fix: address code review findings — batch args, venv path, serve flags

- Fix missing BATCH_ARGS in long-context commands (both benchmark scripts)
- Fix CLAUDE.md stale venv path (data/venv → .venv) and add serve/power docs
- Add -b/--batch to bin/benchmark help text
- Add --no-think flag to serve script (--reasoning-budget 0)
- Sanitize model names in eval run directories
- Simplify agentic setup to use requirements.txt
- Add serve --help test, batch flag assertions to existing tests
- Add requirements.txt for reproducible venv setup (Python 3.13)

.gitignore

@@ -1,4 +1,5 @@
 data/
+.venv/
 *.log
 *.csv
 *.tmp

CLAUDE.md

@@ -41,9 +41,21 @@ make verify         # 9-point optimization checklist
 bin/audit --json | python3 -m json.tool   # Verify JSON output is valid
 ```
 
+## Serving
+
+`scripts/serve/launch.sh` with dispatcher at `bin/serve`. Launches llama-server inside toolbox containers with optimized defaults: Vulkan RADV, q4_0 KV cache, flash attention, no-mmap, full GPU offload. Key flags:
+- `--ngram` — n-gram speculative decoding (~1.1-1.4x tg for repetitive content)
+- `--no-think` — disables thinking/reasoning via `--reasoning-budget 0` (faster for evals)
+- `--ctx N` — context size (default 131072)
+- `--parallel N` — concurrent request slots
+
+## System Tuning
+
+`scripts/optimize/power-profile.sh` applies Phase 2 optimizations: RyzenAdj PPT increase (85W target, HP caps at 70W sustained), sysctl tuning (vm.swappiness=1, vm.max_map_count=500000), THP=always, RADV_PERFTEST=nogttspill. Systemd services for boot/resume persistence at `configs/ryzenadj-llm.service` and `configs/ryzenadj-resume.service`.
+
 ## Agentic Evaluation
 
-Scripts in `scripts/agentic/` with dispatcher at `bin/agentic`. Uses a Python venv at `data/venv/`. Eval frameworks: inspect-ai (all-in-one), evalplus (HumanEval+/MBPP+), bigcodebench. All target an OpenAI-compatible endpoint (ollama or llama.cpp server). Model catalog at `configs/models.conf`.
+Scripts in `scripts/agentic/` with dispatcher at `bin/agentic`. Uses a Python venv at `.venv/` (Python 3.13, dependencies in `requirements.txt`). Eval frameworks: inspect-ai (all-in-one), inspect-evals (task definitions), evalplus (HumanEval+/MBPP+), bigcodebench. All target an OpenAI-compatible endpoint — auto-detects llama-server (port 8080) or ollama (port 11434). Model catalog at `configs/models.conf`.
 
 ## External Resources
 

bin/benchmark

@@ -23,6 +23,7 @@ case "${1:-help}" in
         echo "  --category LIST     Comma-separated: smoke,dense,moe"
         echo "  --skip-longctx      Skip long-context (32K) tests"
         echo "  --reps N            Standard test repetitions (default: 5)"
+        echo "  -b, --batch N       Batch size (default: 2048, try 256 for MoE)"
         echo "  --kv-types LIST     KV cache sweep (e.g. f16,q8_0,q4_0 or q4_0:q8_0)"
         echo ""
         echo "Examples:"

requirements.txt

@@ -0,0 +1,12 @@
+# Agentic evaluation frameworks
+# Install: python3.13 -m venv .venv && source .venv/bin/activate && pip install -r requirements.txt
+# Requires Python >=3.10, <3.14 (bigcodebench constraint)
+
+inspect-ai>=0.3.201
+inspect-evals>=0.6.0
+evalplus>=0.3.1
+bigcodebench>=0.2.5
+openai>=2.26.0
+
+# IFEval dependency (not on PyPI)
+instruction_following_eval @ git+https://github.com/josejg/instruction_following_eval

scripts/agentic/run-eval.sh

@@ -5,7 +5,7 @@ set -euo pipefail
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 source "$SCRIPT_DIR/../../lib/common.sh"
 
-VENV_DIR="$(data_dir venv)"
+VENV_DIR="$PROJECT_ROOT/.venv"
 EVAL_DIR="$(data_dir evals)"
 
 # ── Argument parsing ─────────────────────────────────────
@@ -37,33 +37,59 @@ while [[ $# -gt 0 ]]; do
 done
 
 # ── Validation ───────────────────────────────────────────
-if [[ -z "$MODEL" ]]; then
-    log_error "Model name required. Use --model NAME"
-    log_info "Examples:"
-    log_info "  --model qwen3.5:35b-a3b-q8_0       (ollama)"
-    log_info "  --model Qwen3.5-35B-A3B-Q8_0        (llama.cpp server)"
-    exit 1
-fi
-
 if [[ ! -f "$VENV_DIR/bin/activate" ]]; then
     log_error "Virtual environment not found. Run: make agentic-setup"
     exit 1
 fi
 source "$VENV_DIR/bin/activate"
 
-# Check server is reachable
+# Auto-detect server if no explicit endpoint given
-if ! curl -sf "$ENDPOINT/models" >/dev/null 2>&1; then
+if [[ "$ENDPOINT" == "http://localhost:11434/v1" ]]; then
-    # Try ollama native endpoint
+    if curl -sf "http://localhost:8080/health" >/dev/null 2>&1; then
-    if curl -sf "http://localhost:11434/api/tags" >/dev/null 2>&1; then
+        ENDPOINT="http://localhost:8080/v1"
-        log_info "Ollama detected, using OpenAI-compat endpoint"
+        log_info "Auto-detected llama-server at localhost:8080"
+    elif curl -sf "http://localhost:11434/api/tags" >/dev/null 2>&1; then
+        log_info "Auto-detected ollama at localhost:11434"
     else
-        log_error "No LLM server at $ENDPOINT. Start ollama or llama.cpp server first."
+        log_error "No LLM server found. Start one first:"
+        log_info "  make serve ARGS=\"-m MODEL.gguf\"        (llama-server)"
+        log_info "  ollama serve                            (ollama)"
+        exit 1
+    fi
+else
+    if ! curl -sf "${ENDPOINT%/v1}/health" >/dev/null 2>&1 && \
+       ! curl -sf "$ENDPOINT/models" >/dev/null 2>&1; then
+        log_error "No LLM server at $ENDPOINT"
+        exit 1
+    fi
+fi
+
+# Auto-detect model name from server if not provided
+if [[ -z "$MODEL" ]]; then
+    DETECTED_MODEL=$(curl -sf "$ENDPOINT/models" 2>/dev/null | python3 -c "
+import sys, json
+try:
+    data = json.load(sys.stdin)
+    models = data.get('data', [])
+    if models:
+        print(models[0].get('id', ''))
+except: pass
+" 2>/dev/null || true)
+    if [[ -n "$DETECTED_MODEL" ]]; then
+        MODEL="$DETECTED_MODEL"
+        log_info "Auto-detected model: $MODEL"
+    else
+        log_error "Model name required. Use --model NAME"
+        log_info "Examples:"
+        log_info "  --model qwen3.5:35b-a3b-q8_0       (ollama)"
+        log_info "  --model Qwen3.5-35B-A3B-Q8_0        (llama.cpp server)"
         exit 1
     fi
 fi
 
 TS="$(timestamp)"
-RUN_DIR="$EVAL_DIR/${SUITE}-${MODEL//[:\/]/_}-${TS}"
+SAFE_MODEL="$(echo "$MODEL" | tr -cs 'a-zA-Z0-9._-' '_')"
+RUN_DIR="$EVAL_DIR/${SUITE}-${SAFE_MODEL}-${TS}"
 mkdir -p "$RUN_DIR"
 
 log_header "Agentic Evaluation: $SUITE"
@@ -86,7 +112,11 @@ ENDJSON
 METRICS_FILE="$RUN_DIR/metrics.csv"
 bash "$SCRIPT_DIR/../monitor/log-metrics.sh" --output "$METRICS_FILE" --interval 5 &
 METRICS_PID=$!
-trap 'kill "$METRICS_PID" 2>/dev/null; wait "$METRICS_PID" 2>/dev/null' EXIT
+cleanup() {
+    kill "$METRICS_PID" 2>/dev/null || true
+    wait "$METRICS_PID" 2>/dev/null || true
+}
+trap 'cleanup; exit 0' EXIT
 
 # ── Suite execution ──────────────────────────────────────
 
@@ -113,14 +143,14 @@ run_evalplus() {
 run_inspect_eval() {
     local eval_name="$1"
     local display_name="$2"
+    local safe_name="${eval_name//\//_}"  # inspect_evals/ifeval → inspect_evals_ifeval
     log_info "Running Inspect AI: $display_name..."
-    local out="$RUN_DIR/inspect-${eval_name}.json"
 
     OPENAI_BASE_URL="$ENDPOINT" OPENAI_API_KEY="not-needed" \
     inspect eval "$eval_name" \
         --model "openai/$MODEL" \
         --log-dir "$RUN_DIR/inspect-logs/" \
-        2>&1 | tee "$RUN_DIR/inspect-${eval_name}.log"
+        2>&1 | tee "$RUN_DIR/inspect-${safe_name}.log"
 
     log_success "Inspect $display_name complete"
 }
@@ -138,7 +168,7 @@ run_bigcodebench() {
 case "$SUITE" in
     quick)
         run_evalplus "humaneval"
-        run_inspect_eval "ifeval" "IFEval (instruction following)"
+        run_inspect_eval "inspect_evals/ifeval" "IFEval (instruction following)"
         ;;
     code)
         run_evalplus "humaneval"
@@ -146,13 +176,13 @@ case "$SUITE" in
         run_bigcodebench
         ;;
     tooluse)
-        run_inspect_eval "bfcl" "BFCL (function calling)"
+        run_inspect_eval "inspect_evals/bfcl" "BFCL (function calling)"
         ;;
     full)
         run_evalplus "humaneval"
         run_evalplus "mbpp"
-        run_inspect_eval "ifeval" "IFEval (instruction following)"
+        run_inspect_eval "inspect_evals/ifeval" "IFEval (instruction following)"
-        run_inspect_eval "bfcl" "BFCL (function calling)"
+        run_inspect_eval "inspect_evals/bfcl" "BFCL (function calling)"
         run_bigcodebench
         ;;
     *)

scripts/agentic/setup.sh

@@ -8,91 +8,56 @@ source "$SCRIPT_DIR/../../lib/common.sh"
 log_header "Agentic Evaluation Setup"
 
 # ── Python virtual environment ───────────────────────────
-VENV_DIR="$(data_dir venv)"
+VENV_DIR="$PROJECT_ROOT/.venv"
+REQUIREMENTS="$PROJECT_ROOT/requirements.txt"
+
 if [[ ! -f "$VENV_DIR/bin/activate" ]]; then
-    log_info "Creating Python virtual environment..."
+    # Prefer Python 3.13 (bigcodebench requires <3.14)
-    python3 -m venv "$VENV_DIR"
+    PYTHON_BIN="python3.13"
+    if ! command -v "$PYTHON_BIN" >/dev/null 2>&1; then
+        PYTHON_BIN="python3"
+        log_warn "python3.13 not found, using $(python3 --version). bigcodebench may not install."
+    fi
+    log_info "Creating virtual environment with $($PYTHON_BIN --version)..."
+    "$PYTHON_BIN" -m venv "$VENV_DIR"
     log_success "Virtual environment created at $VENV_DIR"
 fi
 
 source "$VENV_DIR/bin/activate"
 log_info "Python: $(python3 --version) from $VENV_DIR"
 
-# ── Install evaluation frameworks ────────────────────────
+# ── Install from requirements.txt ────────────────────────
-
+log_info "Installing dependencies from requirements.txt..."
-# Inspect AI — the all-in-one eval framework (bundles BFCL, GAIA, HumanEval, IFEval, etc.)
+pip install -r "$REQUIREMENTS" 2>&1 | tail -5
-if python3 -c "import inspect_ai" 2>/dev/null; then
+log_success "Dependencies installed"
-    log_success "inspect-ai already installed"
-else
-    log_info "Installing inspect-ai (main eval framework)..."
-    pip install inspect-ai 2>&1 | tail -3
-    log_success "inspect-ai installed"
-fi
-
-# EvalPlus — HumanEval+ and MBPP+ with native ollama support
-if python3 -c "import evalplus" 2>/dev/null; then
-    log_success "evalplus already installed"
-else
-    log_info "Installing evalplus (code generation benchmarks)..."
-    pip install evalplus 2>&1 | tail -3
-    log_success "evalplus installed"
-fi
-
-# BigCodeBench
-if python3 -c "import bigcodebench" 2>/dev/null; then
-    log_success "bigcodebench already installed"
-else
-    log_info "Installing bigcodebench..."
-    pip install bigcodebench 2>&1 | tail -3
-    log_success "bigcodebench installed"
-fi
 
 # ── Check for local LLM server ──────────────────────────
 log_header "LLM Server Check"
 
-ollama_ok=false
+if curl -sf http://localhost:8080/health >/dev/null 2>&1; then
-llamacpp_ok=false
+    log_success "llama-server running at localhost:8080"
-
+elif curl -sf http://localhost:11434/api/tags >/dev/null 2>&1; then
-if is_cmd ollama; then
+    log_success "ollama running at localhost:11434"
-    if curl -s http://localhost:11434/api/tags >/dev/null 2>&1; then
-        log_success "ollama running at localhost:11434"
-        ollama_ok=true
-        # List available models
-        log_info "Available ollama models:"
-        ollama list 2>/dev/null | head -10 || true
-    else
-        log_warn "ollama installed but not running. Start with: ollama serve"
-    fi
 else
-    log_info "ollama not installed — needed for most agentic benchmarks"
+    log_warn "No local LLM server running. Start one before running evals:"
-    log_info "Install: curl -fsSL https://ollama.com/install.sh | sh"
+    log_info "  make serve ARGS=\"-m MODEL.gguf\"        (llama-server)"
-fi
+    log_info "  ollama serve                            (ollama)"
-
-# Check for llama.cpp server
-if curl -s http://localhost:8080/health >/dev/null 2>&1; then
-    log_success "llama.cpp server running at localhost:8080"
-    llamacpp_ok=true
-else
-    log_info "No llama.cpp server detected at localhost:8080"
-    log_info "Start with: toolbox run -c llama-vulkan-radv -- llama-server -m MODEL -c 8192 -ngl 99 -fa 1 --no-mmap"
-fi
-
-if ! $ollama_ok && ! $llamacpp_ok; then
-    log_warn "No local LLM server running. Agentic benchmarks need one."
 fi
 
 # ── Summary ──────────────────────────────────────────────
 log_header "Setup Complete"
 echo ""
 echo "  Installed tools:"
-echo "    inspect-ai     — All-in-one eval framework (HumanEval, BFCL, IFEval, GAIA, ...)"
+echo "    inspect-ai      — All-in-one eval framework (IFEval, BFCL, GAIA, ...)"
-echo "    evalplus        — HumanEval+ / MBPP+ with native ollama support"
+echo "    inspect-evals   — Task definitions for inspect-ai"
-echo "    bigcodebench    — 1,140 coding tasks across 139 libraries"
+echo "    evalplus         — HumanEval+ / MBPP+ with native ollama support"
+echo "    bigcodebench     — 1,140 coding tasks across 139 libraries"
 echo ""
-echo "  To activate the virtual environment:"
+echo "  Activate venv:  source .venv/bin/activate"
-echo "    source data/venv/bin/activate"
 echo ""
 echo "  Run evaluations:"
-echo "    make agentic-quick      # EvalPlus + IFEval (~1 hour)"
+echo "    make agentic-quick      # EvalPlus HumanEval+ + IFEval (~1 hour)"
-echo "    make agentic-full       # BFCL + BigCodeBench (~3-4 hours)"
+echo "    make agentic-code       # EvalPlus + BigCodeBench (~2-3 hours)"
+echo "    make agentic-tooluse    # BFCL function calling (~1-2 hours)"
+echo "    make agentic-full       # All of the above (~5-6 hours)"
 echo ""

scripts/benchmark/run-baseline.sh

@@ -270,7 +270,7 @@ for MODEL_PATH in "${MODEL_PATHS[@]}"; do
                 CMD_LC=(toolbox run -c "$BACKEND" -- "${ENV_ARGS[@]}" "$BENCH_BIN"
                     -ngl 99 -mmp 0 -m "$TOOLBOX_MODEL_PATH" -fa 1
                     -p "$CTX_PROMPT" -n 32 -d "$CTX_DEPTH" -ub "$UB_SIZE"
-                    -r "$REPS_LONGCTX" "${KV_ARGS[@]}")
+                    -r "$REPS_LONGCTX" "${BATCH_ARGS[@]}" "${KV_ARGS[@]}")
 
                 printf "  cmd: %s\n" "${CMD_LC[*]}"
                 if "${CMD_LC[@]}" > "$OUT_LC" 2>&1; then

scripts/benchmark/run-suite.sh

@@ -252,7 +252,7 @@ for MODEL_PATH in "${MODEL_PATHS[@]}"; do
                 UB_SIZE=2048; [[ "$BACKEND" == *vulkan* ]] && UB_SIZE=512
                 CMD_LC=(toolbox run -c "$BACKEND" -- "${ENV_ARGS[@]}" "$BENCH_BIN"
                     -ngl 99 -mmp 0 -m "$TOOLBOX_MODEL_PATH" -fa 1
-                    -p "$CTX_PROMPT" -n 32 -d "$CTX_DEPTH" -ub "$UB_SIZE" -r "$REPS_LONGCTX" "${KV_ARGS[@]}")
+                    -p "$CTX_PROMPT" -n 32 -d "$CTX_DEPTH" -ub "$UB_SIZE" -r "$REPS_LONGCTX" "${BATCH_ARGS[@]}" "${KV_ARGS[@]}")
                 if "${CMD_LC[@]}" > "$OUT_LC" 2>&1; then
                     log_success "Done"; tail -3 "$OUT_LC"
                 else

scripts/serve/launch.sh

@@ -14,6 +14,7 @@ CTX_SIZE=131072
 PARALLEL=1
 MODEL=""
 NGRAM=false
+NO_THINK=false
 
 while [[ $# -gt 0 ]]; do
     case "$1" in
@@ -23,6 +24,7 @@ while [[ $# -gt 0 ]]; do
         --ctx)          CTX_SIZE="$2"; shift 2 ;;
         --parallel)     PARALLEL="$2"; shift 2 ;;
         --ngram)        NGRAM=true; shift ;;
+        --no-think)     NO_THINK=true; shift ;;
         --help|-h)
             echo "Usage: launch.sh [OPTIONS]"
             echo ""
@@ -33,6 +35,7 @@ while [[ $# -gt 0 ]]; do
             echo "  --ctx N              Context size (default: 131072)"
             echo "  --parallel N         Parallel request slots (default: 1)"
             echo "  --ngram              Enable n-gram speculative decoding (~1.1-1.4x tg)"
+            echo "  --no-think           Disable thinking/reasoning (faster for evals)"
             echo ""
             echo "Presets (pass model filename):"
             echo "  Qwen3.5-35B-A3B-UD-Q4_K_XL.gguf     General purpose daily driver"
@@ -101,7 +104,7 @@ fi
 SERVER_ARGS=(
     -ngl 99                          # Full GPU offload
     --no-mmap                        # Direct load, no mmap overhead
-    -fa                              # Flash attention
+    -fa on                            # Flash attention
     -m "$TOOLBOX_MODEL_PATH"
     -c "$CTX_SIZE"                   # Context size
     --cache-type-k q4_0              # KV cache quantization (fastest on Vulkan)
@@ -110,6 +113,11 @@ SERVER_ARGS=(
     -np "$PARALLEL"                  # Parallel slots
 )
 
+# Disable thinking mode (faster for evals)
+if $NO_THINK; then
+    SERVER_ARGS+=(--reasoning-budget 0)
+fi
+
 # N-gram speculative decoding
 if $NGRAM; then
     SERVER_ARGS+=(
@@ -126,6 +134,7 @@ log_info "Backend: $BACKEND"
 log_info "Context: $CTX_SIZE tokens"
 log_info "KV cache: q4_0/q4_0"
 log_info "Parallel slots: $PARALLEL"
+$NO_THINK && log_info "Thinking mode: DISABLED (--reasoning-budget 0)"
 $NGRAM && log_info "N-gram speculative: enabled (draft-max=64)"
 log_info "Port: $PORT"
 log_info "Endpoint: http://localhost:$PORT"

tests/benchmark_flags.bats

@@ -11,6 +11,7 @@ load test_helper.sh
     assert_output --partial "--category"
     assert_output --partial "--skip-longctx"
     assert_output --partial "--kv-types"
+    assert_output --partial "--batch"
 }
 
 @test "run-suite --help shows usage and exits 0" {
@@ -22,6 +23,7 @@ load test_helper.sh
     assert_output --partial "--skip-longctx"
     assert_output --partial "--tag"
     assert_output --partial "--kv-types"
+    assert_output --partial "--batch"
 }
 
 @test "benchmark dispatcher shows help with no args" {
@@ -31,6 +33,18 @@ load test_helper.sh
     assert_output --partial "--max-size"
     assert_output --partial "--skip-longctx"
     assert_output --partial "--kv-types"
+    assert_output --partial "--batch"
+}
+
+@test "serve --help shows usage and exits 0" {
+    run bash "$PROJECT_ROOT/bin/serve" --help
+    assert_success
+    assert_output --partial "Usage"
+    assert_output --partial "--model"
+    assert_output --partial "--ngram"
+    assert_output --partial "--no-think"
+    assert_output --partial "--ctx"
+    assert_output --partial "--port"
 }
 
 @test "benchmark dispatcher passes --help through to baseline" {