fix(mcp): оживить счётчики кэша getPage + тест «хит пропускает конверсию» (#479, ревью)

Правки по ревью #480.

F1 (мёртвые метрики): счётчики mcp_getpage_cache_hits_total/_misses_total
эмитились через onMetricFn, но серверный синк mcp.service.ts диспатчил по имени
через if/else-if БЕЗ default и знал только 2 имени → мои дропались молча; в
metrics.registry их вообще не было. Починка по существующему 3-частному паттерну
(как collab_connect_timeouts_total): имена-константы в metrics.constants.ts; два
Counter'а + incGetPageCacheHit/Miss в metrics.registry.ts; два else-if в
mcp.service.ts, роутящие ровно эти имена (существующие 2 ветки не тронуты).
Проверено end-to-end: скрейп prom-реестра показывает hits=2/misses=1 после
прогона роутинга.

F2 (нет теста на пропуск конверсии): добавлен overridable seam
convertPageMarkdown в read.ts (идиома проекта для юнит-тестируемости ESM-импортов);
getPage miss-ветка идёт через него. Тест мокает convertPageMarkdown и ассертит
callCount===1 через MISS→HIT одной страницы (конверсия один раз на промахе, ноль
на хите). Мутационно доказан: пропатчил hit-ветку на повторную конверсию → тест
покраснел (callCount=2), откатил → зелёный.

mcp node --test 814/814 (+1); pmd+mcp tsc чисто; серверные metrics-файлы
компилируются изолированно, runtime-тест счётчиков зелёный.

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

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');

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/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);
+});