z4j-bare 1.2.0

z4j framework-free agent bootstrapper (Apache 2.0)

z4j-bare

License: Apache 2.0 Status: v1.0.0 - first public release.

Framework-free agent runtime for z4j. Ships both as a pure-Python supervisor for standalone workers (no Django / Flask / FastAPI) and as proof that z4j-core is genuinely framework-free - z4j-bare is what every framework adapter is built on top of.

Install

pip install z4j-bare

# Optionally add dev-mode filesystem watcher (for live task-registry
# reloads when tasks.py changes on disk; off in production):
pip install "z4j-bare[watcher]"

What it does

• AgentRuntime - pure-Python agent runtime, no framework hooks.
• WebSocket transport with exponential-backoff reconnect.
• HTTPS long-poll fallback for networks that block WebSocket.
• Local SQLite buffer (WAL mode, size-capped, crash-safe event storage).
• Command dispatcher - routes brain commands to registered engine + scheduler adapters.
• Heartbeat loop with liveness + last-seen stamping.
• Orchestrator detection - identifies Docker / Kubernetes / systemd / supervisord hosts for restart_worker safety-gating.
• Per-project HMAC v2 signing using the shared wire-protocol helpers from z4j-core.
• Process singleton guard - prevents duplicate agents from the same project on one host.
• CLI: python -m z4j_bare run --config ...

Use cases

• Celery / RQ / Dramatiq workers not served by a web framework.
• ML pipelines that run Celery but don't use Django.
• Containers where the agent runs as its own process, separate from the app.
• Custom Python scripts that dispatch tasks and want to report lifecycle events to the z4j brain.

Usage - in-process

from celery import Celery
from z4j_bare import install_agent
from z4j_celery import CeleryEngineAdapter

app = Celery("myproject", broker="redis://localhost/0")

agent = install_agent(
    engines=[CeleryEngineAdapter(celery_app=app)],
    brain_url="https://z4j.internal",
    token="z4j_agent_...",
    project_id="ml-pipeline-prod",
)

Usage - standalone process

python -m z4j_bare run \
    --brain-url https://z4j.internal \
    --token $Z4J_TOKEN \
    --project-id ml-pipeline-prod \
    --engine celery \
    --celery-app myproject.celery:app

Reliability contract

Per the z4j safety rule (docs/CLAUDE.md §2.2), the agent must never break the host application:

• No exception inside the runtime propagates to Celery / Django / FastAPI signal handlers.
• Network I/O never blocks signal handlers (events queue to an in-memory ring, flushed from a background thread).
• When the brain is unreachable, events buffer to local SQLite with exponential backoff.
• Buffer drops oldest events when full; the host app is never slowed or blocked.

Documentation

• Adapter guide - how z4j-bare composes with engine + framework adapters.
• Architecture - agent/brain topology, transport layer, reconciliation.
• Security - threat model, HMAC v2 envelope signing, redaction.

License

Apache 2.0. See LICENSE. z4j-bare installs cleanly into proprietary Python workers - there is no copyleft obligation. The brain server it talks to (z4j-brain) is AGPL-3.0; communicating with it over the network does not subject your worker code to AGPL.

Links

• Homepage: https://z4j.com
• Documentation: https://z4j.dev
• Issues: https://github.com/z4jdev/z4j-bare/issues
• Changelog: CHANGELOG.md
• Security: [e-mail] (see SECURITY.md)