Hermes Loop Master
Tvingar fram specifikation, RÖD/GRÖN bevis och adversariell diff-granskning innan kod levereras
Testad · Fungerar
Vad den gör
En coding-discipline loop that makes the agent write a checkable execution contract, implement one slice at a time, prove behavior with RED/GREEN plus positive, negative, preservation and failure paths, then review its own diff against ten fake-done patterns before declaring work finished. Utlöses vid multi-step, production-facing, security-sensitive or cross-session coding tasks. Levereras med en executable harness that scores the loop artifacts, so the process is machine-checked rather than self-reported.
Testrapport
Fixed the same duration-parser bug twice on identical repos, once free-hand and once under the loop. Both arms produced a green test suite, but the skill's mandated failure-path check exposed that the obvious regex fix silently returns 0 for 'banana' and '', 3600 for the typo '1h30', and 300 for '-5m' — a probe of ten inputs against both final versions diverged on five, with the no-skill version returning a wrong number where the skill version raises. The bundled harness_check.py is real and runs: it scored the loop artifacts 32/34, flagged two unchecked Done When boxes I had actually left open, correctly passed the good-loop and critical-loop fixtures and rejected the bad-loop fixture, and the repo's 36 unit tests pass. Costs three extra artifact files and roughly double the tool calls, so it earns its keep on risky work, not on typos. Note it is branded for Nous Research's Hermes Agent and its installer defaults to ~/.hermes, so Claude Code users must pass --target as in the command above.
Testad: 2026-07-30 · Claude Code 2.x (agent harness)
Installation
git clone https://github.com/s0heyl/hermes-loop-master.git mkdir -p ~/.claude/skills cd hermes-loop-master && bash install.sh --target ~/.claude/skills/hermes-loop-master
Kommandon och exempelprompter
/hermes-loop-masterTvingar fram specifikation, RÖD/GRÖN bevis och adversariell diff-granskning innan kod levereras
Skills triggas av vanliga förfrågningar — inga kommandon att memorera. Efter installationen aktiverar prompter som dessa skillen (på engelska):
Fix this auth bypass in production, carefully, with proof it actually worksRefactor the payment webhook handler, this will span a few sessionsContinue yesterday's migration work under the RED/GREEN evidence loop