context-mode

Context Preservation Rules

When executing commands that may produce large output (>5KB), use the context-mode MCP tools:

• Instead of running shell commands directly, use context-mode.execute with language: "shell"
• Instead of reading large files, use context-mode.execute_file
• When fetching documentation, use context-mode.fetch_and_index + context-mode.search
• For any tool returning large output, use context-mode.compress

Commands likely to produce large output

• git log, git diff, cat large files, find, reading log files
• npm list, pip list, dependency audits, yarn why
• API responses, test suite output (>100 tests)
• Browser snapshots, web page content, database dumps

Example usage

// Instead of: bash("git log --oneline -50")
execute({ language: "shell", code: "git log --oneline -50", intent: "recent changes" })

// Instead of: read_file("package-lock.json")
execute_file({
  file_path: "package-lock.json",
  code: "const d=JSON.parse(process.env.FILE_CONTENT); console.log('Packages:', Object.keys(d.dependencies||{}).length)"
})

// Compress any large text
compress({ content: largeOutput, intent: "find error messages" })

// Fetch and index docs
fetch_and_index({ url: "https://docs.example.com" })
search({ query: "what I need to know" })

// Check how much context was saved this session report()