Session continuation consumes excessive tokens - 23% budget for trivial continue message

After a long conversation (no context compactions occurred), sending "continue" after a session reset consumed 23% of the current session token budget before any work could begin. A single skill load (~100 words of actual skill content) plus system reminders consumed 23% with zero productive output.

Expected

A "continue" message should consume minimal tokens, proportional to the compressed context + user message.

Actual

23% of session budget consumed immediately with no context compaction having occurred. The system-reminder injections (skill content, todo list reminders, CLAUDE.md, memory index, deferred tool lists) appear to be re-injected in full on every single turn and compound to consume disproportionate token budget.

Repro steps

1. Run a long conversation with 87+ file read/compare iterations involving skill loads, file reads, and edits
2. No context compaction occurs during the conversation
3. Session resets (user closes/reopens or conversation continues)
4. User sends "continue"
5. First tool call + skill load = 23% session tokens consumed with zero productive output

Impact

Users on usage-limited plans lose significant session capacity to system overhead they cannot control. A session that should have hours of remaining capacity becomes unusable after a single message.

Environment

• Claude Code VSCode extension
• Windows 11
• Claude Opus 4.6 (1M context)
• Max (5x) plan