LightAgent v0.7.5

LightAgent v0.7.5 Release Notes

LightAgent v0.7.5 focuses on safer agent execution and better tool documentation. It adds an opt-in Guardrails layer while keeping default agent.run() behavior backward compatible.

Upgrade

pip install --upgrade lightagent>=0.7.5

Highlights

• Added opt-in input guardrails that run before memory lookup and model requests.
• Added opt-in tool-call guardrails that run after tool arguments are parsed and before Python tool execution.
• Added opt-in non-streaming output guardrails for final response validation or rewriting.
• Added GuardrailDecision and GuardrailManager as public exports.
• Added stable [LA-GUARDRAIL] error code for blocked operations.
• Added guardrail_block trace events when trace=True is enabled.
• Added focused guardrails documentation and FAQ examples.
• Added a comprehensive Tools guide covering tool_info, ToolRegistry, ToolLoader, AsyncToolDispatcher, runtime validation, dynamic loading, built-in tools, and MCP tool registration.
• Linked the Tools guide from README.md, README.zh-CN.md, and docs/FAQ.md.

Compatibility

• Guardrails are disabled by default.
• agent.run("hello") still returns a string by default.
• agent.run(query, stream=True, user_id=user_id) keeps legacy stream behavior unless guardrails are configured.
• Output guardrails currently apply to non-streaming final responses.

Tests

• Guardrails PR verification: python -m compileall -q LightAgent tests/test_v065_core.py tests/test_v070_tracing.py tests/test_memory_policy.py tests/test_guardrails.py
• Guardrails PR verification: PYTHONPATH=. python -m pytest -q tests/test_v065_core.py tests/test_v070_tracing.py tests/test_memory_policy.py tests/test_guardrails.py
• GitHub Actions CI passed for Python 3.10, 3.11, and 3.12 on PR #49.

Included Pull Requests

• #49
• #51

New Contributors

• @Oxygen56 contributed the new Tools guide in PR #51.

Full Changelog: v0.7.0...v0.7.5