Unverified50% confidenceFactExact time
The fixes include: using only the chat split for replay, keeping only answers with reasoning turned off, and setting an upper limit on length; after the changes, the model answers questions in two sentences and the style returns to the customer service agent target, but general evaluation scores drop slightly.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/13/2026
First Seen
Valid until: 10/27/2026
Sources
Related Entities
Related Claims
Unverifiedv2.1.281 fixes multiple session-restoration bugs, including earlier turns being re-sent in altered form upon session resumption causing the API to discard reasoning chains, restored oversized sessions only recovering the last few messages, and broken prompt caching.71% similarUnverifiedThis update fixes multiple crash issues related to --resume, --continue, and session restoration, including crashes when resuming conversations containing malformed message entries.71% similarUnverifiedThe response protocol in the configuration needs to be changed from chat to response; the latest version defaults to Response, but upgrades from older versions may still retain the Chat setting69% similarUnverified切换应用时前端可能重新初始化会话状态，触发新的查询请求，而旧请求未被正确取消导致重复计费，这是一种竞态条件问题66% similarUnverifiedv2.1.291 fixes an issue traced back to version 2.1.288: the last few messages of a session could be lost when a user exits the session65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1018235API
curl https://kongchang.com/api/v1/knowledge/claims/1018235MCP
get_claim(id=1018235)