@meridiona/meridian-darwin-arm64 1.23.1 → 1.23.3

package/VERSION

@@ -1 +1 @@
-1.23.1
+1.23.3

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@meridiona/meridian-darwin-arm64",
-  "version": "1.23.1",
+  "version": "1.23.3",
   "description": "Prebuilt Meridian app for macOS arm64 (daemon binary + dashboard + Python services). Installed via @meridiona/meridian.",
   "homepage": "https://github.com/Meridiona/meridian",
   "repository": {

package/scripts/install-from-bundle.sh

@@ -335,6 +335,25 @@ else
     fi
 fi
 
+# On macOS 26+, install apple-fm-sdk so Apple Intelligence is used instead of
+# downloading a large MLX model. This runs after both venv paths (tarball or uv
+# sync) so the package is available regardless of how the venv was built.
+# apple-fm-sdk only installs on macOS 26+ (links against system frameworks);
+# on older macOS pip will fail gracefully and MLX is used as the fallback.
+_macos_major="$(sw_vers -productVersion 2>/dev/null | cut -d. -f1)"
+if [[ "${_macos_major:-0}" -ge 26 ]]; then
+    if "${VENV}/bin/python" -c "import apple_fm_sdk" 2>/dev/null; then
+        ok "apple-fm-sdk already installed — Apple Intelligence will be used"
+    else
+        info "macOS ${_macos_major} detected — installing apple-fm-sdk for Apple Intelligence (no MLX model download needed)…"
+        if "${VENV}/bin/pip" install --quiet "apple-fm-sdk" 2>/dev/null; then
+            ok "apple-fm-sdk installed — Apple Intelligence will be used"
+        else
+            warn "apple-fm-sdk install failed — MLX model download will be used instead"
+        fi
+    fi
+fi
+
 # ── 5. macOS permissions for screenpipe (manual — can't be automated) ────────
 if [[ "${SKIP_PERMISSIONS}" -eq 0 ]]; then
     echo "→ screenpipe needs 2 macOS permissions: Screen Recording and Accessibility."

package/services/agents/llm_selector.py

@@ -278,6 +278,31 @@ _MANAGED_SERVER_PID_FILE = Path.home() / ".meridian" / "mlx_lm_server.pid"
 APPLE_INTELLIGENCE_ID = "apple-intelligence"
 
 
+def _apple_intelligence_available() -> bool:
+    """Return True only when Apple Intelligence is genuinely usable.
+
+    Checks three things in order (fail-fast):
+      1. macOS 26+ (the minimum for the Foundation Models API)
+      2. apple_fm_sdk is importable (installed in the venv)
+      3. SystemLanguageModel().is_available() — the on-device model is downloaded
+         and Apple Intelligence is enabled in System Settings
+
+    Any failure → False (graceful degradation to smaller MLX or cloud).
+    """
+    try:
+        macos_major = int(platform.mac_ver()[0].split(".")[0] or "0")
+        if macos_major < 26:
+            return False
+        from apple_fm_sdk import SystemLanguageModel  # type: ignore[import]
+        available, reason = SystemLanguageModel().is_available()
+        if not available:
+            log.debug("llm_selector: Apple Intelligence unavailable: %s", reason)
+        return bool(available)
+    except Exception as exc:  # noqa: BLE001
+        log.debug("llm_selector: Apple Intelligence probe skipped: %s", exc)
+        return False
+
+
 def _metal_headroom_gb() -> tuple[float, str]:
     """Primary memory signal — headroom within Metal's recommended working set.
 
@@ -448,8 +473,7 @@ def local_infer(system_prompt: str, user_message: str,
     # Relax budget when screen is locked — user won't feel the latency
     effective_pct = min(0.8, budget_pct * 1.5) if snap.screen_locked else budget_pct
 
-    macos_major = int(platform.mac_ver()[0].split(".")[0] or "0")
-    apple_intelligence = macos_major >= 26
+    apple_intelligence = _apple_intelligence_available()
 
     entry = _select_mlx_entry(snap.metal_headroom_gb, effective_pct,
                               snap.thermal_level, apple_intelligence)
@@ -893,8 +917,7 @@ def select_mlx_model_id(
             span.set_attribute("llm.selected_model", preferred_hf_id or "")
             return preferred_hf_id
 
-        macos_major = int(platform.mac_ver()[0].split(".")[0] or "0")
-        apple_intelligence = macos_major >= 26
+        apple_intelligence = _apple_intelligence_available()
 
         try:
             snap = probe_compute()

package/services/agents/run_task_linker_mlx.py

@@ -272,6 +272,12 @@ def _get_model() -> Any:
     return outlines_model
 
 
+# Apple Foundation Models has a 4096-token combined context window (input + output).
+# Reserve _MAX_TOKENS (1024) for the response; leave ~3072 tokens for the prompt.
+# At ~4 chars/token that is ~12 000 chars. Cap the user message to stay inside it.
+_APPLE_FM_USER_CHARS = 11_000
+
+
 def _classify_apple_fm(messages: list[dict[str, str]]) -> "SessionClassification":
     """Classify via Apple Foundation Models (non-FSM, JSON parsing with one retry)."""
     import asyncio
@@ -280,6 +286,15 @@ def _classify_apple_fm(messages: list[dict[str, str]]) -> "SessionClassification
 
     system = next((m["content"] for m in messages if m["role"] == "system"), "")
     user   = next((m["content"] for m in messages if m["role"] == "user"),   "")
+
+    # Truncate to stay within the 4096-token context window.
+    if len(user) > _APPLE_FM_USER_CHARS:
+        log.debug(
+            "run_task_linker_mlx: truncating Apple FM user message %d → %d chars",
+            len(user), _APPLE_FM_USER_CHARS,
+        )
+        user = user[:_APPLE_FM_USER_CHARS]
+
     user_with_hint = (
         user
         + "\n\nRespond with a JSON object matching the SessionClassification schema. "

package/services/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "meridian-agents"
-version = "1.23.1"
+version = "1.23.3"
 description = "Meridian agents — hermes task linking and Jira progress updates for meridian.db"
 requires-python = ">=3.11"
 authors = [{ name = "Meridiona" }]