Alpha · Open source · Local-first

Pitwall Agent — endurance stint planner

Compare fuel, tyre, driver-rotation and pit-stop strategies before your endurance race.

A local-first, terminal-based strategy tool for sim racers, team captains, and race-day operators. Free and open-source. Optional local AI routes plain language; deterministic tools own the race facts.

Download alpha.2 for Windows See example output View source

MIT licensed · Python 3.11–3.14 · Core planning works offline after installation · No account, API key, or data upload.

What problem it solves

Planning an endurance race is overwhelming — fuel reserves, tyre changes, driver rotations, Safety Car timing. Pitwall Agent turns those variables into a clear, auditable plan you can take to race day. Every number comes from a deterministic engine you can inspect, trace, and rerun.

Who it's for

Four people stand around every pit wall. Pitwall Agent is built for all of them.

Complete beginner

Never planned an endurance race?

Plain-English onboarding and guided setup help you build a race configuration without knowing the terminology up front.

Sim-racing team captain

Auditable strategy, not black-box advice

CSV telemetry in, auditable strategy out. See every assumption, confidence level, and warning before your first stint begins.

Race-day operator

Scan-ready pit sheets

Fuel, tyres, drivers, timing, trigger cards, and uncertainty bounds in one scannable pre-race reference.

Technical contributor

Fork it, extend it, prove it

Typed Python, deterministic engine, 85%+ test coverage enforced in CI, JSON output, documented architecture.

How the numbers stay auditable

Deterministic engine — never AI-guessed

All fuel, tyre, and driver maths is deterministic Python code in engine/. You can read every line and re-run it for the same result.

AI routes; local tools answer

An optional local Ollama model can route plain language into an allowlisted tool. Final race facts are rendered locally from a successful deterministic result; failed or unknown calls cannot ground an answer.

Evidence Levels A / B / C

Each result states its input source, evidence level, confidence, assumptions, and warnings.

Local-first, no telemetry upload

Your CSVs, race config, and reports stay in a visible .pitwall/ folder on your own machine. Nothing is uploaded anywhere.

Honest alpha statement. Pitwall Agent is alpha software. All bundled examples are synthetic (Evidence Level C), and no real-session case study has been published. The Ollama integration is real, but there is no published real-model conformance benchmark; direct deterministic commands remain the recommended operational path.

What it does today

Fuel & stint planning

Stint lengths, fuel loads, and exact fuel additions per stop.

Tyre life & changes

Tyre sets, age at stint end, and change planning per stop.

Driver rotation

Stable driver identities and configured drive-time rule checks.

Strategy comparison

Conservative / Balanced / Fuel Save ranked side by side.

P10 / P90 uncertainty

For projected laps, P10 is the fewer-laps case and P90 the more-laps case. Raw lap-time percentiles reverse because fewer seconds is faster.

Telemetry CSV ingest

Column mapping, outlier checks, and evidence calibration from your laps.

Safety Car what-if

One declared pre-race scenario with explicit multipliers. Pre-race only.

Strategy trigger cards

Pre-agreed HOLD / RECONSIDER thresholds to take to the pit wall.

Markdown pit sheet

One scannable page to print or keep open on race day.

Machine-readable JSON

--json on the commands you script or pipe into your own tools.

Optional local AI

Ollama on loopback only, for plain-language routing. Final race facts remain deterministic, and everything works without AI.

No live data, no lock-in

No live timing, no proprietary APIs, no accounts, no tracking.

What it looks like in the terminal

Captured output from the published alpha.2 CLI running the bundled synthetic 6-hour example in core-only mode.

Pitwall compare output ranks Conservative, Balanced, and Fuel Save strategies, with uncertainty and deterministic trigger cards. — Synthetic example data (Evidence Level C) — not a performance claim.

Open the full stint-by-stint plan screenshot — fuel, tyre, driver, pit-stop, and trigger-card output from the same deterministic run.

The pit sheet you take to race day

Excerpt of an exported Markdown pit sheet. Every figure is traceable to an input and an evidence level.

# Pit sheet: 6h Endurance — GT3
⚠️ PRE-RACE PLAN ONLY — Not live race control

Strategy: Conservative (recommended)
Evidence: C | Confidence: Low | Source: Manual assumptions
Predicted laps: 174 | Pit stops: 6 | Fuel used: 504.6 L
P10: 174.0 laps | Median: 174.0 | P90: 175.0 laps

TRIGGER CARDS
FUEL_BURN_HIGH  [HOLD]  Reconsider above 2.987 L/lap
PACE_SLOW       [HOLD]  Reconsider above 120.96 seconds/lap

STINT 1  Laps 1–28  | Alex Pro     | Fuel start 87.0 L | Tyre set 1
STINT 2  Laps 29–56 | Marco Bronze | Fuel add 81.2 L   | Pit 87.5 s

Install

GitHub-only alpha. Pitwall Agent is not published on PyPI. Download the versioned alpha.2 source ZIP or wheel from GitHub Releases. The core install never downloads Ollama or an AI model.

Windows — easiest

# 1. Install Python 3.11–3.14 from python.org
# 2. Download the source ZIP linked below
# 3. Choose Extract all
# 4. Double-click run.bat

Download the alpha.2 source ZIP. The launcher creates a private environment, runs a health check, and opens Pitwall.

Linux / macOS

git clone https://github.com/Dabi-init/endurance-stint-planner.git
cd endurance-stint-planner
python3 -m venv .venv
. .venv/bin/activate
python -m pip install -e .
pitwall doctor
pitwall init --guided
pitwall compare

Package size and releases

The published alpha.2 wheel is approximately 92 KB and is attached to the versioned GitHub release. SHA-256 checksums are published beside the wheel and source archive.

Measured disk footprint

In a clean Windows/Python 3.14 audit, Pitwall's installed files used about 738 KB, the complete private virtual environment used about 27.4 MB, and an extracted source copy plus its launcher-created environment used about 28.1 MB. The launcher's isolated two-run test retained 0 bytes in its pip-cache directory. A Git clone also retains Git history; totals vary by Python and operating system. App imports are capped at 10 MiB and 50,000 rows per CSV, with 100 MiB total telemetry storage. Reports and validations are capped at 2 MiB each, 25 MiB combined, and 500 files.

Developer install

git clone https://github.com/Dabi-init/endurance-stint-planner
cd endurance-stint-planner
# Windows PowerShell
py -3.14 -m venv .venv
.\.venv\Scripts\python.exe -m pip install -e ".[dev]"

# macOS / Linux (use instead)
python3 -m venv .venv
.venv/bin/python -m pip install -e ".[dev]"

Try the 5-minute demo

1
Source ZIP → run.bat
Download, extract, and launch on Windows.

2
automatic checks
The launcher runs doctor and first-run onboarding.

3
/compare
See three ranked demo strategies.

4
/plan Conservative
Inspect every stint.

5
/export demo
Write a new pit sheet.

6
/setup
Then replace the demo with your race.

The first comparison uses the bundled Evidence C demo. Run /setup before relying on your own race inputs. No API key, cloud account, or data upload.

Choose your model storage. Every deterministic feature works with no AI. Core-only uses 0 GB of model storage; no Ollama install or model opt-in is required.

Core only: keep AI off, or clear a previous selection with .\.venv\Scripts\python.exe -m pitwall model off.
Provisional first try: qwen3:8b, approximately 5.2 GB of model storage.
Smaller unverified candidate: qwen3:4b, approximately 2.5 GB of model storage.

# Provisional first model to try
ollama pull qwen3:8b
.\.venv\Scripts\python.exe -m pitwall model use qwen3:8b

# Smaller unverified candidate
ollama pull qwen3:4b
.\.venv\Scripts\python.exe -m pitwall model use qwen3:4b

pitwall model recommend runs a read-only deterministic core self-check and prints these Ollama-only choices. It never contacts Ollama, downloads a model, creates a workspace, or changes configuration. Core-only is the verified operational path; both model candidates remain unverified until Pitwall publishes a real-model tool-calling conformance benchmark.

Pitwall never downloads a model automatically. Model sizes exclude Ollama and its caches. Ollama's Windows installation requires at least 4 GB, so the 8B setup needs roughly 9.2 GB or more. Remove an opted-in model with .\.venv\Scripts\python.exe -m pitwall model off followed by ollama rm qwen3:8b or ollama rm qwen3:4b; uninstall Ollama separately if you no longer need it.

Alpha status and limitations

Read this before you rely on it for a real event.

Alpha software — not production-validated.
Not validated yet. The accuracy log scores one assumption pack against 2026 post-session results. It is unofficial and not affiliated with Formula 1. Bundled planner examples stay synthetic.
Safety Car support is a pre-race what-if only — it is not live race control.
No live timing, no competitor prediction, no proprietary telemetry APIs.
Ollama is optional and never auto-downloaded — every deterministic feature works with no AI at all. Model recommendations are provisional until a real-model conformance benchmark is published.
Evidence Level C means the numbers came from preset or synthetic assumptions, not measured data.
Sporting regulations are your responsibility. Pitwall Agent checks the rules you configure, not any official rulebook.

Community and feedback

This project gets better the moment real racers tell us what broke.

⭐ Star on GitHub 🐛 Report an issue 💡 Request a feature 📊 Contribute anonymised session data 🤝 Join the contributor community

Feedback checklist

Could you install it without help?
Did you understand the recommendation it gave you?
Did the pit sheet help you on race day?
What was confusing, missing, or wrong?
Would you share an anonymised session so we can validate against real data?