Repository navigation
Lightflow v0.4.0
Lightflow v0.4.0
Breaking: the lightflow-mcp server is removed. Agents drive Lightflow through the CLI (lightflow / python3 -m lightflow) and skills/run_lightflows/SKILL.md.
Why
Lightflow's model is process-boundary suspension: each command runs in its own process, exits with a status code (0 done, 2 paused at a gate, 5 waiting on a poll, …), and leaves state in passport.json. Hosting that engine inside a long-lived server worked against the design and produced bugs the CLI path cannot have:
- #12 — workflows leaked imported modules into one another inside the server process (
sys.pathmutated per manifest and never restored). - #13 — a single polling stage blocked every other tool call, by default for up to an hour, on the single-threaded stdio loop.
The MCP surface also drifted from the CLI more than once (#5 oversized responses, #7 description wording). Every agent host Lightflow targets can run a shell, and the CLI + skill is the path the benchmarks measure.
If you relied on lightflow-mcp, a thin wrapper that shells out to python3 -m lightflow … per call gives the same tools with per-process isolation; open an issue if you want that upstream.
Removed
lightflow/mcp_server.py,tests/test_mcp_server.py,mcp_config.jsonlightflow-mcpconsole script and themcpServersentry inplugin.json- MCP sections of
README.md,AGENTS.md,CLAUDE.md,skills/run_lightflows/SKILL.md
Unchanged
- Python API:
lightflow.engine,lightflow.lib,lightflow.schema,lightflow.visualizer - CLI commands, flags, exit codes, and
passport.jsonformat - 121 standalone tests pass (
PYTHONPATH=. python3 -B -m unittest discover -s tests)