Lightflow v0.2.0

Lightflow v0.2.0

Highlights

1. Stage Re-Runs & Stale Output Scrubbing (#4):
   • Added lightflow resume --rerun=<stage> (and --stage=<stage> --rerun) with cascade=True by default to re-arm a completed/skipped stage and its downstream dependents while preserving upstream stages.
   • Added clear_stage_outputs so stale payload.outputs.<stage> entries are scrubbed whenever a stage is re-armed, re-run, failed, or skipped.
2. Compact MCP Responses & Field Extraction (#5):
   • run_lightflow, resume_lightflow, and get_lightflow_status now return a compact summary by default (completed_count, not_started_count, non-COMPLETED stages, output_keys, and active gate details), with verbose=True for the full passport + Mermaid diagram and fields=[...] for dotted-path extraction (outputs.stage.key or payload.outputs.stage.key).
3. Timeout & Rollback Thread Safety (#3):
   • StageTimeoutError tracks the timed-out daemon worker thread across retries and polling ticks; run_rollback_if_defined skips rollback_action if any timed-out attempt's background thread is still alive (any(w.is_alive() for w in timed_out_workers)).
   • _execute_engine_loop catches StageTimeoutError so independent branches and ALL_DONE cleanup stages still execute before raising.
4. Windows Benchmark Compatibility & Directory --lightflow=<dir> Convention (#6, #7):
   • UTF-8 subprocess decoding (encoding="utf-8", errors="replace"), POSIX path normalization, and directory --lightflow=<dir> resolution across CLI, MCP server, skills, and examples.
5. Neutral operator_action Prompting (#8):
   • Updated operator_action checkpoint footers and skill instructions so human-decision gates require explicit human confirmation while agent-work checkpoints (such as create_lightflow) execute directly.
6. Unified N=5 A/B Consistency Benchmark & Production Recovery Examples (#9):
   • Added examples/blue_green_release (8 stages) and examples/incident_db_failover (9 stages) alongside 7 matched traditional SKILL.md baselines (benchmarks/baseline_skills/<workflow>/SKILL.md).
   • Upgraded benchmarks/run_benchmark.py (--trials=5 default, 40/40 invariant checks) with a non-invasive sys.setprofile tracer and documented 3 controlled N=5 live zero-context LLM subagent cohorts (15 sessions = 30 workflow runs).