Skip to content

v0.1.6 Release

Choose a tag to compare

View all tags
@luyuzhe111 luyuzhe111 released this 13 Oct 18:18
v0.1.6
702b559
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

v0.1.6

This release improves Runtime background execution, unifies rollout-session configuration, and updates the slime integration and training recipes.

Features and improvements

  • Runtime background execution: AgentCoreRuntimeApp(background=True) can now accept ordinary payloads without a protocol envelope and return immediately while the handler runs. Adds a SageMaker RFT integration example and SDK tests. Requests with explicit envelopes retain their per-request execution mode (#179).
  • Rollout-session configuration: Select backends by class import path. AgentCoreHttpSession now works with make_session() and reuses a shared AWS client and connection pool (#177, #183).
  • slime integration: Promote the hook-based integration from backends.experimental.slime to backends.slime, using the toolkit's rollout gateway for trajectory capture (#178).
  • Training recipes: Add SWE agent train/test splits, DAPO group filtering, and multi-region support. Update the migration-agent recipe to use the toolkit's agentcore_sync trainer mode (#173, #180).
  • Gateway and dependencies: Recognize the Qwen3.8 chat template for response parsing and upgrade verl's TransferQueue dependency to 0.1.10 (#174, #180).

Upgrade notes

This release includes breaking changes to rollout-session configuration and training integrations.

  • Backend selection: Replace short backend names such as agentcore_a2a with the full class import path, for example agentcore_rl_toolkit.rollout_session.agentcore_a2a_session.AgentCoreA2ASession. The remaining configuration keys are passed directly to that backend's constructor; remove keys intended for other backends (#177).
  • Direct session construction: AgentCoreA2ASession now uses the keyword agentcore_runtime_arn instead of runtime_arn. DockerA2ASession uses docker_iam_role_arn, docker_log_group, and docker_log_region. AgentCoreS3Session takes agentcore_runtime_arn, rollout_output_s3, and experiment_name instead of a caller-provided client (#177).
  • HTTP sessions: Replace AgentCoreHttpSession(sid, client=client) with AgentCoreHttpSession.with_client(sid, client) when reusing a caller-owned client. For configuration-based construction, use the session's class import path and agentcore_runtime_arn (#183).
  • slime: The old SlimeRunner API and slime extra have been removed. Update experimental module paths to agentcore_rl_toolkit.backends.slime, and use the generate and normalize_episode_rewards hooks. Dataset rows must carry the agent payload in metadata.payload. Follow the updated slime setup guide for installation and configuration (#178).
  • rLLM: The rllm extra has been removed. Follow the upstream rLLM AgentCore documentation for installation and training (#181).

When using the new background mode without an envelope, retrying an ordinary request creates a new invocation and may execute the handler again. Use an explicit protocol envelope with a stable invocation ID within the same Runtime session when deduplication is required.

Full changelog: v0.1.5...v0.1.6

Assets 2
Loading