didev — AI-Powered Development CLI

Персональная команда AI-разработчиков прямо в вашем терминале.
A specialized AI agent family for your codebase, powered by DeepSeek API.

Автор: Симонов Михаил Сергеевич

What is didev?

didev is a CLI tool that brings an entire team of specialized AI agents to your terminal. Unlike generic AI assistants, each agent in didev knows its role — frontend, backend, architecture, security, testing — and works in a coordinated pipeline to complete real development tasks.

Key difference from ChatGPT / Copilot: didev reads your actual project files, understands your codebase structure, and agents pass context to each other — each building on the previous agent's work.

Models

| Model | Use case | |-------|----------| | deepseek-v4-flash (default) | Fast, 1M context window, 384K max output. Daily tasks. | | deepseek-v4-pro | Thinking model — complex architecture decisions |

Note: deepseek-chat and deepseek-reasoner aliases are deprecated as of 2026-07-24.

Installation

npm install -g didev

Requirements: Node.js 18+

Quick Start

# 1. Install globally
npm install -g didev

# 2. Go to your project
cd my-project

# 3. Initialize didev
didev init

# 4. Set your DeepSeek API key
didev config set DEEPSEEK_API_KEY=sk-xxxxxxxxxxxxxxxx

# 5. Start chatting with your codebase
didev chat

# Or run the full AI agent pipeline on a task
didev agent "Add JWT authentication to Express API"

Get a DeepSeek API key at platform.deepseek.com → API Keys.

Running didev with no arguments opens interactive chat immediately.

Commands

didev init

Initialize didev in the current project. Auto-detects language, framework, and project type.

didev init                 # interactive
didev init --frontend      # force frontend type
didev init --backend
didev init --fullstack
didev init -y              # non-interactive

Creates .didev/ with config, context file, sessions dir, and BMad dirs.

didev config

didev config set DEEPSEEK_API_KEY=sk-xxx
didev config set api.model=deepseek-v4-pro
didev config set api.temperature=0.5
didev config show

API key stored in ~/.didev/config.yaml — never committed to the project.

didev chat

Interactive REPL with full project context, token tracking, and shell execution.

didev chat
didev chat -m deepseek-v4-pro
didev chat -f src/api/routes.ts

Slash commands: /menu /agent <task> /review /refactor /apply /files /add <path> /context /context update /save <name> /load <name> /sessions /config /clear /help /exit

Token counter

Every response shows a real-time context usage bar:

ctx [████░░░░░░░░░░░░░░░░] 12.3k/1.00M  +0.8k out  ⚡ cache 73%

• Bar fills as context grows toward the model's limit (1M for v4-flash, 64K for v4-pro)
• ⚡ cache N% shows the DeepSeek prompt cache hit rate — high % = faster + cheaper responses

AI tools available in chat

| Tool | Description | |------|-------------| | read_file | Read any project file | | write_file | Propose file changes (requires confirmation) | | list_files | List files by glob pattern | | search_code | Full-text search across project files | | git_diff | Show git diff | | run_command | Execute shell commands in the project directory |

The run_command tool lets the model run build/test commands directly:

you › проверь запускается ли проект нет ли ошибок компиляции?
▶  Запускаю: npm run build
didev › Build successful — 0 errors, 2 warnings...

Works for any stack: npm run build, mvn compile, ./gradlew test, tsc --noEmit, etc.

File write confirmation

When the AI proposes file changes, you get a 4-choice prompt:

╭─────────────────────────────────────────╮
│ ✏️  Подготовлено 3 файл(а)              │
│   + src/auth/jwt.ts     новый           │
│   ~ src/app.ts          изменён         │
│   ~ src/config.ts       изменён         │
╰─────────────────────────────────────────╯
  [Enter/y] Применить  [d] Diff  [n] Отклонить  [a] Авто-режим

• Enter/y — apply all changes
• d — show colored diff first, then apply
• n — reject; you can then type corrections and the AI will try again
• a — auto-mode: apply all future writes in this session without asking

After applying, didev prompts whether to run tests or code review on the changed files.

didev agent

Run the multi-agent pipeline on a development task.

didev agent "Add Stripe payment integration"
didev agent --type frontend "Build a responsive dashboard"
didev agent --mode light "Fix the login bug"
didev agent --mode developer-only "Add input validation"
didev agent --skip Tester "Quick feature"

Orchestration modes: | Mode | Pipeline | |------|----------| | full | Analyst → Architect → Developer → [Reviewer ∥ SecurityAuditor ∥ Tester ∥ PerformanceAuditor] | | light | Analyst → Developer → Reviewer | | developer-only | Developer only |

В режиме full стадия ревью выполняется параллельно (все 4 агента одновременно) — это сокращает общее время пайплайна.
Reviewer, Tester и PerformanceAuditor автоматически используют быструю модель (api.fastModel), если она задана в конфиге.

Agent feedback loop

When you reject an agent's file writes, you can provide corrections inline. Up to 3 re-run rounds are supported:

• Regular feedback → Developer agent re-runs with your corrections
• Architectural feedback (keywords: architecture, structure, pattern, approach, design) → Architect re-runs first, then Developer gets the updated architecture plan

Context from earlier pipeline stages (Analyst/Architect) is preserved across all feedback rounds.

didev review

didev review               # review git diff
didev review --all         # all source files
didev review --file src/api.ts
didev review --pr          # vs main branch

didev refactor

didev refactor "Extract service layer from controllers"
didev refactor --auto
didev refactor --dry-run "Rename userId to user_id"

didev mcp

Manage MCP (Model Context Protocol) servers. Connected tools become available to all agents and chat.

# From GitHub / GitLab / Bitbucket (clones, installs, builds)
didev mcp add https://github.com/owner/repo.git
didev mcp add https://gitlab.com/owner/repo.git
didev mcp add https://bitbucket.org/owner/repo.git

# From npm
didev mcp add @modelcontextprotocol/server-filesystem

# Interactive wizard
didev mcp add

didev mcp list
didev mcp test
didev mcp remove my-server

didev mcp add <git-url> flow:

1. Clones repo to ~/.didev/mcp/<name>/
2. Detects Node.js / Python
3. npm install + npm run build (or pip install)
4. Detects entry point automatically
5. Reads README → finds required env vars
6. Private repo? → prompts for GitHub/GitLab token or Bitbucket credentials
7. Credentials saved to .didev/.env (gitignored); config stores ${VAR} references
8. Tests MCP connection, lists available tools
9. Saves to .didev/config.yaml

MCP servers auto-start when you run didev — no manual setup needed per session. npx-based servers get 90 seconds on first run to download packages.

didev bmad

BMad Method — Behavior-Motivated Agile Development.

didev bmad init
didev bmad epic
didev bmad story
didev bmad implement
didev bmad review

Build & Test Integration

didev automatically detects your project's build system and offers to run checks after applying changes:

| Project type | Detected by | Commands offered | |---|---|---| | Node.js / React | package.json | npm run build, npm test, npm run lint | | Angular | angular.json | ng build, ng test, ng e2e | | Maven | pom.xml | mvn compile, mvn test | | Gradle | build.gradle | ./gradlew build, ./gradlew test | | Frontend subdir | src/package.json | same as Node.js, in subdirectory |

Build errors are parsed and surfaced with file/line references.

Agent Families

Why specialized agents?

Each agent has a focused system prompt optimized for its role. A frontend agent doesn't need to reason about database internals. A security auditor doesn't write React components. Specialization produces significantly better output than one generic "do everything" agent.

Think of it like a craftsman's workshop: each wall holds tools for a specific trade. The carpenter's wall has chisels and planes; the electrician's wall has testers and crimpers. The master (Orchestrator) knows which specialist to engage and when.

Frontend Family

| Agent | Specialty | |-------|-----------| | FrontendAnalyst | UX requirements, component design | | FrontendArchitect | State management, component architecture | | FrontendDeveloper | React, CSS, TypeScript implementation | | Reviewer | Code quality, accessibility | | Tester | RTL tests, Playwright |

Backend Family

| Agent | Specialty | |-------|-----------| | BackendAnalyst | API design, data modeling | | BackendArchitect | Service architecture, DB schema | | BackendDeveloper | Node.js / Java / Python implementation | | SecurityAuditor | OWASP, auth review | | Reviewer | Error handling, code quality | | Tester | Unit + integration tests |

Fullstack Family

Analyst → Architect → Developer → [SecurityAuditor ∥ Reviewer ∥ Tester ∥ PerformanceAuditor]

MCP Integration

MCP tools are namespaced as serverName__toolName to avoid conflicts:

github__create_issue
postgres__query
filesystem__read_file

MCP connections are managed by a singleton McpManager shared across chat and all agents in the same session. Calling connectAll multiple times is safe — already-connected servers are skipped automatically.

Popular servers

| Server | Command | |--------|---------| | Filesystem | didev mcp add @modelcontextprotocol/server-filesystem | | GitHub | didev mcp add @modelcontextprotocol/server-github | | PostgreSQL | didev mcp add https://github.com/Neovaryag/mcp-pgs-tool.git | | Java CVE scanner | didev mcp add https://github.com/Neovaryag/java-vuln-remediation.git | | Brave Search | didev mcp add @modelcontextprotocol/server-brave-search | | Memory | didev mcp add @modelcontextprotocol/server-memory |

Project Structure

.didev/
  config.yaml      # project config (commit this, no secrets inside)
  context.md       # project context for AI (commit this)
  .env             # MCP credentials (gitignored)
  sessions/        # chat history (gitignored)
  bmad/            # epics, stories, architecture docs

~/.didev/
  config.yaml      # global config — API key lives here
  .env             # global git credentials for private MCP repos
  mcp/             # globally cloned MCP servers

Security Model

| What | Where stored | Committed? | |------|-------------|------------| | DeepSeek API key | ~/.didev/config.yaml or env var | Never | | MCP env vars (DB URL, tokens) | .didev/.env | No (gitignored) | | Git tokens for private MCP repos | ~/.didev/.env | No | | Project config | .didev/config.yaml with ${VAR} placeholders | Yes |

Path traversal protection: all read_file / write_file calls resolve the path against the project root and reject any path that escapes it (e.g. ../../etc/passwd → Access denied).

Configuration Reference

# .didev/config.yaml
api:
  provider: deepseek
  apiKey: ${DEEPSEEK_API_KEY}
  model: deepseek-v4-flash
  # fastModel: deepseek-v4-flash   # optional: cheaper model for Reviewer/Tester/PerformanceAuditor
  maxTokens: 32768
  temperature: 0.3
  baseUrl: https://api.deepseek.com/v1

context:
  maxFiles: 100
  maxTokens: 128000
  autoDiscover: true
  excludePatterns: [node_modules/**, dist/**, .git/**, "*.lock"]

agents:
  family: full        # full | light | custom
  orchestrate: true
  reviewRequired: true
  autoApply: false    # prompt before writing files

mcp:
  servers:
    - name: postgres
      command: node
      args: [/home/user/.didev/mcp/mcp-pgs-tool/dist/index.js]
      env:
        DATABASE_URL: ${DATABASE_URL}

Getting DeepSeek API Key & Billing

1. Go to platform.deepseek.com
2. Sign up / log in
3. API Keys → Create new secret key → copy it
4. Billing → Top up balance (minimum ~$5)
5. Set the key:

   didev config set DEEPSEEK_API_KEY=sk-xxxxxxxxxxxxxxxx

Pricing (2026):

• deepseek-v4-flash: ~$0.27 / 1M input tokens · ~$1.10 / 1M output tokens
• Context window: 1,000,000 tokens
• Max output per response: 384,000 tokens

DeepSeek API is 10–20× cheaper than OpenAI GPT-4o for equivalent tasks.

Architecture

src/
  index.ts              CLI entry (Commander.js) + graceful shutdown (SIGTERM/SIGINT)
  core/
    api.ts              DeepSeek client — chat, stream, tool loop, token usage tracking
    config.ts           Config load/save (global + local merge)
    context.ts          Project detection + file loading
    file-manager.ts     Read/write/diff — assertSafePath, 10MB limit
    mcp.ts              MCP manager — connect, call, tool namespacing, idempotent connectAll
    post-apply.ts       Post-apply build/test detection and prompts
    session.ts          Chat session persistence
  agents/
    base-agent.ts       BaseAgent — tools: read/write/list/search/git_diff/run_command + MCP
    analyst.ts          Analyst variants (frontend / backend / fullstack)
    architect.ts        Architect variants
    developer.ts        Developer variants
    reviewer.ts         Reviewer + SecurityAuditor
    tester.ts           Tester + PerformanceAuditor
    orchestrator.ts     Pipeline builder — parallel review phase, feedback loop
  cli/commands/
    init.ts             didev init
    config.ts           didev config
    chat.ts             didev chat — REPL, token counter, run_command, 4-choice confirm
    agent.ts            didev agent
    review.ts           didev review (read_file + write_file tools only)
    refactor.ts         didev refactor
    mcp.ts              didev mcp (add/list/test/remove)
    menu.ts             interactive menu — initializes MCP if not yet done
  bmad/method.ts        BMad workflow
  utils/
    logger.ts           Styled output (chalk, ora, boxen)
    git.ts              Git diff helpers — with timeout
    resilience.ts       withTimeout + retryWithBackoff
    build-runner.ts     Build script detection (npm/Maven/Gradle/Angular) + error parsing
    token-counter.ts    Token estimation utilities

Отказоустойчивость

| Механизм | Где применяется | Детали | | -------- | --------------- | ------- | | Таймауты | API запросы, MCP, git, run_command | 90s stream, 60s run_command, 20–90s MCP connect (90s для npx) | | Retry с backoff | DeepSeek API | 5 попыток, экспоненциальный backoff + jitter, только транзитные ошибки | | Graceful shutdown | SIGTERM / SIGINT | Закрывает MCP серверы, 5s на очистку, затем force exit | | Параллельный review | Orchestrator full mode | SecurityAuditor + Reviewer + Tester + PerformanceAuditor запускаются одновременно | | Лимит файлов | file-manager | Отказ записи если файл >10MB | | Безопасные пути | file-manager | assertSafePath — блок path traversal (../../) | | Идемпотентный MCP | McpManager | connectAll безопасно вызывается несколько раз — дублей нет | | Feedback loop | Orchestrator | До 3 итераций; Developer + Architect (при архитектурных правках) при отклонении изменений | | Логирование ошибок | config, context, session | Все catch логируют через logger.warn/debug |

License

MIT — Симонов Михаил Сергеевич, 2026