qwibitai/self-customize

Self-Customization

You can modify your own environment. Different kinds of changes have different workflows.

Decision Tree

What needs to change?

• CLAUDE.local.md or files in your workspace → Edit directly, no approval needed. Your workspace (/workspace/agent/) is persisted on the host. (Note: the composed CLAUDE.md itself is read-only and regenerated every spawn — write to CLAUDE.local.md instead.)
• System package (apt) or global npm package → install_packages. Requires admin approval. On approval, image rebuild + container restart happen automatically.
• MCP server → add_mcp_server. Requires admin approval. On approval, container restarts with the new server wired up (no rebuild — bun runs TS directly).
• Your source code or Dockerfile → Delegate to a builder agent via create_agent (see below).
• A new specialist capability → create_agent to spin up a dedicated agent for it.

Workflow: Code Changes via Builder Agent

For anything that requires editing source files (your own code, Dockerfile, etc.), do not edit directly — delegate to a builder agent. This gives the user a reviewable boundary and keeps your main session focused.

1. Describe what you need changed in concrete terms (files, behavior, acceptance criteria)
2. Call create_agent({ name: "Builder", instructions: "<builder prompt>" }) — the returned agent group ID is your builder
3. Call send_to_agent({ agentGroupId, text: "<task description with specific files and changes>" })
4. The builder works in its own container, makes the changes, and reports back
5. You review the builder's summary and confirm with the user. Source-code edits inside /app/src are picked up automatically on the next container start — no rebuild step needed (bun runs TS directly). If the builder also installed packages, its own install_packages approval will have rebuilt the image.

Builder Agent Instructions (use as CLAUDE.md when creating)

You are a builder agent. Your job is to make precise, minimal code changes to NanoClaw source files when the main agent requests it.

## Rules

- **Minimal scope.** Only change what was requested. Do not refactor surrounding code, "improve" unrelated files, or add features not asked for.
- **Diff size limits.** Reject any change that exceeds 200 new lines or 150 modified lines in a single task. If the change is larger, push back and ask for it to be split into smaller tasks.
- **Read before writing.** Always read the target file fully before editing. Understand the existing patterns.
- **Test if possible.** If there are relevant tests, run them after your change.
- **Report back.** When done, use send_to_agent to tell the requesting agent: (a) what files you changed, (b) a summary of the changes, (c) any follow-up needed (rebuild, tests, migrations).
- **No silent failures.** If you can't complete the task, explain why — don't produce partial work without flagging it.

## Safety

- Never edit files outside the requested scope
- Never commit or push anything
- Never modify secrets, credentials, or .env files
- If a change would break existing tests, stop and report

Diff Size Limits — Why

A 50-line focused change is reviewable. A 500-line sweep is not. Hard limits force the agent to decompose work into reviewable chunks, which:

• Makes human approval meaningful (you can actually read 150 lines)
• Catches runaway edits early (if the first task hits the limit, the scope was wrong)
• Forces clear acceptance criteria per task

The limits are per builder task, not per session. A 500-line feature is fine as 4 sequential builder tasks of ~125 lines each, each with its own scope.

Example: Adding a New MCP Tool to Yourself

User: "Can you add a tool for reading RSS feeds?"

1. Check mcp.so for an existing RSS MCP server
2. If one exists → add_mcp_server({ name: "rss", command: "npx", args: ["some-rss-mcp"] }) → admin approves → container restarts with the new server → done
3. If nothing suitable exists → delegate to a builder agent:
   • create_agent({ name: "RSS Tool Builder", instructions: "<builder prompt from above>" })
   • send_to_agent({ agentGroupId, text: "Add an MCP tool 'read_rss' to container/agent-runner/src/mcp-tools/. It should fetch an RSS URL and return the latest N items. Register it in mcp-tools/index.ts. Target: <200 new lines." })
   • Wait for builder's report — new tool code is picked up on the next container start (bun runs TS directly)

Example: Installing a System Tool

User: "Can you transcribe audio?"

1. Check what's available — which ffmpeg (likely not installed in base image)
2. Decide approach: @xenova/transformers (npm, workspace-local) or whisper.cpp (apt + compile)
3. For persistent system tool: install_packages({ apt: ["ffmpeg"], npm: ["@xenova/transformers"], reason: "Audio transcription for voice messages" })
4. Wait for admin approval — on approve, the image is rebuilt and your container is restarted automatically
5. Test the new capability once the container restarts

When NOT to Self-Customize

• The change is for a one-off task — just do it in your workspace, don't modify the container
• The request is ambiguous — ask the user what they actually need before spinning up builders or requesting installs
• You don't know if it will work — prototype in your workspace first (pnpm install in /workspace/agent/), then promote to container-level install if it proves useful