Skip to content
inferred from GitHub

darwin

Durable patch execution for Python agents. LibCST AST-gated speculative repair + fingerprint cache + Laplacian-DP crossfeed. Vendor-neutral. 100% on 50 strict real bugs. MIT.

Code Cerebral Valley

Core trick

Hashes tracebacks into fingerprints and replays cached LibCST AST transformers as deterministic patches with zero LLM calls.

Evidence

README matched "Opus 4.7 hackathon": Darwin was built for the Anthropic Opus 4.7 hackathon (late-apply track, 2026-04-20 submission). The primitive is vendor-neutral — the diagnose adapter accepts Claude Opus or Gemini Flash via env var, same flow, same AST gate, same cache. The repo is public, the benchmark is reproducible from a fres

Tags

dev toolsagents
Miles0sage/darwin★ 00 forksPythonMITCreated Apr 22, 2026 · last push Apr 25, 2026

README excerpt

# Darwin

**AST-level structural patching for failing Python agents.** Deterministic LibCST transformers keyed on traceback fingerprints. Bounded blast radius. Vendor-neutral. MIT.

> Not durable execution (Agentspan, DuraLang, Temporal). Not reasoning verification (Rubric). Not retry-with-backoff. Darwin emits a **surgical AST patch** on cache hit — zero LLM calls, deterministic, auditable.

[![demo](https://asciinema.org/a/vsXu6gddko4rZPgM.svg)](https://asciinema.org/a/vsXu6gddko4rZPgM)

```bash
git clone https://github.com/Miles0sage/darwin && cd darwin && python3 xrepo_proof.py
```

That runs the "same bug, three different repos, one cached transformer, zero LLM calls" proof in ~3 seconds. If it prints `CROSS-REPO TRANSFER: DETERMINISTIC, REAL, AUDITABLE` you have seen the primitive. Everything below is context for what you just watched.

## Numbers (reproducible from `benchmarks/`)

| Corpus | Healed | Notes | `results.json` |
|---|---|---|---|
| **v3 strict real bugs** | **50 / 50 (100%)** | 34 Gemini Flash + 16 Opus rescues | [`benchmarks/v3/results.json`](benchmarks/v3/results.json) |
| **v2 complex real bugs** | **131 / 171 (77%)** | Gemini rate-limited, Pro+Opus carried; $1.09 total LLM spend | [`benchmarks/v2/results.json`](benchmarks/v2/results.json) |
| v1 runnable real bugs | 17 / 18 (94%)* | *17/30 in raw JSON — 12 entries had `no_reproducer` (traceback-only) and were skipped | [`benchmarks/v1/results.json`](benchmarks/v1/results.json) |
| Controlled Opus vs Ge