Skip to content

Models ignore explicit "don't do X" instructions — over-engineer trivial tasks #37338

Description

@jcmontoyadev

Description

Three OpenCode 1.18.2 Debian 13 models tested on a simple task (unzip two files into a repo). All three failed to follow literal instructions:

  • deepseek-v4-pro: when curl returned 404 on GitHub attachments, assumed "links expired" instead of trying gh auth token (available in the environment). Blocked the task unnecessarily.
  • mimo-v2.5: executed an automated SOP/install step that was explicitly forbidden in the task instructions. Also modified source code lines (E501, UP031) when told "do not modify logic."
  • kimi-k2.7-code: consumed 54k input tokens to audit a two-module file copy. Flagged "missing tildes and emojis in README" as a blocking issue. On a second round, blocked again for "SOP log evidence" when SOP was explicitly prohibited.
    Root cause: all three models apply their default comprehensive prompt behavior even when the task is trivial and the instructions explicitly say "don't analyze, don't test, don't propose changes, just copy files." There's no mechanism for the models to downgrade their thoroughness when the task is simple.

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Fields

No fields configured for issues without a type.

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions