NoLetMe · dsh reasoning-trajectory panel

A DeepSeek Harness (dsh) web plugin that adds a real-time reasoning-keyword statistics panel to the right edge of the conversation page.

While the model streams, NoLetMe watches its reasoning blocks and counts keywords that fingerprint how it is thinking:

It is a native-feeling overlay: the panel reuses the harness design system (--dsw-alias-* semantic tokens, the DetailsPanel header/body idiom, CSS Modules, light/dark and reduced-motion preserved) and mounts through the official shell.overlay slot — nothing in the shipped UI is replaced or patched.

数据来源与证据链 (Data sources & evidence chain)

The keyword taxonomy is not invented — every category traces to the public xiaobright/modeltest repository and its investigation of the DeepSeek V4 Pro GA 0813 post-training overfitting event (the RL checkpoints trained on DeepSeek Harness's "Minimal" preset — the "exact RL prompt and schemas" two-tool scaffold — collapse on the wider Standard tool catalog).

The test set. Project2 V4.1b — a real broken ESP-IDF embedded-engineering task, formally frozen (PROJECT_FROZEN.md, frozen 2026-07-23; scoring rules & hidden tests SHA-256-pinned 2026-07-19).

The measured evidence. evaluator/trajectory_evidence/derived/trajectory_stats.json (SHA-256-pinned per run; counting over completed assistant reasoning blocks, streaming chunks excluded):

High-scoring runs (96–99) carry we/let's with let me ≈ 0; low-scoring runs (91–92) carry let me in the hundreds. That clean separation is the source of the 🟢 高效 / 🟠 犹豫 split.

The classifier. The repo ships the exact lexical rules (evaluator/trigger_probe/src/classifier.mjs) that NoLetMe mirrors: first-line We need → minimal-like; we present with no let me → +2; any let me → standard-like; standalone Good./Great./ Excellent. first line → +1. The ⚪ 中性 category covers the remaining ambiguous/reflective frame — the Standard-catalog opening The user wants … Let me … recorded in docs/v4.1/DEEPSEEK_V4_TRIGGER_MECHANISM_EXPERIMENTS_20260814.md, plus general task-description vocabulary.

Honest limits. The matrix warns verbatim: "Lexical trajectory labels are observational fingerprints, not route or identity labels." Word counts fingerprint a reasoning style; they do not identify a backend, a route, or a checkpoint, and V4 Flash shifts style without a score change. NoLetMe is a reasoning-style diagnostic, not a model identity test.

Full event write-up: docs/research.md.

Install

Prerequisite: a dsh CLI (dsh --version), and a profile to install into.

cd /path/to/this/repo/..            # where NoLetMe/ lives
dsh plugin --profile demo add ./NoLetMe
dsh web --profile demo              # or just: dsh --profile demo

The bundle layer (cordis.patch.yml) inserts the dsh-noletme row; the dsh.client block in package.json tells the web shell to serve the browser bundle. The panel appears top-right in the web UI.

Install directly from GitHub (the prepare script builds lib/ on install):

dsh plugin --profile demo add github:Yuer6327/NoLetMe

pnpm ≥10 blocks a git dependency's prepare script until allowlisted — copy the exact package key pnpm prints into the profile's pnpm-workspace.yaml:

allowBuilds:
  dsh-noletme: true

then re-run the add.

Development (local source)

Rebuild after editing, then point the patch at the package:

pnpm install && pnpm build
dsh web --patch 'D:/OneDrive/桌面/play/codes/dsh-plugin/NoLetMe/cordis.patch.yml'

Windows note. The cordis.patch.yml row uses the package name (dsh-noletme), so the overlay works only when the package is installed into the profile (the dsh plugin --profile web add ./NoLetMe step above). Naming the row with a raw absolute Windows path fails — the ESM loader rejects D:\… entry names (ERR_UNSUPPORTED_ESM_URL_SCHEME). On Linux a file:// URL works as an alternative.

Build

pnpm install      # devDependencies: tsdown, lightningcss, typescript, react types
pnpm typecheck    # optional; tsc --noEmit
pnpm build        # tsdown → lib/index.js (node half) + lib/client.js (browser bundle)

The browser bundle is a window.__ModuleLoader__.load(...) closure-factory artifact (same shape as the harness's own clientBundle preset): platform modules resolve through the frozen module table, everything else inlines, and *.module.css compiles to a hashed class map with auto-injected styles.

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.<sessionId>), 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.

Architecture

src/
├── index.ts            # Node (host) half — no-op, satisfies the Loader
└── client/
    ├── index.ts        # browser bundle entry (apply/inject)
    ├── apply.ts        # registers the shell.overlay entry + the stats store
    ├── slots.ts        # inject-face + composed-props contracts
    ├── session-source.ts # current-session ConversationSnapshot observable
    ├── session-store.ts  # stats store: live folding, full-history paging, persistence
    ├── accumulator.ts  # per-session incremental fold + compaction + serialize
    ├── keywords.ts     # the research-backed taxonomy
    ├── stats.ts        # counting engine (longest-match walk, per-block cache)
    ├── NoLetMePanel.tsx / .module.css
    └── locales.ts      # zh + en dictionaries

• Reasoning streams arrive as reasoning-delta chunks; the session layer accumulates them into ConversationSnapshot.partial (published at most once per animation frame). Finalized turns land in snapshot.nodes. The stats store folds both incrementally (per-block counts cached by block identity, new nodes gated by a seq high-water mark) and republishes a ready TrajectoryStats — the panel never recomputes a session from scratch.
• shell.overlay is the layout's frame-wide additive seat; the panel is a click-through-opt-in floating surface that mirrors the DetailsPanel styling.

License

MIT