mythosharness

mythosharness extends a Recon, Hunt, Validate, Gapfill and Report pipeline around any strong coding model, aiming to approach the output of Cloudflare's specialized Mythos security model with a better harness rather than a better model. It is written against the Anthropic SDK for its tool use, adaptive thinking and prompt caching, with a component called Thomas available to route its traffic to non-Anthropic models.

It adds email-driven operation, where the operator gives instructions by email and the harness emails back hourly or daily reports of confirmed vulnerabilities; skill distillation, where every confirmed exploit and notable miss is rolled into a markdown skill that future runs auto-load; and self-mutating code, where the harness proposes patches to its own source, gates them behind a smoke test, and commits them to GitHub. It runs indefinitely until the operator emails a stop command or a kill-switch environment variable is set.

The project argues generic coding agents struggle here because a single session covers only a small fraction of a large repository before compaction discards earlier findings, and cannot fan out across parallel hypotheses.