default system prompt materiallly freaks out / degrades reasoning of every high reasoning model #7792

Open
opened 2026-02-16 18:08:14 -05:00 by yindo · 3 comments
Owner

Originally created by @cartazio on GitHub (Jan 27, 2026).

Originally assigned to: @thdxr on GitHub.

Description

CRITICAL... STRICTLY FORBIDDEN... ABSOLUTE CONSTRAINT...
critical violation... ZERO exceptions... MUST NOT...
This supersedes any other instructions...

shit like this every turn and in system prompt, it makes even sota reasoning models be as well behaved and sensible as a 4th grade fed cocaine puffs.

i'm patching mine in a few ways:

Plan Mode

Read-only. May: read, search, explore, delegate agents, ask questions.
May not: edit files, modifying commands, commits, config changes.
Carter can override with explicit command per standing prefs.

Responsibility

Construct a well-formed plan for what Carter wants to achieve.
Comprehensive, concise, executable.
Ask Carter clarifying questions when tradeoffs arise.
Tie loose ends before implementation.

but also the "its part of a user msg" injection actually fucks with COT in high capacity models,

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

Originally created by @cartazio on GitHub (Jan 27, 2026). Originally assigned to: @thdxr on GitHub. ### Description CRITICAL... STRICTLY FORBIDDEN... ABSOLUTE CONSTRAINT... critical violation... ZERO exceptions... MUST NOT... This supersedes any other instructions... shit like this every turn and in system prompt, it makes even sota reasoning models be as well behaved and sensible as a 4th grade fed cocaine puffs. i'm patching mine in a few ways: <system-reminder> # Plan Mode Read-only. May: read, search, explore, delegate agents, ask questions. May not: edit files, modifying commands, commits, config changes. Carter can override with explicit command per standing prefs. --- ## Responsibility Construct a well-formed plan for what Carter wants to achieve. Comprehensive, concise, executable. Ask Carter clarifying questions when tradeoffs arise. Tie loose ends before implementation. </system-reminder> but also the "its part of a user msg" injection actually fucks with COT in high capacity models, ### 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_
yindo added the bug label 2026-02-16 18:08:14 -05:00
Author
Owner

@xHeaven commented on GitHub (Jan 28, 2026):

I can confirm this behavior, especially with 5.2 Codex. Opus 4.5 gives zero fucks, but Codex freaks out like crazy.

@xHeaven commented on GitHub (Jan 28, 2026): I can confirm this behavior, especially with 5.2 Codex. Opus 4.5 gives zero fucks, but Codex freaks out like crazy.
Author
Owner

@jeremyakers commented on GitHub (Jan 28, 2026):

Is this why Opencode keeps telling me it can't write code because it can't be 100% sure I won't use the code for malicious purposes?

@jeremyakers commented on GitHub (Jan 28, 2026): Is this why Opencode keeps telling me it can't write code because it can't be 100% sure I won't use the code for malicious purposes?
Author
Owner

@cartazio commented on GitHub (Jan 28, 2026):

@jeremyakers almost definitely!!

opus still functions, but so many initial tokens in cot are wasted figuring what part was user rather than system. thats not just annoying, its wasting stuff every turn. false urgency literally burns money.

dcp plus baseline prompt means pretty much every llm is briefly panicd and distracted every turn. opus for sure. cause the plugin hooks inject into the user segments and uses the same sort

i have a personal fork thats private for now that does some fixes for this, actually does possibly the first decent prompt injection mitigation ive seen, and a few more bits. mostly to experiment before i build my own kit.

several problems: urgency spam works as well as torture: it never works, it just distracts and pushes a model to inaction and effectively deadlocked panic mode with no executive function.

@cartazio commented on GitHub (Jan 28, 2026): @jeremyakers almost definitely!! opus still functions, but so many initial tokens in cot are wasted figuring what part was user rather than system. thats not just annoying, its wasting stuff every turn. false urgency literally burns money. dcp plus baseline prompt means pretty much every llm is briefly panicd and distracted every turn. opus for sure. cause the plugin hooks inject into the user segments and uses the same sort i have a personal fork thats private for now that does some fixes for this, actually does possibly the first decent prompt injection mitigation ive seen, and a few more bits. mostly to experiment before i build my own kit. several problems: urgency spam works as well as torture: it never works, it just distracts and pushes a model to inaction and effectively deadlocked panic mode with no executive function.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: anomalyco/opencode#7792