Imported from dawsonblock/M_Fluid (AGENTS.md). Install upstream with npx skills add dawsonblock/M_Fluid. Copyright stays with the author.

M-flow — Developer & Agent Reference

This file is intended for AI coding assistants (Cursor, Copilot, etc.) and human contributors alike. It describes the repository layout, toolchain, and conventions needed to make changes safely.

1. Repository Map

m_flow/                 Python core (FastAPI + pipeline engine)
  api/                  HTTP routers: add, memorize, search, delete, …
  cli/                  CLI entrypoint (`mflow`)
  adapters/             DB adapters (graph, vector, cache)
  llm/                  LLM providers, prompts, structured output
  core/                 Domain models (Episode, Facet, FacetPoint, …)
  memory/               Memory processing (episodic, procedural)
  retrieval/            Search & retrieval algorithms
  pipeline/             Composable pipeline tasks & orchestration
  auth/                 Authentication & multi-tenancy
  eval/                 Evaluation harnesses & adapters
  shared/               Logging, settings, loaders, cross-cutting utilities
  tests/                Pytest suite (unit / integration / CLI)

m_flow-frontend/        Next.js console (pnpm)
m_flow-mcp/             MCP server — exposes memory as tool calls
mflow_workers/          Modal / worker helpers for distributed runs
examples/               Runnable demo scripts (Python, notebooks)
alembic/                SQL migration scripts (Alembic)

Extension Points

• New data source → add a loader under shared/loaders/
• New pipeline step → add a task under pipeline/
• New graph DB → implement the adapter interface in adapters/graph/
• New vector store → implement adapter in adapters/vector/
• New LLM provider → extend llm/LLMGateway.py

2. Local Development

Python backend (requires Python 3.10 – 3.13)

# Bootstrap
uv sync --dev --all-extras --reinstall

# Run the API server
uv run python -m m_flow.api.client

# CLI quick-start
uv run mflow add "M-flow builds structured memory for agents."
uv run mflow memorize
uv run mflow search "How does M-flow work?"
uv run mflow -ui          # launches backend + frontend + MCP

Frontend console (Next.js / pnpm)

cd m_flow-frontend
pnpm install
pnpm dev              # http://localhost:3000
pnpm lint && pnpm build

MCP server

cd m_flow-mcp
uv sync --dev --all-extras
uv run python src/server.py                  # stdio
uv run python src/server.py --transport sse  # SSE mode

See m_flow-mcp/README.md for Docker and API-mode options.

3. Quality Gates

Tests

PYTHONPATH=. uv run pytest m_flow/tests/unit/ -v          # ~963 test cases
PYTHONPATH=. uv run pytest m_flow/tests/integration/ -v   # needs .env keys

Test layout:

Linting & Formatting

uv run ruff check .       # lint (line-length 100, see pyproject.toml)
uv run ruff format .      # auto-format
uv run mypy m_flow/       # optional type-check

CI (.github/workflows/) runs the same commands; passing locally ≈ passing CI.

4. Conventions

Code Style

Commit Messages

Follow Conventional Commits:

feat(graph): add temporal edge weighting
fix(api): handle missing auth cookie
docs: update installation instructions

Pull Requests

1. Branch from dev.
2. Include scope, local test commands run, and any UI/MCP impacts.
3. Sign commits (git commit -s) and affirm the DCO (see CONTRIBUTING.md).
4. PR titles are validated by CI (pr_lint workflow).

5. Docker

docker compose up                       # backend only
docker compose --profile ui up          # backend + frontend
docker compose --profile neo4j up       # backend + Neo4j

Environment variables are documented in .env.template.