Bump reflexio-ai to 0.2.25 (CPU-default reranker) + venv-direct backend with vendored-install self-repair

DEVELOPER.md

@@ -30,6 +30,7 @@ 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. |
 | `CLAUDE_SMART_ENABLE_OPTIMIZER` | enabled unless set to `0` | Hook-side env var that controls shared skill optimization and rollups during `SessionStart`. 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. |

bin/claude-smart.js

@@ -905,7 +905,7 @@ async function installVendoredReflexio(pluginRoot, uv, env) {
   process.stdout.write(`Installing bundled Reflexio source from ${vendorRoot}...\n`);
   let code = await runChecked(
     uv,
-    ["pip", "install", "--project", pluginRoot, "--python", pythonPath, "--quiet", "-e", vendorRoot],
+    ["pip", "install", "--project", pluginRoot, "--python", pythonPath, "--quiet", "--reinstall", "--no-deps", vendorRoot],
     { cwd: pluginRoot, env },
   );
   if (code !== 0) {
@@ -914,7 +914,7 @@ async function installVendoredReflexio(pluginRoot, uv, env) {
     );
     code = await runChecked(
       uv,
-      ["pip", "install", "--project", pluginRoot, "--python", pythonPath, "-e", vendorRoot],
+      ["pip", "install", "--project", pluginRoot, "--python", pythonPath, "--reinstall", "--no-deps", vendorRoot],
       { cwd: pluginRoot, env },
     );
   }

plugin/pyproject.toml

@@ -9,7 +9,7 @@ dependencies = [
     # normal installs; vendor releases may override it after uv sync by
     # installing plugin/vendor/reflexio. Keep in sync via
     # scripts/sync-reflexio-dep.py.
-    "reflexio-ai>=0.2.23",
+    "reflexio-ai>=0.2.25",
     # Needed when Reflexio runs the local embedding provider (ONNXMiniLM_L6_V2).
     # Pulls in onnxruntime + tokenizers; the ~80 MB ONNX model itself is
     # downloaded on first use, not at install time.

plugin/scripts/backend-service.sh

@@ -5,8 +5,8 @@
 #
 # Subcommands:
 #   start         probe /health; if nothing we recognize is on the port,
-#                 spawn `uv run reflexio services start --only backend
-#                 --no-reload` detached. Polls /health briefly so first
+#                 spawn the prepared venv's `reflexio services start --only
+#                 backend --no-reload` detached. Polls /health briefly so first
 #                 use after session start lands on a warm server, then
 #                 returns a continue payload regardless.
 #   stop          SIGTERM the recorded process group, escalating to
@@ -183,6 +183,35 @@ port_holder() {
     | awk 'NR==2 {print $1" (pid "$2")"; exit}'
 }
 
+ensure_vendored_reflexio_active() {
+  vendor="$PLUGIN_ROOT/vendor/reflexio"
+  [ -f "$vendor/pyproject.toml" ] || return 0
+  plugin_python="$(claude_smart_plugin_python "$PLUGIN_ROOT")"
+  [ -x "$plugin_python" ] || return 0
+  if "$plugin_python" - "$vendor/pyproject.toml" <<'PY' >/dev/null 2>&1; then
+import importlib.metadata
+import re
+import sys
+from pathlib import Path
+
+text = Path(sys.argv[1]).read_text(encoding="utf-8")
+match = re.search(r'^version\s*=\s*"([^"]+)"', text, re.MULTILINE)
+if not match:
+    raise SystemExit(1)
+expected = match.group(1)
+try:
+    installed = importlib.metadata.version("reflexio-ai")
+except importlib.metadata.PackageNotFoundError:
+    raise SystemExit(1)
+raise SystemExit(0 if installed == expected else 1)
+PY
+    return 0
+  fi
+  claude_smart_append_capped_log "$LOG_FILE" "$LOG_MAX_BYTES" \
+    "[claude-smart] backend: repairing vendored Reflexio install before start"
+  uv pip install --project "$PLUGIN_ROOT" --python "$plugin_python" --quiet --reinstall --no-deps "$vendor" >&2
+}
+
 # Full shutdown: kill the recorded process group (if any) then sweep
 # both the backend port and the embedding-service port for surviving
 # reflexio listeners. Used by both `stop` and the opt-in `session-end`
@@ -243,6 +272,7 @@ case "$CMD" in
       fi
     fi
     cd "$PLUGIN_ROOT"
+    ensure_vendored_reflexio_active
 
     # Cap local interaction history to keep the SQLite store small for
     # claude-smart users. Reflexio's library defaults are much higher
@@ -266,6 +296,21 @@ case "$CMD" in
     # without touching the file on disk.
     export REFLEXIO_STORAGE="sqlite"
 
+    backend_pythonpath="${PYTHONPATH:-}"
+    if [ -d "$PLUGIN_ROOT/vendor/reflexio/reflexio" ]; then
+      vendor_pythonpath="$PLUGIN_ROOT/vendor/reflexio"
+      pythonpath_sep=":"
+      if claude_smart_is_windows; then
+        # Native Windows Python expects ;-separated Windows-style paths in
+        # PYTHONPATH; MSYS does not auto-convert arbitrary env vars.
+        pythonpath_sep=";"
+        if command -v cygpath >/dev/null 2>&1; then
+          vendor_pythonpath="$(cygpath -w "$vendor_pythonpath")"
+        fi
+      fi
+      backend_pythonpath="$vendor_pythonpath${backend_pythonpath:+$pythonpath_sep$backend_pythonpath}"
+    fi
+
     # (nohup; no process groups). backend-log-runner.sh owns stdout/stderr
     # capture so process output cannot grow backend.log past its cap.
     #
@@ -277,11 +322,12 @@ case "$CMD" in
     # CLAUDE_SMART_BACKEND_WORKERS for power users running concurrent
     # Claude Code sessions or wanting zero-downtime recycling.
     workers="${CLAUDE_SMART_BACKEND_WORKERS:-1}"
+    backend_python="$(claude_smart_plugin_python "$PLUGIN_ROOT")"
     claude_smart_spawn_detached bash "$HERE/backend-log-runner.sh" \
       "$LOG_FILE" "$LOG_MAX_BYTES" -- \
       env PYTHONIOENCODING="${PYTHONIOENCODING:-utf-8}" \
-      uv run --project "$PLUGIN_ROOT" --no-sync --quiet \
-      reflexio services start --only backend --no-reload --workers "$workers"
+      PYTHONPATH="$backend_pythonpath" \
+      "$backend_python" -m reflexio.cli services start --only backend --no-reload --workers "$workers"
     svc_pid=$!
     # Record the spawned pid, not a pgid sampled with ps. On POSIX,
     # setsid/python os.setsid make this pid the new process group leader;

plugin/scripts/cli.sh

@@ -91,4 +91,5 @@ if ! claude_smart_python_imports "$PLUGIN_ROOT" claude_smart.cli; then
   fi
 fi
 
-exec uv run --project "$PLUGIN_ROOT" --no-sync --quiet python -m claude_smart.cli "$@"
+PLUGIN_PYTHON="$(claude_smart_plugin_python "$PLUGIN_ROOT")"
+exec "$PLUGIN_PYTHON" -m claude_smart.cli "$@"

plugin/scripts/ensure-plugin-root.sh

@@ -20,6 +20,7 @@ if [ ! -f "$TARGET/pyproject.toml" ]; then
     echo "[claude-smart] ensure-plugin-root: $TARGET is not a plugin dir (no pyproject.toml)" >&2
     exit 1
 fi
+TARGET="$(cd "$TARGET" && pwd -P)"
 
 LINK="$HOME/.reflexio/plugin-root"
 mkdir -p "$(dirname "$LINK")"

plugin/scripts/hook_entry.sh

@@ -3,8 +3,9 @@
 # CLAUDE_PLUGIN_ROOT points at the plugin dir (dev: <repo>/plugin;
 # installed: ~/.claude/plugins/cache/reflexioai/claude-smart/<version>),
 # which is also the Python project root with pyproject.toml + uv.lock.
-# We invoke via `uv run --project --no-sync` so the prepared env is used without
-# undoing editable Reflexio installs from local development.
+# We invoke the prepared venv Python directly so hooks use the dependency set
+# produced by smart-install.sh, including the vendored Reflexio snapshot it
+# installs non-editably (--reinstall --no-deps).
 #
 # If the Setup hook recorded an install failure at
 # ~/.claude-smart/install-failed, short-circuit with a user-visible
@@ -135,4 +136,5 @@ if ! ensure_hook_package_importable; then
 fi
 
 # Stdin is the hook payload JSON — stream it through to the Python CLI.
-exec uv run --project "$PLUGIN_ROOT" --no-sync --quiet python -m claude_smart.hook "$HOST" "$EVENT"
+PLUGIN_PYTHON="$(claude_smart_plugin_python "$PLUGIN_ROOT")"
+exec "$PLUGIN_PYTHON" -m claude_smart.hook "$HOST" "$EVENT"

plugin/scripts/smart-install.sh

@@ -142,12 +142,12 @@ install_vendored_reflexio() {
   fi
 
   echo "[claude-smart] installing bundled Reflexio source from $VENDORED_REFLEXIO" >&2
-  if uv pip install --project "$PLUGIN_ROOT" --python "$PLUGIN_PYTHON" --quiet -e "$VENDORED_REFLEXIO" >&2; then
+  if uv pip install --project "$PLUGIN_ROOT" --python "$PLUGIN_PYTHON" --quiet --reinstall --no-deps "$VENDORED_REFLEXIO" >&2; then
     return 0
   fi
 
   echo "[claude-smart] warning: quiet vendored Reflexio install failed in $PLUGIN_ROOT; retrying with full output." >&2
-  if ! uv pip install --project "$PLUGIN_ROOT" --python "$PLUGIN_PYTHON" -e "$VENDORED_REFLEXIO" >&2; then
+  if ! uv pip install --project "$PLUGIN_ROOT" --python "$PLUGIN_PYTHON" --reinstall --no-deps "$VENDORED_REFLEXIO" >&2; then
     write_failure "vendored Reflexio install failed in $PLUGIN_ROOT"
   fi
 }