Attributed metadata history

harnessrouter: a Chinese translation of the agent harness router, and why the docs matter more than the codeaiopensourceagentsllm

The original HarnessRouter/harnessrouter repo has 2993 stars and a README written in English. The Chinese version, yangshun2005/harnessrouter-cn, was published today with zero stars. That asymmetry is the whole point of this article. If you have ever tried to evaluate an agent infrastructure project from a non-English-speaking region, you know the failure mode: you skim the README, you understand maybe 70% of it, you miss the sentence that says "do not add --user", and you spend an afternoon debugging something that was documented in plain sight. Translation is not glamorous work. It is, however, the difference between a project being adoptable and being a curiosity. What HarnessRouter actually is Strip away the positioning language and here is the mechanism: HarnessRouter is an infrastructure layer that turns existing agent harnesses — Codex, Claude Code, Hermes, PI, DSH, and others — into pluggable backends behind a single API. The problem it addresses is real. If you are building an agent product, you probably do not want to write separate integration code for every harness you might want to support. You want to run a task, get a result, and swap the underlying harness when cost or latency demands it. HarnessRouter implements the Unified Harness Protocol (UHP), an open standard, and exposes an OpenAI Responses-compatible API on top. It handles persistent sessions, streaming progress, files, artifacts, cancellation, and structured failure reporting. The official one-liner describes it as the self-hosted, Apache-2.0 edition of a unified interface for agent harnesses. Your keys, your infrastructure. That last part matters. There is no bundled model and no trial key. You bring a provider API key. The community edition runs on your own hardware. The Chinese version: what it adds The harnessrouter-cn repo is a translation of the original project documentation. The README itself states the translation date as 2026-10-14 and links to README_en.md as the source. It carries the same license, Apache-2.0, and the original project's star count is listed as 2993. The structure is worth noting because it is not a pure translation. The repo splits its content into two distinct layers: A Chinese quick-start guide at the top, explicitly labeled as supplementary material and not part of the original documentation. Its stated purpose is to let you get the project running without reading the full README. The translated README body below, which aims to stay faithful to the original text. That separation is a small design decision with outsized consequences. Machine-translated technical documentation tends to blur the line between "what the original says" and "what a translator inferred." By marking the quick-start as non-original, the repo makes it possible to trust the translated body as a reference while still using the quick-start as a shortcut. If the translation is wrong somewhere, you can file an issue or a PR — the repo says so explicitly. Getting it running The quick-start path in the Chinese repo is short. You need Docker, roughly 4 GB of disk space, and a provider API key. Start the container: docker run -d --name harnessrouter \ -p 127.0.0.1:3000:3000 \ -v harnessrouter:/data \ harnessrouter/harnessrouter Then watch the logs, because first boot installs the enabled harness CLIs: docker logs -f harnessrouter Wait for this line: [harnessrouter] ready on :3000 At that point you can stop following the logs with Ctrl+C. The container keeps running. Open http://localhost:3000 and log in with the default credentials: username harnessrouter, password harnessrouter. Then go to Bring Your Own Key in the sidebar, click Add Integration, pick a provider, give it a name, and paste the API key. Models supported by that provider become available in the console. That is the whole onboarding flow. No account creation, no hosted dependency. The gotchas th

This record contains metadata and attribution only. It does not host, download, unpack, render, or execute malware samples.