fix(health): restore message field in public health view (SDK contract)

src/serving/api/main.py

@@ -740,12 +740,14 @@ def _public_health_view(payload: dict) -> dict:
 
     `/v1/health` is auth-exempt, yet each component's `message`/`metrics` carry
     recon: `f"Kafka unavailable: {exc}"` (internal hostnames/URLs), broker/topic
-    counts, cluster sizes, and the serving backend's identity. Keep the overall
-    status, each component's name/status/source (the documented contract), and
-    only the allowlisted operational gauges (pool utilization, data freshness);
-    drop `message` and every non-allowlisted metric. Agents still get the
-    per-component status they use to caveat freshness, and pool-utilization
-    observability (its only exposure — not on /metrics) is preserved.
+    counts, cluster sizes, and the serving backend's identity. Neutralise the
+    `message` text (it is where every `{exc}` and count lives) and keep only the
+    allowlisted operational gauges (pool utilization, data freshness); keep the
+    overall status and each component's name/status/source. The field *shape* is
+    unchanged — `message` stays a string and `metrics` a dict — because the SDK's
+    `HealthComponent` model requires them; only the sensitive *content* is
+    dropped. Agents still get per-component status to caveat freshness, and
+    pool-utilization observability (its only exposure — not on /metrics) is kept.
     """
     components = []
     for c in payload.get("components", []):
@@ -754,6 +756,9 @@ def _public_health_view(payload: dict) -> dict:
             {
                 "name": c.get("name"),
                 "status": c.get("status"),
+                # Blanked, not dropped: the SDK HealthComponent requires a string
+                # `message`, but its content (error text, counts) is the leak.
+                "message": "",
                 "source": c.get("source"),
                 "metrics": metrics,
             }

tests/unit/test_health_public_view.py

@@ -9,6 +9,8 @@
 
 from __future__ import annotations
 
+from agentflow.models import HealthStatus
+
 from src.serving.api.main import _public_health_view
 
 _RAW_PAYLOAD = {
@@ -40,10 +42,23 @@
 }
 
 
-def test_view_keeps_only_name_status_source_per_component() -> None:
+def test_view_component_fields_match_sdk_contract() -> None:
     view = _public_health_view(_RAW_PAYLOAD)
     for component in view["components"]:
-        assert set(component) == {"name", "status", "source", "metrics"}
+        assert set(component) == {"name", "status", "message", "source", "metrics"}
+
+
+def test_view_still_validates_against_sdk_health_model() -> None:
+    # The SDK's HealthComponent requires string `message`/`source` and the
+    # top-level HealthStatus requires `checked_at`. The push-only E2E lane
+    # validates the /v1/health payload through this model — dropping `message`
+    # broke it. Pin the contract here so it fails in unit tests, not only e2e.
+    HealthStatus.model_validate(_public_health_view(_RAW_PAYLOAD))
+
+
+def test_view_blanks_message_but_keeps_it_a_string() -> None:
+    for component in _public_health_view(_RAW_PAYLOAD)["components"]:
+        assert component["message"] == ""
 
 
 def test_view_drops_error_strings_counts_and_backend_identity() -> None:
@@ -95,5 +110,5 @@ def test_view_tolerates_missing_optional_keys() -> None:
     view = _public_health_view({"status": "healthy", "components": [{"name": "x", "status": "ok"}]})
     assert view == {
         "status": "healthy",
-        "components": [{"name": "x", "status": "ok", "source": None, "metrics": {}}],
+        "components": [{"name": "x", "status": "ok", "message": "", "source": None, "metrics": {}}],
     }