Add isolated Pi bridge integration and @mmprotest/pi-villani extension

.github/workflows/build-pi-villani-runtime-release.yml

@@ -0,0 +1,77 @@
+name: Build Pi Villani Runtime Release
+permissions:
+  contents: write
+on:
+  workflow_dispatch:
+  push:
+    tags: ["pi-villani-runtime-v*"]
+jobs:
+  build:
+    strategy:
+      matrix:
+        include:
+          - key: win32-x64
+            os: windows-latest
+            ext: zip
+          - key: darwin-arm64
+            os: macos-14
+            ext: tar.gz
+          - key: darwin-x64
+            os: macos-26
+            ext: tar.gz
+          - key: linux-x64
+            os: ubuntu-latest
+            ext: tar.gz
+    runs-on: ${{ matrix.os }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with: {python-version: '3.11'}
+      - run: python -m pip install -e . pyinstaller
+      - run: pyinstaller packaging/villani_runtime.spec --distpath dist-runtime/villani-code
+      - run: python packaging/smoke_test_runtime.py dist-runtime/villani-code/villani-code${{ runner.os == 'Windows' && '.exe' || '' }}
+      - shell: bash
+        run: |
+          name="villani-runtime-v0.1.5-${{ matrix.key }}.${{ matrix.ext }}"
+          cd dist-runtime
+          if [[ "${{ matrix.ext }}" == "zip" ]]; then 7z a ../$name villani-code; else tar -czf ../$name villani-code; fi
+          cd ..
+          python - "$name" > "${name}.sha256" <<'PY'
+          import hashlib
+          import pathlib
+          import sys
+
+          path = pathlib.Path(sys.argv[1])
+          digest = hashlib.sha256(path.read_bytes()).hexdigest()
+          print(f"{digest}  {path.name}")
+          PY
+          mkdir -p smoke-asset
+          python - "$name" smoke-asset <<'PY'
+          import sys, tarfile, zipfile
+          from pathlib import Path
+          archive, dest = Path(sys.argv[1]), Path(sys.argv[2])
+          if archive.suffix == '.zip':
+              with zipfile.ZipFile(archive) as zf:
+                  zf.extractall(dest)
+          else:
+              with tarfile.open(archive, 'r:gz') as tf:
+                  tf.extractall(dest)
+          PY
+          python packaging/smoke_test_runtime.py smoke-asset/villani-code/villani-code${{ runner.os == 'Windows' && '.exe' || '' }}
+      - uses: actions/upload-artifact@v4
+        with:
+          name: runtime-${{ matrix.key }}
+          path: "villani-runtime-v0.1.5-${{ matrix.key }}.${{ matrix.ext }}*"
+  release:
+    needs: build
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          path: artifacts
+      - run: find artifacts -name '*.sha256' -exec cat {} \; > checksums.txt
+      - uses: softprops/action-gh-release@v2
+        with:
+          files: |
+            artifacts/**/villani-runtime-v0.1.5-*
+            checksums.txt

.gitignore

@@ -1,3 +1,11 @@
+# Keep the Pi runtime PyInstaller spec tracked
+!packaging/villani_runtime.spec
+
+# Node
+node_modules/
+npm-debug.log*
+*.tgz
+
 __pycache__/
 *.pyc
 .venv/

README.md

@@ -133,6 +133,21 @@ Autonomous pass:
 villani-code --villani-mode --base-url http://127.0.0.1:1234 --model your-model --repo /path/to/repo
 ```
 
+## Use Villani Code in Pi
+
+# @mmprotest/pi-villani
+Install with:
+```bash
+pi install npm:@mmprotest/pi-villani
+```
+Provides `/villani <task>`.
+
+Runtime version: `v0.1.5`.
+
+## Pi Usage
+
+Run Villani with `/villani <task>`. Active runs are still cleaned up automatically on process/session cancellation or error cleanup.
+
 ## Reports
 
 - [Villani Code Terminal-Bench 2.0 Qwen3.6 27B Report](docs/Villani_Code_Terminal_Bench_2_Qwen27B_Report.pdf)

integrations/pi-villani/README.md

@@ -0,0 +1,86 @@
+# @mmprotest/pi-villani
+Install with:
+```bash
+pi install npm:@mmprotest/pi-villani
+```
+Provides `/villani <task>`.
+
+Runtime version: `v0.1.5`.
+
+## Usage
+
+Run Villani with `/villani <task>`. Active runs are still cleaned up automatically on process/session cancellation or error cleanup.
+
+# Villani Code
+
+**Flagship coding-agent performance from small local models.**
+
+Villani Code is a local-first coding-agent runtime designed to make smaller open models do real repository work: navigate files, run commands, make patches, survive verification, and keep working through messy terminal environments.
+
+The thesis is simple: small models do not just need better weights. They need a better runtime.
+
+## Terminal-Bench 2.0: Qwen3.6 27B full-suite run
+
+Villani Code achieved a **196/445 lower-bound score** on the full Terminal-Bench 2.0 suite using **Qwen3.6 27B**.
+
+That is **44.0%** across **89 tasks** with **5 attempts per task**.
+
+### Headline result
+
+| System | Model | Terminal-Bench 2.0 accuracy |
+|---|---|---:|
+| Codex CLI | GPT-5-Codex | 44.3% |
+| **Villani Code** | **Qwen3.6 27B** | **44.0%** |
+| Mini-SWE-Agent | GPT-5-Codex | 41.3% |
+| Claude Code | Claude Sonnet 4.5 | 40.1% |
+| Dakou Agent | Qwen 3 Coder 480B | 27.2% |
+| little-coder | Qwen3.6-35B-A3B | 24.6% |
+| Bash Agent | TermiGen-32B | 19.3% |
+| little-coder | Qwen3.5-9B | 9.2% |
+
+## Qwen3.5 9B same-model runtime comparison
+
+Villani Code was also tested against Claude Code using the same model: **Qwen3.5 9B**.
+
+Same model. Same tasks. Different agent runtime.
+
+Villani Code won.
+
+
+| Runner | Score | Success rate |
+|---|---:|---:|
+| **Villani Code + Qwen3.5 9B** | **38/60** | **63.3%** |
+| Claude Code + Qwen3.5 9B | 26/60 | 43.3% |
+
+**Villani Code delivered a 46% relative performance improvement over Claude Code.**
+
+This comparison covers **12 overlapping Terminal-Bench tasks**, with **5 runs per task**, for **60 runs per agent**.
+
+Villani Code won **6 tasks**, tied **6 tasks**, and lost **0**.
+
+## What Villani Code is
+
+Villani Code is a terminal-first coding agent for:
+
+- bounded bug fixes
+- repo navigation and localization
+- command-driven debugging
+- test-guided patching
+- local inference setups
+- privacy-sensitive codebases
+- smaller open model backends
+
+It is built for the environment where most coding agents start to fall apart: smaller models, hard verification, constrained context, terminal noise, failed commands, and real repositories.
+
+## What changed in the upgraded runtime
+
+The latest Villani Code upgrade includes:
+
+- new execution loop
+- better local model integration
+- cleaner tool handling
+- improved failure recovery
+- task-scoped memory system
+- better state tracking across long-running coding tasks
+
+The benchmark comparison evaluates the upgraded runtime as a whole.

integrations/pi-villani/docs/release.md

@@ -0,0 +1,5 @@
+# Release checklist
+1. Publish matching GitHub runtime release `pi-villani-runtime-v0.1.5`.
+2. Verify assets and checksums.
+3. Publish `@mmprotest/pi-villani`.
+Install: `pi install npm:@mmprotest/pi-villani`.