Voice Interview Coach 🎙️

A DSH (DeepSeek Harness) plugin that turns the chat into a voice-driven mock interview. Give it your project, language, resume, and target role in natural language; it asks progressively harder questions, listens to your spoken answers, and finishes with a gap-analysis review you can export to Markdown and persist to memory.

Features

🗣️ Voice interview — uses the browser's Web Speech API for both speech recognition and text-to-speech. A text fallback is always available.
🎯 Natural-language difficulty — describe the target role and seniority any way you like (e.g. "高级后端开发工程师，要求高并发和系统设计，偏 P7"). The model adjusts question depth automatically.
🧩 Question type chips — pick any mix of 项目深挖 / 系统设计 / 编程语言基础 / 八股原理 / 行为面试 / 线上故障 / 算法编码 / 工程实践.
📄 Resume upload — supports .txt, .md, .json, .csv natively, plus PDF and DOCX parsed in the Host with no native dependencies (a tiny built-in FlateDecode / zip-XML reader).
📝 Structured review — scores per dimension, strengths, gaps with study plans, better-answer suggestions, and a prioritized action list.
💾 Memory — every review (with transcript) is written to .interview-memory/interview-<timestamp>.md inside the current workspace so future sessions can read it.
📈 History & trend analysis — list past interviews, select any subset, and ask the model to produce a trend report highlighting recurring weak spots.
⬇️ One-click Markdown export.

Requirements

DSH >= 0.1.0-rc.6 (Cordis dynamic-plugin runtime)
A configured llm Provider in your DSH deployment
A Chromium-based browser (Chrome / Edge / Arc) for speech recognition; Firefox/Safari can still use text mode

Installation

From a DSH plugin marketplace

Once the package is published, install it from the Plugins settings page in DSH.

From this repository

# 1. clone or download this repo into your DSH plugin path
git clone https://github.com/<your-org>/dsh-voice-interview-coach.git

# 2. add a row to your agent preset or host composition (cordis.yml):
#    - id: voice-interview-coach
#      apply: ./dsh-voice-interview-coach

Because the plugin is plain JavaScript with zero npm dependencies, there is no npm install step. The Host requires Node's built-in zlib (already available).

Usage

Open the DSH session where the plugin is loaded. Run the cordis_define / cordis_run card (or the plugin's own entry point).
Fill in:
- Target role / difficulty — natural language
- Focus / style — what the interviewer should probe (optional)
- Question types — chip multi-select
- Project, programming language, resume (paste or upload)
- Question count (3–15)
Click 开始语音面试. The interviewer speaks the first question.
Click 🎙 语音回答 (or type in the textarea and press Enter) after each question.
When the count is reached — or you click 结束并复盘 — the review appears.
Use ⬇ 导出 Markdown to download it, 💾 写入记忆 to re-save it, and switch to the 历史对比 tab to compare multiple interviews.

How it works

┌─────────────── Browser (Client) ───────────────┐      ┌──────────── Host (Node) ────────────┐
│  React UI in `tool.view.cordis` slot           │      │  harness.handle():                  │
│  • SpeechRecognition / SpeechSynthesis         │ RPC  │   • startInterview / submitAnswer   │
│  • File upload (base64 for PDF/DOCX)  ─────────┼─────►│   • parseResume (PDF/DOCX via zlib) │
│  • Markdown download, history selection        │      │   • listHistory / compareHistory    │
│                                                │      │   • saveMemory → <workspace>/.interview-memory/ │
└────────────────────────────────────────────────┘      │  ctx.llm.stream(...) for Q&A/review │
                                                        └─────────────────────────────────────┘

The plugin follows the standard Cordis lifecycle — every side effect is tied to the plugin fiber, so stopping/uninstalling removes all listeners and UI registrations cleanly.

Configuration

Field	Where	Default	Description
Memory directory	Host	`<workspace>/.interview-memory/`	Where review Markdown files are stored
Resume size limit	Client	5 MB upload, 20 000 chars after parsing	Larger resumes are truncated before being sent to the model
LLM provider	Host	first registered provider	Selected via `ctx.llm.listProviders()[0]`; change in code or configure your deployment's default
Speech language	Client	`zh-CN`	Change `rec.lang` / `SpeechSynthesisUtterance.lang` in `src/client.js`

Development

# Verify package metadata
npm run lint:meta

# Dry-run the npm tarball
npm run pack

To iterate inside a live DSH session, edit src/host.js / src/client.js and either:

reload the plugin row in your DSH composition, or
copy the function bodies into a cordis_define call to test instantly.

Project layout

.
├── package.json           # plugin manifest (the `dsh` block is the DSH entry)
├── src/
│   ├── host.js            # Node half: LLM, PDF/DOCX parsing, memory, history
│   └── client.js          # Browser half: React UI, speech I/O, file upload
├── assets/icon.svg        # plugin icon
├── scripts/verify-package.js
├── docs/                  # longer design notes
├── README.md
└── LICENSE

Roadmap

Optional cloud ASR (Volcengine/Whisper) for better Chinese recognition
JD/URL ingestion for job description alignment
Multi-session concurrent interviews (currently one per fiber)
Per-skill review trends across weeks/months
English UI

Privacy

Resumes and interview transcripts are sent to whichever llm provider your DSH deployment is configured to use. They are also stored as plain Markdown under .interview-memory/ inside the workspace. Do not use this plugin with sensitive personal data unless you trust both the model provider and the workspace storage.

voice-interview-coach