fix(ai-chat): hydrate crashed mid-run steps for model replay + cover write/read seams (#492)

F1: the server model-replay loaded history via findAllByChat().map(rowToUiMessage)
WITHOUT hydrating parts from ai_chat_run_steps. A HARD crash mid-run (SIGKILL/OOM)
fires no terminal callback, so the assistant row stays parts:[] and its partial
tool-calls/results/text (durable in the steps table) dropped out of the model's
next-turn context. Hydrate needy assistant rows (role==='assistant' &&
!rowHasInlineParts) via findByMessageIds + hydrateAssistantParts before the replay
map — mirroring the controller's withReconstructedParts exactly — guarded on the
optional repo. Fix the now-false interrupt-resume comment.

F2: add a service int-spec that drives the REAL onStep append-persist WRITE branch
through AiChatService.stream with a real AiChatRunStepRepo injected, asserting the
per-step rows' stepIndex + parts slice and the step-marker metadata match a
single-row flush (catches an stepsPersisted-1 off-by-one).

F3: add a controller int-spec that drives withReconstructedParts through getMessages
WITH the repo present (a mid-run marker-only row + its step rows), asserting the
reconstructed metadata.parts and workspace-scoping.

F4: remove the dead countByMessage (zero prod callers; reconstructRunParts derives
stepsPersisted inline) + its now-unused sql import and the redundant test assertion.

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

.github/workflows/develop.yml

@@ -226,6 +226,13 @@ jobs:
       - name: Build mcp
         run: pnpm --filter @docmost/mcp build
 
+      # apps/server imports @docmost/token-estimate at runtime (history-budget.ts,
+      # #490); its dist/ is gitignored and `test:e2e` type-checks + runs the code,
+      # so build it here or tsc fails with TS2307 Cannot find module
+      # '@docmost/token-estimate' (mirrors the editor-ext / mcp build steps above).
+      - name: Build token-estimate
+        run: pnpm --filter @docmost/token-estimate build
+
       - name: Run migrations
         run: pnpm --filter ./apps/server migration:latest
 

.github/workflows/test.yml

@@ -159,6 +159,14 @@ jobs:
       - name: Build prosemirror-markdown
         run: pnpm --filter @docmost/prosemirror-markdown build
 
+      # @docmost/token-estimate is a shared workspace package the client vitest
+      # suite resolves via its dist build (main: ./dist/index.js); dist/ is
+      # gitignored and `pnpm -r test` does NOT honour nx `dependsOn: ^build`, so
+      # build it before the recursive test run or the client suite fails with
+      # "Failed to resolve import '@docmost/token-estimate'" (#490).
+      - name: Build token-estimate
+        run: pnpm --filter @docmost/token-estimate build
+
       - name: Run unit tests
         run: pnpm -r test
 

CHANGELOG.md

@@ -135,6 +135,13 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   new resync path re-reads the live anchor so the suggestion applies against the
   current text, and orphaned anchors (whose marked run was deleted) are
   reconciled rather than left blocking. (#496)
+- **Save intentional page versions.** Press `Cmd/Ctrl+S` (or use the page menu)
+  to save a named version of a page. The history panel now distinguishes
+  intentional versions (a "Saved" / "Agent version" badge) from automatic
+  snapshots, dims autosaves, and offers an "Only versions" filter. Automatic
+  snapshots switched from a fixed interval to a trailing idle-flush with a
+  max-wait ceiling, and a boundary snapshot is pinned whenever the editing source
+  changes (e.g. a person's edits followed by the AI agent). (#370)
 
 - **Place several images side by side in a row.** A new "Inline (side by
   side)" alignment mode in the image bubble menu renders consecutive inline
@@ -385,6 +392,25 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   `@docmost/token-estimate` package) so they can never diverge. Deferred-tool
   activation is also cached in the chat metadata to avoid re-resolving it each
   turn. (#490)
+- **Cyrillic (and any non-ASCII) draw.io labels no longer turn into mojibake
+  when a diagram is opened in the draw.io editor.** Agent-created diagrams
+  (`drawioCreate`) and Confluence-imported diagrams stored their model in the
+  SVG's `content=` attribute as base64; the draw.io editor decodes that via
+  Latin-1 `atob` (no UTF-8 step), so every non-ASCII char (e.g. `Старт-бит`,
+  `ё`, `—`) split into garbage and the editor's autosave then persisted the
+  corrupted model, breaking the page preview too. Both write paths
+  (`buildDrawioSvg`, the import service's `createDrawioSvg`) now write `content=`
+  as XML-entity-escaped mxfile XML — draw.io's own native form, decoded by the
+  DOM as UTF-8 — so labels open intact. The decoder reads both the new
+  entity-encoded form and the old base64 form, so existing diagrams still open.
+  *Healing pre-fix diagrams:* only a diagram that still holds its original
+  (correct-UTF-8) base64 — i.e. one not yet opened/autosaved in the draw.io
+  editor — can be repaired in place by `drawioGet` → `drawioUpdate` with the
+  same XML (rewrites the attachment in the new form); no migration script is
+  needed. A diagram that was already opened in the editor persisted the
+  mojibake at rest, so `drawioGet` reads the already-corrupted text and
+  `drawioUpdate` faithfully rewrites it — that text is lost and is not
+  recoverable by a rewrite. (#507)
 - **A chat with one malformed message part no longer 500s on every turn, and a
   failed send no longer duplicates the user's message.** Incoming client parts
   are now whitelisted to `text` (a forged tool-result part can no longer reach

Dockerfile

@@ -59,6 +59,14 @@ COPY --from=builder /app/packages/mcp/data /app/packages/mcp/data
 COPY --from=builder /app/packages/prosemirror-markdown/build /app/packages/prosemirror-markdown/build
 COPY --from=builder /app/packages/prosemirror-markdown/package.json /app/packages/prosemirror-markdown/package.json
 
+# apps/server imports @docmost/token-estimate (workspace:*) at runtime
+# (history-budget.ts, #490). tsc emits only dist/ and dist/ is gitignored, so the
+# prod install would resolve a broken workspace symlink and the server would die
+# with ERR_MODULE_NOT_FOUND on the first history-budget call. Ship the built
+# package + its manifest, mirroring prosemirror-markdown above.
+COPY --from=builder /app/packages/token-estimate/dist /app/packages/token-estimate/dist
+COPY --from=builder /app/packages/token-estimate/package.json /app/packages/token-estimate/package.json
+
 # Copy root package files
 COPY --from=builder /app/package.json /app/package.json
 COPY --from=builder /app/pnpm*.yaml /app/

apps/client/public/locales/en-US/translation.json

@@ -1418,5 +1418,14 @@
   "The commented text changed since this suggestion was made; it was not applied.": "The commented text changed since this suggestion was made; it was not applied.",
   "Dismiss": "Dismiss",
   "Suggestion dismissed": "Suggestion dismissed",
-  "Failed to dismiss suggestion": "Failed to dismiss suggestion"
+  "Failed to dismiss suggestion": "Failed to dismiss suggestion",
+  "Save version": "Save version",
+  "Ctrl+S": "Ctrl+S",
+  "Version saved": "Version saved",
+  "Already saved as the latest version": "Already saved as the latest version",
+  "Agent version": "Agent version",
+  "Boundary": "Boundary",
+  "Autosave": "Autosave",
+  "Only versions": "Only versions",
+  "No saved versions yet.": "No saved versions yet."
 }

apps/client/public/locales/ru-RU/translation.json

@@ -1433,5 +1433,14 @@
   "The commented text changed since this suggestion was made; it was not applied.": "Прокомментированный текст изменился после создания предложения; оно не было применено.",
   "Dismiss": "Не применять",
   "Suggestion dismissed": "Предложение отклонено",
-  "Failed to dismiss suggestion": "Не удалось отклонить предложение"
+  "Failed to dismiss suggestion": "Не удалось отклонить предложение",
+  "Save version": "Сохранить версию",
+  "Ctrl+S": "Ctrl+S",
+  "Version saved": "Версия сохранена",
+  "Already saved as the latest version": "Уже сохранено как последняя версия",
+  "Agent version": "Версия агента",
+  "Boundary": "Граница",
+  "Autosave": "Автосейв",
+  "Only versions": "Только версии",
+  "No saved versions yet.": "Пока нет сохранённых версий."
 }

apps/client/src/features/ai-chat/components/message-item-memo.test.tsx

@@ -27,6 +27,7 @@ vi.mock("@/features/ai-chat/utils/markdown.ts", async () => {
 
 import MessageItem from "./message-item";
 import { messageSignature } from "@/features/ai-chat/utils/message-signature.ts";
+import { splitPlainChunks } from "./streaming-plain-text";
 
 // matchMedia (read by MantineProvider) is stubbed globally in vitest.setup.ts.
 
@@ -114,3 +115,89 @@ describe("MessageItem markdown memoization", () => {
     expect(queryByText("streamed answer")).not.toBeNull();
   });
 });
+
+// PERF SMOKE (#492): the whole point of the incremental streaming render is that
+// the ANSWER path costs O(number of markdown blocks), NOT O(number of throttled
+// ~20Hz ticks). Pre-#492 the finalized MarkdownPart re-parsed the WHOLE growing
+// answer on every delta — a synthetic ~100 KB stream measured 394 renderChatMarkdown
+// calls (one per tick). With the incremental render each STABILIZED block is parsed
+// exactly once (memoized in MarkdownChunk) and the live tail is cheap plain text, so
+// the call count collapses to ~= the block count regardless of tick granularity.
+describe("MessageItem streaming answer render is O(blocks), not O(ticks)", () => {
+  // ~100 KB answer. Each section is a heading + a paragraph — TWO blank-line
+  // delimited markdown blocks — so the safe-cut block count is ~2× the section
+  // count. The perf claim is about the BLOCK count (the memoization granularity),
+  // measured directly with splitPlainChunks below, not the section count.
+  const buildAnswer = () => {
+    const SECTIONS = 100;
+    const paragraphs: string[] = [];
+    for (let i = 0; i < SECTIONS; i++) {
+      paragraphs.push(`## Section ${i}\n\n` + "lorem ipsum dolor ".repeat(55));
+    }
+    const full = paragraphs.join("\n\n");
+    // The number of memoized markdown blocks the incremental render splits into
+    // (all but the live tail are parsed once each).
+    return { full, blocks: splitPlainChunks(full).length };
+  };
+
+  const streamMsg = (text: string, state: "streaming" | "done"): UIMessage =>
+    ({
+      id: "m1",
+      role: "assistant",
+      parts: [{ type: "text", text, state }],
+    }) as UIMessage;
+
+  it("parses each block ~once over a 100KB stream (≈blocks, ≪ ticks)", () => {
+    renderChatMarkdownSpy.mockClear();
+    const { full, blocks } = buildAnswer();
+    const CHUNK = 128; // a realistic ~20Hz throttled delta size
+    const ticks = Math.ceil(full.length / CHUNK);
+
+    let msg = streamMsg(full.slice(0, CHUNK), "streaming");
+    const { rerender } = render(
+      <MantineProvider>
+        <MessageItem
+          message={msg}
+          signature={messageSignature(msg)}
+          turnStreaming
+        />
+      </MantineProvider>,
+    );
+    for (let end = 2 * CHUNK; end < full.length; end += CHUNK) {
+      msg = streamMsg(full.slice(0, end), "streaming");
+      rerender(
+        <MantineProvider>
+          <MessageItem
+            message={msg}
+            signature={messageSignature(msg)}
+            turnStreaming
+          />
+        </MantineProvider>,
+      );
+    }
+    // Finalize: the streaming→done flip renders the whole answer through ONE
+    // canonical pass (visual parity), so the finished DOM matches the pre-#492
+    // output. This is the single extra parse on top of the per-block ones.
+    const done = streamMsg(full, "done");
+    rerender(
+      <MantineProvider>
+        <MessageItem message={done} signature={messageSignature(done)} />
+      </MantineProvider>,
+    );
+
+    const calls = renderChatMarkdownSpy.mock.calls.length;
+    // Sanity: the stream really had far more ticks than blocks (else the test is
+    // vacuous — the point is that calls scale with blocks, not ticks).
+    expect(ticks).toBeGreaterThan(blocks * 3);
+    // O(blocks): each stabilized block parsed once + the single final whole-text
+    // parse. A small constant absorbs the finalize render and the live-tail block;
+    // the load-bearing claim is the bound below.
+    expect(calls).toBeLessThanOrEqual(blocks + 2);
+    // ≪ ticks — and, non-vacuously, the blocks WERE parsed (not skipped entirely).
+    expect(calls).toBeLessThan(ticks / 3);
+    expect(calls).toBeGreaterThan(blocks / 2);
+    // MUTATION-VERIFY (documented, not run here): dropping the `memo()` wrapper on
+    // MarkdownChunk (so every stable block re-parses each tick) drives `calls`
+    // toward `ticks` (~394), reddening both upper-bound assertions above.
+  });
+});

apps/client/src/features/ai-chat/components/message-item.render.test.tsx

@@ -0,0 +1,112 @@
+import { describe, expect, it, vi } from "vitest";
+import { render } from "@testing-library/react";
+import { MantineProvider } from "@mantine/core";
+import type { UIMessage } from "@ai-sdk/react";
+
+// Stub react-i18next (the component reads `useTranslation`). Mirrors the other
+// message-item specs.
+vi.mock("react-i18next", () => ({
+  useTranslation: () => ({ t: (key: string) => key }),
+}));
+
+import MessageItem from "./message-item";
+import { messageSignature } from "@/features/ai-chat/utils/message-signature.ts";
+// The REAL canonical renderer (NOT the spy the memo test installs): this file
+// exercises the actual markdown output so the visual-regression assertions below
+// compare against genuine HTML (incl. the schema's `<li><p>` wrappers).
+import { renderChatMarkdown } from "@/features/ai-chat/utils/markdown.ts";
+import classes from "./ai-chat.module.css";
+
+const msg = (
+  parts: UIMessage["parts"],
+  extra?: Partial<UIMessage>,
+): UIMessage =>
+  ({ id: "m1", role: "assistant", parts, ...extra }) as UIMessage;
+
+const renderRow = (message: UIMessage, turnStreaming = false) =>
+  render(
+    <MantineProvider>
+      <MessageItem
+        message={message}
+        signature={messageSignature(message)}
+        turnStreaming={turnStreaming}
+      />
+    </MantineProvider>,
+  );
+
+// A rich multi-block answer that exercises headings, a list (the `<li><p>` case
+// the scoped CSS tightens), inline emphasis, and multiple paragraphs.
+const ANSWER = [
+  "# Заголовок",
+  "",
+  "Первый абзац с **жирным** и `кодом`.",
+  "",
+  "- пункт один",
+  "- пункт два",
+  "",
+  "Второй абзац.",
+].join("\n");
+
+describe("MessageItem final render — visual parity with the canonical pipeline", () => {
+  it("a finalized text part renders exactly renderChatMarkdown(text)", () => {
+    const { container } = renderRow(
+      msg([{ type: "text", text: ANSWER, state: "done" }]),
+    );
+    const block = container.querySelector(`.${classes.markdown}`);
+    expect(block).not.toBeNull();
+    // Byte-for-byte the canonical output (the SAME whole-text pass the pre-#492
+    // MarkdownPart produced), including `<li><p>…</p></li>` wrappers.
+    expect(block!.innerHTML).toBe(renderChatMarkdown(ANSWER, {}));
+    // The list wrapper is really present (guards against a vacuous empty render).
+    expect(container.querySelectorAll("li p").length).toBe(2);
+  });
+
+  it("the streaming incremental view CONVERGES to the canonical render on finish", () => {
+    // Mount mid-stream (live tail) — the DOM here is the incremental view.
+    const { container, rerender } = render(
+      <MantineProvider>
+        <MessageItem
+          message={msg([{ type: "text", text: ANSWER, state: "streaming" }])}
+          signature={messageSignature(
+            msg([{ type: "text", text: ANSWER, state: "streaming" }]),
+          )}
+          turnStreaming
+        />
+      </MantineProvider>,
+    );
+    // Finish the turn: state flips to done AND the turn is no longer streaming.
+    const done = msg([{ type: "text", text: ANSWER, state: "done" }]);
+    rerender(
+      <MantineProvider>
+        <MessageItem message={done} signature={messageSignature(done)} />
+      </MantineProvider>,
+    );
+    // After finish there is exactly ONE canonical markdown container whose HTML is
+    // the whole-text render — identical to the non-streaming path above.
+    const blocks = container.querySelectorAll(`.${classes.markdown}`);
+    expect(blocks.length).toBe(1);
+    expect(blocks[0].innerHTML).toBe(renderChatMarkdown(ANSWER, {}));
+  });
+
+  it("neutralizeInternalLinks is honored on the finalized render", () => {
+    const linkAnswer = "См. [страницу](/p/abc).";
+    const { container } = render(
+      <MantineProvider>
+        <MessageItem
+          message={msg([{ type: "text", text: linkAnswer, state: "done" }])}
+          signature={messageSignature(
+            msg([{ type: "text", text: linkAnswer, state: "done" }]),
+          )}
+          neutralizeInternalLinks
+        />
+      </MantineProvider>,
+    );
+    const block = container.querySelector(`.${classes.markdown}`);
+    expect(block!.innerHTML).toBe(
+      renderChatMarkdown(linkAnswer, { neutralizeInternalLinks: true }),
+    );
+    // The internal link was made inert (no href) by the neutralization flag.
+    const a = container.querySelector("a");
+    expect(a?.hasAttribute("href")).toBe(false);
+  });
+});

apps/client/src/features/ai-chat/components/message-item.tsx

@@ -4,6 +4,7 @@ import { useTranslation } from "react-i18next";
 import type { UIMessage } from "@ai-sdk/react";
 import ToolCallCard from "@/features/ai-chat/components/tool-call-card.tsx";
 import ReasoningBlock from "@/features/ai-chat/components/reasoning-block.tsx";
+import { StreamingMarkdownText } from "@/features/ai-chat/components/streaming-markdown-text.tsx";
 import ChatErrorAlert from "@/features/ai-chat/components/chat-error-alert.tsx";
 import ChatStoppedNotice from "@/features/ai-chat/components/chat-stopped-notice.tsx";
 import { ToolUiPart, isToolPart } from "@/features/ai-chat/utils/tool-parts.tsx";
@@ -86,17 +87,39 @@ interface MessageItemProps {
  * One assistant text part rendered as sanitized markdown. Memoized on its inputs
  * so a finalized text part is NOT re-parsed on every streamed delta: during a
  * turn only the actively-growing tail part changes its `text`, so every earlier
- * part hits the memo and skips the expensive marked + DOMPurify pass. Props are
- * primitives, so React.memo's default shallow compare is exactly right (the
- * `text` string is compared by value).
+ * part hits the memo and skips the expensive canonical parse + DOMPurify pass.
+ * Props are primitives, so React.memo's default shallow compare is exactly right
+ * (the `text` string is compared by value).
+ *
+ * Streaming gate (#492) — mirrors ReasoningBlock:
+ *  - `streaming` (this is the live, actively-growing tail part of an in-flight
+ *    turn): render incrementally via StreamingMarkdownText — the stabilized blocks
+ *    go through the canonical pipeline (each parsed ONCE, memoized) and only the
+ *    live tail is cheap plain text. This makes the per-tick cost O(new blocks),
+ *    not the pre-#492 O(ticks) whole-answer re-parse on every ~20Hz delta.
+ *  - finalized (the common case, and the turn-end flip): render the WHOLE text
+ *    through ONE canonical pass — byte-identical to the pre-#492 output (visual
+ *    parity). The row re-renders on the streaming→done flip because
+ *    `messageSignature` tracks each part's `state` (and `turnStreaming` flips at
+ *    turn end), so the incremental view always converges to this single render.
  */
 const MarkdownPart = memo(function MarkdownPart({
   text,
   neutralizeInternalLinks,
+  streaming,
 }: {
   text: string;
   neutralizeInternalLinks: boolean;
+  streaming: boolean;
 }) {
+  if (streaming) {
+    return (
+      <StreamingMarkdownText
+        text={text}
+        neutralizeInternalLinks={neutralizeInternalLinks}
+      />
+    );
+  }
   const html = renderChatMarkdown(text, { neutralizeInternalLinks });
   if (html) {
     return (
@@ -179,47 +202,10 @@ function MessageItem({
         {resolveAssistantName(assistantName) ?? t("AI agent")}
       </Text>
       {message.parts.map((part, index) => {
-        if (part.type === "reasoning") {
-          // Reasoning ("thinking") -> a collapsible block with its own token
-          // count. Empty/whitespace reasoning with no authoritative count carries
-          // nothing to show, so skip it (avoids an empty 0-token block).
-          const text = (part as { text?: string }).text ?? "";
-          if (!text.trim() && !(reasoningTokens && reasoningTokens > 0))
-            return null;
-          // Absent state (persisted rows) and "done" both mean finalized.
-          // `messageSignature` already includes each part's `state`, so the
-          // streaming→done flip changes the row signature and re-renders this
-          // row — which is what lets ReasoningBlock switch from chunked plain
-          // text to its one-time markdown parse (see reasoning-block.tsx).
-          // ALSO require the turn to be live: a part stranded at
-          // `state:"streaming"` after the turn ended (no `reasoning-end` — see
-          // the `turnStreaming` prop doc) must still finalize and parse.
-          const streaming =
-            turnStreaming && (part as { state?: string }).state === "streaming";
-          return (
-            <ReasoningBlock
-              key={index}
-              text={text}
-              tokens={reasoningTokens}
-              streaming={streaming}
-            />
-          );
-        }
-
-        if (part.type === "text") {
-          // Skip empty/whitespace-only text parts (a streaming message often
-          // starts with an empty text part before the first token arrives); the
-          // typing indicator covers that gap until real content streams in.
-          if (!part.text.trim()) return null;
-          return (
-            <MarkdownPart
-              key={index}
-              text={part.text}
-              neutralizeInternalLinks={neutralizeInternalLinks}
-            />
-          );
-        }
-
+        // Tool parts (`tool-*` / `dynamic-tool`) are template-literal kinds, so
+        // they cannot be a `switch` case; the runtime guard handles them, and the
+        // switch below covers every CLOSED (literal-typed) part kind with a
+        // compile-time exhaustiveness check in its default.
         if (isToolPart(part.type)) {
           return (
             <ToolCallCard
@@ -232,7 +218,76 @@ function MessageItem({
           );
         }
 
-        return null;
+        switch (part.type) {
+          case "reasoning": {
+            // Reasoning ("thinking") -> a collapsible block with its own token
+            // count. Empty/whitespace reasoning with no authoritative count
+            // carries nothing to show, so skip it (avoids an empty 0-token block).
+            const text = part.text ?? "";
+            if (!text.trim() && !(reasoningTokens && reasoningTokens > 0))
+              return null;
+            // Absent state (persisted rows) and "done" both mean finalized.
+            // `messageSignature` already includes each part's `state`, so the
+            // streaming→done flip changes the row signature and re-renders this
+            // row — which is what lets ReasoningBlock switch from chunked plain
+            // text to its one-time markdown parse (see reasoning-block.tsx).
+            // ALSO require the turn to be live: a part stranded at
+            // `state:"streaming"` after the turn ended (no `reasoning-end` — see
+            // the `turnStreaming` prop doc) must still finalize and parse.
+            const streaming = turnStreaming && part.state === "streaming";
+            return (
+              <ReasoningBlock
+                key={index}
+                text={text}
+                tokens={reasoningTokens}
+                streaming={streaming}
+              />
+            );
+          }
+
+          case "text": {
+            // Skip empty/whitespace-only text parts (a streaming message often
+            // starts with an empty text part before the first token arrives); the
+            // typing indicator covers that gap until real content streams in.
+            if (!part.text.trim()) return null;
+            // The live, actively-growing tail part of the in-flight turn renders
+            // incrementally (see MarkdownPart); a finalized part (persisted, or
+            // the turn-end flip) renders the whole text through one canonical
+            // pass. Same liveness rule as the reasoning branch above.
+            const streaming = turnStreaming && part.state === "streaming";
+            return (
+              <MarkdownPart
+                key={index}
+                text={part.text}
+                neutralizeInternalLinks={neutralizeInternalLinks}
+                streaming={streaming}
+              />
+            );
+          }
+
+          case "source-url":
+          case "source-document":
+          case "file":
+          case "step-start":
+            // Not surfaced in the chat bubble (v1) — same as the pre-#492 default.
+            return null;
+
+          default: {
+            // Compile-time exhaustiveness over the CLOSED union members: every
+            // literal-typed part kind is handled above, so the only kinds that
+            // can reach here are the OPEN template-literal ones (`tool-*` — caught
+            // by the guard at runtime — and `data-*`) plus `dynamic-tool`. Adding
+            // a NEW closed part kind to UIMessagePart makes this assignment fail
+            // to compile, forcing it to be handled instead of silently ignored
+            // (this replaces the pre-#492 fall-through `return null` + WARNING).
+            const _exhaustive:
+              | `tool-${string}`
+              | "dynamic-tool"
+              | `data-${string}` = part.type;
+            void _exhaustive;
+            return null;
+          }
+        }
       })}
       {/* A persisted turn error (server stored it in metadata.error). Rendered
           here so it survives a thread remount and shows in reopened history. */}

apps/client/src/features/ai-chat/components/streaming-markdown-text.tsx

@@ -0,0 +1,96 @@
+import { memo, useMemo } from "react";
+import { splitPlainChunks } from "@/features/ai-chat/components/streaming-plain-text.tsx";
+import { renderChatMarkdown } from "@/features/ai-chat/utils/markdown.ts";
+import classes from "@/features/ai-chat/components/ai-chat.module.css";
+
+/**
+ * One STABILIZED markdown block, rendered through the canonical pipeline and
+ * memoized on its string prop. During streaming only the TAIL chunk grows (the
+ * `splitPlainChunks` append-only invariant guarantees every earlier chunk is
+ * byte-identical across deltas), so React skips every stable block and each one
+ * is parsed by `renderChatMarkdown` EXACTLY ONCE — turning the pre-#492
+ * "re-parse the whole accumulated answer on every ~20Hz tick" (O(ticks)) into
+ * O(number of blocks). The markup is DOMPurify-sanitized inside renderChatMarkdown
+ * before it reaches `dangerouslySetInnerHTML`.
+ *
+ * NOTE (transient streaming-only artifact): a safe cut is a blank-line boundary,
+ * so a construct that legitimately contains a blank line (e.g. a fenced code block
+ * with an empty line) can be split across chunks and render oddly WHILE it is still
+ * streaming. This is cosmetic and self-heals: the moment the part finalizes,
+ * MarkdownPart renders the WHOLE text through one canonical pass (visual parity
+ * with the pre-#492 output). The reasoning path makes the same trade (plain text
+ * while streaming, one markdown parse at the end).
+ */
+const MarkdownChunk = memo(function MarkdownChunk({
+  text,
+  neutralizeInternalLinks,
+}: {
+  text: string;
+  neutralizeInternalLinks: boolean;
+}) {
+  const html = renderChatMarkdown(text, { neutralizeInternalLinks });
+  if (html) {
+    return (
+      <div
+        className={classes.markdown}
+        // Sanitized by renderChatMarkdown (DOMPurify) before insertion.
+        dangerouslySetInnerHTML={{ __html: html }}
+      />
+    );
+  }
+  // Malformed/unsupported markdown could not render synchronously: raw text.
+  return (
+    <div className={classes.markdown} style={{ whiteSpace: "pre-wrap" }}>
+      {text}
+    </div>
+  );
+});
+
+/**
+ * The cheap streaming-time stand-in for the finalized answer's one-time markdown
+ * parse (see MarkdownPart in message-item.tsx). Mirrors StreamingPlainText's
+ * chunked-memo pattern but renders the STABILIZED prefix as real markdown (each
+ * block parsed once, memoized) and only the LIVE tail as flat plain text — so the
+ * user sees formatted output for everything up to the last safe cut, and the not-
+ * yet-stable tail (which markdown-parsing every tick would make O(ticks)) stays a
+ * single cheap escaped text node until it stabilizes into a new block.
+ *
+ * `splitPlainChunks` yields chunks where, under append-only growth, every chunk
+ * except the LAST is immutable; the last chunk is the live tail. Index keys are
+ * therefore stable (a given index never changes to a different chunk's content).
+ */
+export function StreamingMarkdownText({
+  text,
+  neutralizeInternalLinks,
+}: {
+  text: string;
+  neutralizeInternalLinks: boolean;
+}) {
+  const chunks = useMemo(() => splitPlainChunks(text), [text]);
+  return (
+    <>
+      {chunks.map((chunk, index) =>
+        index < chunks.length - 1 ? (
+          <MarkdownChunk
+            key={index}
+            text={chunk}
+            neutralizeInternalLinks={neutralizeInternalLinks}
+          />
+        ) : (
+          // The live tail: flat, React-escaped plain text (no markdown parse, no
+          // sanitizer, no innerHTML). `pre-wrap` preserves its newlines; trailing
+          // separator newlines are dropped at display time so the block gap comes
+          // from the markdown margins, not a doubled empty line (mirrors
+          // PlainChunk in streaming-plain-text.tsx).
+          <div
+            key={index}
+            className={classes.markdown}
+            style={{ whiteSpace: "pre-wrap" }}
+          >
+            {chunk.replace(/\n+$/, "")}
+          </div>
+        ),
+      )}
+    </>
+  );
+}

apps/client/src/features/editor/atoms/editor-atoms.ts

@@ -3,11 +3,20 @@ import { atom } from "jotai";
 // import would drag the whole @tiptap/core engine into the eager graph of every
 // shell component that reads one of these atoms.
 import type { Editor } from "@tiptap/core";
+import type { HocuspocusProvider } from "@hocuspocus/provider";
 import { PageEditMode } from "@/features/user/types/user.types.ts";
 import type { DictationUnavailableReason } from "@/features/dictation/dictation-status";
 
 export const pageEditorAtom = atom<Editor | null>(null);
 
+// #370 — the active page's collab provider, published by the page editor so the
+// header menu can emit the "save-version" stateless signal (Cmd+S / button).
+// Null when the page is read-only / collab isn't connected. A typed initial
+// value (rather than an explicit generic) keeps jotai's overload resolution on
+// the writable PrimitiveAtom branch.
+const initialCollabProvider: HocuspocusProvider | null = null;
+export const collabProviderAtom = atom(initialCollabProvider);
+
 export const titleEditorAtom = atom<Editor | null>(null);
 
 export const readOnlyEditorAtom = atom<Editor | null>(null);

apps/client/src/features/editor/page-editor.tsx

@@ -31,11 +31,18 @@ import { useAtom, useAtomValue, useSetAtom } from "jotai";
 import useCollaborationUrl from "@/features/editor/hooks/use-collaboration-url";
 import { currentUserAtom } from "@/features/user/atoms/current-user-atom";
 import {
+  collabProviderAtom,
   currentPageEditModeAtom,
   dictationAvailabilityAtom,
   pageEditorAtom,
   yjsConnectionStatusAtom,
 } from "@/features/editor/atoms/editor-atoms";
+import { notifications } from "@mantine/notifications";
+import {
+  VERSION_SAVED_MESSAGE_TYPE,
+  type VersionSavedMessage,
+  saveVersionPending,
+} from "@/features/page-history/version-messages";
 import { asideStateAtom } from "@/components/layouts/global/hooks/atoms/sidebar-atom";
 import {
   activeCommentIdAtom,
@@ -124,6 +131,7 @@ export default function PageEditor({
 
   const [currentUser] = useAtom(currentUserAtom);
   const [, setEditor] = useAtom(pageEditorAtom);
+  const setCollabProvider = useSetAtom(collabProviderAtom);
   const [, setAsideState] = useAtom(asideStateAtom);
   const [, setActiveCommentId] = useAtom(activeCommentIdAtom);
   const [showCommentPopup, setShowCommentPopup] = useAtom(showCommentPopupAtom);
@@ -181,6 +189,24 @@ export default function PageEditor({
       const onStatelessHandler = ({ payload }: onStatelessParameters) => {
         try {
           const message = JSON.parse(payload);
+          // #370 — a version was saved somewhere; live-refresh the history panel
+          // on every client. Only the client that pressed Save (tracked by the
+          // module-level flag) shows the confirmation toast.
+          if (message?.type === VERSION_SAVED_MESSAGE_TYPE) {
+            const versionMsg = message as VersionSavedMessage;
+            queryClient.invalidateQueries({
+              queryKey: ["page-history-list"],
+            });
+            if (saveVersionPending.current) {
+              saveVersionPending.current = false;
+              notifications.show({
+                message: versionMsg.alreadySaved
+                  ? t("Already saved as the latest version")
+                  : t("Version saved"),
+              });
+            }
+            return;
+          }
           if (message?.type !== "page.updated" || !message.updatedAt) return;
           const pageData = queryClient.getQueryData<IPage>(["pages", slugId]);
           if (pageData) {
@@ -238,12 +264,16 @@ export default function PageEditor({
 
       local.on("synced", onLocalSyncedHandler);
       providersRef.current = { socket, local, remote };
+      // #370 — publish the provider so the header menu can emit save-version.
+      setCollabProvider(remote);
       setProvidersReady(true);
     } else {
+      setCollabProvider(providersRef.current.remote);
       setProvidersReady(true);
     }
     // Only destroy on final unmount
     return () => {
+      setCollabProvider(null);
       providersRef.current?.socket.destroy();
       providersRef.current?.remote.destroy();
       providersRef.current?.local.destroy();

apps/client/src/features/page-history/components/history-item.tsx

@@ -1,4 +1,11 @@
-import { Text, Group, UnstyledButton, Avatar, Tooltip } from "@mantine/core";
+import {
+  Text,
+  Group,
+  UnstyledButton,
+  Avatar,
+  Tooltip,
+  Badge,
+} from "@mantine/core";
 import { CustomAvatar } from "@/components/ui/custom-avatar.tsx";
 import { AgentAvatarStack } from "@/components/ui/agent-avatar-stack.tsx";
 import { formattedDate } from "@/lib/time";
@@ -7,36 +14,59 @@ import clsx from "clsx";
 import { IPageHistory } from "@/features/page-history/types/page.types";
 import { memo, useCallback } from "react";
 import { useSetAtom } from "jotai";
+import { useTranslation } from "react-i18next";
 import { historyAtoms } from "@/features/page-history/atoms/history-atoms.ts";
 
 const MAX_VISIBLE_AVATARS = 5;
 
+/**
+ * #370 — map a snapshot's intentionality tier to its badge. `version: true`
+ * marks the intentional points (manual / agent); autosaves (boundary / idle /
+ * legacy null) are non-versions and get dimmed in the list.
+ */
+type HistoryKindMeta = { labelKey: string; color: string; version: boolean };
+export function historyKindMeta(kind?: string | null): HistoryKindMeta {
+  switch (kind) {
+    case "manual":
+      return { labelKey: "Saved", color: "blue", version: true };
+    case "agent":
+      return { labelKey: "Agent version", color: "violet", version: true };
+    case "boundary":
+      return { labelKey: "Boundary", color: "gray", version: false };
+    default: // "idle" | null | undefined (legacy autosave)
+      return { labelKey: "Autosave", color: "gray", version: false };
+  }
+}
+
 interface HistoryItemProps {
   historyItem: IPageHistory;
-  index: number;
-  onSelect: (id: string, index: number) => void;
-  onHover?: (id: string, index: number) => void;
+  // The previous snapshot for diff/restore is resolved by id from the FULL list
+  // in the parent (resolvePrevSnapshotId), so the item only needs to report its
+  // own id — never a list index (which would be the filtered-view index).
+  onSelect: (id: string) => void;
+  onHover?: (id: string) => void;
   onHoverEnd?: () => void;
   isActive: boolean;
 }
 
 const HistoryItem = memo(function HistoryItem({
   historyItem,
-  index,
   onSelect,
   onHover,
   onHoverEnd,
   isActive,
 }: HistoryItemProps) {
   const setHistoryModalOpen = useSetAtom(historyAtoms);
+  const { t } = useTranslation();
+  const kindMeta = historyKindMeta(historyItem.kind);
 
   const handleClick = useCallback(() => {
-    onSelect(historyItem.id, index);
-  }, [onSelect, historyItem.id, index]);
+    onSelect(historyItem.id);
+  }, [onSelect, historyItem.id]);
 
   const handleMouseEnter = useCallback(() => {
-    onHover?.(historyItem.id, index);
-  }, [onHover, historyItem.id, index]);
+    onHover?.(historyItem.id);
+  }, [onHover, historyItem.id]);
 
   const contributors = historyItem.contributors;
   const hasContributors = contributors && contributors.length > 0;
@@ -49,8 +79,20 @@ const HistoryItem = memo(function HistoryItem({
       onMouseEnter={handleMouseEnter}
       onMouseLeave={onHoverEnd}
       className={clsx(classes.history, { [classes.active]: isActive })}
+      // #370 — dim autosnapshots so intentional versions stand out.
+      style={{ opacity: kindMeta.version ? 1 : 0.55 }}
     >
-      <Text size="sm">{formattedDate(new Date(historyItem.createdAt))}</Text>
+      <Group gap={6} wrap="nowrap" justify="space-between">
+        <Text size="sm">{formattedDate(new Date(historyItem.createdAt))}</Text>
+        <Badge
+          size="xs"
+          radius="sm"
+          variant={kindMeta.version ? "filled" : "light"}
+          color={kindMeta.color}
+        >
+          {t(kindMeta.labelKey)}
+        </Badge>
+      </Group>
 
       <Group gap={6} wrap="nowrap" mt={4}>
         {hasContributors ? (

apps/client/src/features/page-history/components/history-list.tsx

@@ -2,14 +2,16 @@ import {
   usePageHistoryListQuery,
   prefetchPageHistory,
 } from "@/features/page-history/queries/page-history-query";
-import HistoryItem from "@/features/page-history/components/history-item";
+import HistoryItem, {
+  historyKindMeta,
+} from "@/features/page-history/components/history-item";
 import {
   activeHistoryIdAtom,
   activeHistoryPrevIdAtom,
   historyAtoms,
 } from "@/features/page-history/atoms/history-atoms";
 import { useAtom, useSetAtom } from "jotai";
-import { useCallback, useEffect, useMemo, useRef } from "react";
+import { useCallback, useEffect, useMemo, useRef, useState } from "react";
 import {
   Button,
   ScrollArea,
@@ -17,9 +19,12 @@ import {
   Divider,
   Loader,
   Center,
+  Switch,
+  Text,
 } from "@mantine/core";
 import { useTranslation } from "react-i18next";
 import { useHistoryRestore } from "@/features/page-history/hooks";
+import { resolvePrevSnapshotId } from "@/features/page-history/utils/resolve-prev-snapshot";
 
 const PREFETCH_DELAY_MS = 150;
 
@@ -47,6 +52,22 @@ function HistoryList({ pageId }: Props) {
     [pageHistoryData],
   );
 
+  // #370 — "only versions" filter: hide autosnapshots (idle/boundary/legacy
+  // null), keep only intentional points (manual/agent). Filtering is over the
+  // already-loaded pages; the diff/restore still targets the true previous
+  // snapshot, so items carry their index within the FULL list.
+  const [onlyVersions, setOnlyVersions] = useState(false);
+  // Reuse historyKindMeta().version — the SAME predicate the badge (HistoryItem)
+  // uses to mark intentional points — so the "Only versions" filter and the badge
+  // can never drift apart when a future intentional kind is added.
+  const visibleItems = useMemo(
+    () =>
+      onlyVersions
+        ? historyItems.filter((item) => historyKindMeta(item.kind).version)
+        : historyItems,
+    [historyItems, onlyVersions],
+  );
+
   const loadMoreRef = useRef<HTMLDivElement>(null);
   const prefetchTimeoutRef = useRef<ReturnType<typeof setTimeout> | null>(null);
 
@@ -60,11 +81,13 @@ function HistoryList({ pageId }: Props) {
   }, []);
 
   const handleHover = useCallback(
-    (historyId: string, index: number) => {
+    (historyId: string) => {
       clearPrefetchTimeout();
       prefetchTimeoutRef.current = setTimeout(() => {
         prefetchPageHistory(historyId);
-        const prevId = historyItems[index + 1]?.id;
+        // The true previous snapshot in the FULL list (not the previous visible
+        // one under the "only versions" filter).
+        const prevId = resolvePrevSnapshotId(historyItems, historyId);
         if (prevId) {
           prefetchPageHistory(prevId);
         }
@@ -78,9 +101,11 @@ function HistoryList({ pageId }: Props) {
   }, [clearPrefetchTimeout]);
 
   const handleSelect = useCallback(
-    (id: string, index: number) => {
+    (id: string) => {
       setActiveHistoryId(id);
-      setActiveHistoryPrevId(historyItems[index + 1]?.id ?? "");
+      // Baseline = true previous snapshot in the FULL list, so the "only
+      // versions" filter never diffs/restores against the wrong item.
+      setActiveHistoryPrevId(resolvePrevSnapshotId(historyItems, id));
     },
     [historyItems, setActiveHistoryId, setActiveHistoryPrevId],
   );
@@ -128,12 +153,27 @@ function HistoryList({ pageId }: Props) {
 
   return (
     <div>
+      <Group px="xs" py={6} justify="flex-end">
+        <Switch
+          size="xs"
+          checked={onlyVersions}
+          onChange={(e) => setOnlyVersions(e.currentTarget.checked)}
+          label={t("Only versions")}
+        />
+      </Group>
+
       <ScrollArea h={620} w="100%" type="scroll" scrollbarSize={5}>
-        {historyItems.map((historyItem, index) => (
+        {onlyVersions && visibleItems.length === 0 && (
+          <Center py="md">
+            <Text size="sm" c="dimmed">
+              {t("No saved versions yet.")}
+            </Text>
+          </Center>
+        )}
+        {visibleItems.map((historyItem) => (
           <HistoryItem
             key={historyItem.id}
             historyItem={historyItem}
-            index={index}
             onSelect={handleSelect}
             onHover={handleHover}
             onHoverEnd={clearPrefetchTimeout}

apps/client/src/features/page-history/types/page.types.ts

@@ -24,6 +24,10 @@ export interface IPageHistory {
   updatedAt: string;
   lastUpdatedBy: IPageHistoryUser;
   contributors?: IPageHistoryUser[];
+  // #370 — intentionality tier: 'manual'/'agent' are versions (intentional
+  // points), 'idle'/'boundary' are autosnapshots; null/undefined = legacy
+  // autosave. Derived server-side, drives the history badge + "versions" filter.
+  kind?: "manual" | "agent" | "idle" | "boundary" | null;
   // Provenance markers copied off the page row when the snapshot was saved.
   // `'agent'` marks a version written by the AI agent; `lastUpdatedAiChatId`
   // (when present) deep-links to the chat that produced the edit.

apps/client/src/features/page-history/utils/resolve-prev-snapshot.test.ts

@@ -0,0 +1,42 @@
+import { describe, it, expect } from "vitest";
+import { resolvePrevSnapshotId } from "./resolve-prev-snapshot";
+
+// #370 F4 — the risky client path: with the "only versions" filter active, diff
+// and restore must still baseline against the TRUE previous snapshot in the FULL
+// list, never the previous VISIBLE version (which would skip the autosnapshots
+// between two versions). These pin that the resolution is by FULL-list order.
+describe("resolvePrevSnapshotId", () => {
+  // Newest-first, as the history list stores it: a version, then two autosaves,
+  // then an older version.
+  const full = [
+    { id: "v2", kind: "manual" },
+    { id: "a2", kind: "idle" },
+    { id: "a1", kind: "boundary" },
+    { id: "v1", kind: "manual" },
+    { id: "a0", kind: null },
+  ];
+
+  it("returns the immediate FULL-list successor, not the previous visible version", () => {
+    // Selecting v2 while filtered to versions-only must baseline against a2 (the
+    // real chronological predecessor), NOT v1 (the previous visible version).
+    expect(resolvePrevSnapshotId(full, "v2")).toBe("a2");
+  });
+
+  it("resolves an autosnapshot's predecessor by full-list order", () => {
+    expect(resolvePrevSnapshotId(full, "a1")).toBe("v1");
+  });
+
+  it("returns '' for the oldest item (no predecessor)", () => {
+    expect(resolvePrevSnapshotId(full, "a0")).toBe("");
+  });
+
+  it("returns '' for an id not in the list", () => {
+    expect(resolvePrevSnapshotId(full, "missing")).toBe("");
+  });
+
+  it("does not depend on a filtered subset — same result whatever is visible", () => {
+    // The helper only ever sees the full list; a filtered view cannot change the
+    // baseline it computes.
+    expect(resolvePrevSnapshotId(full, "v1")).toBe("a0");
+  });
+});

apps/client/src/features/page-history/utils/resolve-prev-snapshot.ts

@@ -0,0 +1,22 @@
+/**
+ * #370 — resolve the TRUE previous snapshot for a history item.
+ *
+ * The history panel can be filtered to "only versions" (manual/agent), but diff
+ * and restore must always compare against the immediately-preceding snapshot in
+ * the FULL, unfiltered list — NOT the previous VISIBLE item. Comparing against
+ * the previous visible version would silently skip the autosnapshots between two
+ * versions and diff/restore the wrong baseline.
+ *
+ * Given the full (newest-first) list and an item id, this returns the id of the
+ * item right after it in the full list (its chronological predecessor), or "" if
+ * it is the oldest / not found. Pure and list-order-preserving so it can be unit
+ * tested without mounting the component.
+ */
+export function resolvePrevSnapshotId(
+  fullItems: ReadonlyArray<{ id: string }>,
+  id: string,
+): string {
+  const index = fullItems.findIndex((item) => item.id === id);
+  if (index === -1) return "";
+  return fullItems[index + 1]?.id ?? "";
+}

apps/client/src/features/page-history/version-messages.ts

@@ -0,0 +1,28 @@
+/**
+ * #370 — page-version stateless wire formats. Kept in one place so the client
+ * emitter (Save hotkey / button) and the client listener (page-editor) agree
+ * with the server (PersistenceExtension) on the message shapes.
+ */
+
+/** Client → server: "save a version now". The server derives the tier
+ * (manual/agent) from the signed connection actor, never from this payload. */
+export const SAVE_VERSION_MESSAGE_TYPE = "save-version";
+
+/** Server → all clients: a version was saved (or promoted / already existed). */
+export const VERSION_SAVED_MESSAGE_TYPE = "version.saved";
+
+export interface VersionSavedMessage {
+  type: typeof VERSION_SAVED_MESSAGE_TYPE;
+  historyId: string;
+  kind: "manual" | "agent";
+  /** True when the latest snapshot was already a manual version (a no-op save). */
+  alreadySaved: boolean;
+}
+
+/**
+ * Cross-component coordination flag so only the client that pressed Save shows
+ * the confirmation toast, while every other client silently refreshes its
+ * history panel on the broadcast. A module-level ref avoids stale-closure
+ * pitfalls in the editor's long-lived stateless handler.
+ */
+export const saveVersionPending = { current: false };

apps/client/src/features/page/components/header/page-header-menu.tsx

@@ -3,6 +3,7 @@ import {
   IconArrowRight,
   IconArrowsHorizontal,
   IconClockHour4,
+  IconDeviceFloppy,
   IconDots,
   IconEye,
   IconEyeOff,
@@ -17,7 +18,7 @@ import {
   IconTrash,
   IconWifiOff,
 } from "@tabler/icons-react";
-import React, { useEffect, useRef, useState } from "react";
+import React, { useCallback, useEffect, useRef, useState } from "react";
 import { useAsideTriggerProps } from "@/hooks/use-toggle-aside.tsx";
 import { useAtom, useAtomValue } from "jotai";
 import { historyAtoms } from "@/features/page-history/atoms/history-atoms.ts";
@@ -39,9 +40,14 @@ import { Trans, useTranslation } from "react-i18next";
 import ExportModal from "@/components/common/export-modal";
 import { convertProseMirrorToMarkdown } from "@docmost/prosemirror-markdown/browser";
 import {
+  collabProviderAtom,
   pageEditorAtom,
   yjsConnectionStatusAtom,
 } from "@/features/editor/atoms/editor-atoms.ts";
+import {
+  SAVE_VERSION_MESSAGE_TYPE,
+  saveVersionPending,
+} from "@/features/page-history/version-messages.ts";
 import { formattedDate } from "@/lib/time.ts";
 import { PageEditModeToggle } from "@/features/user/components/page-state-pref.tsx";
 import MovePageModal from "@/features/page/components/move-page-modal.tsx";
@@ -72,9 +78,34 @@ export default function PageHeaderMenu({ readOnly }: PageHeaderMenuProps) {
   });
   const isDeleted = !!page?.deletedAt;
   const [workspace] = useAtom(workspaceAtom);
+  const collabProvider = useAtomValue(collabProviderAtom);
   // Community public-sharing entry point (replaces the removed EE PageShareModal)
   const workspaceSharingDisabled = workspace?.settings?.sharing?.disabled === true;
 
+  // #370 — explicit "save a version" (Cmd+S / Save button). One path for the
+  // human; the server derives the tier from the signed actor. Readers can't save
+  // (the button is hidden and the collab connection is read-only server-side).
+  const handleSaveVersion = useCallback(() => {
+    if (readOnly || !collabProvider) return;
+    // Flag this client as the initiator so only it shows the confirmation toast;
+    // a safety timeout clears it if no broadcast comes back (e.g. offline).
+    saveVersionPending.current = true;
+    window.setTimeout(() => {
+      saveVersionPending.current = false;
+    }, 5000);
+    collabProvider.sendStateless(
+      JSON.stringify({ type: SAVE_VERSION_MESSAGE_TYPE }),
+    );
+  }, [readOnly, collabProvider]);
+
+  // mod+S must also block the browser's "Save page" dialog. `triggerOnContent-
+  // Editable` + empty ignore-list so it fires while typing in the editor/title.
+  useHotkeys(
+    [["mod+S", handleSaveVersion, { preventDefault: true }]],
+    [],
+    true,
+  );
+
   useHotkeys(
     [
       [
@@ -133,15 +164,16 @@ export default function PageHeaderMenu({ readOnly }: PageHeaderMenuProps) {
         </ActionIcon>
       </Tooltip>
 
-      <PageActionMenu readOnly={readOnly} />
+      <PageActionMenu readOnly={readOnly} onSaveVersion={handleSaveVersion} />
     </>
   );
 }
 
 interface PageActionMenuProps {
   readOnly?: boolean;
+  onSaveVersion?: () => void;
 }
-function PageActionMenu({ readOnly }: PageActionMenuProps) {
+function PageActionMenu({ readOnly, onSaveVersion }: PageActionMenuProps) {
   const { t } = useTranslation();
   const [, setHistoryModalOpen] = useAtom(historyAtoms);
   const clipboard = useClipboard({ timeout: 500 });
@@ -303,6 +335,20 @@ function PageActionMenu({ readOnly }: PageActionMenuProps) {
             </Group>
           </Menu.Item>
 
+          {!readOnly && (
+            <Menu.Item
+              leftSection={<IconDeviceFloppy size={16} />}
+              onClick={onSaveVersion}
+              rightSection={
+                <Text size="xs" c="dimmed">
+                  {t("Ctrl+S")}
+                </Text>
+              }
+            >
+              {t("Save version")}
+            </Menu.Item>
+          )}
+
           <Menu.Item
             leftSection={<IconHistory size={16} />}
             onClick={openHistoryModal}

apps/client/src/features/page/tree/components/space-tree.tsx

@@ -281,10 +281,12 @@ const SpaceTree = forwardRef<SpaceTreeApi, SpaceTreeProps>(function SpaceTree(
       setOpenTreeNodes((prev) => ({ ...prev, [id]: isOpen }));
       if (isOpen) {
         const node = treeModel.find(data, id) as SpaceTreeNode | null;
-        if (
-          node?.hasChildren &&
-          (!node.children || node.children.length === 0)
-        ) {
+        // Same "unloaded branch" predicate the realtime insert paths use
+        // (`isUnloadedBranch`) so the lazy-load gate and the realtime inserts
+        // (`insertByPosition` / `placeByPosition`) can never disagree about what
+        // counts as unloaded (#525). Note: local raw `insert` (DnD/create-page)
+        // does not yet route through it — see #525 follow-up.
+        if (treeModel.isUnloadedBranch(node)) {
           const fetched = await fetchAllAncestorChildren({
             pageId: id,
             spaceId: node.spaceId,

apps/client/src/features/page/tree/model/tree-model.test.ts

@@ -74,6 +74,48 @@ describe("treeModel.isDescendant", () => {
   });
 });
 
+// #525: the single "is this branch unloaded?" predicate shared by the lazy-load
+// gate and the insert paths. Unloaded == server says hasChildren but none are
+// present locally (canonical form `children: []`, also `undefined`). A parent
+// without hasChildren is genuinely empty, not unloaded.
+describe("treeModel.isUnloadedBranch", () => {
+  type PH = TreeNode<{ name: string; hasChildren?: boolean }>;
+  it("true for hasChildren + empty array (canonical unloaded form)", () => {
+    const n: PH = { id: "p", name: "P", hasChildren: true, children: [] };
+    expect(treeModel.isUnloadedBranch(n)).toBe(true);
+  });
+  it("true for hasChildren + undefined children", () => {
+    const n: PH = { id: "p", name: "P", hasChildren: true };
+    expect(treeModel.isUnloadedBranch(n)).toBe(true);
+  });
+  it("false for hasChildren + already-loaded children", () => {
+    const n: PH = {
+      id: "p",
+      name: "P",
+      hasChildren: true,
+      children: [{ id: "c", name: "C" }],
+    };
+    expect(treeModel.isUnloadedBranch(n)).toBe(false);
+  });
+  it("false for a genuinely-empty parent (no hasChildren)", () => {
+    expect(
+      treeModel.isUnloadedBranch({
+        id: "p",
+        name: "P",
+        hasChildren: false,
+        children: [],
+      } as PH),
+    ).toBe(false);
+    expect(
+      treeModel.isUnloadedBranch({ id: "p", name: "P" } as PH),
+    ).toBe(false);
+  });
+  it("false for null/undefined", () => {
+    expect(treeModel.isUnloadedBranch(null)).toBe(false);
+    expect(treeModel.isUnloadedBranch(undefined)).toBe(false);
+  });
+});
+
 describe("treeModel.visible", () => {
   it("returns only root nodes when no openIds", () => {
     const v = treeModel.visible(fixture, new Set());
@@ -197,43 +239,64 @@ describe("treeModel.insertByPosition", () => {
     ]);
   });
 
-  // #159 #1: inserting/moving a node under a parent whose children are NOT
-  // loaded (`children === undefined`, e.g. a collapsed page) must NOT materialize
-  // a partial `[node]` list — that would defeat the lazy-load gate and hide the
-  // parent's other real children. The node is left to be lazy-loaded; only
-  // `hasChildren` is flagged so the chevron appears.
-  it("does NOT materialize a child under an UNLOADED parent (children undefined)", () => {
-    type PH = TreeNode<{
-      name: string;
-      position?: string;
-      hasChildren?: boolean;
-    }>;
+  type PH = TreeNode<{
+    name: string;
+    position?: string;
+    hasChildren?: boolean;
+  }>;
+
+  // #159 #1 / #525: inserting/moving a node under an UNLOADED parent must NOT
+  // materialize a partial `[node]` list — that would defeat the lazy-load gate and
+  // hide the parent's other real children. The canonical unloaded form here is
+  // `children: []` + `hasChildren: true` (from `pageToTreeNode` /
+  // `pruneCollapsedChildren`), which the pre-#525 `=== undefined` guard MISSED.
+  // The node is left to be lazy-loaded; the chevron stays enabled.
+  it("does NOT materialize a child under an UNLOADED parent (children: [], hasChildren: true)", () => {
     const tree: PH[] = [
-      { id: "p", name: "P", position: "a0", hasChildren: false }, // children: undefined
+      { id: "p", name: "P", position: "a0", hasChildren: true, children: [] },
     ];
     const node: PH = { id: "x", name: "X", position: "a1" };
     const t = treeModel.insertByPosition(tree, "p", node);
     const parent = treeModel.find(t, "p");
     // The node was NOT inserted (children stay unloaded -> lazy-load fetches the
-    // full set, including this node, on expand).
-    expect(parent?.children).toBeUndefined();
+    // full set, including this node, on expand). MUTATION: the pre-#525 predicate
+    // `children === undefined` does not fire for `[]`, so it would insert `[x]`
+    // here and reredden this expectation.
+    expect(parent?.children).toEqual([]);
     expect(treeModel.find(t, "x")).toBeNull();
-    // ...but the chevron is enabled so the user can expand to load it.
+    // ...and the chevron stays enabled so the user can expand to load it.
     expect((parent as PH).hasChildren).toBe(true);
   });
 
-  it("DOES insert under a LOADED-but-empty parent (children: [])", () => {
-    type PH = TreeNode<{
-      name: string;
-      position?: string;
-      hasChildren?: boolean;
-    }>;
+  it("does NOT materialize a child under an UNLOADED parent (children undefined, hasChildren: true)", () => {
+    const tree: PH[] = [
+      { id: "p", name: "P", position: "a0", hasChildren: true }, // children: undefined
+    ];
+    const node: PH = { id: "x", name: "X", position: "a1" };
+    const t = treeModel.insertByPosition(tree, "p", node);
+    const parent = treeModel.find(t, "p");
+    expect(parent?.children).toBeUndefined();
+    expect(treeModel.find(t, "x")).toBeNull();
+    expect((parent as PH).hasChildren).toBe(true);
+  });
+
+  it("DOES insert under a genuinely-empty parent (children: [], hasChildren: false)", () => {
     const tree: PH[] = [
       { id: "p", name: "P", position: "a0", hasChildren: false, children: [] },
     ];
     const node: PH = { id: "x", name: "X", position: "a1" };
     const t = treeModel.insertByPosition(tree, "p", node);
-    // A loaded (empty) child list is complete, so the node IS inserted.
+    // No server children (`hasChildren: false`), so materializing the first child
+    // is correct — nothing is hidden.
+    expect(treeModel.find(t, "p")?.children?.map((n) => n.id)).toEqual(["x"]);
+  });
+
+  it("DOES insert under a genuinely-empty parent (children undefined, hasChildren: false)", () => {
+    const tree: PH[] = [
+      { id: "p", name: "P", position: "a0", hasChildren: false }, // children: undefined
+    ];
+    const node: PH = { id: "x", name: "X", position: "a1" };
+    const t = treeModel.insertByPosition(tree, "p", node);
     expect(treeModel.find(t, "p")?.children?.map((n) => n.id)).toEqual(["x"]);
   });
 

apps/client/src/features/page/tree/model/tree-model.ts

@@ -43,6 +43,26 @@ export const treeModel = {
     };
   },
 
+  // A branch is "unloaded" when the server says it HAS children (`hasChildren`)
+  // but none are present locally. The canonical unloaded form in this codebase
+  // is `children: []` (produced by `pageToTreeNode` and by `pruneCollapsedChildren`
+  // resetting collapsed branches), NOT `children: undefined` — so a predicate that
+  // only checks `=== undefined` misses the real case and materializes a misleading
+  // partial list (#525). This is the SINGLE source of truth for "should a
+  // fetch/materialize be deferred?", shared by the lazy-load gate (`handleToggle`)
+  // and the realtime insert paths (`insertByPosition` / `placeByPosition`), so they
+  // can never drift apart again. (The local raw `insert` primitive and its DnD/
+  // create-page callers do NOT yet route through this predicate — see #525
+  // follow-up.) A parent WITHOUT `hasChildren` is genuinely empty
+  // (no server children) — inserting its first child is correct, not deferred.
+  isUnloadedBranch<T extends object>(
+    node: TreeNode<T> | null | undefined,
+  ): boolean {
+    if (!node) return false;
+    const hasChildren = (node as { hasChildren?: boolean }).hasChildren === true;
+    return hasChildren && (node.children == null || node.children.length === 0);
+  },
+
   isDescendant<T extends object>(
     tree: TreeNode<T>[],
     ancestorId: string,
@@ -127,14 +147,15 @@ export const treeModel = {
     }
     const parent = treeModel.find(tree, parentId);
     // The parent is in the tree but its children have NOT been lazy-loaded yet
-    // (`children === undefined`, distinct from a loaded-but-empty `[]`). Inserting
+    // (`hasChildren` set + children absent/empty — see `isUnloadedBranch`; the
+    // canonical unloaded form is `children: []`, NOT just `undefined`). Inserting
     // here would MATERIALIZE a misleading partial child list (`[node]`) that
     // defeats the lazy-load gate — which fetches only when children are
     // absent/empty — so the parent's OTHER real children would never load and the
     // moved/added node would be the only one shown (a silent data loss, #159 #1).
     // Instead, leave the children unloaded and just flag `hasChildren` so the
     // chevron appears; expanding fetches the FULL set (including this node).
-    if (parent && parent.children === undefined) {
+    if (parent && treeModel.isUnloadedBranch(parent)) {
       return treeModel.update(
         tree,
         parentId,

apps/client/src/features/websocket/tree-socket-reducers.test.ts

@@ -82,17 +82,19 @@ describe("applyMoveTreeNode", () => {
     ]);
   });
 
-  it("does NOT create a partial child list when the destination is loaded-but-collapsed (children unloaded) — keeps it lazy-loadable (#159)", () => {
-    // `dstCollapsed` is in the tree but its children were never lazy-loaded
-    // (children === undefined). The OLD behavior inserted `src` as the ONLY
-    // child ([src]), which defeated the lazy-load gate and HID the parent's
-    // other real children. Now the move leaves children unloaded (so expanding
-    // fetches the FULL set, including src) and just flags hasChildren.
+  it("does NOT create a partial child list when the destination is loaded-but-collapsed (children unloaded) — keeps it lazy-loadable (#159 #525)", () => {
+    // `dstCollapsed` is in the tree but its children were never lazy-loaded. The
+    // CANONICAL unloaded form here is `hasChildren: true` + `children: []` (from
+    // `pageToTreeNode` / `pruneCollapsedChildren`), NOT `children: undefined`.
+    // The pre-#525 predicate (`children === undefined`) missed this form and
+    // inserted `src` as the ONLY child ([src]), defeating the lazy-load gate and
+    // HIDING the parent's other real children. Now the move leaves children
+    // unloaded (so expanding fetches the FULL set, including src).
     const tree: SpaceTreeNode[] = [
       node("dstCollapsed", {
         position: "a0",
-        hasChildren: false,
-        children: undefined as unknown as SpaceTreeNode[],
+        hasChildren: true,
+        children: [],
       }),
       node("src", { position: "a9" }),
     ];
@@ -105,9 +107,10 @@ describe("applyMoveTreeNode", () => {
       pageData: {},
     });
     const dst = treeModel.find(next, "dstCollapsed");
-    // Children stay unloaded -> the lazy-load gate fetches the FULL set (incl.
-    // src) on expand, rather than showing a misleading partial [src] list.
-    expect(dst?.children).toBeUndefined();
+    // Children stay unloaded ([] not materialized to [src]) -> the lazy-load gate
+    // fetches the FULL set (incl. src) on expand. MUTATION: the pre-#525
+    // `=== undefined` predicate would insert [src] here and redden this.
+    expect(dst?.children).toEqual([]);
     expect(dst?.hasChildren).toBe(true);
     // src moved away from its old root slot (it lives under dstCollapsed
     // server-side and reappears when the parent is expanded/loaded).

apps/server/src/collaboration/constants.ts

@@ -1,10 +1,36 @@
-export const HISTORY_INTERVAL = 5 * 60 * 1000;
-export const HISTORY_FAST_INTERVAL = 60 * 1000;
-export const HISTORY_FAST_THRESHOLD = 5 * 60 * 1000;
-
 // #348 — debounce window for the per-page RAG re-embed job. Repeated saves
 // within this window collapse to a single delayed job (coalesced by a stable
 // jobId), so active editing does not pile up expensive re-embeds (external API
 // + page_embeddings rewrite, concurrency 1). The worker reads the CURRENT page
 // state at run time, so the last content within the window wins.
 export const EMBED_DEBOUNCE_MS = 30 * 1000;
+
+/**
+ * #370 — page-history intentionality tiers. Domain of `page_history.kind`.
+ *   - 'manual' / 'agent'  → Tier 1 versions (intentional points)
+ *   - 'idle' / 'boundary' → Tier 0 autosnapshots (safety net)
+ * A legacy `null` kind is treated as an autosave.
+ */
+export type PageHistoryKind = 'manual' | 'agent' | 'idle' | 'boundary';
+
+/**
+ * #370 — trailing idle-flush windows. A page's pending idle snapshot is
+ * re-armed on every store and fires this long after edits go quiet, so a burst
+ * of edits collapses into a single autosnapshot instead of one-per-store. Human
+ * sessions are noisier and less risky, so they flush less often than the agent.
+ */
+export const IDLE_INTERVAL_USER = 60 * 60 * 1000; // 60m
+export const IDLE_INTERVAL_AGENT = 15 * 60 * 1000; // 15m
+
+/**
+ * #370 — max-wait ceiling for the idle flush. Pure trailing debounce starves the
+ * safety net: hocuspocus stores at least every ~45s, so a CONTINUOUS editing
+ * session would re-arm the trailing timer forever and never take an idle
+ * snapshot until edits finally go quiet (up to IDLE_INTERVAL_USER = 60m). This
+ * ceiling bounds the actual wait from the FIRST edit of a burst, so an idle
+ * snapshot fires at least this often during a long unbroken session — restoring
+ * a recovery point cadence closer to the old heuristic without one-per-store
+ * noise. Mirrors hocuspocus's own maxDebounce idea.
+ */
+export const IDLE_MAX_WAIT_USER = 10 * 60 * 1000; // 10m
+export const IDLE_MAX_WAIT_AGENT = 5 * 60 * 1000; // 5m

apps/server/src/collaboration/extensions/compute-history-job.spec.ts

@@ -1,84 +1,93 @@
+import { computeHistoryJob, resolveSource } from './persistence.extension';
 import {
-  computeHistoryJob,
-  resolveSource,
-} from './persistence.extension';
-import {
-  HISTORY_FAST_INTERVAL,
-  HISTORY_FAST_THRESHOLD,
-  HISTORY_INTERVAL,
+  IDLE_INTERVAL_AGENT,
+  IDLE_INTERVAL_USER,
+  IDLE_MAX_WAIT_AGENT,
+  IDLE_MAX_WAIT_USER,
 } from '../constants';
 
-// A fixed clock + fixed createdAt make pageAge deterministic.
-const NOW = 1_700_000_000_000;
 const PAGE_ID = '550e8400-e29b-41d4-a716-446655440000';
 
-// Build a minimal page whose age (NOW - createdAt) is exactly `ageMs`.
-const pageAged = (ageMs: number) => ({
-  id: PAGE_ID,
-  createdAt: new Date(NOW - ageMs),
-});
+const page = { id: PAGE_ID };
 
-describe('computeHistoryJob', () => {
-  it('agent edit → delay MUST be 0 and job id is source-keyed', () => {
-    // INVARIANT (§15 H2 / persistence.extension): the agent delay MUST stay 0.
-    // The worker re-reads the page row at run time, so any non-zero delay risks
-    // snapshotting content a later human edit has already overwritten. This is
-    // the load-bearing assertion of this spec — do not relax it.
-    const { jobId, delay } = computeHistoryJob(pageAged(0), 'agent', NOW);
-    expect(delay).toBe(0);
-    expect(jobId).toBe(`${PAGE_ID}-agent`);
-  });
-
-  it('agent edit on an OLD page is still delay 0 (age never applies to agents)', () => {
-    // Even when the page is far older than the fast threshold, the agent path
-    // must short-circuit to 0 — age-based debounce is a human-only concern.
-    const { jobId, delay } = computeHistoryJob(
-      pageAged(HISTORY_FAST_THRESHOLD + 60_000),
-      'agent',
-      NOW,
-    );
-    expect(delay).toBe(0);
-    expect(jobId).toBe(`${PAGE_ID}-agent`);
-  });
-
-  it('human edit on a YOUNG page (age < threshold) → fast interval, bare job id', () => {
-    const { jobId, delay } = computeHistoryJob(
-      pageAged(HISTORY_FAST_THRESHOLD - 1),
-      'user',
-      NOW,
-    );
-    expect(delay).toBe(HISTORY_FAST_INTERVAL);
+describe('computeHistoryJob (#370 — shared trailing idle pipeline)', () => {
+  it('human edit → user idle window, bare page.id job', () => {
+    // Humans and the agent now share ONE idle job per page (jobId = page.id).
+    // The agent's old delay=0 fast path is GONE — intentional agent points now
+    // arrive via the explicit save-version signal, not a zero-delay snapshot.
+    const { jobId, delay } = computeHistoryJob(page, 'user');
+    expect(delay).toBe(IDLE_INTERVAL_USER);
     expect(jobId).toBe(PAGE_ID);
   });
 
-  it('human edit on an OLD page (age > threshold) → standard interval', () => {
-    const { jobId, delay } = computeHistoryJob(
-      pageAged(HISTORY_FAST_THRESHOLD + 1),
-      'user',
-      NOW,
-    );
-    expect(delay).toBe(HISTORY_INTERVAL);
+  it('agent edit → agent idle window (shorter), still the bare page.id job', () => {
+    const { jobId, delay } = computeHistoryJob(page, 'agent');
+    expect(delay).toBe(IDLE_INTERVAL_AGENT);
+    // No `-agent` suffix anymore: the agent joins the common idle pipeline.
     expect(jobId).toBe(PAGE_ID);
   });
 
-  it('boundary: pageAge EXACTLY === threshold takes the slow branch (the `<` is strict)', () => {
-    // Off-by-one guard: the condition is `pageAge < HISTORY_FAST_THRESHOLD`, so
-    // an age of exactly the threshold is NOT "fast" — it must use HISTORY_INTERVAL.
-    const { delay } = computeHistoryJob(
-      pageAged(HISTORY_FAST_THRESHOLD),
-      'user',
-      NOW,
-    );
-    expect(delay).toBe(HISTORY_INTERVAL);
+  it('agent flushes sooner than a human', () => {
+    expect(IDLE_INTERVAL_AGENT).toBeLessThan(IDLE_INTERVAL_USER);
   });
 
-  it('treats any non-"agent" source string as human', () => {
-    // resolveSource only ever yields 'agent' | 'user', but guard the contract:
-    // the agent branch keys strictly on === 'agent'.
-    const { jobId, delay } = computeHistoryJob(pageAged(0), 'user', NOW);
-    expect(delay).toBe(HISTORY_FAST_INTERVAL);
+  it('treats any non-"agent" source string as human (keys strictly on === agent)', () => {
+    const { jobId, delay } = computeHistoryJob(page, 'user');
+    expect(delay).toBe(IDLE_INTERVAL_USER);
     expect(jobId).toBe(PAGE_ID);
   });
+
+  // #370 review round-1 WARNING: the max-wait ceiling prevents autosnapshot
+  // starvation during a continuous editing session (the trailing timer would
+  // otherwise re-arm forever and never fire).
+  describe('max-wait ceiling', () => {
+    const T0 = 1_000_000; // arbitrary fixed epoch for deterministic tests
+
+    it('once a burst is armed, delay clamps to the remaining max-wait budget', () => {
+      // 1 minute into the burst the USER interval (60m) far exceeds the remaining
+      // max-wait budget (10m - 1m = 9m), so the delay is clamped DOWN to that
+      // remaining budget — the full interval is NOT used once a ceiling applies.
+      const { delay } = computeHistoryJob(page, 'user', T0, T0 + 60_000);
+      expect(delay).toBe(IDLE_MAX_WAIT_USER - 60_000);
+    });
+
+    it('never waits longer than the max-wait budget from the burst start', () => {
+      // A store arriving right at the ceiling → delay 0 (fire promptly).
+      const { delay } = computeHistoryJob(
+        page,
+        'user',
+        T0,
+        T0 + IDLE_MAX_WAIT_USER,
+      );
+      expect(delay).toBe(0);
+    });
+
+    it('past the ceiling never returns a negative delay', () => {
+      const { delay } = computeHistoryJob(
+        page,
+        'user',
+        T0,
+        T0 + IDLE_MAX_WAIT_USER + 5 * 60_000,
+      );
+      expect(delay).toBe(0);
+    });
+
+    it('the agent ceiling is shorter than the user ceiling', () => {
+      expect(IDLE_MAX_WAIT_AGENT).toBeLessThan(IDLE_MAX_WAIT_USER);
+      const { delay } = computeHistoryJob(
+        page,
+        'agent',
+        T0,
+        T0 + IDLE_MAX_WAIT_AGENT,
+      );
+      expect(delay).toBe(0);
+    });
+
+    it('without a burstStart there is no ceiling (backward-compatible)', () => {
+      expect(computeHistoryJob(page, 'user').delay).toBe(IDLE_INTERVAL_USER);
+      expect(computeHistoryJob(page, 'agent').delay).toBe(IDLE_INTERVAL_AGENT);
+    });
+  });
 });
 
 describe('resolveSource (truth table)', () => {

apps/server/src/collaboration/extensions/persistence-store.spec.ts

@@ -40,11 +40,12 @@ describe('PersistenceExtension.onStoreDocument — Approach-A boundary snapshot'
   let pageHistoryRepo: {
     saveHistory: jest.Mock;
     findPageLastHistory: jest.Mock;
+    updateHistoryKind: jest.Mock;
   };
   let aiQueue: { add: jest.Mock };
-  let historyQueue: { add: jest.Mock };
+  let historyQueue: { add: jest.Mock; remove: jest.Mock };
   let notificationQueue: { add: jest.Mock };
-  let collabHistory: { addContributors: jest.Mock };
+  let collabHistory: { addContributors: jest.Mock; popContributors: jest.Mock };
   let transclusionService: {
     syncPageTransclusions: jest.Mock;
     syncPageReferences: jest.Mock;
@@ -93,13 +94,22 @@ describe('PersistenceExtension.onStoreDocument — Approach-A boundary snapshot'
     pageHistoryRepo = {
       saveHistory: jest.fn().mockImplementation(async () => {
         callOrder.push('saveHistory');
+        return { id: 'history-1' };
       }),
       findPageLastHistory: jest.fn().mockResolvedValue(null),
+      updateHistoryKind: jest.fn().mockResolvedValue(undefined),
     };
     aiQueue = { add: jest.fn().mockResolvedValue(undefined) };
-    historyQueue = { add: jest.fn().mockResolvedValue(undefined) };
+    historyQueue = {
+      add: jest.fn().mockResolvedValue(undefined),
+      // #370 — enqueuePageHistory now removes any pending idle job before re-adding.
+      remove: jest.fn().mockResolvedValue(undefined),
+    };
     notificationQueue = { add: jest.fn().mockResolvedValue(undefined) };
-    collabHistory = { addContributors: jest.fn().mockResolvedValue(undefined) };
+    collabHistory = {
+      addContributors: jest.fn().mockResolvedValue(undefined),
+      popContributors: jest.fn().mockResolvedValue([]),
+    };
     transclusionService = {
       syncPageTransclusions: jest.fn().mockResolvedValue(undefined),
       syncPageReferences: jest.fn().mockResolvedValue(undefined),
@@ -165,6 +175,50 @@ describe('PersistenceExtension.onStoreDocument — Approach-A boundary snapshot'
     expect(pageRepo.updatePage.mock.calls[0][0].lastUpdatedSource).toBe('user');
   });
 
+  // #370 review round-1 SUGGESTION: the boundary was GENERALIZED from a
+  // user→agent special-case to ANY lastUpdatedSource transition. These pin the
+  // generalized behaviour it was rebuilt for.
+  describe('generalized boundary — any source transition', () => {
+    // Same persisted page but with an explicit prior source.
+    const pageWithPriorSource = (prior: string | null) => ({
+      ...persistedHumanPage('NEW CONTENT'),
+      lastUpdatedSource: prior,
+    });
+
+    it('agent→user transition fires the boundary (pins the prior agent revision)', async () => {
+      const document = ydocFor(doc('NEW CONTENT'));
+      pageRepo.findById.mockResolvedValue(pageWithPriorSource('agent'));
+      pageHistoryRepo.findPageLastHistory.mockResolvedValue(null);
+
+      await ext.onStoreDocument(buildData(document, 'user') as any);
+
+      expect(pageHistoryRepo.saveHistory).toHaveBeenCalledTimes(1);
+      expect(callOrder).toEqual(['saveHistory', 'updatePage']);
+      expect(pageRepo.updatePage.mock.calls[0][0].lastUpdatedSource).toBe('user');
+    });
+
+    it('git→user transition fires the boundary (git-sync overwrite is a source change)', async () => {
+      const document = ydocFor(doc('NEW CONTENT'));
+      pageRepo.findById.mockResolvedValue(pageWithPriorSource('git'));
+      pageHistoryRepo.findPageLastHistory.mockResolvedValue(null);
+
+      await ext.onStoreDocument(buildData(document, 'user') as any);
+
+      expect(pageHistoryRepo.saveHistory).toHaveBeenCalledTimes(1);
+      expect(callOrder).toEqual(['saveHistory', 'updatePage']);
+    });
+
+    it('a null prior source (first-ever edit) does NOT fire the boundary', async () => {
+      const document = ydocFor(doc('NEW CONTENT'));
+      pageRepo.findById.mockResolvedValue(pageWithPriorSource(null));
+
+      await ext.onStoreDocument(buildData(document, 'agent') as any);
+
+      expect(pageHistoryRepo.saveHistory).not.toHaveBeenCalled();
+      expect(pageRepo.updatePage).toHaveBeenCalledTimes(1);
+    });
+  });
+
   it('idempotency: unchanged content → no updatePage, no history, no queues', async () => {
     // The Y.Doc content equals the persisted content deeply → early skip.
     // A Y.Doc round-trip normalizes attrs (e.g. paragraph indent), so derive
@@ -479,4 +533,231 @@ describe('PersistenceExtension.onStoreDocument — Approach-A boundary snapshot'
     // Contributors keyed by the UUID so they match the PAGE_HISTORY job (page.id).
     expect(collabHistory.addContributors.mock.calls[0][0]).toBe(PAGE_ID);
   });
+
+  // #370 — explicit save-version (Cmd+S / agent save tool) over the stateless
+  // seam. The tier is derived from the SIGNED connection actor, the store path
+  // is reused, and promote-not-dup avoids duplicating heavy content rows.
+  describe('save-version (#370)', () => {
+    const emitSave = (document: any, actor: 'user' | 'agent') =>
+      ext.onStateless({
+        connection: {
+          readOnly: false,
+          context: { user: { id: USER_ID, name: 'Alice' }, actor },
+        } as any,
+        documentName: `page.${PAGE_ID}`,
+        document: document as any,
+        payload: JSON.stringify({ type: 'save-version' }),
+      } as any);
+
+    // findById returns a page whose content already equals the live doc, so the
+    // store path is a no-op and we isolate the versioning decision.
+    const pageMatchingDoc = (document: any) => ({
+      ...persistedHumanPage('IGNORED'),
+      content: TiptapTransformer.fromYdoc(document, 'default'),
+    });
+
+    it('human save with no prior snapshot → writes a manual version + broadcasts', async () => {
+      const document = ydocFor(doc('VERSION ME'));
+      pageRepo.findById.mockResolvedValue(pageMatchingDoc(document));
+      pageHistoryRepo.findPageLastHistory.mockResolvedValue(null);
+
+      await emitSave(document, 'user');
+
+      expect(pageHistoryRepo.saveHistory).toHaveBeenCalledTimes(1);
+      expect(pageHistoryRepo.saveHistory.mock.calls[0][1]).toEqual(
+        expect.objectContaining({ kind: 'manual' }),
+      );
+      // The pending idle autosnapshot is cancelled by the explicit version.
+      expect(historyQueue.remove).toHaveBeenCalledWith(PAGE_ID);
+      const msg = JSON.parse(
+        (document as any).broadcastStateless.mock.calls[(document as any).broadcastStateless.mock.calls.length - 1][0],
+      );
+      expect(msg).toMatchObject({
+        type: 'version.saved',
+        kind: 'manual',
+        alreadySaved: false,
+      });
+    });
+
+    it('agent save derives kind=agent from the signed actor', async () => {
+      const document = ydocFor(doc('AGENT VERSION'));
+      pageRepo.findById.mockResolvedValue(pageMatchingDoc(document));
+      pageHistoryRepo.findPageLastHistory.mockResolvedValue(null);
+
+      await emitSave(document, 'agent');
+
+      expect(pageHistoryRepo.saveHistory.mock.calls[pageHistoryRepo.saveHistory.mock.calls.length - 1][1]).toEqual(
+        expect.objectContaining({ kind: 'agent' }),
+      );
+    });
+
+    it('promote-not-dup: latest snapshot is an autosave with identical content → upgrades in place', async () => {
+      const document = ydocFor(doc('SAME'));
+      const page = pageMatchingDoc(document);
+      pageRepo.findById.mockResolvedValue(page);
+      pageHistoryRepo.findPageLastHistory.mockResolvedValue({
+        id: 'auto-1',
+        content: page.content,
+        kind: 'idle',
+      });
+
+      await emitSave(document, 'user');
+
+      // No heavy new content row — the existing autosave is promoted to manual.
+      expect(pageHistoryRepo.updateHistoryKind).toHaveBeenCalledWith(
+        'auto-1',
+        'manual',
+        expect.anything(),
+      );
+      expect(pageHistoryRepo.saveHistory).not.toHaveBeenCalled();
+      const msg = JSON.parse(
+        (document as any).broadcastStateless.mock.calls[(document as any).broadcastStateless.mock.calls.length - 1][0],
+      );
+      expect(msg).toMatchObject({ historyId: 'auto-1', alreadySaved: false });
+    });
+
+    it('no-op when the latest snapshot is already a manual version of this content', async () => {
+      const document = ydocFor(doc('ALREADY SAVED'));
+      const page = pageMatchingDoc(document);
+      pageRepo.findById.mockResolvedValue(page);
+      pageHistoryRepo.findPageLastHistory.mockResolvedValue({
+        id: 'ver-1',
+        content: page.content,
+        kind: 'manual',
+      });
+
+      await emitSave(document, 'user');
+
+      expect(pageHistoryRepo.updateHistoryKind).not.toHaveBeenCalled();
+      expect(pageHistoryRepo.saveHistory).not.toHaveBeenCalled();
+      const msg = JSON.parse(
+        (document as any).broadcastStateless.mock.calls[(document as any).broadcastStateless.mock.calls.length - 1][0],
+      );
+      expect(msg).toMatchObject({ alreadySaved: true, kind: 'manual' });
+    });
+
+    it('a read-only connection cannot save a version', async () => {
+      const document = ydocFor(doc('READER'));
+      pageRepo.findById.mockResolvedValue(pageMatchingDoc(document));
+
+      await ext.onStateless({
+        connection: {
+          readOnly: true,
+          context: { user: { id: USER_ID }, actor: 'user' },
+        } as any,
+        documentName: `page.${PAGE_ID}`,
+        document: document as any,
+        payload: JSON.stringify({ type: 'save-version' }),
+      } as any);
+
+      expect(pageHistoryRepo.saveHistory).not.toHaveBeenCalled();
+      expect(pageHistoryRepo.updateHistoryKind).not.toHaveBeenCalled();
+    });
+
+    // #370 F8-twin — a COMMIT abort (serialization/deadlock/conn-drop) rejects
+    // OUTSIDE the tx callback, AFTER the destructive popContributors (SPOP) and
+    // saveHistory ran but the INSERT rolled back. onStateless has no retry, so
+    // the outer catch MUST re-add (SADD) the popped set or attribution is lost
+    // irrecoverably. MUTATION: drop the outer catch → addContributors is never
+    // called → this reddens.
+    it('restores popped contributors when the commit aborts after the callback', async () => {
+      const document = ydocFor(doc('VERSION ME'));
+      pageRepo.findById.mockResolvedValue(pageMatchingDoc(document));
+      // No matching snapshot → fresh version branch → pops contributors.
+      pageHistoryRepo.findPageLastHistory.mockResolvedValue(null);
+      collabHistory.popContributors.mockResolvedValue(['u1', 'u2']);
+
+      // A db whose commit REJECTS after the callback body resolved: the SPOP and
+      // saveHistory already ran, then the tx aborts. onStoreDocument's flush uses
+      // the same db but its content matches (no-op branch) and its own retry loop
+      // swallows the throw, so only the versioning tx exercises the restore.
+      const commitFailingDb = {
+        transaction: () => ({
+          execute: async (fn: (trx: any) => Promise<any>) => {
+            await fn(trxStub);
+            throw new Error('commit aborted (serialization_failure)');
+          },
+        }),
+      };
+      const ext2 = new PersistenceExtension(
+        pageRepo as any,
+        pageHistoryRepo as any,
+        commitFailingDb as any,
+        aiQueue as any,
+        historyQueue as any,
+        notificationQueue as any,
+        collabHistory as any,
+        transclusionService as any,
+      );
+      jest.spyOn(ext2['logger'], 'debug').mockImplementation(() => undefined);
+      jest.spyOn(ext2['logger'], 'warn').mockImplementation(() => undefined);
+      jest.spyOn(ext2['logger'], 'error').mockImplementation(() => undefined);
+
+      await expect(
+        ext2.onStateless({
+          connection: {
+            readOnly: false,
+            context: { user: { id: USER_ID, name: 'Alice' }, actor: 'user' },
+          } as any,
+          documentName: `page.${PAGE_ID}`,
+          document: document as any,
+          payload: JSON.stringify({ type: 'save-version' }),
+        } as any),
+      ).rejects.toThrow();
+
+      // Attribution preserved: the popped set is SADD-restored, keyed by the page
+      // UUID it was popped under.
+      expect(collabHistory.addContributors).toHaveBeenCalledWith(PAGE_ID, [
+        'u1',
+        'u2',
+      ]);
+    });
+
+    // #370 #260 — for a `page.<slugId>` document the idle job is armed under the
+    // page UUID (computeHistoryJob's jobId = page.id), so the supersede-remove
+    // must target page.id, not the raw slugId doc-name id, or it silently misses.
+    it('cancels the superseded idle job by the page UUID for a slugId doc', async () => {
+      const SLUG = 'slug-1'; // persistedHumanPage.slugId
+      const document = ydocFor(doc('VERSION ME'));
+      pageRepo.findById.mockResolvedValue(pageMatchingDoc(document));
+      pageHistoryRepo.findPageLastHistory.mockResolvedValue(null);
+
+      await ext.onStateless({
+        connection: {
+          readOnly: false,
+          context: { user: { id: USER_ID, name: 'Alice' }, actor: 'user' },
+        } as any,
+        documentName: `page.${SLUG}`,
+        document: document as any,
+        payload: JSON.stringify({ type: 'save-version' }),
+      } as any);
+
+      // remove() keyed by the UUID (the real jobId), never the slugId.
+      expect(historyQueue.remove).toHaveBeenCalledWith(PAGE_ID);
+      expect(historyQueue.remove).not.toHaveBeenCalledWith(SLUG);
+    });
+  });
+
+  // #370 — the in-memory idle-burst marker must be dropped on doc unload (like
+  // its sibling per-document maps) or it grows unbounded for every page that was
+  // edited but never manually saved. MUTATION: drop the afterUnloadDocument
+  // delete → the entry survives → this reddens.
+  describe('idleBurstStart housekeeping', () => {
+    it('afterUnloadDocument clears the idle-burst marker armed by a store', async () => {
+      const document = ydocFor(doc('EDIT'));
+      pageRepo.findById.mockResolvedValue(persistedHumanPage('EDIT'));
+
+      await ext.onStoreDocument(buildData(document, 'user') as any);
+
+      const map = ext['idleBurstStart'] as Map<string, number>;
+      // Keyed by documentName (buildData uses `page.${PAGE_ID}`).
+      expect(map.has(`page.${PAGE_ID}`)).toBe(true);
+
+      await ext.afterUnloadDocument({
+        documentName: `page.${PAGE_ID}`,
+      } as any);
+
+      expect(map.has(`page.${PAGE_ID}`)).toBe(false);
+    });
+  });
 });

apps/server/src/collaboration/extensions/persistence.extension.ts

@@ -37,9 +37,11 @@ import { Page } from '@docmost/db/types/entity.types';
 import { CollabHistoryService } from '../services/collab-history.service';
 import {
   EMBED_DEBOUNCE_MS,
-  HISTORY_FAST_INTERVAL,
-  HISTORY_FAST_THRESHOLD,
-  HISTORY_INTERVAL,
+  IDLE_INTERVAL_AGENT,
+  IDLE_INTERVAL_USER,
+  IDLE_MAX_WAIT_AGENT,
+  IDLE_MAX_WAIT_USER,
+  PageHistoryKind,
 } from '../constants';
 import { TransclusionService } from '../../core/page/transclusion/transclusion.service';
 import {
@@ -56,6 +58,16 @@ import { hasTransclusionFamilyNodes } from '../../core/page/transclusion/utils/t
  */
 export const INTENTIONAL_CLEAR_MESSAGE_TYPE = 'intentional-clear';
 
+/**
+ * #370 — wire format of the client→server "save a version" signal. Sent by the
+ * human (Cmd+S / Save button) and by the agent's explicit save tool over the
+ * SAME stateless channel. The intentionality tier ('manual' vs 'agent') is
+ * derived SERVER-SIDE from the signed connection actor, never from this
+ * payload, so a version's type is unforgeable. The document is taken from the
+ * connection (not the payload), so the signal cannot be aimed at another page.
+ */
+export const SAVE_VERSION_MESSAGE_TYPE = 'save-version';
+
 /**
  * #251 — how long an intentional-clear signal stays "pending" before it is
  * ignored. The signal is set on the clearing keystroke but consumed by the
@@ -92,35 +104,39 @@ export function resolveSource(
 }
 
 /**
- * Compute the BullMQ job id + delay for a page-history snapshot job. Pure so
- * the data-loss-sensitive timing arithmetic is unit-testable; `now` is injected
- * (caller passes `Date.now()`) for determinism.
+ * #370 — compute the BullMQ job id + delay for a page's trailing idle-flush
+ * autosnapshot. Pure so the timing is unit-testable.
  *
- * - Agent edits: delay 0 and a source-keyed job id `${page.id}-agent`. The
- *   delay MUST stay 0 — the worker re-reads the page row at run time, so any
- *   delay risks reading content a later human edit has already overwritten
- *   (mis-tagged snapshot). 0 minimizes that window. The `-agent` suffix keeps
- *   the job from coalescing with the bare-page.id human job.
- * - Human edits: age-based debounce so rapid human edits coalesce into one
- *   snapshot; job id is the bare `page.id`.
- *
- * BullMQ forbids ':' in custom job ids (Redis key separator), so '-' is used;
- * page.id is a UUID, so `${page.id}-agent` cannot collide with a human job.
+ * Both humans and the agent now share ONE idle pipeline (the agent's old
+ * `delay=0` fast path is gone — intentional agent points arrive via the
+ * explicit save-version signal instead). The job id is the bare `page.id`, so a
+ * page has at most one pending idle job; the caller removes-and-re-adds it on
+ * every store to keep it debounced to the trailing edge of an edit burst. The
+ * window differs by source only: the agent flushes sooner than a human.
  */
 export function computeHistoryJob(
-  page: Pick<Page, 'id' | 'createdAt'>,
+  page: Pick<Page, 'id'>,
   source: string,
-  now: number,
+  // Epoch ms of the FIRST edit in the current burst (when the pending idle job
+  // was first armed). Used to enforce the max-wait ceiling so a continuous
+  // editing session cannot re-arm the trailing timer forever. `now` is injectable
+  // for tests; both default to a live clock / no ceiling when omitted.
+  burstStart?: number,
+  now: number = Date.now(),
 ): { jobId: string; delay: number } {
   const isAgent = source === 'agent';
-  const pageAge = now - new Date(page.createdAt).getTime();
-  const delay = isAgent
-    ? 0
-    : pageAge < HISTORY_FAST_THRESHOLD
-      ? HISTORY_FAST_INTERVAL
-      : HISTORY_INTERVAL;
-  const jobId = isAgent ? `${page.id}-agent` : page.id;
-  return { jobId, delay };
+  const interval = isAgent ? IDLE_INTERVAL_AGENT : IDLE_INTERVAL_USER;
+  const maxWait = isAgent ? IDLE_MAX_WAIT_AGENT : IDLE_MAX_WAIT_USER;
+
+  let delay = interval;
+  if (burstStart !== undefined) {
+    // Time already elapsed since the burst's first edit; the snapshot must fire
+    // no later than `maxWait` after that, so shrink the trailing delay to the
+    // remaining budget (never negative, so BullMQ fires it promptly).
+    const remaining = burstStart + maxWait - now;
+    delay = Math.max(0, Math.min(interval, remaining));
+  }
+  return { jobId: page.id, delay };
 }
 
 @Injectable()
@@ -132,6 +148,28 @@ export class PersistenceExtension implements Extension {
   // coalescing window" per document and OR it across all edits in the window,
   // so the snapshot is marked 'agent' regardless of who wrote last.
   private agentTouched: Map<string, boolean> = new Map();
+  // #370 — epoch ms of the FIRST edit in the current idle-flush burst. Keyed by
+  // documentName (like its sibling per-document maps above), NOT by page.id, so
+  // it can be cleaned in afterUnloadDocument alongside `contributors` /
+  // `agentTouched` / `intentionalClear` when the doc unloads — otherwise any page
+  // that was edited but never manually saved (the common case) would keep its
+  // entry forever and the Map would grow unbounded in this long-lived process.
+  // Set when the pending idle job is first armed (empty entry), read to enforce
+  // the max-wait ceiling in computeHistoryJob, and cleared on doc unload or when
+  // a manual save cancels the idle job so the next burst starts a fresh window.
+  //
+  // Single-process assumption (like `contributors` / `agentTouched` above): this
+  // lives only in THIS collab process's memory. A restart, or a page's ownership
+  // moving to another node, loses the burst-start marker. Consequence: a burst
+  // that spans the restart looks like a fresh burst to the surviving process, so
+  // its max-wait ceiling is re-anchored to the first post-restart edit — a single
+  // continuous session straddling a restart can therefore wait up to ~2× the cap
+  // for its idle snapshot (once for the lost pre-restart window, once for the new
+  // one). Bounded and benign (it only DELAYS a safety-net autosnapshot; manual
+  // saves are unaffected and the next quiet period always flushes), but the
+  // assumption and its consequence are recorded here so no one mistakes the
+  // in-memory marker for a durable, cross-process guarantee.
+  private idleBurstStart: Map<string, number> = new Map();
   // #251 — per-document "intentional clear pending" flags. Keyed by
   // documentName, value = expiry timestamp (ms). Set by onStateless when the
   // client reports a deliberate clear; consumed once by the next
@@ -363,20 +401,19 @@ export class PersistenceExtension implements Extension {
             //this.logger.debug('Contributors error:' + err?.['message']);
           }
 
-          // Approach A — boundary snapshot before the agent's first edit.
-          // When this store is the agent's and the page's currently persisted
-          // state was authored by a human, pin that human state as its own
-          // history version BEFORE the agent overwrites it. `page` still holds
-          // the OLD content/provenance here, so saveHistory(page) captures the
-          // pre-agent state tagged 'user'. The agent's new content is
-          // snapshotted later by the debounced PAGE_HISTORY job ('agent'). Skip
-          // if the prior state is already agent-authored (boundary already
-          // pinned on the user->agent transition), if the page is effectively
-          // empty, or if the latest existing snapshot already equals this human
-          // state (avoid duplicates).
+          // #370 — boundary snapshot on ANY source transition. When the store
+          // flips the page's provenance (user↔agent↔git), pin the OUTGOING
+          // state as its own history version BEFORE the incoming source
+          // overwrites it. `page` still holds the OLD content/provenance here,
+          // so saveHistory(page) captures the pre-transition state tagged with
+          // its own source, kind='boundary'. The incoming content is snapshotted
+          // later by the debounced idle job. Skip if the page is effectively
+          // empty or if the latest existing snapshot already equals this state
+          // (the shared isDeepStrictEqual gate — avoids duplicates). Generalizing
+          // beyond the old user→agent special-case also covers git-sync for free.
           if (
-            lastUpdatedSource === 'agent' &&
-            page.lastUpdatedSource !== 'agent'
+            page.lastUpdatedSource &&
+            page.lastUpdatedSource !== lastUpdatedSource
           ) {
             // pageHistory.pageId is uuid-typed; use page.id (never the doc-name
             // slugId) so a `page.<slugId>` doc cannot throw 22P02 here (#260).
@@ -384,15 +421,13 @@ export class PersistenceExtension implements Extension {
               page.id,
               { includeContent: true, trx },
             );
-            const humanBaselineMissing =
+            const baselineMissing =
               !lastHistory ||
               !isDeepStrictEqual(lastHistory.content, page.content);
-            if (
-              !isEmptyParagraphDoc(page.content as any) &&
-              humanBaselineMissing
-            ) {
+            if (!isEmptyParagraphDoc(page.content as any) && baselineMissing) {
               await this.pageHistoryRepo.saveHistory(page, {
                 contributorIds: page.contributorIds ?? undefined,
+                kind: 'boundary',
                 trx,
               });
             }
@@ -522,7 +557,7 @@ export class PersistenceExtension implements Extension {
         { jobId: `embed-${page.id}`, delay: EMBED_DEBOUNCE_MS },
       );
 
-      await this.enqueuePageHistory(page, lastUpdatedSource);
+      await this.enqueuePageHistory(page, documentName, lastUpdatedSource);
     }
 
     // #402 — report the serialized size for the store histogram's size_bucket.
@@ -554,6 +589,14 @@ export class PersistenceExtension implements Extension {
       return; // unrelated / malformed stateless message
     }
 
+    // #370 — explicit "save a version" (human Cmd+S / agent save tool). Edit
+    // rights are already enforced by the readOnly reject above (a reader can't
+    // create a version), exactly as intentional-clear requires.
+    if (message?.type === SAVE_VERSION_MESSAGE_TYPE) {
+      await this.handleSaveVersion(data);
+      return;
+    }
+
     if (message?.type !== INTENTIONAL_CLEAR_MESSAGE_TYPE) return;
 
     this.intentionalClear.set(
@@ -562,6 +605,160 @@ export class PersistenceExtension implements Extension {
     );
   }
 
+  /**
+   * #370 — persist an intentional version from the live in-memory ydoc.
+   *
+   * One stateless path serves BOTH the human and the agent; the tier is derived
+   * SERVER-SIDE from the signed connection actor ('agent' → 'agent', anything
+   * else → 'manual'), so the version type cannot be spoofed by the client. We
+   * take the fresh ydoc from the collab process memory and run it through the
+   * EXISTING store path first (so pages.content/ydoc reflect the exact content
+   * being versioned — a REST endpoint would race the up-to-10s-stale page row),
+   * then snapshot it into page_history with the intentional kind.
+   *
+   * Promote-not-dup: if the latest history row already holds this exact content
+   * and it is an autosave (idle/boundary/legacy-null), upgrade its kind in place
+   * instead of duplicating a heavy content row; if it is already 'manual', it is
+   * a no-op (the client shows an "already saved" toast). Otherwise a fresh
+   * version row is written, popping the aggregated contributors from Redis.
+   */
+  private async handleSaveVersion(data: onStatelessPayload): Promise<void> {
+    const { connection, document, documentName } = data;
+    const context = connection?.context;
+    const pageId = getPageId(documentName);
+    // Unforgeable: 'agent' only for a signed agent connection, else 'manual'.
+    const kind: PageHistoryKind =
+      context?.actor === 'agent' ? 'agent' : 'manual';
+
+    // Flush the live ydoc through the normal store path so the page row + ydoc
+    // hold exactly what we are about to version (also fires the idle enqueue we
+    // supersede below, plus any source-transition boundary). onStoreDocument
+    // only needs document/documentName/context.
+    await this.onStoreDocument({
+      document,
+      documentName,
+      context,
+    } as onStoreDocumentPayload);
+
+    let result:
+      | { historyId: string; kind: PageHistoryKind; alreadySaved: boolean }
+      | undefined;
+
+    // #370 F8-twin — the contributor set popped from Redis (destructive SPOP)
+    // must be restored if the version row does not durably land. The inner
+    // try/catch below only covers a throw INSIDE the callback; but executeTx
+    // COMMITS after the callback, so a commit-abort (serialization/deadlock/
+    // connection drop — the transient class the epic retries in the processor)
+    // rejects OUTSIDE the callback, after saveHistory already ran and the SPOP
+    // already happened, while the INSERT rolls back. onStateless does NOT retry,
+    // so an unrestored pop is a one-shot irrecoverable attribution loss (the
+    // processor got exactly this fix: poppedForRestore + an outer catch). We
+    // track the popped set here (keyed by the page UUID it was popped by — never
+    // the doc-name id, which may be a slugId, #260) and restore it in the outer
+    // catch. addContributors is an idempotent Redis SADD, so a double-restore is
+    // harmless. versionedPageId is also reused below to remove the superseded
+    // idle job by its real jobId (page.id).
+    let poppedForRestore: string[] = [];
+    let versionedPageId: string | undefined;
+
+    try {
+      await executeTx(this.db, async (trx) => {
+        const page = await this.pageRepo.findById(pageId, {
+          withLock: true,
+          includeContent: true,
+          trx,
+        });
+        if (!page) return;
+        versionedPageId = page.id;
+        // Never version an effectively-empty page (mirrors the processor's
+        // first-history guard); there is nothing intentional to pin.
+        if (isEmptyParagraphDoc(page.content as any)) return;
+
+        const lastHistory = await this.pageHistoryRepo.findPageLastHistory(
+          page.id,
+          { includeContent: true, trx },
+        );
+
+        if (
+          lastHistory &&
+          isDeepStrictEqual(lastHistory.content, page.content)
+        ) {
+          // Content is already snapshotted. Promote-not-dup.
+          if (lastHistory.kind === 'manual') {
+            result = {
+              historyId: lastHistory.id,
+              kind: 'manual',
+              alreadySaved: true,
+            };
+            return;
+          }
+          await this.pageHistoryRepo.updateHistoryKind(
+            lastHistory.id,
+            kind,
+            trx,
+          );
+          result = { historyId: lastHistory.id, kind, alreadySaved: false };
+          return;
+        }
+
+        // Fresh version row. Pop the contributors aggregated since the last
+        // snapshot (SPOP); restore them if the write fails so they aren't lost.
+        const contributorIds = await this.collabHistory.popContributors(
+          page.id,
+        );
+        poppedForRestore = contributorIds;
+        try {
+          const saved = await this.pageHistoryRepo.saveHistory(page, {
+            contributorIds,
+            kind,
+            trx,
+          });
+          result = { historyId: saved.id, kind, alreadySaved: false };
+        } catch (err) {
+          await this.collabHistory.addContributors(page.id, contributorIds);
+          poppedForRestore = [];
+          throw err;
+        }
+      });
+    } catch (err) {
+      // A throw here means the tx did NOT commit (callback threw, or the commit
+      // itself failed and rolled back). If we popped contributors and the inner
+      // catch did not already restore them, restore now so attribution is not
+      // lost — onStateless has no retry to recover it. Restore by the page UUID
+      // the pop was keyed under (versionedPageId is always set before the pop).
+      if (poppedForRestore.length && versionedPageId) {
+        await this.collabHistory.addContributors(
+          versionedPageId,
+          poppedForRestore,
+        );
+      }
+      throw err;
+    }
+
+    // Housekeeping: this explicit version supersedes the page's pending idle
+    // autosnapshot, so cancel it and end the current idle burst so the next edit
+    // starts a fresh max-wait window. Remove the idle job by its REAL jobId
+    // (page.id UUID — computeHistoryJob arms it under page.id), not the raw
+    // doc-name id which may be a slugId for a `page.<slugId>` doc (#260), or the
+    // remove silently misses. The burst marker is keyed by documentName (like its
+    // sibling per-document maps), and is also cleaned in afterUnloadDocument.
+    if (versionedPageId) {
+      await this.historyQueue.remove(versionedPageId).catch(() => undefined);
+    }
+    this.idleBurstStart.delete(documentName);
+
+    if (result) {
+      document.broadcastStateless(
+        JSON.stringify({
+          type: 'version.saved',
+          historyId: result.historyId,
+          kind: result.kind,
+          alreadySaved: result.alreadySaved,
+        }),
+      );
+    }
+  }
+
   async onChange(data: onChangePayload) {
     const documentName = data.documentName;
     const userId = data.context?.user?.id;
@@ -586,6 +783,10 @@ export class PersistenceExtension implements Extension {
     this.contributors.delete(documentName);
     this.agentTouched.delete(documentName);
     this.intentionalClear.delete(documentName);
+    // #370 — drop the idle-burst marker with the other per-document maps so it
+    // cannot accumulate across the process lifetime for never-manually-saved
+    // pages. The pending idle job (if any) is a self-expiring BullMQ delayed job.
+    this.idleBurstStart.delete(documentName);
   }
 
   private consumeContributors(documentName: string): string[] {
@@ -617,19 +818,80 @@ export class PersistenceExtension implements Extension {
 
   private async enqueuePageHistory(
     page: Page,
+    documentName: string,
     lastUpdatedSource: string,
   ): Promise<void> {
-    // Job id + delay arithmetic lives in the pure `computeHistoryJob` (see its
-    // doc comment for the agent-delay-0 / age-based-debounce invariants).
+    // #370 — trailing idle debounce with a max-wait ceiling. One pending idle
+    // job per page (jobId = page.id); on every store we remove the pending
+    // delayed job and re-add it, so the snapshot lands `delay` after edits go
+    // quiet rather than once per store (precedent: workspace.service.ts).
+    // remove() on a delayed job simply deletes it (0 if absent, no throw); if the
+    // job is already ACTIVE and the remove is a no-op, the add still de-dups and
+    // the processor's isDeepStrictEqual gate collapses the duplicate content.
+    //
+    // The FIRST arm of a burst records `burstStart`; computeHistoryJob shrinks
+    // the delay to the remaining max-wait budget from that point, so a continuous
+    // session cannot re-arm the trailing timer forever and starve the snapshot.
+    // A burst marker older than THIS TIER's max-wait means the previous idle job
+    // has already fired — start a fresh window instead of firing immediately on
+    // the next edit. Must use the SAME source-specific max-wait computeHistoryJob
+    // uses (agent 5m / user 10m): a hardcoded USER ceiling would leave an agent
+    // burst's marker stale for 5..10m, forcing delay=0 on every store in that
+    // window and writing one idle row per store — exactly the per-store bloat the
+    // debounce exists to prevent, on the continuous-agent path.
+    const maxWait =
+      lastUpdatedSource === 'agent' ? IDLE_MAX_WAIT_AGENT : IDLE_MAX_WAIT_USER;
+    const now = Date.now();
+    // Keyed by documentName (see the map declaration) so afterUnloadDocument can
+    // clean it; the queue jobId stays page.id (computeHistoryJob) as required.
+    let burstStart = this.idleBurstStart.get(documentName);
+    if (burstStart === undefined || now - burstStart >= maxWait) {
+      burstStart = now;
+      this.idleBurstStart.set(documentName, burstStart);
+    }
+
     const { jobId, delay } = computeHistoryJob(
       page,
       lastUpdatedSource,
-      Date.now(),
+      burstStart,
+      now,
     );
 
+    // remove-then-add trailing-debounce idiom, and its ONE race. We delete the
+    // pending delayed job and re-add it under the same jobId so the timer resets
+    // to the trailing edge of the burst. The race is the small window between
+    // these two awaits: if the delayed job's `delay` elapses in that gap it goes
+    // ACTIVE, and then:
+    //   - remove() on an active/locked job is a no-op (BullMQ won't yank a job a
+    //     worker holds), and our `.catch(() => undefined)` swallows that too; and
+    //   - add() with a jobId that already exists (the now-active job's id) is
+    //     DROPPED by BullMQ — a duplicate add is a no-op.
+    // So this store fails to re-arm the trailing job: the just-fired snapshot
+    // captured content up to the moment it went active, and THIS edit is left
+    // without a pending trailing job. It is bounded and self-healing — the NEXT
+    // store re-arms a fresh delayed job (the id is free again once the active job
+    // completes / removeOnComplete frees it), and the processor's
+    // isDeepStrictEqual gate collapses any content-identical duplicate. The only
+    // uncovered case is when the racing store was the LAST in the session: the
+    // tail edits made after the job went active get NO trailing snapshot until
+    // the next edit re-arms one. That is an acceptable safety-net gap (a manual
+    // Save, a source-transition boundary, or simply the next edit all still cover
+    // it), which is why the reviewer accepts documenting it here rather than
+    // adding a post-add "did the add actually arm a job?" re-check.
+    //
+    // NOTE — do NOT "unify" this with the neighbouring embed-debounce idiom
+    // (aiQueue.add of PAGE_CONTENT_UPDATED above): that one uses a STABLE jobId
+    // and NO remove(), relying purely on BullMQ coalescing a repeated add under
+    // the same id, because a re-embed only needs to eventually run once on the
+    // latest content and re-anchoring its delay on every keystroke is undesirable.
+    // THIS idiom deliberately removes-then-adds precisely to PUSH the delay back
+    // to the trailing edge on every store (a true debounce), which coalescing
+    // alone cannot do. Collapsing them would silently change the history cadence.
+    await this.historyQueue.remove(jobId).catch(() => undefined);
+
     await this.historyQueue.add(
       QueueJob.PAGE_HISTORY,
-      { pageId: page.id } as IPageHistoryJob,
+      { pageId: page.id, kind: 'idle' } as IPageHistoryJob,
       { jobId, delay },
     );
   }

apps/server/src/collaboration/processors/history.processor.spec.ts

@@ -66,6 +66,15 @@ describe('HistoryProcessor.process', () => {
     notificationQueue = { add: jest.fn().mockResolvedValue(undefined) };
     generalQueue = { add: jest.fn().mockResolvedValue(undefined) };
 
+    // #370 F3 — the processor now serializes its find+save under a page-row lock
+    // via executeTx. A db whose transaction().execute(fn) runs fn with a trx stub
+    // drives the real executeTx() helper without a database.
+    const db = {
+      transaction: () => ({
+        execute: (fn: (trx: any) => Promise<any>) => fn({ __trx: true }),
+      }),
+    };
+
     // WorkerHost's constructor reads `this.worker`; passing repos positionally
     // matches the constructor and avoids the Nest DI container.
     proc = new HistoryProcessor(
@@ -73,6 +82,7 @@ describe('HistoryProcessor.process', () => {
       pageRepo as any,
       collabHistory as any,
       watcherService as any,
+      db as any,
       notificationQueue as any,
       generalQueue as any,
     );
@@ -126,15 +136,26 @@ describe('HistoryProcessor.process', () => {
     await proc.process(buildJob());
 
     expect(collabHistory.popContributors).toHaveBeenCalledWith(PAGE_ID);
+    // #370 F3/F9 — the snapshot decision runs under a page-row lock. Pin the lock
+    // structurally so a refactor that drops withLock/trx (silently reintroducing
+    // the TOCTOU double-insert) turns this red. The tx stub is { __trx: true }.
+    expect(pageRepo.findById).toHaveBeenCalledWith(
+      PAGE_ID,
+      expect.objectContaining({ withLock: true, trx: { __trx: true } }),
+    );
+    // #370 F7 — addPageWatchers MUST receive the trx, or its FK-check runs on a
+    // separate connection and self-deadlocks against our FOR UPDATE. Asserting
+    // the trx arg here is exactly what would have caught that regression.
     expect(watcherService.addPageWatchers).toHaveBeenCalledWith(
       ['u1', 'u2'],
       PAGE_ID,
       SPACE_ID,
       WORKSPACE_ID,
+      { __trx: true },
     );
     expect(pageHistoryRepo.saveHistory).toHaveBeenCalledWith(
       expect.objectContaining({ id: PAGE_ID }),
-      { contributorIds: ['u1', 'u2'] },
+      { contributorIds: ['u1', 'u2'], kind: 'idle', trx: { __trx: true } },
     );
     expect(generalQueue.add).toHaveBeenCalledWith(
       QueueJob.PAGE_BACKLINKS,
@@ -186,6 +207,48 @@ describe('HistoryProcessor.process', () => {
     ]);
   });
 
+  it('COMMIT failure (throw outside the tx callback) → contributors RESTORED', async () => {
+    // #370 F8 — a commit-time failure throws OUTSIDE the callback, so the inner
+    // try/catch does not run; the outer catch must restore the popped set (else a
+    // BullMQ retry writes an unattributed version). Use a db whose execute() runs
+    // the callback THEN throws, simulating a commit abort.
+    pageHistoryRepo.findPageLastHistory.mockResolvedValue({
+      content: { type: 'doc', content: [] },
+    });
+    const commitFail = {
+      transaction: () => ({
+        execute: async (fn: (trx: any) => Promise<any>) => {
+          await fn({ __trx: true }); // callback succeeds (saveHistory ok)
+          throw new Error('commit aborted'); // ...but the COMMIT fails
+        },
+      }),
+    };
+    const procCommitFail = new HistoryProcessor(
+      pageHistoryRepo as any,
+      pageRepo as any,
+      collabHistory as any,
+      watcherService as any,
+      commitFail as any,
+      notificationQueue as any,
+      generalQueue as any,
+    );
+    jest
+      .spyOn(procCommitFail['logger'], 'error')
+      .mockImplementation(() => undefined);
+
+    await expect(procCommitFail.process(buildJob())).rejects.toThrow(
+      'commit aborted',
+    );
+    // The inner catch did NOT run (save succeeded), so only the outer catch can
+    // restore — assert it did.
+    expect(collabHistory.addContributors).toHaveBeenCalledWith(PAGE_ID, [
+      'u1',
+      'u2',
+    ]);
+    // And the post-snapshot queue work must NOT have run (we rethrew).
+    expect(generalQueue.add).not.toHaveBeenCalled();
+  });
+
   it('backlinks + notification queue failures are swallowed (history still committed)', async () => {
     pageHistoryRepo.findPageLastHistory.mockResolvedValue({
       content: { type: 'doc', content: [] },

apps/server/src/collaboration/processors/history.processor.ts

@@ -19,6 +19,9 @@ import { isDeepStrictEqual } from 'node:util';
 import { CollabHistoryService } from '../services/collab-history.service';
 import { WatcherService } from '../../core/watcher/watcher.service';
 import { isEmptyParagraphDoc } from '../collaboration.util';
+import { InjectKysely } from 'nestjs-kysely';
+import { KyselyDB } from '@docmost/db/types/kysely.types';
+import { executeTx } from '@docmost/db/utils';
 
 @Processor(QueueName.HISTORY_QUEUE)
 export class HistoryProcessor extends WorkerHost implements OnModuleDestroy {
@@ -29,6 +32,7 @@ export class HistoryProcessor extends WorkerHost implements OnModuleDestroy {
     private readonly pageRepo: PageRepo,
     private readonly collabHistory: CollabHistoryService,
     private readonly watcherService: WatcherService,
+    @InjectKysely() private readonly db: KyselyDB,
     @InjectQueue(QueueName.NOTIFICATION_QUEUE) private notificationQueue: Queue,
     @InjectQueue(QueueName.GENERAL_QUEUE) private generalQueue: Queue,
   ) {
@@ -41,6 +45,9 @@ export class HistoryProcessor extends WorkerHost implements OnModuleDestroy {
     try {
       const { pageId } = job.data;
 
+      // Read the page WITHOUT a lock first, only to bail early on the two cheap
+      // no-write cases (page gone / empty first snapshot) without opening a
+      // transaction. The authoritative check-then-write happens locked below.
       const page = await this.pageRepo.findById(pageId, {
         includeContent: true,
       });
@@ -51,40 +58,109 @@ export class HistoryProcessor extends WorkerHost implements OnModuleDestroy {
         return;
       }
 
-      const lastHistory = await this.pageHistoryRepo.findPageLastHistory(
-        pageId,
-        { includeContent: true },
-      );
+      // #370 F3 — the snapshot decision (findPageLastHistory → saveHistory) must
+      // be serialized against manual-save/boundary writers, which run under a
+      // page-row lock in onStoreDocument. Without it, this processor and a
+      // concurrent manual-save each read the same lastHistory (MVCC), both see
+      // content != lastHistory, and both insert — producing two page_history rows
+      // with IDENTICAL content (one 'idle', one 'manual'), defeating
+      // promote-not-dup and the version-vs-autosave split. Taking the same
+      // page-row lock makes the second writer observe the first's committed row so
+      // the isDeepStrictEqual gate collapses the duplicate. Only the read+write
+      // is transacted; the post-snapshot queue work stays outside.
+      let contributorIds: string[] = [];
+      let snapshotWritten = false;
+      let lastHistoryContent: unknown;
+      // #370 F8 — the contributor set popped from Redis (destructive SPOP) must be
+      // restored if the snapshot does not durably land. The inner try/catch only
+      // covers a throw INSIDE the callback; a COMMIT failure (connection drop,
+      // serialization/deadlock abort on commit — the transient class the epic
+      // already retries) throws OUTSIDE it, rolling the snapshot back while the
+      // pop is already gone. We track the popped set here and restore it in the
+      // outer catch so a BullMQ retry re-attributes the version. addContributors
+      // is an idempotent Redis SADD, so a double-restore is harmless.
+      let poppedForRestore: string[] = [];
 
-      if (!lastHistory && isEmptyParagraphDoc(page.content as any)) {
-        this.logger.debug(
-          `Skipping first history for page ${pageId}: empty content`,
-        );
-        await this.collabHistory.clearContributors(pageId);
+      try {
+        await executeTx(this.db, async (trx) => {
+          const lockedPage = await this.pageRepo.findById(pageId, {
+            includeContent: true,
+            withLock: true,
+            trx,
+          });
+          if (!lockedPage) return;
+
+          const lastHistory = await this.pageHistoryRepo.findPageLastHistory(
+            pageId,
+            { includeContent: true, trx },
+          );
+          lastHistoryContent = lastHistory?.content;
+
+          if (!lastHistory && isEmptyParagraphDoc(lockedPage.content as any)) {
+            this.logger.debug(
+              `Skipping first history for page ${pageId}: empty content`,
+            );
+            return;
+          }
+
+          if (
+            lastHistory &&
+            isDeepStrictEqual(lastHistory.content, lockedPage.content)
+          ) {
+            return; // already snapshotted at this content — nothing to write
+          }
+
+          contributorIds = await this.collabHistory.popContributors(pageId);
+          poppedForRestore = contributorIds;
+          try {
+            // Pass `trx` so the watcher insert's FK check (FOR KEY SHARE on
+            // pages[pageId]) runs on the SAME connection that already holds the
+            // FOR UPDATE lock from findById — otherwise it takes the FK lock on a
+            // separate pool connection and self-deadlocks against our own tx.
+            await this.watcherService.addPageWatchers(
+              contributorIds,
+              pageId,
+              lockedPage.spaceId,
+              lockedPage.workspaceId,
+              trx,
+            );
+
+            // #370 — every job on this queue is a trailing idle-flush autosnapshot.
+            await this.pageHistoryRepo.saveHistory(lockedPage, {
+              contributorIds,
+              kind: job.data.kind ?? 'idle',
+              trx,
+            });
+            snapshotWritten = true;
+            this.logger.debug(`History created for page: ${pageId}`);
+          } catch (err) {
+            await this.collabHistory.addContributors(pageId, contributorIds);
+            poppedForRestore = [];
+            throw err;
+          }
+        });
+      } catch (err) {
+        // A throw here means the tx did NOT commit (callback threw, or the commit
+        // itself failed and rolled back). If we popped contributors and the inner
+        // catch did not already restore them, restore now so the retry keeps
+        // attribution. snapshotWritten is irrelevant: it is set before commit, so
+        // it can be true even when the commit rolled the snapshot back.
+        if (poppedForRestore.length) {
+          await this.collabHistory.addContributors(pageId, poppedForRestore);
+        }
+        throw err;
+      }
+
+      // No snapshot written (page vanished / empty-first / unchanged content) →
+      // clear the contributor set for the skip cases and stop.
+      if (!snapshotWritten) {
+        if (!lastHistoryContent && isEmptyParagraphDoc(page.content as any)) {
+          await this.collabHistory.clearContributors(pageId);
+        }
         return;
       }
 
-      if (
-        !lastHistory ||
-        !isDeepStrictEqual(lastHistory.content, page.content)
-      ) {
-        const contributorIds = await this.collabHistory.popContributors(pageId);
-
-        try {
-          await this.watcherService.addPageWatchers(
-            contributorIds,
-            pageId,
-            page.spaceId,
-            page.workspaceId,
-          );
-
-          await this.pageHistoryRepo.saveHistory(page, { contributorIds });
-          this.logger.debug(`History created for page: ${pageId}`);
-        } catch (err) {
-          await this.collabHistory.addContributors(pageId, contributorIds);
-          throw err;
-        }
-
+      {
         const mentions = extractMentions(page.content);
         const pageMentions = extractPageMentions(mentions);
         const internalLinkSlugIds = extractInternalLinkSlugIds(page.content);
@@ -102,7 +178,7 @@ export class HistoryProcessor extends WorkerHost implements OnModuleDestroy {
             );
           });
 
-        if (contributorIds.length > 0 && lastHistory?.content) {
+        if (contributorIds.length > 0 && lastHistoryContent) {
           await this.notificationQueue
             .add(QueueJob.PAGE_UPDATED, {
               pageId,

apps/server/src/core/ai-chat/ai-chat.controller.ts

@@ -35,6 +35,7 @@ import {
 import { PaginationOptions } from '@docmost/db/pagination/pagination-options';
 import { AiChatRepo } from '@docmost/db/repos/ai-chat/ai-chat.repo';
 import { AiChatMessageRepo } from '@docmost/db/repos/ai-chat/ai-chat-message.repo';
+import { AiChatRunStepRepo } from '@docmost/db/repos/ai-chat/ai-chat-run-step.repo';
 import { PageRepo } from '@docmost/db/repos/page/page.repo';
 import { UserThrottlerGuard } from '../../integrations/throttle/user-throttler.guard';
 import { AI_CHAT_THROTTLER } from '../../integrations/throttle/throttler-names';
@@ -43,6 +44,8 @@ import {
   AiChatRunHooks,
   AiChatService,
   AiChatStreamBody,
+  rowHasInlineParts,
+  hydrateAssistantParts,
 } from './ai-chat.service';
 import { AiChatRunService } from './ai-chat-run.service';
 import { AiTranscriptionService } from './ai-transcription.service';
@@ -129,8 +132,39 @@ export class AiChatController {
     // production. Only touched on the resumable-stream (flag-on) path.
     private readonly streamRegistry?: AiChatStreamRegistryService,
     private readonly environment?: EnvironmentService,
+    // #492: reconstruct a #492 mid-run record's parts from the steps table before
+    // returning rows to the client / export. OPTIONAL so positional controller
+    // specs compile unchanged; when absent, hydration is skipped (old-era rows
+    // already carry inline parts, so nothing to reconstruct).
+    private readonly aiChatRunStepRepo?: AiChatRunStepRepo,
   ) {}
 
+  /**
+   * Reconstruct parts for any assistant rows that don't carry them INLINE — a
+   * #492 mid-run record whose per-step parts live in `ai_chat_run_steps` (the
+   * append-persist backend). Every FINISHED row (old-era + #492) and every old-era
+   * streaming snapshot already has inline `metadata.parts`, so the common path
+   * fetches NOTHING and returns the rows untouched; only an actively-streaming
+   * new-style row triggers the batch step fetch. Consumers (seed/poll/export) read
+   * `metadata.parts` off the returned rows exactly as before — the era switch is
+   * invisible to them (reconstructRunParts contract).
+   */
+  private async withReconstructedParts(
+    rows: AiChatMessage[],
+    workspaceId: string,
+  ): Promise<AiChatMessage[]> {
+    if (!this.aiChatRunStepRepo) return rows;
+    const needy = rows.filter(
+      (r) => r.role === 'assistant' && !rowHasInlineParts(r),
+    );
+    if (needy.length === 0) return rows;
+    const stepsByMessage = await this.aiChatRunStepRepo.findByMessageIds(
+      needy.map((r) => r.id),
+      workspaceId,
+    );
+    return hydrateAssistantParts(rows, stepsByMessage);
+  }
+
   /** List the requesting user's chats in this workspace (paginated). */
   @HttpCode(HttpStatus.OK)
   @Post('chats')
@@ -184,11 +218,17 @@ export class AiChatController {
     @AuthWorkspace() workspace: Workspace,
   ) {
     await this.assertOwnedChat(dto.chatId, user, workspace);
-    return this.aiChatMessageRepo.findByChat(
+    const page = await this.aiChatMessageRepo.findByChat(
       dto.chatId,
       workspace.id,
       pagination,
     );
+    // #492: reconstruct parts for any active new-style row so the client seed sees
+    // `metadata.parts` unchanged (a no-op for the finished rows that fill a page).
+    return {
+      ...page,
+      items: await this.withReconstructedParts(page.items, workspace.id),
+    };
   }
 
   /**
@@ -225,7 +265,10 @@ export class AiChatController {
       workspace.id,
     );
     return {
-      rows,
+      // #492: the delta of an actively-streaming new-style row carries its parts
+      // reconstructed from the steps table, so the degraded poll shows persisted
+      // progress exactly as the pre-#492 full-row snapshot did.
+      rows: await this.withReconstructedParts(rows, workspace.id),
       cursor,
       run: run ? { id: run.id, status: run.status } : null,
     };
@@ -247,8 +290,10 @@ export class AiChatController {
     @AuthWorkspace() workspace: Workspace,
   ): Promise<{ markdown: string }> {
     const chat = await this.assertOwnedChat(dto.chatId, user, workspace);
-    const rows = await this.aiChatMessageRepo.findAllByChat(
-      dto.chatId,
+    const rows = await this.withReconstructedParts(
+      await this.aiChatMessageRepo.findAllByChat(dto.chatId, workspace.id),
+      // #492: an interrupted-but-still-active turn exports its persisted steps
+      // (reconstructed from the steps table) just like the pre-#492 full row did.
       workspace.id,
     );
     const markdown = buildChatMarkdown({
@@ -288,7 +333,13 @@ export class AiChatController {
           workspace.id,
         )
       : undefined;
-    return { run, message: message ?? null };
+    // #492: reconnect to an IN-FLIGHT run reconstructs the projection row's parts
+    // from the steps table (the row itself carries only the step marker mid-run);
+    // a finished run's row already has inline parts, so this is a no-op.
+    const [hydrated] = message
+      ? await this.withReconstructedParts([message], workspace.id)
+      : [undefined];
+    return { run, message: hydrated ?? null };
   }
 
   /**

apps/server/src/core/ai-chat/ai-chat.service.ts

@@ -22,6 +22,7 @@ import { AiSettingsService } from '../../integrations/ai/ai-settings.service';
 import { describeProviderError } from '../../integrations/ai/ai-error.util';
 import { AiChatRepo } from '@docmost/db/repos/ai-chat/ai-chat.repo';
 import { AiChatMessageRepo } from '@docmost/db/repos/ai-chat/ai-chat-message.repo';
+import { AiChatRunStepRepo } from '@docmost/db/repos/ai-chat/ai-chat-run-step.repo';
 import { AiChatPageSnapshotRepo } from '@docmost/db/repos/ai-chat/ai-chat-page-snapshot.repo';
 import { AiAgentRoleRepo } from '@docmost/db/repos/ai-agent-roles/ai-agent-roles.repo';
 import { PageRepo } from '@docmost/db/repos/page/page.repo';
@@ -189,10 +190,11 @@ export function stepBudgetWarning(stepNumber: number): string {
 //
 // `system` is the in-scope system prompt; we CONCATENATE so the original
 // persona/context is preserved — a bare `system` override would REPLACE the
-// whole system prompt for the step. `activatedTools` is PER-TURN mutable state
-// owned by the streaming loop (a closure Set grown by loadTools); it is passed
-// in (not module-global, not persisted) so this stays a pure function of its
-// arguments.
+// whole system prompt for the step. `activatedTools` is a closure Set grown by
+// loadTools and owned by the streaming loop; the caller seeds it from and
+// persists it to the chat's metadata across turns (#490), but this function only
+// READS the Set it is handed, so it stays a pure function of its arguments (not
+// module-global).
 //
 // NOTE: at AI SDK v7 the per-step `system` field is renamed to `instructions`.
 // On v6 (`^6.0.134`) `system` is the correct field — adjust when bumping.
@@ -517,6 +519,12 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
     // constructions compile unchanged; Nest always injects the real singleton, so
     // reconcile sees the SAME in-memory active/zombie maps the runner mutates.
     private readonly aiChatRunService?: AiChatRunService,
+    // #492 append-persist: per-step INSERT into the lightweight steps table (the
+    // O(Σ steps) replacement for the O(n²) full-row `metadata.parts` rewrite).
+    // OPTIONAL so existing positional constructions (int-specs) compile unchanged;
+    // Nest injects the real singleton. When ABSENT the per-step path falls back to
+    // the pre-#492 full-row flush (no regression, only no WAL win).
+    private readonly aiChatRunStepRepo?: AiChatRunStepRepo,
   ) {}
 
   // #487: periodic reconcile timer (single-process phase 1). Started in
@@ -1113,8 +1121,34 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
         chatId,
         workspace.id,
       );
+      // #492: HYDRATE needy assistant rows from the steps table BEFORE the replay
+      // map. A #492 mid-run assistant row carries only a step marker
+      // (metadata.parts:[]); its real per-step parts live in `ai_chat_run_steps`.
+      // The graceful terminal callbacks (onFinish/onError/onAbort -> flushAssistant)
+      // assemble the full inline parts, so a normally-ended turn already has them.
+      // But a HARD crash mid-run (SIGKILL/OOM) fires NO terminal callback, so the
+      // row stays parts:[]; without this, rowToUiMessage falls back to an empty
+      // text part and the partial tool-calls/results/text — durable in the steps
+      // table — would DROP OUT of the model's replay context (regressing #183
+      // step-granular durability for the model consumer). Mirrors the controller's
+      // withReconstructedParts EXACTLY (same needy predicate + hydration helper).
+      // Guarded on the optional repo: absent (positional test builds) degrades to
+      // the current behavior rather than crashing.
+      let replayHistory = oldHistory;
+      if (this.aiChatRunStepRepo) {
+        const needy = oldHistory.filter(
+          (r) => r.role === 'assistant' && !rowHasInlineParts(r),
+        );
+        if (needy.length > 0) {
+          const stepsByMessage = await this.aiChatRunStepRepo.findByMessageIds(
+            needy.map((r) => r.id),
+            workspace.id,
+          );
+          replayHistory = hydrateAssistantParts(oldHistory, stepsByMessage);
+        }
+      }
       const uiMessages: Array<Omit<UIMessage, 'id'> & { id: string }> = [
-        ...oldHistory.map(rowToUiMessage),
+        ...replayHistory.map(rowToUiMessage),
         {
           id: 'pending-user',
           role: 'user',
@@ -1153,7 +1187,9 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
       // hint — confirm it against the persisted history (the preceding assistant
       // turn must really be aborted/streaming) so a spoofed flag cannot inject the
       // interrupt note onto an ordinary turn. The partial output the model needs is
-      // already in `messages` (the aborted assistant row replays via findRecent).
+      // already in `messages`: a #492 mid-run row's per-step parts live only in the
+      // `ai_chat_run_steps` table and were hydrated into the replay history above,
+      // so the aborted assistant turn replays WITH its partial parts intact.
       // Append the new user turn (shape-only) so index -2 is the prior assistant.
       const interrupted = isInterruptResume(
         [...oldHistory, { role: 'user', status: null, metadata: null }],
@@ -1410,10 +1446,11 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
       const baseTools = { ...external.tools, ...docmostTools };
 
       // Deferred tool loading state (#332), scoped to THIS streaming loop:
-      //  - `activatedTools` is per-TURN mutable state — a fresh closure Set created
-      //    per streamText call, NOT module-global and NOT persisted, so a new turn
-      //    starts cold. loadTools.execute adds to it; prepareAgentStep reads it to
-      //    widen `activeTools` on the NEXT step.
+      //  - `activatedTools` is a fresh closure Set per streamText call (not
+      //    module-global), SEEDED from the chat's persisted metadata.activatedTools
+      //    (#490, just below) so activation carries across turns. loadTools.execute
+      //    adds to it; prepareAgentStep reads it to widen `activeTools` on the NEXT
+      //    step; turn end persists it back.
       //  - `validDeferredNames` = every tool that is NOT core (the in-app deferred
       //    tools + ALL external MCP tools), computed from the ACTUAL toolset so an
       //    external tool is loadable by its namespaced name. loadTools rejects any
@@ -1557,17 +1594,57 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
         // connection when finalize runs, so the SQL `WHERE status='streaming'`
         // (not this flag) is what prevents it clobbering the terminal row.
         if (finalized) return null;
-        // Build the flush ONCE so the returned count is EXACTLY the persisted
-        // `stepsPersisted` (both derive from capturedSteps.length at this instant).
-        const flushed = flushAssistant(capturedSteps, '', 'streaming', {
-          pageChanged,
-          partsCache,
-        });
-        const stepsPersisted = flushed.metadata.stepsPersisted as number;
+        // The count derives from capturedSteps.length at THIS instant, so the
+        // returned value is EXACTLY the persisted `stepsPersisted` the ring rotates
+        // on (whether we take the append-persist path or the legacy fallback).
+        const stepsPersisted = capturedSteps.length;
         try {
-          await this.aiChatMessageRepo.update(assistantId, workspace.id, flushed, {
-            onlyIfStreaming: true,
-          });
+          if (this.aiChatRunStepRepo) {
+            // #492 APPEND-PERSIST: write only THIS finished step's parts to the
+            // steps table (O(step) WAL), then bump the row's CHEAP step marker —
+            // NO growing `metadata.parts` blob (that O(n²) full-row rewrite is
+            // exactly what this removes). The full `metadata.parts` is assembled
+            // once at finalize; a mid-run resume seed is reconstructed from the
+            // step rows (reconstructRunParts). The INSERT is idempotent
+            // (ON CONFLICT DO NOTHING), so a re-fired step never doubles the parts.
+            const index = stepsPersisted - 1;
+            if (index >= 0) {
+              const stepParts = assistantParts(
+                [capturedSteps[index]],
+                '',
+                partsCache,
+              );
+              await this.aiChatRunStepRepo.insertStep(
+                assistantId,
+                workspace.id,
+                index,
+                stepParts,
+              );
+            }
+            // Marker UPDATE: advance stepsPersisted + keep the toolTrace era marker
+            // (bumps updatedAt so the delta poll observes the step, and carries the
+            // frontier a resuming client attaches from). Scoped onlyIfStreaming so a
+            // late marker never clobbers the terminal finalize.
+            await this.aiChatMessageRepo.update(
+              assistantId,
+              workspace.id,
+              { metadata: stepMarkerMetadata(stepsPersisted) },
+              { onlyIfStreaming: true },
+            );
+          } else {
+            // Legacy fallback (no steps table wired — positional test builds): the
+            // pre-#492 full-row flush, so parts still land inline on the row.
+            const flushed = flushAssistant(capturedSteps, '', 'streaming', {
+              pageChanged,
+              partsCache,
+            });
+            await this.aiChatMessageRepo.update(
+              assistantId,
+              workspace.id,
+              flushed,
+              { onlyIfStreaming: true },
+            );
+          }
           return stepsPersisted;
         } catch (err) {
           this.logger.warn(
@@ -2747,6 +2824,122 @@ export function rowToUiMessage(row: AiChatMessage): Omit<UIMessage, 'id'> & {
   return { id: row.id, role, parts: parts as UIMessage['parts'] };
 }
 
+/**
+ * Cheap step-marker metadata for the #492 per-step UPDATE. Advances
+ * `stepsPersisted` (the resume attach frontier) and keeps the `toolTraceVersion`
+ * era marker, WITHOUT the growing `parts` blob (those live in the steps table
+ * now; the full `metadata.parts` is assembled once at finalize by flushAssistant).
+ * `parts: []` is kept for shape stability — it reads as an empty inline-parts row,
+ * which is exactly the discriminator that routes reconstruction to the steps table.
+ */
+export function stepMarkerMetadata(
+  stepsPersisted: number,
+): Record<string, unknown> {
+  return { parts: [], toolTraceVersion: 2, stepsPersisted };
+}
+
+/**
+ * Whether an assistant row already carries its full UI parts INLINE on the row
+ * (`metadata.parts`). TRUE for every FINISHED row — old-era rows AND #492 rows,
+ * whose full parts are assembled once at finalize — and for old-era streaming
+ * snapshots (the pre-#492 per-step full-row flush). FALSE for a #492 MID-RUN
+ * record, whose per-step parts live in the `ai_chat_run_steps` table. This is the
+ * era discriminator the reconstruct seam branches on — no schema flag needed.
+ */
+export function rowHasInlineParts(row: { metadata?: unknown }): boolean {
+  const meta = (row.metadata ?? {}) as { parts?: unknown };
+  return Array.isArray(meta.parts) && meta.parts.length > 0;
+}
+
+/**
+ * Concatenate persisted per-step parts (in `stepIndex` order) into the turn's UI
+ * parts (#492). Reproduces EXACTLY what flushAssistant → assistantParts would have
+ * written to `metadata.parts` for those finished steps, since each step row stored
+ * `assistantParts([step])` at persist time.
+ */
+export function assembleStepParts(
+  stepRows: ReadonlyArray<{ stepIndex: number; parts: unknown }>,
+): UIMessage['parts'] {
+  const parts: Array<Record<string, unknown>> = [];
+  for (const step of [...stepRows].sort((a, b) => a.stepIndex - b.stepIndex)) {
+    if (Array.isArray(step.parts)) {
+      parts.push(...(step.parts as Array<Record<string, unknown>>));
+    }
+  }
+  return parts as UIMessage['parts'];
+}
+
+/**
+ * reconstructRunParts (#492) — the single backend-switch seam. Given an assistant
+ * ROW and its persisted step rows, return the turn's UI `parts` + the persisted
+ * step count, reading from the ROW when it already carries inline parts (old-era
+ * records AND every finished record) and from the STEPS TABLE otherwise (a #492
+ * mid-run record). The higher-level consumers (attach seed, delta poll, export)
+ * route their row→parts through this / {@link hydrateAssistantParts}, so old and
+ * new records reconstruct identically WITHOUT the consumers branching on the era.
+ */
+export function reconstructRunParts(
+  row: { metadata?: unknown; content?: string | null },
+  stepRows: ReadonlyArray<{ stepIndex: number; parts: unknown }>,
+): { parts: UIMessage['parts']; stepsPersisted: number } {
+  if (rowHasInlineParts(row)) {
+    const meta = row.metadata as {
+      parts: UIMessage['parts'];
+      stepsPersisted?: number;
+    };
+    return {
+      parts: meta.parts,
+      stepsPersisted:
+        typeof meta.stepsPersisted === 'number'
+          ? meta.stepsPersisted
+          : stepRows.length,
+    };
+  }
+  if (stepRows.length > 0) {
+    return {
+      parts: assembleStepParts(stepRows),
+      stepsPersisted: stepRows.length,
+    };
+  }
+  // No inline parts and no step rows: an old-era seed / empty streaming row. Fall
+  // back to a single text part from `content` (mirrors rowToUiMessage).
+  return {
+    parts: textPart(row.content ?? '') as UIMessage['parts'],
+    stepsPersisted: 0,
+  };
+}
+
+/**
+ * Fill each assistant row's `metadata.parts` from its step rows when the row does
+ * not already carry them inline (a #492 mid-run record), so a consumer that reads
+ * `metadata.parts` off the RAW row (the client seed/poll, the Markdown export)
+ * sees the reconstructed parts with NO change to itself. Rows that already have
+ * inline parts (old-era + finished) and non-assistant rows pass through untouched.
+ * Pure: returns new row objects, never mutates the inputs.
+ */
+export function hydrateAssistantParts<
+  T extends { id: string; role?: string; metadata?: unknown },
+>(
+  rows: ReadonlyArray<T>,
+  stepsByMessage: Map<
+    string,
+    ReadonlyArray<{ stepIndex: number; parts: unknown }>
+  >,
+): T[] {
+  return rows.map((row) => {
+    if (row.role !== 'assistant' || rowHasInlineParts(row)) return row;
+    const steps = stepsByMessage.get(row.id);
+    if (!steps || steps.length === 0) return row;
+    return {
+      ...row,
+      metadata: {
+        ...((row.metadata ?? {}) as Record<string, unknown>),
+        parts: assembleStepParts(steps),
+      },
+    };
+  });
+}
+
 /**
  * The persisted-row patch shape produced by {@link flushAssistant}. It is the
  * SAME shape the assistant repo insert/update consume (content + toolCalls +

apps/server/src/database/database.module.ts

@@ -32,6 +32,7 @@ import { TemplateRepo } from '@docmost/db/repos/template/template.repo';
 import { AiChatRepo } from '@docmost/db/repos/ai-chat/ai-chat.repo';
 import { AiChatMessageRepo } from '@docmost/db/repos/ai-chat/ai-chat-message.repo';
 import { AiChatRunRepo } from '@docmost/db/repos/ai-chat/ai-chat-run.repo';
+import { AiChatRunStepRepo } from '@docmost/db/repos/ai-chat/ai-chat-run-step.repo';
 import { AiChatPageSnapshotRepo } from '@docmost/db/repos/ai-chat/ai-chat-page-snapshot.repo';
 import { AiProviderCredentialsRepo } from '@docmost/db/repos/ai-chat/ai-provider-credentials.repo';
 import { AiMcpServerRepo } from '@docmost/db/repos/ai-chat/ai-mcp-server.repo';
@@ -125,6 +126,7 @@ import { firstSqlToken } from '../integrations/metrics/metrics.constants';
     AiChatRepo,
     AiChatMessageRepo,
     AiChatRunRepo,
+    AiChatRunStepRepo,
     AiChatPageSnapshotRepo,
     AiProviderCredentialsRepo,
     AiMcpServerRepo,
@@ -161,6 +163,7 @@ import { firstSqlToken } from '../integrations/metrics/metrics.constants';
     AiChatRepo,
     AiChatMessageRepo,
     AiChatRunRepo,
+    AiChatRunStepRepo,
     AiChatPageSnapshotRepo,
     AiProviderCredentialsRepo,
     AiMcpServerRepo,

apps/server/src/database/migrations/20260707T120000-page-history-kind.ts

@@ -0,0 +1,27 @@
+import { type Kysely } from 'kysely';
+
+/**
+ * #370 — page-versioning intentionality tier on a history snapshot.
+ *
+ * Adds `page_history.kind`, the three-tier "how intentional was this snapshot"
+ * marker that lets versions (intentional points) be told apart from autosaves:
+ *   - 'manual'   — a human explicitly saved a version (Cmd+S / Save button)
+ *   - 'agent'    — the AI agent explicitly saved a version
+ *   - 'idle'     — trailing idle-flush autosnapshot (safety net)
+ *   - 'boundary' — autosnapshot pinned on a source transition (user↔agent↔git)
+ *
+ * Nullable with NO default (mirrors last_updated_source in the agent-provenance
+ * migration): legacy rows predate the marker and read back as `null`, which the
+ * client renders as a plain autosave. Stored as a short varchar to stay
+ * forward-compatible without an enum migration.
+ */
+export async function up(db: Kysely<any>): Promise<void> {
+  await db.schema
+    .alterTable('page_history')
+    .addColumn('kind', 'varchar(20)', (col) => col)
+    .execute();
+}
+
+export async function down(db: Kysely<any>): Promise<void> {
+  await db.schema.alterTable('page_history').dropColumn('kind').execute();
+}

apps/server/src/database/migrations/20260708T120000-ai-chat-run-steps.ts

@@ -0,0 +1,70 @@
+import { type Kysely, sql } from 'kysely';
+
+/**
+ * `ai_chat_run_steps` — append-only per-step persistence for an assistant turn
+ * (#492 wave C). Each finished agent step's UI `parts` (its text part + a part
+ * per tool call, WITH the tool output) is INSERTed as its own lightweight row the
+ * moment the step ends, instead of REWRITING the whole assistant row's growing
+ * `metadata.parts` jsonb on every `onStepFinish`.
+ *
+ * WHY a separate table + INSERT (not a jsonb `||` append on the message row): a
+ * Postgres jsonb UPDATE rewrites the ENTIRE TOASTed row version under MVCC, so
+ * re-persisting a growing `metadata.parts` on every step is O(n²) write volume
+ * (a 50-step run with ~100 KB tool outputs wrote hundreds of MB of WAL / dead
+ * tuples per turn, hammering autovacuum). `||` would only shave the network
+ * payload — the WAL/TOAST rewrite harm remains. An INSERT into a per-step table
+ * writes ONLY that step's bytes, so the per-turn write volume is O(Σ steps).
+ *
+ * The full `metadata.parts` on the message row is assembled ONCE at finalize (the
+ * terminal completed/error/aborted write). Mid-run, a resuming client's seed is
+ * reconstructed by concatenating these step rows in `step_index` order — which
+ * reproduces exactly what the old per-step full-row rewrite persisted. Records
+ * written the OLD way (full `metadata.parts` on the row, no step rows) still
+ * reconstruct from the row unchanged; the two eras are distinguished by whether
+ * the row already carries non-empty `metadata.parts` (see reconstructRunParts /
+ * assembleStepParts in ai-chat.service.ts).
+ *
+ * ON DELETE CASCADE on `message_id`: the step rows are a derived projection of the
+ * assistant message; they must vanish with it (or with its workspace).
+ */
+export async function up(db: Kysely<any>): Promise<void> {
+  await db.schema
+    .createTable('ai_chat_run_steps')
+    .ifNotExists()
+    .addColumn('id', 'uuid', (col) =>
+      col.primaryKey().defaultTo(sql`gen_uuid_v7()`),
+    )
+    // The assistant message row this step belongs to (the #183 projection). The
+    // step rows are a derived, per-step slice of that message, so they cascade.
+    .addColumn('message_id', 'uuid', (col) =>
+      col.references('ai_chat_messages.id').onDelete('cascade').notNull(),
+    )
+    .addColumn('workspace_id', 'uuid', (col) =>
+      col.references('workspaces.id').onDelete('cascade').notNull(),
+    )
+    // 0-based index of the finished step within the turn. Ordering key for
+    // reconstruction; unique per message (idempotent step re-persist).
+    .addColumn('step_index', 'integer', (col) => col.notNull())
+    // The step's UI parts (text part + a `tool-*` part per call, WITH output).
+    // Concatenated in step order to rebuild the turn's `metadata.parts`.
+    .addColumn('parts', 'jsonb', (col) => col.notNull())
+    .addColumn('created_at', 'timestamptz', (col) =>
+      col.notNull().defaultTo(sql`now()`),
+    )
+    .execute();
+
+  // Idempotent per-step persist: a retried INSERT of the same (message, step)
+  // is a no-op (the service uses ON CONFLICT DO NOTHING). This also serves the
+  // reconstruction read (WHERE message_id ORDER BY step_index).
+  await db.schema
+    .createIndex('ai_chat_run_steps_message_step_uidx')
+    .ifNotExists()
+    .on('ai_chat_run_steps')
+    .columns(['message_id', 'step_index'])
+    .unique()
+    .execute();
+}
+
+export async function down(db: Kysely<any>): Promise<void> {
+  await db.schema.dropTable('ai_chat_run_steps').ifExists().execute();
+}

apps/server/src/database/repos/ai-chat/ai-chat-run-step.repo.ts

@@ -0,0 +1,95 @@
+import { Injectable } from '@nestjs/common';
+import { InjectKysely } from 'nestjs-kysely';
+import { KyselyDB, KyselyTransaction } from '../../types/kysely.types';
+import { dbOrTx } from '../../utils';
+import { AiChatRunStep } from '@docmost/db/types/entity.types';
+
+/**
+ * Append-only per-step persistence for an assistant turn (#492). Each finished
+ * agent step's UI `parts` (its text part + a `tool-*` part per call, WITH the
+ * tool output) is INSERTed as its own lightweight row the moment the step ends —
+ * instead of REWRITING the assistant row's growing `metadata.parts` jsonb on every
+ * `onStepFinish` (a Postgres jsonb UPDATE rewrites the whole TOASTed row version
+ * under MVCC, so that was O(n²) WAL/dead-tuple churn per turn).
+ *
+ * The full `metadata.parts` on the message row is assembled ONCE at finalize;
+ * mid-run, a resuming client's seed is rebuilt from these rows in `stepIndex`
+ * order (see `assembleStepParts` / the reconstruct seam in ai-chat.service.ts).
+ * Every method is workspace-scoped as defense-in-depth.
+ */
+@Injectable()
+export class AiChatRunStepRepo {
+  constructor(@InjectKysely() private readonly db: KyselyDB) {}
+
+  /**
+   * Append one finished step's parts. Idempotent: a retried persist of the SAME
+   * (message, stepIndex) is a no-op via ON CONFLICT DO NOTHING — the per-step
+   * writes are fired fire-and-forget + serialized, and a duplicate must never
+   * throw into the stream or double the parts. Returns whether a NEW row landed
+   * (false = the step was already persisted).
+   */
+  async insertStep(
+    messageId: string,
+    workspaceId: string,
+    stepIndex: number,
+    parts: unknown,
+    trx?: KyselyTransaction,
+  ): Promise<boolean> {
+    const db = dbOrTx(this.db, trx);
+    const inserted = await db
+      .insertInto('aiChatRunSteps')
+      .values({
+        messageId,
+        workspaceId,
+        stepIndex,
+        // jsonb column: cast through never (same pattern as the message repo).
+        parts: parts as never,
+      })
+      .onConflict((oc) => oc.columns(['messageId', 'stepIndex']).doNothing())
+      .returning('id')
+      .executeTakeFirst();
+    return inserted !== undefined;
+  }
+
+  /** All persisted steps for ONE assistant message, in step order. */
+  async findByMessage(
+    messageId: string,
+    workspaceId: string,
+  ): Promise<AiChatRunStep[]> {
+    return this.db
+      .selectFrom('aiChatRunSteps')
+      .selectAll('aiChatRunSteps')
+      .where('messageId', '=', messageId)
+      .where('workspaceId', '=', workspaceId)
+      .orderBy('stepIndex', 'asc')
+      .execute();
+  }
+
+  /**
+   * All persisted steps for a SET of assistant messages, grouped by messageId
+   * (each group in step order). One query for the batch — the hydration seam
+   * (getMessages / delta / export) calls this only for the rows that actually
+   * need reconstruction (an active new-style row whose `metadata.parts` is still
+   * empty), which is usually none, so this is skipped on the common path.
+   */
+  async findByMessageIds(
+    messageIds: string[],
+    workspaceId: string,
+  ): Promise<Map<string, AiChatRunStep[]>> {
+    const byMessage = new Map<string, AiChatRunStep[]>();
+    if (messageIds.length === 0) return byMessage;
+    const rows = await this.db
+      .selectFrom('aiChatRunSteps')
+      .selectAll('aiChatRunSteps')
+      .where('messageId', 'in', messageIds)
+      .where('workspaceId', '=', workspaceId)
+      .orderBy('stepIndex', 'asc')
+      .execute();
+    for (const row of rows) {
+      const list = byMessage.get(row.messageId);
+      if (list) list.push(row);
+      else byMessage.set(row.messageId, [row]);
+    }
+    return byMessage;
+  }
+}

apps/server/src/database/repos/page/page-history.repo.ts

@@ -13,6 +13,7 @@ import { jsonArrayFrom, jsonObjectFrom } from 'kysely/helpers/postgres';
 import { ExpressionBuilder, sql } from 'kysely';
 import { DB } from '@docmost/db/types/db';
 import { resolveAgentProvenance } from '../agent-provenance';
+import { PageHistoryKind } from '../../../collaboration/constants';
 
 /**
  * Role-resolution subquery for a page-history row's bound AI chat (#300). Joins
@@ -46,6 +47,9 @@ export class PageHistoryRepo {
     'lastUpdatedById',
     'lastUpdatedSource',
     'lastUpdatedAiChatId',
+    // #370 — intentionality tier ('manual' | 'agent' | 'idle' | 'boundary');
+    // null on legacy rows (= autosave). Selected so callers can read/promote it.
+    'kind',
     'contributorIds',
     'spaceId',
     'workspaceId',
@@ -85,9 +89,15 @@ export class PageHistoryRepo {
 
   async saveHistory(
     page: Page,
-    opts?: { contributorIds?: string[]; trx?: KyselyTransaction },
-  ): Promise<void> {
-    await this.insertPageHistory(
+    opts?: {
+      contributorIds?: string[];
+      // #370 — intentionality tier for this snapshot. Omitted → null (legacy
+      // autosave semantics). Callers derive it server-side, never from a client.
+      kind?: PageHistoryKind;
+      trx?: KyselyTransaction;
+    },
+  ): Promise<PageHistory> {
+    return await this.insertPageHistory(
       {
         pageId: page.id,
         slugId: page.slugId,
@@ -99,6 +109,7 @@ export class PageHistoryRepo {
         // Copy the provenance marker off the page row, as for lastUpdatedById.
         lastUpdatedSource: page.lastUpdatedSource,
         lastUpdatedAiChatId: page.lastUpdatedAiChatId,
+        kind: opts?.kind ?? null,
         contributorIds: opts?.contributorIds,
         spaceId: page.spaceId,
         workspaceId: page.workspaceId,
@@ -107,6 +118,25 @@ export class PageHistoryRepo {
     );
   }
 
+  /**
+   * #370 — promote an existing snapshot's intentionality tier in place. Used by
+   * the manual-save "promote-not-dup" path: when the latest history row already
+   * holds the exact content being versioned, we upgrade its `kind` instead of
+   * duplicating a heavy content row.
+   */
+  async updateHistoryKind(
+    pageHistoryId: string,
+    kind: PageHistoryKind,
+    trx?: KyselyTransaction,
+  ): Promise<void> {
+    const db = dbOrTx(this.db, trx);
+    await db
+      .updateTable('pageHistory')
+      .set({ kind })
+      .where('id', '=', pageHistoryId)
+      .execute();
+  }
+
   async findPageHistoryByPageId(pageId: string, pagination: PaginationOptions) {
     const query = this.db
       .selectFrom('pageHistory')

apps/server/src/database/types/db.d.ts

@@ -280,6 +280,7 @@ export interface PageHistory {
   createdAt: Generated<Timestamp>;
   icon: string | null;
   id: Generated<string>;
+  kind: string | null;
   lastUpdatedAiChatId: string | null;
   lastUpdatedById: string | null;
   lastUpdatedSource: string | null;
@@ -691,6 +692,22 @@ export interface AiChatRuns {
   updatedAt: Generated<Timestamp>;
 }
 
+// Append-only per-step persistence for an assistant turn (#492). Mirrors
+// migration 20260708T120000-ai-chat-run-steps.ts. Each finished agent step's UI
+// `parts` are INSERTed as their own row (instead of rewriting the message row's
+// growing `metadata.parts` jsonb every step — an O(n²) WAL/TOAST churn). The full
+// `metadata.parts` is assembled once at finalize; mid-run a resuming client's seed
+// is rebuilt by concatenating these rows in `stepIndex` order. Cascades with the
+// assistant message row it projects.
+export interface AiChatRunSteps {
+  id: Generated<string>;
+  messageId: string;
+  workspaceId: string;
+  stepIndex: number;
+  parts: Json;
+  createdAt: Generated<Timestamp>;
+}
+
 // Per-(chat,page) snapshot of the open page's Markdown at the END of the agent's
 // previous turn (#274). Mirrors migration 20260702T120000-ai-chat-page-snapshot.ts.
 // The next turn diffs the CURRENT Markdown against `contentMd` to surface edits a
@@ -728,6 +745,7 @@ export interface DB {
   aiChats: AiChats;
   aiChatMessages: AiChatMessages;
   aiChatRuns: AiChatRuns;
+  aiChatRunSteps: AiChatRunSteps;
   aiChatPageSnapshots: AiChatPageSnapshots;
   apiKeys: ApiKeys;
   attachments: Attachments;

apps/server/src/database/types/entity.types.ts

@@ -4,6 +4,7 @@ import {
   AiChats,
   AiChatMessages,
   AiChatRuns,
+  AiChatRunSteps,
   AiChatPageSnapshots,
   Attachments,
   Comments,
@@ -64,6 +65,12 @@ export type InsertableAiChatMessage = Omit<Insertable<AiChatMessages>, 'tsv'>;
 export type AiChatRun = Selectable<AiChatRuns>;
 export type InsertableAiChatRun = Insertable<AiChatRuns>;
 
+// AI Chat Run Step (#492): append-only per-step parts persistence. Each finished
+// agent step's UI parts are stored as their own row; the full turn's parts are
+// assembled from these (in stepIndex order) for a mid-run resume seed.
+export type AiChatRunStep = Selectable<AiChatRunSteps>;
+export type InsertableAiChatRunStep = Insertable<AiChatRunSteps>;
+
 // AI Chat Page Snapshot (#274): per-(chat,page) Markdown snapshot taken at the
 // end of the agent's previous turn, diffed against the current page next turn to
 // detect human edits made between turns.

apps/server/src/integrations/import/services/import-attachment.service.drawio-svg.spec.ts

@@ -0,0 +1,131 @@
+// p-limit and @sindresorhus/slugify are ESM-only and not in jest's transform
+// allowlist; both are irrelevant to createDrawioSvg (a pure fs + string method),
+// so they are mocked out to keep the module graph loadable under ts-jest.
+jest.mock('p-limit', () => ({
+  __esModule: true,
+  default: () => (fn: () => unknown) => fn(),
+}));
+jest.mock('@sindresorhus/slugify', () => ({
+  __esModule: true,
+  default: (input: string) => String(input),
+}));
+
+import { promises as fs } from 'fs';
+import * as os from 'os';
+import * as path from 'path';
+import { ImportAttachmentService } from './import-attachment.service';
+
+/**
+ * Unit test for ImportAttachmentService.createDrawioSvg (issue #507).
+ *
+ * The Confluence import wraps a `.drawio` file into a `.drawio.svg` attachment.
+ * The `content=` payload MUST be the mxfile XML entity-escaped (draw.io's native
+ * form), NOT base64 — draw.io's editor decodes a base64 content= via Latin-1
+ * atob, mangling every non-ASCII char into mojibake. createDrawioSvg touches no
+ * injected dependency, so the service is built with placeholder deps.
+ */
+describe('ImportAttachmentService.createDrawioSvg (#507)', () => {
+  const service = new ImportAttachmentService(
+    {} as any,
+    {} as any,
+    {} as any,
+  );
+  const call = (p: string): Promise<Buffer> =>
+    (service as any).createDrawioSvg(p);
+
+  let tmpDir: string;
+  beforeAll(async () => {
+    tmpDir = await fs.mkdtemp(path.join(os.tmpdir(), 'drawio-507-'));
+  });
+  afterAll(async () => {
+    await fs.rm(tmpDir, { recursive: true, force: true });
+  });
+
+  const writeDrawio = async (name: string, xml: string): Promise<string> => {
+    const p = path.join(tmpDir, name);
+    await fs.writeFile(p, xml, 'utf-8');
+    return p;
+  };
+
+  it('writes content= as entity-encoded XML, not base64', async () => {
+    const drawio =
+      '<mxfile host="Confluence"><diagram name="Схема — ёж">' +
+      '<mxGraphModel><root><mxCell id="0"/>' +
+      '<mxCell id="2" value="Старт-бит" vertex="1" parent="0"/>' +
+      '</root></mxGraphModel></diagram></mxfile>';
+    const p = await writeDrawio('cyrillic.drawio', drawio);
+
+    const svg = (await call(p)).toString('utf-8');
+    const content = /content="([^"]*)"/.exec(svg)?.[1];
+    expect(content).toBeDefined();
+
+    // Entity-encoded XML form, starting with &lt;mxfile — never a base64 blob.
+    expect(content).toMatch(/^&lt;mxfile/);
+    expect(content).toContain('&lt;');
+    // Non-ASCII survives as raw UTF-8, with no Latin-1 mojibake.
+    expect(content).toContain('Старт-бит');
+    expect(content).toContain('Схема — ёж');
+    expect(content).not.toContain('Ð');
+
+    // Decoding the attribute (un-escaping) yields the original drawio file.
+    const decoded = content!
+      .replace(/&lt;/g, '<')
+      .replace(/&gt;/g, '>')
+      .replace(/&quot;/g, '"')
+      .replace(/&amp;/g, '&');
+    expect(decoded).toBe(drawio);
+  });
+
+  it('encodes literal tab/newline/CR as numeric char-refs, not literal control chars (#507 F1)', async () => {
+    // A literal tab/newline/CR inside the mxfile XML would be collapsed to a
+    // single space by XML attribute-value normalization when the draw.io editor
+    // reads content=, silently flattening multi-line labels and tab-bearing
+    // values. They must be emitted as numeric char-refs instead.
+    const drawio =
+      '<mxfile><diagram name="p">' +
+      '<mxGraphModel><root><mxCell id="0"/>' +
+      '<mxCell id="2" value="col1\tcol2" style="html=1;\nshadow=0" vertex="1" parent="0"/>' +
+      '<mxCell id="3" value="Line1\nLine2\rLine3" vertex="1" parent="0"/>' +
+      '</root></mxGraphModel></diagram></mxfile>';
+    const p = await writeDrawio('ctrl.drawio', drawio);
+
+    const svg = (await call(p)).toString('utf-8');
+    const content = /content="([^"]*)"/.exec(svg)?.[1];
+    expect(content).toBeDefined();
+    // No literal control chars survive in the attribute value.
+    expect(content).not.toMatch(/[\t\n\r]/);
+    // They round-trip as numeric char-refs.
+    expect(content).toContain('&#x9;');
+    expect(content).toContain('&#xa;');
+    expect(content).toContain('&#xd;');
+    // Decoding (char-refs back to literal, entities back) recovers the file.
+    const decoded = content!
+      .replace(/&lt;/g, '<')
+      .replace(/&gt;/g, '>')
+      .replace(/&quot;/g, '"')
+      .replace(/&#39;/g, "'")
+      .replace(/&#x9;/gi, '\t')
+      .replace(/&#xa;/gi, '\n')
+      .replace(/&#xd;/gi, '\r')
+      .replace(/&amp;/g, '&');
+    expect(decoded).toBe(drawio);
+  });
+
+  it('escapes XML metacharacters in the drawio payload', async () => {
+    const drawio = '<mxfile><diagram name="a &amp; b">"q" &lt;x&gt;</diagram></mxfile>';
+    const p = await writeDrawio('meta.drawio', drawio);
+
+    const svg = (await call(p)).toString('utf-8');
+    const content = /content="([^"]*)"/.exec(svg)?.[1];
+    expect(content).toBeDefined();
+    // The attribute value must contain no bare `<`, `>` or `"` that would break
+    // out of the content="..." attribute or the SVG element.
+    expect(content).not.toMatch(/[<>"]/);
+    const decoded = content!
+      .replace(/&lt;/g, '<')
+      .replace(/&gt;/g, '>')
+      .replace(/&quot;/g, '"')
+      .replace(/&amp;/g, '&');
+    expect(decoded).toBe(drawio);
+  });
+});

apps/server/src/integrations/import/services/import-attachment.service.ts

@@ -8,6 +8,7 @@ import { createReadStream } from 'node:fs';
 import { promises as fs } from 'fs';
 import { Readable } from 'stream';
 import { getMimeType, sanitizeFileName } from '../../../common/helpers';
+import { htmlEscape } from '../../../common/helpers/html-escaper';
 import { v7 } from 'uuid';
 import { FileTask } from '@docmost/db/types/entity.types';
 import { getAttachmentFolderPath } from '../../../core/attachment/attachment.utils';
@@ -849,7 +850,12 @@ export class ImportAttachmentService {
   ): Promise<Buffer> {
     try {
       const drawioContent = await fs.readFile(drawioPath, 'utf-8');
-      const drawioBase64 = Buffer.from(drawioContent).toString('base64');
+      // Write the mxfile XML XML-entity-escaped (draw.io's native content= form),
+      // NOT base64. draw.io's editor decodes a base64 content= via Latin-1 atob
+      // (no UTF-8 step), turning every non-ASCII char (Cyrillic, ё, —) into
+      // mojibake; the entity-encoded form is decoded by the DOM as UTF-8 and
+      // opens intact. Docmost's own decoder reads both forms.
+      const drawioEscaped = this.xmlEscapeContent(drawioContent);
 
       let imageElement = '';
       // If we have a PNG, include it in the SVG
@@ -875,7 +881,7 @@ export class ImportAttachmentService {
       width="600"
       height="400"
       viewBox="0 0 600 400"
-      content="${drawioBase64}">${imageElement}</svg>`;
+      content="${drawioEscaped}">${imageElement}</svg>`;
 
       return Buffer.from(svgContent, 'utf-8');
     } catch (error) {
@@ -884,6 +890,24 @@ export class ImportAttachmentService {
     }
   }
 
+  /**
+   * Escape a string so it is safe as the value of a double-quoted XML attribute
+   * (the `content=` payload of a `.drawio.svg`). The shared `htmlEscape` covers
+   * `& < > " '` (a strict superset of what this attribute needs; the extra `'`
+   * escape is harmless in a `"`-delimited value). On top of that, the numeric
+   * char-refs for tab/newline/CR are required: a literal tab/newline/CR inside
+   * an attribute value is collapsed to a single space by XML attribute-value
+   * normalization on DOM read (both our decoder and the real draw.io editor),
+   * silently flattening multi-line labels and tab-bearing values. Char-refs
+   * survive that normalization (#507).
+   */
+  private xmlEscapeContent(s: string): string {
+    return htmlEscape(s)
+      .replace(/\t/g, '&#x9;')
+      .replace(/\n/g, '&#xa;')
+      .replace(/\r/g, '&#xd;');
+  }
+
   private async uploadWithRetry(opts: {
     abs: string;
     storageFilePath: string;

apps/server/src/integrations/queue/constants/queue.interface.ts

@@ -20,6 +20,10 @@ export interface IStripeSeatsSyncJob {
 
 export interface IPageHistoryJob {
   pageId: string;
+  // #370 — intentionality tier the worker stamps on the snapshot. All jobs on
+  // this queue are trailing idle-flush autosnapshots, so this is 'idle' (absent
+  // → treated as 'idle' by the processor).
+  kind?: 'idle';
 }
 
 /**

apps/server/test/integration/ai-chat-append-persist-service.int-spec.ts

@@ -0,0 +1,412 @@
+import * as http from 'node:http';
+import { Kysely } from 'kysely';
+import { tool } from 'ai';
+import { z } from 'zod';
+import { MockLanguageModelV3, convertArrayToReadableStream } from 'ai/test';
+import { AiChatRepo } from '@docmost/db/repos/ai-chat/ai-chat.repo';
+import { AiChatMessageRepo } from '@docmost/db/repos/ai-chat/ai-chat-message.repo';
+import { AiChatRunStepRepo } from '@docmost/db/repos/ai-chat/ai-chat-run-step.repo';
+import {
+  AiChatService,
+  assembleStepParts,
+  assistantParts,
+  rowHasInlineParts,
+  stepMarkerMetadata,
+} from 'src/core/ai-chat/ai-chat.service';
+import {
+  getTestDb,
+  destroyTestDb,
+  createWorkspace,
+  createUser,
+  createChat,
+  createMessage,
+} from './db';
+
+/**
+ * #492 append-persist — the REAL onStep WRITE path (F2) and the model-REPLAY
+ * hydration path (F1), driven through `AiChatService.stream` against a LIVE
+ * Postgres with a REAL `AiChatRunStepRepo` INJECTED. The existing append-persist
+ * int-specs hand-roll the insert+marker cycle via the repos directly and build
+ * the service with `aiChatRunStepRepo: undefined` (only the legacy-fallback branch
+ * is covered), so an off-by-one on `stepsPersisted-1`, a wrong `capturedSteps`
+ * slice, or a broken marker payload would pass all of them. These tests exercise
+ * the actual `updateStreaming` append-persist branch end to end.
+ *
+ * The seam is the injected `model` (a seeded `MockLanguageModelV3` from `ai/test`)
+ * plus a REAL Node `ServerResponse` as the hijacked socket — mirrors
+ * ai-chat-stream.int-spec.ts.
+ */
+
+const sleep = (ms: number) => new Promise((r) => setTimeout(r, ms));
+
+async function waitFor(
+  cond: () => Promise<boolean> | boolean,
+  { timeoutMs = 15_000, stepMs = 25 } = {},
+): Promise<void> {
+  const start = Date.now();
+  while (Date.now() - start < timeoutMs) {
+    if (await cond()) return;
+    await sleep(stepMs);
+  }
+  throw new Error('waitFor: condition not met within timeout');
+}
+
+// A real Node ServerResponse wired to a live socket (identical helper to the
+// stream int-spec) so the SDK's pipe/heartbeat writes behave as in prod.
+function makeRealResponse(): Promise<{
+  res: http.ServerResponse;
+  cleanup: () => Promise<void>;
+}> {
+  return new Promise((resolve) => {
+    const server = http.createServer((_req, res) => {
+      resolve({
+        res,
+        cleanup: () =>
+          new Promise<void>((done) => {
+            try {
+              if (!res.writableEnded) res.end();
+            } catch {
+              /* socket already gone */
+            }
+            server.close(() => done());
+          }),
+      });
+    });
+    server.listen(0, () => {
+      const port = (server.address() as any).port;
+      const creq = http.request({ port, method: 'GET' }, (cres) => {
+        cres.resume();
+      });
+      creq.on('error', () => undefined);
+      creq.end();
+    });
+  });
+}
+
+// Stream parts for a normal, successful single-step turn.
+function successStream() {
+  return convertArrayToReadableStream([
+    { type: 'stream-start', warnings: [] },
+    { type: 'text-start', id: 't1' },
+    { type: 'text-delta', id: 't1', delta: 'Hello' },
+    { type: 'text-delta', id: 't1', delta: ' there' },
+    { type: 'text-end', id: 't1' },
+    {
+      type: 'finish',
+      finishReason: 'stop',
+      usage: { inputTokens: 10, outputTokens: 5, totalTokens: 15 },
+    },
+  ] as any);
+}
+
+// A THREE-step turn: steps 0 and 1 each emit text + an `echo` tool call (the SDK
+// runs the tool and continues); step 2 answers and stops. Three steps is
+// deliberate: the LAST finished step's append-persist write races the terminal
+// finalize (which writes the full inline parts anyway, so a lost last-step row is
+// by design), but the NON-final steps 0 and 1 always drain to the steps table
+// before finalize — so those are what the test asserts on deterministically.
+function threeStepModel(): MockLanguageModelV3 {
+  let step = 0;
+  const toolStep = (i: number) => ({
+    stream: convertArrayToReadableStream([
+      { type: 'stream-start', warnings: [] },
+      { type: 'text-start', id: `s${i}` },
+      { type: 'text-delta', id: `s${i}`, delta: `step ${i} ` },
+      { type: 'text-end', id: `s${i}` },
+      {
+        type: 'tool-call',
+        toolCallId: `c${i}`,
+        toolName: 'echo',
+        input: JSON.stringify({ msg: `m${i}` }),
+      },
+      {
+        type: 'finish',
+        finishReason: 'tool-calls',
+        usage: { inputTokens: 5, outputTokens: 3, totalTokens: 8 },
+      },
+    ] as any),
+  });
+  return new MockLanguageModelV3({
+    doStream: async () => {
+      const n = step++;
+      // Realistic inter-step latency. A real model spends seconds per step, so the
+      // fire-and-forget per-step write chain drains to the steps table BETWEEN
+      // steps; the mock otherwise collapses all steps into microseconds and the
+      // terminal finalize wins the race before any but the first step persists.
+      if (n > 0) await sleep(200);
+      if (n < 2) return toolStep(n);
+      return {
+        stream: convertArrayToReadableStream([
+          { type: 'stream-start', warnings: [] },
+          { type: 'text-start', id: 's2' },
+          { type: 'text-delta', id: 's2', delta: 'final answer' },
+          { type: 'text-end', id: 's2' },
+          {
+            type: 'finish',
+            finishReason: 'stop',
+            usage: { inputTokens: 6, outputTokens: 4, totalTokens: 10 },
+          },
+        ] as any),
+      };
+    },
+  } as any);
+}
+
+describe('#492 append-persist service paths [integration]', () => {
+  let db: Kysely<any>;
+  let aiChatRepo: AiChatRepo;
+  let msgRepo: AiChatMessageRepo;
+  let stepRepo: AiChatRunStepRepo;
+  let workspaceId: string;
+  let userId: string;
+
+  let closeCalls: number;
+  const mcpClients = {
+    toolsFor: async () => ({
+      tools: {},
+      clients: [
+        {
+          close: async () => {
+            closeCalls += 1;
+          },
+        },
+      ],
+      outcomes: [],
+      instructions: [],
+    }),
+  };
+
+  // Build the service WITH a REAL AiChatRunStepRepo injected (the property under
+  // test) — unlike the legacy-fallback harness that passes it as undefined.
+  const echoTool = tool({
+    description: 'echo the message back',
+    inputSchema: z.object({ msg: z.string() }),
+    execute: async ({ msg }) => ({ echoed: msg }),
+  });
+
+  function buildService(): AiChatService {
+    return new AiChatService(
+      { getChatModel: async () => null } as any,
+      aiChatRepo,
+      msgRepo,
+      {} as any, // aiChatPageSnapshotRepo
+      { resolve: async () => null } as any, // aiSettings
+      { forUser: async () => ({ echo: echoTool }) } as any, // tools
+      mcpClients as any,
+      {} as any, // aiAgentRoleRepo
+      {} as any, // pageRepo
+      {} as any, // pageAccess
+      {
+        isAiChatDeferredToolsEnabled: () => false,
+        isAiChatFinalStepLockdownEnabled: () => false,
+      } as any, // environment (deferred OFF -> all tools active every step)
+      undefined, // streamRegistry
+      undefined, // aiChatRunService
+      stepRepo, // #492 aiChatRunStepRepo — the append-persist backend
+    );
+  }
+
+  function userUiMessage(text: string) {
+    return {
+      id: `u-${Math.random()}`,
+      role: 'user',
+      parts: [{ type: 'text', text }],
+    };
+  }
+
+  async function runStream(opts: {
+    model: MockLanguageModelV3;
+    chatId: string;
+    body: any;
+  }): Promise<void> {
+    closeCalls = 0;
+    const service = buildService();
+    const { res, cleanup } = await makeRealResponse();
+    try {
+      await service.stream({
+        user: { id: userId, workspaceId } as any,
+        workspace: { id: workspaceId, name: 'WS' } as any,
+        sessionId: 'sess-1',
+        body: opts.body,
+        res: { raw: res } as any,
+        signal: new AbortController().signal,
+        model: opts.model as any,
+        role: null,
+      } as any);
+      await waitFor(async () => {
+        const rows = await msgRepo.findAllByChat(opts.chatId, workspaceId);
+        return rows.some(
+          (r) =>
+            r.role === 'assistant' &&
+            ['completed', 'error', 'aborted'].includes(r.status as string),
+        );
+      });
+      await waitFor(() => closeCalls > 0, { timeoutMs: 5_000 });
+    } finally {
+      await cleanup();
+    }
+  }
+
+  beforeAll(async () => {
+    db = getTestDb();
+    aiChatRepo = new AiChatRepo(db as any);
+    msgRepo = new AiChatMessageRepo(db as any);
+    stepRepo = new AiChatRunStepRepo(db as any);
+    workspaceId = (await createWorkspace(db)).id;
+    userId = (await createUser(db, workspaceId)).id;
+  });
+
+  afterAll(async () => {
+    await destroyTestDb();
+  });
+
+  // --- F2: the real onStep append-persist WRITE branch -----------------------
+  it('drives steps through the real onStep path: per-step rows + marker match a single-row flush', async () => {
+    const chatId = (await createChat(db, { workspaceId, creatorId: userId })).id;
+    const model = threeStepModel();
+
+    // Capture the mid-run step-marker UPDATEs the append-persist branch writes on
+    // the assistant row (a { parts: [], toolTraceVersion, stepsPersisted } patch).
+    const updateSpy = jest.spyOn(msgRepo, 'update');
+    try {
+      await runStream({
+        model,
+        chatId,
+        body: { chatId, messages: [userUiMessage('call the tool then answer')] },
+      });
+
+      const rows = await msgRepo.findAllByChat(chatId, workspaceId);
+      const assistant = rows.find((r) => r.role === 'assistant')!;
+      expect(assistant).toBeDefined();
+      expect(assistant.status).toBe('completed');
+      // The turn finalizes with the FULL inline parts assembled by a single-row
+      // flush (assistantParts over every step) — the baseline the per-step slices
+      // must reproduce.
+      expect(rowHasInlineParts(assistant)).toBe(true);
+      const finalParts = (assistant.metadata as { parts: any[] }).parts;
+
+      // The two NON-final finished steps each landed their own row, in stepIndex
+      // order. (The fire-and-forget write chain drains before the next step, so
+      // poll until both are on disk; the LAST step's write may lose the finalize
+      // race, which is by design — its parts are already in `finalParts`.)
+      await waitFor(async () => {
+        const s = await stepRepo.findByMessage(assistant.id, workspaceId);
+        return s.length >= 2;
+      });
+      const steps = await stepRepo.findByMessage(assistant.id, workspaceId);
+      expect(steps[0].stepIndex).toBe(0);
+      expect(steps[1].stepIndex).toBe(1);
+
+      // Each per-step row carries a NON-trivial slice: this step's text part + its
+      // paired tool part (guards a mutation that persists empty/whole-turn parts).
+      const s0 = steps[0].parts as any[];
+      expect(s0).toContainEqual({ type: 'text', text: 'step 0 ' });
+      expect(s0.some((p) => p.type === 'tool-echo')).toBe(true);
+
+      // The per-step slices are EXACTLY the corresponding prefix of the single-row
+      // flush: assembleStepParts([step0, step1]) === finalParts[0 .. len0+len1].
+      // This is what an off-by-one on `stepsPersisted-1` (a wrong `capturedSteps`
+      // slice) or a shifted stepIndex breaks — the prefix no longer aligns.
+      const prefixLen =
+        (steps[0].parts as any[]).length + (steps[1].parts as any[]).length;
+      expect(assembleStepParts([steps[0], steps[1]] as any)).toEqual(
+        finalParts.slice(0, prefixLen),
+      );
+
+      // The mid-run step markers advanced 1 -> 2 -> ... (the resume frontier), each
+      // a shape-stable empty-parts marker equal to a single-row flush's marker.
+      const markerCounts = updateSpy.mock.calls
+        .map((c) => (c[2] as any)?.metadata)
+        .filter(
+          (m) =>
+            m &&
+            Array.isArray(m.parts) &&
+            m.parts.length === 0 &&
+            typeof m.stepsPersisted === 'number',
+        )
+        .map((m) => m.stepsPersisted);
+      // Monotonic from 1, covering at least the two non-final steps.
+      expect(markerCounts.slice(0, 2)).toEqual([1, 2]);
+      expect(
+        updateSpy.mock.calls
+          .map((c) => (c[2] as any)?.metadata)
+          .find((m) => m && m.stepsPersisted === 2),
+      ).toEqual(stepMarkerMetadata(2));
+    } finally {
+      updateSpy.mockRestore();
+    }
+  }, 60_000);
+
+  // --- F1: model-REPLAY hydrates a hard-crashed mid-run turn from the steps table
+  it('replays a hard-crashed mid-run turn WITH its partial steps hydrated from the steps table', async () => {
+    const chatId = (await createChat(db, { workspaceId, creatorId: userId })).id;
+
+    // Prior turn: a genuine user question...
+    await createMessage(db, {
+      workspaceId,
+      chatId,
+      userId,
+      role: 'user',
+      content: 'What is in the design doc?',
+      createdAt: new Date(Date.now() - 3000),
+    });
+    // ...and an assistant row that a HARD crash (SIGKILL/OOM) left mid-run: only a
+    // step marker on the row (metadata.parts:[] , content:''), NO terminal
+    // callback ever fired, so its real parts live ONLY in ai_chat_run_steps.
+    const crashed = await createMessage(db, {
+      workspaceId,
+      chatId,
+      role: 'assistant',
+      content: '',
+      status: 'aborted',
+      metadata: stepMarkerMetadata(1),
+      createdAt: new Date(Date.now() - 2000),
+    });
+    // The durable partial step: some reasoning text + a completed getPage tool
+    // call (input + output), exactly what #183 step-granular durability preserves.
+    await stepRepo.insertStep(
+      crashed.id,
+      workspaceId,
+      0,
+      assistantParts(
+        [
+          {
+            text: 'HYDRATED_PARTIAL_STEP the doc says',
+            toolCalls: [
+              { toolCallId: 'g1', toolName: 'getPage', input: { id: 'p1' } },
+            ],
+            toolResults: [
+              {
+                toolCallId: 'g1',
+                toolName: 'getPage',
+                output: { id: 'p1', body: 'PARTIAL_TOOL_OUTPUT budget section' },
+              },
+            ],
+          } as any,
+        ],
+        '',
+      ),
+    );
+
+    // The NEXT turn: the model just answers. The service must REPLAY the crashed
+    // assistant turn with its partial parts hydrated from the steps table.
+    const model = new MockLanguageModelV3({
+      doStream: async () => ({ stream: successStream() }),
+    } as any);
+    await runStream({
+      model,
+      chatId,
+      body: { chatId, messages: [userUiMessage('Continue please')] },
+    });
+
+    expect(model.doStreamCalls.length).toBeGreaterThan(0);
+    const prompt = JSON.stringify(model.doStreamCalls[0].prompt);
+    // The partial step's TEXT reached the model context (it would be an empty text
+    // part without hydration — rowToUiMessage falls back to `content:''`).
+    expect(prompt).toContain('HYDRATED_PARTIAL_STEP');
+    // The partial TOOL RESULT survived too (durable in the steps table, replayed).
+    expect(prompt).toContain('PARTIAL_TOOL_OUTPUT');
+    // The genuine prior user turn is present as well (sanity: real history replay).
+    expect(prompt).toContain('What is in the design doc?');
+  }, 60_000);
+});

apps/server/test/integration/ai-chat-append-persist-wal.int-spec.ts

@@ -0,0 +1,173 @@
+import { randomBytes } from 'crypto';
+import { Kysely, sql } from 'kysely';
+import { AiChatRunStepRepo } from '@docmost/db/repos/ai-chat/ai-chat-run-step.repo';
+import { AiChatMessageRepo } from '@docmost/db/repos/ai-chat/ai-chat-message.repo';
+import {
+  assistantParts,
+  flushAssistant,
+  stepMarkerMetadata,
+} from '../../src/core/ai-chat/ai-chat.service';
+import {
+  getTestDb,
+  destroyTestDb,
+  createWorkspace,
+  createUser,
+  createChat,
+} from './db';
+
+/**
+ * #492 append-persist — WRITE-VOLUME regression on a LIVE Postgres, measured via
+ * the `pg_current_wal_lsn()` delta around a realistic multi-step run driven through
+ * the REAL repos (not a mock — a mock cannot observe MVCC/TOAST rewrite volume, the
+ * whole point). Proves the core claim:
+ *
+ *   NEW (per-step INSERT into ai_chat_run_steps + a CHEAP step-marker UPDATE on the
+ *   message row) writes O(Σ steps) of WAL — each step writes only its own bytes.
+ *
+ *   OLD (the pre-#492 full-row rewrite: re-persist the GROWING metadata.parts on
+ *   every onStepFinish) writes O(n²) — step k rewrites the whole TOASTed jsonb of
+ *   all k prior outputs.
+ *
+ * The OLD path here IS the reverted behavior, so this doubles as the mutation
+ * check: swapping the new path back to `flushAssistant` full-row UPDATEs reddens
+ * the assertion (OLD is many times larger).
+ */
+type Step = {
+  text: string;
+  toolCalls: Array<{ toolCallId: string; toolName: string; input: unknown }>;
+  toolResults: Array<{ toolCallId: string; toolName: string; output: unknown }>;
+};
+
+// ~100 KB INCOMPRESSIBLE output per step (a page read). Random base64 so TOAST
+// cannot compress it away and hide the real write volume.
+function makeStep(i: number, outputBytes = 100_000): Step {
+  const body = randomBytes(Math.ceil(outputBytes * 0.75)).toString('base64');
+  return {
+    text: `step ${i} reasoning`,
+    toolCalls: [
+      { toolCallId: `c${i}`, toolName: 'getPage', input: { id: `p${i}` } },
+    ],
+    toolResults: [
+      {
+        toolCallId: `c${i}`,
+        toolName: 'getPage',
+        output: { id: `p${i}`, title: `Page ${i}`, body },
+      },
+    ],
+  };
+}
+
+async function walDelta(
+  db: Kysely<any>,
+  fn: () => Promise<void>,
+): Promise<number> {
+  const before = (
+    await sql<{ l: string }>`select pg_current_wal_lsn() as l`.execute(db)
+  ).rows[0].l;
+  await fn();
+  // NOTE: no pg_switch_wal() — a segment switch pads the LSN to the next 16 MB
+  // boundary and would swamp the delta. The raw LSN advances by the WAL bytes.
+  const after = (
+    await sql<{ l: string }>`select pg_current_wal_lsn() as l`.execute(db)
+  ).rows[0].l;
+  return Number(
+    (
+      await sql<{
+        d: string;
+      }>`select pg_wal_lsn_diff(${after}::pg_lsn, ${before}::pg_lsn) as d`.execute(
+        db,
+      )
+    ).rows[0].d,
+  );
+}
+
+describe('#492 append-persist write volume (pg_current_wal_lsn delta) [integration]', () => {
+  let db: Kysely<any>;
+  let stepRepo: AiChatRunStepRepo;
+  let msgRepo: AiChatMessageRepo;
+  let workspaceId: string;
+  let userId: string;
+  let chatId: string;
+
+  beforeAll(async () => {
+    db = getTestDb();
+    stepRepo = new AiChatRunStepRepo(db as any);
+    msgRepo = new AiChatMessageRepo(db as any);
+    workspaceId = (await createWorkspace(db)).id;
+    userId = (await createUser(db, workspaceId)).id;
+    chatId = (await createChat(db, { workspaceId, creatorId: userId })).id;
+  });
+
+  afterAll(async () => {
+    await destroyTestDb();
+  });
+
+  const seedRow = () =>
+    msgRepo.insert({
+      chatId,
+      workspaceId,
+      userId,
+      role: 'assistant',
+      content: '',
+      status: 'streaming',
+      metadata: stepMarkerMetadata(0) as never,
+    });
+
+  const STEPS = 40;
+
+  it('NEW per-step INSERT is O(Σ steps); OLD full-row rewrite is O(n²)', async () => {
+    const steps: Step[] = [];
+    for (let i = 0; i < STEPS; i++) steps.push(makeStep(i));
+
+    // NEW: per-step INSERT of THIS step's parts + a cheap marker UPDATE.
+    const newRow = await seedRow();
+    const newWal = await walDelta(db, async () => {
+      for (let i = 0; i < STEPS; i++) {
+        await stepRepo.insertStep(
+          newRow.id,
+          workspaceId,
+          i,
+          assistantParts([steps[i]], ''),
+        );
+        await msgRepo.update(
+          newRow.id,
+          workspaceId,
+          { metadata: stepMarkerMetadata(i + 1) },
+          { onlyIfStreaming: true },
+        );
+      }
+    });
+
+    // OLD (the pre-#492 revert): re-persist the GROWING metadata.parts on the
+    // message row on every step.
+    const oldRow = await seedRow();
+    const oldWal = await walDelta(db, async () => {
+      const acc: Step[] = [];
+      for (let i = 0; i < STEPS; i++) {
+        acc.push(steps[i]);
+        await msgRepo.update(
+          oldRow.id,
+          workspaceId,
+          flushAssistant(acc as never, '', 'streaming'),
+          { onlyIfStreaming: true },
+        );
+      }
+    });
+
+    // eslint-disable-next-line no-console
+    console.log(
+      `[#492 WAL] ${STEPS} steps ×100KB: new=${(newWal / 1e6).toFixed(1)}MB ` +
+        `old=${(oldWal / 1e6).toFixed(1)}MB (${(oldWal / newWal).toFixed(
+          1,
+        )}x smaller)`,
+    );
+
+    // O(Σ steps): ~STEPS × (100KB output + marker) of WAL. 40 × ~100KB parts plus
+    // 40 tiny markers is a few tens of MB at most — bounded, linear in step count.
+    expect(newWal).toBeLessThan(30_000_000);
+    // O(n²): step k rewrites ~k × 100KB. Σ over 40 steps ≈ 80+ MB — far larger.
+    expect(oldWal).toBeGreaterThan(30_000_000);
+    // The load-bearing claim: the new path writes a small FRACTION of the old.
+    expect(newWal).toBeLessThan(oldWal * 0.35);
+  }, 120_000);
+});

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

@@ -0,0 +1,169 @@
+import { Kysely } from 'kysely';
+import { AiChatController } from 'src/core/ai-chat/ai-chat.controller';
+import {
+  assembleStepParts,
+  assistantParts,
+  stepMarkerMetadata,
+} from 'src/core/ai-chat/ai-chat.service';
+import { AiChatRepo } from '@docmost/db/repos/ai-chat/ai-chat.repo';
+import { AiChatMessageRepo } from '@docmost/db/repos/ai-chat/ai-chat-message.repo';
+import { AiChatRunStepRepo } from '@docmost/db/repos/ai-chat/ai-chat-run-step.repo';
+import type { User, Workspace } from '@docmost/db/types/entity.types';
+import {
+  getTestDb,
+  destroyTestDb,
+  createWorkspace,
+  createUser,
+  createChat,
+  createMessage,
+} from './db';
+
+/**
+ * #492 controller hydration (crash-before-finalize RESUME) on a LIVE Postgres.
+ * `AiChatController.withReconstructedParts` is wired into getMessages/delta/export/
+ * run, but `aiChatRunStepRepo` is OPTIONAL and every controller unit spec passes it
+ * as `undefined`, so the hydration branch early-returns and NEVER executes in those
+ * tests. This drives the real read path — a mid-run streaming row (marker only,
+ * empty inline parts) PLUS its `ai_chat_run_steps` rows — through getMessages WITH
+ * the repo present, exercising the `role==='assistant' && !rowHasInlineParts`
+ * needy predicate, the workspace-scoped batch step fetch, and the endpoint binding.
+ */
+describe('#492 controller hydration read path [integration]', () => {
+  let db: Kysely<any>;
+  let aiChatRepo: AiChatRepo;
+  let msgRepo: AiChatMessageRepo;
+  let stepRepo: AiChatRunStepRepo;
+  let workspaceId: string;
+  let otherWorkspaceId: string;
+  let userId: string;
+
+  // Build the controller WITH a real AiChatRunStepRepo injected (position 9), the
+  // seam the unit specs leave undefined. Only the read-path deps are real.
+  function buildController(): AiChatController {
+    return new AiChatController(
+      {} as any, // aiChatService
+      {} as any, // aiChatRunService
+      aiChatRepo,
+      msgRepo,
+      {} as any, // aiTranscription
+      {} as any, // pageRepo
+      undefined, // streamRegistry
+      undefined, // environment
+      stepRepo, // #492 aiChatRunStepRepo
+    );
+  }
+
+  beforeAll(async () => {
+    db = getTestDb();
+    aiChatRepo = new AiChatRepo(db as any);
+    msgRepo = new AiChatMessageRepo(db as any);
+    stepRepo = new AiChatRunStepRepo(db as any);
+    workspaceId = (await createWorkspace(db)).id;
+    otherWorkspaceId = (await createWorkspace(db)).id;
+    userId = (await createUser(db, workspaceId)).id;
+  });
+
+  afterAll(async () => {
+    await destroyTestDb();
+  });
+
+  it('getMessages reconstructs a mid-run row from the steps table (finished rows untouched)', async () => {
+    const chatId = (
+      await createChat(db, { workspaceId, creatorId: userId })
+    ).id;
+    const user = { id: userId } as User;
+    const workspace = { id: workspaceId } as Workspace;
+
+    // A prior FINISHED assistant row that already carries inline parts — the needy
+    // predicate must SKIP it (no step fetch), returned untouched.
+    const finishedParts = assistantParts(
+      [{ text: 'done earlier', toolCalls: [], toolResults: [] } as any],
+      '',
+    );
+    await createMessage(db, {
+      workspaceId,
+      chatId,
+      role: 'assistant',
+      content: 'done earlier',
+      status: 'completed',
+      metadata: { parts: finishedParts, toolTraceVersion: 2, stepsPersisted: 1 },
+      createdAt: new Date(Date.now() - 3000),
+    });
+
+    // The mid-run row a crash-before-finalize left behind: a step marker only
+    // (parts:[] , content:''), status 'streaming'. Its real parts live ONLY in the
+    // steps table.
+    const midRun = await createMessage(db, {
+      workspaceId,
+      chatId,
+      role: 'assistant',
+      content: '',
+      status: 'streaming',
+      metadata: stepMarkerMetadata(2),
+      createdAt: new Date(Date.now() - 1000),
+    });
+
+    const step0 = assistantParts(
+      [
+        {
+          text: 'reasoning about the page',
+          toolCalls: [
+            { toolCallId: 'g1', toolName: 'getPage', input: { id: 'p1' } },
+          ],
+          toolResults: [
+            { toolCallId: 'g1', toolName: 'getPage', output: { id: 'p1', body: 'B' } },
+          ],
+        } as any,
+      ],
+      '',
+    );
+    const step1 = assistantParts(
+      [{ text: 'partial synthesis so far', toolCalls: [], toolResults: [] } as any],
+      '',
+    );
+    await stepRepo.insertStep(midRun.id, workspaceId, 0, step0);
+    await stepRepo.insertStep(midRun.id, workspaceId, 1, step1);
+
+    // Workspace-scoping guard: a step row for the SAME message id under a DIFFERENT
+    // workspace must NEVER leak into this workspace's reconstruction.
+    await stepRepo.insertStep(midRun.id, otherWorkspaceId, 99, [
+      { type: 'text', text: 'FOREIGN_WORKSPACE_LEAK' },
+    ]);
+
+    const res = await buildController().getMessages(
+      { chatId } as any,
+      { limit: 50 } as any,
+      user,
+      workspace,
+    );
+
+    const items = res.items as any[];
+    const finished = items.find((r) => r.status === 'completed');
+    const reconstructed = items.find((r) => r.id === midRun.id);
+
+    // The finished row passed through with its inline parts unchanged.
+    expect(finished.metadata.parts).toEqual(finishedParts);
+
+    // The mid-run row's parts were reconstructed from the two step rows, in order,
+    // exactly as assembleStepParts concatenates them — the client seed sees the
+    // persisted progress with no change to itself.
+    const expected = assembleStepParts([
+      { stepIndex: 0, parts: step0 },
+      { stepIndex: 1, parts: step1 },
+    ] as any);
+    expect(reconstructed.metadata.parts).toEqual(expected);
+    // The foreign-workspace step row did NOT leak in.
+    expect(JSON.stringify(reconstructed.metadata.parts)).not.toContain(
+      'FOREIGN_WORKSPACE_LEAK',
+    );
+    // Sanity: reconstruction produced real content (text + the paired tool part +
+    // the second step's text), not an empty fallback.
+    expect(reconstructed.metadata.parts).toContainEqual({
+      type: 'text',
+      text: 'reasoning about the page',
+    });
+    expect(
+      (reconstructed.metadata.parts as any[]).some((p) => p.type === 'tool-getPage'),
+    ).toBe(true);
+  }, 60_000);
+});

apps/server/test/integration/ai-chat-run-step-repo.int-spec.ts

@@ -0,0 +1,163 @@
+import { randomBytes } from 'crypto';
+import { Kysely } from 'kysely';
+import { AiChatRunStepRepo } from '@docmost/db/repos/ai-chat/ai-chat-run-step.repo';
+import { AiChatMessageRepo } from '@docmost/db/repos/ai-chat/ai-chat-message.repo';
+import {
+  assistantParts,
+  reconstructRunParts,
+  hydrateAssistantParts,
+  stepMarkerMetadata,
+  rowHasInlineParts,
+} from '../../src/core/ai-chat/ai-chat.service';
+import {
+  getTestDb,
+  destroyTestDb,
+  createWorkspace,
+  createUser,
+  createChat,
+} from './db';
+
+/**
+ * #492 append-persist — the reconstruct CONTRACT on a live Postgres. Proves that a
+ * turn persisted the NEW way (per-step rows in `ai_chat_run_steps`, only a step
+ * marker on the message row) reconstructs to the SAME UI parts as a turn persisted
+ * the OLD way (full `metadata.parts` inline on the row, no step rows) — so the
+ * era-switch is invisible to attach / delta-poll / export. Real repos + real jsonb
+ * roundtrip, not a mock (a mock cannot prove the parts survive the jsonb column
+ * byte-identical).
+ */
+type Step = {
+  text: string;
+  toolCalls: Array<{ toolCallId: string; toolName: string; input: unknown }>;
+  toolResults: Array<{ toolCallId: string; toolName: string; output: unknown }>;
+};
+
+// A realistic step: some text + a getPage tool call whose ~100 KB body is
+// INCOMPRESSIBLE random base64 (a 'x'.repeat filler would TOAST away and hide the
+// real bytes). Under MAX_TOOL_OUTPUT_BYTES (200 KB) it is stored uncompacted.
+function makeStep(i: number, outputBytes = 4_000): Step {
+  const body = randomBytes(Math.ceil(outputBytes * 0.75)).toString('base64');
+  return {
+    text: `step ${i} text`,
+    toolCalls: [
+      { toolCallId: `c${i}`, toolName: 'getPage', input: { id: `p${i}` } },
+    ],
+    toolResults: [
+      {
+        toolCallId: `c${i}`,
+        toolName: 'getPage',
+        output: { id: `p${i}`, title: `Page ${i}`, body },
+      },
+    ],
+  };
+}
+
+describe('AiChatRunStepRepo + reconstruct contract [integration]', () => {
+  let db: Kysely<any>;
+  let stepRepo: AiChatRunStepRepo;
+  let msgRepo: AiChatMessageRepo;
+  let workspaceId: string;
+  let userId: string;
+  let chatId: string;
+
+  beforeAll(async () => {
+    db = getTestDb();
+    stepRepo = new AiChatRunStepRepo(db as any);
+    msgRepo = new AiChatMessageRepo(db as any);
+    workspaceId = (await createWorkspace(db)).id;
+    userId = (await createUser(db, workspaceId)).id;
+    chatId = (await createChat(db, { workspaceId, creatorId: userId })).id;
+  });
+
+  afterAll(async () => {
+    await destroyTestDb();
+  });
+
+  const seedRow = (metadata: unknown, status: string) =>
+    msgRepo.insert({
+      chatId,
+      workspaceId,
+      userId,
+      role: 'assistant',
+      content: '',
+      status,
+      metadata: metadata as never,
+    });
+
+  it('insertStep is idempotent per (message, stepIndex) and reads back in order', async () => {
+    const row = await seedRow(stepMarkerMetadata(0), 'streaming');
+    const parts0 = assistantParts([makeStep(0)], '');
+    const parts1 = assistantParts([makeStep(1)], '');
+
+    expect(await stepRepo.insertStep(row.id, workspaceId, 0, parts0)).toBe(true);
+    expect(await stepRepo.insertStep(row.id, workspaceId, 1, parts1)).toBe(true);
+    // A retried persist of the SAME step is a no-op (ON CONFLICT DO NOTHING).
+    expect(await stepRepo.insertStep(row.id, workspaceId, 0, parts0)).toBe(
+      false,
+    );
+
+    const steps = await stepRepo.findByMessage(row.id, workspaceId);
+    expect(steps.map((s) => s.stepIndex)).toEqual([0, 1]);
+
+    // Batch fetch groups by message id in step order.
+    const map = await stepRepo.findByMessageIds([row.id], workspaceId);
+    expect(map.get(row.id)!.map((s) => s.stepIndex)).toEqual([0, 1]);
+  });
+
+  it('a NEW-style (step-table) run reconstructs identically to an OLD-style (inline) run', async () => {
+    const steps = [makeStep(10), makeStep(11)];
+    // The inline parts the OLD full-row flush would have written.
+    const fullParts = assistantParts(steps, '');
+
+    // OLD-style record: full parts inline on the row, NO step rows.
+    const oldRow = await seedRow(
+      { parts: fullParts, toolTraceVersion: 2, stepsPersisted: 2 },
+      'completed',
+    );
+
+    // NEW-style record: only a step marker on the row + per-step rows.
+    const newRow = await seedRow(stepMarkerMetadata(2), 'streaming');
+    for (let i = 0; i < steps.length; i++) {
+      await stepRepo.insertStep(
+        newRow.id,
+        workspaceId,
+        i,
+        assistantParts([steps[i]], ''),
+      );
+    }
+
+    // Re-read both from the DB (proves the jsonb roundtrip).
+    const oldFetched = await msgRepo.findById(oldRow.id, workspaceId);
+    const newFetched = await msgRepo.findById(newRow.id, workspaceId);
+    const oldSteps = await stepRepo.findByMessage(oldRow.id, workspaceId);
+    const newSteps = await stepRepo.findByMessage(newRow.id, workspaceId);
+
+    // The discriminator: the old row carries inline parts, the new one does not.
+    expect(rowHasInlineParts(oldFetched!)).toBe(true);
+    expect(rowHasInlineParts(newFetched!)).toBe(false);
+    expect(oldSteps).toHaveLength(0);
+    expect(newSteps).toHaveLength(2);
+
+    const oldRecon = reconstructRunParts(oldFetched!, oldSteps);
+    const newRecon = reconstructRunParts(newFetched!, newSteps);
+
+    // Both reconstruct to the SAME parts + step count — the era is invisible.
+    expect(newRecon.parts).toEqual(fullParts);
+    expect(oldRecon.parts).toEqual(fullParts);
+    expect(newRecon.parts).toEqual(oldRecon.parts);
+    expect(newRecon.stepsPersisted).toBe(2);
+    expect(oldRecon.stepsPersisted).toBe(2);
+
+    // hydrateAssistantParts fills the new row's metadata.parts to match the old
+    // row's inline parts — so a consumer reading `metadata.parts` off the raw row
+    // (the client seed/poll, export) is unchanged across the era.
+    const map = await stepRepo.findByMessageIds([newRow.id], workspaceId);
+    const [hydrated] = hydrateAssistantParts([newFetched!], map);
+    expect((hydrated.metadata as { parts: unknown }).parts).toEqual(fullParts);
+    // A row that already has inline parts passes through untouched (same ref-shape).
+    const [oldPassThrough] = hydrateAssistantParts([oldFetched!], map);
+    expect((oldPassThrough.metadata as { parts: unknown }).parts).toEqual(
+      fullParts,
+    );
+  });
+});

apps/server/test/integration/ai-chat-run-steps-migration.int-spec.ts

@@ -0,0 +1,63 @@
+import { Kysely, sql } from 'kysely';
+import {
+  up,
+  down,
+} from '../../src/database/migrations/20260708T120000-ai-chat-run-steps';
+import { getTestDb, destroyTestDb } from './db';
+
+/**
+ * #492 migration up/down roundtrip on a LIVE Postgres. global-setup already
+ * migrated docmost_test to latest (so the table exists at start); this drives the
+ * migration's own down()/up() and asserts the table presence toggles, then leaves
+ * it PRESENT (up) so the shared test DB is intact for any spec that runs after.
+ */
+async function tableExists(db: Kysely<any>): Promise<boolean> {
+  const row = (
+    await sql<{ t: string | null }>`select to_regclass('ai_chat_run_steps') as t`.execute(
+      db,
+    )
+  ).rows[0];
+  return row.t !== null;
+}
+
+async function uniqueIndexExists(db: Kysely<any>): Promise<boolean> {
+  const row = (
+    await sql<{
+      t: string | null;
+    }>`select to_regclass('ai_chat_run_steps_message_step_uidx') as t`.execute(db)
+  ).rows[0];
+  return row.t !== null;
+}
+
+describe('20260708 ai_chat_run_steps migration roundtrip [integration]', () => {
+  let db: Kysely<any>;
+
+  beforeAll(() => {
+    db = getTestDb();
+  });
+
+  afterAll(async () => {
+    // Belt-and-suspenders: guarantee the table is present for later specs even if
+    // an assertion threw mid-roundtrip.
+    if (!(await tableExists(db))) await up(db);
+    await destroyTestDb();
+  });
+
+  it('down() drops the table+index and up() recreates them (idempotent)', async () => {
+    // Starts applied (global-setup migrated to latest).
+    expect(await tableExists(db)).toBe(true);
+    expect(await uniqueIndexExists(db)).toBe(true);
+
+    await down(db);
+    expect(await tableExists(db)).toBe(false);
+    expect(await uniqueIndexExists(db)).toBe(false);
+
+    await up(db);
+    expect(await tableExists(db)).toBe(true);
+    expect(await uniqueIndexExists(db)).toBe(true);
+
+    // up() is idempotent (ifNotExists) — a second run is a harmless no-op.
+    await expect(up(db)).resolves.not.toThrow();
+    expect(await tableExists(db)).toBe(true);
+  });
+});

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

@@ -322,20 +322,21 @@ describe('AiChatService.stream [integration]', () => {
   });
 
   /**
-   * #332 deferred tool loading, the ON path. The riskiest property is that the
-   * per-turn `activatedTools` Set is created FRESH inside each stream() call, so a
-   * tool a previous turn activated via loadTools is NOT still active when the next
-   * turn starts — the new turn begins "cold" (CORE + loadTools only). The unit
-   * tests only exercise pure prepareAgentStep with hand-fed Sets; this pins the
-   * real wiring end-to-end (loadTools.execute -> activatedTools -> prepareStep ->
-   * per-step activeTools) against the real streamText loop, and proves there is no
-   * cross-turn leak. We drive a MockLanguageModelV3 whose step 1 calls
-   * loadTools(['createPage']) and assert, via the model's recorded per-step
-   * CallOptions.tools (the AI SDK filters the provider tool list by activeTools),
-   * that the deferred tool becomes active on the SAME turn's next step but NOT on a
-   * fresh turn's first step.
+   * #332 + #490 deferred tool loading, the ON path. Turn 1 starts COLD (CORE +
+   * loadTools only) and activates a deferred tool via loadTools; that activation
+   * is PERSISTED into the chat's metadata.activatedTools (#490) so the NEXT turn
+   * SEEDS from it and the tool is active from the fresh turn's FIRST step — the
+   * model never re-runs loadTools to re-activate the same tool. The unit tests
+   * only exercise pure prepareAgentStep with hand-fed Sets; this pins the real
+   * wiring end-to-end (loadTools.execute -> activatedTools -> persist -> next-turn
+   * seed -> prepareStep -> per-step activeTools) against the real streamText loop.
+   * We drive a MockLanguageModelV3 whose step 1 calls loadTools(['createPage'])
+   * and assert, via the model's recorded per-step CallOptions.tools (the AI SDK
+   * filters the provider tool list by activeTools), that the deferred tool becomes
+   * active on the SAME turn's next step AND, seeded from metadata, on the next
+   * turn's first step.
    */
-  describe('deferred tool loading ON — per-turn activation, no leak (#332)', () => {
+  describe('deferred tool loading ON — cross-turn activation persistence (#332 + #490)', () => {
     // A stub deferred (non-core) tool the agent can activate. Its execute is never
     // called — the model only needs to SEE it become active — but it must be a
     // valid AI-SDK tool so the SDK includes it in a step's tool list once active.
@@ -451,7 +452,7 @@ describe('AiChatService.stream [integration]', () => {
       } as any);
     }
 
-    it('activates a deferred tool for the SAME turn, and a NEW turn starts cold (no leak)', async () => {
+    it('activates a deferred tool for the SAME turn, and a NEW turn SEEDS it from persisted chat metadata (#490)', async () => {
       const chatId = (await createChat(db, { workspaceId, creatorId: userId })).id;
 
       // --- Turn 1: loadTools(createPage) on step 1, then answer on step 2. ---
@@ -474,7 +475,7 @@ describe('AiChatService.stream [integration]', () => {
       // Step 2 of the SAME turn sees the just-activated deferred tool.
       expect(step2Tools).toContain('createPage');
 
-      // --- Turn 2 on the SAME chat: must start cold again. ---
+      // --- Turn 2 on the SAME chat: seeds the persisted activation (#490). ---
       const model2 = new MockLanguageModelV3({
         doStream: async () => ({ stream: successStream() }),
       } as any);
@@ -485,9 +486,10 @@ describe('AiChatService.stream [integration]', () => {
 
       const nextTurnFirstStep = toolNames(model2.doStreamCalls[0]);
       expect(nextTurnFirstStep).toContain('loadTools');
-      // The activated set is per-turn: the prior turn's createPage did NOT leak,
-      // so the fresh turn's first step sees it deferred again.
-      expect(nextTurnFirstStep).not.toContain('createPage');
+      // #490: activation PERSISTS across turns — turn 1 wrote createPage into the
+      // chat's metadata.activatedTools, so the next turn seeds from it and the
+      // deferred tool is active from the FIRST step (no need to re-run loadTools).
+      expect(nextTurnFirstStep).toContain('createPage');
     });
   });
 });

apps/server/test/integration/page-history-idle-pipeline.int-spec.ts

@@ -0,0 +1,162 @@
+import { randomUUID } from 'node:crypto';
+import { Queue, Worker } from 'bullmq';
+import { PersistenceExtension } from '../../src/collaboration/extensions/persistence.extension';
+
+/**
+ * #370 — integration property of the idle-snapshot pipeline against REAL BullMQ.
+ *
+ * This is deliberately NOT a unit test of computeHistoryJob (that lives in
+ * compute-history-job.spec.ts). The point here is the OBSERVABLE end-to-end
+ * behaviour of the production `enqueuePageHistory` remove-then-add debounce
+ * driving a real Redis-backed delayed queue + worker (the #431→#439 class: a
+ * locally-correct function whose queue/timer property was never exercised):
+ *
+ *   - a CONTINUOUS burst of stores lasting several caps yields periodic idle
+ *     snapshots — at least one per max-wait cap, NOT one-per-store; and
+ *   - an INTERMITTENT burst (a few stores, then quiet) yields exactly ONE
+ *     trailing snapshot.
+ *
+ * We shrink the idle windows to milliseconds (jest.mock of collaboration
+ * constants) so real BullMQ delayed jobs actually promote within the test —
+ * fake timers cannot advance Redis's own delayed-set clock, so the intervals
+ * must be real but tiny. The production method under test is called verbatim.
+ */
+
+// NOTE: jest.mock is hoisted above the module's const initializers, so its
+// factory cannot close over MAX_WAIT_MS/INTERVAL_MS — the literals are inlined
+// here and MUST stay in sync with the consts below (a single source of truth is
+// impossible across the hoist boundary).
+jest.mock('../../src/collaboration/constants', () => {
+  const actual = jest.requireActual('../../src/collaboration/constants');
+  return {
+    ...actual,
+    IDLE_MAX_WAIT_USER: 300,
+    IDLE_MAX_WAIT_AGENT: 300,
+    IDLE_INTERVAL_USER: 1000,
+    IDLE_INTERVAL_AGENT: 1000,
+  };
+});
+
+// Mirrors the mocked IDLE_MAX_WAIT_* above (IDLE_INTERVAL_* is 1000 > this, so
+// the max-wait ceiling is what actually governs the trailing delay).
+const MAX_WAIT_MS = 300;
+
+const REDIS_CONNECTION = {
+  host: process.env.TEST_REDIS_HOST ?? '127.0.0.1',
+  port: Number(process.env.TEST_REDIS_PORT ?? 6379),
+};
+
+const sleep = (ms: number) => new Promise((r) => setTimeout(r, ms));
+
+describe('#370 idle-snapshot pipeline (real BullMQ)', () => {
+  let queue: Queue;
+  let worker: Worker;
+  let extension: PersistenceExtension;
+  // Every processed snapshot, tagged by pageId so the two scenarios stay isolated.
+  const processed: Array<{ pageId: string; kind: string; at: number }> = [];
+  const queueName = `history-idle-int-${randomUUID()}`;
+
+  beforeAll(async () => {
+    queue = new Queue(queueName, {
+      connection: REDIS_CONNECTION,
+      // Mirror the production default (BullModule.forRoot removeOnComplete): the
+      // enqueue idiom relies on the jobId being freed once a job completes so the
+      // next burst can re-arm the same id.
+      defaultJobOptions: { removeOnComplete: true, removeOnFail: true },
+    });
+    await queue.waitUntilReady();
+
+    worker = new Worker(
+      queueName,
+      async (job) => {
+        processed.push({
+          pageId: job.data?.pageId,
+          kind: job.data?.kind,
+          at: Date.now(),
+        });
+      },
+      { connection: REDIS_CONNECTION },
+    );
+    await worker.waitUntilReady();
+
+    // Construct the real extension; only historyQueue (5th ctor arg) and the
+    // internal idleBurstStart map are exercised by enqueuePageHistory, so the
+    // other collaborators can be null — the constructor only assigns fields.
+    extension = new PersistenceExtension(
+      null as any, // pageRepo
+      null as any, // pageHistoryRepo
+      null as any, // db
+      null as any, // aiQueue
+      queue as any, // historyQueue
+      null as any, // notificationQueue
+      null as any, // collabHistory
+      null as any, // transclusionService
+    );
+  });
+
+  afterAll(async () => {
+    // Force-close and fully drain so no BullMQ background activity (delayed-set
+    // polling, blocking BRPOPLPUSH) bleeds into later suites in this single
+    // shared jest worker (maxWorkers: 1).
+    await worker?.close(true).catch(() => undefined);
+    await queue?.obliterate({ force: true }).catch(() => undefined);
+    await queue?.close();
+    // Let the redis sockets settle before the next suite starts.
+    await sleep(150);
+  });
+
+  const arm = (pageId: string) =>
+    (extension as any).enqueuePageHistory({ id: pageId }, 'user');
+
+  it('continuous burst over several caps → periodic idle snapshots (≥1 per cap, not one-per-store)', async () => {
+    const pageId = randomUUID();
+    const runMs = 6 * MAX_WAIT_MS; // ~6 caps of unbroken editing
+    // Store cadence that does NOT evenly divide the cap: real hocuspocus stores
+    // are not aligned to cap boundaries, so a boundary job promotes in the gap
+    // before the next store's remove(). A cap-aligned cadence would instead land
+    // a store exactly on every boundary and lose the snapshot to the documented
+    // remove-vs-active race — an artefact of the test clock, not the pipeline.
+    const stepMs = 70;
+    const stores = Math.floor(runMs / stepMs);
+
+    const start = Date.now();
+    let count = 0;
+    while (Date.now() - start < runMs) {
+      await arm(pageId);
+      count++;
+      await sleep(stepMs);
+    }
+    // Let the final armed job flush.
+    await sleep(2 * MAX_WAIT_MS);
+
+    const snaps = processed.filter((p) => p.pageId === pageId);
+
+    // Every autosnapshot is an idle-kind row.
+    expect(snaps.every((s) => s.kind === 'idle')).toBe(true);
+    // Periodic: at least one per cap over a multi-cap burst (lower-bounded loosely
+    // to stay robust; the property is "fires at least every cap", not a single
+    // trailing snapshot).
+    expect(snaps.length).toBeGreaterThanOrEqual(3);
+    // But NOT one-per-store: ~`stores` stores were issued; the debounce must
+    // collapse them to a small multiple of the cap count, nowhere near per-store.
+    expect(snaps.length).toBeLessThanOrEqual(Math.ceil(stores / 2));
+  });
+
+  it('intermittent burst (a few stores, then quiet) → exactly ONE trailing snapshot', async () => {
+    const pageId = randomUUID();
+
+    // A short burst well within a single cap window, then silence.
+    await arm(pageId);
+    await sleep(40);
+    await arm(pageId);
+    await sleep(40);
+    await arm(pageId);
+
+    // Wait comfortably past the cap so the single pending trailing job fires.
+    await sleep(4 * MAX_WAIT_MS);
+
+    const snaps = processed.filter((p) => p.pageId === pageId);
+    expect(snaps).toHaveLength(1);
+    expect(snaps[0].kind).toBe('idle');
+  });
+});

docs/features/page-work-time_design.md

@@ -387,7 +387,7 @@ bucketByDay(sessions, tz):
 факты ниже — ground truth, можно дозапросить файлы через gitea MCP по указанному SHA):
 
 - `page_history.kind` — `varchar(20)`, NULLABLE, БЕЗ дефолта (migration
-  `20260705T120000-page-history-kind.ts`). Домен: `manual`/`agent`/`idle`/`boundary`;
+  `20260707T120000-page-history-kind.ts`). Домен: `manual`/`agent`/`idle`/`boundary`;
   legacy `null` = автосейв (`collaboration/constants.ts`, `PageHistoryKind`).
 - `kind` УЖЕ включён в `PageHistoryRepo.baseFields` (`page-history.repo.ts`) — читается всеми
   выборками истории. `saveHistory({kind})` и `updateHistoryKind(id, kind)` существуют.

packages/mcp/src/lib/drawio-xml.ts

@@ -178,10 +178,11 @@ function sliceModel(xml: string): string | null {
 // --- decode chain ----------------------------------------------------------
 
 /**
- * Read the `content=` attribute out of a `.drawio.svg` string. Docmost stores a
- * base64 payload there (createDrawioSvg); draw.io's own SVG export may store the
- * XML entity-encoded instead. The DOM decodes entities for us, so the caller
- * only has to distinguish "starts with '<'" (raw XML) from base64.
+ * Read the `content=` attribute out of a `.drawio.svg` string. Docmost writes
+ * the mxfile XML entity-encoded there (buildDrawioSvg / createDrawioSvg), which
+ * is also how draw.io's own SVG export stores it; older attachments stored a
+ * base64 payload instead. The DOM decodes entities for us, so the caller only
+ * has to distinguish "starts with '<'" (raw XML) from base64.
  */
 export function extractContentAttr(svg: string): string {
   const { doc, error } = parseXml(svg);
@@ -195,12 +196,23 @@ export function extractContentAttr(svg: string): string {
   // value itself never contains a double-quote (base64 / entity-encoded XML).
   const m = /content="([^"]*)"/.exec(svg);
   if (m) {
-    // Decode the handful of XML entities a raw regex would leave encoded.
+    // Decode the handful of XML entities a raw regex would leave encoded. The
+    // numeric char-refs for tab/newline/CR MUST be decoded here too: the DOM
+    // path above turns them back into the literal control chars, so this
+    // regex fallback has to agree or the two decode paths diverge (#507).
+    // `&amp;` is decoded last so an escaped `&amp;#x9;` reads back as the
+    // literal text `&#x9;`, not a tab.
     return m[1]
       .replace(/&lt;/g, "<")
       .replace(/&gt;/g, ">")
       .replace(/&quot;/g, '"')
       .replace(/&#39;/g, "'")
+      .replace(/&#x9;/gi, "\t")
+      .replace(/&#9;/g, "\t")
+      .replace(/&#xa;/gi, "\n")
+      .replace(/&#10;/g, "\n")
+      .replace(/&#xd;/gi, "\r")
+      .replace(/&#13;/g, "\r")
       .replace(/&amp;/g, "&");
   }
   throw new Error("drawio: SVG has no content= attribute to decode");
@@ -307,9 +319,16 @@ export function encodeDrawioFile(modelXml: string, title = "Page-1"): string {
 /**
  * Build the `diagram.drawio.svg` attachment. Mirrors the import service's
  * createDrawioSvg contract exactly:
- *   <svg xmlns=… xmlns:xlink=… content="${base64(drawioFile)}">${inner}</svg>
+ *   <svg xmlns=… xmlns:xlink=… content="${xmlEscape(drawioFile)}">${inner}</svg>
  * plus width/height/viewBox from the diagram bounding box and the schematic
  * preview as the visible children (`inner`).
+ *
+ * The `content=` value is the mxfile XML XML-entity-escaped (draw.io's own
+ * native form), NOT base64. draw.io's editor decodes a base64 content= via
+ * Latin-1 atob (no UTF-8 step), turning every non-ASCII char (e.g. Cyrillic,
+ * ё, —) into mojibake; the entity-encoded form is decoded by the DOM as UTF-8
+ * and opens intact. Our decoder (decodeDrawioSvg) reads both forms, so old
+ * base64 attachments still round-trip.
  */
 export function buildDrawioSvg(
   modelXml: string,
@@ -318,14 +337,14 @@ export function buildDrawioSvg(
   title = "Page-1",
 ): string {
   const file = encodeDrawioFile(modelXml, title);
-  const base64 = Buffer.from(file, "utf-8").toString("base64");
+  const content = xmlEscape(file);
   const w = Math.max(1, Math.round(bbox.width));
   const h = Math.max(1, Math.round(bbox.height));
   return (
     `<svg xmlns="http://www.w3.org/2000/svg" ` +
     `xmlns:xlink="http://www.w3.org/1999/xlink" ` +
     `width="${w}" height="${h}" viewBox="0 0 ${w} ${h}" ` +
-    `content="${base64}">${inner}</svg>`
+    `content="${content}">${inner}</svg>`
   );
 }
 
@@ -334,7 +353,15 @@ function xmlEscape(s: string): string {
     .replace(/&/g, "&amp;")
     .replace(/</g, "&lt;")
     .replace(/>/g, "&gt;")
-    .replace(/"/g, "&quot;");
+    .replace(/"/g, "&quot;")
+    // A literal tab/newline/CR inside an attribute value is collapsed to a
+    // single space by XML attribute-value normalization on DOM read (both jsdom
+    // here and the real draw.io editor), silently flattening multi-line labels
+    // and tab-bearing values. Numeric char-refs survive that normalization, so
+    // emit them the way draw.io's own native export does (#507).
+    .replace(/\t/g, "&#x9;")
+    .replace(/\n/g, "&#xa;")
+    .replace(/\r/g, "&#xd;");
 }
 
 // --- normalization + hash --------------------------------------------------

packages/mcp/test/unit/drawio-xml.test.mjs

@@ -51,6 +51,14 @@ const hasRule = (issues, rule, cellId) =>
     (i) => i.rule === rule && (cellId === undefined || i.cellId === cellId),
   );
 
+// Reverse the attribute-value XML escaping used in the `content=` payload.
+const unescapeAttr = (s) =>
+  s
+    .replace(/&lt;/g, "<")
+    .replace(/&gt;/g, ">")
+    .replace(/&quot;/g, '"')
+    .replace(/&amp;/g, "&");
+
 // --- style parsing ---------------------------------------------------------
 
 test("parseStyle: base stylename + key=value pairs", () => {
@@ -335,16 +343,20 @@ test("encode/build: a title with < > \" & round-trips without corrupting the SVG
 
   // The outer content="..." attribute must not be broken by the title: the raw
   // title metacharacters never appear literally in the SVG markup (they are
-  // base64-encoded inside content=, and escaped inside the file XML).
+  // entity-escaped inside content=, doubly so where the file XML already escaped
+  // them inside name="...").
   const contentMatch = /content="([^"]*)"/.exec(svg);
   assert.ok(contentMatch, "SVG has a single well-formed content= attribute");
 
   // The diagram model still decodes losslessly despite the exotic title.
   assert.equal(decodeDrawioSvg(svg), model);
 
+  // content= is now entity-encoded XML (draw.io's native form), never base64.
+  assert.match(contentMatch[1], /^&lt;mxfile/);
+
   // The file XML is well-formed: the title lives in name="..." as escaped
   // entities, so unescaping recovers the original title byte-for-byte.
-  const fileXml = Buffer.from(contentMatch[1], "base64").toString("utf-8");
+  const fileXml = unescapeAttr(contentMatch[1]);
   const nameMatch = /<diagram id="[^"]*" name="([^"]*)">/.exec(fileXml);
   assert.ok(nameMatch, "the diagram name attribute is intact and quote-safe");
   const decodedTitle = nameMatch[1]
@@ -358,3 +370,112 @@ test("encode/build: a title with < > \" & round-trips without corrupting the SVG
   const file = encodeDrawioFile(model, title);
   assert.match(file, /name="A &lt; B &gt; C &quot; D &amp; E">/);
 });
+
+// --- #507: content= is entity-encoded XML, never base64 --------------------
+
+// A model whose cell values carry Cyrillic, ё and an em dash — exactly the
+// characters draw.io's Latin-1 atob mangles when content= is base64.
+const CYRILLIC_MODEL =
+  "<mxGraphModel><root>" +
+  '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+  '<mxCell id="2" value="Старт-бит — ёж" style="rounded=1;html=1;" vertex="1" parent="1">' +
+  '<mxGeometry x="100" y="100" width="120" height="60" as="geometry"/></mxCell>' +
+  "</root></mxGraphModel>";
+
+test("#507: buildDrawioSvg writes content= as entity-encoded XML (not base64)", () => {
+  const model = normalizeXml(CYRILLIC_MODEL);
+  const svg = buildDrawioSvg(model, "<g/>", { width: 200, height: 120 }, "Диаграмма");
+
+  const contentMatch = /content="([^"]*)"/.exec(svg);
+  assert.ok(contentMatch, "SVG has a single well-formed content= attribute");
+  const content = contentMatch[1];
+
+  // The content= value is the entity-encoded mxfile XML — starts with `&lt;mxfile`.
+  assert.match(content, /^&lt;mxfile/, "content= is entity-encoded mxfile XML");
+  // It must NOT be a base64 blob: base64 has no XML entities and no literal `<`.
+  assert.ok(content.includes("&lt;"), "content= carries XML entities, not base64");
+
+  // Cyrillic / ё / — survive verbatim in the attribute (raw UTF-8, not atob-mangled).
+  assert.ok(content.includes("Старт-бит — ёж"), "non-ASCII value is raw UTF-8 in content=");
+  assert.ok(content.includes("Диаграмма"), "non-ASCII title is raw UTF-8 in content=");
+  // The mojibake that base64+atob would have produced must be absent.
+  assert.ok(!content.includes("Ð"), "no Latin-1 mojibake in content=");
+});
+
+test("#507: Cyrillic model round-trips byte-stable through buildDrawioSvg -> decodeDrawioSvg", () => {
+  const model = normalizeXml(CYRILLIC_MODEL);
+  const svg = buildDrawioSvg(model, "<g/>", { width: 200, height: 120 }, "Заголовок — ё");
+  assert.equal(decodeDrawioSvg(svg), model);
+});
+
+test("#507 back-compat: an OLD base64-form .drawio.svg still decodes losslessly", () => {
+  const model = normalizeXml(CYRILLIC_MODEL);
+  // Reproduce the pre-fix write path: encodeDrawioFile -> base64 in content=.
+  const file = encodeDrawioFile(model, "Старая диаграмма");
+  const base64 = Buffer.from(file, "utf-8").toString("base64");
+  const svg =
+    `<svg xmlns="http://www.w3.org/2000/svg" content="${base64}"><g/></svg>`;
+  // No XML entities, purely base64 alphabet — this is the legacy form.
+  assert.ok(!base64.includes("<") && !base64.includes("&"));
+  assert.equal(decodeDrawioSvg(svg), model);
+});
+
+test("#507 negative: non-ASCII in a cell value AND in the title round-trip clean", () => {
+  const model = normalizeXml(CYRILLIC_MODEL);
+  const title = "Тест — ёмкость № 5";
+  const svg = buildDrawioSvg(model, "<g/>", { width: 200, height: 120 }, title);
+
+  // Model recovered byte-for-byte.
+  assert.equal(decodeDrawioSvg(svg), model);
+
+  // Title lands in the <diagram name="..."> of the decoded file XML, intact.
+  const content = /content="([^"]*)"/.exec(svg)[1];
+  const fileXml = unescapeAttr(content);
+  const nameMatch = /<diagram id="[^"]*" name="([^"]*)">/.exec(fileXml);
+  assert.ok(nameMatch, "diagram name attribute present");
+  assert.equal(unescapeAttr(nameMatch[1]), title);
+});
+
+// --- #507 F1: literal tab/newline/CR in an attribute value survive the
+// content= round-trip. The whole mxfile XML lives in one content="..." attr;
+// XML attribute-value normalization collapses a LITERAL tab/newline/CR to a
+// single space on DOM read, so the escape must emit numeric char-refs instead.
+const CTRL_MODEL =
+  "<mxGraphModel><root>" +
+  '<mxCell id="0"/><mxCell id="1" parent="0"/>' +
+  '<mxCell id="2" value="col1\tcol2" style="html=1;\nshadow=0" vertex="1" parent="1">' +
+  '<mxGeometry x="10" y="10" width="120" height="60" as="geometry"/></mxCell>' +
+  '<mxCell id="3" value="Line1\nLine2\rLine3" vertex="1" parent="1">' +
+  '<mxGeometry x="10" y="100" width="120" height="60" as="geometry"/></mxCell>' +
+  "</root></mxGraphModel>";
+
+test("#507 F1: literal tab/newline/CR in a value round-trip byte-stable (DOM decode)", () => {
+  const svg = buildDrawioSvg(CTRL_MODEL, "<g/>", { width: 200, height: 200 });
+  const content = /content="([^"]*)"/.exec(svg)[1];
+  // The tab/newline/CR must be emitted as numeric char-refs, never as literal
+  // control chars (which DOM attribute-value normalization would eat).
+  assert.ok(
+    !/[\t\n\r]/.test(content),
+    "no literal tab/newline/CR survive in the content= attribute",
+  );
+  assert.ok(
+    content.includes("&#x9;") &&
+      content.includes("&#xa;") &&
+      content.includes("&#xd;"),
+    "tab/newline/CR are emitted as numeric char-refs",
+  );
+  // Full DOM decode recovers the model byte-for-byte, control chars intact.
+  assert.equal(decodeDrawioSvg(svg), CTRL_MODEL);
+});
+
+test("#507 F1: regex fallback decodes tab/newline/CR char-refs (agrees with DOM path)", () => {
+  const svg = buildDrawioSvg(CTRL_MODEL, "<g/>", { width: 200, height: 200 });
+  const content = /content="([^"]*)"/.exec(svg)[1];
+  // A bare `&` makes the SVG wrapper malformed, forcing extractContentAttr onto
+  // its regex fallback branch. That branch must decode the tab/newline/CR
+  // char-refs exactly like the DOM path, or the two decoders diverge.
+  const malformedSvg = `<svg content="${content}">&</svg>`;
+  assert.equal(decodeDrawioSvg(malformedSvg), CTRL_MODEL);
+  // The well-formed (DOM) path yields the identical result.
+  assert.equal(decodeDrawioSvg(svg), CTRL_MODEL);
+});