Writing.
My write-ups on FareWell and LIQ Auctions are linked below.
Jev Router is now reasoning-router: six classifiers, three harnesses
Why I rewrote Jev Router as reasoning-router, which classifiers, harnesses, and models it supports, and how to move an existing Jev Router config over.
My anti-lock-in AI stack
How I separated models, harnesses, workflow and context so I can change AI providers without rebuilding my setup. Updated as the stack changes.
Jev Router: up to 42% fewer tokens, same solve rate
An OpenCode plugin that picks reasoning effort for every GPT-6 or Claude request. On SWE-bench Verified with GPT-6 Astra it solved 44 of 44 attempts, the same as fixed high effort, with 14% fewer output tokens, 20% fewer input tokens, and 9% less time.
Change reasoning effort, keep the prompt cache
OpenAI and Anthropic now document a way to raise or lower reasoning effort mid-conversation and keep the cached prefix. Here is how each one works, and where other providers stand.
Frontier coding models, measured over time
A living comparison of DeepSWE v1.1 scores, cost per task, and complementary coding evals. Official results alongside pending claims, with an interactive Pareto frontier.