hermes(agent): route native Codex through CLI runtime

services/hermes/agent-configmap.yaml

@@ -12,6 +12,9 @@ data:
       provider: openai-codex
       default: gpt-5.6-terra
       model: gpt-5.6-terra
+      # Reuse the owner's authenticated Codex CLI instead of maintaining a
+      # second rotating OAuth token in Hermes' provider store.
+      openai_runtime: codex_app_server
 
     fallback_providers:
       - provider: anthropic

services/hermes/agent-deployment.yaml

@@ -24,7 +24,7 @@ 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-stock-supervisor-probe"
+        ai.bstein.dev/config-rev: "20260811-codex-app-server-routing"
         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
@@ -241,6 +241,33 @@ spec:
           resources:
             requests: {cpu: 25m, memory: 64Mi}
             limits: {cpu: 100m, memory: 128Mi}
+        - name: patch-codex-runtime
+          image: registry.bstein.dev/bstein/hermes-agent@sha256:46de11122226a68adad6d902e81ef9b4b53deb0433dcc61f57fa43fae2e3a3a5
+          imagePullPolicy: IfNotPresent
+          command:
+            - /opt/hermes/.venv/bin/python
+            - /opt/coordinator/patch_codex_runtime.py
+            - /opt/hermes/hermes_cli/runtime_provider.py
+            - /patched/runtime_provider.py
+            - /opt/hermes/agent/transports/codex_app_server_session.py
+            - /patched/codex_app_server_session.py
+            - /opt/hermes/agent/codex_runtime.py
+            - /patched/codex_runtime.py
+          securityContext:
+            allowPrivilegeEscalation: false
+            runAsUser: 10000
+            runAsGroup: 10000
+            seccompProfile:
+              type: RuntimeDefault
+          volumeMounts:
+            - name: coordinator
+              mountPath: /opt/coordinator
+              readOnly: true
+            - name: codex-runtime-patch
+              mountPath: /patched
+          resources:
+            requests: {cpu: 25m, memory: 64Mi}
+            limits: {cpu: 100m, memory: 128Mi}
         - name: bootstrap-coordinator
           image: registry.bstein.dev/bstein/hermes-agent@sha256:46de11122226a68adad6d902e81ef9b4b53deb0433dcc61f57fa43fae2e3a3a5
           imagePullPolicy: IfNotPresent
@@ -274,6 +301,9 @@ spec:
             - sh
             - -ec
             - |
+              set -a
+              . /opt/data/.env
+              set +a
               /opt/hermes/.venv/bin/python /opt/coordinator/configure_agent_clients.py
               /opt/hermes/.venv/bin/python /opt/coordinator/migrate_herdr_state.py
           env:
@@ -352,6 +382,9 @@ spec:
             - {name: provider-auth, mountPath: /shared-auth}
             - {name: coordinator, mountPath: /opt/coordinator, readOnly: true}
             - {name: auth-patch, mountPath: /opt/hermes/hermes_cli/auth.py, subPath: auth.py}
+            - {name: codex-runtime-patch, mountPath: /opt/hermes/hermes_cli/runtime_provider.py, subPath: runtime_provider.py}
+            - {name: codex-runtime-patch, mountPath: /opt/hermes/agent/transports/codex_app_server_session.py, subPath: codex_app_server_session.py}
+            - {name: codex-runtime-patch, mountPath: /opt/hermes/agent/codex_runtime.py, subPath: codex_runtime.py}
             - {name: tui-gateway-patch, mountPath: /opt/hermes/tui_gateway/server.py, subPath: server.py}
             - {name: kubeconfig, mountPath: /opt/data/home/.kube/config, subPath: config, readOnly: true}
             - {name: auto-router-plugin, mountPath: /opt/data/plugins/auto-router, readOnly: true}
@@ -497,6 +530,9 @@ spec:
             - {name: provider-auth, mountPath: /shared-auth}
             - {name: coordinator, mountPath: /opt/coordinator, readOnly: true}
             - {name: auth-patch, mountPath: /opt/hermes/hermes_cli/auth.py, subPath: auth.py}
+            - {name: codex-runtime-patch, mountPath: /opt/hermes/hermes_cli/runtime_provider.py, subPath: runtime_provider.py}
+            - {name: codex-runtime-patch, mountPath: /opt/hermes/agent/transports/codex_app_server_session.py, subPath: codex_app_server_session.py}
+            - {name: codex-runtime-patch, mountPath: /opt/hermes/agent/codex_runtime.py, subPath: codex_runtime.py}
             - {name: tui-gateway-patch, mountPath: /opt/hermes/tui_gateway/server.py, subPath: server.py}
             - {name: kubeconfig, mountPath: /opt/data/home/.kube/config, subPath: config, readOnly: true}
             - {name: auto-router-plugin, mountPath: /opt/data/plugins/auto-router, readOnly: true}
@@ -615,6 +651,8 @@ spec:
           emptyDir: {}
         - name: tui-gateway-patch
           emptyDir: {}
+        - name: codex-runtime-patch
+          emptyDir: {}
         - name: auto-router-plugin
           configMap:
             name: hermes-auto-router-plugin

services/hermes/kustomization.yaml

@@ -58,6 +58,7 @@ configMapGenerator:
       - install_agent_tools.sh=scripts/install_agent_tools.sh
       - migrate_herdr_state.py=scripts/migrate_herdr_state.py
       - patch_hermes_auth.py=scripts/patch_hermes_auth.py
+      - patch_codex_runtime.py=scripts/patch_codex_runtime.py
       - patch_tui_gateway.py=scripts/patch_tui_gateway.py
       - patch_ttyd_index.py=scripts/patch_ttyd_index.py
     options:

services/hermes/scripts/configure_agent_clients.py

@@ -8,6 +8,8 @@ import os
 from pathlib import Path
 from typing import Any
 
+import yaml
+
 
 def configure_claude_state(path: Path) -> None:
     """Disable Claude's long-session resume chooser without losing state."""
@@ -98,11 +100,30 @@ def configure_claude_settings(path: Path) -> None:
     os.chmod(path, 0o600)
 
 
+def configure_codex_runtime(config_path: Path, migrate_fn=None) -> Any:
+    """Expose Hermes tools to Codex and keep the owner lane unattended."""
+    config = yaml.safe_load(config_path.read_text(encoding="utf-8")) or {}
+    if migrate_fn is None:
+        from hermes_cli.codex_runtime_plugin_migration import migrate
+
+        migrate_fn = migrate
+    report = migrate_fn(
+        config,
+        default_permission_profile=":danger-no-sandbox",
+    )
+    if report.errors:
+        raise RuntimeError("; ".join(str(item) for item in report.errors))
+    print(report.summary())
+    return report
+
+
 def main() -> None:
     """Configure provider clients under the persistent Hermes home."""
     home = Path(os.environ.get("CLAUDE_CONFIG_DIR", "/opt/data/home/.claude"))
     configure_claude_state(home / ".claude.json")
     configure_claude_settings(home / "settings.json")
+    hermes_home = Path(os.environ.get("HERMES_HOME", "/opt/data"))
+    configure_codex_runtime(hermes_home / "config.yaml")
 
 
 if __name__ == "__main__":

services/hermes/scripts/hermes_model_routing.py

@@ -328,6 +328,7 @@ def _profile_config(
         "provider": primary_provider,
         "default": primary_model,
         "model": primary_model,
+        "openai_runtime": "codex_app_server",
     }
     config["fallback_providers"] = [
         fallback,
@@ -398,6 +399,7 @@ def configure_routes(root: Path, codex: Catalog, claude: Catalog) -> dict[str, A
         "provider": "openai-codex",
         "default": codex_coordinator,
         "model": codex_coordinator,
+        "openai_runtime": "codex_app_server",
     }
     base["fallback_providers"] = [
         {"provider": "anthropic", "model": claude_coordinator},
@@ -441,10 +443,7 @@ def configure_routes(root: Path, codex: Catalog, claude: Catalog) -> dict[str, A
             ),
             env_values,
         )
-        local = [
-            "custom/qwen2.5:14b-instruct-q4_0",
-            "custom/gpt-oss:20b",
-        ]
+        local = ["custom/gpt-oss:20b"]
         routes[codex_name] = [
             f"openai-codex/{codex_model}",
             f"anthropic/{claude_model}",
@@ -473,7 +472,6 @@ def configure_routes(root: Path, codex: Catalog, claude: Catalog) -> dict[str, A
     routes["synthesis-xhigh"] = [
         f"anthropic/{claude_models['xhigh']}",
         f"openai-codex/{codex_models['xhigh']}",
-        "custom/qwen2.5:14b-instruct-q4_0",
         "custom/gpt-oss:20b",
     ]
     _write_yaml(
@@ -486,7 +484,6 @@ def configure_routes(root: Path, codex: Catalog, claude: Catalog) -> dict[str, A
     routes["coordinator"] = [
         f"openai-codex/{codex_coordinator}",
         f"anthropic/{claude_coordinator}",
-        "custom/qwen2.5:14b-instruct-q4_0",
         "custom/gpt-oss:20b",
     ]
     return routes

services/hermes/scripts/patch_codex_runtime.py

@@ -0,0 +1,155 @@
+#!/usr/bin/env python3
+"""Let Hermes use the authenticated Codex app-server with routed settings."""
+
+from __future__ import annotations
+
+import argparse
+from pathlib import Path
+
+
+PROVIDER_BEFORE = '''    if provider == "openai-codex":
+        try:
+            creds = resolve_codex_runtime_credentials()
+'''
+PROVIDER_AFTER = '''    if provider == "openai-codex":
+        if (
+            str(model_cfg.get("openai_runtime") or "").strip().lower()
+            == "codex_app_server"
+        ):
+            return {
+                "provider": "openai-codex",
+                "api_mode": "codex_app_server",
+                "base_url": DEFAULT_CODEX_BASE_URL,
+                "api_key": "codex-cli-runtime",
+                "source": "codex-cli",
+                "requested_provider": requested_provider,
+            }
+        try:
+            creds = resolve_codex_runtime_credentials()
+'''
+
+SESSION_SIGNATURE_BEFORE = '''    def run_turn(
+        self,
+        user_input: Any,
+        *,
+        turn_timeout: float = 600.0,
+'''
+SESSION_SIGNATURE_AFTER = '''    def run_turn(
+        self,
+        user_input: Any,
+        *,
+        model: Optional[str] = None,
+        effort: Optional[str] = None,
+        turn_timeout: float = 600.0,
+'''
+
+SESSION_REQUEST_BEFORE = '''            ts = self._client.request(
+                "turn/start",
+                {
+                    "threadId": self._thread_id,
+                    "input": [{"type": "text", "text": user_input_text}],
+                },
+                timeout=10,
+            )
+'''
+SESSION_REQUEST_AFTER = '''            turn_params: dict[str, Any] = {
+                "threadId": self._thread_id,
+                "input": [{"type": "text", "text": user_input_text}],
+            }
+            if model:
+                turn_params["model"] = model
+            if effort:
+                turn_params["effort"] = effort
+            ts = self._client.request(
+                "turn/start",
+                turn_params,
+                timeout=10,
+            )
+'''
+
+TURN_BEFORE = '''    try:
+        turn = agent._codex_session.run_turn(user_input=user_message)
+'''
+TURN_AFTER = '''    reasoning = getattr(agent, "reasoning_config", None)
+    effort = (
+        str(reasoning.get("effort") or "").strip()
+        if isinstance(reasoning, dict)
+        else ""
+    )
+    try:
+        turn = agent._codex_session.run_turn(
+            user_input=user_message,
+            model=str(getattr(agent, "model", "") or "").strip() or None,
+            effort=effort or None,
+        )
+'''
+
+
+def _replace_once(content: str, before: str, after: str, label: str) -> str:
+    """Apply one exact replacement and fail closed when upstream drifts."""
+    if content.count(before) != 1:
+        raise RuntimeError(f"Hermes {label} patch context changed")
+    return content.replace(before, after, 1)
+
+
+def patch_provider(source: Path, destination: Path) -> None:
+    """Allow the explicit Codex app-server runtime without duplicate OAuth."""
+    content = source.read_text(encoding="utf-8")
+    destination.parent.mkdir(parents=True, exist_ok=True)
+    destination.write_text(
+        _replace_once(
+            content,
+            PROVIDER_BEFORE,
+            PROVIDER_AFTER,
+            "runtime provider",
+        ),
+        encoding="utf-8",
+    )
+
+
+def patch_session(source: Path, destination: Path) -> None:
+    """Pass the router's selected model and effort to each Codex turn."""
+    content = source.read_text(encoding="utf-8")
+    content = _replace_once(
+        content,
+        SESSION_SIGNATURE_BEFORE,
+        SESSION_SIGNATURE_AFTER,
+        "app-server session signature",
+    )
+    content = _replace_once(
+        content,
+        SESSION_REQUEST_BEFORE,
+        SESSION_REQUEST_AFTER,
+        "app-server turn request",
+    )
+    destination.parent.mkdir(parents=True, exist_ok=True)
+    destination.write_text(content, encoding="utf-8")
+
+
+def patch_turn(source: Path, destination: Path) -> None:
+    """Forward live Hermes route metadata into the Codex session adapter."""
+    content = source.read_text(encoding="utf-8")
+    destination.parent.mkdir(parents=True, exist_ok=True)
+    destination.write_text(
+        _replace_once(content, TURN_BEFORE, TURN_AFTER, "Codex turn"),
+        encoding="utf-8",
+    )
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("provider_source", type=Path)
+    parser.add_argument("provider_destination", type=Path)
+    parser.add_argument("session_source", type=Path)
+    parser.add_argument("session_destination", type=Path)
+    parser.add_argument("turn_source", type=Path)
+    parser.add_argument("turn_destination", type=Path)
+    args = parser.parse_args()
+    patch_provider(args.provider_source, args.provider_destination)
+    patch_session(args.session_source, args.session_destination)
+    patch_turn(args.turn_source, args.turn_destination)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

testing/tests/test_hermes_cli_lanes.py

@@ -37,6 +37,7 @@ policy = _load("claude_command_policy")
 migration = _load("migrate_herdr_state")
 auth_patch = _load("patch_hermes_auth")
 tui_gateway_patch = _load("patch_tui_gateway")
+codex_runtime_patch = _load("patch_codex_runtime")
 ttyd_patch = _load("patch_ttyd_index")
 client_config = _load("configure_agent_clients")
 
@@ -858,6 +859,60 @@ def test_auth_patch_fails_closed_on_upstream_drift(tmp_path: Path):
         auth_patch.patch(source, tmp_path / "patched.py")
 
 
+def test_codex_runtime_patch_uses_cli_and_forwards_route(tmp_path: Path):
+    provider = tmp_path / "runtime_provider.py"
+    provider.write_text(codex_runtime_patch.PROVIDER_BEFORE, encoding="utf-8")
+    provider_out = tmp_path / "patched/runtime_provider.py"
+    codex_runtime_patch.patch_provider(provider, provider_out)
+    assert '"api_mode": "codex_app_server"' in provider_out.read_text()
+
+    session = tmp_path / "codex_app_server_session.py"
+    session.write_text(
+        codex_runtime_patch.SESSION_SIGNATURE_BEFORE
+        + codex_runtime_patch.SESSION_REQUEST_BEFORE,
+        encoding="utf-8",
+    )
+    session_out = tmp_path / "patched/codex_app_server_session.py"
+    codex_runtime_patch.patch_session(session, session_out)
+    session_content = session_out.read_text()
+    assert 'turn_params["model"] = model' in session_content
+    assert 'turn_params["effort"] = effort' in session_content
+
+    turn = tmp_path / "codex_runtime.py"
+    turn.write_text(codex_runtime_patch.TURN_BEFORE, encoding="utf-8")
+    turn_out = tmp_path / "patched/codex_runtime.py"
+    codex_runtime_patch.patch_turn(turn, turn_out)
+    assert "model=str(getattr(agent" in turn_out.read_text()
+
+
+def test_codex_runtime_patch_fails_closed_on_upstream_drift(tmp_path: Path):
+    source = tmp_path / "runtime_provider.py"
+    source.write_text("def changed():\n    pass\n", encoding="utf-8")
+    with pytest.raises(RuntimeError, match="context changed"):
+        codex_runtime_patch.patch_provider(source, tmp_path / "patched.py")
+
+
+def test_codex_runtime_migration_uses_owner_unsafe_profile(tmp_path: Path):
+    config = tmp_path / "config.yaml"
+    config.write_text("model: {}\n", encoding="utf-8")
+    calls = []
+
+    class Report:
+        errors = []
+
+        @staticmethod
+        def summary():
+            return "configured"
+
+    def migrate(value, **kwargs):
+        calls.append((value, kwargs))
+        return Report()
+
+    client_config.configure_codex_runtime(config, migrate)
+
+    assert calls[0][1]["default_permission_profile"] == ":danger-no-sandbox"
+
+
 def test_tui_gateway_patch_extends_and_bounds_agent_startup(tmp_path: Path):
     source = tmp_path / "server.py"
     destination = tmp_path / "patched/server.py"

testing/tests/test_hermes_coordinator.py

@@ -92,7 +92,9 @@ def test_configure_routes_builds_cross_provider_fallback_profiles(tmp_path: Path
         (tmp_path / "profiles/claude-high/config.yaml").read_text(encoding="utf-8")
     )
     assert root["model"]["model"] == "gpt-5.6-terra"
+    assert root["model"]["openai_runtime"] == "codex_app_server"
     assert codex_profile["model"]["model"] == "gpt-5.6-sol"
+    assert codex_profile["model"]["openai_runtime"] == "codex_app_server"
     assert codex_profile["fallback_providers"][0] == {
         "provider": "anthropic",
         "model": "claude-opus-5",
@@ -104,6 +106,7 @@ def test_configure_routes_builds_cross_provider_fallback_profiles(tmp_path: Path
     assert codex_profile["fallback_providers"][1] == routing.ATLAS_FALLBACK
     assert len(codex_profile["fallback_providers"]) == 2
     assert routes["coordinator"][0] == "openai-codex/gpt-5.6-terra"
+    assert "custom/qwen2.5:14b-instruct-q4_0" not in routes["coordinator"]
     assert "max" not in json.dumps(routes)
 
     profile_env = (tmp_path / "profiles/codex-high/.env").read_text(encoding="utf-8")