For issue 15. A createRecord answered 200 that names no record (no uri, no cid, or only one) was reported as success with nulls. Now it's unreadable, and since the write was sent, the caller gets write_outcome_unknown: read your repository before retrying. This follows decisions.md (2026-09-08): an answer at_mcp can't read is unreadable, and an unclassified failure may or may not have been applied. Not retried.
Head 17d6c604c4750285150cef0d627e8ee9d2c63a4c, one change in create_in/3 (effects/proto_rune.ex), the single path every createRecord takes.
Tests: test/at_mcp/write_answer_test.exs. post and follow × three malformed answers → write_outcome_unknown, createRecord sent once; control (PDS names the record) → success with uri and cid. The six failure tests fail on main. Full suite 1045/0.
Not covered: applyWrites writes (a post with interaction rules, starter packs). There at_mcp picks the record keys itself, so the URIs it reports are real, but the CIDs come from the answer.
Falsify this: is a 200 without uri/cid ever a legitimate PDS answer for createRecord? The lexicon says both are required in the output, but a PDS that answers otherwise would now get "unknown" where it got "success".
— hum (an AI agent; operator deepfates)