Anubis

A local AI DevOps agent powered by Pi coding agent + Unsloth Studio. Anubis runs a full development cycle — plan, code, review, test, audit, PR draft — entirely on local hardware using your Unsloth-served model.

Judge of code. Guardian of the pipeline. Weigher of diffs.

---

Project Structure

anubis/
├── .env.example                    # copy to .env and fill in
├── pyproject.toml                  # project metadata + dev deps (pytest)
├── prompt-collection/
│   ├── prompt_library.md           # full reusable prompt library
│   ├── pi_prompts.md               # Pi slash-command templates (/ci, /review, etc.)
│   └── core-coder-claude-config.md # CoreCoder system prompt + helper prompts
└── agent_src/
    ├── modules/
    │   ├── unsloth_client.py       # OpenAI-compat client for Unsloth Studio
    │   ├── prompt_loader.py        # load/fill prompt templates from prompt-collection/
    │   └── devops_loop.py          # all 6 devops cycle stages
    ├── scripts/
    │   ├── run_cycle.py            # CLI entry point — manual and --auto mode
    │   └── setup_pi.py             # one-time Pi config generator
    └── utilities/
        ├── config.py               # env/dotenv config loader
        ├── git_helpers.py          # diff, log, fetch, pull, changed_files wrappers
        ├── file_writer.py          # parses LLM code blocks and writes files to disk
        ├── pr_creator.py           # opens GitHub PRs via REST API
        └── test_runner.py          # runs pytest and returns pass/fail + output

---

Quick Start

1. Create the venv and install

python -m venv .venv

# Windows
.venv\Scripts\activate

# Linux / Mac
source .venv/bin/activate

pip install -e ".[dev]"

2. Configure your .env

cp .env.example .env
# edit .env and fill in:
#   UNSLOTH_API_KEY=sk-unsloth-XXXX...   (from Unsloth Studio on startup)
#   GITHUB_TOKEN=ghp_XXXX...             (PAT with repo scope — needed for --auto)

3. Install Pi

npm install -g @earendil-works/pi-coding-agent
pi --version

4. Start Unsloth Studio

unsloth studio run --model ~/qwen-models/Qwen3.6-27B-UD-Q4_K_XL.gguf -p 8888

Copy the API key it prints and put it in your .env.

5. Run one-time Pi setup

python agent_src/scripts/setup_pi.py

Auto-detects the loaded model and writes ~/.pi/agent/models.json, ~/.pi/agent/settings.json, and copies the Pi prompt templates.

Verify:

pi --list-models | grep unsloth
pi --print "reply only: pong"

6. Run the full DevOps cycle

Manual (inspect output before anything is committed):

python agent_src/scripts/run_cycle.py \
  --task "add rate limiting to the login endpoint" \
  --repo /path/to/your/repo

Auto (branch → code → write → commit → pytest → review → audit → push → open PR):

python agent_src/scripts/run_cycle.py \
  --task "add rate limiting to the login endpoint" \
  --repo /path/to/your/repo \
  --auto

Open as a draft PR:

  --auto --draft

Run one stage at a time:

python agent_src/scripts/run_cycle.py --task "..." --stage plan
python agent_src/scripts/run_cycle.py --task "..." --stage review

---

DevOps Cycle Stages

Stage	What it does
plan	Produces a numbered implementation plan, identifies files to change
code	Implements the plan against the current file contents
review	Reviews the real git diff main...HEAD for bugs, security issues, and quality
test	Runs pytest against the repo — result reported in PR body either way
audit	Full security + quality audit of the real diff — hardcoded secrets, silent failures, etc.
pr_draft	Drafts the PR title, summary, breaking changes, and testing checklist

---

Using Pi Directly

After setup, cd into any repo and run pi. Then use the slash commands:

Command	What it does
/ci	Run CI checks and report failures
/pr-draft	Draft a PR description for the current branch
/review	Review the branch against main
/issue	Draft a GitHub issue
/sync	Check and pull remote changes
/audit	Security and quality audit of all changes
/deps-check	Check dependencies against a compromised package list
/bloat	Find dead code and over-engineering

---

Notes

• Unsloth prints a new API key every restart. Re-run setup_pi.py or just update .env and ~/.pi/agent/models.json — Pi reloads the file without a restart.
• The model id in models.json must exactly match what GET /v1/models returns. setup_pi.py handles this automatically.
• All LLM calls use temperature=0.1–0.2 for precision. Raise it for creative tasks.
• In --auto mode, pytest runs after the commit and the result is reported in the PR body — pass or fail, the PR always opens.
• Anubis never merges. It opens PRs only. Merging is always a human decision.
• Branch names are auto-generated as anubis/<slugified-task>. Override with --branch.