Hive_review: Ruby CLI for multi-agent AI code review

Hive_review is a Ruby CLI and library that automates a looped code-review workflow. The core loop works as follows: an implementor agent writes or fixes code, one or more reviewer agents examine the current diff in parallel, their feedback is collected and sent back to the implementor, and the cycle repeats until all reviewers return the status marker HIVE_REVIEW_STATUS: PASS or until the configured maximum rounds are reached.

The tool is invoked from within a Git repository (or worktree) and accepts several key arguments: --repo-dir specifies the repo path (defaults to current directory), --implementor names the model config for the coder (defaults to GPT-5.5-xhigh-login), --reviewers lists comma-separated reviewer configs (defaults to Claude-Opus-4.8-max-login), and --max-rounds sets the iteration limit (defaults to 5). An optional --prompt provides the initial task; if omitted, hive_review reviews the current Git diff. Sessions can be resumed if the same --session-dir is reused in an interactive terminal.

Model selection is managed through a top-level model_config.json file, where each entry is keyed by a display name (used as a CLI selector and in provider-home directory names) and includes fields for agent type, model name, effort level, and auth method. Default seeded configs include GPT-5.5-xhigh-login, GPT-5.4-xhigh-login, Codex-GPT-5.3-xhigh-api, Claude-Opus-4.8-max-login, and Claude-Sonnet-4.6-max-login. Users can add custom configs for Gemini, Grok, DeepSeek (via OpenCode), Kimi, and other variants by editing the JSON; model names are provider-shaped strings (e.g., gpt-5.5, claude-opus-4-8, deepseek/deepseek-v4-pro), allowing new releases to be configured without code changes.

Provider CLIs are installed per-platform via npm or curl; on macOS, Homebrew alternatives are available for some providers. Runtime state is kept local to the repo under .hive_review_state/provider_homes/, so each model config gets its own provider-home directory for auth tokens and session data. The tool supports macOS, Linux, and Windows via WSL (native Windows shells are not supported). Reviewers run in parallel within each round and their outputs are aggregated in the configured reviewer order; missing status markers are treated as change requests. If all reviewers pass, the session completes; if max rounds are reached and the human does not continue in interactive mode, status becomes max_rounds_reached. Optional flags include --dry-run (validate selections without launching providers), --no-diff (exclude Git diff from prompts), and --initial-implementation (open an interactive implementor session before starting the review loop).

Hive_review addresses a common bottleneck in code-driven workflows: the back-and-forth between implementation and review. By automating the orchestration of multiple AI agents across different providers, it lets teams define a single task (via a prompt or Git diff) and have it automatically cycle through fix-and-review iterations without manual coordination. The tool's design prioritizes isolation—all runtime state (provider homes, transcripts, session IDs, logs) stays within the repository under .hive_review_state, avoiding pollution of user-global provider config directories.

The architecture is flexible: users can choose one implementor and one or more reviewers from a JSON config file, supporting newly released models without code changes. Authentication is handled both via login (interactive) and API-key (unattended), and the tool includes helpers to install missing provider CLIs and verify auth before running. For teams using DeepSeek, the OpenCode wrapper eliminates the need for a separate provider integration, reducing setup friction.

No comments yet. Be the first to share your thoughts!