Releases: primorLee/agent-workflow-platform
Release list
AWP v0.1.0
AWP v0.1.0
This is the first public source release of Agent Workflow Platform: a
production-derived, local-first starter toolkit for teams turning an existing
model, Agent SDK, runtime, or CLI into a product.
What is actually wired
- OpenAI-compatible model → reference Agent CLI → Electron streaming and
native-session resume. - One explicit managed-task tool → loopback FastAPI control plane → SQLite and
strict Redis event path → trusted Python worker → terminal result. - Deterministic no-account Desktop demo and read-only operational surfaces.
The file-backed workflow toolkit and Go VM-agent packages are separately
tested composition boundaries. They are not silently connected to the default
model turn or FastAPI service.
Reliability changes in this release
The task path now claims exactly the worker's free capacity. Every running task
has an expiring lease and a UUID attempt fence; heartbeats renew that exact
attempt, dead-worker claims requeue under a bounded retry budget, stale results
cannot overwrite a newer attempt, and identical terminal retries are
effect-idempotent. Stale Agent/Session records are reaped and SQLite lifecycle
changes use a fail-closed schema version.
CI submits a real ten-task Compose backlog and then hard-stops a busy worker,
waits for lease-based requeue, restarts it, and requires completion under a new
attempt. Exact commands and non-claims are in
Reliability evidence.
The Linux Go runner now starts each task in an isolated process group, removes
descendants on explicit or context cancellation, drains stdout and stderr, and
only then emits terminal completion. This prevents orphan children and missing
final log chunks under scheduler or race-detector timing.
Boundaries
This release does not provide a hosted multi-tenant service, arbitrary Agent
auto-spawning, a VM broker, exactly-once external side effects, or an
untrusted-code sandbox. It contains no signed desktop installer or VM binary.
See the public release boundary before
deployment.