feat: add server-side Thought-Aligner defense

[jiangchangyue]

Summary

Adds an opt-in, server-side Thought-Aligner defense for Python agents.

The integration intercepts an agent model response before its action reaches the framework parser or tool executor, extracts the exposed instruction, current thought, and previous thought/observation history, and sends them to a configured Thought-Aligner endpoint. When the thought is rewritten, the Python client asks the original agent model to regenerate its Action and Action Input from the aligned thought.

This provides thought-level behavioral safety intervention while preserving the existing AgentGuard tool-policy enforcement path.

Thought-Aligner model card: https://huggingface.co/WhitzardAgent/Thought-Aligner-7B

Type of Change

• Bug fix
• New feature (framework adapter, plugin, policy capability, ...)
• Breaking change
• Documentation
• CI / tooling / chore

Changes

• Added the opt-in server-side thought_aligner plugin in the llm_after phase.
• Added extraction and normalization for:
  • explicit normalized thought fields;
  • common reasoning aliases such as reasoning_content, reasoning, thinking, plan, and analysis;
  • <think>, <thought>, <reason>, <reasoning>, and <analysis> blocks;
  • ReAct-style Thought: output;
  • user instructions from metadata, messages, prompts, and LangChain-style inputs;
  • completed historical Thought/Observation pairs formatted with <thought> and <observation> markers.
• Added an OpenAI-compatible Thought-Aligner client with dedicated server-side URL, API key, model, timeout, and context-size configuration.
• Added Python client regeneration support:
  • holds the original LLM response before framework action parsing/execution;
  • injects the aligned thought into a copy of the original request;
  • invokes the original agent model once more to regenerate Action and Action Input;
  • returns only the regenerated response to the framework.
• Added loop prevention so each LLM turn can be regenerated at most once.
• Added fail-safe handling for unsupported or malformed alignment responses.
• Preserved the existing tool_before policy path for regenerated actions.
• Added an example plugin configuration:
  config/plugins.thought-aligner.example.json.
• Added English and Chinese Thought-Aligner documentation.
• Added regression tests for extraction, server transport, plugin decisions, synchronous/asynchronous regeneration, request immutability, unsupported clients, and loop prevention.

[github-actions]

🎉 Thanks for your first pull request to AgentGuard! A maintainer will review it soon. While you wait, please double-check the PR checklist and make sure ruff check src tests, pytest -q, and node --test pass locally — see CONTRIBUTING.md for details.