libaie/onboard-dsh-projects

README excerpt

onboard-dsh-projects English 简体中文 A DeepSeek Harness skill that lets one AI session drive many code repositories — without their contexts bleeding into each other. The scariest failure mode in AI-assisted coding isn't a weak model — it's crossed context. Debug a bug across three repos and repo A's instructions leak into repo B's code; you edit the wrong branch, or write somewhere you shouldn't. onboard-dsh-projects puts every repository in its own isolated entry agent , then adds an optional controller to orchestrate the whole effort. Why it exists Pain What this skill does 😵 Context pollution One isolated entry subagent per repo: instructions, index, evidence, and edits all stay in their own context 🔀 Repo and baseline drift Verifies root, branch, HEAD, and dirty state before every run; drift blocks, it never silently proceeds 💸 Model and task mismatch Three-tier model routing: routine work on v4-flash , engineering and high-stakes work on v4-pro 🔁 Repeating the same mistakes A bounded experience index: proven strategies are reused, deterministic failures reject the same mechanism 🧠 Controller memory bloat The controller keeps only cross-project contracts and dispatch state; …

Category