dsh-assistant-optimization

A DSH plugin for the web profile that makes model output easier to read and your prompts easier to write.

中文文档

Features

Seven capabilities, all plug-and-play. Official rendering is never replaced — the plugin shadows DSH components and delegates back to them, so Markdown, tool cards, images, and tables keep working exactly as shipped.

	Capability	What it does
💭	Reasoning fold	Collapses mis-rendered chain-of-thought into the native "Think" block
📊	Mermaid diagrams	Renders mermaid code blocks as interactive SVG (zoom / pan / touch)
✎	Tool-call grouping	Collapses consecutive tool-call rows into one # bash · N tools header. Reasoning rows between calls do not break the group — merge across them; collapsed groups tuck those reasoning rows in, expanded groups indent them like members
✨	Prompt enhance	Rewrites a rough draft into a clearer instruction in one click
▶	Resume-from-breakpoint	After a manual stop or session error, the send button becomes a play key — hover shows a tooltip, one click resumes from the interruption with the currently selected model. Typing or starting a new turn instantly restores the normal send button.
📁	Turn folding	When a turn finishes, its process (thinking, tool calls, intermediate text) auto-collapses into one "已完成 · 时长" header; the final summary reply stays expanded. Interrupted-and-resumed turns merge into one group; mid-run steering messages collapse with a "· N 条插话" count. The plugin takes over the built-in transcript mode so folds never double up.
🛰️	Semantic search	context_search — locate code from a vague description (Windsurf-backed)

Reasoning fold

Some providers (OpenAI-compatible APIs with no dedicated reasoning_content channel) dump chain-of-thought into the regular text stream:

The user wants to know about X, I should explain Y first...
</thinking>

Here is the answer to your question...

DSH renders that as a single block. This plugin splits it at configurable markers so the reasoning folds into the native "Think" disclosure and only the answer stays visible. Manage the markers under Settings → General → Thinking Tag Markers.

Mermaid diagrams

Mermaid code blocks render inline as interactive SVG. Toolbar buttons zoom and reset, the scroll wheel zooms, and mouse or touch drag pans.

Edit diff counts (retired)

dsh 0.1.2+ renders native diff statistics on Write & Edit rows, so the plugin's own badges were retired in v1.8.0. The module is kept for older hosts but registers nothing on current versions.

Prompt enhance

A sparkle button sits left of the send button and rewrites a rough draft into a clearer instruction using the model already selected in the composer. It sends one plain chat request — no session, nothing logged — and replaces the draft in place, so Ctrl/Cmd+Z undoes it.

While running it shows a spinning arc and "增强中"; on success a green check; on failure the icon flashes red with the reason in its tooltip and the draft is never cleared. The tooltip uses the same CSS styling as DSH's native button tooltips and also reports how much context the last call actually received — the first thing to check when a rewrite is unhelpful.

Resume-from-breakpoint

When a conversation stops abnormally — the user clicks Stop or the session hits an error — the send button transforms into a ▶ play key. A hover tooltip ("断点续发") appears with the same style as DSH's native button tooltips.

• One click sends a resume signal via the host route; the agent picks up from the interruption using whatever model is currently selected in the composer.
• Instant revert: typing in the draft, or the agent starting to run, immediately restores the normal send/stop button — no stuck play icon.
• Gate logic: reads session.chat.timeline for the last closed turn's turn/end reason. Only aborted (user stop) and error (session error) trigger the play button; normal completions and max-tokens do not.
• Empty marker rows: when the resume marker enters the transcript, the plugin replaces the blank bubble + copy button with a subtle "已从中断处继续" hint line.
• Implementation: CSS-overlay approach — the official button's SVG is hidden via data-dsao-resume attribute + a play SVG sibling, so React's re-render cycle is never disrupted.

Turn folding

Replaces the built-in fold. dsh 0.1.2+ ships its own turn-process folding (Settings → Conversation display → Compact). This plugin now drives that setting: while Turn Folding is enabled here, the built-in mode is forced to Normal so the plugin’s own fold (duration + steering counts, resume-chain merge) is the only one acting; disabling Turn Folding here hands the column back to the built-in Compact. Toggling either side realigns both.

While a turn runs, everything stays visible — the native "Deep diving…" status with its clock is the running indicator. The moment the turn completes (the final answer lands), the process content of that turn — thinking rows, tool calls, intermediate text — collapses into a single header line: 已完成 · 时长 (errored/stopped turns show 已出错/已停止 instead). Mid-run steering interjections collapse with the turn; the header shows「· N 条插话」when present. What stays visible: your prompt, the final summary reply, and its action row (copy etc.). Click the header to expand the full process; click again to collapse. Expand/collapse choices live in memory only; a fresh page load starts collapsed.

The plan is computed entirely from the session snapshot: turn grouping via chat.locations, completion via the turn's turn/end reason, the summary reply via the official turn-tail node's closing pointer, duration from the turn's start/end timestamps (same figures the native clock shows). Toggle under Settings → General → Turn Folding.

Semantic search (context_search)

A host-side tool for vague or unclear searches: pass a natural-language query and get back the matching files with their line ranges and code. It runs an agentic search loop built on the fast-context approach and driven by a Windsurf key.

Key gating — the whole point. The tool and its one-line prompt guidance are registered only when a Windsurf key resolves. With no key, nothing is registered, so the model is never told about a tool it cannot call.

The key is resolved in this order (first hit wins): WINDSURF_API_KEY env → manual entry (Settings → General → Windsurf API Key, stored at ~/.dsh/dsao-windsurf-key, 0600) → local auto-read from the logged-in Windsurf/Devin editor's state.vscdb. DSAO_FC_AUTO_KEY=0 disables auto-read. See lib/fast-context/NOTICE.md for the non-official protocol note.

Installation

dsh plugin --profile web add github:qianshe/dsh-assistant-optimization
dsh web

Open http://127.0.0.1:3080 — the plugin activates automatically. A restart is required after install, update, and removal.

From source	dsh plugin --profile web add .
Update	dsh plugin --profile web update dsh-assistant-optimization
Uninstall	dsh plugin --profile web remove dsh-assistant-optimization

Configuration

Setting	Default	Description
Thinking Tag Markers	["</thinking>"]	Strings that split reasoning from body text. Multiple supported. Edit at Settings → General.
Turn Folding	on	Auto-collapse finished turns' process into a one-line header. Edit at Settings → General.
Windsurf API Key	—	Credential for context_search. See key resolution above. No key ⇒ tool not registered.

Requirements

• DSH (DeepSeek Harness) with the web profile
• Internet access only for the mermaid.js CDN, and only when a diagram is present

Development

node test/diff-stats.test.mjs
node test/ensure-badge.test.mjs
node test/context.test.mjs
node test/prompt-enhance.test.mjs
node test/host-prompt-enhance.test.mjs
node test/resume-gate.test.mjs
node test/resume-route.test.mjs
node test/resume-continuity.test.mjs
node test/turn-fold.test.mjs
node test/fast-context-gate.test.mjs
node test/content-embed.test.mjs
node test/turn-fold-sync.test.mjs
node scripts/repro-switch-back.cjs
node scripts/lib-sync-check.cjs
node scripts/build-client.cjs verify
node --check lib/client.js

Tests load shipped modules from lib/client.js / lib/index.js. lib/client.js is a build artifact: edit src/ and regenerate with node scripts/build-client.cjs build — never edit by hand. For slot keys, priorities, the reference-extraction contract, the failure-status map, and the full file structure, see docs/technical-reference.md.

License

MIT