--- a/python/src/captchakraken/pointer/service.py
+++ b/python/src/captchakraken/pointer/service.py
@@ -29,6 +29,8 @@
 from . import vllm_plugin
 
 
+VLLM_VIDEO = "<|vision_start|><|video_pad|><|vision_end|>"
+
 class PointerService:
     def __init__(self, checkpoint: Union[str, Path], detector: Union[str, Path, None] = None,
                  gpu_memory_utilization: float = 0.45, max_model_len: int = 8192, **llm_kwargs):
@@ -41,7 +43,10 @@
         self.llm = LLM(model=str(checkpoint), runner="pooling", max_model_len=max_model_len,
                        gpu_memory_utilization=gpu_memory_utilization,
                        limit_mm_per_prompt={"image": 1 + L.MAX_HISTORY, "video": 1},
-                       mm_processor_kwargs={"min_pixels": 64 * 64, "max_pixels": L.MAX_PIXELS}, **llm_kwargs)
+                       mm_processor_kwargs={"min_pixels": 64 * 64, "max_pixels": L.MAX_PIXELS},
+                       # The heads read every token's hidden state; vLLM returns none for a
+                       # prefix it served from cache, so a repeated prompt came back short.
+                       enable_prefix_caching=False, **llm_kwargs)
         self.detector = None
         if detector is not None:
             from .detector import Detector
@@ -59,7 +64,7 @@
                 marked: bool = True, history: Sequence[Image.Image] = ()) -> Dict:
         """The request vLLM gets, and what decoding its answer needs. `history`:
         the puzzle's earlier screens, oldest first, as they were sent (marked the
-        same way), so the prompt up to them is the one already cached."""
+        same way). The prefix is recomputed each time: the heads need its hidden states."""
         if (image is None) == (frames is None):
             raise ValueError("give exactly one of image or frames")
         if frames is not None and history:
@@ -78,7 +83,12 @@
                 if tracks:
                     fr = np.stack([np.asarray(M.draw(im, E.on_frame(tracks, k)))
                                    for k, im in enumerate(ims)])
-            return {"request": {"prompt": L.chat(self.processor, True),
+            # vLLM (0.25.1) expands a video only where it finds the template's whole
+            # <|vision_start|><|video_pad|><|vision_end|>; `chat` sends transformers
+            # the bare pad. Both then give the same tokens but one: training merges the
+            # instruction's last "." with the first timestamp's "<", vLLM splits them.
+            prompt = L.chat(self.processor, True).replace(L.VIDEO_PAD, VLLM_VIDEO)
+            return {"request": {"prompt": prompt,
                                 "multi_modal_data": {"video": (fr, meta)},
                                 "mm_processor_kwargs": {"do_sample_frames": False}},
                     "size": (w, h), "elements": tracks, "shown": None,
--- a/python/src/captchakraken/pointer/vllm_plugin.py
+++ b/python/src/captchakraken/pointer/vllm_plugin.py
@@ -65,6 +65,9 @@
                     # A real prompt without its markers still fails in Positions.
                     out.append(torch.zeros(L.PACKED + TRAILER, device=h.device))
                     continue
+                if h.shape[0] != ids.shape[0]:
+                    raise RuntimeError(f"{h.shape[0]} hidden states for a {ids.shape[0]}-token prompt: "
+                                       "the heads need them all (is prefix caching on?)")
                 pos = Positions(ids, m.pointer_token_ids)
                 v = m.pointer_heads.decide(h, pos, m.language_model.lm_head.weight)
                 out.append(torch.cat([v, torch.tensor([float(pos.per_moment), float(pos.n_moments)],