2.26.2: transformers prompt cache crops with the negative form (transformers 5.20)

transformers 5.20 removed Cache.crop(n) with a positive n (keep n tokens,
deprecated since 5.x); the call failed, _transformers_crop_cache returned
False, and every warm full-context call rebuilt the cache from scratch
(test_full_context_warm_call_feeds_only_the_suffix: 249 then 258 tokens
fed). Crop with keep - current (drop that many tokens), which every
supported transformers version accepts; the test fake follows the real
semantics. Passes on transformers 5.17.0, 5.19.0 and 5.20.1.

The 2.26.1 release run failed on this test (nothing was published), so
the Windows fixes of 2.26.1 ship in 2.26.2.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Author: lpalbou

CHANGELOG.md

@@ -5,6 +5,18 @@ All notable changes to AbstractCore will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [2.26.2] - 2026-10-18
+
+2.26.1 was not published on PyPI; its Windows fixes ship in this release.
+
+### Fixed
+
+- **Hugging Face (transformers) prompt cache with transformers 5.20 or newer**: a warm call
+  feeds only the new part of the conversation again. transformers 5.20 removed `Cache.crop`
+  with a positive length; the cache is now cropped with the negative form (drop that many
+  tokens), which every supported transformers version accepts. With 5.20, every warm call
+  re-read the whole conversation (correct output, no prefill savings).
+
 ## [2.26.1] - 2026-10-17
 
 Windows fixes.

abstractcore/providers/huggingface_provider.py

@@ -4531,8 +4531,19 @@ def _transformers_crop_cache(self, state: _TransformersPromptCacheValue, keep_to
                 )
             return False
 
+        # `crop(n)` with a positive `n` (keep n tokens) was deprecated and is gone in
+        # transformers 5.20; a negative `n` (drop |n| tokens) works on every version.
+        keep = int(keep_tokens)
         try:
-            crop(int(keep_tokens))
+            lengths = self._transformers_layer_seq_lengths(cache)
+        except Exception:
+            lengths = []
+        current = max(lengths) if lengths else None
+        try:
+            if current is None:
+                crop(keep)  # length unknown: the legacy form is the only one available
+            elif current > keep:
+                crop(keep - current)
         except Exception:
             return False
 

abstractcore/utils/version.py

@@ -10,4 +10,4 @@
 """
 
 # Package version - update this when releasing new versions
-__version__ = "2.26.1"
+__version__ = "2.26.2"

llms-full.txt

@@ -3,7 +3,7 @@
 > Expanded companion to `llms.txt`: setup, API and provider guidance, followed by canonical documentation pages. Use the linked source pages for authoritative documentation and the source code for behavior.
 
 Last updated: 2026-09-30
-Package version: 2.26.1
+Package version: 2.26.2
 Supported Python: 3.9+; GitHub CI tests 3.9, 3.10, 3.11, 3.12, and 3.13.
 
 ## How to use this file (agents)

tests/huggingface/test_transformers_cached_full_context_delta.py

@@ -236,7 +236,10 @@ class _ExactCropCache:
         def __init__(self, length):
             self._len = length
         def crop(self, n):
-            self._len = min(self._len, int(n))
+            # transformers' semantics: a negative n drops |n| tokens (the only form
+            # transformers >= 5.20 accepts); a positive n keeps n tokens (older ones).
+            n = int(n)
+            self._len = max(0, self._len + n) if n < 0 else min(self._len, n)
         def get_seq_length(self):
             return self._len