v0.1.4

This release adds a shared rollout gateway, recoverable AgentCore invocations and sandbox commands, and new training integrations.

Features

Rollout gateway: Capture token IDs, log probabilities, and loss masks through OpenAI- and Anthropic-compatible endpoints. Includes Hugging Face rendering, model-native response schemas, multi-turn trajectory trees, and a threaded gateway server (#80, #93, #94, #97).
HTTP runtime and client: Add AgentCoreRuntimeApp, AgentCoreHttpClient, and invocation handles for foreground/background execution, status polling, and filesystem-backed invocation records (#162, #165).
Sandbox SDK: Run shell commands in AgentCore Runtime through the Python SDK and Go sandboxd, with recoverable execution handles and persisted command results (#88, #153).
Tinker-compatible training: Add synchronous group-relative training through the official Tinker API or a compatible SkyRL endpoint, plus an EC2 deployment example (#155, #158, #165).
verl integration: Use the shared rollout gateway with verl 0.9.0 and vLLM 0.24.0; add variable-trajectory batching for synchronous and asynchronous training, agent metrics, FSDP/Megatron LoRA recipes, and Qwen3.6 support (#116, #118, #129, #131, #139, #140, #144, #147).
Experimental integrations: Add slime rollout-gateway hooks and episode-aware reward normalization, plus A2A rollout-session support (#106, #142, #156).

Improve gateway tokenization throughput and event-loop responsiveness, preserve full generations and full-context trajectories, and strengthen reasoning/tool-call parsing and history matching (#104, #109, #117, #128, #133, #141, #143, #152, #160).
Stabilize verl optimizer steps and reward-metric aggregation for variable-row rollouts; drop failed rollouts that contain no captured trajectory (#122, #131, #136).
Identify toolkit traffic in AWS user-agent metadata (#157).
Improve example input validation and archive extraction, and make SWE agent terminals and grading feedback more reliable (#95, #154).

Add the OpenHands SWE agent and asynchronous verl training recipes (#123, #130).
Add Harbor sandbox evaluation and MigrationBench Megatron LoRA examples (#112, #114).
Expand training, data synthesis, runtime, and sandbox documentation and blog posts.

Python 3.11 or newer is required, up from Python 3.10 in v0.1.3.
The verl integration has changed. Existing users of AgentCoreTrainer, AgentCoreLoopManager, or the old backends.verl.config classes should migrate to AgentCoreAgentLoop and verl's main_ppo entry point. See the verl setup and recipes.
The legacy frameworks.strands.vllm_model wrapper has been removed. Use Strands' standard OpenAIModel with the endpoint's base_url (#82).
Training and gateway dependencies remain optional. Select the extras required by your backend; the tinker_api and tinker_skyrl extras use different Tinker SDK versions.

Full Changelog: v0.1.3...v0.1.4