port2rust

An evidence-first Copilot agent that ports codebases to Rust. Built on Anvil patterns.

Overview

port2rust systematically ports a source repository to Rust with proof, not promises:

• SQL Verification Ledger — Every build, test, and lint result is INSERTed into SQLite. Evidence bundles are SELECTs, not prose.
• Adversarial Multi-Model Review — Up to 3 AI models (GPT, Gemini, Claude) attack every change, each trying to find bugs the others missed.
• Baseline Snapshots — System state captured before every change. Regressions are caught before you see the code.
• Risk Classification — Tasks are tagged 🟢/🟡/🔴. High-risk porting (auth, crypto, parsing) gets 3 adversarial reviewers.

Usage

# Port a repository
python scripts/orchestrate.py https://github.com/example/some-tool

# With explicit name
python scripts/orchestrate.py https://github.com/example/some-tool --name some-tool

# Resume after interruption
python scripts/orchestrate.py --resume

# Run specific phase
python scripts/orchestrate.py --phase porting --resume

With Copilot CLI

@port2rust port https://github.com/example/some-tool

Architecture

port2rust/
├── agents/
│   └── port2rust.agent.md    # Main agent (Anvil-style prompt)
├── .github/agents/           # Sub-agent prompts
│   ├── planner.md            # Planning agent (risk-aware)
│   ├── coder.md              # Coding agent (evidence-first)
│   └── reviewer.md           # Adversarial reviewer
├── scripts/
│   └── orchestrate.py        # Orchestration engine
├── plugin.json               # Plugin metadata
├── agent.yaml                # Agent configuration
└── mcp.json                  # Context7 MCP server

Runtime (created during execution)

├── code2port/                # Source repo submodule
├── src/                      # Ported Rust code
├── test/
│   ├── port_harness/         # Test harness for original code
│   ├── project_harness/      # Test harness for Rust port
│   └── test_cases/           # Shared test cases
├── .port2rust_ledger.db      # SQL verification ledger
└── .port2rust_state.json     # Orchestrator state

Workflow

Phase	Steps	What Happens
Setup	1–4	Branch → folders → submodule → build detect → baseline capture
Test Harness	5–7	Planner analyzes code → Anvil implements harness+tests → sync
Porting	8–11	Strategy → Anvil ports with adversarial review → completeness check
Validation	12–16	Rust test harness → run tests → fix failures → evidence bundles
Review	17–19	3-model adversarial review → fix → re-validate → converge

The Verification Loop (per task)

1. 📸 Baseline — snapshot build/test state into SQL
2. 🔨 Implement — Anvil agent writes code with built-in verification
3. ✓ Verify — build, test, diagnostics → INSERT results
4. ⚔️ Adversarial — 1-3 models attack the code (risk-driven)
5. ⚠️ Regress? — compare baseline vs after, fix if degraded
6. 📋 Evidence — bundle generated from SELECT, not prose
7. ✅ Commit — auto-commit with rollback command

Evidence Bundle Example

────────────────────────────────────────────────────────────
🔨 Port2Rust Evidence Bundle
Phase: porting | Task: port-3 | Risk: 🔴
────────────────────────────────────────────────────────────

  Baseline (before):
    [✓] build: cargo build succeeded
    [✓] tests: 12 passed

  Verification (after):
    [✓] build: cargo build succeeded
    [✓] tests: 14 passed (+2 new)

  Adversarial Review:
    [✓] review-gpt-5.3-codex: No issues
    [✓] review-gemini-3-pro: No issues
    [✓] review-claude-opus-4.6: No issues

  Regressions: None
  Confidence: High
────────────────────────────────────────────────────────────

Requirements

• Python 3.10+
• Git
• Rust toolchain (rustup, cargo)
• GitHub Copilot CLI with agent support

License

MIT — see LICENSE.