Skip to main content
The bundle is your agent project tree — the directory containing dari.yml at its root, plus everything under it. When you publish, the dari CLI archives that tree and uploads it. At runtime, the sandbox receives it as /bundle.

​
What gets uploaded

The CLI archives the entire repo root. The manifest validator enforces rules on known paths:
  • dari.yml at the root
  • tools/<name>/tool.yml for each custom tool
  • skills/<name>/SKILL.md for each declared skill
  • prompts/system.md (or whatever instructions.system points at)
  • Dockerfile at the root, if runtime.dockerfile is set
Everything else in the tree is passed through as-is. There is no .dariignore and no implicit filtering.
Whatever sits at your repo root ends up in the bundle. That includes .git/, node_modules/, local .env files, build artifacts, and anything else you haven’t cleaned up. Be deliberate about what you publish from — secrets in the tree get baked into the sandbox template.

​
Where it lives in the sandbox

During the prepared-runtime build, Dari copies the archive contents to /bundle inside the sandbox image and sets /bundle as the working directory. This holds for both runtime modes:
  • Default base image (no runtime block) — Dari uses e2bdev/base:latest and layers the bundle copy on top.
  • Custom Dockerfile — your image is built first, then Dari copies the bundle to /bundle and sets WORKDIR /bundle on top of it.

​
How the agent sees it

At session time, the harness runs tool calls with cwd = /bundle:
  • Built-ins (bash, read, edit, grep, find, ls, write) resolve relative paths against /bundle, so the agent sees your project layout directly.
  • Custom tool handlers resolve against /bundle/tools/<name>/, so helper modules and data files you ship alongside tool.yml are visible.
  • The system prompt is read from whatever instructions.system points at inside the bundle.
Dari does not install dependencies for files outside the known paths — if you ship a package.json or requirements.txt in your project root and need them installed, do it in a custom Dockerfile.