Skip to content

CLAUDE.md

Latest commit

 

History

History
37 lines (30 loc) · 2.37 KB

CLAUDE.md

File metadata and controls

37 lines (30 loc) · 2.37 KB

Claude Code — project context

@./AGENTS.md

Claude-specific notes

  • Keep instructions modular and prefer progressive disclosure.
  • If deeper scoped behavior is needed, use .claude/rules/ files.
  • Prefer single-responsibility functions. When touching a method that mixes multiple concerns, treat that as a prompt to split it along readability and responsibility boundaries before adding more logic.

Debugging Protocol

  • Always check logs first before using browser automation or code analysis. Look at server logs (~/.steroids/logs/api.log, ~/.steroids/logs/webui.log), runner logs (~/.steroids/runners/logs/), and browser console errors before diving into source code.
  • Never manually trigger recovery. Do not run steroids monitor run, steroids runners wakeup, reset tasks, or invoke any other recovery command to work around a bug. Fix the code so the system self-heals, then let the user restart and verify.
  • Monitor/recovery fixes require two test layers. Any monitor, runner-health, recovery, or orphan-detection fix must add both targeted unit tests and a mock-style integration test that exercises the real DB shape and trigger path.

How to Run Parallel Adversarial Reviews

When a design requires cross-provider review (see AGENTS.md for when), dispatch two reviewers simultaneously by sending a single message with two Task tool calls. Use whichever providers are available: Claude, Codex, Gemini, Mistral, or top reasoning models from Hugging Face.

Call 1 — Claude reviewer:

Task tool:
  subagent_type: superpowers:code-reviewer
  prompt: "Adversarial review of the following design. Look for technical debt,
           architectural regressions, type safety gaps, and logic holes. Be harsh.
           Do not suggest over-engineering. [paste full design doc content here]"

Call 2 — alternate provider (pick whichever is available):

Task tool:
  subagent_type: Bash
  prompt: "Run one of:
    timeout 1800 codex exec \"[adversarial review prompt]\"
    timeout 1800 gemini -p \"[adversarial review prompt]\"
    timeout 1800 mistral \"[adversarial review prompt]\""

Both calls go in a single message so they execute in parallel. When both return, assess each finding independently per AGENTS.md §"How to Conduct a Review" and append a Cross-Provider Review section to the design doc with adopt/defer/reject decisions for each finding.