Objective
运行一个最小 multi-turn tool Agent,导出完整 trajectory,并验证 assistant action、tool observation、mask、reward 和 termination 的边界。
Planned Trace
user query
-> assistant tool call
-> parsed arguments
-> tool execution
-> tool result
-> assistant continuation
-> final answer
-> terminal rewardRequired Artifacts
- raw messages and chat template output;
- token ids grouped by role and turn;
attention_mask、response_mask或等价 loss mask;- tool latency、status 和 truncated response;
- trajectory reward and failure metadata;
- rollout/model version identifiers。
Source Anchors
Acceptance Criteria
- observation token 不产生 policy gradient;
- tool failure 和 normal termination 可区分;
- turn-by-turn tokenization 与完整 conversation tokenization 的差异可观测;
- 结果回填 11 Agentic RL Extension。