feat(llm-services): make the thinking budget a first-class knob

llama.cpp leaves --reasoning-budget unrestricted. A memory prompt already
reaches ~15K of the served window, so an unbounded thinking trace competes with
the transcript and tool results the agent still has to read.

LLAMA_ARG_THINK_BUDGET joins the other server knobs: llama.cpp's own -1 in the
default profile, bounded at 2000 for the Qwen text-agent profile, and surfaced
in the setup summary beside flash attention and KV cache.

This is the layout these settings already have — env vars written by init.py and
read by compose. It existed for flash attention, KV cache type, parallel slots
and the vision projector, but not for this one, which is why a node that needed
it had to hand-edit its committed docker-compose.yml instead.

Author: AnkushMalaker

extras/llm-services/docker-compose.yml

@@ -30,6 +30,9 @@ services:
       LLAMA_ARG_CACHE_TYPE_K: ${LLAMA_ARG_CACHE_TYPE_K:-f16}
       LLAMA_ARG_CACHE_TYPE_V: ${LLAMA_ARG_CACHE_TYPE_V:-f16}
       LLAMA_ARG_MMPROJ_AUTO: ${LLAMA_ARG_MMPROJ_AUTO:-true}
+      # `--reasoning-budget`: -1 is llama.cpp's unrestricted default. Bound it for a
+      # thinking model so a runaway trace cannot eat the context the caller needs.
+      LLAMA_ARG_THINK_BUDGET: ${LLAMA_ARG_THINK_BUDGET:--1}
       LLAMA_ARG_HOST: "0.0.0.0"
     entrypoint:
       - /bin/sh

extras/llm-services/init.py

@@ -86,6 +86,7 @@
     "LLAMA_ARG_CACHE_TYPE_K": "f16",
     "LLAMA_ARG_CACHE_TYPE_V": "f16",
     "LLAMA_ARG_MMPROJ_AUTO": "true",
+    "LLAMA_ARG_THINK_BUDGET": "-1",
 }
 QWEN_TEXT_SERVER_PROFILE = {
     **DEFAULT_SERVER_PROFILE,
@@ -96,6 +97,10 @@
     # `-hf` otherwise auto-downloads and loads Qwen's ~0.93GB vision projector.
     # Chronicle's local memory service is text-only.
     "LLAMA_ARG_MMPROJ_AUTO": "false",
+    # Thinking is unrestricted by default, and a memory prompt already reaches ~15K
+    # of the served window. Cap the trace so an unbounded one cannot crowd out the
+    # transcript and tool results the agent still has to read.
+    "LLAMA_ARG_THINK_BUDGET": "2000",
 }
 
 # Embedding model options
@@ -611,6 +616,11 @@ def show_summary(self, llm_info, embed_info):
             f"K={self.config.get('LLAMA_ARG_CACHE_TYPE_K', 'f16')}, "
             f"V={self.config.get('LLAMA_ARG_CACHE_TYPE_V', 'f16')}",
         )
+        think_budget = str(self.config.get("LLAMA_ARG_THINK_BUDGET", "-1"))
+        table.add_row(
+            "Thinking Budget",
+            "unrestricted" if think_budget == "-1" else f"{think_budget} tokens",
+        )
         table.add_row(
             "Vision Projector",
             (

extras/llm-services/tests/test_init.py

@@ -403,6 +403,8 @@ def test_qwen_uses_a30_text_agent_server_profile(init_module):
         "LLAMA_ARG_CACHE_TYPE_K": "q8_0",
         "LLAMA_ARG_CACHE_TYPE_V": "q8_0",
         "LLAMA_ARG_MMPROJ_AUTO": "false",
+        # Bounded so a runaway thinking trace cannot crowd the served window.
+        "LLAMA_ARG_THINK_BUDGET": "2000",
     }
 
 
@@ -420,6 +422,8 @@ def test_non_qwen_profile_restores_safe_llamacpp_defaults(init_module):
         "LLAMA_ARG_CACHE_TYPE_K": "f16",
         "LLAMA_ARG_CACHE_TYPE_V": "f16",
         "LLAMA_ARG_MMPROJ_AUTO": "true",
+        # llama.cpp's own default: unrestricted.
+        "LLAMA_ARG_THINK_BUDGET": "-1",
     }