hr-onboardlisted

# Onboard a repository ## What this does not do **It does not scan the codebase and describe it.** A scan produces a repository overview, and an overview fails the third condition of the inclusion test below: the agent can read the directory tree itself. Writing down what the code already says is a recurring bill for no return — carrying a context file costs over 20% more inference unconditionally, and context files do not generally improve success rates (arXiv:2602.11988, established). Ablating the overview category directly (Table 7, Appendix B) moved accuracy 68.12%->62.32% (p=0.15) on CTXBench and 54.36%->54.20% (p=0.73) on SWE-bench - **not significant either way**, on one model and LLM-generated files only. **This holds for a documented repository.** The same paper finds context files *do* act as effective overviews when the repo has no documentation. If this repo is thinly documented, an overview may earn its place - judge it, do not assume. It also does not duplicate `/init`. If the host agent has an init command, that is a fine way to get a starting `CLAUDE.md`. This procedure produces something different: the **non-discoverable** knowledge, found by hitting it. ## The inclusion test Nothing enters `AGENTS.md` unless all three hold: 1. **Instruction** — tells an agent to do or not do something, not what the project is 2. **Non-standard** — a competent agent would assume otherwise 3. **Non-obvious** — cannot be read off the code, `package.json`, `Makefile` or `