diff --git a/apps/server/src/core/ai-chat/ai-chat.service.ts b/apps/server/src/core/ai-chat/ai-chat.service.ts
index 908c9200..dfa1bb75 100644
--- a/apps/server/src/core/ai-chat/ai-chat.service.ts
+++ b/apps/server/src/core/ai-chat/ai-chat.service.ts
@@ -56,6 +56,7 @@ import {
 import {
   isDegenerateOutput,
   truncateDegeneratedTail,
+  shouldCheckDegeneration,
 } from './output-degeneration';
 
 // Max agent steps per turn. One step = one model generation; a step that calls
@@ -1101,7 +1102,6 @@ export class AiChatService implements OnModuleInit {
       const degenerationController = new AbortController();
       let degenerationDetected = false;
       let lastDegenerationCheckLen = 0;
-      const DEGENERATION_CHECK_STEP = 2000;
 
       // Step-granular durability (#183): create the assistant row UPFRONT in the
       // 'streaming' state (before any token), then UPDATE it as each step finishes
@@ -1275,8 +1275,10 @@ export class AiChatService implements OnModuleInit {
               // trigger, abort the run ONCE with a distinguishable reason.
               if (
                 !degenerationDetected &&
-                inProgressText.length - lastDegenerationCheckLen >=
-                  DEGENERATION_CHECK_STEP
+                shouldCheckDegeneration(
+                  inProgressText.length,
+                  lastDegenerationCheckLen,
+                )
               ) {
                 lastDegenerationCheckLen = inProgressText.length;
                 if (isDegenerateOutput(inProgressText)) {
@@ -1296,6 +1298,13 @@ export class AiChatService implements OnModuleInit {
             // the in-progress accumulator for the next step.
             capturedSteps.push(step as StepLike);
             inProgressText = '';
+            // Reset the degeneration-check watermark too (#486): it tracks a byte
+            // offset INTO inProgressText, so once that resets to '' a stale (large)
+            // mark makes `inProgressText.length - lastDegenerationCheckLen` go
+            // negative and the throttled detector stays silent until a later step's
+            // text re-grows past the old offset — a whole degenerate step could slip
+            // through undetected. Zeroing it re-arms the check from the next byte.
+            lastDegenerationCheckLen = 0;
             // Step-granular durability (#183): persist this finished step (its text +
             // tool calls + tool RESULTS) the moment it ends, so a process death after
             // this point still recovers the step. Not awaited here (never block the
diff --git a/apps/server/src/core/ai-chat/output-degeneration.spec.ts b/apps/server/src/core/ai-chat/output-degeneration.spec.ts
index a4f6c36e..426f4e40 100644
--- a/apps/server/src/core/ai-chat/output-degeneration.spec.ts
+++ b/apps/server/src/core/ai-chat/output-degeneration.spec.ts
@@ -3,6 +3,8 @@ import {
   hasPeriodicTail,
   isDegenerateOutput,
   truncateDegeneratedTail,
+  shouldCheckDegeneration,
+  DEGENERATION_CHECK_STEP,
   REPEATED_LINES_THRESHOLD,
   MIN_PERIOD_REPEATS,
 } from './output-degeneration';
@@ -180,3 +182,42 @@ describe('truncateDegeneratedTail', () => {
     expect(truncateDegeneratedTail(text)).toBe(text);
   });
 });
+
+/**
+ * Throttle + step-boundary reset (#486). The stream keeps a watermark
+ * (`lastDegenerationCheckLen`) that is an OFFSET into the accumulated step text.
+ * On a step boundary the accumulator resets to '', so the watermark MUST reset to
+ * 0 too — otherwise the throttle goes silent for the whole next step. These tests
+ * pin the pure decision AND the reset property that ai-chat.service.onStepFinish
+ * now enforces.
+ */
+describe('shouldCheckDegeneration (throttle) + step-boundary reset (#486)', () => {
+  it('fires once the text grows a full DEGENERATION_CHECK_STEP past the mark', () => {
+    expect(shouldCheckDegeneration(DEGENERATION_CHECK_STEP, 0)).toBe(true);
+    expect(shouldCheckDegeneration(DEGENERATION_CHECK_STEP - 1, 0)).toBe(false);
+    expect(shouldCheckDegeneration(5000, 3000)).toBe(true); // grew 2000 since mark
+    expect(shouldCheckDegeneration(4000, 3000)).toBe(false); // grew only 1000
+  });
+
+  it('BUG (no reset): a stale large watermark silences the next step', () => {
+    // End of a long step: the watermark sits at 5000. The step ends and the
+    // accumulator resets to '' — but if the watermark is NOT reset, a fresh short
+    // degenerate burst (length 2000) never triggers a check: 2000 - 5000 < STEP.
+    const staleWatermark = 5000;
+    const nextStepLen = DEGENERATION_CHECK_STEP; // a fresh 2KB burst
+    expect(shouldCheckDegeneration(nextStepLen, staleWatermark)).toBe(false);
+  });
+
+  it('FIX (reset to 0): the same short degenerate burst IS checked and detected', () => {
+    // onStepFinish now zeroes the watermark, so the fresh burst re-arms the check.
+    const resetWatermark = 0;
+    const degenerateBurst = 'loadTools.\n'.repeat(300); // real degeneration
+    expect(degenerateBurst.length).toBeGreaterThanOrEqual(DEGENERATION_CHECK_STEP);
+    // The throttle now fires...
+    expect(
+      shouldCheckDegeneration(degenerateBurst.length, resetWatermark),
+    ).toBe(true);
+    // ...and the detector catches the loop that would otherwise stream unchecked.
+    expect(isDegenerateOutput(degenerateBurst)).toBe(true);
+  });
+});
diff --git a/apps/server/src/core/ai-chat/output-degeneration.ts b/apps/server/src/core/ai-chat/output-degeneration.ts
index 2834ce7e..eff4c080 100644
--- a/apps/server/src/core/ai-chat/output-degeneration.ts
+++ b/apps/server/src/core/ai-chat/output-degeneration.ts
@@ -131,6 +131,32 @@ export function isDegenerateOutput(text: string): boolean {
   return hasRepeatedLineRun(text) || hasPeriodicTail(text);
 }
 
+/**
+ * How many bytes the in-progress text must grow before the (amortized) tail
+ * heuristics are re-run. Shared with ai-chat.service so the throttle the stream
+ * applies is the SAME one the unit test drives.
+ */
+export const DEGENERATION_CHECK_STEP = 2000;
+
+/**
+ * Throttle decision for the degeneration guard (#444/#486). Returns true when
+ * the accumulated text has grown at least DEGENERATION_CHECK_STEP bytes past the
+ * last-checked offset, so the pure rules only fire every ~2KB. Pure; the caller
+ * updates its watermark to `textLen` when this returns true.
+ *
+ * The watermark is an offset INTO the accumulator, so when the accumulator is
+ * reset to '' on a step boundary the caller MUST reset the watermark to 0 too
+ * (#486). Otherwise `textLen - lastCheckLen` goes negative after the reset and
+ * this returns false until a later step re-grows past the stale offset — a whole
+ * degenerate step could stream unchecked.
+ */
+export function shouldCheckDegeneration(
+  textLen: number,
+  lastCheckLen: number,
+): boolean {
+  return textLen - lastCheckLen >= DEGENERATION_CHECK_STEP;
+}
+
 /**
  * Truncate a degenerated tail before persist so hundreds of KB of garbage never
  * reach the DB / replay (#444). Keeps everything up to and including the FIRST