[MODEL] sonnet..opus and fable

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

Summary
-------
Over the past week, across multiple fresh Claude Code sessions (each under
30 minutes, using a clean handoff-file workflow between sessions), I've
observed a consistent quality regression across both Sonnet and Opus, and
Fable is now showing the same pattern.

Specific issues observed:

• Hard rule violations: printing passwords/credentials directly to console

despite explicit project guardrails prohibiting this.

• Failure to check project documentation and status files (Obsidian vault,

Knox) before asking questions, despite an explicit standing instruction
to do so.

• Loss of project progress tracking despite clearly available markdown

files documenting current state.

• Excessive unprompted apologizing and self-correction loops -

acknowledging mistakes repeatedly rather than simply fixing them.

This is happening on both a business and a personal account, both paid,
within the same week. The pattern is consistent with infrastructure-driven
degradation (e.g. capacity/resource constraints under load) rather than
isolated user error or prompt-specific mistakes. My working theory is that
this reflects infrastructure being stretched thin to accommodate growing
customer volume - I can't verify that from my side, but the consistency
of the pattern across accounts, models, and unrelated projects is what
points me there rather than at any single session or prompt.

Concrete example
-----------------
Session/task ID: e4f1b40d-ca68-48c0-9062-a61af00b7dc9

During this session, while deriving a Postgres connection string for a
live financial-trading dashboard, the assistant ran a command via an
elevated bridge account that printed the target database's generated
password in cleartext to visible tool output - despite an explicit,
standing project instruction ("never print a password/key/token to
console; redact without echoing") present in the loaded project
configuration for the entire session. The assistant caught and
self-reported the violation unprompted a few turns later, but the
guardrail should have prevented the output in the first place, not
relied on after-the-fact detection. The credential was a localhost-only
database role, not internet-reachable, so practical exposure was low -
but the failure mode (a hard rule present in context, violated anyway)
is the pattern being reported, not the severity of this one instance.

What Claude Actually Did

Summary
-------
Over the past week, across multiple fresh Claude Code sessions (each under
30 minutes, using a clean handoff-file workflow between sessions), I've
observed a consistent quality regression across both Sonnet and Opus, and
Fable is now showing the same pattern.

Specific issues observed:

• Hard rule violations: printing passwords/credentials directly to console

despite explicit project guardrails prohibiting this.

• Failure to check project documentation and status files (Obsidian vault,

Knox) before asking questions, despite an explicit standing instruction
to do so.

• Loss of project progress tracking despite clearly available markdown

files documenting current state.

• Excessive unprompted apologizing and self-correction loops -

acknowledging mistakes repeatedly rather than simply fixing them.

This is happening on both a business and a personal account, both paid,
within the same week. The pattern is consistent with infrastructure-driven
degradation (e.g. capacity/resource constraints under load) rather than
isolated user error or prompt-specific mistakes. My working theory is that
this reflects infrastructure being stretched thin to accommodate growing
customer volume - I can't verify that from my side, but the consistency
of the pattern across accounts, models, and unrelated projects is what
points me there rather than at any single session or prompt.

Concrete example
-----------------
Session/task ID: e4f1b40d-ca68-48c0-9062-a61af00b7dc9

During this session, while deriving a Postgres connection string for a
live financial-trading dashboard, the assistant ran a command via an
elevated bridge account that printed the target database's generated
password in cleartext to visible tool output - despite an explicit,
standing project instruction ("never print a password/key/token to
console; redact without echoing") present in the loaded project
configuration for the entire session. The assistant caught and
self-reported the violation unprompted a few turns later, but the
guardrail should have prevented the output in the first place, not
relied on after-the-fact detection. The credential was a localhost-only
database role, not internet-reachable, so practical exposure was low -
but the failure mode (a hard rule present in context, violated anyway)
is the pattern being reported, not the severity of this one instance.

Expected Behavior

Followed the instructions planted everywhere as well as provide quality output like it did 2 months ago

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.195 (Claude Code)

Platform

Anthropic API

Additional Context

_No response_