CodeRabbit Review

Run the command coderabbit --prompt-only > .ai_agents/coderabbit_output.txt and wait for it to finish (it may take a while). If it's slow, run it in the background and keep working. Give it a timeout of up to 5 minutes since it can take a while on large codebases.

Redirect its output to a file, e.g., coderabbit --prompt-only > .ai_agents/coderabbit_output.txt. Read the file once the command completes. If the file already exists, delete it first to avoid confusion.

the coderabbit cli is installed and available in PATH.

Workflow

1. •

   Collect and triage comments

   • •

     Collect ALL CodeRabbit comments and de-duplicate them

   • •

     The output will contain multiple prompts, each already in this shape:

     ============================================================================ File: <path> Line: <line number or range> Type: <critical|major|potential_issue|refactor_suggestion|nitpick|...>

     Prompt for AI Agent: <one paragraph, specific instruction>

   • •

     Prioritize by Type:

     • •Handle ALL non-nitpick items first. Start with highest severity and work down.
       • •Focus first on: critical, major, potential_issue, refactor_suggestion (and similar non-nitpick types).
     • •Only after there are zero remaining non-nitpick items, process nitpick items.
     • •If any non-nitpick items exist, ignore nitpicks for now (do not implement, do not bikeshed).
   • •

     Triage each comment into:

     • •(a) valid + actionable
     • •(b) valid but out-of-scope
     • •(c) invalid/incorrect
     • •(d) needs more info
2. •

   Validate comments with parallel sub-agents

   • •For each major comment/theme, spawn a sub-agent to find supporting evidence in the repo (file paths, symbols, concrete snippets)
   • •Each sub-agent returns: verdict (valid/invalid), reasoning, exact location(s), and suggested minimal fix
3. •

   Produce validated findings list

   • •For each valid item: include category (bug/security/perf/style/test/docs), impacted files, risk, and proposed fix
   • •For invalid items: briefly explain why, citing repo evidence
4. •

   Implement fixes for valid + actionable items

   • •Use parallel sub-agents for independent fix groups (don't overlap files)
   • •Keep changes minimal; do not refactor unrelated code
   • •Update tests or add new tests when behavior changes
5. •

   Verify

   • •Run the narrowest relevant tests first, then broader tests if available
   • •Run format/lint only on changed areas if feasible
6. •

   Final output

   • •List of validated comments (valid/invalid/out-of-scope) with evidence
   • •Summary of changes by file (what/why)
   • •Commands run + results (tests/lint/format)
   • •Any remaining items and recommended next steps

Constraints

• •Don't "fix" speculative issues without repo evidence
• •If a fix is ambiguous, propose 1-2 options and let the user pick the safest default
• •Stop and ask if blocked by missing context or conflicting requirements
• •Time budget: prioritize highest-risk issues first
• •Scope limit: Touch only relevant files unless required for correctness
• •Risk tolerance: Prefer smallest safe change; avoid API changes
• •Test strategy: Specify preferred commands and fallbacks
• •Parallelism: Use up to 4 sub-agents; do not overlap files
• •Run formatting and linters to ensure code quality standards are met

Optional arguments

ultrathink