diff --git a/docs/ADR/ADR-022-Workspace-First-Paradigm.md b/docs/ADR/ADR-022-Workspace-First-Paradigm.md
index c8bf427..1173855 100644
--- a/docs/ADR/ADR-022-Workspace-First-Paradigm.md
+++ b/docs/ADR/ADR-022-Workspace-First-Paradigm.md
@@ -285,6 +285,59 @@ mlxk run whisper-large-v3
 
 ---
 
+## Name Resolution Semantics
+
+### Fuzzy Matching by Context
+
+| Resolution Type | Example | Fuzzy? | Rationale |
+|-----------------|---------|--------|-----------|
+| **Name** in namespace | `mlxk run whisper` | ✅ Yes | Namespace search (MLXK_WORKSPACE_HOME, then HF cache) |
+| **Explicit path** | `mlxk run /path/whisper` | ❌ No | User points to concrete location |
+| **Query** (list) | `mlxk list /path/pix` | ✅ Yes | Search/discovery, not execution |
+
+**Security rationale:** Explicit paths (`/`, `./`, `../`) have exact semantics, analogous to `exec()` vs shell globbing. User intent is explicit → resolution is exact.
+
+```bash
+MLXK_WORKSPACE_HOME=~/mlx-models
+
+# Name → Fuzzy in MLXK_WORKSPACE_HOME (then HF cache)
+mlxk run whisper        # → ~/mlx-models/whisper-large-v3-mlx ✅
+mlxk run pixtral        # → Error: Ambiguous (pixtral-12b-8bit, pixtral-12b-4bit)
+
+# Explicit path → Exact match required
+mlxk run ~/mlx-models/whisper          # → Error: not found
+mlxk run ~/mlx-models/whisper-large-v3-mlx  # → OK ✅
+
+# Query → Fuzzy (discovery)
+mlxk list ~/mlx-models/whisper         # → shows whisper-large-v3-mlx
+```
+
+---
+
+## Command Scope
+
+### Which commands work with workspaces?
+
+| Command | Cache | Workspace | Notes |
+|---------|-------|-----------|-------|
+| `list` | ✅ | ✅ | Shows both with `source` column |
+| `show` | ✅ | ✅ | Includes workspace metadata |
+| `health` | ✅ | ✅ | Workspace-specific checks |
+| `run` | ✅ | ✅ | Primary use case |
+| `serve` | ✅ | ✅ | Via `--model ./path` |
+| `pull` | ✅ | ❌ | Cache only (by design) |
+| `clone` | ❌ | ✅ | HF Hub → workspace (direct download) |
+| `push` | ❌ | ✅ | Workspace → HF Hub |
+| `rm` | ✅ | ❌ | **Cache only** — use `rm -rf ./workspace` |
+
+**Why no `mlxk rm` for workspaces?**
+- Workspaces are user-managed directories (like any project folder)
+- User has full filesystem control — standard `rm -rf` is appropriate
+- Avoids accidental deletion of user data vs. cache (which is regenerable)
+- Principle: mlx-knife manages cache, user manages workspaces
+
+---
+
 ## UX Details
 
 ### list: Source Column
diff --git a/mlxk2/operations/clone.py b/mlxk2/operations/clone.py
index be3daf1..af0d3a7 100644
--- a/mlxk2/operations/clone.py
+++ b/mlxk2/operations/clone.py
@@ -23,7 +23,7 @@ from typing import Optional, Dict, Any, Tuple
 
 from .pull import pull_to_cache
 from .workspace import write_workspace_sentinel
-from ..core.cache import hf_to_cache_dir, get_current_cache_root
+from ..core.cache import hf_to_cache_dir
 from mlxk2 import __version__
 
 logger = logging.getLogger(__name__)
diff --git a/mlxk2/operations/serve.py b/mlxk2/operations/serve.py
index d0e098a..602a883 100644
--- a/mlxk2/operations/serve.py
+++ b/mlxk2/operations/serve.py
@@ -134,6 +134,20 @@ def start_server(
     # Suppress tqdm progress bars in server mode (must be set before tqdm import)
     os.environ["TQDM_DISABLE"] = "1"
     if model:
+        # Pre-validate model specification before starting server (consistency with run.py)
+        from ..core.model_resolution import resolve_model_for_operation
+        from .workspace import is_explicit_path
+        resolved_name, _, ambiguous = resolve_model_for_operation(model)
+        if ambiguous:
+            raise ValueError(
+                f"Ambiguous model specification '{model}'. Could be: {ambiguous}"
+            )
+        if not resolved_name:
+            # Model not found - give appropriate error message
+            if is_explicit_path(model):
+                raise ValueError(f"Workspace not found: {model}")
+            else:
+                raise ValueError(f"Model not found in cache: {model}")
         os.environ["MLXK2_PRELOAD_MODEL"] = model
     if max_tokens is not None:
         os.environ["MLXK2_MAX_TOKENS"] = str(max_tokens)