Skip to content
unfault

Use with AI Agents

AI coding assistants can write code fast. The failure mode is rarely syntax. It is misunderstanding: wrong entry point, wrong pattern, hidden blast radius.

Unfault helps with that. It brings context to developers, and it can also surface runtime risks when they matter.

1. Add instructions your agent will read

Section titled “1. Add instructions your agent will read”

Put this snippet in AGENTS.md at the repo root.

If your agent does not load AGENTS.md, mirror the same content into the instruction file it does read:

  • GitHub Copilot: .github/copilot-instructions.md
  • Cursor: .cursorrules or .cursor/rules/*.mdc
  • Claude Code: CLAUDE.md
# Unfault (project context)


Use Unfault to get codebase context before and after edits.


Find symbols (semantic search):
  unfault graph find "<symbol>"
  unfault graph find "<Pattern>*" -t class
  unfault graph find "*" -t route


Estimate blast radius before refactoring:
  unfault graph impact <file>
  unfault graph function-impact <file>:<function>


Review the patch you just wrote:
  unfault review --uncommitted --llm


Ask for deeper codebase context:
  unfault ask "<question>"

2. Use the context loop while coding

Section titled “2. Use the context loop while coding”
Terminal window
unfault graph find "validate_token"
unfault graph impact src/utils/auth.py
unfault review --uncommitted --llm

3. Read review output like a colleague

Section titled “3. Read review output like a colleague”

When unfault review --llm returns:

  • Start with next_steps.
  • Treat findings[].message as a discussion starter, not an order.
  • Stay within files_allowed_to_change.
  • Use decision_level to avoid architecture changes inside a small patch.

If a finding is in code you did not touch, acknowledge it and move on.

4. If you want review to feel less like a risk list

Section titled “4. If you want review to feel less like a risk list”

Alongside findings, the most useful context to attach to a review is:

  • change_intent: what the patch tries to do
  • blast_radius: key callers, routes, and entry points affected
  • touched_symbols: functions/classes/routes changed or added
  • existing_patterns: similar code locations to follow
  • relevant_tests: existing coverage and tests worth adding

This keeps the conversation about meaning and fit, not only production readiness.

Section titled “Related guides”