Pitwall Agent v0.4.0-alpha.2

Pitwall Agent v0.4.0-alpha.2

Hardening release focused on safe failure, bounded storage, trustworthy local
Ollama routing, reproducible installation, and clear model choices.

Hardened

• Corrupt race or configuration files fail closed with recovery guidance.
• JSON failures return valid error payloads and nonzero exit status.
• Telemetry, history, model responses, and generated reports have explicit size limits.
• App-owned control files use atomic writes; reports remain exclusive-create.
• Unknown or failed model tool calls cannot ground a displayed race answer.
• The Ollama client ignores environment proxies and rejects redirects away from
  its validated loopback endpoint.
• Python 3.14 joins the tested CI matrix.

Changed

• Ollama is a bounded local intent router; final race facts are rendered locally.
• pitwall model recommend performs a read-only core self-check. Core-only is
  the verified zero-model-storage path; qwen3:8b is a provisional first try,
  and qwen3:4b is a smaller unverified candidate.
• Windows one-click startup caches installation and exposes /setup,
  /compare, /plan, and /export.
• Synthetic telemetry stays at Evidence Level C after renaming or line-ending
  conversion.
• Narrow terminal output is more compact and actionable.

Verification

• 226 automated tests pass with the 85% coverage gate.
• Python 3.11–3.14, wheel installation, Ruff/compile, Windows launcher, and
  CodeQL checks pass.
• Clean installed-wheel and repeat-launcher smoke tests pass.

Known limitations

• Alpha software, not production-validated.
• All bundled examples are synthetic (Evidence Level C).
• No live timing, competitor data, or official regulation database.
• No published real-model conformance benchmark; direct deterministic commands
  remain the verified operational path.
• Model recommendations are advisory only and never contact Ollama, download a
  model, create a workspace, or change configuration.

Install

Windows users can download the versioned source ZIP, extract it, and
double-click run.bat:

https://github.com/Dabi-init/endurance-stint-planner/archive/refs/tags/v0.4.0-alpha.2.zip

The wheel, source distribution, and SHA256SUMS.txt are attached to this
release. Pitwall Agent is not published on PyPI.

v0.4.0-alpha.1 — Trigger Cards, Guided Setup & Launch

Pitwall Agent v0.4.0-alpha.1

⚠️ Alpha software. All case studies use synthetic data (Evidence Level C). No real-session validation has been published.

What's New

🃏 Strategy Trigger Cards (Issue #13)

Deterministic pre-race trigger cards for the recommended strategy. Each card shows:

• Metric, unit, and threshold — fuel burn, pace, tyre life, Safety Car window
• Status: HOLD / RECONSIDER — crosses threshold = reconsider your plan
• Source — manual assumptions or telemetry ranges
• Affected decision — exactly what changes if the trigger fires
• Cards appear in pit sheet Markdown and --json output

🧭 Guided Beginner Setup (pitwall init --guided)

Step-by-step interactive race configuration for users with no endurance-racing knowledge:

• Explains every field in plain English with units and safe ranges
• Shows whether values came from presets or user input
• Validates impossible constraints (fuel burn > tank, negative values)
• Produces clear recovery messages

👋 Beginner Welcome (pitwall welcome)

New command explaining all key concepts: stints, fuel reserve, tyre life, P10/P90, Evidence Levels, parallel/sequential pit service, Safety Car scenarios, and what the tool cannot know.

📋 Pit Sheet Improvements

• P10 (pessimistic) / P90 (optimistic) uncertainty with clear labels
• Evidence Level and confidence with plain-English explanation
• Uncertainty source: manual assumptions vs telemetry-calibrated
• Tyre age at stint end
• Per-stop pit duration
• Warnings, infeasibilities, and assumptions lists
• Explicit pre-race-only header on every export

🔬 Validation Report (pitwall validate)

Compare your planned strategy against user-supplied actual results:

• Laps, stops, fuel burn, stint lengths
• Clear provenance: user-supplied comparison, not an independent validation claim
• Evidence stays Level C for synthetic/user-supplied data

🔒 Security & Safety Hardening

• Driver names/labels from telemetry CSV are redacted before appearing in model-visible tool output (shown as Driver_1, Driver_2) — your real driver names stay in your pit sheet
• Report overwrite protection moved to the workspace layer
• Safety Car output includes explicit pre-race what-if disclaimer on all surfaces
• Manual vs telemetry input source shown in all comparison output

📖 Documentation

• README glossary: Stint, Fuel Reserve, Tyre Life, P10/P90, Evidence Levels, Trigger Cards, and more
• docs/PROJECT_HANDOFF.md — durable architecture and agent handoff
• docs/LAUNCH.md — launch material and community copy
• AGENTS.md — guidance for future AI agents
• Public landing page: https://dabi-init.github.io/endurance-stint-planner/

Upgrade

pip install --upgrade pitwall-agent
# or from source:
git pull && pip install -e .

Verification

pitwall doctor
pitwall welcome
pitwall compare
pitwall plan
pitwall export --name race_day
pitwall validate --actual-laps 72 --actual-stops 3

Known Limits (alpha)

• No real-session validation — all case studies are synthetic (Evidence C)
• Safety Car is pre-race what-if only
• No live timing, no competitor data, no proprietary APIs
• Ollama AI is optional — all planning works without it

What's Next (0.5 roadmap)

• Real anonymised session case studies
• Documented simulator CSV adapter
• Read-only MCP server
• Versioned import/export schema

Full changelog: CHANGELOG.md

---

Pitwall Agent: AI may explain the plan, but it must never invent, change, or own the race mathematics.

Pitwall Agent 0.3.0-alpha.1

Pitwall Agent 0.3.0-alpha.1

The endurance stint-planner prototype is now local-first terminal software.

Highlights

• installable pitwall command and interactive race strategist;
• editable per-event car, race, driver, service, and regulation configuration;
• deterministic plan, comparison, telemetry audit, driver-rule, Safety Car, and pit-sheet tools;
• optional local Ollama tool calling with no-model fallback;
• path, prompt-injection, model-loop, unknown-tool, remote-host, and overwrite guards;
• exact fuel/service accounting, stable driver identities, seeded uncertainty, and evidence levels;
• Windows PowerShell launcher plus JSON automation;
• 56 tests, 86.8% core coverage, Python 3.11–3.13 CI, CodeQL, and clean wheel installation.

Start on Windows

git clone https://github.com/Dabi-init/endurance-stint-planner.git
Set-Location .\endurance-stint-planner
powershell -ExecutionPolicy Bypass -File .\run.ps1

Scope

This is alpha pre-race decision support. It does not claim live timing, live race control, traffic/weather modelling, championship rule packs, or real-session validation.