- Fewer wasted turns. The small model sometimes answered its first turn in prose ("no repository was provided") or called Claude Code's tools by name. Now a prose answer is asked again with a tool call required (on servers that allow it), shell commands that only look at files (
grep,cat,sed -n,ls,find) run as Otōto's own tools, and a search that finds nothing says why and which kinds of file the repository has. - Safe from a model server's stale cache. On 27 September our vLLM server's shared prompt cache held wrong data for part of Otōto's instructions, and every question that reused it went wrong until a restart. Each question now keeps its own cache on the server (
cache_salt), andototo doctorchecks whether a server's cache answers as a fresh request would, and says to restart it if not. - Running your own vLLM?
vllm/README.mdin this package has the flags we run, a chat template for coding clients (Qwen's own, plus what those clients send that Qwen's refuses) with a script that checks it, and what to do when the prompt cache goes bad. - Plugins, sealed. Compiled plugins are cached with a key of your own (
~/.config/ototo/cache.key, made on first use), so code put in the cache by anything else is compiled over, never run. The first start after upgrading compiles them once, in about a second. - Reporting: if your team collects Otōto's counts, they now carry your Claude account's e-mail, as Claude Code's own do, so the dashboard shows the two side by side per person without any setting.
user.email=inotlp_attributessends none, andototo doctorsays which address goes out. - For organisations:
ototo managed --installrun twice no longer loses the backup of the original settings, andototo doctorno longer gives managed users advice they cannot take.
source & further reading
ototo.dev — original article
Otōto 2026.15
Show HN: Ototo: A tool to offload code exploration to a smaller model
Otōto 2.11