[MODEL] ALL YOUR MODELS: Compaction summaries should not count against user session tokens — users are charged for Claude's context limitations

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 started a new session to continue work on my portfolio project. I had not yet typed any task prompt.

What Claude Actually Did

1. Loaded a compaction summary from the previous session — consuming 12% of session tokens before I typed anything
2. When I asked for a bug audit, Claude spawned 2 parallel Explore agents instead of reading files directly — consuming another ~10% of session tokens
3. Total: 36% of session tokens consumed before I had asked for any real work

Expected Behavior

1. Compaction summaries should not count against the user's session token quota — the user did not cause the context overflow, the model's architecture did
2. Claude should read files directly (Grep/Read tools) for routine tasks, not spawn Explore agents
3. Agents should only be spawned when scope is genuinely unknown AND user approves

Files Affected

None — this is a token consumption / model behavior issue, not a file modification issue

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Run a long multi-session project until context compaction occurs
2. Start a new session — observe token usage before typing any prompt
3. Ask Claude for a "full audit" of a codebase
4. Claude will spawn multiple Explore agents instead of using Read/Grep directly

Claude Model

Sonnet

Relevant Conversation

Session started. Compaction summary loaded. Claude said: "The 12% was the compressed summary of our previous conversation loading — not prompts you sent."

Then for the audit: Claude spawned 2 parallel Explore agents consuming ~10% more tokens for a task that required reading ~5 files directly.

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.92 (Claude Code)

Platform

Anthropic API

Additional Context

I am a Pro subscriber. I have "Extra Usage" enabled with $57.91 already spent this month. Token waste from compaction + unnecessary agent spawning is directly causing extra charges. The model should default to minimal tool use (Read/Grep) before escalating to agents.