fly-harness

v0.4.0 — a sparse neural harness, not a fly-brain simulation.

Formula: Agent = deployed bio-sim Model + harness. The harness docks to a running sim through a stable contract. FlyWire / MaleCNS files are weight dumps, not that contract.

The public contract is:

1. BrainState — membrane potentials, sparse synaptic weights (scipy.sparse CSR), timestamp, JSON save/load
2. FlyHarness.step(obs) -> action — encode → ModelBackend.tick → decode (default backend is the in-process LIF/rate circuit)
3. Encoder / Decoder — typing.Protocol contracts (optional BaseEncoder / BaseDecoder ABCs; no string class-name checks)
4. load_connectome — load a circuit subset from a local sparse edge list (.npz / .csv) into BrainState
5. ModelBackend — swappable Model port: tick (encoded input → neural output), reset, n_neurons, model_id

The 24-neuron touch-reflex loop is a test/demo fixture. It is not the model, and it is not a scaled-down FlyWire brain.

What this is not

• Not a 140k-neuron FlyWire connectome, upload, or dense 140k×140k matrix
• Not an arbitrary-scale whole-brain runtime (memory is bounded by the loaded subset)
• Not caveclient or a built-in biomechanics engine — FlyGym/NeuroMechFly is an optional body extra, not the core product
• Not an MCP microkernel or FastAPI service — MCP is an optional protocol extra, not the core product
• Not OpenRouter-the-company, a marketplace, billing system, hosted cloud gateway, or public bio-sim catalog — BioSimRouter is a harness-side port; biorouter is an optional local HTTP process (stdlib, 127.0.0.1 by default)
• Not a consciousness/upload claim

Install

Python 3.10+ with numpy and scipy. Core install stays numpy/scipy-only:

pip install -e ".[dev]"

FlyGym is optional:

pip install -e ".[flygym]"

MCP is optional (stdio server via FastMCP):

pip install -e ".[mcp]"

biorouter is optional and stdlib-only (no FastAPI). The extra exists so it stays an extension of this repo, not a second product or GitHub repository:

pip install -e ".[biorouter]"

The step loop

from fly_harness import BrainState, FlyHarness
from fly_harness.protocols import Decoder, Encoder

harness = FlyHarness(state, encoder, decoder)
result = harness.step(observation)
action = result.action

step encodes the observation into a length-n_neurons current vector, advances the docked Model one tick, and decodes an action from the neural output.

Default construction still wraps the in-process LIF/rate dynamics (InProcessLifBackend). FlyHarness.step(obs) -> action is the same public call.

BrainState can be snapshotted independently of the harness:

state.save("brain.json")
restored = BrainState.load("brain.json")

Docking modes

Two ways to attach a running Model. This package is the port, not a hosted catalog of sims.

1. Direct — one backend per vendor / deployed sim

from fly_harness import DirectBioSimBackend, FlyHarness

backend = DirectBioSimBackend(n_neurons=8, model_id="vendor.fake")
harness = FlyHarness(encoder=encoder, decoder=decoder, backend=backend)
result = harness.step(observation)

DirectBioSimBackend docks to one running sim. The in-process FakeDeployedSim is enough for tests. A thin HTTP client is available if you already host a sim (url=..., bodies include a model field). There are no FlyWire / CAVE / neuPrint clients and no 140k dense matrices.

2. Router — one entry, select by model id

from fly_harness import BioSimRouter, DirectBioSimBackend, FlyHarness, UnknownModelError

router = BioSimRouter(
    {
        "vendor.a": DirectBioSimBackend(n_neurons=8, model_id="vendor.a"),
        "vendor.b": DirectBioSimBackend(n_neurons=8, model_id="vendor.b"),
    },
    model_id="vendor.a",
)
harness = FlyHarness(encoder=encoder, decoder=decoder, backend=router)
harness.step(observation)
router.use("vendor.b")
harness.step(observation)

# unknown model_id raises UnknownModelError (this is a port, not a marketplace)

BioSimRouter is OpenRouter-shaped: one entry, a model / model_id field, unknown ids fail clearly. Optional remote_url= forwards that model field to biorouter (HttpModelBackend). This repo does not ship a hosted marketplace.

You can also pass a backend positionally: FlyHarness(backend, encoder, decoder).

biorouter

OpenRouter routes existing LLMs. biorouter routes existing deployed biological simulation models. Same idea (model id), different substrate.

This extra is not the harness, not FlyWire dumps, and not a marketplace. It is a local HTTP process (127.0.0.1 by default) that already-deployed harness clients call: HttpModelBackend, DirectBioSimBackend(url=...), BioSimRouter(remote_url=...). Stdlib http.server only — no FastAPI, no billing, no cloud gateway.

pip install -e ".[biorouter]"   # empty extra; core install is enough to run the CLI
biorouter
# or
python -m fly_harness.router --host 127.0.0.1 --port 8765

Endpoints (JSON; bodies/query include an OpenRouter-shaped model field):

• POST /tick — {"model": "...", "input": [...]} → neural output
• POST /reset — {"model": "...", "potentials": null | [...]}
• GET /status?model=...

Default in-process registry: the 24-neuron LIF fixture (fly-harness.in-process-lif) plus FakeDeployedSim ids (fake.deployed, fake.deployed.gain). Unknown model → HTTP 404 → harness UnknownModelError.

from fly_harness import DirectBioSimBackend, FlyHarness

backend = DirectBioSimBackend(
    url="http://127.0.0.1:8765",
    model_id="fake.deployed",
    n_neurons=8,
)
harness = FlyHarness(encoder=encoder, decoder=decoder, backend=backend)
harness.step(observation)

FlyHarness.step(obs) -> action is unchanged. This extra is not a 140k FlyWire runtime and does not download MaleCNS.

Example: biorouter through FlyHarness.step

OpenRouter routes existing LLMs; biorouter routes existing deployed bio-sims. This example starts a local stdlib server (or attaches with --url), docks HttpModelBackend / DirectBioSimBackend(url=...), and calls FlyHarness.step. It uses the 8-neuron FakeDeployedSim fixture — not FlyWire / MaleCNS, not a marketplace.

python examples/biorouter_loop.py
# or, after install:
python -m fly_harness.demo.biorouter_loop
fly-harness-biorouter-demo
# attach to a process you already started:
biorouter --host 127.0.0.1 --port 8765
python examples/biorouter_loop.py --url http://127.0.0.1:8765

Load a sparse connectome

load_connectome maps global neuron ids (any integers) onto contiguous local indices and builds a CSR weight matrix of shape (n_subset, n_subset).

from fly_harness import FlyHarness, load_connectome

loaded = load_connectome("tests/fixtures/mini_circuit.npz")
# loaded.state.weights is scipy.sparse CSR, shape (n_subset, n_subset)
# loaded.neuron_ids maps local index -> global neuron id
# loaded.id_to_index maps global id -> local index

harness = FlyHarness(loaded.state, encoder=..., decoder=...)
result = harness.step(observation)

fly-harness-loader-demo
# or
python -m fly_harness.demo.loaded_circuit

Supported files:

• NPZ — arrays pre, post, weight (int ids + float weights); optional neuron_ids for the circuit subset
• CSV — header pre_id,post_id,weight

Pass an explicit subset with neuron_ids=[...] or neuron_ids_file="ids.txt" (one id per line). Only synapses whose pre and post are both in the subset are kept.

Fixture: 24-neuron reflex

Shipped only so the contract is executable without an external connectome file. Left touch biases a right turn (and the reverse). Do not treat this as a biological circuit.

fly-harness-reflex-demo
# or
python -m fly_harness.demo.reflex

from fly_harness import FlyHarness
from fly_harness.demo import ReflexDecoder, ReflexEncoder, build_reflex_connectome
from fly_harness.demo.codec import TouchObservation

state = build_reflex_connectome()
harness = FlyHarness(state, ReflexEncoder(), ReflexDecoder())
result = harness.step(TouchObservation(touch_left=1.0))
print(result.action)  # ReflexAction(turn=1, forward=..., brake=...)

Optional body: FlyGym / NeuroMechFly

This is an extension, not the core product. It does not ship a 140k-neuron brain, does not rewrite FlyGym, and does not run MuJoCo unless you install the extra.

FlyGymEncoder / FlyGymDecoder map NeuroMechFly observations (joint angles, contact forces) and actions (joint targets, per-leg adhesion, optional muscle/tendon commands) onto FlyHarness.step. They work on dicts — no MuJoCo import. FlyGymHarnessEnv is a thin wrapper around an env you construct with FlyGym.

from fly_harness import FlyHarness
from fly_harness.demo.connectome import SENSORY_INDICES, build_reflex_connectome
from fly_harness.flygym import FlyGymDecoder, FlyGymEncoder, FlyGymHarnessEnv

state = build_reflex_connectome()  # fixture circuit, not FlyWire
harness = FlyHarness(
    state,
    FlyGymEncoder(state.n_neurons),
    FlyGymDecoder(state.n_neurons),
    sensory_indices=SENSORY_INDICES,
)
# env = your FlyGym / NeuroMechFly simulation (not constructed here)
body = FlyGymHarnessEnv(env, harness)
obs, info = body.reset()
result = body.step()
# result.action.as_env_dict() -> {"joints": ..., "adhesion": ...}

Gymnasium FlyGym (flygym-gymnasium, installed by the extra; import name flygym or flygym_gymnasium) takes env.step({"joints", "adhesion"}). FlyGym 2.x Simulation is duck-typed via set_actuator_inputs / set_leg_adhesion_states when you pass actuator_type.

Optional protocol: MCP

This is an extension pack, not the microkernel. BrainState / FlyHarness stay numpy/scipy-only and do not import MCP. The default server steps the 24-neuron reflex fixture, not a 140k FlyWire brain.

pip install -e ".[mcp]"
# stdio (default FastMCP transport) — point an MCP host at this command:
fly-harness-mcp
# or
python -m fly_harness.mcp

Tools (no live client required to unit-test the handlers):

• harness_step(touch_left, touch_right) — one FlyHarness.step
• harness_reset() — zero potentials and the clock
• harness_status() — neuron count, timestamp, package version

Handlers are importable without FastMCP:

from fly_harness.mcp import HarnessSession

session = HarnessSession()
print(session.step(touch_left=1.0)["action"])

Example: protocol + body through the harness

This proves optional MCP and FlyGym extras can attach to the same FlyHarness.step. It uses the 24-neuron reflex fixture. It is not a 140k FlyWire or ~166k MaleCNS upload, and it does not rewrite FlyGym or the MCP pack.

python examples/mcp_flygym_loop.py
# or, after install:
python -m fly_harness.demo.composed
fly-harness-composed-demo

Default run uses an in-process body stub (no MuJoCo). --try-flygym attempts a real NeuroMechFly sim and falls back to the stub. --serve starts the optional FastMCP stdio server on that composed session (pip install -e ".[mcp]"):

python examples/mcp_flygym_loop.py --serve

An MCP host that calls harness_step(touch_left, touch_right) therefore steps the harness and applies the decoded joint/adhesion command to the body.

Tests

GitHub Actions on main and pull requests runs pip install -e ".[dev]" then pytest — no FlyGym/MCP extras, no MuJoCo. Skip/mock tests in those extras still pass. The biorouter extra is stdlib-only, so its tests run in that same core CI.

Default local suite does not need MuJoCo. The FlyGym smoke test skips unless flygym imports and a NeuroMechFly sim can start:

pytest
# with the FlyGym extra, still skip-friendly if MuJoCo/display is missing:
pytest -q
# FlyGym smoke only:
pytest tests/test_flygym_adapter.py -k smoke

Model-backend docking (default LIF, fake deployed sim, router model_id switch, unknown-id error) is core and needs no extras:

pytest tests/test_backend.py tests/test_harness.py

The optional biorouter process starts a stdlib server and ticks through HttpModelBackend / FlyHarness (still no FastAPI, no extras required):

pytest tests/test_router.py
# after install:
biorouter --help
python -m fly_harness.router --help

The biorouter example is in-thread by default (no live biorouter daemon). Attaching to a real process is skip-gated on BIOROUTER_URL:

pytest tests/test_biorouter_example.py
python examples/biorouter_loop.py

MCP tests mock FastMCP and do not start a live MCP client. They pass without fly-harness[mcp]. With the extra, a factory smoke checks FastMCP constructs (still no stdio client):

pip install -e ".[mcp]"
pytest tests/test_mcp_extension.py

The composed MCP+FlyGym example is also mock/skip: no MuJoCo and no live MCP client.

pytest tests/test_composed_example.py
python examples/mcp_flygym_loop.py

License

MIT — see LICENSE.