SWAAG

(S)elfhosted (W)orking (A)utonomous (AG)ent

SWAAG is a local-first autonomous agent core for a direct llama.cpp server. It is built to stay inspectable, replayable, and reproducible on one machine.

Core properties:

• strict context-budget control on every model call
• append-only canonical history with replayable state rebuild
• explicit runtime loop with planning, tools, verification, and recovery
• self-hosted operation against a local llama.cpp HTTP server
• session persistence, control messages, checkpoints, and benchmark plumbing

What is in this repo

This repository contains the full working agent project, including:

• the main swaag package under src/swaag
• CLI entrypoints for agent use, devcheck, lanes, benchmarks, and final proof
• benchmark fixtures and local benchmark wrappers
• full repo-level test suite under tests/
• additional package-level smoke tests under src/swaag/tests
• detailed documentation under doc/

Install

Clone the repo and install it in editable mode:

cd /data/src/github/swaag
python3 -m pip install -e .[test]

Optional benchmark dependencies:

python3 -m pip install -e .[official-benchmarks]

Optional packaging/publish tooling:

python3 -m pip install -e .[publish]

If your current Python environment is not writable for publish extras, use a throwaway build venv instead:

python3 -m venv /tmp/swaag-build
/tmp/swaag-build/bin/python -m pip install build twine
/tmp/swaag-build/bin/python -m build

Once the package is published, the intended install command is:

pip install swaag

Local llama.cpp server setup

SWAAG expects a local llama.cpp server that exposes at least:

• /health
• /tokenize
• /completion

Official llama.cpp repository:

• https://github.com/ggml-org/llama.cpp

A solid general-purpose example model for local testing:

• Qwen2.5-7B-Instruct-GGUF
• https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-GGUF

Example server launch:

llama-server \
  -m /absolute/path/to/Qwen2.5-7B-Instruct-Q5_K_M.gguf \
  --host 127.0.0.1 \
  --port 14829 \
  -c 2048

Default SWAAG config expects the server here:

[model]
base_url = "http://127.0.0.1:14829"
context_limit = 2048

If your server lives elsewhere, override it with either config/local.toml or env vars:

export SWAAG__MODEL__BASE_URL=http://127.0.0.1:14829
export SWAAG__MODEL__CONTEXT_LIMIT=2048

Quickstart

Basic checks:

cd /data/src/github/swaag
python3 -m swaag doctor --json
python3 -m swaag tools

Single-turn ask:

python3 -m swaag ask "Use the calculator tool to compute 6 * 7. Reply with the numeric result only."

Interactive chat:

python3 -m swaag chat

Useful session commands:

python3 -m swaag sessions
python3 -m swaag rename latest my-session
python3 -m swaag control "Keep the current task running, but also answer with plain digits." --session latest
python3 -m swaag checkpoint create --session latest --label before-edit
python3 -m swaag checkpoint restore --session latest
python3 -m swaag history detail latest "What exact command copied src.txt to dst.txt?"

Config

Main defaults:

• src/swaag/assets/defaults.toml

Example local override:

• config/local.example.toml

Environment override prefix:

• SWAAG__...

Config path override:

• SWAAG_CONFIG=/path/to/file.toml

Examples:

export SWAAG__MODEL__BASE_URL=http://127.0.0.1:14829
export SWAAG__SESSIONS__ROOT=/tmp/swaag-sessions
export SWAAG__TOOLS__READ_ROOTS='["/safe/root"]'
export SWAAG__TOOLS__ALLOW_SIDE_EFFECT_TOOLS=true

Testing

Installed-package smoke test:

python3 -m swaag.tests

Full repo test suite:

cd /data/src/github/swaag
python3 run_tests.py
pytest -q

Changed-area developer loop:

python3 -m swaag.devcheck

Fast lane:

python3 -m swaag.testlane fast

System lane:

python3 -m swaag.testlane system

Integration lane:

python3 -m swaag.testlane integration

Live lane against a real local server:

SWAAG_RUN_LIVE=1 python3 -m swaag.testlane live

Final proof loop:

python3 -m swaag.finalproof

Build and publish

Build a source distribution and wheel:

python3 -m build

Or use the provided helper:

./build.sh

The helper builds the package, uploads it with twine, and cleans local build artifacts.

Benchmarks

Main benchmark entrypoints:

python3 -m swaag.benchmark external list
python3 -m swaag.benchmark external smoke --all --output /tmp/swaag-external-smoke --json
python3 -m swaag.benchmark system --all --output /tmp/swaag-system-bench --json

Reproducible bounded SWE-bench fixtures live in:

• src/swaag/benchmark/fixtures/swebench/

Local Terminal-Bench task fixtures live in:

• src/swaag/benchmark/terminal_tasks/

Documentation

Start here:

• doc/installation.md
• doc/testing.md
• doc/architecture.md
• doc/runtime_loop.md
• doc/context_budgeting.md
• doc/history_and_projections.md
• doc/memory_and_editing.md
• doc/live_runtime_profiles.md

License

MIT. See LICENSE.