The ledger blocked a harmless ls, because the answer was already known
The guard only fires when the answer is already known, and its deny message is the fix. We have not measured whether it stopped the repeats.
After the 22 October terminal crash, the recovery list held ten sessions whose last recorded inputs were mostly one-word lines with no task in them, so the task had to be recovered from each session's transcript.
On 22 October, at 00:21, our terminal workspace crashed. The workspace is herdr, the tool we run coding sessions in, one per pane. A session is one conversation with the coding assistant. Afterwards we wrote down the last thing each session had been given, so we could resume the work.
The list holds ten sessions. Most of their last entries are one word: "ok", "go", "k", "do". A one-word message names no task. Whatever the session was doing lives somewhere else.
The ten session ids sit in eight working folders. A lane, our word for one parallel work stream, shows up under two ids in two of them: a terminal lane and a project folder. Counting by folder undercounts sessions, and counting by id overstates the number of folders. We report both numbers.
Eight of the ten sessions were last written between 00:15 and 00:21 on 22 October. The other two were last written at 23:56 the evening before. So the list covers a window of about twenty-five minutes, and the last lines are the last things typed or injected in that window.
A reader who wants the full state of the day has to go to the transcripts, not the list.
"ok go" and "k" are the last entries in one of the terminal lane's sessions. "do 1 2" appears in the same list. "ok" closes another. A reader who sees only the last line learns that someone was acknowledging something. The request that prompted it is gone from the list. A reader can tell from the words that a session was at a decision point, but not which decision.
The list also mixes what people typed with lines the tool injected. Some sessions show a compaction summary, which is text the assistant writes when it shortens a long conversation. Others show a notification instruction the tool added. The list does not mark which lines a person typed. So "last human input" is only a best guess per session.
Several of the lists include a compaction summary, the assistant's short record written when a long conversation is shortened. The summary text in these lists says the messages after it were kept verbatim, and that the summary was written without seeing them. So something the summary calls pending may already have happened in those later lines.
That is a second reason the last line is not the state. The summary is older than the messages after it, and the list shows only the newest ones.
A transcript is the full record of a session, stored as a file. Two of the ten entries carry a path to one: a project session and one terminal lane session. The others carry no path. The summaries of those two sessions say the full transcript holds details the summary leaves out.
So the recovery path is the transcript, but only for the sessions that left a pointer. The rest need a search of the transcript folder by session id.
A separate entry in our herdr ledger, dated 25 July, describes the same family of failure. A message pasted into a pane's input line was never submitted. The tool reported it as delivered, and the session sat idle until someone noticed. The fix recorded there is to send messages with a command that types, submits and confirms on screen, and to read the pane back before trusting it.
That entry is from July. It is not the cause of the October crash. We put it here because it is the same kind of loss: a last line that looked like progress and was not.
Komentarze
Brak komentarzy.
Zaloguj się aby skomentować.