dsh-survey

Ask the user ten questions at once instead of ten times in a row.

do_a_survey is a tool plugin for DeepSeek Harness. The model sends a whole questionnaire; the user answers it in one card and submits once.

When you'd want this

• The agent needs five decisions before it can start. Without this it asks one, waits, asks the next — five round trips before any work happens. Here it asks once and you answer the lot in a single card.
• You want the answers back as data, not prose. Each reply arrives as { id, selected, custom?, skipped? }, so the model never has to parse "yeah the first one, and skip the last question" out of a sentence.
• Some of it you genuinely don't care about. Every question has a skip, so a survey that asks more than you want to answer costs you a click, not a negotiation.

When the model calls do_a_survey, the Web UI renders the survey by mode:

• compact — single-question card
• inline — embedded in the conversation column
• overlay — fullscreen, for compare questions
• grid — matrix of many simple questions

All text supports Markdown (code blocks, blockquotes, inline code, bold) and color ({color:red}text{/color}). A readable two-column recap follows the submit.

Preview

Screenshots of the real toolview, captured from the shipped bundle.

Grid matrix — fullscreen overlay for many simple questions, one card each:

Inline — survey embedded in the conversation column, filled and submitted together:

Overlay compare — fullscreen side-by-side comparison with Markdown:

Compact — single-question card with rich question types:

Install

From npm (the scoped name; the unscoped dsh-survey on the registry is not this package):

dsh plugin --profile web add @jinhuang712/dsh-survey
# restart dsh web, then refresh the page

Or pinned to a release — build artifacts are committed, so there is nothing to build:

dsh plugin --profile web add "github:jinhuang712/dsh-survey#v1.1.4"
# restart dsh web, then refresh the page

Or track main to pick up unreleased commits:

dsh plugin --profile web add "github:jinhuang712/dsh-survey#main"

Or from a local checkout, linked so edits show up on refresh:

git clone https://github.com/jinhuang712/dsh-survey.git
dsh plugin --profile web add "link:$PWD/dsh-survey"

If your dsh profile directory is a pnpm workspace, pnpm asks for -w before touching its root — pass it through: dsh plugin --profile web add -w ….

After install, the do_a_survey tool and its survey UI are available permanently.

The companion skill dsh-survey registers with the install (dsh.skills declaration):

• a usage guide covering the four modes and five question types
• a dynamic-plugin fallback recipe (references/dynamic-plugin-fallback.md) for environments without the bundle

Usage

Tell the model what you want to collect; it calls do_a_survey(mode, questions). mode is required:

Example:

{
  "mode": "inline",
  "questions": [
    {
      "id": "q1",
      "header": "Runtime",
      "question": "Which runtime does this ship on?",
      "options": [
        { "label": "Node 20 LTS (recommended)", "description": "Matches the CI base image." },
        { "label": "Bun 1.1", "description": "Faster boot, fewer native addons." }
      ]
    },
    {
      "id": "q2",
      "header": "Scope",
      "question": "Which surfaces need the rewrite?",
      "multi_select": true,
      "options": [{ "label": "Web client" }, { "label": "CLI" }]
    },
    { "id": "q3", "kind": "boolean", "question": "Ship behind a flag first?" },
    { "id": "q4", "header": "Notes", "question": "Anything the changelog should call out?" }
  ]
}

header is an optional short heading shown beside the question number. A trailing (recommended) on a label becomes a badge.

Question types

Each answer is { id, selected, custom?, skipped? }. Question ids must be unique within one survey, and a survey nobody submits within 30 minutes fails on timeout rather than hanging the call.

Features

• Full Markdown — question text, option labels/descriptions, compare blocks and recap all render through the official safe renderer
  • micromark + protocol allowlist + shiki highlighting
  • code blocks, blockquotes, inline code, bold
• Color — {color:red}text{/color} (named / #hex / rgb()), usable in questions, options, compare blocks
• Skip / restore — per-question ✕ grays out, ↺ restores; submitted as skipped: true
• Fullscreen overlay — mode: "overlay" centers fullscreen (mask + 1180px), breaking past the 748px conversation column
• Grid matrix — mode: "grid" fullscreen grid of many simple questions
  • every card is the same size, controls sit under the question, per-card skip
  • card text is inline-only: a fenced block collapses to inline code and hard breaks to spaces, so one long question cannot blow open the whole matrix
  • compare questions degrade to a left/right choice
• Readable recap — strict two-column grid, one "question → answer" row each
• Follows your language — card copy ships in English and Chinese, tracking the Web UI's language setting
  • falls back to the browser's language where that setting is unavailable
  • answers stay language-neutral, so switching never changes what the model receives
• Accessibility — radio/checkbox semantics + keyboard focus rings

Architecture

• Host half (lib/index.mjs): Cordis entry
  • defineTool registers do_a_survey with a 30-minute timeoutMs
  • webServer.register serves /api/dsh-survey/submit|cancel
  • execute suspends on the answer, correlated by exec.callId; submit, cancel, abort, timeout and unload each release it
• Client half (src/ → lib/client.js): __ModuleLoader__.load bundle registering tool.call.toolview key=do_a_survey
  • runtime.js binds the host's React and UI primitives from the loader's require — they are never bundled, so the plugin shares the host's React instance
  • styles.css the stylesheet, injected once per page
  • i18n.js zh/en copy, markdown.js Markdown and colour, answers.js pick ↔ answer mapping
  • controls.js the answer controls, model.js draft state and submit/cancel
  • modes/ one file per presentation: compact, survey (inline + overlay), grid, recap
• Skill (skills/dsh-survey/SKILL.md): usage guide + dynamic-plugin fallback recipe (references/dynamic-plugin-fallback.md)

Develop

lib/client.js and its source map are build output, committed so the GitHub install stays one line. Edit src/, then:

pnpm install
pnpm build      # esbuild src/index.js -> lib/client.js + lib/client.js.map

The host serves the map at /plugins/dsh-survey/client.js.map, so breakpoints land in src/.

Verify

It loaded at all:

• __DSH_BOOT__ includes the dsh-survey client row, and /plugins/dsh-survey/client.js returns 200
• cordis_inspect_query (Tool.listTools) lists do_a_survey

Then run a survey covering all four modes and check what the card does:

Uninstall

• Remove the dsh-survey insert row from the web profile's cordis.patch.yml
• Remove the dsh-survey dependency from the web profile's dsh.profile.bundles and run pnpm remove

License

MIT