allfly-crew

allfly-crew is Allfly's unified AI toolset: a home for company-wide automation built on crewAI and AWS Bedrock, designed so new capabilities can be added over time without restructuring.

The first feature is PR/code review: crewAI agents that read a project's review standards and a git diff, then produce structured findings (file, line, severity, comment). It ships in two modes:

• diff — a single generalist reviewer, driven by review_criteria.md.
• categorized — one specialist agent per rule category (one subdirectory of .md rule files = one category), run concurrently via asyncio.gather, each scoped to only its own category's rules.

Layout

src/allfly_crew/
├── main.py          # top-level entrypoint, dispatches to sub-crews
├── reviews/         # the PR/code review sub-crew
│   ├── crew.py               # diff mode (single generalist reviewer)
│   ├── categorized_review.py # categorized mode (specialist agents)
│   ├── config/                # agents.yaml / tasks.yaml for diff mode
│   ├── standards/              # universal rule content (future)
│   └── tools/
│       └── git_diff_tool.py
├── architecture/    # placeholder for a future sub-crew
└── devops/          # placeholder for a future sub-crew

Usage

uv sync

# Single-reviewer diff mode
uv run python -m allfly_crew.main reviews diff <repo_path> <base_ref> [head_ref]

# Categorized mode: one specialist reviewer per rules_dir subdirectory
uv run python -m allfly_crew.main reviews categorized <rules_dir> <repo_path> <base_ref> [head_ref]

Requires AWS Bedrock credentials and a BEDROCK_MODEL_ID environment variable (see .env.example).

More sub-crews (architecture, devops, ...) and a universal rule-selection model (allfly-crew.toml) are planned; see prompt.md for the full roadmap.

Installing as a CLI

Once published to PyPI, install as a standalone tool (no venv activation needed):

uv tool install allfly-crew
# or
pipx install allfly-crew

allfly-crew reviews diff <repo_path> <base_ref> [head_ref]

Release Procedures

• This project uses "trunk" based development off of master.
• PRs are allowed but always folded back to master.
• To release: go to the repo's Actions tab, run the Publish workflow, and choose a bump type (patch, minor, major). It bumps the version in pyproject.toml, tags the release, builds with uv build, publishes to PyPI, and creates a GitHub release.