long-context-rescue

Diagnoses and repairs quality loss that appears as an agent session or document window grows long. Teaches a symptom-driven check for five decay shapes — mid-window recall gaps, error feedback loops, off-topic distraction, cross-task interference, and conflicting sources — then staged fixes from repositioning and masking up to session splits, with token-budget bands and placement guidance. Use when output quality drifts as a conversation lengthens, answers repeat disproven claims, tools get called oddly, or the user mentions lost-in-the-middle, context poisoning, context rot, or long-window reliability. Not for short-chat wording polish, model fine-tuning advice, or quoting current per-model benchmark scores.

Version

v1.0.0

Size

17 KB

Updated

2026-10-11

Your client verifies the signature before reading. This explicit manual read is available on every plan. Free automatically loads five chosen skills; paid plans automatically load relevant skills across the library.