AI Buddies v3.1 — Help & Reference

Available Skills

Skill	Engine	Description
`/codex "prompt"`	Codex	Ask Codex anything — brainstorm, delegate, second opinion
`/codex-review`	Codex	Code review via Codex (uncommitted, branch, commit)
`/gemini "prompt"`	Gemini	Ask Gemini anything — brainstorm, delegate, second opinion
`/gemini-review`	Gemini	Code review via Gemini (uncommitted, branch, commit)
`/opencode "prompt"`	OpenCode	Ask OpenCode anything — multi-provider, configurable model
`/opencode-review`	OpenCode	Code review via OpenCode (uncommitted, branch, commit)
`/brainstorm "topic"`	All	Multi-AI roundtable — all available buddies assess the task
`/forge "task" --fitness "cmd"`	All	Competitive build — buddies build, automated scoring picks winner
`/tribunal "question"`	2+	Adversarial debate — FOR vs AGAINST with evidence
`/tribunal --socratic`	2+	Socratic — probe assumptions, Claude synthesizes
`/tribunal --steelman`	2+	Steelman — argue other side's strongest case
`/tribunal --red-team`	2+	Red-team — attack from multiple angles
`/tribunal --synthesis`	2+	Synthesis — propose solutions, then hybridize
`/tribunal --postmortem`	2+	Postmortem — investigate failure from different angles
`/leaderboard`	—	Show ELO ratings leaderboard
`/add-buddy`	—	Register a new AI CLI as a buddy
`/buddy-help`	—	This help reference

Dynamic Buddy Registry (v3)

Any CLI-based AI tool can become a buddy. Builtin buddies (Claude, Codex, Gemini, OpenCode) are auto-detected. Register custom buddies with /add-buddy:

/add-buddy --id aider --binary aider --display "Aider" --modes exec

Buddy definitions are stored as JSON:

Builtin: <plugin>/buddies/builtin/*.json
User: ~/.claudes-ai-buddies/buddies/*.json

Configuration

Config file: ~/.claudes-ai-buddies/config.json

General

Key	Default	Description
`timeout`	`120`	Max seconds per call
`sandbox`	`full-auto`	Sandbox mode (`full-auto` or `suggest`)
`debug`	`false`	Enable debug logging

Engine overrides

Key	Default	Description
`codex_model`	(CLI default)	Override Codex model
`gemini_model`	(CLI default)	Override Gemini model
`claude_model`	(CLI default)	Override Claude model
`opencode_model`	`opencode/minimax-m2.5-free`	Override OpenCode model (format: `provider/model`)
`codex_path`	(auto-detected)	Explicit path to codex binary
`gemini_path`	(auto-detected)	Explicit path to gemini binary
`claude_path`	(auto-detected)	Explicit path to claude binary
`opencode_path`	(auto-detected)	Explicit path to opencode binary

Forge

Key	Default	Description
`forge_timeout`	`600`	Engine timeout in seconds
`forge_auto_accept_score`	`88`	Stage 1 auto-accept threshold
`forge_clear_winner_spread`	`8`	Points spread to skip synthesis
`forge_enable_synthesis`	`true`	Enable critique-based synthesis
`forge_max_critiques`	`3`	Max critique hunks per loser
`forge_starter_strategy`	`fixed`	`fixed` or `rotate`
`forge_fixed_starter`	`claude`	Default starter
`forge_require_baseline_check`	`true`	Run fitness on base before forging

Tribunal

Key	Default	Description
`tribunal_rounds`	`2`	Cross-examination rounds
`tribunal_max_buddies`	`2`	Max debaters

ELO

Key	Default	Description
`elo_enabled`	`true`	Enable ELO tracking
`elo_k_factor`	`32`	ELO K-factor (higher = more volatile)

Example config

{
  "codex_model": "gpt-5.4-codex",
  "gemini_model": "gemini-2.5-pro",
  "timeout": "180",
  "sandbox": "full-auto",
  "debug": "false",
  "elo_enabled": "true"
}

Review Targets

/codex-review                        # uncommitted changes (Codex)
/codex-review branch:main            # diff from main to HEAD
/gemini-review                       # uncommitted changes (Gemini)
/gemini-review commit:abc1234        # specific commit
/opencode-review                     # uncommitted changes (OpenCode)
/opencode-review branch:feature      # diff from branch to HEAD

Requirements

Codex CLI v0.100.0+ (npm install -g @openai/codex)
Gemini CLI v0.30.0+ (npm install -g @google/gemini-cli)
OpenCode CLI v1.2.0+ (brew install opencode)
Auth — codex auth login / gemini auth login / opencode providers login (or API key env vars)
jq (required for registry, ELO, tribunal)
git (required for review skills, forge, tribunal)

Debug Logging

Enable debug mode:

mkdir -p ~/.claudes-ai-buddies
echo '{"debug": "true"}' > ~/.claudes-ai-buddies/config.json

Logs are written to ~/.claudes-ai-buddies/debug.log (auto-rotated at 1MB).

How It Works

User → Claude → /skill → Bash(wrapper.sh) → CLI call → output file → Claude reads → presents

All engines run in non-interactive/headless mode — stateless, no interactive prompts, no persistent state. Output is captured to a temp file, which Claude reads and synthesizes.

buddy-help