diff --git a/docs/plans/2026-02-19-phase1-core-implementation-design.md b/docs/plans/2026-02-19-phase1-core-implementation-design.md
new file mode 100644
index 0000000..3efc54c
--- /dev/null
+++ b/docs/plans/2026-02-19-phase1-core-implementation-design.md
@@ -0,0 +1,52 @@
+# Phase 1: Core Implementation Design
+
+Approved 19 Feb 2026.
+
+## Component Structure
+
+4 files, layered bottom-up:
+
+```
+internal/llamacpp/client.go   <- HTTP + SSE (no inference types)
+server.go                     <- process lifecycle (spawn, health, kill)
+model.go                      <- inference.TextModel (wraps server + client)
+backend.go                    <- inference.Backend (fill in stubs)
+```
+
+Each layer only knows about the one below it. `internal/llamacpp/` is a pure HTTP client with zero go-inference dependency — it speaks llama-server's OpenAI-compatible API and returns plain Go types.
+
+## Data Flow
+
+```
+LoadModel("/path/to/model.gguf")
+  -> backend.go: find llama-server binary, apply LoadConfig
+  -> server.go:  spawn process (HIP_VISIBLE_DEVICES=0, random port)
+                  poll GET /health until {"status":"ok"} (timeout 60s)
+  -> model.go:   return &rocmModel{server, client}
+
+Generate(ctx, "prompt")
+  -> model.go:   build request body from GenerateConfig
+  -> client.go:  POST /v1/completions, stream=true
+                  parse SSE lines: data: {json}
+                  yield Token{Text: chunk} via iter.Seq[Token]
+  -> on ctx cancel: close HTTP response body (stops SSE)
+
+Close()
+  -> server.go:  SIGTERM -> wait 5s -> SIGKILL if needed
+```
+
+## Key Decisions
+
+- **Server lifecycle**: Long-lived. Spawn on LoadModel(), reuse for all Generate/Chat calls, kill on Close().
+- **Port selection**: `net.Listen(":0")` for kernel-assigned free port. Close listener, pass port to llama-server.
+- **SSE parsing**: Hand-rolled with `bufio.Scanner`. llama-server's SSE format is simple (`data: {json}\n\n`), ~20 lines.
+- **Health polling**: Simple loop, 100ms interval, 60s timeout. Returns error if exceeded.
+- **Error handling**: `Err()` stores last error. Process death mid-stream detected via response body EOF + process state.
+- **Thread safety**: `rocmModel` safe for concurrent calls (llama-server handles via slots). `lastErr` protected by mutex.
+- **iGPU mitigation**: `HIP_VISIBLE_DEVICES=0` set in process env unconditionally.
+
+## Testing
+
+- **Unit tests** (no GPU): mock HTTP responses for SSE parsing in client.go
+- **Integration test** (GPU + model): LoadModel -> Generate -> Close, `t.Skip()` when unavailable
+- **Test model**: LEK-Gemma3-1B-layered-v2-Q5_K_M.gguf (smallest, fast load)