chat-ui

2a61e090 - ML Intern: don't end a turn before its plan's steps have started (#2753)

ML Intern: don't end a turn before its plan's steps have started (#2753) * ML Intern: don't end a turn before its plan's steps have started A round with no tool calls is final, so when GLM-class models reason about Qwen's <think></think> switch, the literal tags end their reasoning at the provider and the leftover half-thought finalizes the turn after step one. Retry once when an answer ends the turn while a step of this turn's plan was never started, naming the open steps. Also stop telling a think-only round in a tool loop to give its final answer; tell it to continue with its next tool call instead. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * Skip the plan retry without the plan tool or after an output-limit recovery Let the nudge leave a step pending while it waits on the user. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>