slots und kontextfenster angepasst

client/continue/config.yaml

@@ -29,7 +29,7 @@ models:
     roles: [chat, edit, apply, summarize]
     capabilities: [tool_use]
     defaultCompletionOptions:
-      contextLength: 65536
+      contextLength: 122880
       maxTokens: 8192
 
   - name: GLM-4.7 Flash
@@ -38,7 +38,7 @@ models:
     roles: [chat, edit, apply]
     capabilities: [tool_use]
     defaultCompletionOptions:
-      contextLength: 65536
+      contextLength: 122880
       maxTokens: 8192
 
   - name: Qwen3.5 9B (schnell)
@@ -47,7 +47,7 @@ models:
     roles: [chat, edit, apply]
     capabilities: [tool_use]
     defaultCompletionOptions:
-      contextLength: 32768
+      contextLength: 28672
       maxTokens: 4096
 
   # --- Tab-Autocomplete (Fill-in-the-Middle) ------------------------------

client/opencode.json

@@ -18,17 +18,17 @@
         "qwen3.5-35b-a3b": {
           "name": "Qwen3.5 35B-A3B",
           "tool_call": true,
-          "limit": { "context": 65536, "output": 16384 }
+          "limit": { "context": 122880, "output": 16384 }
         },
         "glm-4.7-flash": {
           "name": "GLM-4.7 Flash",
           "tool_call": true,
-          "limit": { "context": 65536, "output": 16384 }
+          "limit": { "context": 122880, "output": 16384 }
         },
         "qwen3.5-9b": {
           "name": "Qwen3.5 9B (schnell)",
           "tool_call": true,
-          "limit": { "context": 32768, "output": 8192 }
+          "limit": { "context": 28672, "output": 8192 }
         }
       }
     }

llama-swap/config.yaml

@@ -23,8 +23,9 @@ macros:
   server: >-
     /app/llama-server --host 127.0.0.1 --port ${PORT}
     --jinja -fa on --threads 12 --no-webui
-  # Agent-Betrieb: großer Kontext, KV-Cache quantisiert, VRAM-Reserve für Autocomplete
-  agent-ctx: -c 65536 -ctk q8_0 -ctv q8_0 --fit-target 3072
+  # Agent-Betrieb: großer Kontext, KV-Cache quantisiert, VRAM-Reserve für Autocomplete;
+  # -np 1 = eine Sitzung gleichzeitig, sie bekommt den ganzen Kontext (sonst teilen sich 4 Slots)
+  agent-ctx: -c 131072 -np 1 -ctk q8_0 -ctv q8_0 --fit-target 3072
 
 models:
   # --- Chat / Agent (opencode, Continue Chat/Agent/Edit) -----------------
@@ -45,7 +46,7 @@ models:
   qwen3.5-9b:
     name: Qwen3.5 9B (schnell, komplett im VRAM)
     cmd: >-
-      ${server} -c 32768 -ctk q8_0 -ctv q8_0 --fit-target 3072
+      ${server} -c 32768 -np 1 -ctk q8_0 -ctv q8_0 --fit-target 3072
       -hf unsloth/Qwen3.5-9B-GGUF:Q4_K_M
     ttl: 3600