Two recent Claude Code releases - 2.1.124 and 2.1.126 - add 166 tokens to the system prompt then claw back 87, leaving a net gain of 79. The changes are small in size but tell you something useful about how Anthropic is solving a real agent problem.
The headliner addition in 2.1.124 is a "File modification detected" system reminder. When a user or a linter edits a file during a session but the diff (the record of what changed) gets cut because other modified files already consumed the available context budget, Claude Code now receives an explicit notice saying the change happened - and an instruction to re-read the file if it needs the current content. Without this, the agent could keep working from a stale version of a file it never knew had changed. That's a silent failure mode that causes real bugs, and this reminder patches it.
The 2.1.126 update restructures the harness instructions - the scaffolding that tells the model how to behave as a coding agent. The old approach used a core-identity function call; the new one replaces it with explicit insertion points for an introductory line and a security note. This is internal plumbing more than user-facing behavior, but it suggests Anthropic is making the prompt architecture easier to modify modularly rather than updating one monolithic block.
Neither release changes what Claude Code can do. They tighten how it tracks state during long sessions where files are moving around - which is exactly when coding agents tend to go sideways.