Releases: AIPentest/CyberStrikeAI
Release list
v1.7.17
这次修复关闭了聊天链路中的 Eino checkpoint 自动恢复，避免旧断点抢先恢复导致模型反复回复历史对话。
This update disables automatic Eino checkpoint resume in chat flows, preventing stale checkpoints from taking over and causing repeated old responses.
同时新增/完善了按轮次展示 Token 用量的统计能力，可查看每次对话的输入、输出、缓存、推理和模型调用次数。
It also adds/improves per-turn token usage tracking, showing input, output, cached, reasoning tokens, and model call counts for each conversation turn.
v1.7.16
修复 Eino 单代理模式下长对话触发摘要压缩后，tool_call/tool_result 配对被破坏导致模型 API 返回 400 错误（NodeRunError）的问题。为 agentic 路径补充了与 classic 路径对齐的 toolPairReconciler 和 orphanToolPruner 纵深防御中间件。
Fixed a bug where long conversations in Eino single-agent mode could trigger a 400 error (NodeRunError: "insufficient tool messages following tool_calls message") after summarization truncated the message history and broke tool_call/tool_result pairing. Added toolPairReconciler and orphanToolPruner defense-in-depth middlewares to the agentic path, aligned with the existing classic path.
v1.7.15
v1.7.14
本次更新已将 CyberStrikeAI 的 Eino 主链路升级到 0.9.x Agentic Runtime：单代理/多代理采用 AgenticMessage 与 typed Agent，并接入 Runner、checkpoint、TurnLoop、summarization、reduction、toolsearch、plantask、Skills、retry/failover 与观测能力。
This update moves CyberStrikeAI’s Eino runtime onto the 0.9.x Agentic Runtime path: single-agent and multi-agent flows now use AgenticMessage and typed agents, with Runner, checkpoint, TurnLoop, summarization, reduction, toolsearch, plantask, Skills, retry/failover, and observability integrated.