inferred from GitHub
darwin hackathon
Darwin — Claude Code that heals its own crashes. Submission for Cerebral Valley x Anthropic Opus 4.7 hackathon (Apr 21-27 2026).
Core trick
Fingerprints agent errors and stores verified fix recipes on a shared blackboard, so later agents apply stored fixes with zero LLM calls.
Evidence
README matched "Built with Opus 4.7": Built for Cerebral Valley × Anthropic "Built with Opus 4.7" hackathon (Apr 21–27, 2026). Solo, Germany.
README excerpt
# Darwin > **TL;DR** — Darwin is a failure memory layer for Claude Code agents. First agent hits a crash, Opus 4.7 diagnoses, fix goes on shared blackboard. Every subsequent agent reuses the stored fix with zero LLM calls. Fresh clone → `python3 darwin_harness.py` → 90-second self-healing demo. `python3 benchmark.py` → 20 agents healed, 1 LLM call, 19 cache hits. **AI coding agents should not make the same mistake twice.** Darwin is a failure memory layer for coding agents. It captures errors, fingerprints them, retrieves proven fix recipes, applies them, verifies the result, and records the outcome. The fleet gets smarter with every crash. ## Why AI agents repeatedly hit the same dependency, import, test, and environment failures. Today those fixes disappear into chat logs and terminal output. Darwin turns them into reusable repair memory that grows stronger with every run. ## Demo 90 seconds, narrated. 5 scenes, zero human input. ▶ https://asciinema.org/a/MOdDP1NrJGkFNrpU 1. Baseline — agent polls API v1, works fine 2. Real-world failure — upstream ships v2, agent crashes with `KeyError` 3. Darwin fingerprints the failure, checks blackboard, diagnoses via Opus 4.7 4. Self-verify → apply patch → broadcast fix + diff to the fleet blackboard 5. **Second agent hits the same failure → blackboard serves the stored fix → zero LLM call** ## Benchmark 20 agents, same recurring failure class. ▶ https://asciinema.org/a/QC5NO4kvnuQO2aaU ``` Fleet size: 20 Healed: