diff --git a/packages/engine/src/config.ts b/packages/engine/src/config.ts
index ea8055c4..a07d3a13 100644
--- a/packages/engine/src/config.ts
+++ b/packages/engine/src/config.ts
@@ -72,6 +72,30 @@ export interface EngineConfig {
   /** Custom manifest path for Hyperframe runtime. */
   runtimeManifestPath?: string;
 
+  // ── Cache ────────────────────────────────────────────────────────────
+  /**
+   * Directory where the content-addressed extraction cache persists frame
+   * bundles keyed on (path, mtime, size, mediaStart, duration, fps, format).
+   * Undefined disables caching — extraction runs into the render's workDir
+   * and cleanup removes it when the render ends, preserving the pre-cache
+   * behaviour.
+   *
+   * **Single-writer.** The cache is not safe for concurrent renders pointing
+   * at the same directory. A `.hf-complete` sentinel prevents another render
+   * from serving an entry that hasn't finished extracting, but individual
+   * frame files are written non-atomically — a second render reading during
+   * the write window can observe a truncated frame. Give each concurrent
+   * render pipeline its own `extractCacheDir`, or gate with an external mutex.
+   *
+   * **Network filesystems.** `mtime` resolution on NFS/SMB mounts can be
+   * coarser than expected (seconds rather than nanoseconds), which may
+   * produce spurious cache hits if a source file is overwritten within the
+   * same mtime tick. Local filesystems are the intended deployment target.
+   *
+   * Env fallback: `HYPERFRAMES_EXTRACT_CACHE_DIR`.
+   */
+  extractCacheDir?: string;
+
   // ── Debug ────────────────────────────────────────────────────────────
   debug: boolean;
 }
@@ -203,6 +227,8 @@ export function resolveConfig(overrides?: Partial<EngineConfig>): EngineConfig {
 
     verifyRuntime: env("PRODUCER_VERIFY_HYPERFRAME_RUNTIME") !== "false",
     runtimeManifestPath: env("PRODUCER_HYPERFRAME_MANIFEST_PATH"),
+
+    extractCacheDir: env("HYPERFRAMES_EXTRACT_CACHE_DIR"),
   };
 
   // Remove undefined values so they don't override defaults
diff --git a/packages/engine/src/services/extractionCache.test.ts b/packages/engine/src/services/extractionCache.test.ts
new file mode 100644
index 00000000..82cd09a6
--- /dev/null
+++ b/packages/engine/src/services/extractionCache.test.ts
@@ -0,0 +1,199 @@
+import { afterEach, beforeEach, describe, expect, it } from "vitest";
+import { existsSync, mkdirSync, mkdtempSync, rmSync, writeFileSync } from "node:fs";
+import { join } from "node:path";
+import { tmpdir } from "node:os";
+
+import {
+  COMPLETE_SENTINEL,
+  FRAME_FILENAME_PREFIX,
+  SCHEMA_PREFIX,
+  cacheEntryDirName,
+  computeCacheKey,
+  ensureCacheEntryDir,
+  lookupCacheEntry,
+  markCacheEntryComplete,
+  readKeyStat,
+  type CacheKeyInput,
+} from "./extractionCache.js";
+
+const keyFor = (videoPath: string, overrides: Partial<CacheKeyInput> = {}): CacheKeyInput => {
+  const stat = readKeyStat(videoPath);
+  if (!stat) throw new Error(`keyFor fixture missing on disk: ${videoPath}`);
+  return {
+    videoPath,
+    mtimeMs: stat.mtimeMs,
+    size: stat.size,
+    mediaStart: 0,
+    duration: 3,
+    fps: 30,
+    format: "jpg",
+    ...overrides,
+  };
+};
+
+describe("extractionCache constants", () => {
+  it("exposes the v2 schema prefix", () => {
+    expect(SCHEMA_PREFIX).toBe("hfcache-v2-");
+  });
+
+  it("exposes the frame filename prefix shared with the extractor", () => {
+    expect(FRAME_FILENAME_PREFIX).toBe("frame_");
+  });
+
+  it("uses a dotfile sentinel so ls-without-A hides it", () => {
+    expect(COMPLETE_SENTINEL.startsWith(".")).toBe(true);
+  });
+});
+
+describe("computeCacheKey", () => {
+  let tmpRoot: string;
+  let sourceFile: string;
+
+  beforeEach(() => {
+    tmpRoot = mkdtempSync(join(tmpdir(), "hf-extract-cache-test-"));
+    sourceFile = join(tmpRoot, "clip.mp4");
+    writeFileSync(sourceFile, "fake-video-bytes", "utf-8");
+  });
+
+  afterEach(() => {
+    if (existsSync(tmpRoot)) rmSync(tmpRoot, { recursive: true, force: true });
+  });
+
+  const base = (videoPath: string): CacheKeyInput => keyFor(videoPath);
+
+  it("returns the same key for identical inputs", () => {
+    const a = computeCacheKey(base(sourceFile));
+    const b = computeCacheKey(base(sourceFile));
+    expect(a).toBe(b);
+  });
+
+  it("produces a 64-char hex SHA-256 digest", () => {
+    const key = computeCacheKey(base(sourceFile));
+    expect(key).toMatch(/^[0-9a-f]{64}$/);
+  });
+
+  it("changes when path changes (moved files re-extract)", () => {
+    const other = join(tmpRoot, "other.mp4");
+    writeFileSync(other, "fake-video-bytes", "utf-8");
+    const a = computeCacheKey(base(sourceFile));
+    const b = computeCacheKey(base(other));
+    expect(a).not.toBe(b);
+  });
+
+  it("changes when mediaStart changes", () => {
+    const a = computeCacheKey(base(sourceFile));
+    const b = computeCacheKey({ ...base(sourceFile), mediaStart: 1 });
+    expect(a).not.toBe(b);
+  });
+
+  it("changes when duration changes", () => {
+    const a = computeCacheKey(base(sourceFile));
+    const b = computeCacheKey({ ...base(sourceFile), duration: 5 });
+    expect(a).not.toBe(b);
+  });
+
+  it("changes when fps changes (different frame count invalidates key)", () => {
+    const a = computeCacheKey(base(sourceFile));
+    const b = computeCacheKey({ ...base(sourceFile), fps: 60 });
+    expect(a).not.toBe(b);
+  });
+
+  it("changes when format changes", () => {
+    const a = computeCacheKey(base(sourceFile));
+    const b = computeCacheKey({ ...base(sourceFile), format: "png" });
+    expect(a).not.toBe(b);
+  });
+
+  it("normalizes non-finite duration so Infinity doesn't produce unstable keys", () => {
+    const a = computeCacheKey({ ...base(sourceFile), duration: Infinity });
+    const b = computeCacheKey({ ...base(sourceFile), duration: Infinity });
+    expect(a).toBe(b);
+  });
+
+  it("changes when file content changes (mtime+size bump)", () => {
+    const before = computeCacheKey(base(sourceFile));
+    // Force an mtime change by waiting 5ms then overwriting with different bytes.
+    // 5ms is well above the Linux mtime resolution (typically nanoseconds) and
+    // below any Windows cache coherency window. Using a longer sleep pads against
+    // coarse filesystem mtime granularity without slowing the suite.
+    const start = Date.now();
+    while (Date.now() - start < 5) {
+      /* spin */
+    }
+    writeFileSync(sourceFile, "different-bytes-longer-than-before", "utf-8");
+    const after = computeCacheKey(base(sourceFile));
+    expect(after).not.toBe(before);
+  });
+
+  it("readKeyStat returns null for a missing source (callers skip the cache)", () => {
+    // Previously readKeyStat returned a `{mtimeMs: 0, size: 0}` sentinel for
+    // missing files; two unrelated missing paths then shared the same cache
+    // key tuple and polluted the cache. The contract now returns null so
+    // callers can explicitly skip the cache path and let the extractor
+    // surface the real file-not-found error.
+    const missing = join(tmpRoot, "does-not-exist.mp4");
+    expect(readKeyStat(missing)).toBeNull();
+  });
+});
+
+describe("cacheEntryDirName", () => {
+  it("prefixes with the schema and truncates to 16 hex chars", () => {
+    const full = "a".repeat(64);
+    expect(cacheEntryDirName(full)).toBe(`${SCHEMA_PREFIX}${"a".repeat(16)}`);
+  });
+});
+
+describe("lookupCacheEntry / markCacheEntryComplete", () => {
+  let tmpRoot: string;
+  let sourceFile: string;
+
+  beforeEach(() => {
+    tmpRoot = mkdtempSync(join(tmpdir(), "hf-extract-cache-test-"));
+    sourceFile = join(tmpRoot, "clip.mp4");
+    writeFileSync(sourceFile, "fake-video-bytes", "utf-8");
+  });
+
+  afterEach(() => {
+    if (existsSync(tmpRoot)) rmSync(tmpRoot, { recursive: true, force: true });
+  });
+
+  const base = (videoPath: string): CacheKeyInput => keyFor(videoPath);
+
+  it("misses on an empty cache root", () => {
+    const lookup = lookupCacheEntry(tmpRoot, base(sourceFile));
+    expect(lookup.hit).toBe(false);
+    expect(lookup.entry.dir.startsWith(tmpRoot)).toBe(true);
+  });
+
+  it("hits after ensureCacheEntryDir + markCacheEntryComplete", () => {
+    const first = lookupCacheEntry(tmpRoot, base(sourceFile));
+    ensureCacheEntryDir(first.entry);
+    markCacheEntryComplete(first.entry);
+
+    const second = lookupCacheEntry(tmpRoot, base(sourceFile));
+    expect(second.hit).toBe(true);
+    expect(second.entry.dir).toBe(first.entry.dir);
+  });
+
+  it("treats an in-progress dir without the sentinel as a miss", () => {
+    const lookup = lookupCacheEntry(tmpRoot, base(sourceFile));
+    ensureCacheEntryDir(lookup.entry);
+    // Simulate abandoned extraction — frames written but sentinel never marked.
+    writeFileSync(join(lookup.entry.dir, "frame_00001.jpg"), "x", "utf-8");
+    const again = lookupCacheEntry(tmpRoot, base(sourceFile));
+    expect(again.hit).toBe(false);
+  });
+
+  it("places entries under the cache root, not the source parent", () => {
+    const subroot = join(tmpRoot, "cache-root");
+    mkdirSync(subroot, { recursive: true });
+    const lookup = lookupCacheEntry(subroot, base(sourceFile));
+    expect(lookup.entry.dir.startsWith(subroot)).toBe(true);
+  });
+
+  it("uses the same directory for identical inputs across lookups", () => {
+    const a = lookupCacheEntry(tmpRoot, base(sourceFile));
+    const b = lookupCacheEntry(tmpRoot, base(sourceFile));
+    expect(a.entry.dir).toBe(b.entry.dir);
+  });
+});
diff --git a/packages/engine/src/services/extractionCache.ts b/packages/engine/src/services/extractionCache.ts
new file mode 100644
index 00000000..53859624
--- /dev/null
+++ b/packages/engine/src/services/extractionCache.ts
@@ -0,0 +1,216 @@
+/**
+ * Content-Addressed Extraction Cache
+ *
+ * Video frame extraction is the single most expensive phase of a render
+ * after capture. Repeat renders of the same composition (preview → final,
+ * studio iteration) re-extract identical frames from the same source file,
+ * burning ffmpeg time that adds no value. This module keys extracted frame
+ * bundles on the (path, mtime, size, mediaStart, duration, fps, format)
+ * tuple so re-renders resolve to a pre-extracted directory instead of
+ * re-invoking ffmpeg.
+ *
+ * ### Scheme
+ *
+ * - The key is the SHA-256 of a stable JSON encoding of the tuple above.
+ * - Cache entries live under `<rootDir>/<SCHEMA_PREFIX><key[0..16]>/` so
+ *   `ls` output and tracing logs stay short. Truncation to 16 hex chars
+ *   leaves 64 bits of entropy — collision risk at cache scale is negligible.
+ * - A completed entry is marked by writing the `.hf-complete` sentinel file
+ *   after all frames are on disk. A dir without the sentinel is treated as
+ *   absent (stale/abandoned) and re-extracted into a fresh key (the old dir
+ *   is left for external gc — the cache owns keys, not deletion policy).
+ *
+ * ### Versioning
+ *
+ * `SCHEMA_PREFIX` bumps when the cache-contents invariant changes (e.g.
+ * extraction format, frame layout). Old entries under the previous prefix
+ * become inert and can be gc'd by the caller.
+ */
+
+import { createHash } from "node:crypto";
+import { mkdirSync, readdirSync, statSync, writeFileSync } from "node:fs";
+import { existsSync } from "node:fs";
+import { join } from "node:path";
+import type { VideoMetadata } from "../utils/ffprobe.js";
+
+/** Filename prefix for extracted frames. Shared with the extractor. */
+export const FRAME_FILENAME_PREFIX = "frame_";
+
+/** Sentinel filename written after a cache entry is fully populated. */
+export const COMPLETE_SENTINEL = ".hf-complete";
+
+/** Current schema version. Bump when cache-entry layout changes. */
+export const SCHEMA_PREFIX = "hfcache-v2-";
+
+/** Truncated hex chars of SHA-256 used for the entry directory name. */
+const KEY_HEX_CHARS = 16;
+
+export type CacheFrameFormat = "jpg" | "png";
+
+export interface CacheKeyInput {
+  /** Absolute path to the source video file. Part of the key so moved files
+   *  re-extract rather than match by (size, mtime) alone. */
+  videoPath: string;
+  /** Source file modification time in ms (floored). Invalidates the key on edit. */
+  mtimeMs: number;
+  /** Source file size in bytes. Invalidates the key on content change. */
+  size: number;
+  /** Seconds into source the composition starts reading (video.mediaStart). */
+  mediaStart: number;
+  /** Seconds of source the composition uses. Infinity is normalized to -1
+   *  so callers that pass an unresolved "natural duration" still produce a
+   *  stable key across invocations. */
+  duration: number;
+  /** Target output frames-per-second. */
+  fps: number;
+  /** Output image format. */
+  format: CacheFrameFormat;
+}
+
+export interface CacheEntry {
+  /** Absolute path to the cache entry directory. */
+  dir: string;
+  /** Full 64-char SHA-256 hex digest (parent of the truncated key). */
+  keyHash: string;
+}
+
+export interface CacheLookup {
+  /** Cache entry information — always returned even on a miss so the caller
+   *  can extract directly into `dir` then call `markCacheEntryComplete`. */
+  entry: CacheEntry;
+  /** True when the entry exists AND carries the completion sentinel. */
+  hit: boolean;
+}
+
+/**
+ * Read `(mtimeMs, size)` for a path. Returns `null` if the file is missing —
+ * callers should skip the cache path for that entry so the extractor surfaces
+ * the real file-not-found error. Returning a zero-stat sentinel would let two
+ * missing files share the same `(0, 0)` tuple and pollute the cache with an
+ * orphaned entry.
+ */
+export function readKeyStat(videoPath: string): { mtimeMs: number; size: number } | null {
+  try {
+    const stat = statSync(videoPath);
+    return { mtimeMs: Math.floor(stat.mtimeMs), size: stat.size };
+  } catch {
+    return null;
+  }
+}
+
+function canonicalKeyBlob(input: CacheKeyInput): string {
+  const durationForKey = Number.isFinite(input.duration) ? input.duration : -1;
+  return JSON.stringify({
+    p: input.videoPath,
+    m: input.mtimeMs,
+    s: input.size,
+    ms: input.mediaStart,
+    d: durationForKey,
+    f: input.fps,
+    fmt: input.format,
+  });
+}
+
+/**
+ * Compute the SHA-256 hex digest for a cache key input.
+ */
+export function computeCacheKey(input: CacheKeyInput): string {
+  return createHash("sha256").update(canonicalKeyBlob(input)).digest("hex");
+}
+
+/**
+ * Derive the truncated cache-entry directory name from a full key hash.
+ * Exposed so tests and the entry dir resolver share one truncation rule.
+ */
+export function cacheEntryDirName(keyHash: string): string {
+  return SCHEMA_PREFIX + keyHash.slice(0, KEY_HEX_CHARS);
+}
+
+/**
+ * Look up a cache entry by key input. Returns the resolved entry path plus a
+ * `hit` flag. On miss, callers should extract frames into `entry.dir`
+ * (after calling `ensureCacheEntryDir`) and then call `markCacheEntryComplete`
+ * once the extraction succeeds.
+ */
+export function lookupCacheEntry(rootDir: string, input: CacheKeyInput): CacheLookup {
+  const keyHash = computeCacheKey(input);
+  const dir = join(rootDir, cacheEntryDirName(keyHash));
+  const complete = existsSync(join(dir, COMPLETE_SENTINEL));
+  return { entry: { dir, keyHash }, hit: complete };
+}
+
+/**
+ * Ensure a cache entry's directory exists so the extractor can write into it.
+ * Idempotent: `mkdirSync({recursive:true})` is a no-op when the dir exists.
+ */
+export function ensureCacheEntryDir(entry: CacheEntry): void {
+  mkdirSync(entry.dir, { recursive: true });
+}
+
+/**
+ * Write the completion sentinel so subsequent lookups treat this entry as a
+ * hit. Must be called only after every frame has been written.
+ *
+ * Concurrency: lookup→populate→mark is non-atomic. Two concurrent renders of
+ * the same key may both miss, both extract into the same dir, and the later
+ * writer's frames win. The result is correct (identical inputs yield identical
+ * frames) but wasteful. Acceptable for a single-process render pipeline;
+ * anyone running concurrent renders against a shared cache root should front
+ * it with an external lock.
+ */
+export function markCacheEntryComplete(entry: CacheEntry): void {
+  writeFileSync(join(entry.dir, COMPLETE_SENTINEL), "", "utf-8");
+}
+
+/**
+ * Rebuild the in-memory frame index for a cached entry. Called on cache hits
+ * so the extractor's caller receives the same `ExtractedFrames` shape it
+ * would get from a fresh extraction — without re-running ffmpeg or ffprobe.
+ *
+ * The `metadata` argument is the `VideoMetadata` probed in the extractor's
+ * Phase 2 (pre-preflight). Passing it here avoids an extra ffprobe on the
+ * hit path.
+ */
+export interface RehydrateOptions {
+  videoId: string;
+  srcPath: string;
+  fps: number;
+  format: CacheFrameFormat;
+  metadata: VideoMetadata;
+}
+
+export interface RehydratedFrames {
+  videoId: string;
+  srcPath: string;
+  outputDir: string;
+  framePattern: string;
+  fps: number;
+  totalFrames: number;
+  metadata: VideoMetadata;
+  framePaths: Map<number, string>;
+}
+
+export function rehydrateCacheEntry(
+  entry: CacheEntry,
+  options: RehydrateOptions,
+): RehydratedFrames {
+  const framePattern = `${FRAME_FILENAME_PREFIX}%05d.${options.format}`;
+  const framePaths = new Map<number, string>();
+  const suffix = `.${options.format}`;
+  const files = readdirSync(entry.dir)
+    .filter((f) => f.startsWith(FRAME_FILENAME_PREFIX) && f.endsWith(suffix))
+    .sort();
+  files.forEach((file, idx) => {
+    framePaths.set(idx, join(entry.dir, file));
+  });
+  return {
+    videoId: options.videoId,
+    srcPath: options.srcPath,
+    outputDir: entry.dir,
+    framePattern,
+    fps: options.fps,
+    totalFrames: framePaths.size,
+    metadata: options.metadata,
+    framePaths,
+  };
+}
diff --git a/packages/engine/src/services/videoFrameExtractor.test.ts b/packages/engine/src/services/videoFrameExtractor.test.ts
index c63e0e93..65436f3a 100644
--- a/packages/engine/src/services/videoFrameExtractor.test.ts
+++ b/packages/engine/src/services/videoFrameExtractor.test.ts
@@ -196,6 +196,137 @@ describe.skipIf(!HAS_FFMPEG)("extractAllVideoFrames on a VFR source", () => {
     expect(result.phaseBreakdown.vfrPreflightMs).toBeGreaterThan(0);
   }, 60_000);
 
+  it("reuses extracted frames on a warm cache hit", async () => {
+    const CACHE_DIR = mkdtempSync(join(tmpdir(), "hf-extract-cache-test-"));
+    const SRC = join(FIXTURE_DIR, "cache-src.mp4");
+
+    // Synthesize a clean CFR SDR clip — bypasses VFR preflight so the cache
+    // key is stable across the two runs.
+    const synth = await runFfmpeg([
+      "-y",
+      "-hide_banner",
+      "-loglevel",
+      "error",
+      "-f",
+      "lavfi",
+      "-i",
+      "testsrc2=s=320x180:d=2:rate=30",
+      "-c:v",
+      "libx264",
+      "-preset",
+      "ultrafast",
+      "-pix_fmt",
+      "yuv420p",
+      SRC,
+    ]);
+    if (!synth.success) {
+      throw new Error(`Cache fixture synthesis failed: ${synth.stderr.slice(-400)}`);
+    }
+
+    const video: VideoElement = {
+      id: "cv1",
+      src: SRC,
+      start: 0,
+      end: 2,
+      mediaStart: 0,
+      hasAudio: false,
+    };
+
+    const outDirA = join(FIXTURE_DIR, "out-cache-miss");
+    mkdirSync(outDirA, { recursive: true });
+    const miss = await extractAllVideoFrames(
+      [video],
+      FIXTURE_DIR,
+      { fps: 30, outputDir: outDirA },
+      undefined,
+      { extractCacheDir: CACHE_DIR },
+    );
+    expect(miss.errors).toEqual([]);
+    expect(miss.phaseBreakdown.cacheHits).toBe(0);
+    expect(miss.phaseBreakdown.cacheMisses).toBe(1);
+
+    const outDirB = join(FIXTURE_DIR, "out-cache-hit");
+    mkdirSync(outDirB, { recursive: true });
+    const hit = await extractAllVideoFrames(
+      [video],
+      FIXTURE_DIR,
+      { fps: 30, outputDir: outDirB },
+      undefined,
+      { extractCacheDir: CACHE_DIR },
+    );
+    expect(hit.errors).toEqual([]);
+    expect(hit.phaseBreakdown.cacheHits).toBe(1);
+    expect(hit.phaseBreakdown.cacheMisses).toBe(0);
+    // extractMs on a hit is only the cache-lookup bookkeeping; asserting <50ms
+    // is loose enough to survive CI jitter but tight enough to catch a
+    // regression that accidentally triggered ffmpeg again.
+    expect(hit.phaseBreakdown.extractMs).toBeLessThan(50);
+    expect(hit.extracted).toHaveLength(1);
+    expect(hit.extracted[0]!.totalFrames).toBe(miss.extracted[0]!.totalFrames);
+
+    rmSync(CACHE_DIR, { recursive: true, force: true });
+  }, 60_000);
+
+  it("invalidates the cache when fps changes", async () => {
+    const CACHE_DIR = mkdtempSync(join(tmpdir(), "hf-extract-cache-test-"));
+    const SRC = join(FIXTURE_DIR, "cache-fps-src.mp4");
+
+    const synth = await runFfmpeg([
+      "-y",
+      "-hide_banner",
+      "-loglevel",
+      "error",
+      "-f",
+      "lavfi",
+      "-i",
+      "testsrc2=s=320x180:d=1:rate=30",
+      "-c:v",
+      "libx264",
+      "-preset",
+      "ultrafast",
+      "-pix_fmt",
+      "yuv420p",
+      SRC,
+    ]);
+    if (!synth.success) {
+      throw new Error(`Cache-fps fixture synthesis failed: ${synth.stderr.slice(-400)}`);
+    }
+
+    const video: VideoElement = {
+      id: "cv2",
+      src: SRC,
+      start: 0,
+      end: 1,
+      mediaStart: 0,
+      hasAudio: false,
+    };
+
+    const outA = join(FIXTURE_DIR, "out-cache-fps-30");
+    mkdirSync(outA, { recursive: true });
+    const first = await extractAllVideoFrames(
+      [video],
+      FIXTURE_DIR,
+      { fps: 30, outputDir: outA },
+      undefined,
+      { extractCacheDir: CACHE_DIR },
+    );
+    expect(first.phaseBreakdown.cacheMisses).toBe(1);
+
+    const outB = join(FIXTURE_DIR, "out-cache-fps-60");
+    mkdirSync(outB, { recursive: true });
+    const second = await extractAllVideoFrames(
+      [video],
+      FIXTURE_DIR,
+      { fps: 60, outputDir: outB },
+      undefined,
+      { extractCacheDir: CACHE_DIR },
+    );
+    expect(second.phaseBreakdown.cacheMisses).toBe(1);
+    expect(second.phaseBreakdown.cacheHits).toBe(0);
+
+    rmSync(CACHE_DIR, { recursive: true, force: true });
+  }, 60_000);
+
   // Regression test for the segment-scope HDR preflight fix: pre-fix,
   // convertSdrToHdr re-encoded the entire source, so a 30-minute SDR source
   // contributing a 2-second clip took ~200× longer than needed. Post-fix the
diff --git a/packages/engine/src/services/videoFrameExtractor.ts b/packages/engine/src/services/videoFrameExtractor.ts
index c4ff8253..8f81c6de 100644
--- a/packages/engine/src/services/videoFrameExtractor.ts
+++ b/packages/engine/src/services/videoFrameExtractor.ts
@@ -18,6 +18,15 @@ import {
 import { downloadToTemp, isHttpUrl } from "../utils/urlDownloader.js";
 import { runFfmpeg } from "../utils/runFfmpeg.js";
 import { DEFAULT_CONFIG, type EngineConfig } from "../config.js";
+import {
+  FRAME_FILENAME_PREFIX,
+  ensureCacheEntryDir,
+  lookupCacheEntry,
+  markCacheEntryComplete,
+  readKeyStat,
+  rehydrateCacheEntry,
+  type CacheFrameFormat,
+} from "./extractionCache.js";
 
 export interface VideoElement {
   id: string;
@@ -37,6 +46,13 @@ export interface ExtractedFrames {
   totalFrames: number;
   metadata: VideoMetadata;
   framePaths: Map<number, string>;
+  /**
+   * True when the extractor owns `outputDir` and cleanup should rm it when
+   * the render ends. Cache hits set this to false so the shared entry isn't
+   * deleted by a single render's cleanup — the cache dir is owned by the
+   * caller's gc policy, not any one render.
+   */
+  ownedByLookup?: boolean;
 }
 
 export interface ExtractionOptions {
@@ -181,15 +197,22 @@ export async function extractVideoFramesRange(
   options: ExtractionOptions,
   signal?: AbortSignal,
   config?: Partial<Pick<EngineConfig, "ffmpegProcessTimeout">>,
+  /**
+   * Override the output directory for this extraction. When provided, frames
+   * are written directly into `outputDirOverride` (no per-videoId subdir).
+   * Used by the cache layer to materialize frames straight into the keyed
+   * cache entry directory.
+   */
+  outputDirOverride?: string,
 ): Promise<ExtractedFrames> {
   const ffmpegProcessTimeout = config?.ffmpegProcessTimeout ?? DEFAULT_CONFIG.ffmpegProcessTimeout;
   const { fps, outputDir, quality = 95, format = "jpg" } = options;
 
-  const videoOutputDir = join(outputDir, videoId);
+  const videoOutputDir = outputDirOverride ?? join(outputDir, videoId);
   if (!existsSync(videoOutputDir)) mkdirSync(videoOutputDir, { recursive: true });
 
   const metadata = await extractMediaMetadata(videoPath);
-  const framePattern = `frame_%05d.${format}`;
+  const framePattern = `${FRAME_FILENAME_PREFIX}%05d.${format}`;
   const outputPattern = join(videoOutputDir, framePattern);
 
   // When extracting from HDR source, tone-map to SDR in FFmpeg rather than
@@ -253,7 +276,7 @@ export async function extractVideoFramesRange(
 
       const framePaths = new Map<number, string>();
       const files = readdirSync(videoOutputDir)
-        .filter((f) => f.startsWith("frame_") && f.endsWith(`.${format}`))
+        .filter((f) => f.startsWith(FRAME_FILENAME_PREFIX) && f.endsWith(`.${format}`))
         .sort();
       files.forEach((file, index) => {
         framePaths.set(index, join(videoOutputDir, file));
@@ -353,6 +376,21 @@ async function convertSdrToHdr(
   }
 }
 
+/**
+ * Resolve the used-segment duration for a video, falling back to the source's
+ * natural duration when the caller hasn't specified bounds (end=Infinity) or
+ * the bounds are nonsensical (end<=start).
+ */
+function resolveSegmentDuration(
+  requested: number,
+  mediaStart: number,
+  metadata: VideoMetadata,
+): number {
+  if (Number.isFinite(requested) && requested > 0) return requested;
+  const sourceRemaining = metadata.durationSeconds - mediaStart;
+  return sourceRemaining > 0 ? sourceRemaining : metadata.durationSeconds;
+}
+
 /**
  * Re-encode a VFR (variable frame rate) video segment to CFR so the downstream
  * fps filter can extract frames reliably. Screen recordings, phone videos, and
@@ -414,7 +452,7 @@ export async function extractAllVideoFrames(
   baseDir: string,
   options: ExtractionOptions,
   signal?: AbortSignal,
-  config?: Partial<Pick<EngineConfig, "ffmpegProcessTimeout">>,
+  config?: Partial<Pick<EngineConfig, "ffmpegProcessTimeout" | "extractCacheDir">>,
   compiledDir?: string,
 ): Promise<ExtractionResult> {
   const startTime = Date.now();
@@ -469,6 +507,28 @@ export async function extractAllVideoFrames(
 
   breakdown.resolveMs = Date.now() - phase1Start;
 
+  // Snapshot the pre-preflight key inputs so the extraction cache keys on the
+  // user-visible source (original path, original mediaStart, original segment
+  // bounds) rather than the workDir-local normalized file produced by
+  // Phase 2a/2b preflight. Without this, every render would write a new
+  // normalized file with a fresh mtime → fresh cache key → perpetual misses.
+  const cacheKeyInputs = resolvedVideos.map(({ video, videoPath }) => {
+    const stat = readKeyStat(videoPath);
+    // Missing files return null — skip the cache path for that entry. The
+    // extractor will surface the real file-not-found error downstream, and we
+    // avoid polluting the cache with a `(mtimeMs: 0, size: 0)` tuple that two
+    // unrelated missing paths would otherwise share.
+    if (!stat) return null;
+    return {
+      videoPath,
+      mtimeMs: stat.mtimeMs,
+      size: stat.size,
+      mediaStart: video.mediaStart,
+      start: video.start,
+      end: video.end,
+    };
+  });
+
   // Phase 2: Probe color spaces and normalize if mixed HDR/SDR
   const phase2ProbeStart = Date.now();
   const videoMetadata = await Promise.all(
@@ -565,6 +625,10 @@ export async function extractAllVideoFrames(
         resolvedVideos.splice(i, 1);
         videoMetadata.splice(i, 1);
         videoColorSpaces.splice(i, 1);
+        // Added by the extraction-cache commit: keep cacheKeyInputs aligned
+        // with the other parallel arrays so Phase 3's `cacheKeyInputs[i]`
+        // lookup doesn't point at a stale slot after the splice.
+        cacheKeyInputs.splice(i, 1);
       }
     }
   }
@@ -615,25 +679,85 @@ export async function extractAllVideoFrames(
   }
   breakdown.vfrPreflightMs = Date.now() - vfrPreflightStart;
 
-  // Phase 3: Extract frames (parallel)
   const phase3Start = Date.now();
+  const cacheRootDir = config?.extractCacheDir;
+  const cacheFormat: CacheFrameFormat = options.format ?? "jpg";
+
+  async function tryCachedExtract(
+    video: VideoElement,
+    videoPath: string,
+    videoDuration: number,
+    i: number,
+  ): Promise<ExtractedFrames | null> {
+    if (!cacheRootDir) return null;
+    const keyInput = cacheKeyInputs[i];
+    const probedMeta = videoMetadata[i];
+    if (!keyInput || !probedMeta) return null;
+
+    const keyDuration = resolveSegmentDuration(
+      keyInput.end - keyInput.start,
+      keyInput.mediaStart,
+      probedMeta,
+    );
+    const lookup = lookupCacheEntry(cacheRootDir, {
+      videoPath: keyInput.videoPath,
+      mtimeMs: keyInput.mtimeMs,
+      size: keyInput.size,
+      mediaStart: keyInput.mediaStart,
+      duration: keyDuration,
+      fps: options.fps,
+      format: cacheFormat,
+    });
+
+    if (lookup.hit) {
+      breakdown.cacheHits += 1;
+      const rehydrated = rehydrateCacheEntry(lookup.entry, {
+        videoId: video.id,
+        srcPath: keyInput.videoPath,
+        fps: options.fps,
+        format: cacheFormat,
+        metadata: probedMeta,
+      });
+      return { ...rehydrated, ownedByLookup: true };
+    }
+
+    breakdown.cacheMisses += 1;
+    ensureCacheEntryDir(lookup.entry);
+    const result = await extractVideoFramesRange(
+      videoPath,
+      video.id,
+      video.mediaStart,
+      videoDuration,
+      options,
+      signal,
+      config,
+      lookup.entry.dir,
+    );
+    // Mark complete only AFTER frames are on disk — a crash mid-extract
+    // leaves the entry un-sentineled so the next lookup re-extracts over it.
+    markCacheEntryComplete(lookup.entry);
+    return { ...result, ownedByLookup: true };
+  }
+
   const results = await Promise.all(
-    resolvedVideos.map(async ({ video, videoPath }) => {
+    resolvedVideos.map(async ({ video, videoPath }, i) => {
       if (signal?.aborted) {
         throw new Error("Video frame extraction cancelled");
       }
       try {
-        let videoDuration = video.end - video.start;
-
-        // Fallback: if no data-duration/data-end was specified (end is Infinity or 0),
-        // probe the actual video file to get its natural duration.
-        if (!Number.isFinite(videoDuration) || videoDuration <= 0) {
-          const metadata = await extractMediaMetadata(videoPath);
-          const sourceDuration = metadata.durationSeconds - video.mediaStart;
-          videoDuration = sourceDuration > 0 ? sourceDuration : metadata.durationSeconds;
+        const probedMeta = videoMetadata[i] ?? (await extractMediaMetadata(videoPath));
+        const videoDuration = resolveSegmentDuration(
+          video.end - video.start,
+          video.mediaStart,
+          probedMeta,
+        );
+        if (video.end - video.start !== videoDuration) {
           video.end = video.start + videoDuration;
         }
 
+        const cached = await tryCachedExtract(video, videoPath, videoDuration, i);
+        if (cached) return { result: cached };
+
         const result = await extractVideoFramesRange(
           videoPath,
           video.id,
@@ -800,6 +924,10 @@ export class FrameLookupTable {
 
   cleanup(): void {
     for (const video of this.videos.values()) {
+      // Cache-hit / cache-write entries are owned by the extraction cache —
+      // a single render must not delete them, or the next render's lookup
+      // would miss and re-extract unnecessarily.
+      if (video.extracted.ownedByLookup) continue;
       if (existsSync(video.extracted.outputDir)) {
         rmSync(video.extracted.outputDir, { recursive: true, force: true });
       }
diff --git a/packages/producer/src/services/renderOrchestrator.ts b/packages/producer/src/services/renderOrchestrator.ts
index 115ae13c..64cd5b02 100644
--- a/packages/producer/src/services/renderOrchestrator.ts
+++ b/packages/producer/src/services/renderOrchestrator.ts
@@ -1427,7 +1427,7 @@ export async function executeRenderJob(
         projectDir,
         { fps: job.config.fps, outputDir: join(workDir, "video-frames") },
         abortSignal,
-        undefined,
+        { extractCacheDir: cfg.extractCacheDir },
         compiledDir,
       );
       assertNotAborted();