dsh-think-phrases

DeepSeek models silently turn on reasoning blocks during gray-test phases, yet the visible reply alone gives almost no clue whether a thought chain was actually produced. dsh-think-phrases solves that blind spot: it walks every assistant reasoning block in the current session, tallies four high-frequency phrases (let me, we, let s, I ll), and pins a compact counter onto the conversation header utility slot. Click the badge and you see per-phrase counts plus a grand total at a glance, with no log digging or ad-hoc regex required.

The matching engine respects word boundaries carefully. The we pattern captures only first-person plural usage (we are, we will, pre-we, and so on) and will not false-positive on words like weird or sweet. let s handles straight quotes, curly quotes, and the no-apostrophe variant. I ll covers both the short form and the full I will expression. All matching is case-insensitive, and only reasoning text is counted while visible body text and user messages are excluded, keeping the numbers clean and meaningful.

The plugin is lightweight, carries zero external dependencies, and works in DSH Desktop 2.x as well as plain dsh web and dsh console. It is ideal for developers and power users who run DeepSeek models under DSH and want an always-visible, low-interference signal of whether the model is truly producing a thinking chain. Once installed and the host is restarted, the counter appears in the header without disturbing any other plugin or service.