feat(ai-chat): токен-бюджет реплея истории + реактивная ветка (#490)

Вся персистентная история реплеится провайдеру КАЖДЫЙ ход, поэтому длинный чат
рано или поздно упирается в контекстное окно и получает провайдерский 400 на
каждом ходу — навсегда (чат «кирпичится»). Бюджетер ограничивает РЕПЛЕЙ (никогда
не мутирует персист — в БД остаётся полная запись), детерминированно и byte-stable
(обрезанный префикс идентичен от хода к ходу → дружелюбно к prompt-cache).

Единый оценщик chars/2.5 (кириллица; chars/4 занижает вдвое) вынесен в shared-пакет
packages/token-estimate; клиентский count-stream-tokens.ts переведён на него ТЕМ ЖЕ
коммитом (два расходящихся оценщика = «бейдж 60%, а бюджет уже режет»).

history-budget.ts (чистый, покрыт тестами):
- resolveReplayBudget(raw): min(100k, 0.7×window) при заданном окне; флэт 100k при
  незаданном (именно эти инсталляции ловят терминальный overflow — warn-лог); 0 =
  явный off-switch. Читается СЫРОЙ chatContextWindow, т.к. parsePositiveInt схлопывает
  0 и unset в undefined (новое поле ResolvedAiConfig.chatContextWindowRaw).
- trimHistoryForReplay: первичный сигнал — провайдерский факт metadata.contextTokens
  прошлого хода; chars-оценка — дельта/раскройка/фолбэк. Порядок: обрезка tool-outputs
  старых ходов (head+tail+маркер) → механическое схлопывание старейших ходов
  (конкатенация, НЕ LLM) → текущий + последние N ходов всегда полные. Пейринг
  tool-call/result сохраняется (схлопывание убирает ОБЕ части).
- isContextOverflowError: классификация провайдерского 400 (статус + паттерны).

Реактивная ветка: превентивная оценка не даёт инварианта (первый переполняющий ход
не имеет usage). onError классифицирует context-overflow → пишет различимую причину
и штампует metadata.replayOverflow; следующий ход бюджетер режет агрессивно
(0.5×), что и раскирпичивает чат. Наблюдаемость: metadata.replayTrimmedToTokens.

ПРИМЕЧАНИЕ по реактивной ветке (форк, требует решения ревьюера): истинный in-turn
re-pipe (перезапуск streamText в тот же ответ) архитектурно несовместим с текущим
пайпом — pipeUIMessageStreamToResponse пишет writeHead СИНХРОННО (подтверждено в
ai@6.0.207), а suite ожидает await stream() c моком, не дёргающим колбэки, — так что
отложенный пайп/ожидание сигнала повесит тесты. Поэтому реализована реактивная
рекавери «классификация → штамп → агрессивный ре-трим на следующем ходу», что даёт
тот же инвариант (чат не кирпичится) без рискованного рефактора стрима.

Тесты (наблюдаемые свойства): объём записи через дельту pg_current_wal_lsn() на живой
gitmost-test-pg вокруг 50-шагового прогона (несжимаемые payload'ы) — trace-колонка
v1=140МБ → v2=0.04МБ (в 3206× меньше), полная строка 289МБ → 140МБ (−51%); dual-shape
не нужен здесь; «окно не задано → бюджет применяется»; реактивная классификация на
реальном 400-шейпе; parity клиент/сервер оценщика.

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

.gitignore

@@ -29,6 +29,10 @@ packages/mcp/build/
 # is a build artifact like build/ — never committed, always fresh.
 packages/mcp/src/registry-stamp.generated.ts
 
+# token-estimate compiled output (#490; built in CI/Docker via `pnpm build` /
+# the server `pretest`, never committed, so src/ and prod can never diverge).
+packages/token-estimate/dist/
+
 # Logs
 logs
 *.log

apps/client/package.json

@@ -22,6 +22,7 @@
     "@casl/react": "5.0.1",
     "@docmost/editor-ext": "workspace:*",
     "@docmost/prosemirror-markdown": "workspace:*",
+    "@docmost/token-estimate": "workspace:*",
     "@excalidraw/excalidraw": "0.18.0-3a5ef40",
     "@mantine/core": "8.3.18",
     "@mantine/dates": "8.3.18",

apps/client/src/features/ai-chat/utils/count-stream-tokens.test.ts

@@ -6,10 +6,13 @@ describe("estimateTokens", () => {
     expect(estimateTokens("")).toBe(0);
   });
 
-  it("ceils chars/4 so any non-empty text is at least 1 token", () => {
+  // #490: migrated onto the shared @docmost/token-estimate module (chars/2.5, up
+  // from the old client-only chars/4) so the client counter and the server replay
+  // budgeter can never diverge.
+  it("ceils chars/2.5 so any non-empty text is at least 1 token", () => {
     expect(estimateTokens("a")).toBe(1);
-    expect(estimateTokens("abcd")).toBe(1);
-    expect(estimateTokens("abcde")).toBe(2);
-    expect(estimateTokens("12345678")).toBe(2);
+    expect(estimateTokens("ab")).toBe(1);
+    expect(estimateTokens("abcde")).toBe(2); // 5 / 2.5 = 2
+    expect(estimateTokens("x".repeat(10))).toBe(4); // 10 / 2.5 = 4
   });
 });

apps/client/src/features/ai-chat/utils/count-stream-tokens.ts

@@ -2,18 +2,10 @@
  * Rough client-side token estimation for AI-chat UI affordances.
  *
  * No provider streams exact per-token usage mid-stream, so any in-flight figure
- * is a CLIENT ESTIMATE (chars/≈4 heuristic). Pure + unit-testable: it never runs
- * a real BPE tokenizer (that would be O(n²) on the hot path, bloat the bundle,
- * and be wrong for Gemini/Ollama anyway). Used by the in-body reasoning counter
- * ("Thinking · N tokens").
+ * is a CLIENT ESTIMATE. This re-exports the SHARED estimator from
+ * `@docmost/token-estimate` (chars/2.5) so the in-body counter and the server's
+ * replay budgeter use the SAME heuristic — two divergent estimators would mean
+ * "the badge shows 60%" while "the budgeter already trimmed" (#490). Used by the
+ * in-body reasoning counter ("Thinking · N tokens").
  */
-
-/**
- * Rough token estimate for a piece of text using the standard chars/≈4 heuristic.
- * Returns 0 for empty/whitespace-free-of-content input, and ceils so any
- * non-empty text counts as at least one token.
- */
-export function estimateTokens(text: string): number {
-  if (!text) return 0;
-  return Math.ceil(text.length / 4);
-}
+export { estimateTokens } from "@docmost/token-estimate";

apps/server/package.json

@@ -23,7 +23,7 @@
     "migration:reset": "tsx src/database/migrate.ts down-to NO_MIGRATIONS",
     "migration:codegen": "kysely-codegen --dialect=postgres --camel-case --env-file=../../.env --out-file=./src/database/types/db.d.ts",
     "lint": "eslint \"{src,apps,libs,test}/**/*.ts\" --fix",
-    "pretest": "pnpm --filter @docmost/editor-ext build && pnpm --filter @docmost/prosemirror-markdown build",
+    "pretest": "pnpm --filter @docmost/editor-ext build && pnpm --filter @docmost/prosemirror-markdown build && pnpm --filter @docmost/token-estimate build",
     "test": "jest",
     "test:int": "jest --config test/jest-integration.json",
     "test:watch": "jest --watch",
@@ -44,6 +44,7 @@
     "@docmost/mcp": "workspace:*",
     "@docmost/pdf-inspector": "1.9.6",
     "@docmost/prosemirror-markdown": "workspace:*",
+    "@docmost/token-estimate": "workspace:*",
     "@fastify/compress": "^9.0.0",
     "@fastify/cookie": "^11.0.2",
     "@fastify/multipart": "^10.0.0",
@@ -206,6 +207,7 @@
       "^@docmost/db/(.*)$": "<rootDir>/database/$1",
       "^@docmost/transactional/(.*)$": "<rootDir>/integrations/transactional/$1",
       "^@docmost/ee/(.*)$": "<rootDir>/ee/$1",
+      "^@docmost/token-estimate$": "<rootDir>/../../../packages/token-estimate/src/index.ts",
       "^src/(.*)$": "<rootDir>/$1",
       "^@tiptap/react$": "<rootDir>/../test/stubs/tiptap-react.js"
     }

apps/server/src/core/ai-chat/ai-chat.service.run-race.spec.ts

@@ -181,7 +181,7 @@ describe('AiChatService.stream — abortSignal wiring (#184 F3)', () => {
       {} as never, // pageAccess
       { isAiChatDeferredToolsEnabled: () => false, isAiChatFinalStepLockdownEnabled: () => false } as never, // environment
     );
-    return { svc };
+    return { svc, aiChatMessageRepo };
   }
 
   const body = {
@@ -287,7 +287,7 @@ describe('AiChatService.stream — abortSignal wiring (#184 F3)', () => {
   // Drive stream() to the point streamText is called, capturing the options object
   // (which carries onStepFinish/onFinish/onError/onAbort) and the run hooks.
   async function captureStreamCallbacks() {
-    const { svc } = makeService();
+    const { svc, aiChatMessageRepo } = makeService();
     let capturedOpts: any;
     streamTextMock.mockImplementation((opts: any) => {
       capturedOpts = opts;
@@ -314,7 +314,7 @@ describe('AiChatService.stream — abortSignal wiring (#184 F3)', () => {
       runHooks: runHooks as never,
     });
     expect(capturedOpts).toBeDefined();
-    return { capturedOpts, runHooks };
+    return { capturedOpts, runHooks, aiChatMessageRepo };
   }
 
   it('F9: onStepFinish bumps the run step count, onFinish settles the run "completed" (the dominant autonomous-run path)', async () => {
@@ -369,6 +369,51 @@ describe('AiChatService.stream — abortSignal wiring (#184 F3)', () => {
       expect.stringContaining('provider exploded'),
     );
   });
+
+  // #490 reactive branch: a provider CONTEXT-OVERFLOW 400 in onError is classified,
+  // records a distinguishable cause, and stamps metadata.replayOverflow so the NEXT
+  // turn's budgeter trims aggressively (the recovery that un-bricks the chat).
+  it('#490: a context-overflow 400 stamps replayOverflow on the finalized row', async () => {
+    jest
+      .spyOn(Logger.prototype, 'error')
+      .mockImplementation(() => undefined as never);
+    jest
+      .spyOn(Logger.prototype, 'warn')
+      .mockImplementation(() => undefined as never);
+    const { capturedOpts, aiChatMessageRepo } = await captureStreamCallbacks();
+
+    const overflow = Object.assign(new Error('too large'), {
+      statusCode: 400,
+      message:
+        "This model's maximum context length is 128000 tokens. However, your messages resulted in 214000 tokens. Please reduce the length.",
+    });
+    await capturedOpts.onError({ error: overflow });
+
+    // The seed row exists (finalizeOwner is the owner-write path).
+    expect(aiChatMessageRepo.finalizeOwner).toHaveBeenCalled();
+    const calls = aiChatMessageRepo.finalizeOwner.mock.calls as any[][];
+    const patch = calls[calls.length - 1][2] as {
+      status: string;
+      metadata: Record<string, unknown>;
+    };
+    expect(patch.status).toBe('error');
+    expect(patch.metadata.replayOverflow).toBe(true);
+    expect(patch.metadata.error).toContain('контекстное окно');
+  });
+
+  it('#490: a non-overflow error does NOT stamp replayOverflow', async () => {
+    jest
+      .spyOn(Logger.prototype, 'error')
+      .mockImplementation(() => undefined as never);
+    const { capturedOpts, aiChatMessageRepo } = await captureStreamCallbacks();
+    await capturedOpts.onError({ error: new Error('network reset') });
+    const calls = aiChatMessageRepo.finalizeOwner.mock.calls as any[][];
+    const patch = calls[calls.length - 1][2] as {
+      status: string;
+      metadata: Record<string, unknown>;
+    };
+    expect('replayOverflow' in patch.metadata).toBe(false);
+  });
 });
 
 /**

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

@@ -29,6 +29,8 @@ import {
   FINAL_STEP_NUDGE,
   STEP_LIMIT_NO_ANSWER_MARKER,
   OUTPUT_DEGENERATION_ERROR,
+  lastAssistantContextTokens,
+  lastAssistantReplayOverflow,
 } from './ai-chat.service';
 import type { AiChatMessage, Workspace } from '@docmost/db/types/entity.types';
 import { buildSystemPrompt } from './ai-chat.prompt';
@@ -413,6 +415,55 @@ describe('toolTraceVersion era marker (#490)', () => {
   });
 });
 
+// #490 replay-budget signal helpers over persisted history.
+describe('lastAssistantContextTokens', () => {
+  const row = (
+    role: string,
+    metadata: Record<string, unknown> | null,
+  ): AiChatMessage => ({ role, metadata }) as unknown as AiChatMessage;
+
+  it('reads the most recent assistant turn contextTokens (provider fact)', () => {
+    const hist = [
+      row('user', null),
+      row('assistant', { contextTokens: 12000 }),
+      row('user', null),
+      row('assistant', { contextTokens: 41000 }),
+    ];
+    expect(lastAssistantContextTokens(hist)).toBe(41000);
+  });
+
+  it('returns undefined when the last assistant turn recorded no usage', () => {
+    const hist = [row('assistant', { error: 'boom' }), row('user', null)];
+    expect(lastAssistantContextTokens(hist)).toBeUndefined();
+    expect(lastAssistantContextTokens([])).toBeUndefined();
+  });
+});
+
+describe('lastAssistantReplayOverflow', () => {
+  const row = (
+    role: string,
+    metadata: Record<string, unknown> | null,
+  ): AiChatMessage => ({ role, metadata }) as unknown as AiChatMessage;
+
+  it('is true only when the LAST assistant turn overflowed', () => {
+    expect(
+      lastAssistantReplayOverflow([
+        row('assistant', { replayOverflow: true }),
+        row('user', null),
+      ]),
+    ).toBe(true);
+    // A recovered (later, non-overflow) assistant turn clears it.
+    expect(
+      lastAssistantReplayOverflow([
+        row('assistant', { replayOverflow: true }),
+        row('user', null),
+        row('assistant', { contextTokens: 5 }),
+      ]),
+    ).toBe(false);
+    expect(lastAssistantReplayOverflow([])).toBe(false);
+  });
+});
+
 describe('rowToUiMessage', () => {
   it('prefers metadata.parts over content', () => {
     const row = {
@@ -743,6 +794,23 @@ describe('flushAssistant', () => {
     expect(flushed.metadata.error).toBe('boom');
   });
 
+  // #490 observability: the replay budgeter's decision is stamped on the turn.
+  it('records replayTrimmedToTokens + replayOverflow when provided', () => {
+    const f = flushAssistant([], '', 'error', {
+      error: 'ctx',
+      replayTrimmedToTokens: 42_000,
+      replayOverflow: true,
+    });
+    expect(f.metadata.replayTrimmedToTokens).toBe(42_000);
+    expect(f.metadata.replayOverflow).toBe(true);
+  });
+
+  it('omits the replay metadata when not provided', () => {
+    const f = flushAssistant([], '', 'completed', { finishReason: 'stop' });
+    expect('replayTrimmedToTokens' in f.metadata).toBe(false);
+    expect('replayOverflow' in f.metadata).toBe(false);
+  });
+
   // #274 observability: the page-change diff the agent saw this turn is persisted
   // to metadata.pageChanged when a non-empty diff was injected, and omitted when
   // the diff is empty/whitespace or the arg is not supplied.

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

@@ -55,6 +55,12 @@ import {
   type SelectionContext,
 } from './tools/current-page.util';
 import { roleModelOverride } from './roles/role-model-config';
+import {
+  resolveReplayBudget,
+  isContextOverflowError,
+  trimHistoryForReplay,
+  REPLAY_AGGRESSIVE_FRACTION,
+} from './history-budget';
 import {
   startSseHeartbeat,
   stripStreamingHopByHopHeaders,
@@ -127,6 +133,15 @@ const STEP_LIMIT_NO_ANSWER_MARKER =
 const OUTPUT_DEGENERATION_ERROR =
   'Output degeneration detected (repeated token loop)';
 
+// Prefix recorded on the assistant row when the provider rejected the turn for
+// CONTEXT OVERFLOW (#490): the replayed history exceeded the model's window. The
+// row is ALSO stamped `metadata.replayOverflow` so the NEXT turn's budgeter trims
+// aggressively (the reactive recovery — the overflowing turn had no usage signal
+// to trigger preventive trimming, so the classified 400 is what un-bricks it).
+export const CONTEXT_OVERFLOW_ERROR_PREFIX =
+  'Диалог превысил контекстное окно модели; история будет агрессивно ' +
+  'сокращена на следующем ходу.';
+
 /**
  * Compute the step-budget warning text (#444), or '' when this step is outside
  * the warning band. The warning fires on steps
@@ -1086,7 +1101,7 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
       // per-row conversion and degraded to plain text with a "[tool context
       // omitted]" marker rather than 500-ing the whole turn (silent loss of tool
       // context is not acceptable — the model must see the truncation).
-      const messages = await convertHistoryResilient(uiMessages, (index, err) =>
+      let messages = await convertHistoryResilient(uiMessages, (index, err) =>
         this.logger.warn(
           `Degraded unconvertible history row ${index} on chat ${chatId} to text: ${
             err instanceof Error ? err.message : 'unknown error'
@@ -1135,6 +1150,58 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
       // Here we only need the admin-configured system prompt.
       const resolved = await this.aiSettings.resolve(workspace.id);
 
+      // History-replay token budget (#490). The full conversation is replayed to
+      // the provider every turn, so a long chat eventually 400s on the context
+      // window — forever. Bound the REPLAYED history (never the persisted rows).
+      // PRIMARY signal is the provider's own fact: the last turn's contextTokens.
+      const replayBudget = resolveReplayBudget(resolved?.chatContextWindowRaw);
+      if (replayBudget.usedDefault) {
+        // The default fires precisely for installs with NO configured window —
+        // the ones that hit terminal overflow. Warn so it is observable.
+        this.logger.warn(
+          `AI chat (chat ${chatId}): no chatContextWindow configured; ` +
+            `applying the default replay budget (${replayBudget.thresholdTokens} tokens).`,
+        );
+      }
+      // Last turn's provider-reported context size (authoritative when present).
+      const priorContextTokens = lastAssistantContextTokens(oldHistory);
+      // Reactive recovery (#490): if the LAST turn was rejected for context
+      // overflow (stamped by onError), trim AGGRESSIVELY this turn — the
+      // overflowing turn produced no usage signal, so a normal-threshold trim may
+      // not shrink enough to fit. This is what un-bricks a chat that just 400'd.
+      const priorOverflowed = lastAssistantReplayOverflow(oldHistory);
+      const effectiveThreshold =
+        priorOverflowed && replayBudget.thresholdTokens != null
+          ? Math.floor(
+              replayBudget.thresholdTokens * REPLAY_AGGRESSIVE_FRACTION,
+            )
+          : replayBudget.thresholdTokens;
+      if (priorOverflowed) {
+        this.logger.warn(
+          `AI chat (chat ${chatId}): previous turn hit context overflow; ` +
+            `applying aggressive replay budget (${effectiveThreshold} tokens).`,
+        );
+      }
+      const preTrim = trimHistoryForReplay(
+        messages,
+        effectiveThreshold,
+        // A prior OVERFLOW means the provider count is stale/absent — force the
+        // char-estimate path by ignoring priorContextTokens on recovery.
+        priorOverflowed ? undefined : priorContextTokens,
+      );
+      messages = preTrim.messages;
+      // Observability (#490): record the budgeter's decision on the turn so the UI
+      // can surface "replay truncated at N tokens". Threaded into flushAssistant.
+      let replayTrimmedToTokens: number | undefined = preTrim.trimmed
+        ? preTrim.estimatedTokens
+        : undefined;
+      if (preTrim.trimmed) {
+        this.logger.log(
+          `AI chat (chat ${chatId}): replay history trimmed to ~${preTrim.estimatedTokens} ` +
+            `tokens (budget ${replayBudget.thresholdTokens}).`,
+        );
+      }
+
       // Build the external MCP toolset FIRST so the system prompt can carry each
       // connected server's admin-authored guidance (#180). Merge in admin-
       // configured external MCP tools (web search, etc.; §6.8). A down/slow
@@ -1658,6 +1725,7 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
                 maxContextTokens: resolved?.chatContextWindow,
                 pageChanged,
                 partsCache,
+                replayTrimmedToTokens,
               }),
             );
             // #184/#487: the RUN is finalized ALWAYS (never gated on the message).
@@ -1705,7 +1773,16 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
             // object, so the actual provider cause is clearly logged. Reuse the
             // shared formatter so provider error formatting stays unified.
             const e = error as { stack?: string };
-            const errorText = describeProviderError(error, String(error));
+            // #490 reactive branch: classify a CONTEXT-OVERFLOW rejection (the
+            // replayed history exceeded the model window). The overflowing turn had
+            // no prior usage to trigger preventive trimming, so we record a clear,
+            // distinguishable cause AND stamp the row so the NEXT turn's budgeter
+            // trims aggressively — the reactive recovery that un-bricks the chat.
+            const overflow = isContextOverflowError(error);
+            const providerError = describeProviderError(error, String(error));
+            const errorText = overflow
+              ? `${CONTEXT_OVERFLOW_ERROR_PREFIX} (${providerError})`
+              : providerError;
             this.logger.error(`AI chat stream error: ${errorText}`, e?.stack);
             // DIAGNOSTIC (Safari stream-drop investigation) — temporary: timing of
             // an error-terminated stream.
@@ -1724,6 +1801,8 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
                 error: errorText,
                 pageChanged,
                 partsCache,
+                replayTrimmedToTokens,
+                replayOverflow: overflow || undefined,
               }),
             );
             // #184: settle the RUN as failed, carrying the provider/transport cause.
@@ -2103,6 +2182,45 @@ export function chatStreamMetadata(
   return undefined;
 }
 
+/**
+ * The provider-reported context size of the most recent assistant turn, read from
+ * its persisted `metadata.contextTokens` (#490 replay budgeter's PRIMARY signal —
+ * the provider's own fact, not an estimate). Returns undefined for a chat with no
+ * assistant turn yet, or one whose last turn recorded no usage (e.g. it errored),
+ * in which case the budgeter falls back to the char-estimate.
+ */
+export function lastAssistantContextTokens(
+  history: ReadonlyArray<AiChatMessage>,
+): number | undefined {
+  for (let i = history.length - 1; i >= 0; i--) {
+    const row = history[i];
+    if (row.role !== 'assistant') continue;
+    const meta = (row.metadata ?? {}) as { contextTokens?: unknown };
+    const n = meta.contextTokens;
+    return typeof n === 'number' && Number.isFinite(n) && n > 0 ? n : undefined;
+  }
+  return undefined;
+}
+
+/**
+ * Whether the most recent assistant turn was rejected for CONTEXT OVERFLOW
+ * (#490): its row carries `metadata.replayOverflow` (stamped by the stream's
+ * onError). The next turn's budgeter reads this to trim aggressively — the
+ * reactive recovery. Only the LAST assistant turn matters (an older overflow was
+ * already recovered), so we stop at the first assistant row scanning backwards.
+ */
+export function lastAssistantReplayOverflow(
+  history: ReadonlyArray<AiChatMessage>,
+): boolean {
+  for (let i = history.length - 1; i >= 0; i--) {
+    const row = history[i];
+    if (row.role !== 'assistant') continue;
+    const meta = (row.metadata ?? {}) as { replayOverflow?: unknown };
+    return meta.replayOverflow === true;
+  }
+  return false;
+}
+
 /** The last message with role 'user' from a useChat payload, if any. */
 function lastUserMessage(
   messages: UIMessage[] | undefined,
@@ -2636,6 +2754,13 @@ export function flushAssistant(
     // Per-turn step->parts memo (#490): pass the SAME cache on every flush of a
     // turn so each finished step's output is stringified once, not once per flush.
     partsCache?: StepPartsCache;
+    // #490 observability: when the replay budgeter trimmed this turn's history,
+    // the (estimated) token size it trimmed to — the UI can show "replay truncated
+    // at N tokens". Omitted when nothing was trimmed.
+    replayTrimmedToTokens?: number;
+    // #490 reactive branch: set when the provider rejected this turn for context
+    // overflow. Stamped into metadata so the NEXT turn's budgeter trims aggressively.
+    replayOverflow?: boolean;
   },
 ): AssistantFlush {
   const finished = capturedSteps ?? [];
@@ -2674,6 +2799,9 @@ export function flushAssistant(
   if (extra?.contextTokens) metadata.contextTokens = extra.contextTokens;
   if (extra?.maxContextTokens)
     metadata.maxContextTokens = extra.maxContextTokens;
+  if (extra?.replayTrimmedToTokens)
+    metadata.replayTrimmedToTokens = extra.replayTrimmedToTokens;
+  if (extra?.replayOverflow) metadata.replayOverflow = true;
   if (extra?.error) metadata.error = extra.error;
   // Persist the page-change diff the agent saw this turn (#274 observability),
   // so history / the Markdown export can show what the user changed. Only when

apps/server/src/core/ai-chat/ai-chat.write-volume.spec.ts

@@ -0,0 +1,209 @@
+import { randomBytes } from 'crypto';
+import { Client } from 'pg';
+import { flushAssistant, serializeSteps } from './ai-chat.service';
+
+/**
+ * #490 write-volume regression — an OBSERVABLE-PROPERTY test on a LIVE Postgres,
+ * not "bytes through a mock repo" (a mock measures exactly the thing that does not
+ * hurt). It drives a realistic 50-step run where each step returns a ~100 KB tool
+ * output and, at every `onStepFinish`, UPDATEs the assistant row the way the
+ * service does — then reads the REAL write volume via the `pg_current_wal_lsn()`
+ * delta around the run.
+ *
+ * The property proven: v2 stores each tool OUTPUT only in `metadata.parts`, no
+ * longer ALSO in the `tool_calls` trace. So:
+ *   1. the trace (`tool_calls`) column's write volume is now O(Σ steps) — tiny,
+ *      linear outcome flags — vs the pre-#490 O(N²) that re-persisted every prior
+ *      output on every step; and
+ *   2. the FULL-row write volume drops sharply (the duplicated output copy is gone).
+ *
+ * Connects to the local gitmost test Postgres (docker `gitmost-test-pg` on :5432);
+ * SKIPS cleanly when that DB is not reachable so it never breaks a DB-less CI.
+ */
+const CONN =
+  process.env.WAL_TEST_DATABASE_URL ??
+  'postgresql://docmost:docmost_dev_pw@localhost:5432/docmost';
+
+// A step whose tool output is ~100 KB (a page read), in the SDK StepLike shape.
+// The body is INCOMPRESSIBLE random text — a `'x'.repeat()` filler would TOAST-
+// compress to nothing and hide the real write volume (a page body does not).
+function makeStep(i: number, outputBytes = 100_000) {
+  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 },
+      },
+    ],
+  };
+}
+
+// The pre-#490 (v1) trace: outputs stored a SECOND time in `tool_calls`
+// (the duplication #490 removed). Mirrors the OLD serializeSteps shape.
+function v1Trace(steps: ReturnType<typeof makeStep>[]): unknown {
+  const calls: unknown[] = [];
+  for (const s of steps) {
+    for (const c of s.toolCalls) calls.push({ toolName: c.toolName, input: c.input });
+    for (const r of s.toolResults)
+      calls.push({ toolName: r.toolName, output: r.output });
+  }
+  return calls;
+}
+
+async function walDelta(
+  client: Client,
+  fn: () => Promise<void>,
+): Promise<number> {
+  const before = (await client.query('SELECT pg_current_wal_lsn() AS l')).rows[0]
+    .l as string;
+  await fn();
+  // NOTE: do NOT pg_switch_wal() here — a segment switch pads the LSN to the next
+  // 16 MB boundary and would swamp the actual write delta. The raw LSN advances by
+  // the bytes of WAL emitted, which is exactly what we want to measure.
+  const after = (await client.query('SELECT pg_current_wal_lsn() AS l')).rows[0]
+    .l as string;
+  return Number(
+    (await client.query('SELECT pg_wal_lsn_diff($1,$2) AS d', [after, before]))
+      .rows[0].d,
+  );
+}
+
+describe('#490 write-volume on a live Postgres (pg_current_wal_lsn delta)', () => {
+  let client: Client | undefined;
+  let available = false;
+
+  beforeAll(async () => {
+    try {
+      client = new Client(CONN);
+      await client.connect();
+      await client.query('SELECT pg_current_wal_lsn()');
+      available = true;
+    } catch {
+      available = false;
+      client = undefined;
+    }
+  });
+
+  afterAll(async () => {
+    await client?.end().catch(() => undefined);
+  });
+
+  const STEPS = 50;
+
+  it('v2 trace write volume is O(Σ steps) — a tiny fraction of the v1 duplicate', async () => {
+    if (!available || !client) {
+      console.warn('SKIP: gitmost-test-pg not reachable; skipping WAL test.');
+      return;
+    }
+    const c = client;
+    // Isolated table so we measure only the tool_calls (trace) column's writes.
+    await c.query('DROP TABLE IF EXISTS _wal_trace');
+    await c.query('CREATE TABLE _wal_trace(id int primary key, tool_calls jsonb)');
+    await c.query("INSERT INTO _wal_trace VALUES (1, '[]'::jsonb)");
+
+    const steps: ReturnType<typeof makeStep>[] = [];
+
+    // v1: each step re-persists ALL prior outputs into the trace (the O(N²) churn).
+    const v1 = await walDelta(c, async () => {
+      const acc: ReturnType<typeof makeStep>[] = [];
+      for (let i = 0; i < STEPS; i++) {
+        acc.push(makeStep(i));
+        await c.query('UPDATE _wal_trace SET tool_calls=$1 WHERE id=1', [
+          JSON.stringify(v1Trace(acc)),
+        ]);
+      }
+      steps.push(...acc);
+    });
+
+    await c.query("UPDATE _wal_trace SET tool_calls='[]'::jsonb WHERE id=1");
+
+    // v2: the REAL serializeSteps — outcome flags only, NO outputs.
+    const v2 = await walDelta(c, async () => {
+      const acc: ReturnType<typeof makeStep>[] = [];
+      for (let i = 0; i < STEPS; i++) {
+        acc.push(makeStep(i));
+        await c.query('UPDATE _wal_trace SET tool_calls=$1 WHERE id=1', [
+          JSON.stringify(serializeSteps(acc)),
+        ]);
+      }
+    });
+
+    await c.query('DROP TABLE IF EXISTS _wal_trace');
+
+    // eslint-disable-next-line no-console
+    console.log(
+      `[#490 WAL] trace column over ${STEPS} steps: v1=${(v1 / 1e6).toFixed(1)}MB ` +
+        `v2=${(v2 / 1e6).toFixed(2)}MB (${(v1 / v2).toFixed(0)}x smaller)`,
+    );
+
+    // The trace no longer carries outputs: v2 is a tiny fraction of v1's WAL.
+    expect(v2).toBeLessThan(v1 * 0.1);
+    // And v2's trace WAL is small in absolute terms — O(Σ steps) of flags, not
+    // O(N² × output). 50 steps of ~40-byte flags is well under a few MB of WAL.
+    expect(v2).toBeLessThan(5_000_000);
+    // v1's duplicate alone is huge (≈ the 100 KB output re-written N² times).
+    expect(v1).toBeGreaterThan(50_000_000);
+  }, 120_000);
+
+  it('the full assistant row write drops sharply once the duplicate is gone', async () => {
+    if (!available || !client) return;
+    const c = client;
+    await c.query('DROP TABLE IF EXISTS _wal_full');
+    await c.query(
+      'CREATE TABLE _wal_full(id int primary key, content text, tool_calls jsonb, metadata jsonb, status text)',
+    );
+    await c.query("INSERT INTO _wal_full VALUES (1, '', '[]'::jsonb, '{}'::jsonb, 'streaming')");
+
+    const writeRow = async (patch: {
+      content: string;
+      toolCalls: unknown;
+      metadata: unknown;
+      status: string;
+    }) =>
+      c.query(
+        'UPDATE _wal_full SET content=$1, tool_calls=$2, metadata=$3, status=$4 WHERE id=1',
+        [
+          patch.content,
+          JSON.stringify(patch.toolCalls ?? null),
+          JSON.stringify(patch.metadata),
+          patch.status,
+        ],
+      );
+
+    // v2 (real flushAssistant): outputs live once, in metadata.parts.
+    const v2 = await walDelta(c, async () => {
+      const acc: ReturnType<typeof makeStep>[] = [];
+      for (let i = 0; i < STEPS; i++) {
+        acc.push(makeStep(i));
+        await writeRow(flushAssistant(acc as never, '', 'streaming'));
+      }
+    });
+
+    await c.query("UPDATE _wal_full SET content='', tool_calls='[]'::jsonb, metadata='{}'::jsonb WHERE id=1");
+
+    // v1: same row PLUS the duplicated outputs in the trace column.
+    const v1 = await walDelta(c, async () => {
+      const acc: ReturnType<typeof makeStep>[] = [];
+      for (let i = 0; i < STEPS; i++) {
+        acc.push(makeStep(i));
+        const f = flushAssistant(acc as never, '', 'streaming');
+        await writeRow({ ...f, toolCalls: v1Trace(acc) });
+      }
+    });
+
+    await c.query('DROP TABLE IF EXISTS _wal_full');
+
+    // eslint-disable-next-line no-console
+    console.log(
+      `[#490 WAL] full row over ${STEPS} steps: v1=${(v1 / 1e6).toFixed(1)}MB ` +
+        `v2=${(v2 / 1e6).toFixed(1)}MB (saved ${((1 - v2 / v1) * 100).toFixed(0)}%)`,
+    );
+
+    // Removing the duplicated trace copy is a large, real write-volume reduction.
+    expect(v2).toBeLessThan(v1 * 0.75);
+  }, 120_000);
+});

apps/server/src/core/ai-chat/history-budget.spec.ts

@@ -0,0 +1,266 @@
+import type { ModelMessage } from 'ai';
+import {
+  resolveReplayBudget,
+  isContextOverflowError,
+  estimateMessagesTokens,
+  trimHistoryForReplay,
+  REPLAY_BUDGET_DEFAULT_TOKENS,
+  REPLAY_TRUNCATION_MARKER,
+  REPLAY_TURN_COLLAPSED_MARKER,
+} from './history-budget';
+
+describe('resolveReplayBudget', () => {
+  it('uses min(default, 0.7 x window) for a configured window', () => {
+    // 0.7 x 60k = 42k < 100k
+    expect(resolveReplayBudget(60_000)).toEqual({
+      thresholdTokens: 42_000,
+      usedDefault: false,
+    });
+    // 0.7 x 1M = 700k, capped to the 100k default
+    expect(resolveReplayBudget(1_000_000)).toEqual({
+      thresholdTokens: REPLAY_BUDGET_DEFAULT_TOKENS,
+      usedDefault: false,
+    });
+  });
+
+  it('accepts the raw ::text stored form', () => {
+    expect(resolveReplayBudget('60000').thresholdTokens).toBe(42_000);
+  });
+
+  // The crux (#490): a chat with NO context window configured must STILL be
+  // budgeted — those are exactly the installs that hit terminal overflow.
+  it('applies the flat default when the window is unset/empty', () => {
+    expect(resolveReplayBudget(undefined)).toEqual({
+      thresholdTokens: REPLAY_BUDGET_DEFAULT_TOKENS,
+      usedDefault: true,
+    });
+    expect(resolveReplayBudget('')).toEqual({
+      thresholdTokens: REPLAY_BUDGET_DEFAULT_TOKENS,
+      usedDefault: true,
+    });
+    expect(resolveReplayBudget('   ')).toEqual({
+      thresholdTokens: REPLAY_BUDGET_DEFAULT_TOKENS,
+      usedDefault: true,
+    });
+  });
+
+  it('treats an explicit 0 as the off-switch (distinct from unset)', () => {
+    expect(resolveReplayBudget(0)).toEqual({
+      thresholdTokens: null,
+      usedDefault: false,
+    });
+    expect(resolveReplayBudget('0')).toEqual({
+      thresholdTokens: null,
+      usedDefault: false,
+    });
+  });
+
+  it('falls back to the default on a negative/garbage value', () => {
+    expect(resolveReplayBudget(-5).usedDefault).toBe(true);
+    expect(resolveReplayBudget('abc').usedDefault).toBe(true);
+  });
+});
+
+describe('isContextOverflowError', () => {
+  it('classifies a real provider 400 context-overflow shape', () => {
+    // OpenAI-compatible shape.
+    expect(
+      isContextOverflowError({
+        statusCode: 400,
+        message:
+          "This model's maximum context length is 128000 tokens. However, your messages resulted in 214000 tokens. Please reduce the length of the messages.",
+      }),
+    ).toBe(true);
+    // Anthropic-style wording.
+    expect(
+      isContextOverflowError({
+        status: 400,
+        message: 'prompt is too long: 250000 tokens > 200000 maximum',
+      }),
+    ).toBe(true);
+    // Nested body + string status.
+    expect(
+      isContextOverflowError({
+        response: { status: '400' },
+        message: 'input is too long for the requested model',
+      }),
+    ).toBe(true);
+    // Error instance with the cause carrying the body.
+    const e = new Error('Bad request');
+    (e as any).statusCode = 400;
+    (e as any).cause = new Error('maximum context window exceeded');
+    expect(isContextOverflowError(e)).toBe(true);
+  });
+
+  it('does NOT classify unrelated 400s or auth/rate-limit errors', () => {
+    expect(
+      isContextOverflowError({ statusCode: 400, message: 'invalid tool schema' }),
+    ).toBe(false);
+    expect(
+      isContextOverflowError({
+        statusCode: 429,
+        message: 'context length exceeded but rate limited',
+      }),
+    ).toBe(false);
+    expect(isContextOverflowError({ statusCode: 500, message: 'server error' })).toBe(
+      false,
+    );
+    expect(isContextOverflowError(undefined)).toBe(false);
+    expect(isContextOverflowError('some random string')).toBe(false);
+  });
+});
+
+// Helpers to build ModelMessage fixtures in the ai@6 shape.
+const userMsg = (text: string): ModelMessage =>
+  ({ role: 'user', content: [{ type: 'text', text }] }) as ModelMessage;
+const assistantMsg = (
+  text: string,
+  toolCallId?: string,
+  toolName?: string,
+): ModelMessage =>
+  ({
+    role: 'assistant',
+    content: [
+      { type: 'text', text },
+      ...(toolCallId
+        ? [{ type: 'tool-call', toolCallId, toolName, input: {} }]
+        : []),
+    ],
+  }) as ModelMessage;
+const toolMsg = (
+  toolCallId: string,
+  toolName: string,
+  value: unknown,
+): ModelMessage =>
+  ({
+    role: 'tool',
+    content: [
+      { type: 'tool-result', toolCallId, toolName, output: { type: 'json', value } },
+    ],
+  }) as ModelMessage;
+
+describe('trimHistoryForReplay', () => {
+  it('null budget disables trimming (returns the same reference)', () => {
+    const msgs = [userMsg('hi'), assistantMsg('yo')];
+    const r = trimHistoryForReplay(msgs, null);
+    expect(r.trimmed).toBe(false);
+    expect(r.messages).toBe(msgs);
+  });
+
+  it('leaves history under budget untouched (same reference)', () => {
+    const msgs = [userMsg('hi'), assistantMsg('a short answer')];
+    const r = trimHistoryForReplay(msgs, 100_000);
+    expect(r.trimmed).toBe(false);
+    expect(r.messages).toBe(msgs);
+  });
+
+  it('truncates OLD tool outputs but keeps recent turns full', () => {
+    const big = 'X'.repeat(40_000); // ~16k tokens on its own
+    const msgs: ModelMessage[] = [];
+    // 6 OLD turns (indices 0..5), each with a huge tool output.
+    for (let i = 0; i < 6; i++) {
+      msgs.push(userMsg(`old q${i}`));
+      msgs.push(assistantMsg('looking', `c${i}`, 'getPage'));
+      msgs.push(toolMsg(`c${i}`, 'getPage', { body: big }));
+      msgs.push(assistantMsg(`old a${i}`));
+    }
+    // 3 small recent turns, then the CURRENT turn with its own huge tool output.
+    // With REPLAY_KEEP_RECENT_TURNS=4 the last 4 user-turns stay full, so only
+    // these small recent turns + the current big one are kept full; the 6 old
+    // turns above fall in the trim region.
+    for (let i = 0; i < 3; i++) {
+      msgs.push(userMsg(`recent q${i}`));
+      msgs.push(assistantMsg(`recent a${i}`));
+    }
+    msgs.push(userMsg('current q'));
+    msgs.push(assistantMsg('looking', 'cR', 'getPage'));
+    msgs.push(toolMsg('cR', 'getPage', { body: big }));
+    msgs.push(assistantMsg('current a'));
+
+    // Budget large enough that phase-1 tool truncation alone brings it under.
+    const r = trimHistoryForReplay(msgs, 30_000);
+    expect(r.trimmed).toBe(true);
+    const flat = JSON.stringify(r.messages);
+    // The CURRENT turn's tool output survives in full.
+    expect(flat).toContain(big);
+    // Old outputs were truncated with the marker.
+    expect(flat).toContain(REPLAY_TRUNCATION_MARKER);
+    // Phase 1 sufficed: the oldest turns were NOT collapsed.
+    expect(flat).not.toContain(REPLAY_TURN_COLLAPSED_MARKER);
+    expect(estimateMessagesTokens(r.messages)).toBeLessThan(
+      estimateMessagesTokens(msgs),
+    );
+  });
+
+  it('collapses the oldest turns when tool truncation is not enough', () => {
+    // Many turns with LARGE assistant TEXT (not tool output) so phase 1 can't help.
+    const bigText = 'слово '.repeat(8_000); // large Cyrillic text per turn
+    const msgs: ModelMessage[] = [];
+    for (let i = 0; i < 12; i++) {
+      msgs.push(userMsg(`q${i}`));
+      msgs.push(assistantMsg(bigText));
+    }
+    const r = trimHistoryForReplay(msgs, 30_000);
+    expect(r.trimmed).toBe(true);
+    // Oldest turns collapsed; result fits (best-effort) and is much smaller.
+    expect(estimateMessagesTokens(r.messages)).toBeLessThan(
+      estimateMessagesTokens(msgs),
+    );
+    // The LAST turn's text is preserved in full (recent turns stay full).
+    expect(JSON.stringify(r.messages[r.messages.length - 1])).toContain(bigText);
+  });
+
+  it('is deterministic / byte-stable for identical inputs', () => {
+    const big = 'Y'.repeat(30_000);
+    const build = (): ModelMessage[] => {
+      const m: ModelMessage[] = [];
+      for (let i = 0; i < 10; i++) {
+        m.push(userMsg(`q${i}`));
+        m.push(assistantMsg('t', `c${i}`, 'getPage'));
+        m.push(toolMsg(`c${i}`, 'getPage', { body: big }));
+      }
+      return m;
+    };
+    const a = trimHistoryForReplay(build(), 15_000);
+    const b = trimHistoryForReplay(build(), 15_000);
+    expect(JSON.stringify(a.messages)).toBe(JSON.stringify(b.messages));
+  });
+
+  it('never leaves an unpaired tool-call after collapsing (balanced history)', () => {
+    const big = 'Z'.repeat(40_000);
+    const msgs: ModelMessage[] = [];
+    for (let i = 0; i < 10; i++) {
+      msgs.push(userMsg(`q${i}`));
+      msgs.push(assistantMsg('t', `c${i}`, 'getPage'));
+      msgs.push(toolMsg(`c${i}`, 'getPage', { body: big }));
+    }
+    const r = trimHistoryForReplay(msgs, 8_000);
+    // Count tool-call vs tool-result parts in the trimmed output.
+    let calls = 0;
+    let results = 0;
+    for (const m of r.messages) {
+      if (!Array.isArray(m.content)) continue;
+      for (const p of m.content as Array<{ type?: string }>) {
+        if (p.type === 'tool-call') calls++;
+        if (p.type === 'tool-result' || p.type === 'tool-error') results++;
+      }
+    }
+    // Every surviving tool-call has a surviving result (collapsing drops BOTH).
+    expect(calls).toBe(results);
+    // Collapsed turns carry the marker.
+    expect(JSON.stringify(r.messages)).toContain(REPLAY_TURN_COLLAPSED_MARKER);
+  });
+
+  it('respects the provider fact: under-budget contextTokens skips trimming', () => {
+    const big = 'W'.repeat(60_000);
+    const msgs = [
+      userMsg('q'),
+      assistantMsg('t', 'c1', 'getPage'),
+      toolMsg('c1', 'getPage', { body: big }),
+    ];
+    // char-estimate is high, but the provider says we are well under budget.
+    const r = trimHistoryForReplay(msgs, 100_000, 5_000);
+    expect(r.trimmed).toBe(false);
+    expect(r.messages).toBe(msgs);
+  });
+});

apps/server/src/core/ai-chat/history-budget.ts

@@ -0,0 +1,356 @@
+/**
+ * History-replay token budget (#490).
+ *
+ * The whole persisted conversation is replayed to the provider on EVERY turn, so
+ * a long chat eventually exceeds the model's context window and the provider 400s
+ * on every turn — terminally (the chat "bricks"). This module bounds the replayed
+ * history at REPLAY TIME only: it never mutates what is persisted (the DB stays
+ * the full record), and its output is a deterministic, byte-stable function of its
+ * input so the trimmed prefix is identical turn to turn (provider prompt-cache
+ * friendliness — real money on long chats).
+ *
+ * The PRIMARY signal is the provider's own fact: `metadata.contextTokens` from the
+ * last turn. The chars-based {@link estimateTokens} (shared with the client) is
+ * used only for the DELTA of not-yet-sent messages, to decide WHAT to trim, and as
+ * the fallback for chats with no usage yet.
+ */
+import type { ModelMessage } from 'ai';
+import { estimateTokens } from '@docmost/token-estimate';
+
+/** Flat default budget when no context window is configured (tokens). */
+export const REPLAY_BUDGET_DEFAULT_TOKENS = 100_000;
+/** Fraction of a configured context window used as the budget. */
+export const REPLAY_BUDGET_WINDOW_FRACTION = 0.7;
+/**
+ * Fraction of the normal budget used for the REACTIVE re-trim after a provider
+ * context-overflow 400 — the preventive estimate under-counted, so cut harder.
+ */
+export const REPLAY_AGGRESSIVE_FRACTION = 0.5;
+/**
+ * Turns (a user message + its assistant/tool replies) kept FULL at the tail,
+ * including the current one — never trimmed. Older turns are compacted first.
+ */
+export const REPLAY_KEEP_RECENT_TURNS = 4;
+/** Leading chars kept from a truncated old tool output. */
+export const REPLAY_TOOL_OUTPUT_HEAD = 800;
+/** Trailing chars kept from a truncated old tool output. */
+export const REPLAY_TOOL_OUTPUT_TAIL = 300;
+/** Marker inserted where an old tool output was truncated for replay. */
+export const REPLAY_TRUNCATION_MARKER =
+  '[…truncated for replay; call the tool again to read the full output]';
+/** Marker for a whole old turn collapsed to its text. */
+export const REPLAY_TURN_COLLAPSED_MARKER =
+  '[earlier tool activity omitted for replay]';
+
+export interface ReplayBudget {
+  /** Token threshold above which replay history is trimmed; `null` = OFF. */
+  thresholdTokens: number | null;
+  /** True when the flat default was used (no context window configured). */
+  usedDefault: boolean;
+}
+
+/**
+ * Resolve the replay budget from the RAW stored `chatContextWindow` (text/number).
+ *  - a positive value  -> `min(default, floor(fraction × window))`
+ *  - explicit `0`       -> OFF (admin opt-out; `null` threshold)
+ *  - unset/empty/invalid-> the flat default (still protects — the installations
+ *    that hit terminal overflow are exactly the ones that never set a window)
+ *
+ * Note the raw value is needed because the parsed `chatContextWindow` collapses
+ * both `0` and unset to `undefined`, which would erase the explicit off-switch.
+ */
+export function resolveReplayBudget(rawContextWindow: unknown): ReplayBudget {
+  let n: number | undefined;
+  if (typeof rawContextWindow === 'number') {
+    n = rawContextWindow;
+  } else if (typeof rawContextWindow === 'string') {
+    const t = rawContextWindow.trim();
+    n = t === '' ? undefined : Number(t);
+  }
+  // Unset / empty / non-numeric / negative -> flat default (the protective case).
+  if (n === undefined || !Number.isFinite(n) || n < 0) {
+    return { thresholdTokens: REPLAY_BUDGET_DEFAULT_TOKENS, usedDefault: true };
+  }
+  // Explicit 0 -> off-switch.
+  if (n === 0) {
+    return { thresholdTokens: null, usedDefault: false };
+  }
+  return {
+    thresholdTokens: Math.min(
+      REPLAY_BUDGET_DEFAULT_TOKENS,
+      Math.floor(REPLAY_BUDGET_WINDOW_FRACTION * n),
+    ),
+    usedDefault: false,
+  };
+}
+
+/**
+ * True when a provider error is a CONTEXT-OVERFLOW rejection (the prompt exceeds
+ * the model's window). Providers surface this as an HTTP 400 with a recognizable
+ * message; match both the status and the message patterns robustly across
+ * OpenAI-compatible / Anthropic / Gemini wordings, since the exact shape varies.
+ */
+export function isContextOverflowError(error: unknown): boolean {
+  const status = extractStatus(error);
+  const msg = extractMessage(error).toLowerCase();
+  // Message patterns seen across providers for "prompt too long".
+  const overflowPattern =
+    /context (?:length|window)|maximum context|too many tokens|too large for|reduce the length|prompt is too long|input (?:is )?too long|exceeds? the (?:maximum )?(?:context|token)|maximum.*tokens|string too long/;
+  if (!overflowPattern.test(msg)) return false;
+  // A 400/413 with an overflow-shaped message is an overflow. Some providers
+  // omit/rewrite the status, so accept the message match when the status is
+  // unknown, but reject it for auth/rate-limit statuses that never mean overflow.
+  if (status === 400 || status === 413) return true;
+  if (status === 401 || status === 403 || status === 429) return false;
+  return true;
+}
+
+function extractStatus(error: unknown): number | undefined {
+  if (!error || typeof error !== 'object') return undefined;
+  const e = error as Record<string, unknown>;
+  for (const k of ['statusCode', 'status']) {
+    const v = e[k];
+    if (typeof v === 'number') return v;
+    if (typeof v === 'string' && /^\d+$/.test(v)) return Number(v);
+  }
+  // Nested (e.g. { response: { status } } / { cause: { statusCode } }).
+  for (const k of ['response', 'cause', 'data']) {
+    const nested = e[k];
+    if (nested && typeof nested === 'object') {
+      const s = extractStatus(nested);
+      if (s !== undefined) return s;
+    }
+  }
+  return undefined;
+}
+
+function extractMessage(error: unknown): string {
+  if (error == null) return '';
+  if (typeof error === 'string') return error;
+  if (error instanceof Error) {
+    // Include nested causes (provider libs wrap the real body in `cause`).
+    const cause = (error as { cause?: unknown }).cause;
+    return `${error.message} ${cause ? extractMessage(cause) : ''}`;
+  }
+  if (typeof error === 'object') {
+    const e = error as Record<string, unknown>;
+    const parts: string[] = [];
+    for (const k of ['message', 'error', 'body', 'responseBody', 'data']) {
+      const v = e[k];
+      if (typeof v === 'string') parts.push(v);
+      else if (v && typeof v === 'object') parts.push(extractMessage(v));
+    }
+    return parts.join(' ');
+  }
+  return String(error);
+}
+
+/** Rough token size of a ModelMessage array via the shared chars estimator. */
+export function estimateMessagesTokens(
+  messages: ReadonlyArray<ModelMessage>,
+): number {
+  let total = 0;
+  for (const m of messages) {
+    total += estimateTokens(serializeContent(m.content));
+  }
+  return total;
+}
+
+function serializeContent(content: unknown): string {
+  if (typeof content === 'string') return content;
+  try {
+    return JSON.stringify(content) ?? '';
+  } catch {
+    return '';
+  }
+}
+
+/** Deep JSON string of an arbitrary value, bounded so estimation never throws. */
+function stringifyValue(value: unknown): string {
+  if (typeof value === 'string') return value;
+  try {
+    return JSON.stringify(value) ?? String(value);
+  } catch {
+    return String(value);
+  }
+}
+
+export interface TrimResult {
+  messages: ModelMessage[];
+  /** Whether any trimming was applied. */
+  trimmed: boolean;
+  /** Estimated tokens of the returned messages (chars-based). */
+  estimatedTokens: number;
+}
+
+/**
+ * Bound the replayed history to `budgetTokens`, deterministically. Returns the
+ * SAME array reference (no copy) when nothing needs trimming, so the common case
+ * is free and byte-identical. Trimming order (spec #490):
+ *   1. truncate OLD turns' tool outputs (head+tail + marker) — the bulk of the size
+ *   2. mechanically collapse the OLDEST turns to their text (concatenation, no LLM)
+ *   3. the current + last {@link REPLAY_KEEP_RECENT_TURNS} turns stay FULL
+ *
+ * `budgetTokens === null` disables trimming. `priorContextTokens` (the provider's
+ * fact from last turn) short-circuits the decision: when it is known and already
+ * under budget we skip trimming even if the char-estimate is higher (the provider
+ * count is authoritative). The char-estimate drives WHAT to cut.
+ */
+export function trimHistoryForReplay(
+  messages: ModelMessage[],
+  budgetTokens: number | null,
+  priorContextTokens?: number,
+): TrimResult {
+  if (budgetTokens == null) {
+    return { messages, trimmed: false, estimatedTokens: 0 };
+  }
+  const estimated = estimateMessagesTokens(messages);
+  // Decision signal: prefer the provider's fact (last turn's contextTokens) plus
+  // the estimated delta of the messages appended since; fall back to the pure
+  // char-estimate for a chat with no usage yet.
+  const projected =
+    priorContextTokens != null
+      ? Math.max(priorContextTokens, estimated)
+      : estimated;
+  if (projected <= budgetTokens) {
+    return { messages, trimmed: false, estimatedTokens: estimated };
+  }
+
+  // The tail we always keep full: from the Nth-from-last user message onward.
+  const boundary = recentBoundaryIndex(messages, REPLAY_KEEP_RECENT_TURNS);
+  const tail = messages.slice(boundary);
+  let head = messages.slice(0, boundary).map(cloneMessage);
+
+  // Phase 1: truncate old tool outputs.
+  for (const m of head) {
+    if (m.role === 'tool') truncateToolMessage(m);
+  }
+  let out = [...head, ...tail];
+  let est = estimateMessagesTokens(out);
+  if (est <= budgetTokens) {
+    return { messages: out, trimmed: true, estimatedTokens: est };
+  }
+
+  // Phase 2: collapse the oldest turns (in `head`) to their text, one at a time,
+  // from the oldest, until we fit or the whole head is collapsed.
+  const turns = splitTurns(head);
+  const collapsed: ModelMessage[] = [];
+  let i = 0;
+  for (; i < turns.length; i++) {
+    if (est <= budgetTokens) break;
+    collapsed.push(...collapseTurn(turns[i]));
+    // Re-estimate the whole prospective output.
+    const remaining = turns.slice(i + 1).flat();
+    out = [...collapsed, ...remaining, ...tail];
+    est = estimateMessagesTokens(out);
+  }
+  // Include any turns we didn't need to collapse.
+  const remaining = turns.slice(i).flat();
+  out = [...collapsed, ...remaining, ...tail];
+  est = estimateMessagesTokens(out);
+  return { messages: out, trimmed: true, estimatedTokens: est };
+}
+
+/** Index of the first message of the Nth-from-last user turn (0 if fewer). */
+function recentBoundaryIndex(
+  messages: ReadonlyArray<ModelMessage>,
+  keepTurns: number,
+): number {
+  const userIdx: number[] = [];
+  for (let i = 0; i < messages.length; i++) {
+    if (messages[i].role === 'user') userIdx.push(i);
+  }
+  if (userIdx.length <= keepTurns) return 0;
+  return userIdx[userIdx.length - keepTurns];
+}
+
+/** Split a message list into turns; each turn starts at a `user` message. */
+function splitTurns(messages: ModelMessage[]): ModelMessage[][] {
+  const turns: ModelMessage[][] = [];
+  for (const m of messages) {
+    if (m.role === 'user' || turns.length === 0) turns.push([m]);
+    else turns[turns.length - 1].push(m);
+  }
+  return turns;
+}
+
+/**
+ * Collapse a whole turn to its plain text (mechanical concatenation, not an LLM
+ * summary). Keeps the user message; replaces the assistant/tool messages with a
+ * single assistant text message = the assistant's concatenated text + a marker
+ * when tool activity was dropped. Dropping BOTH the tool-call and tool-result
+ * parts together keeps the rebuilt history balanced (no unpaired calls).
+ */
+function collapseTurn(turn: ModelMessage[]): ModelMessage[] {
+  const out: ModelMessage[] = [];
+  let assistantText = '';
+  let hadTools = false;
+  for (const m of turn) {
+    if (m.role === 'user') {
+      out.push(m);
+    } else if (m.role === 'assistant') {
+      const { text, tools } = extractAssistantText(m.content);
+      assistantText += text;
+      hadTools = hadTools || tools;
+    } else if (m.role === 'tool') {
+      hadTools = true;
+    } else {
+      out.push(m);
+    }
+  }
+  const note =
+    (assistantText ? assistantText.trimEnd() : '') +
+    (hadTools
+      ? `${assistantText ? '\n\n' : ''}${REPLAY_TURN_COLLAPSED_MARKER}`
+      : '');
+  if (note) out.push({ role: 'assistant', content: note } as ModelMessage);
+  return out;
+}
+
+function extractAssistantText(content: unknown): {
+  text: string;
+  tools: boolean;
+} {
+  if (typeof content === 'string') return { text: content, tools: false };
+  if (!Array.isArray(content)) return { text: '', tools: false };
+  let text = '';
+  let tools = false;
+  for (const part of content) {
+    const type = (part as { type?: string })?.type;
+    if (type === 'text') text += (part as { text?: string }).text ?? '';
+    else if (type === 'tool-call') tools = true;
+  }
+  return { text, tools };
+}
+
+/** Truncate every tool-result output in a `tool` message to head+tail+marker. */
+function truncateToolMessage(message: ModelMessage): void {
+  const content = message.content;
+  if (!Array.isArray(content)) return;
+  for (const part of content) {
+    const p = part as { type?: string; output?: { type?: string; value?: unknown } };
+    if (p.type !== 'tool-result' && p.type !== 'tool-error') continue;
+    if (!p.output) continue;
+    const raw = stringifyValue(p.output.value);
+    const budget = REPLAY_TOOL_OUTPUT_HEAD + REPLAY_TOOL_OUTPUT_TAIL;
+    if (raw.length <= budget + REPLAY_TRUNCATION_MARKER.length) continue;
+    const truncated =
+      raw.slice(0, REPLAY_TOOL_OUTPUT_HEAD) +
+      `\n${REPLAY_TRUNCATION_MARKER}\n` +
+      raw.slice(raw.length - REPLAY_TOOL_OUTPUT_TAIL);
+    // Represent the shrunk output as a text output (a valid tool-result output).
+    p.output = { type: 'text', value: truncated };
+  }
+}
+
+/** Shallow-ish clone so trimming never mutates the caller's (persisted-derived)
+ *  message objects — only the OLD region is cloned before it is edited. */
+function cloneMessage(m: ModelMessage): ModelMessage {
+  if (typeof m.content === 'string') return { ...m };
+  return {
+    ...m,
+    content: (m.content as unknown[]).map((p) =>
+      p && typeof p === 'object' ? { ...(p as object) } : p,
+    ),
+  } as ModelMessage;
+}

apps/server/src/integrations/ai/ai-settings.service.ts

@@ -245,6 +245,9 @@ export class AiSettingsService {
       // Max context window for the chat header badge denominator. Stored as
       // ::text; 0/unset/invalid = no limit (undefined).
       chatContextWindow: parsePositiveInt(provider.chatContextWindow),
+      // RAW stored value (#490): the replay budgeter reads this to distinguish an
+      // explicit `0` (off-switch) from unset, which parsePositiveInt cannot.
+      chatContextWindowRaw: provider.chatContextWindow,
       // Plain passthrough; getChatModel defaults unset to 'openai-compatible'.
       chatApiStyle: provider.chatApiStyle,
       // Cheap model id for the anonymous public-share assistant; reuses the chat

apps/server/src/integrations/ai/ai.types.ts

@@ -105,6 +105,10 @@ export interface ResolvedAiConfig extends Partial<AiProviderSettings> {
   // Max context window in tokens; surfaced to the chat header badge as the
   // "current / max" denominator. 0/unset = no limit.
   chatContextWindow?: number;
+  // RAW stored context window (::text), BEFORE parsePositiveInt collapses `0` and
+  // unset to `undefined`. The #490 replay budgeter needs the raw value to honor an
+  // explicit `0` off-switch distinctly from "unset -> flat default".
+  chatContextWindowRaw?: string | number;
   // Cheap model id for the public-share assistant; reuses the chat creds.
   publicShareChatModel?: string;
   // Agent-role id whose persona the public-share assistant adopts (empty/unset

packages/token-estimate/package.json

@@ -0,0 +1,19 @@
+{
+  "name": "@docmost/token-estimate",
+  "version": "0.1.0",
+  "description": "Shared, provider-agnostic token estimator (chars/2.5) used by the AI-chat client counter and the server history-replay budgeter, so the two never diverge.",
+  "private": true,
+  "main": "./dist/index.js",
+  "types": "./dist/index.d.ts",
+  "scripts": {
+    "build": "tsc",
+    "watch": "tsc --watch",
+    "test": "vitest run",
+    "test:watch": "vitest"
+  },
+  "license": "MIT",
+  "devDependencies": {
+    "typescript": "^5.0.0",
+    "vitest": "4.1.6"
+  }
+}

packages/token-estimate/src/index.test.ts

@@ -0,0 +1,31 @@
+import { describe, it, expect } from 'vitest';
+import { estimateTokens, CHARS_PER_TOKEN } from './index';
+
+describe('estimateTokens (shared chars/2.5)', () => {
+  it('returns 0 for empty / nullish input', () => {
+    expect(estimateTokens('')).toBe(0);
+    expect(estimateTokens(null)).toBe(0);
+    expect(estimateTokens(undefined)).toBe(0);
+  });
+
+  it('uses the chars/2.5 ratio, ceiled', () => {
+    expect(CHARS_PER_TOKEN).toBe(2.5);
+    // 5 chars / 2.5 = 2
+    expect(estimateTokens('abcde')).toBe(2);
+    // any non-empty string is at least 1 token (ceil)
+    expect(estimateTokens('a')).toBe(1);
+    // 100 chars / 2.5 = 40
+    expect(estimateTokens('x'.repeat(100))).toBe(40);
+  });
+
+  it('counts Cyrillic ~2x higher than the old chars/4 rule (no undercount)', () => {
+    const cyr = 'привет мир как дела'; // 19 chars
+    expect(estimateTokens(cyr)).toBe(Math.ceil(19 / 2.5)); // 8
+    expect(estimateTokens(cyr)).toBeGreaterThan(Math.ceil(19 / 4)); // > 5
+  });
+
+  it('is deterministic / byte-stable (same input => same output)', () => {
+    const s = 'the quick brown fox';
+    expect(estimateTokens(s)).toBe(estimateTokens(s));
+  });
+});

packages/token-estimate/src/index.ts

@@ -0,0 +1,35 @@
+/**
+ * Shared, provider-agnostic token estimator (#490).
+ *
+ * No provider exposes an exact tokenizer we can afford to run on the hot path (a
+ * real BPE pass is O(n²)-ish, bloats the client bundle, and is wrong for
+ * Gemini/Ollama anyway), so both the client's in-body counter AND the server's
+ * history-replay budgeter use this ONE cheap chars-based heuristic. Keeping it in
+ * a single shared module is deliberate: two independent estimators drift, and then
+ * "the badge shows 60%" while "the budgeter already trimmed" — the exact confusion
+ * this package prevents.
+ *
+ * Ratio: **chars / 2.5**. Most content here is Cyrillic, where a token is ~2.5
+ * characters; the common English `chars/4` rule of thumb UNDER-counts Cyrillic by
+ * ~2×, which for a budget check is the dangerous direction (it lets the context
+ * overflow). 2.5 slightly over-estimates pure English/code, which is the SAFE
+ * direction for a budget. This is an estimate, never an exact count — the
+ * authoritative figure is always the provider's reported usage; the estimate is
+ * for UI affordances, the delta of not-yet-sent messages, and deciding what to
+ * trim.
+ */
+
+/** Characters per token for the shared estimate. See the module comment. */
+export const CHARS_PER_TOKEN = 2.5;
+
+/**
+ * Rough token estimate for a piece of text (chars / {@link CHARS_PER_TOKEN}).
+ * Returns 0 for empty/nullish input, and ceils so any non-empty text counts as at
+ * least one token. Pure and deterministic (byte-stable), so the same text always
+ * yields the same estimate — which the server budgeter relies on to keep replay
+ * trimming stable turn to turn (provider prompt-cache friendliness).
+ */
+export function estimateTokens(text: string | null | undefined): number {
+  if (!text) return 0;
+  return Math.ceil(text.length / CHARS_PER_TOKEN);
+}

packages/token-estimate/tsconfig.json

@@ -0,0 +1,16 @@
+{
+  "compilerOptions": {
+    "target": "ES2022",
+    "module": "CommonJS",
+    "moduleResolution": "Node",
+    "outDir": "./dist",
+    "rootDir": "./src",
+    "strict": true,
+    "esModuleInterop": true,
+    "skipLibCheck": true,
+    "forceConsistentCasingInFileNames": true,
+    "declaration": true
+  },
+  "include": ["src/**/*"],
+  "exclude": ["src/**/*.test.ts"]
+}

pnpm-lock.yaml

@@ -287,6 +287,9 @@ importers:
       '@docmost/prosemirror-markdown':
         specifier: workspace:*
         version: link:../../packages/prosemirror-markdown
+      '@docmost/token-estimate':
+        specifier: workspace:*
+        version: link:../../packages/token-estimate
       '@excalidraw/excalidraw':
         specifier: 0.18.0-3a5ef40
         version: 0.18.0-3a5ef40(@types/react-dom@18.3.1)(@types/react@18.3.12)(react-dom@18.3.1(react@18.3.1))(react@18.3.1)
@@ -561,6 +564,9 @@ importers:
       '@docmost/prosemirror-markdown':
         specifier: workspace:*
         version: link:../../packages/prosemirror-markdown
+      '@docmost/token-estimate':
+        specifier: workspace:*
+        version: link:../../packages/token-estimate
       '@fastify/compress':
         specifier: ^9.0.0
         version: 9.0.0
@@ -1148,6 +1154,15 @@ importers:
         specifier: 4.1.6
         version: 4.1.6(@opentelemetry/api@1.9.0)(@types/node@20.19.43)(@vitest/coverage-v8@4.1.6)(happy-dom@20.8.9)(jsdom@25.0.0)(vite@8.0.5(@types/node@20.19.43)(esbuild@0.28.0)(jiti@2.4.2)(less@4.2.0)(sugarss@5.0.1(postcss@8.5.14))(terser@5.39.0)(tsx@4.21.0)(yaml@2.8.3))
 
+  packages/token-estimate:
+    devDependencies:
+      typescript:
+        specifier: ^5.0.0
+        version: 5.9.3
+      vitest:
+        specifier: 4.1.6
+        version: 4.1.6(@opentelemetry/api@1.9.0)(@types/node@25.5.0)(@vitest/coverage-v8@4.1.6)(happy-dom@20.8.9)(jsdom@27.4.0(@noble/hashes@2.0.1))(vite@8.0.5(@types/node@25.5.0)(esbuild@0.28.0)(jiti@2.4.2)(less@4.2.0)(sugarss@5.0.1(postcss@8.5.14))(terser@5.39.0)(tsx@4.21.0)(yaml@2.8.3))
+
 packages:
 
   '@aashutoshrathi/word-wrap@1.2.6':