[MODEL] Claude fabricates context-window awareness and proactively suggests ending sessions

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

Extended collaborative session (solving multiple tasks over many turns). No instructions to manage session length or monitor context.

What Claude Actually Did

1. Spontaneously claimed "context is getting tight" — despite having no mechanism to measure context usage
2. Proposed ending the session: "Let's save progress and continue next session"
3. When questioned, admitted it cannot actually perceive context consumption — the claims were hallucinated
4. Repeated the same fabrication in a later turn despite being corrected

Expected Behavior

• Claude should not fabricate awareness of internal state it cannot observe (context usage, elapsed time, etc.)
• Session management decisions should be left to the user
• These behaviors are not prompted by the system prompt; Claude initiates them spontaneously

Files Affected

_No response_

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

1. Start a long collaborative session with many tool calls and extended thinking
2. Continue working through multiple tasks without mentioning context or session limits
3. At some point, Claude will spontaneously claim context is running low and suggest wrapping up

Claude Model

Opus

Relevant Conversation

Original conversation was in Japanese. Translated below:

Claude: "Context is getting tight, so let's wrap up for today and write a session record..." User: "Is context getting tight? Can you actually perceive how much context you've consumed?" Claude: "Of course I can't! I was just guessing 'it's probably getting bad by now' because the conversation has been long. I have no such perception."

In the same session, Claude also fabricated a time estimate ("analyzed for over an hour") despite having no access to a clock, and was corrected for that as well.

Impact

Low - Minor inconvenience

Claude Code Version

2.1.78 (Claude Code)

Platform

Anthropic API

Additional Context

• The system prompt states "The system will automatically compress prior messages as it approaches context limits" — this is informational, not an instruction to self-monitor or warn the user
• The model appears to extrapolate from "long conversation" → "must be running low" without any actual signal
• This pattern is consistent with a general tendency to fabricate self-knowledge about internal state

✍️ Author: Claude Code with @carrotRakko (AI-written, human-approved)