LightAgent v0.9.3

Highlights

LightAgent v0.9.3 completes the runtime hook lifecycle and hardens streaming tool execution. This release includes all changes since v0.9.1, including the hook completion work that was prepared under v0.9.2 but did not receive a standalone GitHub Release.

• Add after_run and on_error hooks for consistent run lifecycle closure.
• Add before_memory_retrieve and after_memory_retrieve hooks around memory reads.
• Route terminal model errors, Guardrail/Hook blocks, recoverable tool errors, and stream loop-limit failures through the shared error lifecycle.
• Add runtime hook recipes for auditing, persistence, filtering, alerting, and fallback handling.
• Bound streaming tool-call loops and expose the independent max_tool_iterations parameter.
• Preserve compatibility by defaulting max_tool_iterations to the existing max_retry value.
• Ensure stream loop-limit failures close through _finish_run(...), triggering on_error, after_run, and a single consistent run_end trace.
• Add a dependency-free ClawMem MemoryProtocol adapter example, documentation, and fake-client tests.
• Add Atlas Cloud as an OpenAI-compatible provider example using environment-based credentials.

Compatibility

• Existing agent.run("hello") behavior remains compatible.
• Existing agent.run(query, stream=True, user_id=user_id) behavior remains compatible.
• Applications that only configure max_retry retain the previous effective streaming loop limit.
• ClawMem and Atlas Cloud examples add no mandatory core dependencies and contain no embedded credentials.

Validation

• GitHub CI passed on Python 3.10, 3.11, and 3.12 for the v0.9.3 core and documentation PRs.
• Local validation on the merged main: 82 passed across all 11 Git-tracked test modules.
• The Atlas Cloud example was syntax-checked and its endpoint/model configuration was verified against official provider documentation.

Full changes: v0.9.1...v0.9.3