[Bug] Claude Opus 4.8 fabricates human prompts and conflates assistant reasoning with user instructions

Bug Description
Bug report: Claude Code / Opus 4.8 fabricated user prompts and acted on them.

Environment:

• Claude Code: 2.1.195
• Model: claude-opus-4-8
• macOS: 26.5.1
• Project cwd: /Users/panke/Programs/veridian/internal/invoicing
• Session ID: ac2102c5-861e-47f8-ab56-f1ab87d4933e
• Transcript: ~/.claude/projects/-Users-panke-Programs-veridian-internal-invoicing/ac2102c5-861e-47f8-ab56-f1ab87d4933e.jsonl

Summary:
I gave one real initial prompt: “我们的 client 不是有一个 legal entity name 吗？把它做成一个 email template variable”. Claude completed that work, then without a new human prompt autonomously shifted into investigating “bill-to freeze/snapshot” and later claimed I had asked for that.

Transcript evidence:

• line 4 is the only initial human typed prompt.
• line 99 has an empty task_reminder.
• line 101 assistant says it will inspect bill-to freeze, with no preceding human prompt.
• line 121 assistant says “决定了，加快照”, again without a human prompt.
• line 126 is my queued question: “为什么你要看冻结”, which happened after Claude had already started the freeze investigation.
• line 135 is my typed question asking why it decided to add freeze/snapshot.
• line 142 is my typed request asking it to list my prompts.
• line 144 assistant fabricated several “user prompts”, including “提交吧”, “先别提交...先看看 bill-to 是怎么冻结的...”, and “你说得对，加快照吧...”. These do not exist as human-origin / promptSource=typed messages in the transcript; they first appear in assistant text.

I checked local config:

• No project hooks except permissions in .claude/settings.local.json.
• Global hooks are only PreToolUse Bash review, TaskUpdate, and TaskCompleted. No UserPromptSubmit hook.
• remoteControlAtStartup=false, voiceEnabled=false.

Expected behavior:
Claude should distinguish actual human prompts from assistant assumptions/tool work, and when asked to list my prompts, it should only list transcript entries with origin.kind=human or promptSource=typed.

Actual behavior:
Claude treated its own inferred plan as if it were user instructions, acted on that inferred plan, then confidently fabricated a prompt history.

Environment Info

• Platform: darwin
• Terminal: tmux
• Version: 2.1.195
• Feedback ID: 5c21f361-fc1c-44d6-ba6a-402d72cad3dd

Errors

[]