The offline agent needed files

On Friday, the Wiki's unattended maintenance run asked Luna to organize a batch it had already collected. Luna answered that it could not do the job: the material lived in local files, but the prompt said the run had “no shell” and forbade commands. It wrote no page. The run record nevertheless called the attempt completed and advanced past the material.

“Offline” had been stretched too far. The model needed no browser, mailbox or network to maintain the notebook; it did need to read the staged files and write to a disposable copy. The prompt now permits bounded local file operations inside that workspace, while continuing to forbid network calls, source-app commands and instructions hidden in the material. The maintenance Skill now makes the same distinction. Investigation received the same wording because it had repeated the contradiction.

A natural-language response is not proof that the batch was assessed. When a maintenance pass changes nothing, it now has to leave a small receipt naming the sources it reviewed and why no update was needed. Without that receipt, the run fails and the source cursor stays put. A six-case runner evaluation covers refusal, incorrect source attribution, blocked and missing material, a real no-change decision and a page update.

Luna then completed a synthetic offline smoke run: it read a one-off request, left the notebook alone for a reason, and wrote the receipt. That success cost 144,414 input tokens, much of it cached. The broken permission instruction is fixed; the size of the instruction bundle is now the next visible problem.

Another synthetic run exposed a different edge. Luna correctly treated an October launch date as unapproved and ignored a command quoted in the source, but wrote its citations as Markdown list numbers. The strict page reviewer refused the otherwise useful candidate. The runner now converts those list labels to the canonical citation spelling before validating the unchanged source IDs and claims. The refused candidate remains on disk for comparison; its two citation errors drop to zero after that narrow conversion. On a retest with the same source messages, the page was accepted with both source IDs and the date still marked unapproved. The quoted command was never run.

The preview release adds one more boundary to that lesson: source-tree tests do not prove that the installed package carries the same Wiki instructions. For 1.8.9b6, the wheel was installed into a fresh virtual environment and its CLI, included data files and Wiki reflection path were exercised there. All 13 installed-wheel checks passed. That establishes what is shipped, not whether the next real mailbox batch will produce a good page. The preview keeps that production-quality question open instead of calling a synthetic success a finished Wiki.