Skip to content

Repository files navigation

Flow

Local, bounded planner → coder → reviewer workflows, with real file tools and a host-enforced verification gate. Ollama is the default backend. CLI, terminal UI and the MCP server share one safe runtime; none uses the old unrestricted shell or auto-loaded plugin path.

Flow can work alone or bidirectionally with Rose: Rose calls Flow over MCP stdio; Flow calls Rose's native editor tools over a private Neovim socket.

Get started

Requirements: Python 3.11+, uv, and Ollama with an installed tool-capable model. POSIX is required for Flow's secure local file I/O; the primary targets are Linux and macOS.

From this checkout:

uv venv .venv
uv pip install --python .venv/bin/python -e '.[editor,dev]'

# In a separate terminal, if not already running:
ollama serve
ollama pull qwen2.5-coder:7b

# No model request or command execution; this also works without Ollama running:
.venv/bin/flow status --workspace /absolute/path/to/project

The editor extra installs optional pynvim. dev installs pytest, Ruff and the wheel builder. Omit extras for the basic CLI/TUI runtime. Dependencies belong in the project environment, not a global Python installation.

Approve checks, then run

Create an operator-reviewed TOML file, for example flow-operator.toml. Workspace values must be at the root, before any TOML table header.

workspace_dir = "/absolute/path/to/project"

[ollama]
base_url = "http://127.0.0.1:11434"
model = "qwen2.5-coder:7b"
timeout = 120 # seconds

[models]
planner = "qwen2.5-coder:7b"
coder = "qwen2.5-coder:7b"
reviewer = "qwen2.5-coder:7b"

[agent]
max_iterations = 6      # model turns per role invocation
max_cycles = 2          # coder → verification → reviewer cycles
max_tool_calls = 64     # total across the entire run
max_context_chars = 100000

[checks.tests]
cmd = [".venv/bin/python", "-m", "pytest", "-q"]
timeout = 60000         # milliseconds, as in Rose
required = true
kind = "test"
filetypes = ["python"]

[checks.lint]
cmd = [".venv/bin/ruff", "check", "."]
timeout = 30000
required = true
kind = "lint"
filetypes = ["python"]

Those executables must exist in the target project's environment; Flow does not install project dependencies or select commands for you. For this Flow checkout, the included config.toml supplies its own test and lint checks.

.venv/bin/flow check --config /absolute/path/to/flow-operator.toml --trusted
.venv/bin/flow run --config /absolute/path/to/flow-operator.toml --trusted \
  "Fix the failing parser tests and explain the changes"
.venv/bin/flow status --workspace /absolute/path/to/project
.venv/bin/flow tui --config /absolute/path/to/flow-operator.toml --trusted

run, check, and status print one JSON object. Exit codes: 0 for an ok result, 1 for failed/unverified results, 2 for configuration/usage errors, and 130 for an interrupted session. Common flags work before or after the command. --workspace overrides configuration before runtime/tool construction; --model overrides every role. With no subcommand, flow starts the TUI.

TUI commands: /help, /status, /check [name], /tools, /models, /model <installed-model>, /build <request>, /clear, /quit. Ctrl-D exits. Natural-language tasks and /build use the same verified runtime. Every task has fresh, separate role contexts. Legacy /plugins, /evolve, /memory, and /feedback are explicitly unavailable, not silently wired to unsafe helpers.

What “verified” means

A run is verified: true only when:

  1. The planner/coder complete within their budgets.
  2. At least one required named check runs, and every required check passes.
  3. Every source file changed through Flow has successful relevant static analysis.
  4. A separate read-only reviewer returns a structured approval.
  5. Edited files have not changed during verification/review and attached editor buffers are not dirty.

Static evidence is an explicitly configured required check with kind = "lint", "typecheck", or "diagnostics" and matching filetypes, or a completed native Rose linter returning status="ok", verified=true. kind="test", "build", and "check" do not by themselves establish static language coverage. Use language names ("python"), extensions (".py"), or an explicit "*" only when a check really covers every language you edit. Unknown source extensions can use their extension or suffix as the coverage key. Documentation/data extensions such as .md, .json, and .toml still require the named gate but do not require a source-language linter.

All required checks run: filetypes describes coverage, never silently skips a required command. Optional checks are reported but cannot replace a required gate. Checks are language-agnostic exact argv: Rust, Go, TypeScript, C, Nix, or another language can use the same mechanism. There are no inferred npx downloads or automatic package-manager installs.

Missing executables/checks, timeout, invalid reviewer output, absent static coverage and exhausted budgets are not success. Failed checks and review issues feed the next bounded repair cycle. Reports retain checks, verification.coverage, verification_history, changed files, role/model/turn counts and tool-call IDs/errors. Empty cached LSP diagnostics are only a snapshot, not verification; a missing LSP remains explicitly unavailable even when a real linter independently verifies its configured scope. Actual diagnostic errors or stale results veto native coverage.

This is evidence within configured checks' scope, not a proof that all behavior, all languages, or an entire application are correct.

Trust and security

  • Default mode is read-only. --trusted allows project edits and explicitly configured named checks. It does not approve new model-selected commands.
  • --config FILE explicitly approves loading that operator file. Workspace config.toml/.flow.toml are never auto-loaded, even with --trusted. Only $XDG_CONFIG_HOME/flow/config.toml (default ~/.config/flow/config.toml) is loaded automatically as user configuration.
  • No model-controlled shell, Python/Bash code execution, executable arguments, cwd, environment overrides, deletion tool, plugins, Lua, or Ex commands. A named check may invoke an interpreter, but its entire argv is fixed by operator configuration and snapshotted before the run.
  • File tools require relative paths. Parent traversal, absolute/Windows-style paths, .git, symlinks, multiply-linked files and outside-root paths are rejected. Writes are atomic, size-limited, descriptor-relative and no-follow. Root replacement is detected as stale. Operator configuration cannot be edited through the model file tools.
  • File reads/writes are limited to 256 KiB. Listings are capped and exclude generated directories. Tool output and model response sizes are bounded.
  • The default Ollama URL is loopback. HTTP redirects and environment proxies are disabled. Non-loopback endpoints require explicit ollama.allow_remote=true; no cloud fallback, API key lookup or automatic external search exists.
  • Trusted project tests/builds/linters execute code and can access the OS, network and files as your user. Named commands can run project configuration, plugins or hooks. This is not an OS sandbox: review the project and the exact check command, use a container/VM for untrusted projects, and do not run Flow with elevated privileges. Concurrent hostile same-user processes are outside the security guarantee.

Legacy executable factories now fail closed. The former application generator and orchestrator names require an explicit Runtime; their raw-write/text-to-shell paths are gone. Bundled prompts/profiles are installed inside flow/skills and read as package resources, never implicitly from a project's skills/ directory.

Rose integration

Configure Rose's native entrypoint with the same absolute workspace and trust:

require('rose').setup({
  workspace = '/absolute/path/to/project',
  trusted = true,
  ollama = { base_url = 'http://127.0.0.1:11434', model = 'qwen2.5-coder:7b' },
  flow = {
    cmd = { '/absolute/path/to/flow/.venv/bin/flow', 'serve',
            '--config', '/absolute/path/to/flow-operator.toml' },
    bridge = true,
    timeout = 300000, -- milliseconds; budget for the complete bounded run
  },
})

Rose appends the workspace/trust flags and starts/passes a private socket for the reverse bridge. Manual invocation has the same interface:

flow serve --workspace /absolute/root --trusted \
  --config /absolute/operator.toml --nvim /private/directory/nvim.sock \
  --editor-timeout 120

The socket must be explicitly supplied and same-user/private: on POSIX use a 0700 parent directory or a socket inaccessible to group/others. TCP is rejected. A Neovim socket grants arbitrary code execution to other same-user clients; keep it private. Flow itself only sends the fixed expressions return require('rose.tools').schemas() and return require('rose.tools').call(...), with names/args as parameters.

Editor schemas are discovered dynamically and filtered to editor_context, editor_diagnostics, editor_symbols, editor_references, editor_lint, editor_check, editor_scip, and editor_debug. Debug launch/probes remain manual: the agent sees only editor_debug status. Planner/reviewer receive read-only tools, not checks/lint or writes.

For buffer safety the reverse bridge also allowlists fixed file_read and file_write calls. Flow validates its paths, trust, role, size and protected configuration first, verifies Rose's resolved workspace matches, then uses Rose's buffer-aware I/O. Unsaved/read-only buffers are not overwritten; loaded buffers are synchronized with disk writes. An unavailable/mismatched bridge does not fall back to unsafe disk writes. Named checks refuse unsaved buffers. Rose's editor-context freshness API v1 is required: Flow checks the global dirty buffer list and compares loaded/observed workspace snapshots across checks and review, including noncurrent buffers. Missing freshness fields fail closed.

pynvim is imported on the main thread before attachment (pynvim records the signal-owner thread at import time); connection creation, requests and close stay on one owning worker. Editor request timeout defaults to 120 seconds and is configurable from 0.1–660 seconds. Model and check timeouts have separate units. SCIP support is decoded JSON supplied by Rose, not raw protobuf decoding. No reverse flow_run calls are exposed.

MCP wire contract

One UTF-8 JSON-RPC 2.0 object per newline; not Content-Length framing. Protocol 2025-11-25; compatible dates 2025-06-18 and 2025-03-26 are accepted. Send initialize, then notifications/initialized, before tools/list or tools/call. ping is available throughout.

Tool Arguments
flow_status {}
flow_run {"task":"..."}
flow_check {} or {"name":"configured-name"}

Tool responses contain content=[{type:"text",text:"<JSON report>"}], structuredContent=report, and isError for non-ok outcomes. The MCP tool list is intentionally only these three tools. Stdout is protocol only; stderr is for diagnostics. Parse/method/parameter errors preserve JSON-RPC IDs where valid. Oversized frames close the session after an error.

The server is deliberately sequential to fit on device. It does not process notifications/cancelled cooperatively during a running request. To cancel, terminate the Flow session (RoseStop): SIGTERM/KeyboardInterrupt unwind active check process-group cleanup and close the editor client; reconnect afterward. Already saved edits are not rolled back. EOF closes an idle session; a busy request finishes within its bounds unless the supervisor terminates it. This stdio transport needs a client supervisor, not stdout redirected to a systemd journal as a pretend network daemon.

Development and verification

uv pip install --python .venv/bin/python -e '.[editor,dev]'
.venv/bin/pytest -q
.venv/bin/ruff check flow tests
.venv/bin/python -m compileall -q flow
.venv/bin/python -m build --wheel --no-isolation

# Include real headless editor tests when Neovim is not on PATH:
NVIM=/absolute/path/to/nvim .venv/bin/pytest -q

Tests use deterministic fake backends, a local fake HTTP Ollama endpoint, real CLI/MCP subprocesses, and optional real headless Neovim. They cover role separation, repair bounds, verification, tool IDs/errors, trust/config/path regressions, timeouts/process cleanup, buffer bridge and wheel resources loaded outside the checkout. Neovim-specific tests report skips if no executable is available. No live Ollama inference/GPU quality test is claimed. A real model may choose poor edits or exhaust its budget; inspect the resulting report and diff.

The Nix expression is retained but was not built in this environment.

MIT — Amor Fati Labs.

About

AI Flywheel

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages