fix(mcp): adapt e2e node-ops calls to the #413 XOR input of patchNode/insertNode

#413 changed patchNode/insertNode to take { markdown? | node? } (exactly
one), but test-e2e.mjs still passed the raw ProseMirror node directly,
so the e2e-mcp CI job died with the XOR guard error right after the
node_ops seed step. Wrap both call sites in { node: ... }.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

.github/workflows/develop.yml

@@ -157,6 +157,12 @@ jobs:
       - name: Build prosemirror-markdown
         run: pnpm --filter @docmost/prosemirror-markdown build
 
+      # docmost-client.loader.ts type-imports from @docmost/mcp (issue #446); its
+      # build/ is gitignored and `test:e2e` type-checks, so build it here or tsc
+      # fails with TS2307 (mirrors the e2e-mcp / mcp-server-parity jobs).
+      - name: Build mcp
+        run: pnpm --filter @docmost/mcp build
+
       - name: Run migrations
         run: pnpm --filter ./apps/server migration:latest
 

apps/server/src/core/ai-chat/tools/tool-tiers.spec.ts

@@ -27,10 +27,12 @@ import type { DocmostClientLike } from './docmost-client.loader';
  */
 
 describe('tool tier metadata (#332)', () => {
-  it('core set is the documented 13 + searchInPage + insertFootnote (15)', () => {
-    expect(CORE_TOOL_KEYS).toHaveLength(15);
+  it('core set is the documented 13 + searchInPage + insertFootnote + getTree + getPageContext (17, #443)', () => {
+    expect(CORE_TOOL_KEYS).toHaveLength(17);
     expect(CORE_TOOL_SET.has('searchInPage')).toBe(true); // #330, promoted to core
     expect(CORE_TOOL_SET.has('insertFootnote')).toBe(true); // #410, promoted to core
+    expect(CORE_TOOL_SET.has('getTree')).toBe(true); // #443, promoted to core
+    expect(CORE_TOOL_SET.has('getPageContext')).toBe(true); // #443, promoted to core
     // loadTools is a meta-tool, not a normal core key.
     expect(CORE_TOOL_SET.has(LOAD_TOOLS_NAME)).toBe(false);
   });

apps/server/src/core/ai-chat/tools/tool-tiers.ts

@@ -39,12 +39,14 @@ export interface ToolCatalogEntry {
 
 /**
  * CORE (always-active) in-app tool keys — 13 frequent/tiny tools + `searchInPage`
- * (#330) + `insertFootnote` (#410). `searchInPage` is core because it is frequent
- * for the editorial roles this feature targets; `insertFootnote` is core so the
- * footnote tool is NOT hidden while its natural sibling `editPageText` is always
- * active (that asymmetry is exactly what pushed the agent to write literal
- * `^[...]`). `loadTools` is active too but is not a normal tool key (it is added
- * to activeTools separately).
+ * (#330) + `insertFootnote` (#410) + `getTree`/`getPageContext` (#443).
+ * `searchInPage` is core because it is frequent for the editorial roles this
+ * feature targets; `insertFootnote` is core so the footnote tool is NOT hidden
+ * while its natural sibling `editPageText` is always active (that asymmetry is
+ * exactly what pushed the agent to write literal `^[...]`). `getTree` and
+ * `getPageContext` are the single-call navigation/lookup tools — core so the
+ * agent never has to loadTools just to orient itself. `loadTools` is active too
+ * but is not a normal tool key (it is added to activeTools separately).
  */
 export const CORE_TOOL_KEYS = [
   'searchPages',
@@ -66,6 +68,11 @@ export const CORE_TOOL_KEYS = [
   // #410 insertFootnote — core so pinpoint citations to already-written text
   // don't degrade into literal `^[...]`; kept symmetric with editPageText.
   'insertFootnote',
+  // #443 getTree + getPageContext — cheap single-call navigation/lookup tools
+  // (the core listPages even points to getTree); core so the agent never has
+  // to loadTools just to orient itself.
+  'getTree',
+  'getPageContext',
 ] as const;
 
 /** O(1) membership test for the core tier. */

apps/server/test/integration/ai-chat-attach.int-spec.ts

@@ -182,6 +182,7 @@ describe('AiChatService run-stream attach [integration]', () => {
       {
         isAiChatDeferredToolsEnabled: () => false,
         isAiChatResumableStreamEnabled: () => true,
+        isAiChatFinalStepLockdownEnabled: () => false,
       } as any,
       registry,
     );
@@ -499,6 +500,7 @@ describe('AiChatService run-stream attach [integration]', () => {
       {
         isAiChatDeferredToolsEnabled: () => false,
         isAiChatResumableStreamEnabled: () => true,
+        isAiChatFinalStepLockdownEnabled: () => false,
       } as any,
       registry,
     );

apps/server/test/integration/ai-chat-stream.int-spec.ts

@@ -150,7 +150,7 @@ describe('AiChatService.stream [integration]', () => {
       {} as any, // pageAccess (idem)
       // environment (#332): keep deferred tool loading OFF for this lifecycle
       // harness so the toolset/behavior is exactly as before.
-      { isAiChatDeferredToolsEnabled: () => false } as any,
+      { isAiChatDeferredToolsEnabled: () => false, isAiChatFinalStepLockdownEnabled: () => false } as any,
     );
   }
 
@@ -378,7 +378,7 @@ describe('AiChatService.stream [integration]', () => {
         {} as any,
         {} as any,
         // #332: deferred tool loading ON — the property under test.
-        { isAiChatDeferredToolsEnabled: () => true } as any,
+        { isAiChatDeferredToolsEnabled: () => true, isAiChatFinalStepLockdownEnabled: () => false } as any,
       );
     }
 

packages/mcp/src/lib/diff.ts

@@ -14,7 +14,19 @@
  * signature.
  *
  * If recreateTransform / the changeset throws on a pathological document pair,
- * we fall back to a coarse block-level text diff so the tool never hard-fails.
+ * OR the pair is too large to diff cheaply (see the size guard below), we fall
+ * back to a coarse block-level text diff so the tool never hard-fails and never
+ * pins the event loop.
+ *
+ * SIZE GUARD (issue #464 — prod CPU-DoS). recreateTransform computes its diff via
+ * rfc6902.createPatch, whose array diff is O(n·m) Levenshtein per array pair and
+ * whose per-run word diff is O(w²); on a large/heavily-changed doc this runs for
+ * seconds-to-hours and starves the whole process (BullMQ, Redis lock renewals,
+ * embeddings). It never THROWS — it just never finishes — so the try/catch below
+ * cannot save us. Because diffDocs runs on EVERY in-app/MCP content edit's verify
+ * report, we PRE-FLIGHT the doc size and route anything above a cheap cap straight
+ * to the coarse fallback (the same shape the catch produces). Same cap+fallback
+ * pattern as the ELK-layout DoS fix (#440 / c917dcc3).
  */
 
 import { Node } from "@tiptap/pm/model";
@@ -72,6 +84,56 @@ function countNodes(doc: any, pred: (node: any) => boolean): number {
   return n;
 }
 
+// --- Issue #464: pre-flight size guard for the precise diff ------------------
+// Defaults are BENCHMARK-derived on the recreateTransform(complexSteps:false,
+// wordDiffs:true, simplifyDiff:true) pipeline, chosen so the WORST case (a fully
+// re-written doc — the adversarial shape that drove the incident) keeps the
+// synchronous block under ~200ms REGARDLESS of input:
+//   - 150 total nodes: worst-case pair ~176ms; the O(node²) array diff crosses
+//     200ms at ~170 nodes and then explodes super-linearly (400 nodes ~1.3s,
+//     800 ~5.5s), so cap just below the crossover.
+//   - 12 KiB serialized JSON: an independent axis, because the per-run word diff
+//     is O(words²) — a FEW nodes with very long text runs is dangerous even at a
+//     low node count (17 nodes / ~11 KiB ~176ms, / ~14 KiB ~290ms). A node-light
+//     but byte-heavy doc is still refused.
+// Either metric over its cap routes to the coarse fallback. Both are env-tunable
+// for operators who accept more CPU in exchange for exact diffs on larger docs.
+const DEFAULT_MAX_NODES = 150;
+const DEFAULT_MAX_BYTES = 12 * 1024;
+
+/**
+ * Read a positive-integer env override, falling back to `dflt`. Garbage / unset /
+ * non-finite / non-positive all fall back (so the guard can never be accidentally
+ * disabled by a malformed value). Read fresh on every call so a test / operator
+ * can flip the knob without a restart.
+ */
+function readPositiveIntEnv(name: string, dflt: number): number {
+  const raw = parseInt(process.env[name] ?? "", 10);
+  return Number.isFinite(raw) && raw > 0 ? raw : dflt;
+}
+
+/**
+ * True when the pair is too large for the precise (recreateTransform) diff and
+ * must degrade to the coarse fallback. Takes the MAX of the two docs on each
+ * metric so an ASYMMETRIC pair (a small new doc vs a huge old doc, or vice
+ * versa) — which still explodes rfc6902 — is caught. Cheap: one node walk +
+ * one JSON.stringify per doc, both O(size).
+ */
+function exceedsDiffSizeGuard(oldDoc: any, newDoc: any): boolean {
+  const maxNodes = readPositiveIntEnv("MCP_DIFF_MAX_NODES", DEFAULT_MAX_NODES);
+  const maxBytes = readPositiveIntEnv("MCP_DIFF_MAX_BYTES", DEFAULT_MAX_BYTES);
+  const nodes = Math.max(
+    countNodes(oldDoc, () => true),
+    countNodes(newDoc, () => true),
+  );
+  if (nodes > maxNodes) return true;
+  const bytes = Math.max(
+    JSON.stringify(oldDoc)?.length ?? 0,
+    JSON.stringify(newDoc)?.length ?? 0,
+  );
+  return bytes > maxBytes;
+}
+
 /**
  * Count UNIQUE links in a JSON doc by their `href`. A single link can be split
  * across several adjacent text runs (e.g. a "link+bold" run followed by a "link"
@@ -226,6 +288,81 @@ function coarseDiff(oldDoc: any, newDoc: any): DiffChange[] {
   return changes;
 }
 
+/** Accumulated textual changes plus their derived char/block tallies. */
+interface DiffTally {
+  changes: DiffChange[];
+  inserted: number;
+  deleted: number;
+  changedBlocks: Set<string>;
+}
+
+/**
+ * Produce the coarse-fallback tally for a pair. This is the SINGLE source of the
+ * `fellBack:true` result shape, shared by BOTH degrade paths in diffDocs (the
+ * pre-flight size guard and the recreateTransform catch) so they behave and
+ * report identically.
+ */
+function coarseDiffTally(oldDoc: any, newDoc: any): DiffTally {
+  const changes = coarseDiff(oldDoc, newDoc);
+  let inserted = 0;
+  let deleted = 0;
+  const changedBlocks = new Set<string>();
+  for (const c of changes) {
+    if (c.op === "insert") inserted += c.text.length;
+    else deleted += c.text.length;
+    if (c.block) changedBlocks.add(c.op[0] + ":" + c.block);
+  }
+  return { changes, inserted, deleted, changedBlocks };
+}
+
+/**
+ * Compute the PRECISE tally via the recreateTransform pipeline. Callers MUST
+ * gate this behind the size guard (it can block the event loop for a large pair)
+ * and wrap it in try/catch (a pathological pair can throw); on either the guard
+ * or a throw, use `coarseDiffTally` instead. Kept as a sibling of
+ * `coarseDiffTally` so both produce the same `DiffTally` shape.
+ */
+function preciseDiffTally(oldDocJson: any, newDocJson: any): DiffTally {
+  const oldNode = Node.fromJSON(docmostSchema, oldDocJson);
+  const newNode = Node.fromJSON(docmostSchema, newDocJson);
+  const tr = recreateTransform(oldNode, newNode, {
+    complexSteps: false,
+    wordDiffs: true,
+    simplifyDiff: true,
+  });
+  const changeSet = ChangeSet.create(oldNode).addSteps(tr.doc, tr.mapping.maps, []);
+  const simplified = simplifyChanges(changeSet.changes, newNode);
+
+  const changes: DiffChange[] = [];
+  let inserted = 0;
+  let deleted = 0;
+  const changedBlocks = new Set<string>();
+
+  for (const change of simplified) {
+    // Deleted text lives in the OLD doc coordinate range [fromA, toA).
+    if (change.toA > change.fromA) {
+      const text = oldNode.textBetween(change.fromA, change.toA, "\n", " ");
+      if (text.length > 0) {
+        deleted += text.length;
+        const block = blockContextAt(oldNode, change.fromA);
+        changes.push({ op: "delete", block, text });
+        if (block) changedBlocks.add("d:" + block);
+      }
+    }
+    // Inserted text lives in the NEW doc coordinate range [fromB, toB).
+    if (change.toB > change.fromB) {
+      const text = newNode.textBetween(change.fromB, change.toB, "\n", " ");
+      if (text.length > 0) {
+        inserted += text.length;
+        const block = blockContextAt(newNode, change.fromB);
+        changes.push({ op: "insert", block, text });
+        if (block) changedBlocks.add("i:" + block);
+      }
+    }
+  }
+  return { changes, inserted, deleted, changedBlocks };
+}
+
 /** Build the human-readable unified-ish markdown summary. */
 function renderMarkdown(
   result: Omit<DiffResult, "markdown">,
@@ -276,66 +413,39 @@ export function diffDocs(
   newDocJson: any,
   notesHeading: string = "Примечания переводчика",
 ): DiffResult {
+  // computeIntegrity is cheap (linear node walks) and its counts are needed in
+  // BOTH the precise and coarse paths, so it always runs first.
   const integrity = computeIntegrity(oldDocJson, newDocJson, notesHeading);
 
-  let changes: DiffChange[] = [];
-  let inserted = 0;
-  let deleted = 0;
   let fellBack = false;
-  const changedBlocks = new Set<string>();
+  let tally: DiffTally;
 
-  try {
-    const oldNode = Node.fromJSON(docmostSchema, oldDocJson);
-    const newNode = Node.fromJSON(docmostSchema, newDocJson);
-    const tr = recreateTransform(oldNode, newNode, {
-      complexSteps: false,
-      wordDiffs: true,
-      simplifyDiff: true,
-    });
-    const changeSet = ChangeSet.create(oldNode).addSteps(
-      tr.doc,
-      tr.mapping.maps,
-      [],
-    );
-    const simplified = simplifyChanges(changeSet.changes, newNode);
-
-    for (const change of simplified) {
-      // Deleted text lives in the OLD doc coordinate range [fromA, toA).
-      if (change.toA > change.fromA) {
-        const text = oldNode.textBetween(change.fromA, change.toA, "\n", " ");
-        if (text.length > 0) {
-          deleted += text.length;
-          const block = blockContextAt(oldNode, change.fromA);
-          changes.push({ op: "delete", block, text });
-          if (block) changedBlocks.add("d:" + block);
-        }
-      }
-      // Inserted text lives in the NEW doc coordinate range [fromB, toB).
-      if (change.toB > change.fromB) {
-        const text = newNode.textBetween(change.fromB, change.toB, "\n", " ");
-        if (text.length > 0) {
-          inserted += text.length;
-          const block = blockContextAt(newNode, change.fromB);
-          changes.push({ op: "insert", block, text });
-          if (block) changedBlocks.add("i:" + block);
-        }
-      }
-    }
-  } catch {
-    // Pathological pair: degrade to a coarse block-level diff so we never throw.
+  // Pre-flight size guard (#464): a too-large pair would make recreateTransform
+  // block the event loop for seconds-to-hours WITHOUT throwing, so route it to
+  // the coarse fallback BEFORE calling recreateTransform at all. Both this path
+  // and the catch below go through coarseDiffTally for an identical `fellBack`
+  // result shape.
+  if (exceedsDiffSizeGuard(oldDocJson, newDocJson)) {
     fellBack = true;
-    changes = coarseDiff(oldDocJson, newDocJson);
-    for (const c of changes) {
-      if (c.op === "insert") inserted += c.text.length;
-      else deleted += c.text.length;
-      if (c.block) changedBlocks.add(c.op[0] + ":" + c.block);
+    tally = coarseDiffTally(oldDocJson, newDocJson);
+  } else {
+    try {
+      tally = preciseDiffTally(oldDocJson, newDocJson);
+    } catch {
+      // Pathological pair: degrade to a coarse block-level diff so we never throw.
+      fellBack = true;
+      tally = coarseDiffTally(oldDocJson, newDocJson);
     }
   }
 
   const partial: Omit<DiffResult, "markdown"> = {
-    summary: { inserted, deleted, blocksChanged: changedBlocks.size },
+    summary: {
+      inserted: tally.inserted,
+      deleted: tally.deleted,
+      blocksChanged: tally.changedBlocks.size,
+    },
     integrity,
-    changes,
+    changes: tally.changes,
   };
   return { ...partial, markdown: renderMarkdown(partial, fellBack) };
 }

packages/mcp/test-e2e.mjs

@@ -316,7 +316,8 @@ async function main() {
         const [idA, idB, idC] = seedIds;
 
         // patchNode: replace the middle paragraph; siblings' ids must be unchanged.
-        await client.patchNode(nid, idB, mkPara(idB, "Bravo PATCHED."));
+        // #413 XOR input: the raw ProseMirror node goes under the `node` key.
+        await client.patchNode(nid, idB, { node: mkPara(idB, "Bravo PATCHED.") });
         await new Promise((r) => setTimeout(r, 16000));
         const afterPatch = (await client.getPageJson(nid)).content;
         const patchText = JSON.stringify(afterPatch);
@@ -327,7 +328,7 @@ async function main() {
         // insertNode: place a new block after the first paragraph.
         await client.insertNode(
           nid,
-          mkPara("nodeops-ins", "Inserted paragraph."),
+          { node: mkPara("nodeops-ins", "Inserted paragraph.") },
           { position: "after", anchorNodeId: idA },
         );
         await new Promise((r) => setTimeout(r, 16000));

packages/mcp/test/unit/diff-guard-skips-recreate.test.mjs

@@ -0,0 +1,86 @@
+// Issue #464 — prove the size guard SKIPS the recreateTransform pipeline over
+// the cap, not merely that it returns "coarse". node:test's mock.module needs an
+// experimental flag the suite does not pass, so instead of a module spy we use a
+// deterministic BEHAVIORAL proxy that isolates the one variable — the guard:
+//
+//   Same over-cap pair, run twice:
+//     (a) default caps            -> guard trips  -> recreateTransform skipped,
+//     (b) caps raised above the doc -> guard OFF -> recreateTransform DOES run.
+//
+// The only code path that differs between (a) and (b) is whether
+// recreateTransform executes. recreateTransform on this pair is O(n²) and takes
+// SECONDS; the guarded path is a linear coarse diff taking milliseconds. So a
+// large (a)≪(b) time ratio can ONLY be explained by (a) skipping the transform.
+// This asserts the skip without depending on mock.module.
+import { test } from "node:test";
+import assert from "node:assert/strict";
+
+import { diffDocs } from "../../build/lib/diff.js";
+
+const t = (text) => ({ type: "text", text });
+const para = (text) => ({ type: "paragraph", content: text ? [t(text)] : [] });
+const doc = (children) => ({ type: "doc", content: children });
+function buildDoc(n, seed) {
+  return doc(
+    Array.from({ length: n }, (_, i) =>
+      para(Array.from({ length: 8 }, (_, w) => `${seed}${i}_${w}`).join(" ")),
+    ),
+  );
+}
+function clearEnv() {
+  delete process.env.MCP_DIFF_MAX_NODES;
+  delete process.env.MCP_DIFF_MAX_BYTES;
+}
+function timed(fn) {
+  const s = performance.now();
+  const out = fn();
+  return { out, ms: performance.now() - s };
+}
+
+// A 300-para (~600-node) pair: comfortably over the 150-node default, yet small
+// enough that the un-guarded recreateTransform still FINISHES (~1-3s) so the
+// test can time the contrast without hanging.
+const OLD = buildDoc(300, "a");
+const NEW = buildDoc(300, "b");
+
+test("guard skips recreateTransform over-cap (guarded run is far faster than un-guarded)", () => {
+  // (a) Guarded: default caps -> should short-circuit to coarse, near-instant.
+  clearEnv();
+  const guarded = timed(() => diffDocs(OLD, NEW));
+  assert.match(
+    guarded.out.markdown,
+    /coarse block-level diff/,
+    "guarded run must be coarse (guard tripped)",
+  );
+
+  // (b) Un-guarded: raise both caps above the doc so the precise path runs.
+  process.env.MCP_DIFF_MAX_NODES = "1000000";
+  process.env.MCP_DIFF_MAX_BYTES = "100000000";
+  let unguarded;
+  try {
+    unguarded = timed(() => diffDocs(OLD, NEW));
+  } finally {
+    clearEnv();
+  }
+  assert.doesNotMatch(
+    unguarded.out.markdown,
+    /coarse block-level diff/,
+    "with caps raised, the precise recreateTransform path runs",
+  );
+
+  // The precise run executed recreateTransform (O(n²)); the guarded run did not.
+  // Require a large speedup so the ONLY explanation is the skipped transform.
+  assert.ok(
+    guarded.ms * 5 < unguarded.ms,
+    `guarded (${guarded.ms.toFixed(1)}ms) must be >=5x faster than un-guarded ` +
+      `(${unguarded.ms.toFixed(1)}ms); a small gap would mean the transform still ran`,
+  );
+});
+
+test("guarded over-cap call stays within the ~200ms event-loop budget", () => {
+  clearEnv();
+  // Best-of-3 to shed GC/JIT noise; the guarded coarse path is a linear walk.
+  let best = Infinity;
+  for (let i = 0; i < 3; i++) best = Math.min(best, timed(() => diffDocs(OLD, NEW)).ms);
+  assert.ok(best < 200, `guarded over-cap diff must be <200ms, was ${best.toFixed(1)}ms`);
+});

packages/mcp/test/unit/diff-size-guard.test.mjs

@@ -0,0 +1,229 @@
+// Issue #464 — prod CPU-DoS pre-flight size guard for diffDocs.
+//
+// diffDocs synchronously calls recreateTransform (rfc6902) which is O(n·m) in
+// node count and O(w²) in per-run word count; on a large/heavily-changed doc it
+// pins the event loop for seconds-to-hours WITHOUT throwing. A pre-flight size
+// guard routes any doc over MCP_DIFF_MAX_NODES / MCP_DIFF_MAX_BYTES straight to
+// the coarse fallback (`fellBack:true`), so the sync block stays ~<200ms.
+//
+// These tests assert the BEHAVIOR of the guard (fast + coarse-mode + asymmetry +
+// env knobs). A sibling test (diff-guard-skips-recreate.test.mjs) proves
+// recreateTransform is skipped over the cap via a behavioral proxy (guarded run
+// is orders of magnitude faster than the same pair with the caps raised).
+import { test } from "node:test";
+import assert from "node:assert/strict";
+
+import { diffDocs } from "../../build/lib/diff.js";
+
+// ---------------------------------------------------------------------------
+// Builders
+// ---------------------------------------------------------------------------
+const t = (text) => ({ type: "text", text });
+const para = (text) => ({ type: "paragraph", content: text ? [t(text)] : [] });
+const doc = (children) => ({ type: "doc", content: children });
+
+/** A doc of `n` paragraphs whose words are seeded from `seed` (fully changeable). */
+function buildDoc(n, wordsPerPara, seed) {
+  const blocks = [];
+  for (let i = 0; i < n; i++) {
+    const words = [];
+    for (let w = 0; w < wordsPerPara; w++) words.push(`${seed}${i}_${w}`);
+    blocks.push(para(words.join(" ")));
+  }
+  return doc(blocks);
+}
+
+/** Reset the env knobs to their unset default between tests. */
+function clearEnv() {
+  delete process.env.MCP_DIFF_MAX_NODES;
+  delete process.env.MCP_DIFF_MAX_BYTES;
+}
+
+// ---------------------------------------------------------------------------
+// Over-threshold (by node count) -> FAST + coarse mode.
+// A fully re-written 600-para doc is the worst case that drove the incident;
+// with the guard it must return in well under the ~200ms budget and in coarse
+// mode. Without the guard this single call takes multiple SECONDS.
+// ---------------------------------------------------------------------------
+test("over-threshold doc falls back to coarse mode and returns fast", () => {
+  clearEnv();
+  // 600 paragraphs -> ~1200 nodes, far over the 150-node default.
+  const oldDoc = buildDoc(600, 8, "a");
+  const newDoc = buildDoc(600, 8, "b");
+
+  const start = performance.now();
+  const r = diffDocs(oldDoc, newDoc);
+  const elapsed = performance.now() - start;
+
+  // Coarse mode is signalled in the markdown note (fellBack path).
+  assert.match(
+    r.markdown,
+    /coarse block-level diff/,
+    "over-threshold pair must use the coarse fallback",
+  );
+  // Budget: the guard makes this near-instant. Generous 1s ceiling to avoid CI
+  // flake while still being ~10x under the multi-second un-guarded cost.
+  assert.ok(
+    elapsed < 1000,
+    `expected fast coarse fallback, took ${elapsed.toFixed(0)}ms`,
+  );
+  // Coarse diff still detects the wholesale change.
+  assert.ok(r.summary.inserted > 0 || r.summary.deleted > 0, "reports changes");
+});
+
+// ---------------------------------------------------------------------------
+// Under-threshold (small) doc -> precise diff, NOT coarse mode. No regression.
+// ---------------------------------------------------------------------------
+test("under-threshold doc uses the precise diff (no fallback note)", () => {
+  clearEnv();
+  const oldDoc = doc([para("Hello world")]);
+  const newDoc = doc([para("Hello brave world")]);
+  const r = diffDocs(oldDoc, newDoc);
+
+  assert.doesNotMatch(
+    r.markdown,
+    /coarse block-level diff/,
+    "a small doc must take the precise path",
+  );
+  // Precise word diff finds exactly the inserted word.
+  const ins = r.changes.find((c) => c.op === "insert");
+  assert.ok(ins && /brave/.test(ins.text), "precise diff isolates the inserted word");
+});
+
+// ---------------------------------------------------------------------------
+// Asymmetry: a small NEW doc vs a huge OLD doc (and vice versa) still explodes
+// rfc6902, so max(old,new) must trip the guard in BOTH directions.
+// ---------------------------------------------------------------------------
+test("asymmetric pair (huge old, tiny new) falls back to coarse", () => {
+  clearEnv();
+  const hugeOld = buildDoc(600, 8, "a");
+  const tinyNew = doc([para("just one line")]);
+  const r = diffDocs(hugeOld, tinyNew);
+  assert.match(r.markdown, /coarse block-level diff/, "huge-old side must trip the guard");
+});
+
+test("asymmetric pair (tiny old, huge new) falls back to coarse", () => {
+  clearEnv();
+  const tinyOld = doc([para("just one line")]);
+  const hugeNew = buildDoc(600, 8, "b");
+  const r = diffDocs(tinyOld, hugeNew);
+  assert.match(r.markdown, /coarse block-level diff/, "huge-new side must trip the guard");
+});
+
+// ---------------------------------------------------------------------------
+// Byte axis: a FEW nodes but a very large serialized size (long text runs) is
+// dangerous too (per-run word diff is O(words²)), so the byte cap must trip
+// independently of the node count.
+// ---------------------------------------------------------------------------
+test("node-light but byte-heavy doc falls back on the byte cap", () => {
+  clearEnv();
+  // 5 paragraphs (~11 nodes, well under the node cap) but each a very long run,
+  // pushing the serialized size far over the 12 KiB byte default.
+  const bigRun = (seed) =>
+    doc(
+      Array.from({ length: 5 }, (_, i) =>
+        para(Array.from({ length: 800 }, (_, w) => `${seed}${i}_${w}`).join(" ")),
+      ),
+    );
+  const oldDoc = bigRun("a");
+  const newDoc = bigRun("b");
+  // Sanity: node count is under the default node cap, so ONLY the byte cap can
+  // be what trips the guard here.
+  const nodeCount = (d) => {
+    let n = 0;
+    const v = (x) => {
+      if (!x || typeof x !== "object") return;
+      n++;
+      if (Array.isArray(x.content)) for (const c of x.content) v(c);
+    };
+    v(d);
+    return n;
+  };
+  assert.ok(nodeCount(oldDoc) < 150, "node count is under the node cap");
+  assert.ok(JSON.stringify(oldDoc).length > 12 * 1024, "serialized size is over the byte cap");
+
+  const r = diffDocs(oldDoc, newDoc);
+  assert.match(r.markdown, /coarse block-level diff/, "byte cap must trip independently");
+});
+
+// ---------------------------------------------------------------------------
+// Env override: a very low MCP_DIFF_MAX_NODES forces fallback on a tiny doc,
+// proving the knob is read fresh and actually gates the diff.
+// ---------------------------------------------------------------------------
+test("MCP_DIFF_MAX_NODES override forces fallback on a small doc", () => {
+  clearEnv();
+  const oldDoc = doc([para("Hello world")]);
+  const newDoc = doc([para("Hello brave world")]);
+
+  // Baseline: default caps -> precise diff.
+  assert.doesNotMatch(diffDocs(oldDoc, newDoc).markdown, /coarse block-level diff/);
+
+  // Knob set absurdly low -> even this 4-node doc trips the guard.
+  process.env.MCP_DIFF_MAX_NODES = "1";
+  try {
+    const r = diffDocs(oldDoc, newDoc);
+    assert.match(r.markdown, /coarse block-level diff/, "low node cap forces fallback");
+  } finally {
+    clearEnv();
+  }
+});
+
+test("MCP_DIFF_MAX_BYTES override forces fallback on a small doc", () => {
+  clearEnv();
+  const oldDoc = doc([para("Hello world")]);
+  const newDoc = doc([para("Hello brave world")]);
+
+  process.env.MCP_DIFF_MAX_BYTES = "1";
+  try {
+    const r = diffDocs(oldDoc, newDoc);
+    assert.match(r.markdown, /coarse block-level diff/, "low byte cap forces fallback");
+  } finally {
+    clearEnv();
+  }
+});
+
+// ---------------------------------------------------------------------------
+// Garbage / unset env values fall back to the DEFAULT (the guard can never be
+// accidentally disabled by a malformed knob). A small doc must still diff
+// precisely under a garbage cap.
+// ---------------------------------------------------------------------------
+test("garbage env values fall back to the default cap (guard not disabled)", () => {
+  clearEnv();
+  const oldDoc = doc([para("Hello world")]);
+  const newDoc = doc([para("Hello brave world")]);
+
+  for (const bad of ["not-a-number", "0", "-5", "", "NaN", "1e999"]) {
+    process.env.MCP_DIFF_MAX_NODES = bad;
+    process.env.MCP_DIFF_MAX_BYTES = bad;
+    // Under the DEFAULT caps this small doc is precise (garbage did not raise
+    // OR disable the cap). "1e999" -> parseInt yields 1 (finite) which is a
+    // valid low cap and would fall back; exclude that from the precise check.
+    const r = diffDocs(oldDoc, newDoc);
+    if (bad === "1e999") {
+      // parseInt("1e999",10) === 1 -> a legit low cap -> fallback. Guard active.
+      assert.match(r.markdown, /coarse block-level diff/);
+    } else {
+      assert.doesNotMatch(
+        r.markdown,
+        /coarse block-level diff/,
+        `garbage value ${JSON.stringify(bad)} must fall back to the default cap`,
+      );
+    }
+  }
+  clearEnv();
+});
+
+// ---------------------------------------------------------------------------
+// A large doc that trips the guard must still return the correct INTEGRITY
+// counts (computeIntegrity runs before the diff and is unaffected by fallback).
+// ---------------------------------------------------------------------------
+test("integrity counts are still correct on a guard-tripped (coarse) doc", () => {
+  clearEnv();
+  const image = { type: "image", attrs: { src: "/api/files/a.png" } };
+  const oldDoc = doc([image, ...buildDoc(600, 8, "a").content]);
+  const newDoc = doc([...buildDoc(600, 8, "b").content]); // image removed
+
+  const r = diffDocs(oldDoc, newDoc);
+  assert.match(r.markdown, /coarse block-level diff/, "large pair fell back");
+  assert.deepEqual(r.integrity.images, [1, 0], "integrity is computed regardless of fallback");
+});