fix(tests): fix golden run a fixtures

Taskfile.yml

@@ -111,7 +111,7 @@ tasks:
     deps: [install-dev]
     cmds:
       - task: verify-lab-smoke
-      #- task: verify-golden-runs
+      - task: verify-golden-runs
       - task: verify-tutorials
 
   ci:

ctl/src/mas/ctl/workspace/config.py

@@ -184,14 +184,21 @@ def runtime_profile_path(self) -> Path | None:
         return candidate if candidate.is_file() else None
 
     def resolve_library_path(self, lib_ref: str) -> Path | None:
-        """Resolve ``team:bundle/sub`` via manifest_libraries."""
+        """Resolve ``team:bundle/sub`` via manifest_libraries.
+
+        The ``base`` path in manifest_libraries may be:
+        - Absolute: ``/abs/path`` → used directly
+        - Home-relative: ``~/.config/…`` → expanded via Path.expanduser()
+        - Relative: ``./infra`` → resolved relative to the workspace root
+        """
         if ":" not in lib_ref:
             return None
         lib, rest = lib_ref.split(":", 1)
         base = self.manifest_libraries.get(lib)
         if not base or self._path is None:
             return None
-        root = (self._path / base).resolve()
+        base_path = Path(base).expanduser()
+        root = (base_path if base_path.is_absolute() else self._path / base_path).resolve()
         candidate = (root / rest).with_suffix(".yaml")
         if candidate.is_file():
             return candidate

examples/sample-workspace/config.yaml

@@ -8,6 +8,14 @@
 #
 # For live OpenAI/production infra, see:
 #   docs/tutorials/00-environment-setup/config.openai.example.yaml
+#
+# For private/team proxy bundles (e.g. claris:llm-proxy), add manifest_libraries:
+#
+#   manifest_libraries:
+#     claris: ~/.config/mas/infra   # resolves claris:llm-proxy → ~/.config/mas/infra/llm-proxy.yaml
+#
+# Then set MAS_INFRA_REFS=claris:llm-proxy for real-LLM runs.
+# Template: config/infra/claris-proxy.example.yaml → copy to ~/.config/mas/infra/llm-proxy.yaml
 
 mas_ctl:
   flavour: local

scripts/capture_golden_run.py

@@ -73,17 +73,21 @@ def _capture_one(
     # cache entry with stale event counts and return cached results instead of
     # running the agent.  The test fixture does the same via monkeypatch.
     os.environ["MAS_TRACE_CACHE"] = str(trace_cache)
-    # Isolate from the user's personal ~/.config/mas/config.yaml so internal
-    # infra refs (e.g. claris:llm-proxy) don't bleed into the OSS capture.
+    # Isolate XDG so the user's personal ~/.config/mas/config.yaml (which may
+    # point at a private claris:llm-proxy) cannot bleed into the OSS capture.
     os.environ["XDG_CONFIG_HOME"] = str(tmp / "xdg-config")
-    # Default to standard:llm-proxy, which reads LLM_PROXY_API_BASE from the
-    # environment (falling back to https://api.openai.com/v1). The caller can
-    # override by setting MAS_INFRA_REFS before running.
-    if not os.environ.get("MAS_INFRA_REFS"):
-        os.environ["MAS_INFRA_REFS"] = "standard:llm-proxy"
+    # Pin the workspace so find_workspace_file() always finds the OSS workspace
+    # config (infra_refs: [standard:mock-llm]).  Without this, the walk stops at
+    # .git (no config.yaml at repo root), falls back to the now-isolated XDG
+    # path, finds nothing, and resolve_infra_refs falls through to
+    # standard:production — hitting a real LLM and causing a 401 or wrong event
+    # count.  MAS_INFRA_REFS can still override this for real-LLM captures.
+    if "MAS_WORKSPACE_ROOT" not in os.environ:
+        os.environ["MAS_WORKSPACE_ROOT"] = str(ROOT / "examples" / "sample-workspace")
     print(
-        f"  infra:   MAS_INFRA_REFS={os.environ.get('MAS_INFRA_REFS')!r}",
-        f"  api_base: LLM_PROXY_API_BASE={os.environ.get('LLM_PROXY_API_BASE', '<not set — will use OpenAI default>')!r}",
+        f"  workspace: MAS_WORKSPACE_ROOT={os.environ['MAS_WORKSPACE_ROOT']!r}",
+        f"  infra:     MAS_INFRA_REFS={os.environ.get('MAS_INFRA_REFS', '<not set — workspace infra_refs used>')!r}",
+        f"  api_base:  LLM_PROXY_API_BASE={os.environ.get('LLM_PROXY_API_BASE', '<not set>')!r}",
         sep="\n",
         file=sys.stderr,
     )

tests/fixtures/golden-runs/design-space/cache-backup/design-space/240c01f5b2a41ca79519/result.json

@@ -1,6 +1,6 @@
 {
   "status": "ok",
-  "elapsed_ms": 1135.0,
+  "elapsed_ms": 1097.7,
   "error": "",
-  "executed_at": "2026-07-06T01:02:36.135611Z"
+  "executed_at": "2026-07-06T11:50:37.912704Z"
 }

tests/fixtures/golden-runs/design-space/cache-backup/design-space/240c01f5b2a41ca79519/run.json

@@ -1,6 +1,6 @@
 {
   "run_hash": "240c01f5b2a41ca79519",
-  "created_at": "2026-07-06T01:02:36.134972Z",
+  "created_at": "2026-07-06T11:50:37.912082Z",
   "item_id": "1",
   "run_idx": 0,
   "model": "",

tests/fixtures/golden-runs/design-space/cache-backup/design-space/240c01f5b2a41ca79519/run_info.json

@@ -4,7 +4,7 @@
   "scenario": "mock-tools",
   "item_id": "1",
   "run_idx": 0,
-  "referenced_at": "2026-07-06T01:02:36.135723Z",
+  "referenced_at": "2026-07-06T11:50:37.912797Z",
   "mas": {
     "app": "qa-agent",
     "mas_ref": "/Users/augjorda/repos/outshift-open/mas-lab/docs/tutorials/01-building-an-agent/agent.yaml",