[opus 4.1 + opus plan mode] - intelligence variation between two claude code instances

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

Other unexpected behavior

What You Asked Claude to Do

i was resuming a session, and asked for claude to read the claude file and the plan we were following the day before. audit the particular area of the codebase that we were working on to verify progress (this also serves to gain context) and then update the plan.

What Claude Actually Did

went off on a tangent, about state management when we were doing singleton conversions. the analysis was incorrect. the audit was incorrect. the update to the plan was incorrect.

Expected Behavior

exactly what i instructed it to do.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

_No response_

Claude Model

Opus

Relevant Conversation

Impact

Medium - Extra work to undo changes

Claude Code Version

1.0.106

Platform

Anthropic API

Additional Context

this happens sometimes.

two claude code terminals running simultaneously, resuming previous days work (new day new session 8 hour sleep gap). one is performing brilliantly, and the other just terribly. going off on a tangent and lacking context retention.

it almost feels like the model develops different personas for inference and somewhere in there is the eager persona, the over engineer, the dumb (almost like adhd, all over the place with context and temperature) persona (showing up a lot recently), the quality engineer. I could go on. this volatility is very problematic and needs to be streamlined.