wndaasa/newbie-harness is a Windows desktop AI agent that handles the OpenAI and Anthropic APIs plus local LLMs such as Ollama, LM Studio, llama.cpp, and vLLM on a single screen. It supports file editing, shell, git, MCP, and reading HWP documents. The October 25, 2026 commit aa0ad823d89eda451c77a3b70f2894e9cf834657 fixed two things at once. The problem where the token gauge appeared empty after turning the app back on and opening a conversation, and the problem where the Anthropic cache has a default 5-minute lifetime, so long conversations were re-read from the beginning every time.
On the gauge side, it newly added storedUsage() to app/main/controller.ts, which reads the last measured usage (cached, input, output) left in the session file and restores that value when no turn is in progress. In other words, it replaced the existing behavior of showing only the value while a turn was running with the session file value. On the cache side, cacheMark() in src/providers/anthropic.ts was changed so that the default cache marker carries a 1-hour lifetime (ttl 1h). Because Anthropic, unlike OpenAI, does not apply the cache automatically, the code attaches the cache_control marker itself.
The commit message states that the 5-minute cache disappeared between reading the answer and writing the next message, so a 490,000-token conversation was re-read from the beginning, and the code comment explains that it took close to 50 seconds until the first response. The same comment notes that the cache write price is twice the input price (1.25 times for 5 minutes) but the read price stays the same. The comment in the test file, based on Anthropic's rule that tokens read from the cache are not counted toward the per-minute input limit (ITPM), says that caching conversation history reduces not only cost but hitting the limit itself.
However, the effect of this commit applies only to the Anthropic path. Local inference paths such as Ollama or llama.cpp have no such concept as a cache lifetime, so the gauge-side change only covers the flow of opening a conversation after restarting the app. The tests check only that the stored measured value reappears on the gauge and the ttl value of the cache marker, and do not include any measurement showing that latency or cost actually went down. The size of the change is 76 added lines and 12 deleted lines, combining app/main/controller.ts, src/providers/anthropic.ts, and 2 test files. The 1-hour cache is an option that doubles the write price, so in environments where short conversations are frequent it is more reasonable to revert to NEWBIE_CACHE_TTL=5m. It is a substantive improvement for those who repeat long conversations, but there is still no basis for declaring it an improvement for environments centered on short conversations.