Skip to content
Notifly Notifly

Context Close to Limit

Most bugs in long sessions with an AI agent start precisely when the context window is filled to 90%+. The model begins to “lose” system prompt instructions and tools. The signal is usage.input_tokens relative to the model limit:

LIMITS = {
    "claude-sonnet-4-5": 200_000,
    "claude-opus-4-5":   200_000,
    "gpt-4.1":           1_000_000,
    "gpt-4o":            128_000,
}


def observe(model, used_tokens):
    limit = LIMITS.get(model, 200_000)
    ratio = used_tokens / limit
    flag  = f"/tmp/ctx-{model}.flag"
    if ratio >= 0.85 and not os.path.exists(flag):
        push(f"🧠 Контекст {int(ratio*100)}% у {model}",
             f"Использовано {used_tokens:,} из {limit:,} токенов.\n"
             "Сделайте summary или начните новый thread.",
             priority=7)
        open(flag, "w").close()
    elif ratio < 0.5 and os.path.exists(flag):
        os.unlink(flag)            # reset flag after summary

It’s convenient to call observe() in a common wrapper over the SDK — this control then works automatically for any long dialogue.

Section titled “Related recipes”