v0.1.0 — Initial Public Release

Lightflow v0.1.0 — Lightweight DAG Workflows for AI Agents

Lightflow turns multi-step AI agent skills into deterministic, checkpointed DAG workflows (lightflow.yaml + actions.py) that pause cleanly at human-in-the-loop approval gates (EXIT_CODE=2), keep intermediate state on disk (passport.json) out of the LLM context window (69–92% token savings), and recover surgically from stage failures.

Highlights

• Zero-Context Entry & Staggered-Turn Execution: Agents invoke lightflow start directly and yield on EXIT_CODE=2 operator gates without burning context tokens polling or re-reading workflow code.
• Checkpointed Passport State: Per-stage outputs (payload.outputs.<stage>.<key>), audit stamps, operator provenance, and file-locked persistence (fcntl on POSIX, msvcrt on Windows).
• Resilience & Safety: Safe-subset CEL expressions (when), exponential-backoff retry_policy, polling_policy loops, automatic rollback_action on failure, and ALL_DONE cleanup stages.
• Cross-Platform CLI & MCP Server: Standalone lightflow CLI (with --payload=@file.json and Windows UTF-8/PowerShell support), lightflow-mcp JSON-RPC stdio server with automatic local module hot-reload, and self-contained interactive HTML DAG visualizer.
• 5 End-to-End Example Workflows & Reproducible Benchmark Suite: hn_digest, usgs_seismic_alert, pypi_upgrade_guard, async_job_watcher, and create_lightflow (meta-workflow for authoring new Lightflows), plus benchmarks/run_benchmark.py.