AppSprout-dev · jkbennitt · Jun 11, 2026 · Jun 11, 2026
diff --git a/.github/workflows/release-please.yml b/.github/workflows/release-please.yml
@@ -0,0 +1,22 @@
+name: release-please
+
+on:
+  push:
+    branches: [master]
+
+permissions:
+  contents: write
+  pull-requests: write
+
+jobs:
+  release-please:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: googleapis/release-please-action@v4
+        with:
+          config-file: release-please-config.json
+          manifest-file: .release-please-manifest.json
+          # Uses the default GITHUB_TOKEN. Note: PRs opened by this action with
+          # GITHUB_TOKEN do not re-trigger ci.yml. If you want CI to run on the
+          # release PR, create a fine-grained PAT (contents+PRs write) as the
+          # RELEASE_PLEASE_TOKEN secret and add `token: ${{ secrets.RELEASE_PLEASE_TOKEN }}`.
diff --git a/.release-please-manifest.json b/.release-please-manifest.json
@@ -0,0 +1,3 @@
+{
+  ".": "0.1.0"
+}
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -0,0 +1,35 @@
+# Changelog
+
+## 0.1.0 (2026-06-11)
+
+First tagged release. RLE is a multi-agent benchmark where 7 Felix Agent SDK
+role-specialized LLM agents manage a RimWorld colony — multi-agent coordination
+under uncertainty.
+
+### Features
+
+* **7-agent architecture** — MapAnalyst (spatial, runs first) + ResourceManager,
+  DefenseCommander, ResearchDirector, SocialOverseer, ConstructionPlanner,
+  MedicalOfficer, coordinating through Felix CentralPost hub-spoke.
+* **Deterministic spatial analysis** — MAP_SUMMARY with verified build/farm/stockpile
+  coordinates injected into every agent; agents may not invent coordinates.
+* **6 scenarios** — Crashlanded, First Winter, Toxic Fallout, Raid Defense,
+  Plague Response, Ship Launch, each with victory/failure conditions and weight overrides.
+* **10-metric weighted composite scoring** with bootstrap confidence intervals and
+  per-tick CSV export.
+* **Provider-agnostic** via felix-agent-sdk 0.3.0 native async — local (LM Studio),
+  OpenRouter, Anthropic, OpenAI, and the claude-code subscription provider.
+* **RIMAPI integration** — async REST client + SSE event stream (raids, deaths,
+  mental breaks injected into agent context).
+* **Pinned no-agent baseline** — N=4, colony dead by day 8, committed as a provenance
+  sidecar; every agent run is measured against it.
+* **Helix phase adaptation** — exploration → analysis → synthesis as the colony progresses.
+* **Docker headless mode**, React dashboard with 5 RLE widgets, W&B/HuggingFace loggers.
+* **Post-run analysis & media toolkit** — cross-model leaderboard, story mining, footage
+  indexing, and the `/run-analysis` skill.
+
+### Notes
+
+* First public benchmark data: a 6-model spread (Crashlanded, N=1 content-first).
+  Grok 4.3 led on mean composite; full caveats in the README.
+* 458 tests, mypy strict, ruff clean.
diff --git a/CLAUDE.md b/CLAUDE.md
@@ -377,9 +377,14 @@ src/rle/
 scripts/
 ├── run_scenario.py        # Single scenario CLI (auto-loads save, unforbids items)
 ├── run_benchmark.py       # Full benchmark suite CLI (--docker, --smoke-test, --runs)
+├── run_spread_n1.sh       # N=1 multi-model spread runner (sequential, continue-on-error)
 ├── compare_benchmarks.py  # Paired statistical comparison of benchmark runs
+├── analyze_spread.py      # Cross-model leaderboard vs baseline + failure taxonomy
 ├── visualize_results.py   # Matplotlib CSV plotter
-└── serve_dashboard.py     # CORS-enabled file server for dashboard
+├── serve_dashboard.py     # CORS-enabled file server for dashboard
+├── replay_ticks.py        # Replay a run's tick snapshots into the dashboard (for capture)
+├── record_dashboard.mjs   # Headless Playwright dashboard recording (run with node)
+└── capture_run_media.sh   # During-run game-window stills + tick snapshots
 docker/
 ├── Dockerfile             # HeadlessRim + Xvfb (debian:bookworm-slim)
 ├── docker-compose.yml     # Volume mounts for game files, mods, saves

diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
@@ -6,7 +6,7 @@
 git clone https://github.com/AppSprout-dev/RLE.git
 cd RLE
 uv sync --extra dev
-pytest  # should pass 326+ tests
+pytest  # should pass 458+ tests
 ```
 
 Requires Python 3.14+ and [uv](https://docs.astral.sh/uv/).
@@ -16,11 +16,24 @@ Requires Python 3.14+ and [uv](https://docs.astral.sh/uv/).
 1. Create a branch from `master`
 2. Make changes
 3. Run `pytest`, `ruff check src/ tests/ scripts/`, and `mypy src/`
-4. Commit with a descriptive message
-5. Open a PR against `master`
+4. Commit using [Conventional Commits](https://www.conventionalcommits.org/) (see below)
+5. Open a PR against `master` — **regular merge commits, never squash** (preserves per-commit history)
 
 CI runs lint + type check + tests + smoke test on every push/PR.
 
+### Commit messages drive versioning
+
+Versioning is automated by [release-please](https://github.com/googleapis/release-please): commit prefixes determine the version bump and changelog entry. Don't hand-edit `version` in `pyproject.toml`.
+
+| Prefix | Example | Effect (pre-1.0) |
+|--------|---------|------------------|
+| `feat:` | `feat: add toxic fallout scenario` | minor bump |
+| `fix:` | `fix: correct food_security divisor` | patch bump |
+| `feat!:` / `BREAKING CHANGE:` | breaking API change | minor bump (stays <1.0) |
+| `docs:` `chore:` `refactor:` `test:` `ci:` `perf:` | tidy-ups | no release |
+
+release-please keeps an open "release PR" accumulating changes; merging it tags the version and publishes a GitHub release with generated notes.
+
 ## Running locally
 
 ### Prerequisites
@@ -130,7 +143,7 @@ docker/                    # HeadlessRim Dockerfile, compose, entrypoint
 
 ## Key dependencies
 
-- [felix-agent-sdk](https://github.com/AppSprout-dev/felix-agent-sdk) >= 0.2.1
+- [felix-agent-sdk](https://github.com/AppSprout-dev/felix-agent-sdk) >= 0.3.0
 - [RIMAPI](https://github.com/IlyaChichkov/RIMAPI) C# mod
 - httpx, pydantic >= 2.0, pyyaml
 - Optional: wandb, huggingface-hub (`uv sync --extra tracking`)

diff --git a/README.md b/README.md
@@ -86,7 +86,7 @@ PROVIDER_BASE_URL=http://localhost:1234/v1
 # OpenRouter (cloud)
 OPENAI_API_KEY=sk-or-v1-your-key-here
 PROVIDER=openai
-MODEL=nvidia/nemotron-3-super-120b-a12b:free
+MODEL=nvidia/nemotron-3-super-120b-a12b
 PROVIDER_BASE_URL=https://openrouter.ai/api/v1
 ```
 
@@ -149,25 +149,32 @@ cd ../rimapi-dashboard && bun run start
 
 ```bash
 # Full benchmark (mock game state, real LLM)
-python scripts/run_benchmark.py --dry-run --ticks 10
+python scripts/run_benchmark.py --smoke-test --ticks 10
 
 # List scenarios
 python scripts/run_scenario.py --list
 
 # Visualize CSV results
 python scripts/visualize_results.py results/ --all
+
+# Cross-model leaderboard vs the pinned baseline (after a spread)
+python scripts/analyze_spread.py --spread-dir results/spread
 ```
 
 ## Benchmark Results
 
-Live game, 10 ticks, Crashlanded scenario:
+**6-model spread** — Crashlanded, 10 ticks, seed 42. **N=1, content-first — not statistically significant; top models advance to N=4.** Ranked by mean composite across the run (endpoint scores are single-event noisy):
 
-| Model | Composite | Survival | Food | Mood | Efficiency |
-|-------|----------|----------|------|------|------------|
-| Nemotron 30B (OpenRouter) | **0.808** | 1.000 | 0.950 | 0.441 | 0.754 |
-| Nemotron 30B (OpenRouter) | **0.794** | 1.000 | 0.850 | 0.510 | 0.894 |
+| # | Model | Mean composite | Cost |
+|---|-------|---------------|------|
+| 1 | Grok 4.3 | **0.831** | $0.38 |
+| 2 | Opus 4.8 | 0.826 | subscription |
+| 3 | Gemini 3.5 Flash | 0.825 | $2.41 |
+| 4 | Fable 5 | 0.803 | subscription |
+| 5 | GPT-5.5 | 0.722 | $3.79 |
+| 6 | DeepSeek V4 Pro | 0.704 | $0.55 |
 
-All colonists alive, buildings on solid ground, no water placement.
+Measured against a pinned no-agent baseline (N=4, colony dead by day 8). Notable finding: all six models repeatedly re-issued research targets that had just failed, despite explicit per-tick error feedback.
 
 ## Scenarios
 

diff --git a/release-please-config.json b/release-please-config.json
@@ -0,0 +1,14 @@
+{
+  "$schema": "https://raw.githubusercontent.com/googleapis/release-please/main/schemas/config.json",
+  "packages": {
+    ".": {
+      "release-type": "python",
+      "package-name": "rimworld-learning-environment",
+      "changelog-path": "CHANGELOG.md",
+      "bump-minor-pre-major": true,
+      "extra-files": [
+        "src/rle/__init__.py"
+      ]
+    }
+  }
+}
diff --git a/src/rle/__init__.py b/src/rle/__init__.py
@@ -3,7 +3,7 @@
 Multi-agent benchmark where Felix Agent SDK agents play RimWorld.
 """
 
-__version__ = "0.1.0"
+__version__ = "0.1.0"  # x-release-please-version
 
 from rle.agents import (
     Action,