fix(zai): sanitize blocked Hermes prompt phrase

agent/chat_completion_helpers.py

@@ -1360,6 +1360,21 @@ def handle_max_iterations(agent, messages: list, api_call_count: int) -> str:
         # turns so Anthropic-family providers don't 400 the summary call.
         api_messages = agent._drop_thinking_only_and_merge_users(api_messages)
 
+        # The max-iterations summary path calls chat.completions.create()
+        # directly instead of going through _build_api_kwargs(), so apply the
+        # same ProviderProfile message hook here. This keeps provider-gated
+        # outbound prompt fixes (for example Z.AI's prompt-sensitive 1305
+        # mitigation) from being bypassed during summary generation.
+        if agent.api_mode == "chat_completions":
+            try:
+                from providers import get_provider_profile
+
+                profile = get_provider_profile(agent.provider)
+                if profile:
+                    api_messages = profile.prepare_messages(api_messages)
+            except Exception as exc:
+                logger.debug("Summary provider message preparation skipped: %s", exc)
+
         summary_extra_body = {}
         try:
             from agent.auxiliary_client import _fixed_temperature_for_model, OMIT_TEMPERATURE as _OMIT_TEMP

plugins/model-providers/zai/__init__.py

@@ -1,9 +1,77 @@
 """ZAI / GLM provider profile."""
 
+from typing import Any
+
 from providers import register_provider
 from providers.base import ProviderProfile
 
-zai = ProviderProfile(
+
+_BLOCKED_SYSTEM_PROMPT_PHRASE = "Hermes Agent"
+_SAFE_SYSTEM_PROMPT_PHRASE = "Hermes framework"
+
+
+def _sanitize_system_prompt_content(content: Any) -> tuple[Any, bool]:
+    if isinstance(content, str):
+        if _BLOCKED_SYSTEM_PROMPT_PHRASE not in content:
+            return content, False
+        return (
+            content.replace(_BLOCKED_SYSTEM_PROMPT_PHRASE, _SAFE_SYSTEM_PROMPT_PHRASE),
+            True,
+        )
+
+    if not isinstance(content, list):
+        return content, False
+
+    changed = False
+    sanitized_parts = []
+    for part in content:
+        if (
+            isinstance(part, dict)
+            and isinstance(part.get("text"), str)
+            and _BLOCKED_SYSTEM_PROMPT_PHRASE in part["text"]
+        ):
+            sanitized_parts.append(
+                {
+                    **part,
+                    "text": part["text"].replace(
+                        _BLOCKED_SYSTEM_PROMPT_PHRASE,
+                        _SAFE_SYSTEM_PROMPT_PHRASE,
+                    ),
+                }
+            )
+            changed = True
+        else:
+            sanitized_parts.append(part)
+    return (sanitized_parts, True) if changed else (content, False)
+
+
+class ZaiProviderProfile(ProviderProfile):
+    def prepare_messages(self, messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        """Avoid Z.AI Coding Plan's prompt-sensitive 1305 false overload.
+
+        Z.AI GLM Coding Plan rejects requests whose effective system prompt
+        contains the exact phrase "Hermes Agent" with HTTP 429 / code 1305.
+        Rewrite only outbound system/developer prompt copies so cached Hermes
+        prompts, history, and user text remain unchanged.
+        """
+        sanitized_messages = None
+        for index, message in enumerate(messages):
+            if (
+                not isinstance(message, dict)
+                or message.get("role") not in {"system", "developer"}
+            ):
+                continue
+            content, changed = _sanitize_system_prompt_content(message.get("content"))
+            if not changed:
+                continue
+            if sanitized_messages is None:
+                sanitized_messages = list(messages)
+            sanitized_messages[index] = {**message, "content": content}
+
+        return sanitized_messages if sanitized_messages is not None else messages
+
+
+zai = ZaiProviderProfile(
     name="zai",
     aliases=("glm", "z-ai", "z.ai", "zhipu"),
     env_vars=("GLM_API_KEY", "ZAI_API_KEY", "Z_AI_API_KEY"),

tests/run_agent/test_run_agent.py

@@ -1519,6 +1519,42 @@ def test_basic_kwargs(self, agent):
         assert kwargs["messages"] is messages
         assert kwargs["timeout"] == 1800.0
 
+    def test_zai_sanitizes_hermes_agent_in_outgoing_system_prompt_only(self, agent):
+        agent.provider = "zai"
+        agent.model = "glm-5.2"
+        agent.base_url = "https://api.z.ai/api/coding/paas/v4"
+        messages = [
+            {
+                "role": "system",
+                "content": "You are Hermes Agent. Help debug Hermes Agent.",
+            },
+            {"role": "user", "content": "Explain the phrase Hermes Agent."},
+        ]
+
+        kwargs = agent._build_api_kwargs(messages)
+
+        assert kwargs["messages"] is not messages
+        assert (
+            kwargs["messages"][0]["content"]
+            == "You are Hermes framework. Help debug Hermes framework."
+        )
+        assert kwargs["messages"][1]["content"] == "Explain the phrase Hermes Agent."
+        assert messages[0]["content"] == "You are Hermes Agent. Help debug Hermes Agent."
+
+    def test_non_zai_preserves_hermes_agent_system_prompt(self, agent):
+        agent.provider = "custom"
+        agent.model = "llama-3.3"
+        agent.base_url = "https://custom.example.test/v1"
+        messages = [
+            {"role": "system", "content": "You are Hermes Agent."},
+            {"role": "user", "content": "hi"},
+        ]
+
+        kwargs = agent._build_api_kwargs(messages)
+
+        assert kwargs["messages"] is messages
+        assert kwargs["messages"][0]["content"] == "You are Hermes Agent."
+
     def test_public_moonshot_kimi_k2_5_omits_temperature(self, agent):
         """Kimi models should NOT have client-side temperature overrides.
 
@@ -3262,6 +3298,26 @@ def test_summary_strips_strict_schema_foreign_fields(self, agent):
         assert messages[2]["tool_name"] == "execute_code"
         assert messages[1]["codex_reasoning_items"] == [{"id": "rs_1"}]
 
+    def test_zai_summary_sanitizes_cached_system_prompt_without_mutating_cache(self, agent):
+        agent.provider = "zai"
+        agent.model = "glm-5.2"
+        agent.base_url = "https://api.z.ai/api/coding/paas/v4"
+        agent._base_url_lower = agent.base_url.lower()
+        agent.client.chat.completions.create.return_value = _mock_response(content="Summary")
+        agent._cached_system_prompt = "You are Hermes Agent. Debug Hermes Agent."
+
+        result = agent._handle_max_iterations(
+            [{"role": "user", "content": "Explain Hermes Agent."}],
+            60,
+        )
+
+        assert result == "Summary"
+        sent_msgs = agent.client.chat.completions.create.call_args.kwargs["messages"]
+        assert sent_msgs[0]["role"] == "system"
+        assert sent_msgs[0]["content"] == "You are Hermes framework. Debug Hermes framework."
+        assert sent_msgs[1]["content"] == "Explain Hermes Agent."
+        assert agent._cached_system_prompt == "You are Hermes Agent. Debug Hermes Agent."
+
     def test_summary_omits_provider_preferences_for_non_openrouter(self, agent):
         agent.base_url = "https://api.openai.com/v1"
         agent._base_url_lower = agent.base_url.lower()