From 8b7b074df9506d512b80ab6855f9773041314e0e Mon Sep 17 00:00:00 2001
From: Teknium <127238744+teknium1@users.noreply.github.com>
Date: Thu, 30 Apr 2026 19:59:36 -0700
Subject: [PATCH] test(context_compressor): regression test for PR #17025
 tail-protection off-by-one

When len(messages) <= protect_tail_count and a token budget is set, the
previous formula min(protect_tail_count, len(result) - 1) under-protected
the tail by one, allowing the oldest message to be summarized.

The test fails on the buggy formula (pruned == 1) and passes on the fix
(pruned == 0, tool content preserved verbatim).
---
 tests/agent/test_context_compressor.py | 28 ++++++++++++++++++++++++++
 1 file changed, 28 insertions(+)

diff --git a/tests/agent/test_context_compressor.py b/tests/agent/test_context_compressor.py
index 8f5d6c4d1..8d1de377b 100644
--- a/tests/agent/test_context_compressor.py
+++ b/tests/agent/test_context_compressor.py
@@ -1143,6 +1143,34 @@ class TestTokenBudgetTailProtection:
         # At least one old tool result should have been pruned
         assert pruned >= 1
 
+    def test_prune_short_conv_protects_entire_tail(self, budget_compressor):
+        """Regression guard for PR #17025.
+
+        When ``len(messages) <= protect_tail_count`` and a token budget is
+        also set, every message must be protected. The previous code used
+        ``min(protect_tail_count, len(result) - 1)`` which capped the floor
+        one below the full length, leaving the oldest message eligible for
+        pruning.
+        """
+        c = budget_compressor
+        # 4 messages, protect_tail_count=4 -- nothing should be pruned.
+        # Oldest message is a large tool result; on the buggy path it falls
+        # outside the protected window and gets summarized.
+        messages = [
+            {"role": "tool", "content": "x" * 5000, "tool_call_id": "c0"},
+            {"role": "assistant", "content": "ack"},
+            {"role": "user", "content": "recent"},
+            {"role": "assistant", "content": "reply"},
+        ]
+        result, pruned = c._prune_old_tool_results(
+            messages,
+            protect_tail_count=4,
+            protect_tail_tokens=1_000_000,  # budget large enough to protect all
+        )
+        assert pruned == 0
+        # Tool result at index 0 must be preserved verbatim
+        assert result[0]["content"] == "x" * 5000
+
     def test_prune_without_token_budget_uses_message_count(self, budget_compressor):
         """Without protect_tail_tokens, falls back to message-count behavior."""
         c = budget_compressor