root
0f3a9a977e
Desktop Client Build / Resolve Build Metadata (push) Has been cancelled
Desktop Client Build / Build Desktop Client (push) Has been cancelled
Desktop Client Build / Publish Client Artifacts to NAS (push) Has been cancelled
Frontend CI / Frontend (push) Successful in 6m6s
K2.6 introduced new UI surfaces that polluted the answer extraction and
the busy-signal probe, breaking monitor runs even though Kimi rendered a
correct answer:
- collectCandidates now drops elements that contain a chat composer child
(<textarea>, non-hidden <input>, [contenteditable]). This filters
wrapper containers like .chat-detail-content that the wide
[class*="content"] fallback would otherwise pick, pulling editor
placeholder text ("可快捷使用技能", "Kimi K2.6 已就位") into the answer.
- isAuxiliaryPanelElement skips elements with role="dialog" /
"alertdialog", aria-modal="true", or the <dialog> tag, so the K2.6
launch-announcement popover ("Kimi K2.6, 已就位 / 一键部署 OpenClaw")
no longer counts as answer content.
- Drop the class-based loading-indicator probe entirely. The previous
global [class*="loading"]/"stream"/"typing"/"spinner" sweep matched
K2.6's always-on text-stream wrapper and held busy=true forever, which
caused kimi_query_timeout even after the assistant turn had fully
rendered. The textual busySignalsPattern (思考中/搜索中/生成中…) is
the actual semantic signal and survives any UI rename.
- Surface a capacityNotice from the assistant turn when Kimi shows its
short capacity-exhausted message ("算力不够 / 请稍后再试"). The top
level falls back to it when answerText is empty so the SaaS pipeline
records the user-visible message instead of an unknown row, with
capacity_limited / capacity_notice flagged in raw_response_json.
All filters lean on HTML form / ARIA semantics and visible text rather
than Kimi-specific BEM classes, so they survive future redesigns.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>