Merge remote-tracking branch 'gitea/develop' into fix/520-recovery-escalation

Остаточный кирпич (#490/#510): при незаданном chatContextWindow база = плоский
дефолт 100k, а реальное окно модели маленькое (<50k). Фиксированный одинарный
cut 0.5×100k=50k всё равно превышал реальное окно → провайдер снова 400,
строка переставлялась replayOverflow, но булев priorOverflowed уже был true →
второго ужатия не происходило. Чат навсегда застревал на 50k и не восстанавливался.

Фикс (без парсинга тел 400):
- Сигнал из булева переведён в счётчик `metadata.replayOverflowCount` = число
  ПОДРЯД идущих overflow-ходов: инкремент (prior+1) на каждом overflow, сброс в 0
  на любом чистом финализе (чистая строка не пишет поле → читается как 0).
  BACK-COMPAT: старая строка с булевым `replayOverflow:true` читается как k=1.
- resolveEffectiveReplayThreshold(threshold, k) = max(floor(threshold·0.5**k),
  min(REPLAY_MIN_FLOOR_TOKENS, threshold)). k=0 → база; k=1 → 0.5×; k=2 → 0.25×;
  большой k → упирается в пол 8k (сходимость). null-база (trimming OFF) не трогается;
  пол никогда не поднимает легитимно малый настроенный бюджет выше него самого.
- Пол REPLAY_MIN_FLOOR_TOKENS=8k: ниже него чат не несёт осмысленный недавний
  контекст, и даже малое реальное окно его вмещает; keep-recent-turns сверху.

Тесты: таблица эскалации (k=0/1/2/большой/null), регрессия остаточного кирпича
(база 100k, окно ~40k → сходится ниже 40k, чего фиксированный 0.5× никогда не мог),
жизненный цикл счётчика на реальном pg (инкремент/сброс/back-compat через jsonb),
мутация **k→**1 краснит convergence-тест.

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

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

@@ -370,10 +370,12 @@ describe('AiChatService.stream — abortSignal wiring (#184 F3)', () => {
     );
   });
 
-  // #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 () => {
+  // #490/#520 reactive branch: a provider CONTEXT-OVERFLOW 400 in onError is
+  // classified, records a distinguishable cause, and stamps the consecutive-overflow
+  // COUNTER (metadata.replayOverflowCount) so the NEXT turn's budgeter trims with
+  // escalating aggression (the recovery that un-bricks the chat). This is a fresh
+  // chat (empty history -> prior streak 0), so the first overflow stamps count 1.
+  it('#490/#520: a context-overflow 400 stamps replayOverflowCount=1 on the finalized row', async () => {
     jest
       .spyOn(Logger.prototype, 'error')
       .mockImplementation(() => undefined as never);
@@ -397,11 +399,14 @@ describe('AiChatService.stream — abortSignal wiring (#184 F3)', () => {
       metadata: Record<string, unknown>;
     };
     expect(patch.status).toBe('error');
-    expect(patch.metadata.replayOverflow).toBe(true);
+    // First overflow on a fresh chat -> k = prior(0) + 1 = 1.
+    expect(patch.metadata.replayOverflowCount).toBe(1);
+    // The legacy boolean is no longer written (the counter supersedes it).
+    expect('replayOverflow' in patch.metadata).toBe(false);
     expect(patch.metadata.error).toContain('контекстное окно');
   });
 
-  it('#490: a non-overflow error does NOT stamp replayOverflow', async () => {
+  it('#490/#520: a non-overflow error does NOT stamp the overflow counter', async () => {
     jest
       .spyOn(Logger.prototype, 'error')
       .mockImplementation(() => undefined as never);
@@ -412,6 +417,7 @@ describe('AiChatService.stream — abortSignal wiring (#184 F3)', () => {
       status: string;
       metadata: Record<string, unknown>;
     };
+    expect('replayOverflowCount' in patch.metadata).toBe(false);
     expect('replayOverflow' in patch.metadata).toBe(false);
   });
 });

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

@@ -30,13 +30,16 @@ import {
   STEP_LIMIT_NO_ANSWER_MARKER,
   OUTPUT_DEGENERATION_ERROR,
   lastAssistantContextTokens,
-  lastAssistantReplayOverflow,
+  lastAssistantReplayOverflowCount,
   seedActivatedTools,
 } from './ai-chat.service';
 import type { AiChatMessage, Workspace } from '@docmost/db/types/entity.types';
 import { buildSystemPrompt } from './ai-chat.prompt';
 import type { McpClientsService } from './external-mcp/mcp-clients.service';
-import { resolveEffectiveReplayThreshold } from './history-budget';
+import {
+  resolveEffectiveReplayThreshold,
+  REPLAY_MIN_FLOOR_TOKENS,
+} from './history-budget';
 
 /**
  * Unit tests for compactToolOutput: the pure helper that shrinks tool outputs
@@ -554,49 +557,123 @@ describe('seedActivatedTools', () => {
   });
 });
 
-describe('lastAssistantReplayOverflow', () => {
+describe('lastAssistantReplayOverflowCount', () => {
   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', () => {
+  it('reads the consecutive-overflow count from the LAST assistant turn', () => {
     expect(
-      lastAssistantReplayOverflow([
-        row('assistant', { replayOverflow: true }),
+      lastAssistantReplayOverflowCount([
+        row('assistant', { replayOverflowCount: 3 }),
         row('user', null),
       ]),
-    ).toBe(true);
-    // A recovered (later, non-overflow) assistant turn clears it.
+    ).toBe(3);
+    // A recovered (later, non-overflow) assistant turn resets it to 0 — the read
+    // stops at the most recent assistant row, which carries no count.
     expect(
-      lastAssistantReplayOverflow([
-        row('assistant', { replayOverflow: true }),
+      lastAssistantReplayOverflowCount([
+        row('assistant', { replayOverflowCount: 3 }),
         row('user', null),
         row('assistant', { contextTokens: 5 }),
       ]),
-    ).toBe(false);
-    expect(lastAssistantReplayOverflow([])).toBe(false);
+    ).toBe(0);
+    expect(lastAssistantReplayOverflowCount([])).toBe(0);
   });
 
-  // #490 reactive recovery: a prior turn stamped `replayOverflow` must make the
-  // NEXT turn's effective budget the AGGRESSIVE 0.5x cut — that harder trim is
-  // what un-bricks a chat that just 400'd on the context window. This exercises
-  // the exact wiring the service uses: read the stamp, then scale the threshold.
-  it('#490: a prior replayOverflow drives the next turn to the 0.5x aggressive budget', () => {
+  // BACK-COMPAT (#520): an in-flight row written by the pre-#520 boolean stamp
+  // (`replayOverflow: true`, no count) reads as k=1 — the old single 0.5× behavior —
+  // so a chat mid-recovery across the deploy does not regress.
+  it('#520 back-compat: a legacy boolean replayOverflow reads as k=1', () => {
+    expect(
+      lastAssistantReplayOverflowCount([
+        row('assistant', { replayOverflow: true }),
+        row('user', null),
+      ]),
+    ).toBe(1);
+    // A legacy row with the flag absent/false is k=0.
+    expect(
+      lastAssistantReplayOverflowCount([row('assistant', { contextTokens: 5 })]),
+    ).toBe(0);
+  });
+
+  // A corrupt/negative persisted count never yields a negative k.
+  it('clamps a corrupt negative count to 0', () => {
+    expect(
+      lastAssistantReplayOverflowCount([
+        row('assistant', { replayOverflowCount: -4 }),
+      ]),
+    ).toBe(0);
+  });
+
+  // #490/#520 reactive recovery: the prior consecutive-overflow count `k` drives
+  // the next turn's effective budget to an ESCALATING cut (0.5**k) — each further
+  // consecutive 400 tightens it, which is what un-bricks a chat that keeps
+  // overflowing. This exercises the exact wiring the service uses: read the count,
+  // then scale the threshold.
+  it('#490/#520: the prior count drives the next turn to the escalating aggressive budget', () => {
     const history = [
-      row('assistant', { replayOverflow: true }),
+      row('assistant', { replayOverflowCount: 1 }),
       row('user', null),
     ];
-    const priorOverflowed = lastAssistantReplayOverflow(history);
-    expect(priorOverflowed).toBe(true);
-    // Base budget 100k -> aggressive recovery halves it to 50k this turn.
-    expect(resolveEffectiveReplayThreshold(100_000, priorOverflowed)).toBe(50_000);
+    const k = lastAssistantReplayOverflowCount(history);
+    expect(k).toBe(1);
+    // Base budget 100k -> first-overflow recovery halves it to 50k this turn.
+    expect(resolveEffectiveReplayThreshold(100_000, k)).toBe(50_000);
+    // A second consecutive overflow (k=2) quarters it.
+    expect(resolveEffectiveReplayThreshold(100_000, 2)).toBe(25_000);
     // Odd base floors, not rounds.
-    expect(resolveEffectiveReplayThreshold(99_999, true)).toBe(49_999);
-    // No prior overflow -> the base budget is used verbatim (no aggressive cut).
-    expect(resolveEffectiveReplayThreshold(100_000, false)).toBe(100_000);
+    expect(resolveEffectiveReplayThreshold(99_999, 1)).toBe(49_999);
+    // No prior overflow (k=0) -> the base budget is used verbatim (no cut).
+    expect(resolveEffectiveReplayThreshold(100_000, 0)).toBe(100_000);
     // An explicit off-switch (null) is never overridden, even on recovery.
-    expect(resolveEffectiveReplayThreshold(null, true)).toBeNull();
+    expect(resolveEffectiveReplayThreshold(null, 3)).toBeNull();
+  });
+
+  // #520 escalation table + convergence: the cut deepens each consecutive overflow
+  // and is CLAMPED at the floor so it converges (un-bricks even against a small
+  // real window), instead of the old fixed single 0.5× that stuck at 50k forever.
+  it('#520: escalates and converges to the floor, un-bricking a small real window', () => {
+    const base = 100_000;
+    expect(resolveEffectiveReplayThreshold(base, 0)).toBe(base);
+    expect(resolveEffectiveReplayThreshold(base, 1)).toBe(50_000);
+    expect(resolveEffectiveReplayThreshold(base, 2)).toBe(25_000);
+    expect(resolveEffectiveReplayThreshold(base, 3)).toBe(12_500);
+
+    // Residual-brick regression (#520): with the flat-default base (100k) and a real
+    // model window of ~40k, the OLD fixed 0.5× stuck at 50k forever (> 40k -> 400s
+    // again, never recovers). The escalating cut drops BELOW 40k after enough
+    // consecutive overflows -> the history finally fits -> the chat un-bricks.
+    const realWindow = 40_000;
+    // k=1 (50k) still exceeds the window — the old behavior's terminal state.
+    expect(resolveEffectiveReplayThreshold(base, 1)).toBeGreaterThan(realWindow);
+    // But escalation converges under the window within a couple more turns.
+    const converged = [2, 3, 4, 5].some(
+      (k) => (resolveEffectiveReplayThreshold(base, k) as number) < realWindow,
+    );
+    expect(converged).toBe(true);
+
+    // Convergence is bounded BELOW by the floor: a large k never trims below it.
+    for (const k of [4, 8, 20, 100]) {
+      expect(resolveEffectiveReplayThreshold(base, k)).toBe(REPLAY_MIN_FLOOR_TOKENS);
+      expect(
+        resolveEffectiveReplayThreshold(base, k) as number,
+      ).toBeGreaterThanOrEqual(REPLAY_MIN_FLOOR_TOKENS);
+    }
+  });
+
+  // The floor never RAISES a legitimately small configured budget above itself —
+  // that would re-overflow the very window it was configured for.
+  it('#520: never inflates a small configured budget above itself', () => {
+    const small = 5_000; // below the floor
+    expect(resolveEffectiveReplayThreshold(small, 0)).toBe(small);
+    // Even under escalation the effective threshold never exceeds the base.
+    for (const k of [1, 2, 3, 10]) {
+      expect(
+        resolveEffectiveReplayThreshold(small, k) as number,
+      ).toBeLessThanOrEqual(small);
+    }
   });
 });
 
@@ -930,21 +1007,29 @@ 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', () => {
+  // #490/#520 observability: the replay budgeter's decision is stamped on the turn,
+  // now including the consecutive-overflow COUNTER (#520) the next turn escalates on.
+  it('records replayTrimmedToTokens + replayOverflowCount when provided', () => {
     const f = flushAssistant([], '', 'error', {
       error: 'ctx',
       replayTrimmedToTokens: 42_000,
-      replayOverflow: true,
+      replayOverflowCount: 2,
     });
     expect(f.metadata.replayTrimmedToTokens).toBe(42_000);
-    expect(f.metadata.replayOverflow).toBe(true);
+    expect(f.metadata.replayOverflowCount).toBe(2);
   });
 
   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);
+    expect('replayOverflowCount' in f.metadata).toBe(false);
+  });
+
+  // A clean finalize (no overflow -> count 0/omitted) leaves NO counter, which the
+  // next turn reads as k=0 — the reset that ends a recovery streak.
+  it('omits replayOverflowCount for a zero/absent count (reset semantics)', () => {
+    const zero = flushAssistant([], '', 'completed', { replayOverflowCount: 0 });
+    expect('replayOverflowCount' in zero.metadata).toBe(false);
   });
 
   // #274 observability: the page-change diff the agent saw this turn is persisted

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

@@ -140,9 +140,10 @@ const OUTPUT_DEGENERATION_ERROR =
 
 // 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).
+// row is ALSO stamped `metadata.replayOverflowCount` (the consecutive-overflow
+// counter, #520) so the NEXT turn's budgeter trims with escalating aggression (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.';
@@ -1192,19 +1193,23 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
       }
       // 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);
+      // Reactive recovery (#490/#520): `k` = how many CONSECUTIVE preceding turns
+      // were rejected for context overflow (stamped by onError). Each consecutive
+      // overflow trims MORE aggressively (resolveEffectiveReplayThreshold scales the
+      // budget by 0.5**k, clamped at the floor) so recovery ESCALATES until the
+      // history fits — the overflowing turn produced no usage signal, so a single
+      // fixed cut may not shrink enough when the real model window is small. This is
+      // what un-bricks a chat that keeps 400'ing on the context window.
+      const priorOverflowCount = lastAssistantReplayOverflowCount(oldHistory);
       const effectiveThreshold = resolveEffectiveReplayThreshold(
         replayBudget.thresholdTokens,
-        priorOverflowed,
+        priorOverflowCount,
       );
-      if (priorOverflowed) {
+      if (priorOverflowCount > 0) {
         this.logger.warn(
-          `AI chat (chat ${chatId}): previous turn hit context overflow; ` +
-            `applying aggressive replay budget (${effectiveThreshold} tokens).`,
+          `AI chat (chat ${chatId}): ${priorOverflowCount} consecutive context ` +
+            `overflow(s); applying escalated aggressive replay budget ` +
+            `(${effectiveThreshold} tokens).`,
         );
       }
       const preTrim = trimHistoryForReplay(
@@ -1212,7 +1217,7 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
         effectiveThreshold,
         // A prior OVERFLOW means the provider count is stale/absent — force the
         // char-estimate path by ignoring priorContextTokens on recovery.
-        priorOverflowed ? undefined : priorContextTokens,
+        priorOverflowCount > 0 ? undefined : priorContextTokens,
       );
       messages = preTrim.messages;
       // Observability (#490): record the budgeter's decision on the turn so the UI
@@ -1904,7 +1909,12 @@ export class AiChatService implements OnModuleInit, OnModuleDestroy {
                 pageChanged,
                 partsCache,
                 replayTrimmedToTokens,
-                replayOverflow: overflow || undefined,
+                // #520: escalate the consecutive-overflow counter so the NEXT turn
+                // trims MORE aggressively (0.5**k). k grows by 1 each consecutive
+                // overflow; a clean finalize omits the field, resetting it to 0.
+                replayOverflowCount: overflow
+                  ? priorOverflowCount + 1
+                  : undefined,
               }),
             );
             // #184: settle the RUN as failed, carrying the provider/transport cause.
@@ -2336,22 +2346,39 @@ export function seedActivatedTools(
 }
 
 /**
- * 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.
+ * How many CONSECUTIVE recent turns were rejected for CONTEXT OVERFLOW (#490/#520):
+ * `k`, read from the most recent assistant row's `metadata.replayOverflowCount`
+ * (stamped by the stream's onError, incremented each consecutive overflow and reset
+ * to 0 on any clean finalize). The next turn's budgeter feeds this to
+ * {@link resolveEffectiveReplayThreshold} to trim with ESCALATING aggression — the
+ * reactive recovery. Only the LAST assistant turn matters (its count already carries
+ * the consecutive streak; an older overflow followed by a clean turn was recovered),
+ * so we stop at the first assistant row scanning backwards.
+ *
+ * BACK-COMPAT: a row written by the pre-#520 boolean stamp (`replayOverflow: true`,
+ * no count) is read as k=1 — the old single 0.5× behavior — so in-flight chats do
+ * not regress across the deploy.
  */
-export function lastAssistantReplayOverflow(
+export function lastAssistantReplayOverflowCount(
   history: ReadonlyArray<AiChatMessage>,
-): boolean {
+): number {
   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;
+    const meta = (row.metadata ?? {}) as {
+      replayOverflowCount?: unknown;
+      replayOverflow?: unknown;
+    };
+    if (typeof meta.replayOverflowCount === 'number') {
+      // Guard against a corrupt/negative persisted value.
+      return meta.replayOverflowCount > 0
+        ? Math.floor(meta.replayOverflowCount)
+        : 0;
+    }
+    // Back-compat: legacy boolean stamp -> one overflow (0.5× cut).
+    return meta.replayOverflow === true ? 1 : 0;
   }
-  return false;
+  return 0;
 }
 
 /** The last message with role 'user' from a useChat payload, if any. */
@@ -2891,9 +2918,11 @@ export function flushAssistant(
     // 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;
+    // #490/#520 reactive branch: the consecutive context-overflow count for THIS
+    // turn (prior streak + 1) when the provider rejected it for context overflow.
+    // Stamped into metadata so the NEXT turn's budgeter trims with escalating
+    // aggression (0.5**k). Omitted (undefined) on a clean turn, which resets k to 0.
+    replayOverflowCount?: number;
   },
 ): AssistantFlush {
   const finished = capturedSteps ?? [];
@@ -2946,7 +2975,8 @@ export function flushAssistant(
     metadata.maxContextTokens = extra.maxContextTokens;
   if (extra?.replayTrimmedToTokens)
     metadata.replayTrimmedToTokens = extra.replayTrimmedToTokens;
-  if (extra?.replayOverflow) metadata.replayOverflow = true;
+  if (extra?.replayOverflowCount && extra.replayOverflowCount > 0)
+    metadata.replayOverflowCount = extra.replayOverflowCount;
   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.int-spec.ts

@@ -1,6 +1,11 @@
 import { randomBytes } from 'crypto';
 import { Client } from 'pg';
-import { flushAssistant, serializeSteps } from './ai-chat.service';
+import {
+  flushAssistant,
+  serializeSteps,
+  lastAssistantReplayOverflowCount,
+} from './ai-chat.service';
+import type { AiChatMessage } from '@docmost/db/types/entity.types';
 
 /**
  * #490 write-volume regression — an OBSERVABLE-PROPERTY test on a LIVE Postgres,
@@ -207,3 +212,111 @@ describe('#490 write-volume on a live Postgres (pg_current_wal_lsn delta)', () =
     expect(v2).toBeLessThan(v1 * 0.75);
   }, 120_000);
 });
+
+/**
+ * #520 reactive-recovery COUNTER lifecycle on a LIVE Postgres — proves the
+ * consecutive-overflow count survives a real jsonb metadata round-trip (the persist
+ * path), not just an in-memory object. flushAssistant BUILDS the row metadata, we
+ * WRITE it to a jsonb column, READ it back, then reconstruct the assistant row and
+ * run lastAssistantReplayOverflowCount over it — exactly the read the next turn does.
+ *
+ * The lifecycle proven end-to-end through pg:
+ *   - consecutive overflows INCREMENT k (1 -> 2 -> 3);
+ *   - a CLEAN finalize omits the field, which the reader treats as a RESET to 0;
+ *   - a legacy boolean row (`replayOverflow: true`) reads back as k=1 (back-compat).
+ */
+describe('#520 overflow-counter lifecycle on a live Postgres (jsonb round-trip)', () => {
+  let client: Client | undefined;
+  let available = false;
+
+  beforeAll(async () => {
+    try {
+      client = new Client(CONN);
+      await client.connect();
+      await client.query('SELECT 1');
+      available = true;
+    } catch {
+      available = false;
+      client = undefined;
+    }
+  });
+
+  afterAll(async () => {
+    await client?.end().catch(() => undefined);
+  });
+
+  // Round-trip an arbitrary metadata object through a real jsonb column and read it
+  // back as the reconstructed assistant row the next turn would load.
+  async function roundTrip(
+    c: Client,
+    metadata: unknown,
+  ): Promise<AiChatMessage> {
+    await c.query('UPDATE _wal_counter SET metadata=$1 WHERE id=1', [
+      JSON.stringify(metadata),
+    ]);
+    const back = (await c.query('SELECT metadata FROM _wal_counter WHERE id=1'))
+      .rows[0].metadata as Record<string, unknown>;
+    return { role: 'assistant', metadata: back } as unknown as AiChatMessage;
+  }
+
+  it('increments across consecutive overflows, resets on a clean turn, and honors the legacy boolean', async () => {
+    if (!available || !client) {
+      console.warn('SKIP: gitmost-test-pg not reachable; skipping counter test.');
+      return;
+    }
+    const c = client;
+    await c.query('DROP TABLE IF EXISTS _wal_counter');
+    await c.query('CREATE TABLE _wal_counter(id int primary key, metadata jsonb)');
+    await c.query("INSERT INTO _wal_counter VALUES (1, '{}'::jsonb)");
+
+    // Turn 1 overflow: prior streak 0 -> stamp k=1 (as the service does: prior+1).
+    let prior = lastAssistantReplayOverflowCount([]); // fresh chat
+    expect(prior).toBe(0);
+    let row = await roundTrip(
+      c,
+      flushAssistant([], '', 'error', {
+        error: 'ctx',
+        replayOverflowCount: prior + 1,
+      }).metadata,
+    );
+    prior = lastAssistantReplayOverflowCount([row]);
+    expect(prior).toBe(1);
+
+    // Turn 2 overflow: prior 1 -> stamp k=2.
+    row = await roundTrip(
+      c,
+      flushAssistant([], '', 'error', {
+        error: 'ctx',
+        replayOverflowCount: prior + 1,
+      }).metadata,
+    );
+    prior = lastAssistantReplayOverflowCount([row]);
+    expect(prior).toBe(2);
+
+    // Turn 3 overflow: prior 2 -> stamp k=3.
+    row = await roundTrip(
+      c,
+      flushAssistant([], '', 'error', {
+        error: 'ctx',
+        replayOverflowCount: prior + 1,
+      }).metadata,
+    );
+    prior = lastAssistantReplayOverflowCount([row]);
+    expect(prior).toBe(3);
+
+    // Turn 4 CLEAN finalize: no overflow -> the field is omitted -> reset to 0.
+    row = await roundTrip(
+      c,
+      flushAssistant([], 'all good', 'completed', { finishReason: 'stop' })
+        .metadata,
+    );
+    expect('replayOverflowCount' in (row.metadata as object)).toBe(false);
+    expect(lastAssistantReplayOverflowCount([row])).toBe(0);
+
+    // Back-compat: a row persisted by the pre-#520 boolean stamp reads back as k=1.
+    row = await roundTrip(c, { replayOverflow: true });
+    expect(lastAssistantReplayOverflowCount([row])).toBe(1);
+
+    await c.query('DROP TABLE IF EXISTS _wal_counter');
+  }, 60_000);
+});

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

@@ -1,14 +1,79 @@
 import type { ModelMessage } from 'ai';
 import {
   resolveReplayBudget,
+  resolveEffectiveReplayThreshold,
   isContextOverflowError,
   estimateMessagesTokens,
   trimHistoryForReplay,
   REPLAY_BUDGET_DEFAULT_TOKENS,
+  REPLAY_MIN_FLOOR_TOKENS,
   REPLAY_TRUNCATION_MARKER,
   REPLAY_TURN_COLLAPSED_MARKER,
 } from './history-budget';
 
+describe('resolveEffectiveReplayThreshold (#520 iterative escalation)', () => {
+  // The escalation table: each consecutive overflow (k) deepens the cut by 0.5×.
+  it('scales the base by 0.5**k, flooring (not rounding) fractional tokens', () => {
+    const base = 100_000;
+    expect(resolveEffectiveReplayThreshold(base, 0)).toBe(base); // k=0: unchanged
+    expect(resolveEffectiveReplayThreshold(base, 1)).toBe(50_000); // 0.5×
+    expect(resolveEffectiveReplayThreshold(base, 2)).toBe(25_000); // 0.25×
+    expect(resolveEffectiveReplayThreshold(base, 3)).toBe(12_500); // 0.125×
+    // Floors, not rounds.
+    expect(resolveEffectiveReplayThreshold(99_999, 1)).toBe(49_999);
+  });
+
+  it('passes a null base (trimming OFF) through unchanged for any k', () => {
+    for (const k of [0, 1, 2, 5, 100]) {
+      expect(resolveEffectiveReplayThreshold(null, k)).toBeNull();
+    }
+  });
+
+  // The crux of #520: convergence. A large k is clamped at REPLAY_MIN_FLOOR_TOKENS,
+  // so the escalation CONVERGES to a small-but-usable budget instead of trimming to
+  // zero — and, unlike the old fixed 0.5× that stuck at 50k, it drops far enough to
+  // fit a small real model window.
+  it('clamps a large k at the floor (converges, never below)', () => {
+    const base = 100_000;
+    for (const k of [4, 6, 10, 50, 200]) {
+      const t = resolveEffectiveReplayThreshold(base, k) as number;
+      expect(t).toBe(REPLAY_MIN_FLOOR_TOKENS);
+      expect(t).toBeGreaterThanOrEqual(REPLAY_MIN_FLOOR_TOKENS);
+    }
+  });
+
+  // Residual-brick regression (#520): flat-default base 100k, real window ~40k. The
+  // OLD fixed single 0.5× stuck at 50k > 40k forever (re-overflows every turn — the
+  // brick). The iterative cut drops BELOW 40k after a couple more consecutive
+  // overflows, so the history finally fits and the chat un-bricks.
+  it('un-bricks: escalation drops below a small real window the fixed 0.5× never could', () => {
+    const base = 100_000;
+    const realWindow = 40_000;
+    // The old terminal state: 0.5× = 50k, still above the window.
+    expect(resolveEffectiveReplayThreshold(base, 1)).toBeGreaterThan(realWindow);
+    // Escalation converges under the window.
+    const converged = [2, 3, 4, 5].some(
+      (k) => (resolveEffectiveReplayThreshold(base, k) as number) < realWindow,
+    );
+    expect(converged).toBe(true);
+    // MUTATION SENTINEL: reverting `** k` to `** 1` (fixed 0.5×) makes every k yield
+    // 50k, so `converged` above would be FALSE and this test reddens. Removing the
+    // floor reddens the clamp test instead.
+  });
+
+  // The floor never RAISES a legitimately small configured budget above itself
+  // (min(floor, base)); doing so would re-overflow the very small window it was set
+  // for. So a base BELOW the floor is passed through unchanged and never inflated.
+  it('never inflates a small configured budget above itself', () => {
+    const small = 5_000; // below REPLAY_MIN_FLOOR_TOKENS
+    expect(resolveEffectiveReplayThreshold(small, 0)).toBe(small);
+    for (const k of [1, 2, 3, 10]) {
+      const t = resolveEffectiveReplayThreshold(small, k) as number;
+      expect(t).toBeLessThanOrEqual(small);
+    }
+  });
+});
+
 describe('resolveReplayBudget', () => {
   it('uses floor(0.7 x window) for a configured window (no cap)', () => {
     // 0.7 x 60k = 42k

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

@@ -22,10 +22,25 @@ 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.
+ * Per-step fraction of the normal budget applied on the REACTIVE re-trim after a
+ * provider context-overflow 400 — the preventive estimate under-counted, so cut
+ * harder. This is now applied ITERATIVELY: with `k` consecutive overflow turns the
+ * budget is scaled by `fraction ** k` (k=1 -> 0.5×, k=2 -> 0.25×, …), so recovery
+ * ESCALATES turn over turn until the replayed history finally fits, instead of the
+ * old single fixed 0.5× cut that could never un-brick a chat whose real model
+ * window is smaller than 0.5 × the (unconfigured, flat-default) base budget (#520).
  */
 export const REPLAY_AGGRESSIVE_FRACTION = 0.5;
+/**
+ * Lower bound (tokens) on the escalating reactive budget: the iterative cut is
+ * clamped here so it CONVERGES (a fixed floor, not an ever-shrinking value that
+ * would eventually trim everything). Rationale: below ~8k tokens a chat cannot
+ * carry meaningful recent context, and even a small real model window comfortably
+ * fits this much — keep-recent-turns still applies on top, so a handful of recent
+ * turns survive. It is never applied so as to RAISE a legitimately small configured
+ * budget (that would re-overflow a tiny window); see resolveEffectiveReplayThreshold.
+ */
+export const REPLAY_MIN_FLOOR_TOKENS = 8_000;
 /**
  * Turns (a user message + its assistant/tool replies) kept FULL at the tail,
  * including the current one — never trimmed. Older turns are compacted first.
@@ -85,22 +100,34 @@ export function resolveReplayBudget(rawContextWindow: unknown): ReplayBudget {
 }
 
 /**
- * The effective replay threshold for THIS turn, given the base budget and whether
- * the PREVIOUS turn hit a context-overflow 400 (the reactive-recovery signal,
- * `metadata.replayOverflow`). On recovery the base budget is scaled down by
- * {@link REPLAY_AGGRESSIVE_FRACTION}: the overflowing turn produced no usage
- * signal, so the preventive estimate under-counted and a normal-threshold trim may
- * not shrink enough to fit — this harder cut is what un-bricks the chat.
+ * The effective replay threshold for THIS turn, given the base budget and `k` — the
+ * number of CONSECUTIVE preceding turns that hit a context-overflow 400 (the
+ * reactive-recovery signal `metadata.replayOverflowCount`, read from the last
+ * assistant row). On recovery the base budget is scaled down ITERATIVELY by
+ * {@link REPLAY_AGGRESSIVE_FRACTION} ** k and clamped at {@link REPLAY_MIN_FLOOR_TOKENS}:
+ *   - k=0 -> base unchanged (no overflow: nothing to recover from).
+ *   - k=1 -> floor(0.5 × base); k=2 -> floor(0.25 × base); … each further consecutive
+ *     overflow tightens the cut, so recovery ESCALATES until the history fits.
+ *   - the escalation is clamped at the floor so it CONVERGES — this is what un-bricks
+ *     a chat whose real model window is smaller than a single 0.5× cut of the base
+ *     (e.g. an unconfigured window: flat-default base 100k, real window <50k) (#520).
+ *
+ * The overflowing turn produced no usage signal, so the preventive estimate
+ * under-counted and a normal-threshold (or single fixed 0.5×) trim may not shrink
+ * enough to fit; the escalating cut is what recovers such a chat.
  *
  * A `null` base budget (trimming OFF) is passed through unchanged: an explicit
- * off-switch is never overridden by the recovery path.
+ * off-switch is never overridden by the recovery path. The floor is applied as
+ * `min(floor, base)` so it never RAISES a legitimately small configured budget
+ * above itself (which would re-overflow the same small window it was set for).
  */
 export function resolveEffectiveReplayThreshold(
   thresholdTokens: number | null,
-  priorOverflowed: boolean,
+  k: number,
 ): number | null {
-  if (!priorOverflowed || thresholdTokens == null) return thresholdTokens;
-  return Math.floor(thresholdTokens * REPLAY_AGGRESSIVE_FRACTION);
+  if (thresholdTokens == null || k <= 0) return thresholdTokens;
+  const scaled = Math.floor(thresholdTokens * REPLAY_AGGRESSIVE_FRACTION ** k);
+  return Math.max(scaled, Math.min(REPLAY_MIN_FLOOR_TOKENS, thresholdTokens));
 }
 
 /**