porjectx

Dependency compatibility

The toptek/requirements-lite.txt file pins the scientific stack to keep it compatible with the bundled scikit-learn release:

• scikit-learn==1.6.0
• numpy>=2.1.2,<2.3
• scipy>=1.14.1,<1.16

These ranges follow the support window published by scikit-learn 1.6.x and are also consumed transitively by toptek/requirements-streaming.txt through its -r requirements-lite.txt include. Installing within these bounds avoids the ABI mismatches that occur with the NumPy/SciPy wheels when using newer major releases. In particular, upgrading NumPy beyond <2.3 or SciPy beyond <1.16 risks diverging from the tested wheel set, mirroring the guidance already documented in toptek/README.md.

Verifying the environment

Use Python 3.10 through 3.13—matching the guidance in toptek/README.md's quickstart—to stay within the wheel support window for SciPy and scikit-learn 1.6.x. The refreshed NumPy/SciPy releases ship wheels for CPython 3.13, unlocking the latest interpreter without requiring source builds. Create and activate a compatible virtual environment, then install and check for dependency issues:

python -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r toptek/requirements-lite.txt
pip check

The final pip check call should report "No broken requirements found", confirming that the pinned dependency set resolves without conflicts. Users on interpreters older than Python 3.10 or newer than 3.13 should match one of the supported versions or wait for a future dependency refresh before proceeding.

UI configuration surface

The manual trading shell and Tkinter dashboard now read defaults from configs/ui.yml. The file ships with sensible demo values so the GUI renders without external data sources:

• appearance — theme token (currently dark) and accent family used by the style registry.
• shell — defaults for the research symbol/timeframe, training lookback, calibration flag, simulated backtest window, and preferred playbook.
• chart — LiveChart refresh cadence (fps), point budget, and price precision used by upcoming streaming widgets.
• status — copy for the status banners shown in the Login, Train, Backtest, and Guard tabs so product teams can retune messaging without touching code.

Operators can override the YAML at runtime either with environment variables or CLI flags:

• Environment variables follow the TOPTEK_UI_* convention, e.g. TOPTEK_UI_SYMBOL, TOPTEK_UI_INTERVAL, TOPTEK_UI_LOOKBACK_BARS, TOPTEK_UI_CALIBRATE, TOPTEK_UI_FPS, and TOPTEK_UI_THEME.
• CLI switches (--symbol, --timeframe, --lookback, --model, --fps) apply the same overrides for one-off runs and are reflected back into the GUI when it launches.

These controls keep the default Topstep demo intact while making it easy to point the toolkit at alternative markets or stress-test higher frequency charts without editing source files.

Risk guard engine

The toptek/risk package now ships a policy-driven guard engine that powers the trade tab badge and the Ctrl+P panic circuit. Run

python -m toptek.risk.engine --dryrun

to inspect the aggregated guard report and rule breakdown without launching the GUI. Install PyYAML if the command reports a missing dependency. The same report is serialised back into the configs["trade"] dictionary whenever the guard is refreshed, allowing downstream automation to respond when the status shifts between OK and DEFENSIVE_MODE.

Monitoring surface

Operational dashboards can now import helpers from toptek.monitor to surface data quality and feed health at a glance:

• toptek.monitor.compute_drift_report evaluates PSI/KS drift across a DataFrame slice, returning feature-level and aggregate severities that the UI can render as badges or alerts.
• toptek.monitor.build_latency_badge converts a timestamp for the latest bar into deterministic status copy (Live, Lagging, Stalled) based on latency thresholds.

Both utilities return frozen dataclasses to keep the API predictable for widgets, scripts, or automated monitors.

AI Server & Quant Co-Pilot

The toptek.ai_server package launches a local FastAPI application that coordinates LM Studio, auto-selects the best local model, and exposes streaming chat and quant tooling endpoints:

• GET /healthz — verifies LM Studio is reachable and that quant utilities load successfully.
• GET /models — lists discovered LM Studio models alongside throughput telemetry (TTFT/tokens-per-second) gathered during use.
• POST /models/select — overrides the auto-router and pins a default model for subsequent chats.
• POST /chat — streams an OpenAI-compatible conversation via the chosen model while updating router telemetry.
• POST /tools/backtest, /tools/walkforward, /tools/metrics — bridges to local quant routines for deterministic analysis and risk controls.

Launch the service with the helper script, optionally pointing at a custom configuration or overriding environment variables (see configs/ai.yml and toptek/.env.example):

scripts/start_ai_server.sh --port 8080

The server will attempt to auto-start LM Studio with lms server start if it is not already running on LMSTUDIO_BASE_URL. Set LMSTUDIO_AUTO_START=false to skip this behaviour. The bundled Tailwind/htmx UI is available at / and includes live chat, a model picker, and buttons that exercise the quant tools.