[nuriyoo]

CONTRIBUTING.md says where a new test case goes when an existing test already sends its request, and AGENTS.md has an agent look for that test before writing a new function.

What changed

CONTRIBUTING.md. Item 3 of Pull requests gains the rule: the case is a row of the existing test's table, a refusal and its successful counterpart on one URL are one table, and a new function is kept for a state or an assertion no test has yet.

AGENTS.md. One bullet under Tests: the search to run before writing a new test function, then the rule in one clause with a pointer to item 3, as the bullet on titles and descriptions points to items 5 and 6.

.claude/agents/behaviour-reviewer.md (item 4, "Tests did not sprawl"), step 6 of .claude/skills/backlot-loop/SKILL.md and item 7 of .claude/skills/backlot-debloat/SKILL.md already ask this of a pull request. This pull request leaves those three files as they are.

Why this is what the real API does

No vendor behaviour changes. The rule was written in the reviewer agent's checklist and two skills, and not in CONTRIBUTING.md or AGENTS.md, where a contributor and their agent look before writing a test. Reviews here often ask a contributor to fold a test into an existing table after it was written as a new function. Stated in those two files, it can be followed the first time.

Verification

• Tests — PYTHONWARNDEFAULTENCODING=1 pytest -rs on a uv sync --all-extras --locked install, plus llama-index-readers-hubspot<0.6 as CONTRIBUTING.md says: 8810 passed, 1 skipped.
• Optional surfaces that ran rather than skipped — every extra in the lock and the HubSpot reader; the skip is tests/test_mcp.py:102, mcp-atlassian, with no Docker daemon running.
• Lint — ruff check . && ruff format --check .: All checks passed, 193 files already formatted.

At 831e94d9, on main at f420eda2.

pytest tests/test_docs.py                    20 passed
python scripts/prose_check.py origin/main    no findings

• A test fails without this change — a change to two Markdown files; tests/test_docs.py checks their links and passes.
• A new endpoint serving corpus content is ACL-scoped, proved for both an admin and a scoped token — no endpoint changes.
• Comments state what was measured, not the history of the fix — no code comments change.