Test branch

.brev/setup.sh

@@ -0,0 +1,134 @@
+#!/usr/bin/env bash
+# .brev/setup.sh
+# Runs once on instance first boot (or launchable deploy).
+# Idempotent: safe to re-run.
+set -euo pipefail
+
+# ---------------------------------------------------------------------------
+# Paths
+# ---------------------------------------------------------------------------
+REPO_DIR="${REPO_DIR:-$HOME/cuopt-opt}"
+VENV="$HOME/coo-venv"
+NV_PYPI="https://pypi.nvidia.com"
+
+echo ">>> [1/5] Creating Python virtual environment at $VENV"
+python3 -m venv "$VENV" --system-site-packages
+# shellcheck source=/dev/null
+source "$VENV/bin/activate"
+
+# Upgrade pip/wheel first so binary wheels resolve correctly
+pip install --quiet --upgrade pip wheel
+
+# ---------------------------------------------------------------------------
+# Install cuOpt (binary wheel — no build from source required)
+# ---------------------------------------------------------------------------
+echo ">>> [2/5] Installing cuOpt and CUDA runtime libraries"
+pip install \
+  --extra-index-url "$NV_PYPI" \
+  --quiet \
+  cuopt-cu12 \
+  nvidia-nvjitlink-cu12
+
+# ---------------------------------------------------------------------------
+# Install the cuoptopt-agent + dashboard
+# ---------------------------------------------------------------------------
+echo ">>> [3/5] Installing cuoptopt-agent + dashboard dependencies"
+pip install --quiet -e "$REPO_DIR/cuoptopt-agent[dev]"
+
+# ---------------------------------------------------------------------------
+# nvJitLink LD_PRELOAD fix
+# Without this, libcuopt.so fails to import due to symbol version mismatch
+# between the system CUDA toolkit's libnvJitLink and the pip-installed one.
+# ---------------------------------------------------------------------------
+echo ">>> [4/5] Configuring LD_PRELOAD for libnvJitLink"
+JITLINK=$(python3 - <<'PYEOF'
+import pathlib
+try:
+    import nvidia.nvjitlink as m
+    base = pathlib.Path(m.__file__).parent
+    # Find the versioned .so (e.g. libnvJitLink.so.12)
+    candidates = sorted(base.glob("libnvJitLink.so.*"))
+    print(candidates[-1])
+except Exception as e:
+    print("")
+PYEOF
+)
+
+if [[ -n "$JITLINK" && -f "$JITLINK" ]]; then
+    MARKER="export LD_PRELOAD=$JITLINK"
+    if ! grep -qF "$JITLINK" "$HOME/.bashrc"; then
+        echo "${MARKER}\${LD_PRELOAD:+:\$LD_PRELOAD}" >> "$HOME/.bashrc"
+        echo "    Added LD_PRELOAD to ~/.bashrc: $JITLINK"
+    else
+        echo "    LD_PRELOAD already set in ~/.bashrc — skipping"
+    fi
+else
+    echo "    WARNING: could not locate libnvJitLink.so — skipping LD_PRELOAD fix"
+fi
+
+# Also export for the remainder of this script session
+if [[ -n "$JITLINK" && -f "$JITLINK" ]]; then
+    export LD_PRELOAD="$JITLINK${LD_PRELOAD:+:$LD_PRELOAD}"
+fi
+
+# ---------------------------------------------------------------------------
+# Write .env.template (API keys — never committed with real values)
+# ---------------------------------------------------------------------------
+echo ">>> [5/5] Writing $REPO_DIR/.env.template"
+cat > "$REPO_DIR/.env.template" << 'ENVEOF'
+# Copy this file to .env and fill in your keys.
+# Source it before starting the agent or dashboard:
+#   source .env
+#
+# Only the key for the model you use is required.
+# GITHUB_TOKEN is required for the branch sidebar and PR creation.
+
+export ANTHROPIC_API_KEY=""       # https://console.anthropic.com
+export OPENAI_API_KEY=""          # https://platform.openai.com
+export NVIDIA_API_KEY=""          # https://build.nvidia.com
+
+export GITHUB_TOKEN=""            # GitHub → Settings → Developer settings → PAT (scope: repo)
+
+# Optional: change the dashboard port (default 8080)
+# export DASHBOARD_PORT=8080
+ENVEOF
+
+# ---------------------------------------------------------------------------
+# Smoke-test: verify cuOpt imports cleanly
+# ---------------------------------------------------------------------------
+echo ""
+echo ">>> Verifying cuOpt import..."
+if python3 -c "import cuopt; print('  cuopt', cuopt.__version__, '— OK')" 2>/dev/null; then
+    :
+else
+    echo "  WARNING: cuopt import failed — check CUDA driver compatibility."
+    echo "  Run: python3 -c \"import cuopt\" to diagnose."
+fi
+
+# ---------------------------------------------------------------------------
+# Done
+# ---------------------------------------------------------------------------
+cat << 'BANNER'
+
+=============================================================
+  cuoptopt-agent setup complete!
+
+  Next steps:
+    1. Copy and fill in API keys:
+         cp .env.template .env && nano .env
+    2. Source the keys:
+         source .env
+    3. Start the dashboard:
+         source ~/coo-venv/bin/activate
+         cuoptopt-dashboard
+    4. Open the dashboard:
+         Use the Brev Secure Link for port 8080, or
+         http://localhost:8080 if connecting via brev port-forward.
+
+  Connect Cursor:
+    On your local machine:
+         brev refresh
+         brev open cursor <instance-name>
+=============================================================
+
+BANNER

AGENTS.md

@@ -32,6 +32,9 @@ AI agent skills for NVIDIA cuOpt optimization engine. Skills live in **`skills/`
 - `skills/cuopt-qp-api-cli/`
 - `skills/cuopt-server-api-python/` (deploy + client)
 
+### Optimization Agent
+- `skills/cuoptopt-agent/` — Autonomous optimization agent; reads this skill when the user asks to speed up, optimize, or improve cuOpt's solver performance or solution quality. Orchestrates literature research, LLM-driven code changes, benchmarking, and GitHub PR creation. Also launchable as a CLI via `cuoptopt-agent "query" --model claude|gpt|nvidia` or via Cursor tasks (`Ctrl+Shift+P` → **Tasks: Run Task** → **cuoptopt-agent: Run (interactive)**).
+
 ## Resources
 
 ### Documentation