strain

Strain

A long session degrades before it fails. The context window fills, the same problem comes back a third time, an earlier answer turns out to be wrong — and the work keeps going, because nothing in the loop is watching the loop. Strain is the thing that watches: a small, countable reading of how loaded this conversation has become, reported on a schedule, in a form the user can act on.

The failure this exists to prevent is not a crash. It is silence — the agent that never mentions the session has gone bad, and lets the user find out from the output.

The one rule

One session, one reading. Strain measures a single conversation, identified by its session id. Two agents working the same project in two windows are two sessions and their counts never add up. If you cannot tell which session a number belongs to, it is not a measurement.

When to run the check

Run it when any of these happen — not on a feeling that it might be time:

1. A strain tick fires. On hosts with hooks, a tick arrives every N tool calls (default 10) and says so explicitly. Run the check before continuing the work.
2. A hard signal lands (see the table below) — a factual error caught, a regression introduced, a revert of your own work, a context compaction.
3. The task list changes shape — a new side task, a goal switch, a task that balloons past the one it was supposed to serve.
4. The user asks how the session is doing, or whether to wrap.

On a host with no hooks, 2–4 still work. That is the cooperative half, and it is weaker: say so rather than implying the check is firing on its own when it is not.

What to count

Two inputs, and they are not equally strong.

Context occupancy — measured, when the host allows it

Some hosts publish a per-session transcript carrying token usage. Where that exists, the context reading is a real number, not an impression:

• current = the input side of the most recent turn (input + cache_read + cache_creation — cached tokens are still context the model is carrying)
• baseline = the same sum on the first turn: what the boot alone cost before any work happened. System prompt, tool schemas, project instructions, skills. It is the floor the session can never get back under, and it is usually larger than people expect.