big-pickle got stuck in an autoregressive loop during a brainstorming task

Description

Model: opencode/big-pickle
Trigger: Brainstorming/generative task producing lists of similar items
What happened:
The model entered an autoregressive loop during a brainstorming session, generating the same block of content repeatedly. The output consisted of a list of name suggestions with descriptions, and the model kept producing near-identical list items without recognising it was repeating. After several iterations, Chinese characters began appearing in the output mixed with English text.
(note these were apparently visible internally to the model and not by me on the screen)

Reproduction steps:

1. Start a conversation with some context (in this case, discussing a search engine project)
2. Ask the model to brainstorm a list of creative/abstract names (or any generative task requiring many similar items)
3. The model produces a long list, then continues generating additional items that are structurally identical to what it already produced
4. The loop persists until the user intervenes or the context becomes very long

What was observed:

• Output repeated the same content multiple times in a consistent pattern
• Chinese characters appeared in the output (the conversation was entirely in English)
• The model did not self-correct or recognise the repetition
• Both loops occurred during the same session with the same task type (brainstorming names)
• First loop was ~10x longer than the second, but both had the same repetitive structure
  Possible contributing factors:
• Long context window with prior similar content to pattern-match against
• Task structure (lists with repeated format: "name — description — assessment") creating a template the model fills repeatedly
• No natural stopping point in the generation

Plugins

none

OpenCode version

1.18.18

Steps to reproduce

1. Start a conversation with some context (in this case, discussing a search engine project)
2. Ask the model to brainstorm a list of creative/abstract names (or any generative task requiring many similar items)
3. The model produces a long list, then continues generating additional items that are structurally identical to what it already produced
4. The loop persists until the user intervenes or the context becomes very long

Screenshot and/or share link

Operating System

MacOS 14.8.5

Terminal

iTerm2