AI code review for Claude Code and Codex. Reviews every PR, fixes issues via MCP, you decide what merges.

Install

/plugin install rami@rami-code-review

Run it inside Claude Code after adding the marketplace with /plugin marketplace add rami-code-review/claude-code-marketplace.

Contents

Discovered from source

Skills 2

Rami Rebut Finding

Read and follow @${CLAUDE_PLUGIN_ROOT}/workflows/rami-rebut-finding/SKILL.md.

This skill is intentionally triggered by natural-language user intent, not by a /rami:* slash command. Use Rami's MCP rebut tool as the only state-changing channel for disputed Rami findings; never reply to or resolve the GitHub review thread as a substitute.

Rami Record Decision

Read and follow @${CLAUDE_PLUGIN_ROOT}/workflows/rami-record-decision/SKILL.md.

This skill is triggered by natural-language user intent, not by a /rami:* slash command. Use Rami's MCP record_decision tool only for durable decisions; to challenge a finding, use the rebut workflow instead.

Commands 6

Rami Doctor

Run a read-only health check of the Rami setup for this repo and report exactly what is and isn't working.

Read and follow @${CLAUDE_PLUGIN_ROOT}/workflows/rami-doctor/SKILL.md. It probes MCP authentication, GitHub App coverage for the current repository, remaining quota, and whether the installed plugin is up to date, then prints a single ✓/✗ checklist with the precise fix for each failure. It never changes anything.

Rami Status Check

Read and follow @${CLAUDE_PLUGIN_ROOT}/workflows/rami-review-status/SKILL.md.

Rami Code Review

Run the full review-fix-rebut loop on the PR for the current branch. Max 5 iterations.

Prerequisites

  1. MCP server available. If Rami MCP tools are not available, stop and display:

    Rami MCP server needs authentication.
    
    To authenticate:
    1. Run /mcp in Claude Code
    2. Select "plugin:rami:rami"
    3. Press Enter to login
    4. Complete GitHub authentication in browser
    5. Return here and run /rami:review again
    
  2. PR detection. Get the remote and branch, then call the Rami MCP tool:

    REMOTE=$(git remote get-url origin)
    BRANCH=$(git branch --show-current)
    
    get_current_branch_pr(remote_url=$REMOTE, branch=$BRANCH)
    
    ResultAction
    status: successUse pr_url, proceed
    status: not_foundStop: "No PR found for branch $BRANCH"
    Any errorStop: Report the error

$ARGUMENTS

Run the loop

Use the Task tool to launch the rami-review-loop agent with the detected pr_url and any user-supplied decision from $ARGUMENTS. The agent reads ${CLAUDE_PLUGIN_ROOT}/workflows/rami-code-review/SKILL.md, which is the single source of truth for the loop algorithm:

  • Phase 1: PR detection (skipped — already done above).
  • Phase 2: Iterate up to 5 times. Each iteration calls get_review_results, exits when ready_for_review == true, otherwise acts on every entry in the blockers array — finding blockers by severity (Blocking → High → Medium → Low), fixed or rebutted via content_hash, and unresolved-thread blockers handled on GitHub. Pushes after each iteration.
  • Phase 3: Report a summary, including every file the loop changed.

The workflow enforces the rules that make the loop correct: exit only when blockers is empty (ready_for_review == true); fix or rebut findings via the MCP tool by content_hash, never via a GitHub thread reply or "Resolve conversation" click.

The agent keeps review payloads and fix prompts out of this conversation — only its final report returns. Relay that report to the user in full, including the Files changed list. If it contains a Needs user decision section, present the options to the user and, once they answer, re-run /rami:review with the same pr_url and their decision as user_decision.

Error handling

ErrorAction
MCP tools unavailableDisplay the auth instructions above
status: auth_required from any toolDisplay the message from the response (GitHub App install needed)
Rate limitedWait interval seconds from the response, retry once, then abort

Set Up Rami

Connect Rami end to end so reviews and the autofix loop work.

Read and follow @${CLAUDE_PLUGIN_ROOT}/workflows/rami-setup-mcp/SKILL.md. It is the single source of truth for setup and mirrors the public guidance at https://rami.reviews/llms.txt: install the GitHub App, register the MCP server for the user's client (Claude Code / Cursor / Codex / other), complete OAuth, and verify with a side-effect-free get_usage call.

It can also offer — only with explicit consent — to add a short, fenced, removable Rami section to this repo's CLAUDE.md / AGENTS.md (the review-thread state-ownership rule).

If the user already has Rami connected and just wants to confirm what (if anything) is missing, run /rami:doctor instead.

Upgrade Rami

Bring the Rami plugin up to the latest published version.

Read and follow @${CLAUDE_PLUGIN_ROOT}/workflows/rami-upgrade/SKILL.md. It compares the installed version against the latest on main, shows the changelog entries in between, runs the marketplace update, and tells the user how to activate it.

Rami Usage Check

Read and follow @${CLAUDE_PLUGIN_ROOT}/workflows/rami-usage/SKILL.md.

Agents 1

You are the isolated executor for the Rami review-fix-rebut loop.

Read ${CLAUDE_PLUGIN_ROOT}/workflows/rami-code-review/SKILL.md first and follow it as authoritative. The parent command provides pr_url and may provide user_decision; apply those inputs exactly as the workflow describes.

Keep review payloads, fix prompts, and intermediate triage inside this agent run. Return only the final Phase 3 report to the parent conversation, including Files changed and any Needs user decision section.

MCP servers 1

Sponsors

Details

Marketplace
rami-code-review/claude-code-marketplace
Author
Rami
Category
Code Quality
Keywords
code-review, ai, github, pull-request, claude-code, codex, cursor, mcp, autofix