diff --git a/apps/pii/server.py b/apps/pii/server.py
index fdd7c1b2604..05b23e9b0b1 100644
--- a/apps/pii/server.py
+++ b/apps/pii/server.py
@@ -176,23 +176,50 @@ def build_analyzer() -> AnalyzerEngine:
 batch_analyzer = BatchAnalyzerEngine(analyzer_engine=analyzer)
 anonymizer = AnonymizerEngine()
 
-# Every entity the spaCy NER recognizers can produce. A request touching any of
-# these must run spaCy; a request naming only non-NER (regex/checksum) entities can
-# skip it. Derived from the live registry so it stays authoritative if Presidio's
-# default entity set changes (e.g. ORGANIZATION), unioned with a known floor so an
-# unexpectedly empty derivation can never let an NER request skip the NLP pass.
+# Every entity the spaCy NER recognizers can actually produce. A request touching
+# any of these must run spaCy; a request naming only non-NER (regex/checksum)
+# entities can skip it. The registry's SpacyRecognizers CLAIM every entity in
+# Presidio's default NER-model mapping — including PHONE_NUMBER/AGE/ID/EMAIL,
+# which exist for transformer models and which no spaCy model can emit — so the
+# claimed set is intersected with the entities the loaded models' NER labels map
+# onto. Without that filter, selecting PHONE_NUMBER (present in nearly every
+# redaction rule) silently forces the full spaCy pass and the regex-only fast
+# path never fires. The floor guarantees the core NER entities always take the
+# full pass even if label introspection ever derives empty.
 _SPACY_NER_FLOOR = frozenset({"PERSON", "LOCATION", "NRP", "DATE_TIME", "ORGANIZATION"})
-NER_ENTITIES = _SPACY_NER_FLOOR | frozenset(
-    entity
-    for recognizer in analyzer.registry.recognizers
-    if isinstance(recognizer, SpacyRecognizer)
-    for entity in recognizer.supported_entities
+
+
+def _producible_ner_entities() -> frozenset:
+    """Presidio entities some loaded spaCy model's NER labels map onto."""
+    configuration = getattr(analyzer.nlp_engine, "ner_model_configuration", None)
+    mapping = configuration.model_to_presidio_entity_mapping if configuration else {}
+    producible = set()
+    for nlp in getattr(analyzer.nlp_engine, "nlp", {}).values():
+        if "ner" not in nlp.pipe_names:
+            continue
+        for label in nlp.get_pipe("ner").labels:
+            entity = mapping.get(label)
+            if entity:
+                producible.add(entity)
+    return frozenset(producible)
+
+
+NER_ENTITIES = _SPACY_NER_FLOOR | (
+    frozenset(
+        entity
+        for recognizer in analyzer.registry.recognizers
+        if isinstance(recognizer, SpacyRecognizer)
+        for entity in recognizer.supported_entities
+    )
+    & _producible_ner_entities()
 )
 
 # One blank NlpArtifacts per language, built once at startup. Passing these to
 # analyze() skips nlp_engine.process_text (the spaCy tok2vec+ner pass) entirely:
-# the pattern recognizers still match on the raw text, SpacyRecognizer is excluded
-# by the entity filter, and score_threshold is unset so detection is identical.
+# the pattern recognizers still match on the raw text, SpacyRecognizer finds
+# nothing in blank artifacts (and can only be consulted at all for claimed-but-
+# unproducible entities like PHONE_NUMBER), and score_threshold is unset so
+# detection is identical.
 # Only context-based score boosting (which needs real tokens) is unavailable — an
 # accepted trade for skipping NER on the hot block-output path. Read-only, so it
 # is safe to share across requests and workers.
diff --git a/apps/sim/lib/guardrails/mask-client.test.ts b/apps/sim/lib/guardrails/mask-client.test.ts
index 3fcb1894353..f8ee031545c 100644
--- a/apps/sim/lib/guardrails/mask-client.test.ts
+++ b/apps/sim/lib/guardrails/mask-client.test.ts
@@ -3,13 +3,15 @@
  */
 import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
 
-const { mockToken, mockBaseUrl } = vi.hoisted(() => ({
+const { mockToken, mockBaseUrl, mockSleep } = vi.hoisted(() => ({
   mockToken: vi.fn(),
   mockBaseUrl: vi.fn(),
+  mockSleep: vi.fn(),
 }))
 
 vi.mock('@/lib/auth/internal', () => ({ generateInternalToken: mockToken }))
 vi.mock('@/lib/core/utils/urls', () => ({ getInternalApiBaseUrl: mockBaseUrl }))
+vi.mock('@sim/utils/helpers', () => ({ sleep: mockSleep }))
 
 import { maskPIIBatchViaHttp } from '@/lib/guardrails/mask-client'
 
@@ -20,6 +22,7 @@ describe('maskPIIBatchViaHttp', () => {
     vi.clearAllMocks()
     mockToken.mockResolvedValue('tok')
     mockBaseUrl.mockReturnValue('http://app.internal:3000')
+    mockSleep.mockResolvedValue(undefined)
     fetchMock = vi.fn(async (_url: string, init: { body: string }) => {
       const { texts } = JSON.parse(init.body) as { texts: string[] }
       return new Response(JSON.stringify({ masked: texts.map((t) => `M(${t})`) }), {
@@ -52,10 +55,80 @@ describe('maskPIIBatchViaHttp', () => {
     expect(fetchMock).toHaveBeenCalledTimes(3) // 2000-per-request cap
   })
 
-  it('throws on a non-2xx response so the caller can scrub', async () => {
-    fetchMock.mockResolvedValueOnce(new Response('boom', { status: 500 }))
+  it('throws immediately on a deterministic 4xx without retrying', async () => {
+    fetchMock.mockResolvedValueOnce(new Response('bad request', { status: 400 }))
 
     await expect(maskPIIBatchViaHttp(['a'], [])).rejects.toThrow(/mask-batch request failed/)
+    expect(fetchMock).toHaveBeenCalledTimes(1)
+    expect(mockSleep).not.toHaveBeenCalled()
+  })
+
+  it('retries a transient 5xx with backoff and then succeeds', async () => {
+    fetchMock.mockResolvedValueOnce(new Response('deploying', { status: 503 }))
+
+    const out = await maskPIIBatchViaHttp(['a'], [])
+
+    expect(out).toEqual(['M(a)'])
+    expect(fetchMock).toHaveBeenCalledTimes(2)
+    expect(mockSleep).toHaveBeenCalledTimes(1)
+  })
+
+  it('retries a rejected fetch (network error) and then succeeds', async () => {
+    fetchMock.mockRejectedValueOnce(new TypeError('fetch failed'))
+
+    const out = await maskPIIBatchViaHttp(['a'], [])
+
+    expect(out).toEqual(['M(a)'])
+    expect(fetchMock).toHaveBeenCalledTimes(2)
+  })
+
+  it('retries a runtime-level request timeout and then succeeds', async () => {
+    const timeout = new Error('The operation timed out.')
+    timeout.name = 'TimeoutError'
+    fetchMock.mockRejectedValueOnce(timeout)
+
+    const out = await maskPIIBatchViaHttp(['a'], [])
+
+    expect(out).toEqual(['M(a)'])
+    expect(fetchMock).toHaveBeenCalledTimes(2)
+  })
+
+  it('gives up after the retry budget is exhausted on a persistent 5xx', async () => {
+    fetchMock.mockImplementation(async () => new Response('down', { status: 503 }))
+
+    await expect(maskPIIBatchViaHttp(['a'], [])).rejects.toThrow(/mask-batch request failed/)
+    expect(fetchMock).toHaveBeenCalledTimes(8)
+    expect(mockSleep).toHaveBeenCalledTimes(7)
+  })
+
+  it('mints a fresh internal token per attempt', async () => {
+    fetchMock.mockResolvedValueOnce(new Response('deploying', { status: 503 }))
+
+    await maskPIIBatchViaHttp(['a'], [])
+
+    expect(mockToken).toHaveBeenCalledTimes(2)
+  })
+
+  it('does not retry a null 200 body (deterministic, not a transient TypeError)', async () => {
+    fetchMock.mockResolvedValueOnce(
+      new Response('null', { status: 200, headers: { 'content-type': 'application/json' } })
+    )
+
+    await expect(maskPIIBatchViaHttp(['a'], [])).rejects.toThrow(/unexpected result/)
+    expect(fetchMock).toHaveBeenCalledTimes(1)
+    expect(mockSleep).not.toHaveBeenCalled()
+  })
+
+  it('does not retry a shape mismatch (deterministic server bug)', async () => {
+    fetchMock.mockResolvedValueOnce(
+      new Response(JSON.stringify({ nope: true }), {
+        status: 200,
+        headers: { 'content-type': 'application/json' },
+      })
+    )
+
+    await expect(maskPIIBatchViaHttp(['a'], [])).rejects.toThrow(/unexpected result/)
+    expect(fetchMock).toHaveBeenCalledTimes(1)
   })
 
   it('returns [] without any request for empty input', async () => {
diff --git a/apps/sim/lib/guardrails/mask-client.ts b/apps/sim/lib/guardrails/mask-client.ts
index 7f5776b8524..7972e0e55cc 100644
--- a/apps/sim/lib/guardrails/mask-client.ts
+++ b/apps/sim/lib/guardrails/mask-client.ts
@@ -1,3 +1,5 @@
+import { sleep } from '@sim/utils/helpers'
+import { backoffWithJitter, parseRetryAfter } from '@sim/utils/retry'
 import type { GuardrailsMaskBatchResult } from '@/lib/api/contracts'
 import { generateInternalToken } from '@/lib/auth/internal'
 import { env } from '@/lib/core/config/env'
@@ -18,6 +20,60 @@ import type { CustomPiiPattern } from '@/lib/guardrails/pii-entities'
  */
 const CHUNK_CONCURRENCY = env.PII_MASK_CHUNK_CONCURRENCY ?? 64
 
+/**
+ * Per-chunk retry budget for transient failures (network errors, 408/429/5xx).
+ * A large payload fans out into many chunk requests, so a single blip — an ALB
+ * 502 during a deploy, a Presidio pod restart — must not fail the whole
+ * redaction (and, on the execution-altering stages, abort the run). With the
+ * default 500ms→30s jittered backoff this rides out ~2 minutes of outage per
+ * chunk before giving up. Deterministic failures (4xx, shape mismatches) throw
+ * immediately.
+ */
+const MAX_CHUNK_ATTEMPTS = 8
+
+const RETRYABLE_STATUSES = new Set([408, 429, 500, 502, 503, 504])
+
+class MaskChunkHttpError extends Error {
+  constructor(
+    message: string,
+    readonly status: number,
+    readonly retryAfterMs: number | null
+  ) {
+    super(message)
+    this.name = 'MaskChunkHttpError'
+  }
+}
+
+function isRetryableChunkError(error: unknown): boolean {
+  if (error instanceof MaskChunkHttpError) {
+    return RETRYABLE_STATUSES.has(error.status)
+  }
+  // A rejected fetch (connection refused/reset, DNS, socket drop) is transient —
+  // Node wraps these in TypeError('fetch failed'). Runtime-level request
+  // timeouts (undici's default 300s headers/body timeout, Bun's TimeoutError)
+  // and mid-flight socket closes surface with their own names/codes per runtime;
+  // all are congestion or connection churn, not a deterministic failure: a chunk
+  // queued behind a saturated Presidio fleet must retry, not fail the payload.
+  if (error instanceof TypeError) {
+    return true
+  }
+  const { name, code } = (error ?? {}) as { name?: unknown; code?: unknown }
+  if (name === 'TimeoutError' || name === 'HeadersTimeoutError' || name === 'BodyTimeoutError') {
+    return true
+  }
+  return (
+    typeof code === 'string' &&
+    [
+      'ECONNRESET',
+      'ECONNREFUSED',
+      'EPIPE',
+      'ETIMEDOUT',
+      'ConnectionClosed',
+      'ConnectionRefused',
+    ].includes(code)
+  )
+}
+
 /**
  * Mask PII across many strings via the internal app-container endpoint.
  *
@@ -29,8 +85,10 @@ const CHUNK_CONCURRENCY = env.PII_MASK_CHUNK_CONCURRENCY ?? 64
  * concurrency, so a large payload fans out rather than serializing; order is
  * preserved, so the returned array matches `texts` length.
  *
- * Rejects on any non-2xx, timeout, or shape mismatch so the caller can apply
- * its own fail-safe (scrubbing rather than leaking).
+ * Transient chunk failures (network errors, 408/429/5xx) retry with jittered
+ * backoff (see {@link MAX_CHUNK_ATTEMPTS}); only a deterministic failure or an
+ * exhausted retry budget rejects, so the caller can apply its own fail-safe
+ * (scrubbing rather than leaking).
  */
 export async function maskPIIBatchViaHttp(
   texts: string[],
@@ -64,8 +122,29 @@ async function postChunk(
   language: string | undefined,
   customPatterns: CustomPiiPattern[] | undefined
 ): Promise<string[]> {
-  // Mint per request: a single token (5min TTL) can expire mid-batch when a
-  // large execution fans out into many sequential chunk requests.
+  for (let attempt = 1; ; attempt++) {
+    try {
+      return await postChunkOnce(url, texts, entityTypes, language, customPatterns)
+    } catch (error) {
+      if (attempt >= MAX_CHUNK_ATTEMPTS || !isRetryableChunkError(error)) {
+        throw error
+      }
+      const retryAfterMs = error instanceof MaskChunkHttpError ? error.retryAfterMs : null
+      await sleep(backoffWithJitter(attempt, retryAfterMs))
+    }
+  }
+}
+
+async function postChunkOnce(
+  url: string,
+  texts: string[],
+  entityTypes: string[],
+  language: string | undefined,
+  customPatterns: CustomPiiPattern[] | undefined
+): Promise<string[]> {
+  // Mint per attempt: a single token (5min TTL) can expire mid-batch when a
+  // large execution fans out into many sequential chunk requests or a chunk
+  // spends its retry budget waiting out an outage.
   const token = await generateInternalToken()
 
   // boundary-raw-fetch: internal server-to-server call to the app container (internal JWT auth, configurable base URL)
@@ -80,11 +159,15 @@ async function postChunk(
 
   if (!response.ok) {
     const detail = await response.text().catch(() => '')
-    throw new Error(`PII mask-batch request failed (${response.status}): ${detail.slice(0, 200)}`)
+    throw new MaskChunkHttpError(
+      `PII mask-batch request failed (${response.status}): ${detail.slice(0, 200)}`,
+      response.status,
+      parseRetryAfter(response.headers.get('retry-after'))
+    )
   }
 
-  const data = (await response.json()) as GuardrailsMaskBatchResult
-  if (!Array.isArray(data.masked)) {
+  const data = (await response.json()) as GuardrailsMaskBatchResult | null
+  if (!data || !Array.isArray(data.masked)) {
     throw new Error('PII mask-batch returned an unexpected result')
   }
   return data.masked
diff --git a/apps/sim/lib/logs/execution/pii-large-values.test.ts b/apps/sim/lib/logs/execution/pii-large-values.test.ts
index 15292ef1419..3176b1131e3 100644
--- a/apps/sim/lib/logs/execution/pii-large-values.test.ts
+++ b/apps/sim/lib/logs/execution/pii-large-values.test.ts
@@ -1,33 +1,41 @@
 /**
  * @vitest-environment node
  */
+import { sleep } from '@sim/utils/helpers'
 import { beforeEach, describe, expect, it, vi } from 'vitest'
 
-const { mockMaterializeRef, mockCompact, mockMaterializeManifest, mockMaskBatch } = vi.hoisted(
-  () => ({
-    mockMaterializeRef: vi.fn(),
-    mockCompact: vi.fn(),
-    mockMaterializeManifest: vi.fn(),
-    mockMaskBatch: vi.fn(),
-  })
-)
+const { mockMaterializeRef, mockStoreLargeValue, mockCompact, mockMaskBatch } = vi.hoisted(() => ({
+  mockMaterializeRef: vi.fn(),
+  mockStoreLargeValue: vi.fn(),
+  mockCompact: vi.fn(),
+  mockMaskBatch: vi.fn(),
+}))
 
 vi.mock('@/lib/execution/payloads/store', () => ({
   materializeLargeValueRef: mockMaterializeRef,
+  storeLargeValue: mockStoreLargeValue,
 }))
 vi.mock('@/lib/execution/payloads/serializer', () => ({
   compactExecutionPayload: mockCompact,
 }))
-vi.mock('@/lib/execution/payloads/large-array-manifest', () => ({
-  materializeLargeArrayManifest: mockMaterializeManifest,
-}))
-vi.mock('@/lib/execution/payloads/large-array-manifest-metadata', () => ({
-  isLargeArrayManifest: () => false,
-}))
+vi.mock('@/lib/execution/payloads/materialization.server', () => {
+  const MAX_INLINE_MATERIALIZATION_BYTES = 16 * 1024 * 1024
+  return {
+    MAX_INLINE_MATERIALIZATION_BYTES,
+    MAX_DURABLE_LARGE_VALUE_BYTES: 64 * 1024 * 1024,
+    assertInlineMaterializationSize: (size: number, maxBytes?: number) => {
+      if (size > (maxBytes ?? MAX_INLINE_MATERIALIZATION_BYTES)) {
+        throw new Error('Execution memory limit exceeded. Reduce payload size and try again.')
+      }
+    },
+  }
+})
 vi.mock('@/lib/guardrails/mask-client', () => ({
   maskPIIBatchViaHttp: mockMaskBatch,
 }))
 
+import type { LargeArrayManifest } from '@/lib/execution/payloads/large-array-manifest'
+import { isLargeArrayManifest } from '@/lib/execution/payloads/large-array-manifest-metadata'
 import {
   redactLargeValueRefs,
   redactLargeValueRefsInValue,
@@ -44,16 +52,58 @@ const REF = {
 
 const STORE = { workspaceId: 'ws-1', workflowId: 'wf-1', executionId: 'ex-1', userId: 'u-1' }
 
-describe('redactLargeValueRefs', () => {
-  beforeEach(() => {
-    vi.clearAllMocks()
-    mockMaskBatch.mockImplementation(async (texts: string[]) => texts.map((t) => `MASKED(${t})`))
-    // compact echoes its input so we can assert the masked content is what's re-stored.
-    mockCompact.mockImplementation(async (value: unknown) => value)
+/** Chunk arrays keyed by ref id, served by mockMaterializeRef and captured by mockStoreLargeValue. */
+const chunkData = new Map<string, unknown>()
+let storedRefCounter = 0
+
+function makeChunkRef(id: string, size: number, kind = 'array') {
+  return { __simLargeValueRef: true, version: 1, id, kind, size } as const
+}
+
+function makeManifest(
+  chunks: Array<{ id: string; size: number; items: unknown[] }>
+): LargeArrayManifest {
+  for (const chunk of chunks) {
+    chunkData.set(chunk.id, chunk.items)
+  }
+  return {
+    __simLargeArrayManifest: true,
+    version: 2,
+    kind: 'array',
+    totalCount: chunks.reduce((sum, c) => sum + c.items.length, 0),
+    chunkCount: chunks.length,
+    byteSize: chunks.reduce((sum, c) => sum + c.size, 0),
+    chunks: chunks.map((c) => ({
+      ref: makeChunkRef(c.id, c.size),
+      count: c.items.length,
+      byteSize: c.size,
+    })),
+    preview: [{ note: 'raw unmasked preview' }],
+  }
+}
+
+function installDefaultMocks() {
+  mockMaskBatch.mockImplementation(async (texts: string[]) => texts.map((t) => `MASKED(${t})`))
+  // compact echoes its input so we can assert the masked content is what's re-stored.
+  mockCompact.mockImplementation(async (value: unknown) => value)
+  mockMaterializeRef.mockImplementation(async (ref: { id: string }) => chunkData.get(ref.id))
+  mockStoreLargeValue.mockImplementation(async (value: unknown, _json: string, size: number) => {
+    const id = `lv_${String(storedRefCounter++).padStart(12, '0')}`
+    chunkData.set(id, value)
+    return { __simLargeValueRef: true, version: 1, id, kind: 'array', size }
   })
+}
+
+beforeEach(() => {
+  vi.clearAllMocks()
+  chunkData.clear()
+  storedRefCounter = 0
+  installDefaultMocks()
+})
 
+describe('redactLargeValueRefs', () => {
   it('hydrates, masks, and re-stores a large-value ref (content preserved, PII masked)', async () => {
-    mockMaterializeRef.mockResolvedValue({ note: 'contact bob', id: 42 })
+    chunkData.set(REF.id, { note: 'contact bob', id: 42 })
 
     const result = await redactLargeValueRefs(
       { finalOutput: REF },
@@ -62,7 +112,11 @@ describe('redactLargeValueRefs', () => {
 
     expect(mockMaterializeRef).toHaveBeenCalledWith(
       REF,
-      expect.objectContaining({ executionId: 'ex-1', trackReference: false })
+      expect.objectContaining({
+        executionId: 'ex-1',
+        trackReference: false,
+        maxBytes: 64 * 1024 * 1024,
+      })
     )
     expect(result.finalOutput).toEqual({ note: 'MASKED(contact bob)', id: 42 })
     expect(mockCompact).toHaveBeenCalledTimes(1)
@@ -79,7 +133,7 @@ describe('redactLargeValueRefs', () => {
   })
 
   it('falls back to the marker when re-store throws (never leaks)', async () => {
-    mockMaterializeRef.mockResolvedValue({ note: 'secret@x.com' })
+    chunkData.set(REF.id, { note: 'secret@x.com' })
     mockCompact.mockRejectedValueOnce(new Error('s3 down'))
     const result = await redactLargeValueRefs(
       { finalOutput: REF },
@@ -91,9 +145,8 @@ describe('redactLargeValueRefs', () => {
   it('hydrates+masks refs across multiple payload keys (parallel, cross-key)', async () => {
     const refA = { ...REF, id: 'lv_aaaaaaaaaaaa' }
     const refB = { ...REF, id: 'lv_bbbbbbbbbbbb' }
-    mockMaterializeRef.mockImplementation(async (ref: { id: string }) =>
-      ref.id === 'lv_aaaaaaaaaaaa' ? { note: 'call bob' } : { note: 'email amy' }
-    )
+    chunkData.set(refA.id, { note: 'call bob' })
+    chunkData.set(refB.id, { note: 'email amy' })
 
     const result = await redactLargeValueRefs(
       { finalOutput: refA, traceSpans: [{ output: refB }] },
@@ -110,9 +163,7 @@ describe('redactLargeValueRefs', () => {
     const refA = { ...REF, id: 'lv_aaaaaaaaaaaa' }
     const refB = { ...REF, id: 'lv_bbbbbbbbbbbb' }
     // refA materializes fine; refB can't — in throw mode the whole redaction must abort.
-    mockMaterializeRef.mockImplementation(async (ref: { id: string }) =>
-      ref.id === 'lv_aaaaaaaaaaaa' ? { note: 'ok' } : undefined
-    )
+    chunkData.set(refA.id, { note: 'ok' })
 
     await expect(
       redactLargeValueRefs(
@@ -132,17 +183,246 @@ describe('redactLargeValueRefs', () => {
     expect(result).toEqual(payload)
     expect(mockMaterializeRef).not.toHaveBeenCalled()
   })
+
+  it('masks a single ref past the inline ceiling with the raised durable budget', async () => {
+    const bigRef = { ...REF, id: 'lv_bigbigbigbig', size: 30 * 1024 * 1024 }
+    chunkData.set(bigRef.id, { note: 'ssn 123-45-6789' })
+
+    const result = await redactLargeValueRefs(
+      { finalOutput: bigRef },
+      { entityTypes: [], language: 'en', store: STORE, onFailure: 'throw' }
+    )
+
+    expect(mockMaterializeRef).toHaveBeenCalledWith(
+      bigRef,
+      expect.objectContaining({ maxBytes: 64 * 1024 * 1024 })
+    )
+    expect(result.finalOutput).toEqual({ note: 'MASKED(ssn 123-45-6789)' })
+  })
+
+  it('processes oversized single refs serially, after the pooled refs', async () => {
+    const smallRef = { ...REF, id: 'lv_smallsmall12' }
+    const bigRef = { ...REF, id: 'lv_bigbigbigbig', size: 30 * 1024 * 1024 }
+    chunkData.set(smallRef.id, { note: 'small' })
+    chunkData.set(bigRef.id, { note: 'big' })
+
+    await redactLargeValueRefs(
+      { finalOutput: { a: bigRef, b: smallRef } },
+      { entityTypes: [], language: 'en', store: STORE }
+    )
+
+    const order = mockMaterializeRef.mock.calls.map(([ref]) => (ref as { id: string }).id)
+    expect(order).toEqual(['lv_smallsmall12', 'lv_bigbigbigbig'])
+  })
+
+  it('completes an oversized ref whose content nests another oversized ref (no gate deadlock)', async () => {
+    const outer = { ...REF, id: 'lv_outerbig1234', size: 30 * 1024 * 1024 }
+    const inner = { ...REF, id: 'lv_innerbig1234', size: 20 * 1024 * 1024 }
+    chunkData.set(outer.id, { note: 'outer bob', deep: inner })
+    chunkData.set(inner.id, { note: 'inner amy' })
+
+    const result = await redactLargeValueRefs(
+      { finalOutput: outer },
+      { entityTypes: [], language: 'en', store: STORE, onFailure: 'throw' }
+    )
+
+    expect(result.finalOutput).toEqual({
+      note: 'MASKED(outer bob)',
+      deep: { note: 'MASKED(inner amy)' },
+    })
+  })
+
+  it('serializes nested oversized refs discovered under different pooled parents', async () => {
+    const parentA = { ...REF, id: 'lv_parentaaaaaa' }
+    const parentB = { ...REF, id: 'lv_parentbbbbbb' }
+    const nestedA = { ...REF, id: 'lv_nestedaaaaaa', size: 30 * 1024 * 1024 }
+    const nestedB = { ...REF, id: 'lv_nestedbbbbbb', size: 30 * 1024 * 1024 }
+    chunkData.set(parentA.id, { deep: nestedA })
+    chunkData.set(parentB.id, { deep: nestedB })
+    let inFlight = 0
+    let maxInFlight = 0
+    mockMaterializeRef.mockImplementation(async (ref: { id: string; size: number }) => {
+      if (ref.size > 16 * 1024 * 1024) {
+        inFlight++
+        maxInFlight = Math.max(maxInFlight, inFlight)
+        await sleep(1)
+        inFlight--
+      }
+      return chunkData.get(ref.id)
+    })
+    chunkData.set(nestedA.id, { note: 'a' })
+    chunkData.set(nestedB.id, { note: 'b' })
+
+    const result = await redactLargeValueRefs(
+      { finalOutput: { a: parentA, b: parentB } },
+      { entityTypes: [], language: 'en', store: STORE }
+    )
+
+    expect((result.finalOutput as any).a).toEqual({ deep: { note: 'MASKED(a)' } })
+    expect((result.finalOutput as any).b).toEqual({ deep: { note: 'MASKED(b)' } })
+    expect(maxInFlight).toBe(1)
+  })
+
+  it('processes a manifest containing an oversized chunk serially, after the pooled refs', async () => {
+    const smallRef = { ...REF, id: 'lv_smallsmall12' }
+    chunkData.set(smallRef.id, { note: 'small' })
+    const bigChunkManifest = makeManifest([
+      { id: 'lv_hugechunk000', size: 20 * 1024 * 1024, items: [{ note: 'one giant item' }] },
+    ])
+
+    await redactLargeValueRefs(
+      { finalOutput: { a: bigChunkManifest, b: smallRef } },
+      { entityTypes: [], language: 'en', store: STORE }
+    )
+
+    const order = mockMaterializeRef.mock.calls.map(([ref]) => (ref as { id: string }).id)
+    expect(order).toEqual(['lv_smallsmall12', 'lv_hugechunk000'])
+  })
 })
 
-describe('redactLargeValueRefsInValue (arbitrary blockStates)', () => {
-  beforeEach(() => {
-    vi.clearAllMocks()
-    mockMaskBatch.mockImplementation(async (texts: string[]) => texts.map((t) => `MASKED(${t})`))
-    mockCompact.mockImplementation(async (value: unknown) => value)
+describe('redactManifest — chunk-wise', () => {
+  it('masks a >16MB manifest chunk-by-chunk without tripping the inline ceiling', async () => {
+    const manifest = makeManifest([
+      { id: 'lv_chunk0chunk0', size: 9_000_000, items: [{ note: 'alpha' }, { note: 'beta' }] },
+      { id: 'lv_chunk1chunk1', size: 9_000_000, items: [{ note: 'gamma' }, { note: 'delta' }] },
+      { id: 'lv_chunk2chunk2', size: 9_000_000, items: [{ note: 'omega' }] },
+    ])
+
+    const result = await redactLargeValueRefs(
+      { finalOutput: manifest },
+      { entityTypes: ['PERSON'], language: 'en', store: STORE, onFailure: 'throw' }
+    )
+
+    const masked = result.finalOutput as LargeArrayManifest
+    expect(isLargeArrayManifest(masked)).toBe(true)
+    expect(masked.totalCount).toBe(5)
+
+    const maskedItems = masked.chunks.flatMap((chunk) => chunkData.get(chunk.ref.id) as unknown[])
+    expect(maskedItems).toEqual([
+      { note: 'MASKED(alpha)' },
+      { note: 'MASKED(beta)' },
+      { note: 'MASKED(gamma)' },
+      { note: 'MASKED(delta)' },
+      { note: 'MASKED(omega)' },
+    ])
+    // Manifests re-store through the manifest writer, never the whole-value serializer.
+    expect(mockCompact).not.toHaveBeenCalled()
+    // One hydration per source chunk, read-only and with the raised per-chunk budget.
+    const chunkReads = mockMaterializeRef.mock.calls.filter(([ref]) =>
+      String((ref as { id: string }).id).startsWith('lv_chunk')
+    )
+    expect(chunkReads).toHaveLength(3)
+    for (const [, context] of chunkReads) {
+      expect(context).toMatchObject({ trackReference: false, maxBytes: 64 * 1024 * 1024 })
+    }
+  })
+
+  it('derives the rebuilt preview from masked items, never the raw source preview', async () => {
+    const manifest = makeManifest([
+      { id: 'lv_chunk0chunk0', size: 9_000_000, items: [{ note: 'bob smith' }] },
+      { id: 'lv_chunk1chunk1', size: 9_000_000, items: [{ note: 'amy jones' }] },
+    ])
+
+    const result = await redactLargeValueRefs(
+      { finalOutput: manifest },
+      { entityTypes: ['PERSON'], language: 'en', store: STORE }
+    )
+
+    const masked = result.finalOutput as LargeArrayManifest
+    expect(masked.preview).toEqual([{ note: 'MASKED(bob smith)' }])
+  })
+
+  it('returns an empty manifest unchanged in shape for a zero-item manifest', async () => {
+    const empty: LargeArrayManifest = {
+      __simLargeArrayManifest: true,
+      version: 2,
+      kind: 'array',
+      totalCount: 0,
+      chunkCount: 0,
+      byteSize: 0,
+      chunks: [],
+      preview: [],
+    }
+
+    const result = await redactLargeValueRefs(
+      { finalOutput: empty },
+      { entityTypes: [], language: 'en', store: STORE }
+    )
+
+    const masked = result.finalOutput as LargeArrayManifest
+    expect(isLargeArrayManifest(masked)).toBe(true)
+    expect(masked.totalCount).toBe(0)
+    expect(mockStoreLargeValue).not.toHaveBeenCalled()
+  })
+
+  it('recursively masks a nested large-value ref inside a chunk item', async () => {
+    const nestedRef = { ...REF, id: 'lv_nestednested' }
+    chunkData.set(nestedRef.id, { note: 'nested bob' })
+    const manifest = makeManifest([
+      { id: 'lv_chunk0chunk0', size: 9_000_000, items: [{ note: 'top', deep: nestedRef }] },
+    ])
+
+    const result = await redactLargeValueRefs(
+      { finalOutput: manifest },
+      { entityTypes: [], language: 'en', store: STORE }
+    )
+
+    const masked = result.finalOutput as LargeArrayManifest
+    const [item] = chunkData.get(masked.chunks[0].ref.id) as Array<Record<string, unknown>>
+    expect(item.note).toBe('MASKED(top)')
+    expect(item.deep).toEqual({ note: 'MASKED(nested bob)' })
+  })
+
+  describe('partial chunk failure', () => {
+    it('scrubs the whole manifest to the marker in scrub mode', async () => {
+      const manifest = makeManifest([
+        { id: 'lv_chunk0chunk0', size: 9_000_000, items: [{ note: 'ok' }] },
+        { id: 'lv_chunk1chunk1', size: 9_000_000, items: [{ note: 'lost' }] },
+      ])
+      chunkData.delete('lv_chunk1chunk1')
+
+      const result = await redactLargeValueRefs(
+        { finalOutput: manifest },
+        { entityTypes: [], language: 'en', store: STORE }
+      )
+
+      expect(result.finalOutput).toBe('[REDACTION_FAILED]')
+    })
+
+    it('throws PiiRedactionError in throw mode', async () => {
+      const manifest = makeManifest([
+        { id: 'lv_chunk0chunk0', size: 9_000_000, items: [{ note: 'ok' }] },
+        { id: 'lv_chunk1chunk1', size: 9_000_000, items: [{ note: 'lost' }] },
+      ])
+      chunkData.delete('lv_chunk1chunk1')
+
+      await expect(
+        redactLargeValueRefs(
+          { finalOutput: manifest },
+          { entityTypes: [], language: 'en', store: STORE, onFailure: 'throw' }
+        )
+      ).rejects.toBeInstanceOf(PiiRedactionError)
+    })
+  })
+
+  it('fails fast when a chunk materializes with the wrong item count', async () => {
+    const manifest = makeManifest([
+      { id: 'lv_chunk0chunk0', size: 9_000_000, items: [{ note: 'a' }, { note: 'b' }] },
+    ])
+    chunkData.set('lv_chunk0chunk0', [{ note: 'a' }])
+
+    const result = await redactLargeValueRefs(
+      { finalOutput: manifest },
+      { entityTypes: [], language: 'en', store: STORE }
+    )
+
+    expect(result.finalOutput).toBe('[REDACTION_FAILED]')
   })
+})
 
+describe('redactLargeValueRefsInValue (arbitrary blockStates)', () => {
   it('hydrates + re-stores a ref nested in a non-RedactablePayload shape', async () => {
-    mockMaterializeRef.mockResolvedValue({ note: 'contact bob' })
+    chunkData.set(REF.id, { note: 'contact bob' })
     const blockStates = { 'block-1': { output: REF }, 'block-2': { output: { plain: 'hi' } } }
 
     const result = await redactLargeValueRefsInValue(blockStates, {
@@ -167,7 +447,7 @@ describe('redactLargeValueRefsInValue (arbitrary blockStates)', () => {
   })
 
   it('rethrows a re-store failure as PiiRedactionError under throw mode', async () => {
-    mockMaterializeRef.mockResolvedValue({ note: 'secret@x.com' })
+    chunkData.set(REF.id, { note: 'secret@x.com' })
     mockCompact.mockRejectedValueOnce(new Error('s3 down'))
 
     await expect(
diff --git a/apps/sim/lib/logs/execution/pii-large-values.ts b/apps/sim/lib/logs/execution/pii-large-values.ts
index 1797964a065..dd323eb6b46 100644
--- a/apps/sim/lib/logs/execution/pii-large-values.ts
+++ b/apps/sim/lib/logs/execution/pii-large-values.ts
@@ -3,9 +3,17 @@ import { getErrorMessage } from '@sim/utils/errors'
 import { env } from '@/lib/core/config/env'
 import { mapWithConcurrency } from '@/lib/core/utils/concurrency'
 import type { LargeArrayManifest } from '@/lib/execution/payloads/large-array-manifest'
-import { materializeLargeArrayManifest } from '@/lib/execution/payloads/large-array-manifest'
+import {
+  appendLargeArrayManifest,
+  createLargeArrayManifest,
+  readLargeArrayManifestSlice,
+} from '@/lib/execution/payloads/large-array-manifest'
 import { isLargeArrayManifest } from '@/lib/execution/payloads/large-array-manifest-metadata'
 import { isLargeValueRef, type LargeValueRef } from '@/lib/execution/payloads/large-value-ref'
+import {
+  MAX_DURABLE_LARGE_VALUE_BYTES,
+  MAX_INLINE_MATERIALIZATION_BYTES,
+} from '@/lib/execution/payloads/materialization.server'
 import { compactExecutionPayload } from '@/lib/execution/payloads/serializer'
 import type { LargeValueStoreContext } from '@/lib/execution/payloads/store'
 import { materializeLargeValueRef } from '@/lib/execution/payloads/store'
@@ -35,6 +43,41 @@ export interface RedactLargeValueRefsOptions {
    * than feeding a marker (or leaking raw bytes) downstream.
    */
   onFailure?: PiiRedactionFailureMode
+  /**
+   * @internal Shared across every nested invocation spawned from one entry-point
+   * call, so oversized hydrations serialize globally — not just within their own
+   * {@link resolveReplacements} pass. Created by the entry points; callers never
+   * set it.
+   */
+  oversizedGate?: OversizedHydrationGate
+  /**
+   * @internal True while running under the gate. Nested oversized work inside a
+   * gated ref runs directly instead of re-acquiring — the holder would otherwise
+   * wait on itself (deadlock). Safe because the holder is globally serialized.
+   */
+  withinOversizedGate?: boolean
+}
+
+/** Promise-chain mutex for oversized hydrations (see {@link requiresSerialHydration}). */
+interface OversizedHydrationGate {
+  chain: Promise<void>
+}
+
+/** Queue `fn` behind every previously gated task; failures don't poison the chain. */
+async function runSerially<T>(gate: OversizedHydrationGate, fn: () => Promise<T>): Promise<T> {
+  const run = gate.chain.then(fn)
+  gate.chain = run.then(
+    () => undefined,
+    () => undefined
+  )
+  return run
+}
+
+/** Ensure one shared gate exists for the whole entry-point call tree. */
+function withOversizedGate(options: RedactLargeValueRefsOptions): RedactLargeValueRefsOptions {
+  return options.oversizedGate
+    ? options
+    : { ...options, oversizedGate: { chain: Promise.resolve() } }
 }
 
 /**
@@ -57,6 +100,7 @@ export async function redactLargeValueRefs(
   payload: RedactablePayload,
   options: RedactLargeValueRefsOptions
 ): Promise<RedactablePayload> {
+  const gatedOptions = withOversizedGate(options)
   // Collect refs across the WHOLE payload first (shared `seen`), so every ref is
   // hydrated+masked+re-stored in one bounded-concurrency pass instead of one
   // sequential pass per key. A ref shared across keys is walked/masked once.
@@ -67,7 +111,7 @@ export async function redactLargeValueRefs(
   }
   if (refs.length === 0) return payload
 
-  const replacements = await resolveReplacements(refs, options)
+  const replacements = await resolveReplacements(refs, gatedOptions)
   const result: RedactablePayload = { ...payload }
   for (const key of Object.keys(payload) as (keyof RedactablePayload)[]) {
     if (payload[key] !== undefined) result[key] = substituteRefs(payload[key], replacements)
@@ -85,7 +129,7 @@ export async function redactLargeValueRefsInValue<T>(
   value: T,
   options: RedactLargeValueRefsOptions
 ): Promise<T> {
-  return (await redactValueRefs(value, options)) as T
+  return (await redactValueRefs(value, withOversizedGate(options))) as T
 }
 
 /** Sync-collect every ref/manifest in `value`, then async-replace each, then sync-substitute. */
@@ -108,10 +152,30 @@ async function redactValueRefs(
  */
 const REF_CONCURRENCY = env.PII_REF_CONCURRENCY ?? 4
 
+/**
+ * A single ref past the default inline ceiling hydrates its whole blob at once
+ * (~2-3× its serialized size in transient heap) — and a manifest whose packer
+ * emitted an oversized chunk (a single item larger than the chunk target)
+ * hydrates that chunk the same way. Those run serially instead of in the
+ * {@link REF_CONCURRENCY} pool — one 64MB hydration is fine, four at once is an
+ * OOM on the small trigger.dev machines. Normally-chunked manifests page ~8MB
+ * at a time and stay pooled regardless of total size.
+ */
+function requiresSerialHydration(ref: object): boolean {
+  if (isLargeValueRef(ref)) {
+    return ref.size > MAX_INLINE_MATERIALIZATION_BYTES
+  }
+  if (isLargeArrayManifest(ref)) {
+    return ref.chunks.some((chunk) => chunk.byteSize > MAX_INLINE_MATERIALIZATION_BYTES)
+  }
+  return false
+}
+
 /**
  * Dedupe the collected refs by identity, then replace each in parallel (bounded by
- * {@link REF_CONCURRENCY}). `Map.set` is synchronous, so concurrent workers writing
- * the shared map do not race.
+ * {@link REF_CONCURRENCY}); refs needing an oversized hydration run serially after
+ * the pool drains (see {@link requiresSerialHydration}). `Map.set` is synchronous,
+ * so concurrent workers writing the shared map do not race.
  *
  * `mapWithConcurrency`'s `fn` must not reject (a rejection fails the pool
  * non-deterministically), so the mapper is total: it catches per-ref errors and
@@ -124,15 +188,28 @@ async function resolveReplacements(
   options: RedactLargeValueRefsOptions
 ): Promise<Map<object, unknown>> {
   const unique = [...new Set(refs)]
+  const pooled = unique.filter((ref) => !requiresSerialHydration(ref))
+  const oversized = unique.filter(requiresSerialHydration)
   const replacements = new Map<object, unknown>()
   let firstError: unknown
-  await mapWithConcurrency(unique, REF_CONCURRENCY, async (ref) => {
+  const resolveOne = async (ref: object, opts: RedactLargeValueRefsOptions): Promise<void> => {
     try {
-      replacements.set(ref, await replaceRef(ref, options))
+      replacements.set(ref, await replaceRef(ref, opts))
     } catch (error) {
       if (firstError === undefined) firstError = error
     }
-  })
+  }
+  await mapWithConcurrency(pooled, REF_CONCURRENCY, (ref) => resolveOne(ref, options))
+  const gate = options.oversizedGate
+  for (const ref of oversized) {
+    if (!gate || options.withinOversizedGate) {
+      // Already serialized by a gated ancestor (or no gate: direct internal
+      // call) — re-acquiring would deadlock on the ancestor's own hold.
+      await resolveOne(ref, options)
+    } else {
+      await runSerially(gate, () => resolveOne(ref, { ...options, withinOversizedGate: true }))
+    }
+  }
   if (firstError !== undefined) throw firstError
   return replacements
 }
@@ -176,21 +253,32 @@ async function replaceRef(ref: object, options: RedactLargeValueRefsOptions): Pr
 }
 
 /**
- * Mask a materialized large value and re-offload it: handle any nested refs
- * first, then mask inline strings, then re-store. `redactObjectStrings` skips
- * refs, so the nested re-stored refs are left intact while their siblings mask.
+ * Mask a materialized value: mask inline strings first (`redactObjectStrings`
+ * treats refs as opaque), then replace any nested refs with their own masked
+ * results. Strings-first matters — a nested ref whose masked value shrinks
+ * below the offload threshold comes back inline, and running the string pass
+ * after substitution would mask that already-masked content a second time
+ * (non-idempotent for custom patterns).
  */
-async function maskAndReStore(
+async function maskMaterializedValue(
   value: unknown,
   options: RedactLargeValueRefsOptions
 ): Promise<unknown> {
-  const nested = await redactValueRefs(value, options)
-  const masked = await redactObjectStrings(nested, {
+  const masked = await redactObjectStrings(value, {
     entityTypes: options.entityTypes,
     language: options.language,
     customPatterns: options.customPatterns,
     onFailure: options.onFailure ?? 'scrub',
   })
+  return redactValueRefs(masked, options)
+}
+
+/** Mask a materialized large value and re-offload it as a fresh durable ref. */
+async function maskAndReStore(
+  value: unknown,
+  options: RedactLargeValueRefsOptions
+): Promise<unknown> {
+  const masked = await maskMaterializedValue(value, options)
   return compactExecutionPayload(masked, { ...options.store, requireDurable: true })
 }
 
@@ -217,6 +305,11 @@ async function redactRef(
     const materialized = await materializeLargeValueRef(ref, {
       ...options.store,
       trackReference: false,
+      // Redaction must hydrate refs past the default inline ceiling to mask
+      // them; refs above that ceiling are scheduled serially (see
+      // isOversizedSingleRef) so the raised budget never multiplies across the
+      // concurrency pool.
+      maxBytes: MAX_DURABLE_LARGE_VALUE_BYTES,
     })
     if (materialized === undefined) {
       return onRefFailure(
@@ -231,13 +324,47 @@ async function redactRef(
   }
 }
 
+/**
+ * Mask a large-array manifest chunk-by-chunk: page one stored chunk at a time,
+ * mask nested refs + strings in that slice, and append the masked items to a
+ * fresh manifest. Peak memory stays ~one chunk regardless of the manifest's
+ * total `byteSize`, so a large offloaded array never trips the inline
+ * materialization ceiling. The rebuilt manifest re-chunks by byte target,
+ * recomputes `count`/`byteSize` bookkeeping from the masked items, and derives
+ * `preview` from masked content — the source manifest's preview holds raw items
+ * and must never be carried forward.
+ */
 async function redactManifest(
   manifest: LargeArrayManifest,
   options: RedactLargeValueRefsOptions
 ): Promise<unknown> {
   try {
-    const materialized = await materializeLargeArrayManifest(manifest, { ...options.store })
-    return await maskAndReStore(materialized, options)
+    const readContext = {
+      ...options.store,
+      trackReference: false,
+      // Chunks target ~8MB, but a single item larger than the target still
+      // occupies its own chunk — allow those up to the durable cap since they
+      // hydrate one at a time.
+      maxBytes: MAX_DURABLE_LARGE_VALUE_BYTES,
+    }
+    let masked: LargeArrayManifest | undefined
+    let cursor = 0
+    for (const chunk of manifest.chunks) {
+      const slice = await readLargeArrayManifestSlice(manifest, cursor, chunk.count, readContext)
+      cursor += chunk.count
+      const items = (await maskMaterializedValue(slice, options)) as unknown[]
+      masked =
+        masked === undefined
+          ? await createLargeArrayManifest(items, { ...options.store })
+          : await appendLargeArrayManifest(masked, items, { ...options.store })
+    }
+    if (masked === undefined) {
+      return createLargeArrayManifest([], { ...options.store })
+    }
+    if (masked.totalCount !== manifest.totalCount) {
+      throw new Error('Masked manifest item count does not match the source manifest.')
+    }
+    return masked
   } catch (error) {
     return onRefFailure(error, options, 'Failed to redact large array manifest')
   }
diff --git a/apps/sim/lib/logs/execution/pii-redaction.ts b/apps/sim/lib/logs/execution/pii-redaction.ts
index d29c4cb864b..cfca0c8ea0e 100644
--- a/apps/sim/lib/logs/execution/pii-redaction.ts
+++ b/apps/sim/lib/logs/execution/pii-redaction.ts
@@ -145,7 +145,9 @@ function transformUnit(
  *
  * There is no total-size ceiling — the batching layer chunks the request and
  * fans out with bounded concurrency, so payloads of any size are masked properly
- * rather than scrubbed. The per-chunk request timeout is the real backstop.
+ * rather than scrubbed. Transient chunk failures retry with backoff inside the
+ * mask client; a failure surfacing here means the retry budget is exhausted or
+ * the failure is deterministic.
  */
 async function maskCollected(
   collected: string[],