fix(agent): sanitize surrogate characters from Ollama model output

[tmchow]

What does this PR do?

Fixes three gaps in the existing surrogate character sanitization that cause UnicodeEncodeError crashes when using Ollama-hosted models (Kimi K2.5, GLM-5, Qwen). These models can return invalid surrogate code points (U+D800-U+DFFF) that crash json.dumps() during UTF-8 serialization.

The existing _sanitize_surrogates() and _sanitize_messages_surrogates() functions work correctly, but they're not called in the right places:

1. No proactive sanitization before the API call (line 6848). Surrogates from prior model output or tool results survive in api_messages and crash on the next turn. Fixed by calling _sanitize_messages_surrogates(api_messages) after _sanitize_api_messages().

2. API responses stored without sanitization (line 5242). Surrogates from model output persist in session history via _build_assistant_message(). Fixed by sanitizing content and reasoning_text before building the message dict.

3. Retry path sanitized the wrong variable (line 7368). The UnicodeEncodeError handler called _sanitize_messages_surrogates(messages) but the API call uses api_messages (a separate list). The retry would fail again with the same error. Fixed by sanitizing both api_messages and messages.

Related Issue

Fixes #5059

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• run_agent.py:6848: Added _sanitize_messages_surrogates(api_messages) call after _sanitize_api_messages() for proactive surrogate cleaning before the API call
• run_agent.py:5242: Sanitize assistant_message.content and reasoning_text with _sanitize_surrogates() before storing in message history
• run_agent.py:7368: Changed _sanitize_messages_surrogates(messages) to also sanitize api_messages so the retry actually cleans the list sent to the API

How to Test

1. Set up Hermes with an Ollama-hosted model (GLM-5 or Kimi K2.5)
2. Ask the model to process a large file or have a long conversation that produces surrogate characters
3. Verify no UnicodeEncodeError crash occurs
4. Run pytest tests/test_surrogate_sanitization.py to verify existing tests still pass

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: macOS 15 (Darwin 25.3.0)

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — or N/A
• I've updated cli-config.yaml.example if I added/changed config keys — or N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — or N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — or N/A
• I've updated tool descriptions/schemas if I changed tool behavior — or N/A

This contribution was developed with AI assistance (Claude Code).

[tmchow]

Closing in favor of #5074 which addresses the same issue and was submitted earlier. Nice fix @ygd58. Missed yours before i started my fix!