Skip to content

README.md

Latest commit

 

History

History
131 lines (95 loc) · 3.73 KB

README.md

File metadata and controls

131 lines (95 loc) · 3.73 KB

LLM Context Generator (llm-context)

Generates ~/CLAUDE.md context file for LLMs (Claude Code, Gemini CLI, etc.) with Workbench workspace information. Claude Code auto-discovers this file on startup.

Example Usage

"features": {
    "ghcr.io/verily-src/workbench-app-devcontainers/llm-context:1": {
        "username": "jupyter",
        "userHomeDir": "/home/jupyter"
    }
}

Or for local development:

"features": {
    "./.devcontainer/features/llm-context": {
        "username": "jupyter",
        "userHomeDir": "/home/jupyter"
    }
}

Options

Options Id Description Type Default Value
username Username of the container user string root
userHomeDir Home directory of the container user string /root

What It Does

When installed, this feature:

  1. Generates ~/CLAUDE.md - Claude Code auto-discovers this file on startup
  2. Provides workspace context - Name, ID, role, resources, cloud paths
  3. Includes skill files - Detailed guides (e.g., custom app creation) in ~/.claude/skills/
  4. Sets up aliases - generate-llm-context, refresh-context

What's in ~/CLAUDE.md

  • Quick Rules - When to use this file vs. MCP/CLI
  • Current Workspace - Name, ID, description, role, cloud platform
  • Resource Paths - JSON lookup for all resources (GCS, BigQuery, etc.)
  • Data Persistence - Warning + save commands
  • Data Exploration - Common BigQuery/GCS commands
  • MCP Tools - Available tools and CLI equivalents
  • Skills - Links to detailed guides

When Context Gets Generated

  1. Automatically on app start - Via postStartCommand (after bucket mounting completes)
  2. Manually - Run generate-llm-context or refresh-context

Important: Add the context generation to your postStartCommand in .devcontainer.json:

"postStartCommand": [
  "bash",
  "-c",
  "./startupscript/remount-on-restart.sh jupyter /home/jupyter \"${templateOption:cloud}\" \"${templateOption:login}\" && /opt/llm-context/generate-context.sh /home/jupyter"
]

Note: Pass the user home directory (e.g., /home/jupyter) as an argument because postStartCommand runs as root, not as the container user.

This ensures context is generated AFTER authentication and workspace setup complete.

MCP Integration

This feature works well alongside the wb-mcp-server feature:

  • llm-context provides static context (workspace info, resource paths)
  • wb-mcp-server provides dynamic tools (search, create, modify)

For optimal LLM experience, use both:

"features": {
    "./.devcontainer/features/llm-context": {},
    "./.devcontainer/features/wb-mcp-server": {}
}

Troubleshooting

Context not generating?

# Check if workspace is set
wb workspace describe

# If not authenticated:
wb auth login --mode=APP_DEFAULT_CREDENTIALS
wb workspace set <workspace-id>

# Then generate manually:
generate-llm-context

Claude Code not seeing context?

# Check file exists
ls -la ~/CLAUDE.md

# Check it's not empty
head ~/CLAUDE.md

File Locations

File Purpose
/opt/llm-context/generate-context.sh Main generation script
/opt/llm-context/run-context-generator.sh Auto-run wrapper
~/.claude/CLAUDE.md Generated context (primary)
~/CLAUDE.md Symlink for auto-discovery
~/.claude/skills/ Skill files (e.g., CUSTOM_APP.md)

Notes

  • This feature requires the Workbench CLI (wb) to be installed
  • jq is automatically installed if not present
  • Context is only generated if a workspace is set (wb workspace describe succeeds)

Note: This feature is automatically configured to work with the wb-mcp-server feature if both are installed.