Python Reference
Manifest, hooks, parameters, packages, and output in Furge scripting.
Script structure
script = {
"name": "my-signal",
"version": "2",
"lookback": 20,
"params": {"window": {"type": "number", "default": 5}},
}
SOURCE = "btc@candles@hyperliquidf:timeframe=60"
def on_data(ctx, history):
candle = history.source(SOURCE, 0)
if candle is not None:
return {"metrics": {"close": candle["c"]}}script and on_data(ctx, history) are required. No package import or base class is required. version must be "2"; lookback retains records per source (2 through 5,000).
Parameters
Define defaults in script["params"], then override them in the editor's Parameters field:
window=10Read the resolved value inside a hook:
window = int(ctx.params["window"])Supported parameter types are number, boolean, and string. Literal defaults are prefilled in the editor. Computed defaults are resolved by the runtime. Changing parameters affects the next run, not an existing deployment.
Hooks and context
| API | Purpose |
|---|---|
on_data(ctx, history) | Runs after an accepted source event has been added to history. |
on_execution(ctx) | Receives lifecycle events through ctx.execution. |
on_finish(ctx, history) | Final reports on normal completion/stop; not guaranteed after a crash or resource kill. Execution methods are disabled here. |
ctx.source | Triggering startup selector or dynamic subscription handle. |
ctx.subscribe(request) | Add a live market feed during a run. |
ctx.unsubscribe(handle) | Release a dynamic market feed. |
ctx.timestamp_ms | Triggering event time in Unix milliseconds. |
ctx.state | JSON dictionary checkpointed after successful hooks for recovery of the same live job; at most 1 MiB. New deployments start with new state. |
ctx.params | Resolved parameters for this run. |
ctx.study | Native studies such as SMA/EMA; they do not place orders. |
history.source(SOURCE) | Retained records, oldest to newest. |
history.source(SOURCE, 0) | Newest record or None. |
See Data Sources for records and Execution for trading actions. Multiple sources update independently; guard missing data and do not assume synchronized freshness.
Editor types
New starters include annotations for the Python editor. Add them to an existing script to see hook inputs, request keys, and return types.
from typing import TYPE_CHECKING
if TYPE_CHECKING:
from furge_scripting import Context, History
def on_data(ctx: "Context", history: "History"):
...The runtime supplies these arguments. The type import is used by the editor.
Artifacts
Use ctx.artifact_path(...) for run-owned files. Do not write to arbitrary filesystem paths.
import json
def on_finish(ctx, history):
with open(ctx.artifact_path("summary.json"), "w") as output:
json.dump({"candles": len(history.source(SOURCE))}, output)Save figures instead of opening windows. Exported files appear under Artifacts after the run stops. Image/text previews are available for supported files up to 2 MiB; larger files can be downloaded. HTML output is downloaded, not executed inline.
Available packages
The runtime image bundles these packages; no per-user installation is needed:
| Package | Import | Use |
|---|---|---|
| NumPy | import numpy as np | Arrays and numerical calculations |
| pandas | import pandas as pd | Tables, rolling windows and time series |
| SciPy | import scipy | Statistics, optimization and signal processing |
| statsmodels | import statsmodels.api as sm | Regression and time-series models |
| scikit-learn | import sklearn | Preprocessing and machine learning |
| Matplotlib | import matplotlib.pyplot as plt | Save charts as images |
| Seaborn | import seaborn as sns | Statistical charts |
| Plotly | import plotly.graph_objects as go | Interactive HTML charts |
Available versions follow the installed runtime image. Upgrade the runtime to receive package updates. Additional package installation, arbitrary network calls, and child processes/threads are blocked by the hosted sandbox; use a single worker where a library accepts n_jobs.
Runtime safety
User Python runs separately from execution credentials. Signing stays inside the daemon. The hosted sandbox restricts filesystem access, networking, processes, memory, and hook execution time. It does not make a strategy profitable or eliminate trading risk.
A failed check or exception can stop a run. Read the error and Logs before retrying. Never store secrets in code, parameters, output, or artifacts.