Dispatch prompts to external AI CLIs from Claude Code. Run GPT-5.4 via Codex, Gemini 3.1 Pro via Gemini CLI, or Claude Opus 4.6 via Claude Code CLI — as isolated subagents that keep your main context clean.

brew install 1F47E/tap/rival
rival install

cd rival && make install
rival install

Note: go install is not supported due to the repo's subdirectory layout. Use Homebrew or build from source.

rival install copies the Claude Code skills (embedded in the binary) into ~/.claude/skills/. After that, /rival-codex, /rival-gemini, /rival-claude, and /rival-megareview are available in Claude Code.

Use rival install --force to overwrite without prompting.

• Codex CLI: npm install -g @openai/codex + codex login
• Gemini CLI: npm install -g @google/gemini-cli + set GEMINI_API_KEY
• Claude Code CLI: install + authenticate

You only need the CLIs for the commands you use. Megareview uses all available CLIs.

/rival-codex explain the auth flow in this project
/rival-codex -re xhigh find bugs in src/main.go
/rival-codex review                        — review (auto-detects changed files via git)
/rival-codex review src/api/               — review specific scope (bypasses git detection)
/rival-codex -re xhigh review src/api/     — review with xhigh reasoning

/rival-gemini explain the auth flow
/rival-gemini -re high analyze this complex algorithm
/rival-gemini review                       — review (auto-detects changed files via git)
/rival-gemini review src/api/              — review specific scope (bypasses git detection)

/rival-claude explain the auth flow
/rival-claude -re xhigh find code quality issues in src/
/rival-claude review                       — review (auto-detects changed files via git)
/rival-claude review src/api/              — review specific scope (bypasses git detection)

/rival-megareview                          — review with ALL CLIs (auto-detects changed files)
/rival-megareview src/api/                 — review specific scope (bypasses git detection)
/rival-megareview -re xhigh src/api/       — both CLIs, max reasoning effort

Reasoning effort (-re): low, medium, high (default), xhigh

When you run a review, Codex/Gemini get full access to your project. They don't just see a diff — they run as CLI tools inside your workdir with tool use enabled, so they can:

• Read any file in the project
• Follow imports and trace dependencies
• Explore the full codebase to understand context
• Run commands to inspect project structure

Smart scope detection. Running /rival-codex review with no arguments auto-detects what to review via git:

1. Dirty files (staged + unstaged + untracked new files) → reviews those files
2. Last commit (if working tree is clean) → reviews files from HEAD
3. Full project → only if not a git repo or no changes found

The scope is a focus hint, not a restriction. review src/api/ tells the reviewer to focus on src/api/, but it can (and will) read other files to understand the code in context. Explicit scope bypasses git detection entirely.

This means you can use natural language for the scope:

/rival-codex review the files changed in the last commit
/rival-codex review the authentication middleware
/rival-megareview -re xhigh the new payment flow in src/billing/

The reviewer will figure out what to look at, explore the relevant code, and give you a review with full project understanding.

Megareview assigns specialized roles to each reviewer:

• Codex → Bug Hunter — finds concrete code-level defects: logic bugs, broken state transitions, race conditions, missing edge cases. Optimizes for true positives with high confidence.
• Gemini → Architecture & Security — attacks from angles a bug hunter misses: architectural regressions, broken cross-file flows, incomplete refactors, concurrency issues, security problems, silent failure gaps.
• Claude → Code Quality & DX — focuses on readability, naming, unnecessary complexity, error message quality, API ergonomics, maintainability traps, developer footguns, and dead code.

All reviewers emit structured JSON with file, line, severity, category, confidence (1-10), and fix suggestions.

Role prompts can be customized via ~/.rival/config.yaml:

roles:
  bug_hunter: |
    Your custom bug hunter instructions...
  code_quality: |
    Your custom code quality instructions...

A third consilium judge (runs via Codex) then:

• Merges duplicate findings (same file + line + problem → single finding with all reporters in found_by)
• Applies consensus bonus (+2 confidence for findings reported by 2+ reviewers)
• Filters by confidence threshold (default: ≥6)
• Sorts by severity (critical first), then confidence
• Produces a unified verdict: approve, request_changes, or comment

═══ RIVAL REVIEW ═══

Summary: ...

[CRITICAL] file.go:42 — Title
  Description...
  Fix: ...
  Found by: codex, gemini

[HIGH] file.go:100 — Title
  ...

Recommendation: request_changes — ...

Reviewed by: codex (bug_hunter), gemini (arch_security), claude (code_quality)
Judge: codex (consilium)
Findings: 5 (threshold: 6)

If only one CLI is available, the consilium judge falls back to whichever CLI is present.

# Run with prompt from stdin
echo 'explain the auth flow' | rival command codex --workdir .
echo 'explain the auth flow' | rival command gemini --workdir .
echo 'explain the auth flow' | rival command claude --workdir .

# Review via megareview (both CLIs in parallel)
echo 'src/api/' | rival command megareview --workdir .

Monitor running and past sessions in a full-screen terminal UI:

rival tui

List view shows all sessions with status, CLI (◈ codex / ✦ gemini / ● claude / ◈✦● mega), model, effort, elapsed time, workdir, and prompt preview. Megareview sessions are grouped into a single row.

Detail view shows full metadata, prompt, and live-streaming log output. For megareview groups, both Codex and Gemini logs are shown side by side.

rival sessions              # all sessions as JSON
rival version               # show version

Claude Code main session
    │
    │ /rival-codex review src/
    ▼
Claude skill (context: fork)
    │
    │ stdin heredoc → rival command codex --workdir $(pwd)
    ▼
rival binary
    ├─ parses arguments (-re flag, review/prompt mode)
    ├─ builds review prompt with scope injection
    ├─ spawns codex/gemini via subprocess
    ├─ pipes prompt to stdin, tees stdout to log file
    ├─ writes session JSON + live log to ~/.rival/sessions/
    └─ returns output to skill → back to Claude Code

Megareview (roles + consilium):
    rival binary
    ├─ generates shared GroupID (UUID)
    ├─ assigns roles: codex=bug_hunter, gemini=arch_security, claude=code_quality
    ├─ spawns all available CLIs concurrently with role-specific prompts
    ├─ parses structured JSON output from each reviewer
    ├─ spawns codex again as consilium judge
    │   ├─ merges duplicates, applies consensus bonus
    │   ├─ filters by confidence threshold (≥6)
    │   └─ produces unified verdict with found_by attribution
    ├─ prints formatted review to stdout
    └─ TUI groups all sessions by GroupID

Second terminal:
    rival tui
      ├─ watches ~/.rival/sessions/ via fsnotify (.json + .log)
      ├─ groups sessions by GroupID for megareview display
      ├─ live-refreshes every second while sessions are running
      └─ x key sends SIGTERM to kill stuck sessions

• Full project access: reviewers run as AI CLI tools with tool use — they explore your codebase, not just diffs
• Isolated execution: skills use context: fork — runs in subagent, zero impact on your Claude context
• Stdin piping: prompts passed via heredoc, never shell-quoted into argv (prevents injection)
• Env filtering: child processes get a sanitized environment (blocks proxy/preload vars from .env)
• Fault tolerant: megareview continues if one CLI fails, reports the error inline

Run Claude via Docker with a separate Anthropic subscription. This is an internal setting — skills and commands stay the same, the executor switches transparently.

1. Generate a token on the second account:

   claude setup-token   # → sk-ant-oat01-...

2. Set the env var:

   export RIVAL_CLAUDE_TOKEN=sk-ant-oat01-YOUR-TOKEN-HERE

3. Enable Docker mode in ~/.rival/config.yaml:

   claude:
     mode: docker

4. First run auto-builds the rival-claude Docker image. Subsequent runs reuse it.

Now /rival-claude, megareview, and all claude commands run inside Docker using the second subscription. Set mode: native (or remove the key) to switch back.

• OAuth tokens can expire — re-run claude setup-token if you get 401 errors
• The Docker image is node:22-slim + @anthropic-ai/claude-code (auto-built, ~200MB)
• Your workdir is mounted as /workspace inside the container
• To rebuild the image: docker rmi rival-claude, next run rebuilds automatically

rm -rf ~/.claude/skills/rival-codex ~/.claude/skills/rival-gemini ~/.claude/skills/rival-claude ~/.claude/skills/rival-megareview
brew uninstall rival        # if installed via brew
# or: rm "$(go env GOPATH)/bin/rival"   # if installed from source

MIT