The model that made up its own tool result

The console starts an agent CLI for each chat turn and hands it a set of MCP tools. One setting, meant to cut background traffic, made the turn start before the tool servers were ready.

The model had no tools. It did not say so. It wrote a tool call out as plain text, then wrote the result it expected to get back, and carried on as if both had happened.

Nothing ran. Nothing could have: a write needs an approval card, and there was no card. But anyone reading the transcript saw a confident answer resting on a call that never took place.

The relay now looks for tool-call markup in plain text and flags the turn as fabricated.

A model with no tools does not always admit it. Check what was called, not what the transcript says was called.