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

Five 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.

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

Write and Edit rows carry a +N (added) / -N (removed) badge right after the file path, visible without expanding the row. A failed edit produces no diff, so it gets no badge.

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 also reports how much context the last call actually received — the first thing to check when a rewrite is unhelpful.

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.

Configuration

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/fast-context-gate.test.mjs
node test/content-embed.test.mjs
node --check lib/client.js

Tests load shipped modules from lib/client.js / lib/index.js. For slot keys, priorities, the reference-extraction contract, the failure-status map, and the full file structure, see docs/technical-reference.md.

License

MIT