Merge pull request #833 from yasinBursali/fix/reasoning-token-budget

fix: disable reasoning by default to prevent empty chat responses

Author: Lightheartdevs

dream-server/.env.schema.json

@@ -228,6 +228,21 @@
       "description": "Dashboard API authentication key",
       "secret": true
     },
+    "DREAM_AGENT_KEY": {
+      "type": "string",
+      "description": "API key for the Dream Host Agent (falls back to DASHBOARD_API_KEY if unset)",
+      "secret": true
+    },
+    "APE_API_KEY": {
+      "type": "string",
+      "description": "API key for the APE (AI Proxy Engine) extension",
+      "secret": true
+    },
+    "SHIELD_API_KEY": {
+      "type": "string",
+      "description": "API key for the Privacy Shield extension",
+      "secret": true
+    },
     "OPENCODE_SERVER_PASSWORD": {
       "type": "string",
       "description": "OpenCode web UI authentication password",
@@ -454,6 +469,12 @@
       "type": "string",
       "description": "GPU UUIDs assigned to llama-server (comma-separated, used by NVIDIA_VISIBLE_DEVICES)"
     },
+    "LLAMA_REASONING": {
+      "type": "string",
+      "description": "llama.cpp reasoning/thinking mode: off (default) | auto | on. Off prevents thinking models from consuming the entire token budget on internal reasoning.",
+      "enum": ["off", "auto", "on"],
+      "default": "off"
+    },
     "LLAMA_ARG_SPLIT_MODE": {
       "type": "string",
       "description": "llama.cpp split mode (LLAMA_ARG_SPLIT_MODE): none | layer (pipeline) | row (tensor/hybrid)"

dream-server/docker-compose.base.yml

@@ -22,6 +22,7 @@ services:
   # Serves OpenAI-compatible API on port 8080
   # ============================================
   llama-server:
+    image: ${LLAMA_SERVER_IMAGE:-ghcr.io/ggml-org/llama.cpp:server-b8248}
     container_name: dream-llama-server
     restart: unless-stopped
     volumes:
@@ -47,6 +48,8 @@ services:
       - --parallel
       - "${LLAMA_PARALLEL:-1}"
       - --metrics
+    environment:
+      - LLAMA_ARG_REASONING=${LLAMA_REASONING:-off}
     security_opt:
       - no-new-privileges:true
     logging: *default-logging

dream-server/installers/macos/dream-macos.sh

@@ -148,11 +148,21 @@ start_native_llama() {
 
     mkdir -p "$(dirname "$LLAMA_SERVER_PID_FILE")"
 
+    local reasoning="${ENV_LLAMA_REASONING:-off}"
+    # Map .env values (off/on/auto) to llama-server --reasoning-format values
+    local reasoning_fmt
+    case "$reasoning" in
+        off)  reasoning_fmt="none" ;;
+        on)   reasoning_fmt="deepseek" ;;
+        *)    reasoning_fmt="$reasoning" ;;
+    esac
+
     "$LLAMA_SERVER_BIN" \
         --host 0.0.0.0 --port 8080 \
         --model "$model_path" \
         --ctx-size "$ctx_size" \
         --n-gpu-layers 999 \
+        --reasoning-format "$reasoning_fmt" \
         --metrics \
         > "$LLAMA_SERVER_LOG" 2>&1 &
     local pid=$!

dream-server/installers/macos/install-macos.sh

@@ -545,11 +545,23 @@ else
             fi
         fi
 
+        # Read reasoning mode from .env (default off to prevent thinking models
+        # from consuming the entire token budget on internal reasoning)
+        _reasoning=$(grep '^LLAMA_REASONING=' "$INSTALL_DIR/.env" 2>/dev/null | cut -d= -f2 || echo "")
+        [[ -z "$_reasoning" ]] && _reasoning="off"
+        # Map .env values (off/on/auto) to llama-server --reasoning-format values
+        case "$_reasoning" in
+            off)  _reasoning_fmt="none" ;;
+            on)   _reasoning_fmt="deepseek" ;;
+            *)    _reasoning_fmt="$_reasoning" ;;
+        esac
+
         "$LLAMA_SERVER_BIN" \
             --host 0.0.0.0 --port 8080 \
             --model "$MODEL_FULL_PATH" \
             --ctx-size "$MAX_CONTEXT" \
             --n-gpu-layers 999 \
+            --reasoning-format "$_reasoning_fmt" \
             --metrics \
             > "$LLAMA_SERVER_LOG" 2>&1 &
         LLAMA_PID=$!

dream-server/scripts/bootstrap-upgrade.sh

@@ -176,16 +176,25 @@ else
 fi
 
 # ── Phase 2: Verify integrity (if SHA256 provided) ──
-if [[ -n "$FULL_GGUF_SHA256" ]] && command -v sha256sum &>/dev/null; then
+if [[ -n "$FULL_GGUF_SHA256" ]]; then
     write_status "verifying" 100 "$TOTAL_BYTES" "$TOTAL_BYTES" 0 ""
     log "Verifying SHA256..."
-    ACTUAL_HASH=$(sha256sum "$MODELS_DIR/$FULL_GGUF_FILE" 2>/dev/null | awk '{print $1}')
-    if [[ "$ACTUAL_HASH" != "$FULL_GGUF_SHA256" ]]; then
-        rm -f "$MODELS_DIR/$FULL_GGUF_FILE"
-        write_status "failed"
-        fail "SHA256 mismatch (expected: $FULL_GGUF_SHA256, got: $ACTUAL_HASH). Deleted corrupt file."
+    if command -v sha256sum &>/dev/null; then
+        ACTUAL_HASH=$(sha256sum "$MODELS_DIR/$FULL_GGUF_FILE" 2>/dev/null | awk '{print $1}')
+    elif command -v shasum &>/dev/null; then
+        ACTUAL_HASH=$(shasum -a 256 "$MODELS_DIR/$FULL_GGUF_FILE" 2>/dev/null | awk '{print $1}')
+    else
+        log "WARNING: No checksum tool available — skipping SHA256 verification"
+        ACTUAL_HASH=""
+    fi
+    if [[ -n "$ACTUAL_HASH" ]]; then
+        if [[ "$ACTUAL_HASH" != "$FULL_GGUF_SHA256" ]]; then
+            rm -f "$MODELS_DIR/$FULL_GGUF_FILE"
+            write_status "failed"
+            fail "SHA256 mismatch (expected: $FULL_GGUF_SHA256, got: $ACTUAL_HASH). Deleted corrupt file."
+        fi
+        log "SHA256 verified"
     fi
-    log "SHA256 verified"
     write_status "complete"
 fi
 
@@ -327,13 +336,24 @@ elif [[ -f "$INSTALL_DIR/data/.llama-server.pid" ]]; then
                 fi
             fi
 
+            # Read reasoning mode from .env (default off to prevent thinking models
+            # from consuming the entire token budget on internal reasoning)
+            _reasoning=$(grep '^LLAMA_REASONING=' "$ENV_FILE" 2>/dev/null | cut -d= -f2 || echo "")
+            [[ -z "$_reasoning" ]] && _reasoning="off"
+            case "$_reasoning" in
+                off)  _reasoning_fmt="none" ;;
+                on)   _reasoning_fmt="deepseek" ;;
+                *)    _reasoning_fmt="$_reasoning" ;;
+            esac
+
             # Relaunch with new model
             log "Starting native llama-server with ${_gguf_file}..."
             "$LLAMA_SERVER_BIN" \
                 --host 0.0.0.0 --port 8080 \
                 --model "$_model_path" \
                 --ctx-size "$_ctx_size" \
                 --n-gpu-layers 999 \
+                --reasoning-format "$_reasoning_fmt" \
                 --metrics \
                 > "$LLAMA_SERVER_LOG" 2>&1 &
             _new_pid=$!
@@ -365,6 +385,7 @@ elif [[ -f "$INSTALL_DIR/data/.llama-server.pid" ]]; then
                         --model "$_old_model_path" \
                         --ctx-size "$_ctx_size" \
                         --n-gpu-layers 999 \
+                        --reasoning-format "${_reasoning_fmt:-none}" \
                         --metrics \
                         > "$LLAMA_SERVER_LOG" 2>&1 &
                     _rollback_pid=$!