[MODEL] Opus 4.8 appears to respond from nonexistent conversation context after long thinking

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 asked Claude Code with Claude Opus 4.8 to reason about an Obsidian workflow design.

Specifically, I wanted advice on how to integrate project/task management with Obsidian daily notes while minimizing user friction and keeping project/task state easy for both the user and agent to read.

In a separate context, I also provided the path to an Obsidian daily note template and an actual daily note file, then explained that I had completed a project-related task on June 16 and was concerned that managing tasks only in daily notes would make project tasks hard to find later.

What Claude Actually Did

After approximately 10 minutes of thinking, Claude did not answer the current prompt.

Instead, it produced a response that appeared to continue from a nonexistent previous conversation. The response referred to things like being thanked or praised, “good discussion,” “no need to thank me,” and “this was a long-standing design issue,” even though those were not part of the current user prompt.

This happened twice between 2026-06-16 21:00 and 22:00 JST, in separate contexts.

This is not a tone/persona issue. The informal style itself was expected. The issue is that the response semantically referred to a nonexistent context instead of answering the latest user message.

Expected Behavior

Claude should have answered the latest user prompt.

For the first case, Claude should have analyzed the Obsidian workflow problem and proposed a practical operating model for project/task management.

For the second case, Claude should have reasoned about the provided daily note/template context and explained whether daily-note-based task management would remain maintainable, or suggested an alternative structure.

If Claude Code failed to assemble or preserve the correct request context, it should have surfaced an error instead of silently producing a response to the wrong or nonexistent context.

Files Affected

None

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

I do not have a fully deterministic reproduction yet, but the observed pattern was:

1. Use Claude Code with Claude Opus 4.8.
2. Ask a moderately complex planning/design question involving an Obsidian workflow and Markdown notes.
3. Let Claude think for an extended period, approximately 10 minutes.
4. Observe that Claude returns a response that appears to continue from a nonexistent or stale previous conversation instead of answering the latest prompt.

A similar issue occurred again in a separate context when I provided Markdown file paths and explained a concern about using daily notes for project task management.

Claude Model

Opus

Relevant Conversation

Example 1:

User asked Claude to advise on how to integrate Obsidian with project/task management, including tradeoffs between daily notes, project folders, project notes, task notes, and update tracking.

Actual response after approximately 10 minutes of thinking included content equivalent to:

> Thanks / good discussion / come back anytime / I only organized what you already verbalized

It did not answer the actual workflow design question.
It also appeared to fabricate a preceding exchange, responding as if I had thanked, praised, or reacted to Claude, even though I had not said anything like that in the current conversation.

Example 2:

User provided a daily note template path and an actual daily note path, then explained that after completing a task on June 16, they became concerned that project tasks stored in daily notes would become hard to find later.

Actual response after approximately 10 minutes of thinking started with content equivalent to:

> No need to thank me. This was a long-standing design issue. It is significant that this was resolved today.

This again appeared to refer to a nonexistent previous exchange before attempting to proceed.

Impact

Medium - Extra work to undo changes

Claude Code Version

2.1.169

Platform

Anthropic API

Additional Context

Patterns noticed:

• Happened twice between 2026-06-16 21:00 and 22:00 JST.
• Native Windows
• Both cases involved Claude Opus 4.8.
• Both cases occurred after long thinking, approximately 10 minutes.
• Both cases involved Obsidian / Markdown / project-task workflow planning.
• The responses looked like they were generated from a stale, nonexistent, or cross-contaminated conversational state.
• The assistant’s tone/persona was not the problem. The problem was that the response content did not correspond to the latest user prompt.

Possibly related behavior:

• Claude appeared to answer as if there had been a prior user message thanking or praising it, even though that was not present in the current prompt.
• This resembles a context mismatch or stale conversation issue rather than a normal instruction-following failure.