[codex] fix(zai): sanitize GLM system prompt at runtime

[RyderFreeman4Logos]

Summary

• sanitize the per-request system prompt copy for Z.AI / GLM requests
• strip or neutralize Hermes/Nous branding strings that trigger Z.AI Coding overload-style 429/code-1305 failures
• add focused tests that verify Z.AI routing is sanitized without mutating the original cached messages

Root Cause

Z.AI GLM-5.x can reject the normal Hermes-branded system prompt at the provider boundary, returning overloaded/rate-limit style failures even for tiny non-streaming requests. Direct testing showed the request succeeds when the triggering Hermes Agent branding phrase is removed, so this patch rewrites only the API-boundary copy sent to Z.AI/GLM endpoints.

This preserves per-conversation prompt caching because the cached system prompt and conversation history are not modified.

This is intentionally scoped to the prompt sanitizer path. It does not duplicate the retry/backoff work from #59523.

Validation

• /usr/local/share/uv/tools/bin/pytest tests/agent/test_zai_prompt_policy.py
• /home/obj/.hermes/hermes-agent/venv/bin/python -m py_compile agent/zai_prompt_policy.py agent/chat_completion_helpers.py agent/conversation_loop.py tests/agent/test_zai_prompt_policy.py

[teknium1]

Thanks for isolating the outbound-prompt approach and for preserving the cached message objects.

Problems

• agent/zai_prompt_policy.py:28 rewrites bare Hermes case-insensitively. Current prompts contain operational ~/.hermes/... profile paths at agent/system_prompt.py:397-409; this rule changes those paths for Z.AI users.
• agent/zai_prompt_policy.py:40 enables the policy for every glm-* model. Current main explicitly treats arbitrary local/private GLM endpoints differently from direct provider identity (run_agent.py:1527-1547), so model-name-only gating is too broad.
• Current chat-completions conversion is centralized at agent/transports/chat_completions.py:331-338; move normal-request policy there and retain explicit coverage for the direct summary call at agent/chat_completion_helpers.py:1761-1764.

Suggested changes

• Preserve bare CLI/path/environment identifiers and test ~/.hermes/... content.
• Gate on explicit provider identity or parsed direct Z.AI hostnames, with a localhost GLM regression test.

Automated hermes-sweeper review.

[teknium1]

Please keep this replacement set narrow. With re.IGNORECASE, this rule rewrites the hermes segment of current system-prompt paths such as ~/.hermes/profiles/... (agent/system_prompt.py:397-409), making profile-isolation instructions point to a nonexistent path. Preserve bare operational identifiers and add a regression test for them.

[teknium1]

A glm-* model name does not establish a direct Z.AI endpoint. Current main deliberately distinguishes Ollama-hosted GLM from arbitrary local/private endpoints (run_agent.py:1527-1547). Restrict this to explicit Z.AI/Zhipu provider identity or parsed direct Z.AI hostnames, and cover a custom localhost GLM case.