From 86d020f99cf0b9f918e421044191c60017d30a5a Mon Sep 17 00:00:00 2001
From: ARC <261443955+arc-uri-el@users.noreply.github.com>
Date: Tue, 8 Sep 2026 12:54:31 +0200
Subject: [PATCH 1/2] server : retain replaced branches with a shared prompt
 prefix

Preserve and consult the existing RAM prompt cache when any cached suffix is replaced. The half-prefix-loss heuristic drops independent conversations that share a large system or tool prefix. Extend the existing slot tests with two interleaved shared-prefix branches and exact restored token checks.

Assisted-by: GPT 6 Astra
---
 tools/server/server-context.cpp           |  4 +--
 tools/server/tests/unit/test_slot_save.py | 30 +++++++++++++++++++++++
 2 files changed, 32 insertions(+), 2 deletions(-)

diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp
index 7d636c7923b8..af28bacdea62 100644
--- a/tools/server/server-context.cpp
+++ b/tools/server/server-context.cpp
@@ -1808,8 +1808,8 @@ struct server_context_impl {
                             f_sim_best, slot_prompt_similarity, f_keep);
                 }
 
-                // if we are about to lose a large portion of the existing context - save it in the prompt cache
-                if (f_keep < 0.5f) {
+                // A replaced branch can share most of a tool schema and still need its own cached suffix.
+                if (ret->prompt.tokens.get_common_prefix(task.tokens) < ret->prompt.tokens.size()) {
                     update_cache = true;
                 }
             }
diff --git a/tools/server/tests/unit/test_slot_save.py b/tools/server/tests/unit/test_slot_save.py
index 5eca46cb292d..c8d95aa47144 100644
--- a/tools/server/tests/unit/test_slot_save.py
+++ b/tools/server/tests/unit/test_slot_save.py
@@ -158,6 +158,36 @@ def test_slot_erase():
     assert res.body["timings"]["prompt_n"] == 21  # all tokens are processed
 
 
+def test_ram_cache_interleaved_shared_prefix():
+    server.n_slots = 1
+    server.cache_ram = 16
+    server.n_predict = 8
+    server.start()
+
+    prefix = [1] + [10] * 192
+    prompts = [prefix + [20] * 48, prefix + [30] * 48]
+
+    def complete(prompt):
+        res = server.make_request("POST", "/completion", data={
+            "prompt": prompt,
+            "cache_prompt": True,
+            "n_predict": 8,
+            "ignore_eos": True,
+            "return_tokens": True,
+            "temperature": 0.0,
+        })
+        assert res.status_code == 200
+        return res.body
+
+    first = [complete(prompt) for prompt in prompts]
+    assert first[0]["timings"]["prompt_n"] == len(prompts[0])
+    for prompt, expected in zip(prompts, first):
+        restored = complete(prompt)
+        assert restored["timings"]["cache_n"] == len(prompt) - 1
+        assert restored["timings"]["prompt_n"] == 1
+        assert restored["tokens"] == expected["tokens"]
+
+
 #
 # Multimodal server (mmproj loaded) slot save/restore.
 #

From 9cbefdcf068ca21cf57d81cc153bdcede67ea510 Mon Sep 17 00:00:00 2001
From: ARC <261443955+arc-uri-el@users.noreply.github.com>
Date: Tue, 8 Sep 2026 13:57:18 +0200
Subject: [PATCH 2/2] tests : compare branch restoration against the same
 replay shape

Take a one-token replay control before displacement. A full-prefill comparison also changes kernel shape and is not an isolation test for RAM restoration.

Assisted-by: GPT 6 Astra
---
 tools/server/tests/unit/test_slot_save.py | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/tools/server/tests/unit/test_slot_save.py b/tools/server/tests/unit/test_slot_save.py
index c8d95aa47144..6e07bdc43c5c 100644
--- a/tools/server/tests/unit/test_slot_save.py
+++ b/tools/server/tests/unit/test_slot_save.py
@@ -179,8 +179,11 @@ def complete(prompt):
         assert res.status_code == 200
         return res.body
 
-    first = [complete(prompt) for prompt in prompts]
-    assert first[0]["timings"]["prompt_n"] == len(prompts[0])
+    first = []
+    for prompt in prompts:
+        complete(prompt)
+        # Compare the same one-token replay shape before and after displacement.
+        first.append(complete(prompt))
     for prompt, expected in zip(prompts, first):
         restored = complete(prompt)
         assert restored["timings"]["cache_n"] == len(prompt) - 1