Agent Bash calls that repeat with byte-identical output are now warned about, then denied

What changed

In Claude Code agent sessions, repeated identical Bash calls are now guarded in two tiers. When the exact same command has run 3 times in a row with byte-identical output, the next identical call is allowed but the agent receives a loop warning so it can self-correct. When the run reaches 5, the next identical call is denied with an explanation. Any output change, a completed file edit, or the user pressing Esc resets the count. The denial applies in every permission mode, including bypass-permissions runs.

Why this matters to the user

Unattended agent runs could previously burn tokens retrying a command that provably returns the same result. Those loops now surface a warning at the 4th identical call and stop at the 6th, with the agent told to diagnose, vary the command, or report the blocker. The one visible narrowing: a deliberate workflow that polls a command expecting identical output (for example waiting on a health endpoint that never changes) is interrupted and must vary the invocation or ask the user.

What the user should do

Nothing — automatic. If an agent legitimately needs to poll, any edit, pressing Esc once, or a trivially varied invocation (such as adding && true) starts a fresh count.

Notes for release manager

Only affects the Claude Code runtime's agent sessions; terminal usage by the user directly is untouched.