rimworld-learning-environment: v0.4.0

0.4.0 (2026-06-12)

Features

• add export_site_data.py website fold-in exporter (aa71f15)
• add export_site_data.py website fold-in exporter (2c27cc4)
• HuggingFace dataset integration (8a20c18)
• HuggingFace dataset integration (closes #45) (82911be)
• spread capture polish — model-labeled ticker, dashboard tick mirror (53995be)
• spread capture polish + post-spread harness fixes (6a84657)

Bug Fixes

• billed-cost reconciliation + no-pause speed keepalive (209b275)
• billed-cost reconciliation + no-pause speed keepalive (ce97a6c)
• stockpile overlap guard, semantic priority parsing, window preflight (6b33521)

v0.1.0 — RimWorld Learning Environment

RLE v0.1.0 — the first tagged release of the RimWorld Learning Environment: a multi-agent benchmark where 7 Felix Agent SDK role-specialized LLM agents manage a RimWorld colony. Multi-agent coordination under uncertainty — think FLE, but for a team that has to talk to each other.

Highlights

• 7-agent architecture — MapAnalyst (spatial reasoning, runs first) feeds a verified MAP_SUMMARY to ResourceManager, DefenseCommander, ResearchDirector, SocialOverseer, ConstructionPlanner, and MedicalOfficer, all coordinating through Felix CentralPost hub-spoke.
• 6 scenarios from Crashlanded (easy) to Ship Launch (extreme), each with victory/failure conditions and scoring-weight overrides.
• 10-metric weighted composite scoring with bootstrap confidence intervals, a pinned no-agent baseline (N=4, colony dead by day 8), and per-tick CSV export.
• Provider-agnostic via felix-agent-sdk 0.3.0 native async — LM Studio (local), OpenRouter, Anthropic, OpenAI, and the claude-code subscription provider.
• Live RIMAPI integration — async REST + SSE event stream feeding raids, deaths, and mental breaks into agent context.
• Helix phase adaptation, Docker headless mode, a React dashboard with 5 RLE widgets, and a post-run analysis + media toolkit (the /run-analysis skill).

First benchmark data

A 6-model spread (Crashlanded, 10 ticks, N=1 content-first — not statistically significant; top models advance to N=4): Grok 4.3 led on mean composite (0.831) and was the cheapest cloud model ($0.38). Every model repeatedly re-issued research targets that had just failed, despite explicit per-tick error feedback. Full table and caveats in the README.

Under the hood

458 tests, mypy strict, ruff clean. Versioning is now automated via release-please — 0.1.1+ will roll from conventional-commit prefixes.

🤖 Generated with Claude Code

rimworld-learning-environment: v0.3.0

0.3.0 (2026-06-11)

Features

• OBS production studio automation for benchmark capture (6af463d)
• OBS production studio automation for benchmark capture (8bc321d)

rimworld-learning-environment: v0.2.0

0.2.0 (2026-06-11)

Features

• reasoning-token costs, growing-zone overlap guard, auto-dismiss dialogs, camera director (#33, #34) (c48b408)
• reasoning-token costs, growing-zone overlap guard, auto-dismiss dialogs, camera director (#33, #34) (563f976)