dsh-noletme
实时推理轨迹关键词统计插件:在 Web 聊天界面跟踪并展示模型流式推理块中的语气倾向与犹豫度指标。
AI 分析
核心用途是分析和监控大模型推理过程中的思维倾向。适合对模型提示词工程、推理行为模式感兴趣的研究者。
安装
$
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:Yuer6327/NoLetMe说明文档
阅读完整 README ↗Usage
- The panel docks below the session header at the top-right (clears the "Session log" download action), floating over the conversation frame.
- Collapsed it is a compact rounded-rectangle chip (dot + "NoLetMe" +
current mode); click it to expand into the card. Expanded it shows the
live status strip (streaming/syncing dot, reasoning blocks / chars, visible
replies), the trajectory-mode badge on the same row as its label,
per-category share bars, raw research metrics (
we · let's · let me · I), the keyword breakdown, and a hesitation-pressure health note. - The chip↔card transition is a single surface that morphs on a
critically-damped spring (interruptible, anchored at the right dock), and
degrades to an instant swap under
prefers-reduced-motion. The open state is remembered. - The panel body's scrollbar is hover-revealed through the harness's own
--dsh-scrollbar-*tokens (invisible until hovered).
Data freshness, persistence, and privacy
- Reasoning-only, like the evidence: the keyword trajectory is counted over
reasoning blocks only — the exact scope of
analyze_trajectory_exports.py. If a model streams its output as visible text with little or no reasoning, the panel shows a reasoning-health alert (with the raw reasoning/text totals) instead of fabricating counts from the text. - Real-time: stats fold incrementally on every streamed reasoning delta (at most once per animation frame) — nothing re-walks the conversation.
- Session switching: switching conversations immediately repaints the new session's stats from its local cache, then pages in the complete history (a "syncing" indicator shows meanwhile) so the count covers the whole conversation.
- Local persistence: each session's folded counts are stored in
localStorage(dsh-noletme.stats.), so reopening a conversation does not recount it — only new messages are folded. - Robustness: compaction rewrites reset the count once; history paging is capped and aborts when you switch away; storage failures are swallowed.
- No data leaves your browser.