Visible Technical Details

Agent supports Plan and Auto execution modes. Plan mode generates a plan for your approval first. Auto mode automatically chains tool calls, up to 15 iterative rounds: think → act → observe → re-think.

• 6 built-in tools: read_file / write_file / edit_file / run_command / search_code / list_directory
• RAG context auto-injection + project tree auto-loading
• package.json auto-parsing, tech stack recognition

After writing code, Agent automatically runs validation commands (tsc / eslint / lint / build). Detects errors and auto-fixes them, up to 3 rounds. Stops automatically when progress stalls to save Tokens.

• Auto-detects project config: tsconfig / eslint / package.json scripts
• Stagnation detection: stops when errors stop decreasing

Pub-sub based usage tracking system. Automatically records Prompt Tokens, Completion Tokens, and costs after each AI request. Persisted in localStorage + Electron Store dual storage.

• 26+ model pricing tables (DeepSeek / GPT-4o / Claude / Gemini, etc.)
• Character-based Token approximation with CN/EN distinction
• Cumulative + session dual-dimension stats, auto UI updates

TF-IDF similarity-based engine. During indexing, code files are automatically split by function and class boundaries. Search uses camelCase tokenization + CN-EN synonym expansion.

• 500+ CN-EN synonym pairs: login ↔ auth ↔ 登录 ↔ 认证
• 30+ file types: ts / tsx / js / py / rs / go and more
• Auto-rebuild index on file changes + import/require dependency extraction

Skills are packaged AI assistant configurations. Each skill has its own Prompt, available tools, and model parameters. 4 built-in skills, supports import/export.

• Built-in: Cipher Agent / Novel Assistant / PPT Designer / Excel Data
• Import: .zip / .json / .md, Export: .zip / .json
• Skill attachments: extra prompts / scripts / config files

7 built-in provider presets, each independently toggleable. Custom providers support custom API URLs, Keys, and model lists. One-click default switch with built-in connection test.

• Presets: Xiaomi / NVIDIA NIM / DeepSeek / OpenAI / Anthropic / OpenRouter
• Config: maxTokens / temperature / topP / toolCallRounds

Global rules apply across all projects, project rules are specific to the current project. Auto-reads CLAUDE.md, .cursorrules, .windsurfrules from the project directory. Each rule can be independently enabled or disabled.

Electron main and renderer processes communicate via 13 IPC handlers covering file R/W, terminal execution, config persistence, and security trust confirmation. Sandbox mode isolates processes.

• Path trust confirmation: user must confirm first folder open
• Trust list persistence, auto-saves trusted paths

4 theme modes (Dark / Dark Blue / Light / System), ~50 semantic CSS variables. Glassmorphism frosted glass via backdrop-filter + semi-transparent backgrounds, 300ms smooth transitions.

Working memory tracks file operations (create / modify / delete) within a session, 24-hour TTL auto-expiry. Project notes persist across sessions. Memory auto-formatted and injected into Agent context.