.gitignore

*.zip
__pycache__
dist_release/

# local runtime artifacts
data/
backend/dataflow_cache/
backend/cache_local/
backend/data/dataflow_core/

# editor / OS
.DS_Store
*.swp
*.pyc
.idea/
.vscode/

# ---------------------------------------------------------------------------
# Generated agent assets. The only hand-edited skill source is skills/canonical/;
# everything below is produced from it by installers/generate_agent_assets.py.
#
# These are deliberately NOT tracked. They used to be, which made .claude/skills/
# both a committed artifact and the destination of `--scope project` installs —
# so a `skills`-profile install (rendered without MCP guidance) conflicted with
# the committed `webui` rendering of the same file. One source, generated on
# demand, removes the conflict. See docs/architecture/adr-001-source-of-truth.md.
#
# After cloning:  make skills   (or: python3 installers/generate_agent_assets.py)
.claude/skills/
.cursor/skills/
.codex/skills/
AGENTS.md
.cursor/rules/generating-dataflow-pipeline.mdc
.cursor/rules/dataflow-dev.mdc
.cursor/rules/dataflow-operator-builder.mdc
.cursor/rules/prompt-template-builder.mdc

# stale "* copy" skill dirs accidentally committed in earlier branches
.claude/skills/* copy/

# installer artifacts
.dataflow-install/
.dataflow-install-receipt
.backend.pid
.backend.log

# Project-scoped agent MCP config written by `install.sh configure-agent`:
# .mcp.json and .cursor/mcp.json are tracked. Codex MCP config is never written
# here — Codex only reads ~/.codex/config.toml. (.codex/skills/ IS written, and
# is ignored above as generated output.)