diff --git a/agents/completed.md b/agents/completed.md
index 74f07ea..42664b3 100644
--- a/agents/completed.md
+++ b/agents/completed.md
@@ -8,6 +8,18 @@
 
 ---
 
+# #378: discovery drops variants= from the manifest; model= shorthand breaks with ServerHandle setup param
+
+**Completed:** yes
+**Status:** DONE (2026-07-04, Claude) — two @endpoint API bugs found by the inference-endpoints #343 port. (1) `discover_functions` deduped manifest entries by (declared_module, class_name, python_name) — variants= rows share all three with their base function, so every variant was silently dropped from endpoint.lock; the stamped `name` is now part of the dedup key. (2) The `model=` shorthand counted a `ServerHandle`-annotated setup() param as a slot candidate, so the documented runtime= shape `setup(self, model: str, server: ServerHandle)` failed at decoration; ServerHandle params are now excluded from slot resolution. Regression tests for both; suite 223 passed, 1 skipped.
+
+## Tasks
+- [x] Include entry name in the discovery dedup key (discovery/discover.py).
+- [x] Skip ServerHandle-annotated setup params in `_resolve_single_slot` (api/decorators.py).
+- [x] Regression tests (manifest keeps variant entries; documented runtime= shape decorates).
+
+---
+
 # #377: HF binding files= not honored on every download path (14GB instead of ~3GB)
 
 **Completed:** yes
diff --git a/agents/progress.md b/agents/progress.md
index 9071c82..56bb4ef 100644
--- a/agents/progress.md
+++ b/agents/progress.md
@@ -7,7 +7,7 @@
 > Source for all findings: /home/fidika/cozy/python-gen-worker/AUDIT.md (2026-07-03 full-stack audit,
 > file:line evidence for every claim). Counterpart protocol/orchestrator issues: tensorhub #503-#510.
 
-next_id: 378
+next_id: 379
 
 ---
 
diff --git a/src/gen_worker/api/decorators.py b/src/gen_worker/api/decorators.py
index e5734e0..045f3e2 100644
--- a/src/gen_worker/api/decorators.py
+++ b/src/gen_worker/api/decorators.py
@@ -255,6 +255,19 @@ def _setup_params(cls: type) -> Optional[dict[str, inspect.Parameter]]:
     return {p.name: p for p in params}
 
 
+def _is_server_handle_param(p: inspect.Parameter) -> bool:
+    """True when the param is annotated ``ServerHandle`` (a runtime= injection
+    target, not a model slot)."""
+    ann = p.annotation
+    if ann is inspect.Parameter.empty:
+        return False
+    from ..runtimes.server import ServerHandle
+
+    if ann is ServerHandle:
+        return True
+    return isinstance(ann, str) and ann.split(".")[-1] == "ServerHandle"
+
+
 def _resolve_single_slot(
     cls: type,
     models: dict[str, Binding],
@@ -269,6 +282,7 @@ def _resolve_single_slot(
         named = [
             n for n, p in setup_kwargs.items()
             if p.kind in (inspect.Parameter.POSITIONAL_OR_KEYWORD, inspect.Parameter.KEYWORD_ONLY)
+            and not _is_server_handle_param(p)
         ]
         if len(named) == 1:
             models[named[0]] = binding
diff --git a/src/gen_worker/discovery/discover.py b/src/gen_worker/discovery/discover.py
index 1a64f48..9ded91e 100644
--- a/src/gen_worker/discovery/discover.py
+++ b/src/gen_worker/discovery/discover.py
@@ -390,13 +390,16 @@ def discover_functions(root: Optional[Path] = None, *, main_module: str | None =
         ) from e
 
     functions: List[Dict[str, Any]] = []
-    seen: Set[Tuple[str, str, str]] = set()
+    seen: Set[Tuple[str, str, str, str]] = set()
     for f in found:
         for entry in _extract_entries(f.obj, f.walked_module):
+            # name is part of the key: variants= share (module, class,
+            # python_name) with their base function but stamp distinct names.
             key = (
                 entry.get("declared_module", entry.get("module", "")),
                 entry.get("class_name", ""),
                 entry.get("python_name", ""),
+                entry.get("name", ""),
             )
             if key in seen:
                 continue
diff --git a/tests/test_discovery_and_decorators.py b/tests/test_discovery_and_decorators.py
index 6cc4512..46426df 100644
--- a/tests/test_discovery_and_decorators.py
+++ b/tests/test_discovery_and_decorators.py
@@ -422,3 +422,56 @@ def test_from_scratch_example_uses_publish_contract() -> None:
     assert s.output_mode == "single"
     assert not s.is_async_gen
     assert s.output_type is mod.FromScratchResult
+
+
+def test_manifest_keeps_variant_entries(tmp_pkg: Path) -> None:
+    """variants= rows share (module, class, python_name) with their base
+    function; the manifest dedup must key on the stamped name too."""
+    from gen_worker.discovery.discover import discover_functions
+
+    pkg = tmp_pkg / "ep_variants"
+    pkg.mkdir()
+    (pkg / "__init__.py").write_text("")
+    (pkg / "main.py").write_text(textwrap.dedent("""
+        import msgspec
+        from gen_worker import HF, RequestContext, Resources, endpoint
+
+        class In_(msgspec.Struct):
+            x: str = ""
+
+        class Out_(msgspec.Struct):
+            y: str
+
+        @endpoint(
+            model=HF("o/base", dtype="bf16"),
+            resources=Resources(vram_gb=24),
+            variants={"generate_fp8": (HF("o/base-fp8"), Resources(vram_gb=14))},
+        )
+        class Gen:
+            def setup(self, model: str) -> None:
+                self.model = model
+
+            def generate(self, ctx: RequestContext, data: In_) -> Out_:
+                return Out_(y="ok")
+    """))
+
+    fns = discover_functions(tmp_pkg, main_module="ep_variants.main")
+    assert sorted(f["name"] for f in fns) == ["generate", "generate-fp8"]
+
+
+def test_model_shorthand_skips_server_handle_setup_param() -> None:
+    """runtime= endpoints inject a ServerHandle into setup(); the model=
+    shorthand must resolve the slot from the remaining parameter."""
+    from gen_worker.runtimes.server import ServerHandle
+
+    @endpoint(model=HF("o/llm"), resources=Resources(vram_gb=40), runtime="vllm")
+    class Chat:
+        def setup(self, model: str, server: ServerHandle) -> None:
+            self.base_url = server.base_url
+
+        def complete(self, ctx: RequestContext, data: _In) -> _Out:
+            return _Out(result="")
+
+    (s,) = extract_specs(Chat)
+    assert list(s.models) == ["model"]
+    assert s.models["model"].ref == "o/llm"