feat: disable reranking in bundled backend

DEVELOPER.md

@@ -32,7 +32,7 @@ Tunables read by the plugin at runtime. Most users don't need to touch these —
 | `REFLEXIO_EMBEDDING_PROVIDER` | `local_service` when local embedding is enabled | Embedding provider mode: `cloud`, `local_service`, `internal_service`, `inprocess`, or `off`. |
 | `REFLEXIO_EMBEDDING_SERVICE_URL` | `http://127.0.0.1:$EMBEDDING_PORT` | OpenAI-compatible embedding endpoint used by `local_service` and `internal_service`. |
 | `EMBEDDING_PORT` | `8072` | Local embedding daemon port. Shared by Claude Code and Codex on the same machine. |
-| `REFLEXIO_RERANK_DEVICE` | `cpu` | Torch device for the in-process cross-encoder reranker used by search. Defaults to CPU: auto-selected MPS on Apple Silicon accumulates gigabytes of GPU memory in the backend process for no latency win at this model size. Set to `cuda` only for GPU deployments. |
+| `REFLEXIO_RERANK_ENABLED` | `false` (hard-set by `backend-service.sh`) | claude-smart disables Reflexio's shared-service cross-encoder completely, so the local inference process skips reranker model loading and prewarm. This is intentionally not user-overridable in the bundled backend. |
 | `CLAUDE_SMART_ENABLE_OPTIMIZER` | auto: local Reflexio only | Hook-side env var that controls shared skill optimization and rollups during `SessionStart`. Set to `1` to force optimizer defaults for hosted Reflexio, or `0` to disable everywhere. Set it in Claude Code settings, not `~/.reflexio/.env`. |
 | `CLAUDE_SMART_CITATIONS` | `on` | Controls the final `✨ claude-smart rule applied` marker instruction. `on` (default) injects a compact instruction asking the assistant to cite only memories that materially changed the answer. `off` skips injection of the citation instruction entirely and strips any stray marker line from the assistant text before publishing. Legacy values `auto` and `marker-only` are accepted as enabled aliases. Set in Claude Code settings (hook env), not `~/.reflexio/.env`. |
 | `CLAUDE_SMART_CITATION_LINK_STYLE` | `markdown` | Controls the visible links in the final `✨ claude-smart rule applied` marker. `markdown` asks the model to emit `[human title](dashboard URL)` links and is the Codex default because Codex renders markdown links in assistant messages. `osc8` asks it to emit terminal-native OSC 8 hyperlinks so the human title can be clickable without showing the URL; unsupported terminals should fall back to markdown links. |

plugin/dashboard/app/configure/server/page.tsx

@@ -229,9 +229,10 @@ export default function ConfigureServerPage() {
               <div className="min-w-0">
                 <Label htmlFor="search-reranker">Search reranker</Label>
                 <p className="text-xs text-muted-foreground mt-0.5">
-                  Use the local cross-encoder relevance floor for injected
-                  memory search. The first enabled search loads the model into
-                  the backend process.
+                  Use the shared-service cross-encoder relevance floor for
+                  injected memory search. The bundled claude-smart backend
+                  disables this model; the toggle applies to connected
+                  deployments that enable reranking.
                 </p>
               </div>
               <Switch

plugin/scripts/backend-service.sh

@@ -51,6 +51,11 @@ export REFLEXIO_DEFAULT_ORG_ID="claude-smart"
 # the source path in _lib.sh (claude_smart_source_reflexio_env).
 export REFLEXIO_ENV_FILE="$HOME/.claude-smart/.env"
 
+# claude-smart does not use the cross-encoder reranker. Hard-disable the whole
+# feature so its shared local inference service does not load or prewarm the
+# reranker model, even if a shared Reflexio env file enables it.
+export REFLEXIO_RERANK_ENABLED=false
+
 # Default: route extraction through the active host CLI + ONNX embedder
 # so claude-smart works without any LLM API key. Users can opt out by
 # pre-exporting these to 0.

tests/test_codex_support.py

@@ -73,6 +73,7 @@ def test_codex_hook_loads_managed_reflexio_env_and_skips_backend_start() -> None
     assert "backend-service.sh" in script
     backend = (REPO_ROOT / "plugin" / "scripts" / "backend-service.sh").read_text()
     assert "remote REFLEXIO_URL configured; skipping local backend start" in backend
+    assert "export REFLEXIO_RERANK_ENABLED=false" in backend
 
 
 def test_codex_marketplace_points_at_plugin_root() -> None: