hermes: keep quality routes on hosted models

services/hermes/switchyard-configmap.yaml

@@ -303,7 +303,7 @@ data:
     classifier_target = "classifier"
     # Switchyard falls through this list after a request-local target failure.
     # Keep both xhigh providers first so recovery can escalate, never downgrade.
-    targets = ["claude_opus_xhigh", "codex_sol_xhigh", "claude_sonnet_high", "codex_sol_high", "claude_opus_high", "codex_terra_high", "claude_sonnet_medium", "codex_sol_medium", "codex_terra_medium", "local_qwen_medium", "claude_haiku_low", "codex_luna_low", "codex_terra_low", "local_qwen_low"]
+    targets = ["claude_opus_xhigh", "codex_sol_xhigh", "claude_sonnet_high", "codex_sol_high", "claude_opus_high", "codex_terra_high", "claude_sonnet_medium", "codex_sol_medium", "codex_terra_medium", "claude_haiku_low", "codex_luna_low", "codex_terra_low"]
     default_target = "claude_sonnet_high"
     session_affinity = false
     recent_turn_window = 12
@@ -331,23 +331,24 @@ data:
     the fastest target at the safety floor, never below it. An explicit
     available provider or model wins at the safety floor.
 
-    3. Use local Qwen only at low or medium for low-risk summaries, formatting,
-    lookup, and continuity. Prefer Codex for implementation, debugging, tests,
-    commands, and repository work. Prefer Claude for diagnosis, architecture,
-    ambiguity, synthesis, risk analysis, adversarial analysis, and independent
-    review. Anthropic means Claude; OpenAI means Codex. Never select a provider
-    stated to be unavailable, failed, exhausted, rate-limited, or out of
-    capacity; use the other provider at the same floor.
+    3. This quality-first triage route must use a hosted Codex or Claude target.
+    The local Qwen model remains the routing classifier, but its 32K context and
+    tool protocol are not eligible for foreground triage execution. Prefer
+    Codex for implementation, debugging, tests, commands, and repository work.
+    Prefer Claude for diagnosis, architecture, ambiguity, synthesis, risk
+    analysis, adversarial analysis, and independent review. Anthropic means
+    Claude; OpenAI means Codex. Never select a provider stated to be
+    unavailable, failed, exhausted, rate-limited, or out of capacity; use the
+    other provider at the same floor.
 
     4. Map exactly: Codex low=codex_luna_low, medium=codex_terra_medium,
     high=codex_sol_high, xhigh=codex_sol_xhigh; Claude low=claude_haiku_low,
     medium=claude_sonnet_medium, high=claude_sonnet_high,
-    xhigh=claude_opus_xhigh; local low=local_qwen_low and local
-    medium=local_qwen_medium. Re-evaluate every boundary and resolve "continue"
+    xhigh=claude_opus_xhigh. Re-evaluate every boundary and resolve "continue"
     or "do it" from recent context.
     """
     response_schema = '''
-    {"type":"object","properties":{"decision":{"type":"object","properties":{"target":{"type":"string","enum":["claude_sonnet_high","codex_sol_high","codex_terra_medium","claude_sonnet_medium","codex_terra_high","codex_sol_medium","claude_opus_high","codex_sol_xhigh","claude_opus_xhigh","codex_luna_low","claude_haiku_low","local_qwen_medium","local_qwen_low","codex_terra_low"]}},"required":["target"],"additionalProperties":false}},"required":["decision"],"additionalProperties":false}
+    {"type":"object","properties":{"decision":{"type":"object","properties":{"target":{"type":"string","enum":["claude_sonnet_high","codex_sol_high","codex_terra_medium","claude_sonnet_medium","codex_terra_high","codex_sol_medium","claude_opus_high","codex_sol_xhigh","claude_opus_xhigh","codex_luna_low","claude_haiku_low","codex_terra_low"]}},"required":["target"],"additionalProperties":false}},"required":["decision"],"additionalProperties":false}
     '''
 
     [routes.auto_deep.policy]
@@ -361,7 +362,7 @@ data:
     classifier_target = "classifier"
     # Switchyard falls through this list after a request-local target failure.
     # Keep both xhigh providers first so recovery can escalate, never downgrade.
-    targets = ["codex_sol_xhigh", "claude_opus_xhigh", "codex_sol_high", "claude_opus_high", "claude_sonnet_high", "codex_terra_high", "codex_sol_medium", "claude_sonnet_medium", "codex_terra_medium", "local_qwen_medium", "codex_luna_low", "claude_haiku_low", "codex_terra_low", "local_qwen_low"]
+    targets = ["codex_sol_xhigh", "claude_opus_xhigh", "codex_sol_high", "claude_opus_high", "claude_sonnet_high", "codex_terra_high", "codex_sol_medium", "claude_sonnet_medium", "codex_terra_medium", "codex_luna_low", "claude_haiku_low", "codex_terra_low"]
     default_target = "codex_sol_high"
     session_affinity = false
     recent_turn_window = 16
@@ -389,23 +390,24 @@ data:
     the fastest target at the safety floor, never below it. An explicit
     available provider or model wins at the safety floor.
 
-    3. Use local Qwen only at low or medium for low-risk lookup, formatting, and
-    continuity. Prefer Codex for implementation, debugging, tests, commands,
-    and repository changes. Prefer Claude for architecture, ambiguity,
-    synthesis, risk analysis, adversarial analysis, and independent review.
-    Anthropic means Claude; OpenAI means Codex. Never select a provider stated
-    to be unavailable, failed, exhausted, rate-limited, or out of capacity; use
-    the other provider at the same floor.
+    3. This owner-only engineering route must use a hosted Codex or Claude
+    target. The local Qwen model remains the routing classifier, but its 32K
+    context and tool protocol are not eligible for foreground agent execution.
+    Prefer Codex for implementation, debugging, tests, commands, and repository
+    changes. Prefer Claude for architecture, ambiguity, synthesis, risk
+    analysis, adversarial analysis, and independent review. Anthropic means
+    Claude; OpenAI means Codex. Never select a provider stated to be
+    unavailable, failed, exhausted, rate-limited, or out of capacity; use the
+    other provider at the same floor.
 
     4. Map exactly: Codex low=codex_luna_low, medium=codex_terra_medium,
     high=codex_sol_high, xhigh=codex_sol_xhigh; Claude low=claude_haiku_low,
     medium=claude_sonnet_medium, high=claude_sonnet_high,
-    xhigh=claude_opus_xhigh; local low=local_qwen_low and local
-    medium=local_qwen_medium. Re-evaluate every boundary and resolve "continue"
+    xhigh=claude_opus_xhigh. Re-evaluate every boundary and resolve "continue"
     or "do it" from recent context.
     """
     response_schema = '''
-    {"type":"object","properties":{"decision":{"type":"object","properties":{"target":{"type":"string","enum":["codex_sol_high","claude_opus_high","claude_sonnet_high","codex_terra_high","codex_sol_xhigh","claude_opus_xhigh","codex_terra_medium","claude_sonnet_medium","codex_sol_medium","codex_luna_low","claude_haiku_low","local_qwen_medium","local_qwen_low","codex_terra_low"]}},"required":["target"],"additionalProperties":false}},"required":["decision"],"additionalProperties":false}
+    {"type":"object","properties":{"decision":{"type":"object","properties":{"target":{"type":"string","enum":["codex_sol_high","claude_opus_high","claude_sonnet_high","codex_terra_high","codex_sol_xhigh","claude_opus_xhigh","codex_terra_medium","claude_sonnet_medium","codex_sol_medium","codex_luna_low","claude_haiku_low","codex_terra_low"]}},"required":["target"],"additionalProperties":false}},"required":["decision"],"additionalProperties":false}
     '''
 
     [routes.auto_maximum.policy]

services/hermes/switchyard-deployment.yaml

@@ -22,7 +22,7 @@ spec:
       labels:
         app: hermes-switchyard
       annotations:
-        ai.bstein.dev/config-rev: "20260812-auto-response-budget"
+        ai.bstein.dev/config-rev: "20260812-quality-route-local-exclusion"
         prometheus.io/scrape: "true"
         prometheus.io/port: "9005"
         prometheus.io/path: /metrics

testing/tests/test_hermes_chat_quality.py

@@ -855,6 +855,16 @@ def test_local_flux_runtime_and_gpu_handoff_are_flux_managed():
         leading_targets = set(routes[route_name]["targets"][:2])
         assert leading_targets == {"codex_sol_xhigh", "claude_opus_xhigh"}
         assert "max_output_tokens" not in routes[route_name]
+    for route_name in ("auto_deep", "auto_maximum"):
+        targets = routes[route_name]["targets"]
+        selector_targets = routes[route_name]["response_schema"]
+        assert not any(target.startswith("local_") for target in targets)
+        assert "local_qwen" not in selector_targets
+        assert "not eligible for foreground" in routes[route_name]["prompt"]
+    for route_name in ("auto_fast", "auto_balanced", "manual_local_qwen"):
+        assert any(
+            target.startswith("local_") for target in routes[route_name]["targets"]
+        )
     assert "max_output_tokens" not in routes["worker_auto_maximum"]
     model_gate = _documents(HERMES / "model-gate-configmap.yaml")[0]["data"][
         "model_gate.py"