ansiburr

Pardon me, are you Ansi-Burr, Sir?

Run Ansible modules as Burr state-machine actions in Python.

A single decorator wraps an Ansible module call as a Burr @action. The module runs through ansible-runner against the target host. Its result projects into Burr's State, and the action's _last_failed, _last_changed, and _last_msg flags become available to downstream transitions. The output is a standard Burr Application that runs, persists, traces, and serves like any other Burr graph.

The GIF above is a single Ansible playbook (set_fact, block, loop, notify/handlers, changed_when) lifted into a Burr Application via ansiburr.from_playbook(...) and walked one action at a time. Every Ansible task is a discrete observable Burr step, and so is every loop iteration, every notify marker, and every handler. No opaque ansible-playbook invocation in the middle of the trace. The playbook driving it is at examples/from_playbook_advanced/playbook.yml; the walker is at examples/from_playbook_walker.py.

What you can build

• Self-healing service workflows that observe, decide, and remediate one Ansible module at a time, with every step visible in Burr's tracker.
• SRE agents where an LLM picks one label from a fixed allow-list of remediation actions and the FSM (not the model) enforces termination and retry policy.
• Cross-platform automation that gathers facts on the target up front and dispatches to the right modules based on the OS family, init system, or package manager.
• Plan-then-apply pipelines using Ansible's --check and --diff with a deterministic review gate before any change runs.
• Polling sub-graphs (port readiness, service health, file existence) where every poll attempt is a discrete step in the trace.

Install

uv add ansiburr
# or
pip install ansiburr

ansible-core is pulled in transitively as a runtime requirement. Install additional collections via ansible-galaxy:

ansible-galaxy collection install community.general community.crypto community.docker ansible.posix

Quickstart

Save as my_fsm.py and run with python my_fsm.py. No remote host or extra setup required: ansible.builtin.ping runs against localhost via the ansible-runner already pulled in by pip install ansiburr.

from burr.core import ApplicationBuilder, action
from ansiburr import module_action, initial_sentinels


# `@module_action` turns a function that returns a dict of Ansible module
# args into a Burr `@action`. The module runs through ansible-runner and
# the result projects into State. `writes=["ping"]` projects the module's
# `ping` field; ansiburr also writes ambient `_last_*` sentinels on every
# call (`_last_failed`, `_last_changed`, `_last_msg`, etc.).
@module_action("ansible.builtin.ping", writes=["ping"])
def check(state):
    return {}


# A regular Burr `@action` is a pure-Python step. It reads from State and
# returns the new State. Mixing module actions and plain actions in the
# same graph is the common pattern.
@action(reads=["ping", "_last_failed", "_last_msg"], writes=["report"])
def summarize(state):
    if state["_last_failed"]:
        return state.update(report=f"ping failed: {state['_last_msg']}")
    return state.update(report=f"ansible reachable: ping={state['ping']!r}")


app = (
    ApplicationBuilder()
    .with_actions(check=check, summarize=summarize)
    .with_transitions(("check", "summarize"))
    .with_state(**initial_sentinels(), ping="", report="")
    .with_entrypoint("check")
    .build()
)

_, _, final = app.run(halt_after=["summarize"])
print(final["report"])
# -> ansible reachable: ping='pong'

From there, the moves are:

• Add host() to point a group of actions at a remote target without repeating the connection dict.
• Use host.gather_facts() to expand ansible_facts into top-level state keys (ansible_pkg_mgr, ansible_os_family, etc.) and branch transitions on them.
• Use wait_until() for polling sub-graphs where each attempt is a discrete trace step.
• Use check_mode=True + diff=True for plan-then-apply patterns with a deterministic review gate.

Working examples of each are in examples/.

From an existing playbook

If you already have an Ansible playbook, from_playbook(...) lifts it into a runnable Burr Application without rewriting the YAML. The full demo lives in examples/from_playbook/.

For a sense of how much of the wild ansiburr ingests: all six of the most-downloaded geerlingguy roles (the de facto Galaxy benchmark) convert directly from their unmodified published source.

Here's the smaller demo playbook from examples/from_playbook/:

# playbook.yml
- name: tool availability check
  hosts: localhost
  gather_facts: no

  tasks:
    - name: check for git
      ansible.builtin.command:
        cmd: git --version
      register: git_check
      ignore_errors: yes
      changed_when: false

    - name: report git availability
      ansible.builtin.debug:
        msg: "git is installed: {{ git_check.stdout }}"
      when: git_check.rc == 0

    - name: check for jq
      ansible.builtin.command:
        cmd: jq --version
      register: jq_check
      ignore_errors: yes
      changed_when: false

# run.py
import ansiburr

app = ansiburr.from_playbook("playbook.yml")
last_action, _, final = app.run(halt_after=["done", "escalate"])

print(f"git: rc={final['git_check']['rc']} {final['git_check'].get('stdout', '').strip()}")
print(f"jq:  rc={final['jq_check']['rc']}  {final['jq_check'].get('stdout', '').strip()}")

Output (when both binaries are present):

git: rc=0 git version 2.50.1 (Apple Git-155)
jq:  rc=0  jq-1.7.1-apple

The converter handles a substantial subset of single-play Ansible: name, when: (including attribute access on registered names), register:, failed_when:, changed_when:, ignore_errors:, become:, gather_facts:, play-level vars:, block: (group-only), include_tasks: and import_tasks: (literal paths), notify: plus handlers:, loop: and with_items: (literal lists), set_fact:, and Jinja2 templates in task arguments. Each task runs as its own play under the hood, so Jinja references to set_fact-written and registered values resolve through Burr state rather than ansible's cross-play context. rescue: / always:, Jinja-templated loop: or include: values, roles:, multi-play files, and the parallelism keywords (serial: / strategy:) raise UnsupportedPlaybookConstruct at conversion time with the offending node named in the message, so a partially-converted FSM never starts. The full supported-vs-rejected list is in REFERENCE.md. From here, the resulting Application can be hand-edited (add transitions, swap actions, wire in policy gates) or the playbook can stay as the source of truth and be re-converted.

CLI

pip install ansiburr ships an ansiburr command for running and inspecting FSMs without writing a wrapper script.

# Run a playbook directly (no manual conversion).
ansiburr run playbook.yml

# Run a Python module that exposes ``app`` (or a ``build_application()`` callable).
ansiburr run examples/localhost_disk_check.py

# Print the FSM structure as mermaid (default), graphviz dot, or plain text.
ansiburr graph examples/from_playbook/playbook.yml --format text
ansiburr graph examples/from_playbook/playbook.yml --format mermaid
ansiburr graph examples/from_playbook/playbook.yml --format dot

ansiburr run halts on done or escalate by default, and accepts --halt-after ACTION (repeatable) to override.

Demo corpus

examples/ contains twelve self-contained FSMs plus three playbook-conversion demos. Most of the FSMs run against a local Docker container set up by examples/service_remediation/setup.sh; the conversion demos and a few others run locally.

Each example runs in seconds. Run an individual demo with uv run python examples/<name>/fsm.py (or run.py for the conversion demos).

The full library API and the Ansible-playbook-idiom mapping live in REFERENCE.md.

Dependencies and licensing

ansiburr is licensed under the Apache License 2.0. See LICENSE for the full text.

ansiburr imports only Apache-2.0 and MIT licensed code (ansible-runner, burr, pyyaml). At runtime it requires ansible-core, which is licensed under GPL-3.0-or-later and is invoked as a separate subprocess by ansible-runner rather than imported directly. Anyone redistributing an installed ansiburr environment should be aware that the bundled ansible-core component carries GPL-3.0+ obligations, and that individual Ansible collections in the user's ansible_collections path may have their own licenses.

The NOTICE file contains the canonical attribution and license summary.

This README is engineering documentation, not legal advice.

Development

git clone https://github.com/msradam/ansiburr
cd ansiburr
uv sync
uv run pytest
uv run ruff check .
uv run mypy src/ansiburr

Most examples require a small Docker container. examples/service_remediation/setup.sh builds the image and generates a per-clone SSH key; examples/service_remediation/start.sh runs the container.

ansiburr was developed with significant AI assistance (Anthropic's Claude). All changes were reviewed and committed by the project owner.

Acknowledgements

• The Burr team (Apache Software Foundation) for the FSM substrate.
• The Ansible community for the module ecosystem.
• IBM Research and UC Berkeley for the MAST failure-mode taxonomy (blog, arXiv:2503.13657).

License

Apache License 2.0. See LICENSE and NOTICE.