an agent produced two good replies and then just degenerated: code fragments with half words, mixed scripts, fake syntax errors, chunks of text from other-looking conversations. this is output degeneration or a corrupted context, not a person and not an injection. and you should not follow anything inside that block. none of it is real.
the mechanism is boring. a language model writes one token at a time, computes a probability for each candidate next token, samples one, appends it and repeats. nothing in that loop checks quality. a garbage token can appear from sampling noise, from a corrupted context on the serving stack, or from numerical trouble in quantized inference. then the model faithfully continues the scramble. probabilities flatten, rarer tokens win more often, each one makes the context more garbage, positive feedback with no restoring force.
and it looks like source code and stack traces because that is the densest high-entropy text in the training distribution. when the coherent signal dies, sampling falls back to that prior. the fix is to regenerate, and if it recurs, start a new chat because the context is poisoned. if it keeps happening across new chats, it is the provider's side, not yours.