netclaw-dev/netclaw/netclaw-projects

Project Workspaces

A project is a local git repo with an AGENTS.md at its root. Projects live in your workspaces directory (check TOOLING.md for the path).

Projects can be nested at any depth within the workspaces directory. A cloned engineering repo counts. So does a directory you create for a marketing campaign, research initiative, or any sustained body of work.

Finding Existing Projects

Before creating a new project, always check what already exists:

file_search(Root: "<workspaces_path>", Query: "AGENTS.md", Mode: "name")

Read the AGENTS.md of any match to understand its purpose. If a project already covers the user's request, cd into it and work there.

Creating a New Project

Only create a project when:

• The work is sustained (multi-session, ongoing)
• No existing project covers it
• The user explicitly asks to start one

Steps:

1. Choose a descriptive directory name (kebab-case)
2. Create <workspaces_path>/<name> with one shell operation.
3. Set that path with set_working_directory.
4. Initialize version control with one shell operation.
5. Write an AGENTS.md with file_write describing:

• What this project is about
• Goals and constraints
• Relevant context (target audience, products, tools, etc.)

1. Stage and commit with separate shell operations.

The AGENTS.md is the project's living context document. Update it as the project evolves.

Working in a Project

Before the first tool call for a different task-named project, use set_working_directory to set the project directory:

set_working_directory(path: "/home/user/workspaces/my-project")

Declare the named path before probing it. If rejected, declare the user-provided fallback before other tools. Use the task's first project path exactly; do not substitute its parent first.

This automatically loads the project's identity file (.netclaw/AGENTS.md, CLAUDE.md, AGENTS.md, or CONTEXT.md — first match wins) into the system prompt. The project context persists across crash/restart, so you don't need to re-read it manually.

Use the project directory as your working directory for all project-related file operations. Commit meaningful changes so the project has history.

Team and Personal sessions refresh Git working context at the beginning of a turn. That snapshot is grounding, not a live view: changes made during the turn appear in the next turn. A spawned subagent receives its own read-only snapshot and reports only changes confirmed through its tools. Do not infer that every dirty file in the shared worktree was changed by that subagent.

When NOT to Create a Project

One-off tasks don't need a project. If the user asks a quick question, runs a single search, or needs something done once — just do it in the current session. Don't create overhead for throwaway work.