feat(llm): oMLX-backed chat client + server control + model-id resolution

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016DvdRZkyCy9NJpHYDdPqmS

llm/server.py

@@ -0,0 +1,87 @@
+"""Ensure the oMLX server is up for a run.
+
+oMLX runs as a managed background service (auto-starts on login) and loads models
+on-demand with LRU eviction, so there's usually nothing to do but health-check. For the
+nightly launchd job we still `omlx start` defensively and warm the active model so the
+first agent turn isn't paying cold-load latency.
+"""
+
+from __future__ import annotations
+
+import os
+import shutil
+import subprocess
+import time
+
+import httpx
+from loguru import logger
+
+from config.settings import get_settings
+
+_OMLX_CANDIDATES = ["omlx", os.path.expanduser("~/.omlx/bin/omlx")]
+
+
+def _omlx_bin() -> str | None:
+    for cand in _OMLX_CANDIDATES:
+        path = shutil.which(cand) if os.path.sep not in cand else (cand if os.path.exists(cand) else None)
+        if path:
+            return path
+    return None
+
+
+def server_up() -> bool:
+    s = get_settings()
+    try:
+        r = httpx.get(
+            f"{s.llm_base_url}/models",
+            headers={"Authorization": f"Bearer {s.llm_api_key}"},
+            timeout=5,
+        )
+        return r.status_code == 200
+    except Exception:
+        return False
+
+
+def start_server() -> None:
+    if server_up():
+        logger.info("oMLX already up at {}", get_settings().llm_base_url)
+        return
+    bin_ = _omlx_bin()
+    if not bin_:
+        logger.warning("omlx CLI not found; cannot auto-start. Start the oMLX app manually.")
+        return
+    proc = subprocess.run([bin_, "start"], capture_output=True, text=True, timeout=120)
+    logger.info("omlx start -> {} {}", proc.returncode, (proc.stdout + proc.stderr).strip()[:300])
+
+
+def wait_healthy(timeout_s: int = 120) -> bool:
+    deadline = time.time() + timeout_s
+    while time.time() < deadline:
+        if server_up():
+            return True
+        time.sleep(3)
+    return False
+
+
+def warm_model() -> None:
+    """Fire a tiny completion so the active model is resident before the first agent turn."""
+    from llm.client import build_llm
+
+    try:
+        build_llm().invoke("ok")
+        logger.info("warmed model")
+    except Exception as e:
+        logger.warning("warm_model failed (non-fatal): {}", e)
+
+
+def ensure_ready(warm: bool = True) -> bool:
+    start_server()
+    ok = wait_healthy(120)
+    if ok and warm:
+        warm_model()
+    logger.info("oMLX ready: {}", ok)
+    return ok
+
+
+if __name__ == "__main__":
+    ensure_ready()