Release LightAgent v0.11.0 · wanxingai/LightAgent

Dynamic DAG Multi-Agent And Unified Security Context

Added the opt-in LightDAG runtime for persistent dynamic task graphs, runtime decomposition, bounded concurrent workers, and parent-task integration.
Added transactional SQLite graph mutations, cycle checks, task attempts, durable events, budget reservations, leases, fencing, and restart recovery.
Added application-owned verification gates and immutable content-addressed artifacts. Model output alone cannot mark a task complete or publish a trusted artifact.
Added narrowing-only SecurityContext, capability gates, and identity-, argument-, resource-, expiry-, and policy-bound approval tokens.
Fixed async-generator tool consumption in both streaming and non-streaming agent runs, and documented MCP Streamable HTTP configuration.
Hardened Python executor parsing with narrow JSON exception boundaries, bounded nested code extraction, cycle handling, and cross-version regression tests.
Added an optional dependency-free-at-import Memcode memory adapter example with aligned agent identity, namespaced users, provenance checks, and offline policy/hook round-trip tests.

Existing agent.run(...), streaming entry points, and fixed-workflow LightFlow usage remain available. LightDAG is opt-in.
Async-generator tools are buffered into a final result; this is not live per-chunk async streaming.
Legacy execute_python_code, execute_python_file, and execute_python_code_stream are blocked in model, dispatcher, and tool-provider invocation, even with enable_unsafe_python=True or a registered sandbox provider. Trusted direct utility calls remain available under application-managed isolation. Use safe_expression for data-only calculations or a custom tool backed by a real sandbox for agent code execution.
LightDAG is single-host, uses local SQLite, and supports one active scheduler per run with multiple async workers. Execution is at-least-once; external side effects still require idempotency or dedicated isolation.
Distributed workers, a Web UI, automatic deployment, and built-in code sandboxing are outside this release. Real-backend security acceptance remains opt-in; this release does not claim certification of external memory services.

Full tracked test suite: 365 passed, 1 skipped.
The skipped test requires an explicitly configured isolated, disposable Mem0 Graph backend.
Python 3.10, 3.11, 3.12, and 3.13 CI passed.
Offline LightDAG decomposition, failed verification, repair, and integration example passed.
Package build, isolated wheel import/version check, compileall, and git diff --check passed.

Core Dynamic DAG and security-context implementation: #105.
Thanks to long6177 for the async-generator fix (#108), ct-jaryn for the MCP documentation (#109), harshadkhetpal for the Python parsing improvements (#95, completed in #111), and Vivek Gupta / vivekgupta-memcode for the optional Memcode adapter (#106, completed in #112).
Additional executor dispatch hardening: #110.