Walkieware: turn off model thinking for visible speed, kidlisp stop button master
A desktop replay of the slow wwGihap edit through the same relay took 23.7 s over three rounds, and two thirds of all output tokens were DeepSeek's hidden thinking: 11.8 s and 6.3 s before the first visible line, while headers took 0.6 s and code streamed at about 320 tokens/s. Thinking also counts against max_tokens, which is why the phone run hit the output limit. Walkieware now asks for reasoning {effort:"none"}. AcServer forwards a reasoning object when given one; the relay validates OpenRouter's documented shapes (effort, max_tokens, enabled, exclude) and passes them upstream. The deployed relay still drops the field, so the phone benefits after the next lith deploy. The pending row's stop control is kidlisp.com's red disc with a white rounded square. run.sh picks only the physical paired iPhone, since a booted simulator also reports as an iPhone. Builds 70 and 71 recorded in checks.json. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>