[BUG] Unrelated, unrequested conversation content appeared mid-session in an otherwise unrelated conversation

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

While working through an unrelated task (AWS MediaConvert cost/break-even analysis for a video transcoding service), a completely unrelated question-and-answer exchange about EIP-4844 ("proto-danksharding," blob transactions, KZG commitments) appeared in the conversation transcript. Neither I (the user) nor the assistant authored or requested this content — it simply appeared inline between my message and the assistant's message, as if it were part of the conversation history.

The assistant confirmed it did not generate the EIP-4844 answer and has no way to trace where it came from from within the conversation itself.

This is concerning because the session involved sensitive material (AWS account IDs, IAM role ARNs, internal cost/pricing data). If this is cross-session content bleeding into the wrong conversation, it could indicate a session isolation issue.

What Should Happen?

A conversation should only ever contain messages that the user actually sent and responses the assistant actually generated for that session. No unrelated, unrequested Q&A content from another topic/session should ever appear in the transcript.

Error Messages/Logs

No explicit error message was shown. The unrelated content appeared as a fully-formed user question followed by a fully-formed, well-structured assistant answer (multiple sections, code blocks, headers) about EIP-4844 blob transactions — styled consistently with a normal assistant response, which made it initially easy to mistake for part of the real conversation.

Steps to Reproduce

Not reliably reproducible on demand. Observed once during a long-running session that included:

• Multiple background Bash tasks (AWS CLI calls, polling MediaConvert job status)
• Several Agent/Explore subagent dispatches
• An Artifact publish
• Many turns over an extended period (session spanned tasks from AWS resource setup through cost analysis)

The unrelated content appeared right before a user message reading "환율을 1450원 기준으로 하자" ("let's use a 1450 won exchange rate"), i.e., it was interleaved into the transcript at that point rather than sent as a real user turn.

Claude Model

Sonnet (default)

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.220 ( Claude Code for VS Code )

Platform

Other

Operating System

Ubuntu/Debian Linux

Terminal/Shell

VS Code integrated terminal

Additional Information

• The session had been running for a long time with heavy background task usage (multiple run_in_background: true Bash calls polling AWS MediaConvert job status over several hours of wall-clock time, plus async Agent dispatches).
• The unrelated content was thematically and stylistically coherent (a detailed, well-formatted technical explanation of EIP-4844) but had zero relation to anything discussed in the session — no prior mention of Ethereum, blockchain, rollups, or any related topic anywhere else in the conversation.
• Suspect this may be a transcript/context-merging issue possibly related to long sessions with many interleaved background task notifications, but this is speculation — I have no visibility into server-side session handling.

• Possible related report: I've seen another user-submitted bug report describing a very similar pattern — during a long conversation, a fully-formed, unrelated block of content (in their case, a multi-round medical triage roleplay transcript) appeared with a clean boundary after/within the assistant's normal reply. Their report: model claude-fable-5, Claude Code CLI 2.1.216, macOS, Anthropic API. Despite the different model/platform/version, the shared pattern (long session, unrelated coherent content with intact structure, clean language/topic boundary) suggests this may be the same underlying issue rather than two unrelated bugs — possibly a session/context-routing issue above the model or client layer.