = makeClient(baseURL, metrics);
+
+  // Spy on the seam that wraps convertProseMirrorToMarkdown; it still delegates,
+  // so output stays real and byte-identical — we only count invocations.
+  const spy = mock.method(client, "convertPageMarkdown");
+
+  await client.getPage(PAGE_UUID); // MISS -> converts once
+  assert.equal(spy.mock.callCount(), 1, "the miss converts exactly once");
+
+  await client.getPage(PAGE_UUID); // HIT -> must NOT convert again
+  assert.equal(
+    spy.mock.callCount(),
+    1,
+    "the hit skips the conversion: still exactly one call across MISS+HIT",
+  );
+  assert.equal(metrics["mcp_getpage_cache_hits_total"], 1, "and it was recorded as a hit");
+
+  spy.mock.restore();
+});

packages/mcp/test/unit/getpage-cache.test.mjs

@@ -0,0 +1,113 @@
+// Unit tests for the getPage content-addressed conversion cache (issue #479).
+// Exercises the GetPageConversionCache class in isolation: key composition,
+// hit/miss, LRU recency on read, and eviction by BOTH the count cap and the
+// byte cap. The getPage integration (counter emission, byte-identical output)
+// is covered separately in test/mock/getpage-conversion-cache.test.mjs.
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import {
+  GetPageConversionCache,
+  hashConvertOptions,
+} from "../../build/client/getpage-cache.js";
+
+test("key: identical (pageId, updatedAt, optionsHash) collapse to one entry", () => {
+  const c = new GetPageConversionCache();
+  const k1 = GetPageConversionCache.key("uuid-1", "2026-01-01T00:00:00Z", "{}");
+  const k2 = GetPageConversionCache.key("uuid-1", "2026-01-01T00:00:00Z", "{}");
+  assert.equal(k1, k2, "same parts -> same key");
+  c.set(k1, "MD-A");
+  assert.equal(c.get(k2), "MD-A", "hit on the identical key");
+  assert.equal(c.size, 1);
+});
+
+test("miss after updatedAt changes (precise invalidation)", () => {
+  const c = new GetPageConversionCache();
+  const oldK = GetPageConversionCache.key("uuid-1", "v1", "{}");
+  const newK = GetPageConversionCache.key("uuid-1", "v2", "{}");
+  c.set(oldK, "OLD-MD");
+  assert.equal(c.get(newK), undefined, "a new updatedAt is a fresh key -> miss");
+  assert.equal(c.get(oldK), "OLD-MD", "the old snapshot is still addressable");
+});
+
+test("miss after options change (dropResolvedCommentAnchors true vs false)", () => {
+  const c = new GetPageConversionCache();
+  const hDrop = hashConvertOptions({ dropResolvedCommentAnchors: true });
+  const hKeep = hashConvertOptions({ dropResolvedCommentAnchors: false });
+  assert.notEqual(hDrop, hKeep, "different options hash to different keys");
+  const kDrop = GetPageConversionCache.key("uuid-1", "v1", hDrop);
+  const kKeep = GetPageConversionCache.key("uuid-1", "v1", hKeep);
+  c.set(kDrop, "AGENT-MD");
+  assert.equal(
+    c.get(kKeep),
+    undefined,
+    "the export variant must NOT be served the agent variant",
+  );
+});
+
+test("hashConvertOptions is order-insensitive and handles empty", () => {
+  assert.equal(
+    hashConvertOptions({ a: 1, b: 2 }),
+    hashConvertOptions({ b: 2, a: 1 }),
+    "key order does not change the hash",
+  );
+  assert.equal(hashConvertOptions(undefined), "{}");
+  assert.equal(hashConvertOptions(null), "{}");
+  assert.equal(hashConvertOptions({}), "{}");
+});
+
+test("LRU eviction by COUNT cap evicts the least-recently-used entry", () => {
+  const c = new GetPageConversionCache({ maxEntries: 2, maxBytes: 10 * 1024 * 1024 });
+  c.set("k1", "a");
+  c.set("k2", "b");
+  c.set("k3", "c"); // over the count cap -> evict k1 (oldest)
+  assert.equal(c.size, 2);
+  assert.equal(c.get("k1"), undefined, "k1 evicted");
+  assert.equal(c.get("k2"), "b");
+  assert.equal(c.get("k3"), "c");
+});
+
+test("a read refreshes recency so the OTHER entry is evicted next", () => {
+  const c = new GetPageConversionCache({ maxEntries: 2, maxBytes: 10 * 1024 * 1024 });
+  c.set("k1", "a");
+  c.set("k2", "b");
+  // Touch k1 so k2 becomes the least-recently-used.
+  assert.equal(c.get("k1"), "a");
+  c.set("k3", "c"); // evicts the LRU, which is now k2 (not k1)
+  assert.equal(c.get("k2"), undefined, "k2 was LRU and got evicted");
+  assert.equal(c.get("k1"), "a", "k1 survived because it was read");
+  assert.equal(c.get("k3"), "c");
+});
+
+test("LRU eviction by BYTE cap evicts oldest until under the cap", () => {
+  // Byte cap of 100; each value is 40 bytes -> at most 2 fit (80 < 100 < 120).
+  const big = "x".repeat(40);
+  const c = new GetPageConversionCache({ maxEntries: 50, maxBytes: 100 });
+  c.set("k1", big); // 40
+  c.set("k2", big); // 80
+  assert.equal(c.size, 2);
+  c.set("k3", big); // 120 > 100 -> evict k1 -> 80
+  assert.equal(c.size, 2, "byte cap forced an eviction despite the count cap");
+  assert.equal(c.get("k1"), undefined, "oldest evicted by bytes");
+  assert.equal(c.get("k2"), big);
+  assert.equal(c.get("k3"), big);
+  assert.ok(c.bytes <= 100, "total bytes stays within the cap");
+});
+
+test("re-setting an existing key updates value+recency and keeps bytes exact", () => {
+  const c = new GetPageConversionCache({ maxEntries: 5, maxBytes: 10 * 1024 * 1024 });
+  c.set("k1", "short");
+  const b1 = c.bytes;
+  assert.equal(b1, Buffer.byteLength("short", "utf8"));
+  c.set("k1", "a much longer value");
+  assert.equal(c.size, 1, "no duplicate entry");
+  assert.equal(c.bytes, Buffer.byteLength("a much longer value", "utf8"));
+  assert.equal(c.get("k1"), "a much longer value");
+});
+
+test("an oversized single entry is still stored (never a permanent miss)", () => {
+  const c = new GetPageConversionCache({ maxEntries: 5, maxBytes: 10 });
+  const big = "y".repeat(1000);
+  c.set("k1", big);
+  assert.equal(c.get("k1"), big, "the page bigger than the cap is still served");
+  assert.equal(c.size, 1);
+});