diff --git a/src/ucode/cli.py b/src/ucode/cli.py
index 3d4fb87..4668e5d 100644
--- a/src/ucode/cli.py
+++ b/src/ucode/cli.py
@@ -1109,6 +1109,8 @@ def codex_router_hook_cmd(
                     }
                 )
             )
+        # When matched is None the harness didn't report the subagent model —
+        # the PreToolUse hook already injected the routed model, so emit nothing.
         return
     if event != "route-subagent" or not host:
         return
@@ -1180,6 +1182,8 @@ def claude_router_hook_cmd(
                     }
                 )
             )
+        # When matched is None the harness didn't report the subagent model —
+        # the PreToolUse hook already injected the routed model, so emit nothing.
         return
     if event != "route-subagent" or not host:
         return
diff --git a/src/ucode/databricks.py b/src/ucode/databricks.py
index c912e2c..f7acfd8 100644
--- a/src/ucode/databricks.py
+++ b/src/ucode/databricks.py
@@ -1374,6 +1374,20 @@ def list_model_services(
     return [], last_reason or "model-services listing returned no models"
 
 
+def _prefer_opus_4_8(models: dict[str, str], all_ids: list[str]) -> None:
+    """Swap the opus slot to claude-opus-4-8 when it's available.
+
+    Discovery picks the newest opus (opus-5) but smart routing's
+    CLAUDE_ROUTE_ARMS require claude-opus-4-8. Pin to 4-8 when both
+    exist so the routing availability check passes.
+    """
+    opus = models.get("opus")
+    if opus and "claude-opus-5" in opus:
+        opus_48 = next((m for m in all_ids if "claude-opus-4-8" in m), None)
+        if opus_48:
+            models["opus"] = opus_48
+
+
 def discover_model_services(
     workspace: str, token: str
 ) -> tuple[dict[str, str], list[str], list[str], list[str], str | None]:
@@ -1404,6 +1418,12 @@ def discover_model_services(
         )
         if candidates:
             claude_models[family] = candidates[0]
+    # Smart routing's CLAUDE_ROUTE_ARMS require claude-opus-4-8, but the
+    # newest-wins sort above picks opus-5 when both exist — making the
+    # routing availability check fail. Pin opus-4-8 when it's available so
+    # routing works with the currently-deployed task_v1 router. Revert to
+    # newest-wins once the router accepts opus-5 (PR databricks-eng/universe#2365446).
+    _prefer_opus_4_8(claude_models, ids)
 
     codex_models = [m for m in ids if "gpt-" in m]
     gemini_models = sorted([m for m in ids if "gemini-" in m], key=model_version_sort_key)
@@ -2211,6 +2231,8 @@ def discover_claude_models(workspace: str, token: str) -> tuple[dict[str, str],
         )
         if candidates:
             result[family] = candidates[0]
+    # Same opus-4-8 pin as discover_model_services — see comment there.
+    _prefer_opus_4_8(result, raw_ids)
     if result:
         return result, None
     if not raw_ids:
diff --git a/src/ucode/smart_routing/claude_routing.py b/src/ucode/smart_routing/claude_routing.py
index 005eb28..79c9506 100644
--- a/src/ucode/smart_routing/claude_routing.py
+++ b/src/ucode/smart_routing/claude_routing.py
@@ -187,8 +187,13 @@ def clear_routing_artifacts() -> None:
 def _claude_model_id(model: str) -> str:
     """The model id Claude Code should launch the subagent with.
 
-    The router returns a routable workspace id (e.g.
-    ``system.ai.claude-opus-4-8``); Claude Code's ``Agent`` tool ``model`` field
-    accepts that id verbatim through the gateway, so pass it through unchanged.
+    Claude Code's ``Agent`` tool ``model`` field accepts only short family
+    names (``sonnet``, ``opus``, ``haiku``, ``fable``), not full workspace
+    ids. Map the router's pick (e.g. ``system.ai.claude-sonnet-5``) back to
+    its family name.
     """
+    normalized = _normalize_model(model)
+    for family in ("fable", "opus", "sonnet", "haiku"):
+        if f"claude-{family}-" in normalized:
+            return family
     return model
diff --git a/src/ucode/smart_routing/routing.py b/src/ucode/smart_routing/routing.py
index 594fc0e..80df358 100644
--- a/src/ucode/smart_routing/routing.py
+++ b/src/ucode/smart_routing/routing.py
@@ -202,16 +202,16 @@ def route_spawn_tool(
     tool_input = payload.get("tool_input")
     if not isinstance(tool_input, dict):
         return None
-    # Derive the routing task from the first available plaintext field. `message`
-    # carries the actual subagent task content — prefer it when present and a
-    # plaintext string (Codex encrypts it at send-time, but the PreToolUse hook
-    # fires before that, so it may be readable here). When `message` is an
-    # encrypted dict (or absent), fall back to `task_name` / `agent_name`
-    # (weaker labels), then the generic default.
+    # Derive the routing task from the first available plaintext field. The
+    # harness-specific names are tried in order: `prompt`/`description` (Claude
+    # Code's Agent tool), `message` (Codex's spawn_agent — encrypted at
+    # send-time but readable here because the PreToolUse hook fires before
+    # that), then `task_name` / `agent_name` (weaker labels), then the generic
+    # default.
     task = next(
         (
             value
-            for field in ("message", "task_name", "agent_name")
+            for field in ("prompt", "description", "message", "task_name", "agent_name")
             if isinstance(value := tool_input.get(field), str) and value
         ),
         default_task_label,
@@ -269,12 +269,17 @@ def record_subagent_start(
         "at": time.time(),
     }
     if decision is not None:
+        # When the harness doesn't report the subagent's model (actual_model is
+        # None), we can't verify the match — record None rather than a false
+        # mismatch. The PreToolUse hook already injected the routed model, so
+        # routing still worked; the reconciliation is observability, not enforcement.
+        matches = None if actual_model is None else decision.get("requested_model") == actual_model
         record.update(
             {
                 "decision_id": decision.get("decision_id"),
                 "router_model": decision.get("router_model"),
                 "requested_model": decision.get("requested_model"),
-                "matches_router_decision": decision.get("requested_model") == actual_model,
+                "matches_router_decision": matches,
             }
         )
     _append_jsonl(audit_path, record)
diff --git a/tests/test_claude_routing.py b/tests/test_claude_routing.py
index 49f77af..d30ee66 100644
--- a/tests/test_claude_routing.py
+++ b/tests/test_claude_routing.py
@@ -155,21 +155,20 @@ def test_spawn_rewrite_injects_routed_model(monkeypatch):
 
     hook = output["hookSpecificOutput"]
     # The rationale is surfaced in the systemMessage (shown to the user), not
-    # only in permissionDecisionReason.
+    # only in permissionDecisionReason. The model field is the short family
+    # name ("opus") that Claude Code's Agent tool schema accepts.
     assert output["systemMessage"] == (
-        "Using Smart Routing. Routing to system.ai.claude-opus-4-8. "
-        "Deep exploration needs the strongest model."
+        "Using Smart Routing. Routing to opus. Deep exploration needs the strongest model."
     )
     assert hook["permissionDecision"] == "allow"
     assert hook["updatedInput"] == {
         "subagent_type": "Explore",
         "prompt": "map the codebase",
         "description": "explore",
-        "model": "system.ai.claude-opus-4-8",
+        "model": "opus",
     }
     assert hook["permissionDecisionReason"] == (
-        "Using Smart Routing. Routing to system.ai.claude-opus-4-8. "
-        "Deep exploration needs the strongest model."
+        "Using Smart Routing. Routing to opus. Deep exploration needs the strongest model."
     )
 
 
@@ -192,7 +191,41 @@ def test_task_tool_alias_is_routed(monkeypatch):
         available_models=["system.ai.claude-opus-4-8", "system.ai.claude-sonnet-5"],
     )
 
-    assert output["hookSpecificOutput"]["updatedInput"]["model"] == "system.ai.claude-sonnet-5"
+    assert output["hookSpecificOutput"]["updatedInput"]["model"] == "sonnet"
+
+
+def test_spawn_routes_on_prompt_not_generic_label(monkeypatch):
+    # Claude Code's Agent tool puts the subagent task in `prompt`/`description`,
+    # not `message`. The router should receive the real prompt, not the generic
+    # "Claude Code subagent task" fallback.
+    captured = {}
+
+    def fake_decision(workspace, token, task, models, **kwargs):
+        captured["task"] = task
+        return (
+            claude_routing.RoutingDecision(
+                model="system.ai.claude-sonnet-5", raw_model="claude-sonnet-5"
+            ),
+            None,
+        )
+
+    monkeypatch.setattr(claude_routing, "request_routing_decision", fake_decision)
+
+    claude_routing.route_pre_tool_use(
+        {
+            "tool_name": "Agent",
+            "tool_input": {
+                "subagent_type": "Explore",
+                "prompt": "find all Python entry points",
+                "description": "explore the repo",
+            },
+        },
+        workspace=WS,
+        token="token",
+        available_models=["system.ai.claude-opus-4-8", "system.ai.claude-sonnet-5"],
+    )
+
+    assert captured["task"] == "find all Python entry points"
 
 
 def test_non_spawn_tool_has_no_opinion():
@@ -255,14 +288,52 @@ def test_decision_is_reconciled_with_actual_subagent_model(tmp_path, monkeypatch
         audit_decision=True,
     )
     record = claude_routing.record_subagent_start(
-        {"session_id": "s1", "agent_id": "a1", "model": "system.ai.claude-opus-4-8"}
+        {"session_id": "s1", "agent_id": "a1", "model": "opus"}
     )
 
     assert record["router_model"] == "claude-opus-4-8"
-    assert record["requested_model"] == "system.ai.claude-opus-4-8"
+    assert record["requested_model"] == "opus"
     assert record["matches_router_decision"] is True
 
 
+def test_subagent_start_without_model_reports_unknown_not_mismatch(tmp_path, monkeypatch):
+    # Claude Code's SubagentStart event does not include the subagent's model,
+    # so reconciliation can't verify the match — record None (unknown) rather
+    # than a false mismatch. The PreToolUse hook already injected the model.
+    decisions = tmp_path / "decisions.jsonl"
+    audit = tmp_path / "audit.jsonl"
+    monkeypatch.setattr(claude_routing, "DECISIONS_PATH", decisions)
+    monkeypatch.setattr(claude_routing, "AUDIT_PATH", audit)
+    monkeypatch.setattr(
+        claude_routing,
+        "request_routing_decision",
+        lambda *args, **kwargs: (
+            claude_routing.RoutingDecision(
+                model="system.ai.claude-sonnet-5", raw_model="claude-sonnet-5"
+            ),
+            None,
+        ),
+    )
+
+    claude_routing.route_pre_tool_use(
+        {
+            "session_id": "s2",
+            "tool_name": "Agent",
+            "tool_input": {"subagent_type": "Explore", "prompt": "x"},
+        },
+        workspace=WS,
+        token="token",
+        available_models=["system.ai.claude-opus-4-8", "system.ai.claude-sonnet-5"],
+        audit_decision=True,
+    )
+    record = claude_routing.record_subagent_start(
+        {"session_id": "s2", "agent_id": "a2", "model": None}
+    )
+
+    assert record["requested_model"] == "sonnet"
+    assert record["matches_router_decision"] is None
+
+
 def test_launch_task_uses_positional_prompt():
     assert claude_routing._launch_routing_task(["fix the parser"]) == "fix the parser"