diff --git a/docs/plans/2026-02-19-phase2-robustness-design.md b/docs/plans/2026-02-19-phase2-robustness-design.md
new file mode 100644
index 0000000..910ba88
--- /dev/null
+++ b/docs/plans/2026-02-19-phase2-robustness-design.md
@@ -0,0 +1,46 @@
+# Phase 2: Robustness Design
+
+Approved 19 Feb 2026.
+
+## 1. Graceful Shutdown (context cancellation)
+
+Already works in Phase 1. Context cancellation closes the HTTP response body, stops SSE streaming, but leaves llama-server alive. Generate/Chat can be called again with a new context.
+
+Only change: add an integration test that cancels mid-stream then generates again on the same model to verify the server survives.
+
+## 2. Port Conflict Handling
+
+Retry loop in startServer(): if the process fails (port taken), call freePort() again and retry up to 3 attempts.
+
+## 3. Server Crash Recovery
+
+Add server.alive() method (non-blocking check on exited channel). Generate/Chat check alive() before making HTTP calls. If dead, return error immediately. No auto-restart — consumer must Close() + LoadModel() again.
+
+## 4. VRAM Monitoring
+
+Read sysfs directly (no subprocess spawn):
+- `/sys/class/drm/cardN/device/mem_info_vram_total`
+- `/sys/class/drm/cardN/device/mem_info_vram_used`
+
+Find dGPU by picking the card with the largest VRAM total (avoids hardcoding card numbers). On this machine: card0 = iGPU (2GB), card1 = dGPU (16GB).
+
+Expose via:
+```go
+type VRAMInfo struct {
+    Total uint64
+    Used  uint64
+    Free  uint64
+}
+
+func GetVRAMInfo() (VRAMInfo, error)
+```
+
+## 5. Concurrent Requests
+
+Integration test only. 3 goroutines calling Generate() on the same model simultaneously. Verify all get results. Document concurrency limits in FINDINGS.md.
+
+## Testing
+
+- Tasks 1-3: unit tests (mock servers, process helpers)
+- Task 4: unit test (sysfs on real hardware)
+- Task 5: integration test (GPU + model, //go:build rocm)