fix: redact brand string for z.ai provider to avoid content-filter 429 (#60118)

[AlexFucuson9]

Summary

z.ai (GLM Coding Plan) silently rejects system prompts containing the literal string "Hermes Agent", returning HTTP 429 with code 1305 (mislabeled content-filter rejection). Every Hermes request injects this brand string, so every request fails with a fake "rate limit" error.

Root Cause

z.ai's content filter trips on "Hermes Agent" in the system prompt. The 429/1305 response is mislabeled — it's not a rate limit, it's a content filter. The account has quota, the key works, and removing the brand string makes requests succeed.

Changes

• Add _is_zai provider detection flag in build_api_kwargs()
• When z.ai is detected, sanitize system messages to replace:
  • "Hermes Agent" → "the assistant"
  • "Hermes agent" → "the assistant"
  • "Nous Research" → "the development team"
• Only system messages are affected; user/assistant messages are untouched

Precedent

The Anthropic adapter already does the same thing (line 2545): it replaces "Hermes Agent" → "Claude Code" for Anthropic's content filters. This is the same pattern applied to z.ai.

Test Plan

• Syntax check passes
• Requests to z.ai with the brand string no longer return 429/1305
• Requests to other providers are unaffected (sanitization is z.ai-only)
• System prompt functionality is preserved (agent identity is generic but functional)

Fixes #60118

[alt-glitch]

This was generated by AI during triage.

Duplicate of #59975 (earliest open, same site agent/chat_completion_helpers.py, same mechanism: redacting the Hermes Agent/Nous Research brand strings from z.ai system messages at the chat/completions API boundary). #59975 is the more comprehensive version (dedicated agent/zai_prompt_policy.py module, preserves prompt caching). Related: #53006 (same fix at the build_system_prompt() chokepoint — different site), #56823 (different trigger substring — plain-text skill_view() syntax), and issues #60118 / #53002 (the z.ai 1305 false-overload family). Human to consolidate the z.ai sanitizer cluster.

[teknium1]

Thanks for isolating the Z.AI prompt trigger. The reported source condition is still present on current main, but this implementation does not reach the normal direct Z.AI path.

Problems

• The new if _is_zai block is below the legacy no-profile branch in the PR diff. zai is a registered ZaiProfile (plugins/model-providers/zai/__init__.py:111-127), and build_api_kwargs() returns from the provider-profile path before the added block (agent/chat_completion_helpers.py:998-1028).
• Non-Codex iteration-limit summaries bypass _build_api_kwargs() and send api_messages directly (agent/chat_completion_helpers.py:1836-1847), leaving that request path uncovered.
• The PR has no regression tests for the registered Z.AI path or input/cache preservation.

Suggested changes

• Sanitize a per-request message copy before provider-profile dispatch, and add the same handling to the direct summary request path.
• Add focused tests for registered zai, system-only replacement, unchanged source messages, and non-Z.AI passthrough.

Automated hermes-sweeper review.

[teknium1]

This block is in the legacy no-profile branch. Direct zai resolves to the registered ZaiProfile, whose earlier return bypasses this code entirely; move a copy-only sanitizer before provider-profile dispatch or to the outbound request-construction boundary.