hermes: normalize Codex response budgets

services/hermes/agent-deployment.yaml

@@ -25,7 +25,7 @@ spec:
         ai.bstein.dev/execution: Hermes Kanban with durable direct Codex and Claude Code CLI workers
         ai.bstein.dev/model-policy: Jetson-assisted AUTO routing, low through xhigh, cross-provider fallback
         ai.bstein.dev/placement: rpi5 preferred; Jetson deferred until state storage is available
-        ai.bstein.dev/config-rev: "20260811-switchyard-responses-output"
+        ai.bstein.dev/config-rev: "20260812-codex-budget-normalization"
         vault.hashicorp.com/agent-inject: "true"
         vault.hashicorp.com/role: hermes-agent
         vault.hashicorp.com/agent-inject-secret-anthropic-token: kv/data/atlas/hermes/agent-tokens

services/hermes/scripts/codex_broker.py

@@ -124,6 +124,17 @@ def _validate_payload(payload: Any) -> dict[str, Any]:
         payload["input"] = [response_input]
     elif not isinstance(response_input, list) or not response_input:
         raise ValueError("non-empty Responses input list required")
+    # Switchyard uses ``max_output_tokens`` to bound the tiny classifier call,
+    # but its Responses translation can also copy that internal option onto the
+    # selected provider request.  The first-party subscription Codex endpoint
+    # does not accept any of the public API token-budget aliases.  Let Codex use
+    # its own response budget instead of turning a healthy fallback into a 400.
+    for token_budget_key in (
+        "max_output_tokens",
+        "max_completion_tokens",
+        "max_tokens",
+    ):
+        payload.pop(token_budget_key, None)
     # Tenant conversations must not enter the owner's server-side history.
     payload["store"] = False
     payload["stream"] = True

testing/tests/test_hermes_chat_quality.py

@@ -525,10 +525,16 @@ def test_codex_broker_auth_and_request_contract(tmp_path: Path, monkeypatch):
             "input": "route this chat turn",
             "store": True,
             "stream": False,
+            "max_output_tokens": 96,
+            "max_completion_tokens": 96,
+            "max_tokens": 96,
         }
     )
     assert payload["store"] is False
     assert payload["stream"] is True
+    assert "max_output_tokens" not in payload
+    assert "max_completion_tokens" not in payload
+    assert "max_tokens" not in payload
     assert payload["input"] == [
         {
             "type": "message",