[BUG] Conversations compacting prematurely at ~65% token capacity instead of near limit

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?

Preflight Checklist

• I am using the latest version of Claude Code
• I have searched existing issues to ensure this hasn't been reported
• I can reproduce this issue consistently
• I have run /clear to start fresh and the issue persists

What's Wrong?

Bug Description:
Claude Code is compacting conversations prematurely, well before reaching the 200,000 token limit. Compacting occurs at approximately 65-75% token capacity (~130k-150k tokens) instead of near the actual limit.

Steps to Reproduce:

1. Run /clear to start a fresh conversation
2. Engage in a conversation that uses approximately 130k-150k tokens
3. Observe conversation being compacted despite having 50k+ tokens remaining

Expected Behavior:
Conversations should only be compacted when approaching the 200,000 token limit (e.g., at 90-95% capacity or higher), not when significant token budget remains available.

Actual Behavior:

• Conversations are compacted at approximately 64.9% capacity (129,795/200,000 tokens in current session)
• Compacting occurs even after using /clear to start fresh
• 50k+ tokens of available budget are unused when compacting triggers

Environment:

• Claude Code CLI
• Model: Claude Sonnet 4.5 (claude-sonnet-4-5-20250929)
• Platform: macOS (Darwin 25.1.0)

Impact:
This premature compacting reduces the effective context window by 30-35%, limiting the ability to work on complex tasks that require maintaining extensive conversation context.

What Should Happen?

Preflight Checklist

• I am using the latest version of Claude Code
• I have searched existing issues to ensure this hasn't been reported
• I can reproduce this issue consistently
• I have run /clear to start fresh and the issue persists

What's Wrong?

Bug Description:
Claude Code is compacting conversations prematurely, well before reaching the 200,000 token limit. Compacting occurs at approximately 65-75% token capacity (~130k-150k tokens) instead of near the actual limit.

Steps to Reproduce:

1. Run /clear to start a fresh conversation
2. Engage in a conversation that uses approximately 130k-150k tokens
3. Observe conversation being compacted despite having 50k+ tokens remaining

Expected Behavior:
Conversations should only be compacted when approaching the 200,000 token limit (e.g., at 90-95% capacity or higher), not when significant token budget remains available.

Actual Behavior:

• Conversations are compacted at approximately 64.9% capacity (129,795/200,000 tokens in current session)
• Compacting occurs even after using /clear to start fresh
• 50k+ tokens of available budget are unused when compacting triggers

Environment:

• Claude Code CLI
• Model: Claude Sonnet 4.5 (claude-sonnet-4-5-20250929)
• Platform: macOS (Darwin 25.1.0)

Impact:
This premature compacting reduces the effective context window by 30-35%, limiting the ability to work on complex tasks that require maintaining extensive conversation context.

Error Messages/Logs

Steps to Reproduce

1. Start Claude Code session
2. Run /clear to begin new conversation
3. Activate a Serena project (note: this may be relevant)
4. Continue normal conversation
5. Observe compacting occurs at ~129,795 tokens (64.9% of 200k budget)

Expected Behavior

Conversations should only compact when approaching the 200k token limit (e.g., at 180k-190k tokens), leaving minimal buffer for continuation.

Actual Behavior

Compacting occurs at approximately 129,795/200,000 tokens (64.9% capacity), with 70,205 tokens (35%+ of budget) still remaining. User receives "This session is being continued from a previous conversation that ran out of context" message despite substantial remaining capacity.

Environment

• Platform: macOS (Darwin 25.1.0)
• Claude Code Version: Latest (as of 2025-12-25)
• Model: claude-sonnet-4-5-20250929
• Token Budget: 200,000

Additional Context

The premature compacting appears to correlate with Serena project activation, though the root cause is likely in Claude Code's conversation management. There may be an interaction between project activation and conversation state that triggers early compacting. Worth investigating whether Serena integration affects conversation lifecycle decisions.

Impact

This significantly reduces effective conversation length and forces unnecessary context loss, degrading the development experience when working on complex tasks that require extended context.

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.0.76 (Claude Code)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_