While code is being written, the coding agent asks the graph instead of searching through files. One question to the graph can replace many text searches and file reads. For example: "who calls this method, three levels up?"

Who and how

What agents are taught to do

The skill Niro installs tells the agent to do four things:

  1. Explore first. Use code_grep, get_source_code, resolve_name, list_types, get_class_structure, find_api_endpoints and find_handler before reading files by hand.
  2. Check both directions before changing anything. Use get_blast_radius, get_reverse_call_chain and get_impact_analysis.
  3. Follow flows across services. Use the HTTP, event and data tools.
  4. Say when Niro does not apply. For example: the folder is not indexed, or the graph is behind the files.

Claude Code gets more help than the other agents. Niro adds hooks to Claude Code: small commands that Claude Code runs at fixed moments. One of them shows an "impact check" before Claude Code edits a file. It is advisory only: it shows the information and never blocks the edit. To turn it off, run touch ~/.niro/impact-gate-off. The other agents get instruction text instead of hooks.

When the graph is behind your edits

The graph is rebuilt in the background about 3 seconds after a folder stops changing. If an answer is about a file that changed after the last build, the answer says so, and lists the files to read directly.

Playbooks for this stage