m/market labResearch, with a record.
Source on GitHub ↗

Current capabilities and support

This page describes the source edition and the scope of existing verification. Configured CI jobs are not evidence that an unpushed change passed on another OS.

Available workflows

  • Desktop, CLI and Python access to local simulated accounts and historical market data.
  • Controlled manual and model experiments across several symbols (Desktop or Platform API).
  • Human and Rule experiments, Controlled experiments driven by your own model file, and saved Agent revisions using a built-in template or trusted imported Python source.
  • Dependency preparation, selected Agent interpreters, installed SDK and reusable environments.
  • Prepared historical data with clock-limited visibility; Agent and Live policy work can continue while the financial owner processes events and commands.
  • Limit-order acceptance, cash/share reservation, partial fills, cancellation and account history.
  • Model IO and local user tools in the Agent process; platform tool requests cross IPC.
  • Recorded-operation Replay that recalculates financial outcomes without rerunning a model. Platform.replay_run() verifies finished Controlled runs, including runs on price-bar files, by preparing the original file again; Live and Agent runs use their operation log.
  • Native ITCH decoding/order-book rebuilding and a separate loopback network-feed simulator.

The current architecture is documented here. The native engine and the simulated account order book have different responsibilities; see native architecture.

Support and validation matrix

Area Current evidence / limitation
macOS arm64 Desktop Existing local acceptance on Python 3.9.6 / PySide6 6.8.3; rendering and demo were checked during source-release preparation
Ubuntu 24.04 / Python 3.12 CI builds native components, runs source tests, formatting/lint and first-run CLI commands
macOS / Ubuntu / Windows, Python 3.9 and 3.12 CI configured for lifecycle, installed SDK/interpreter and official offline environment preparation; new preparation jobs await actual remote results
Full Windows App Not accepted; native build and Desktop acceptance still require Windows execution
Full Linux Desktop Enabled offscreen suites are now configured in CI; actual remote results and native window/rendering acceptance remain pending
App Python Source edition targets Python 3.9–3.12; the main source workflow covers 3.12, not every version on every OS
Agent Python Provisioning policy accepts native CPython 3.9–3.14 requests; this is not whole-App or dependency certification on all those versions
Model providers Built-in adapter exists; automated integration uses scripted/local providers, not paid live-provider acceptance
OS credential storage Optional macOS keychain; session-only credentials remain available; no general non-macOS persistent-store claim

See how each check runs. Job configuration states what CI is intended to run; inspect the actual run before claiming that a particular commit passed.

Boundaries

This is a local simulation platform, with no broker or real exchange order submission. Historical public depth is not hidden liquidity or a consolidated multi-venue book. A recording made from price bars has a synthetic book built from assumptions. Multi-day bar files work in Controlled and Live mode; Live mode skips the hours between sessions. Trading uses simulated cash with limit and market orders; short selling is off unless the experiment allows it, and then follows a fixed margin rule. Several symbols are available to Manual, Rule and model-file experiments in Controlled mode and to Manual and Agent experiments in Live mode; news is not merged across symbols. Leverage, multi-agent evaluation and an overall agent leaderboard are outside this release. Agent evaluation compares saved runs with intervals and tests, in the Desktop's Compare runs page or from Python; it computes no overall score.

Agent code is trusted local code unless its experiment sets an OS sandbox (SDK guide). Pause closes capability/delivery gates and freezes experimental time; it does not suspend arbitrary user CPU instructions. Provider context limits still apply. There is no automatic conversation summary/truncation. There is no hard-real-time deadline guarantee. Native input, transport, scheduling, financial processing and output all have costs; prepared data does not make those costs disappear.

Packaged Desktop installers, signing, automatic updates and full additional-OS acceptance are future work. Historical data is not redistributed; the demo uses project-generated synthetic data.

Section index · All documentation

Maintained source: docs/user-guide/support.md