fix(cookbook): reset MLX env vars before each serve run

routes/cookbook_routes.py

@@ -2496,9 +2496,20 @@ async def model_serve(request: Request, req: ServeRequest):
                 runner_lines.append('        p = os.path.expanduser(p)')
                 runner_lines.append('        if p not in roots: roots.append(p)')
                 runner_lines.append('    add(os.environ.get("HUGGINGFACE_HUB_CACHE"))')
+                runner_lines.append('    # Best-effort: huggingface_hub knows the platform default')
+                runner_lines.append('    # (macOS uses ~/Library/Caches/huggingface, not ~/.cache) - #5978.')
+                runner_lines.append('    try:')
+                runner_lines.append('        import huggingface_hub')
+                runner_lines.append('        _hc = getattr(getattr(huggingface_hub, "constants", None), "HF_HUB_CACHE", None)')
+                runner_lines.append('        if _hc: add(_hc)')
+                runner_lines.append('    except Exception:')
+                runner_lines.append('        pass')
                 runner_lines.append('    hf_home = os.environ.get("HF_HOME")')
                 runner_lines.append('    if hf_home: add(os.path.join(hf_home, "hub"))')
-                runner_lines.append('    add("~/.cache/huggingface/hub")')
+                runner_lines.append('    if sys.platform == "darwin":')
+                runner_lines.append('        add("~/Library/Caches/huggingface/hub")')
+                runner_lines.append('    else:')
+                runner_lines.append('        add("~/.cache/huggingface/hub")')
                 runner_lines.append('    cache_name = "models--" + model.replace("/", "--")')
                 runner_lines.append('    best = ""')
                 runner_lines.append('    best_mtime = -1.0')
@@ -2640,6 +2651,15 @@ async def model_serve(request: Request, req: ServeRequest):
                 runner_lines.append('    echo "Install with: $ODYSSEUS_MLX_IMAGE_CMD_PY -m pip install -U fastapi uvicorn python-multipart pillow huggingface_hub"')
                 runner_lines.append('    ODYSSEUS_PREFLIGHT_EXIT=127')
                 runner_lines.append('  fi')
+                # Reset any lingering MLX environment variables before each run to ensure a clean environment
+                runner_lines.append('    # Begin MLX env reset')
+                runner_lines.append('    for var in MLX_ROOT MLX_CACHE; do')
+                runner_lines.append('        if [ -n "${!var}" ]; then')
+                runner_lines.append('            echo "[odysseus] Unsetting $var for clean MLX run"')
+                runner_lines.append('            unset $var')
+                runner_lines.append('        fi')
+                runner_lines.append('    done')
+                runner_lines.append('    # End MLX env reset')
                 runner_lines.append('  if ! command -v odysseus-mlx-inpaint >/dev/null 2>&1 && ! command -v mlx-lama-serve >/dev/null 2>&1; then')
                 runner_lines.append('    echo "ERROR: LaMa / MI-GAN MLX serving requires the Odysseus mlx-lama-swift bridge on PATH: odysseus-mlx-inpaint or mlx-lama-serve."')
                 runner_lines.append('    echo "Build it from swift/odysseus-mlx-image-bridge in Cookbook Dependencies."')

tests/test_cookbook_mlx_cache_roots.py

@@ -0,0 +1,113 @@
+"""Runner-side MLX HF-cache resolution (#5978 follow-up).
+
+The cookbook runner's MLX model-path resolver must find locally cached
+`mlx-community/...` snapshots. It previously listed only the Linux cache root
+(`~/.cache/huggingface/hub`), so on macOS — where `huggingface_hub` defaults to
+`~/Library/Caches/huggingface/hub` — a downloaded model could not be resolved at
+serve time. Same defect class as issue #5978 (scan path), fixed here on the
+serve path.
+
+The resolver ships as generated Python lines inside
+`routes/cookbook_routes.py` (the tmux runner script), so the test extracts the
+candidate-list block from the SHIPPED string literals via AST and executes it —
+asserting on the executed behavior (the resulting `roots` list), not on source
+text. Extraction is required because the block only exists as emitted runner
+content; this mirrors `tests/test_cookbook_stale_shim_recovery.py`, which
+extracts and runs shipped shell lines for the same reason.
+"""
+import ast
+import textwrap
+from pathlib import Path
+
+import pytest
+
+REPO = Path(__file__).resolve().parent.parent
+ROUTES = REPO / "routes" / "cookbook_routes.py"
+
+
+def _shipped_runner_lines() -> list:
+    """Every `runner_lines.append("<literal>")` string in the route source."""
+    tree = ast.parse(ROUTES.read_text(encoding="utf-8"))
+    out = []
+    for node in ast.walk(tree):
+        if (
+            isinstance(node, ast.Call)
+            and isinstance(node.func, ast.Attribute)
+            and node.func.attr == "append"
+            and getattr(node.func.value, "id", "") == "runner_lines"
+            and node.args
+        ):
+            arg = node.args[0]
+            if isinstance(arg, ast.Constant) and isinstance(arg.value, str):
+                out.append(arg.value)
+    return out
+
+
+def _cache_block() -> list:
+    """The MLX resolver's candidate-list block, in shipped order."""
+    lines = _shipped_runner_lines()
+    start = next(i for i, ln in enumerate(lines) if ln.strip() == "roots = []")
+    end = next(
+        i for i, ln in enumerate(lines)
+        if i > start and ln.strip().startswith("cache_name =")
+    )
+    block = lines[start:end]
+    assert any("HUGGINGFACE_HUB_CACHE" in ln for ln in block)
+    assert any("add(" in ln for ln in block)
+    return block
+
+
+def _roots(monkeypatch, platform: str, home, hf_home=None, hub_cache=None) -> list:
+    """Execute the shipped block with a controlled platform + HOME."""
+    monkeypatch.setenv("HOME", str(home))
+    for var in ("HF_HOME", "HUGGINGFACE_HUB_CACHE"):
+        monkeypatch.delenv(var, raising=False)
+    if hf_home:
+        monkeypatch.setenv("HF_HOME", str(hf_home))
+    if hub_cache:
+        monkeypatch.setenv("HUGGINGFACE_HUB_CACHE", str(hub_cache))
+    monkeypatch.setattr("sys.platform", platform)
+
+    import os
+    import sys
+
+    ns = {"os": os, "sys": sys}
+    # The shipped block lives inside the runner script's `if` body, so it
+    # carries that indentation; dedent before compiling it standalone.
+    program = textwrap.dedent("\n".join(_cache_block())) + "\n__ROOTS = list(roots)\n"
+    exec(compile(program, "<shipped-mlx-cache-block>", "exec"), ns)
+    return ns["__ROOTS"]
+
+
+def test_macos_includes_library_caches_root(monkeypatch, tmp_path):
+    """macOS resolves the platform default cache (#5978 class)."""
+    roots = _roots(monkeypatch, "darwin", tmp_path)
+    assert any("Library/Caches/huggingface" in r for r in roots), roots
+
+
+def test_linux_keeps_dotcache_root(monkeypatch, tmp_path):
+    """Linux behavior is unchanged (no regression for the common case)."""
+    roots = _roots(monkeypatch, "linux", tmp_path)
+    assert any(r.endswith(".cache/huggingface/hub") for r in roots), roots
+    assert not any("Library/Caches" in r for r in roots), roots
+
+
+def test_explicit_cache_env_wins_first(monkeypatch, tmp_path):
+    """An explicit HUGGINGFACE_HUB_CACHE is honored ahead of the defaults."""
+    explicit = tmp_path / "custom-hub"
+    roots = _roots(monkeypatch, "linux", tmp_path, hub_cache=explicit)
+    assert roots[0] == str(explicit), roots
+
+
+def test_hf_home_hub_is_included(monkeypatch, tmp_path):
+    """HF_HOME-based caches are still searched."""
+    hf_home = tmp_path / "hfhome"
+    roots = _roots(monkeypatch, "linux", tmp_path, hf_home=hf_home)
+    assert str(hf_home / "hub") in roots, roots
+
+
+def test_darwin_and_linux_sets_are_not_identical(monkeypatch, tmp_path):
+    """Guards against a regression back to a single hard-coded root."""
+    mac = _roots(monkeypatch, "darwin", tmp_path)
+    lin = _roots(monkeypatch, "linux", tmp_path)
+    assert mac != lin, (mac, lin)