[FEATURE] Optional per-session activity summary the agent can read for self-monitoring

What

An optional, read-only summary of the current session's own activity that the agent can consult: counts like tool-call volume, error and retry rate, repeated identical actions, and whether a recurring task is taking more steps than it did earlier in the session.

Why

A long session accumulates real signal about its own health, but the agent never sees it because each turn starts fresh from the conversation, not from a view of its behavior over time. A slow drift (climbing retry rate, the same failing action attempted again and again, a loop producing no new result) is invisible until a human notices. If the agent could read a short, grounded summary of its own recent activity, it could act on it: stop when it is looping on identical outputs, surface a rising error rate instead of pushing on, or flag that it is stuck.

This is deliberately different from two things that already exist:

• Compaction summarizes conversation content. This is process telemetry — derived from the tool/turn log, not from what was said.
• Hooks fire on a single action. This is the aggregate view across the session that no single hook sees.

Rough shape

It could surface as a value the agent can read on demand, or as a periodic event (every N turns) carrying the summary, so behavior is opt-in and cheap. The numbers come from data the harness already has.

Happy to sketch the specific signals (the three most useful in my experience are retry-rate trend, repeated-identical-action detection, and steps-per-recurring-task) if this is worth pursuing.