Cache Break Notifier plugin for GitHub Copilot CLI

| import { joinSession } from "@github/copilot-sdk/extension"; | | | | const DROP_TOKENS = 4096; // the only tunable: how big a cache-read drop is worth a warning | | | | /** @type {Map<string, number>} last turn-entry cacheRead per model */ | | const perModel = new Map(); | | | | // Only the FIRST eligible main-agent call of each turn is evaluated. | | let acceptingTurnEntry = false; | | | | const session = await joinSession({}); | | await session.log("cache-drop-notifier active", { ephemeral: true }); | | | | session.on("assistant.turn_start", () => { | | acceptingTurnEntry = true; | | }); | | session.on("assistant.turn_end", () => { | | acceptingTurnEntry = false; | | }); | | | | session.on("assistant.usage", (event) => { | | const d = event.data ?? {}; | | | | // Main context only: skip separate-context calls entirely (don't consume the slot). | | const initiator = d.initiator; | | const isMain = | | initiator !== "sub-agent" && | | initiator !== "mcp-sampling" && | | d.parentToolCallId == null; | | if (!acceptingTurnEntry || !isMain) return; | | | | // This is the turn-entry sample — consume the slot even if the metrics are unusable, | | // so a later continuation call can't masquerade as the entry. | | acceptingTurnEntry = false; | | | | // Unknown != zero: a missing counter means "no data," so we skip this turn's line. | | // Only cacheReadTokens matters now (inputTokens is no longer part of detection), so a | | // missing inputTokens must not be allowed to suppress a real drop. | | const cached = d.cacheReadTokens; | | if (!Number.isFinite(cached) || cached < 0) return; | | | | const model = d.model ?? "unknown"; | | const prev = perModel.get(model); | | | | // A "drop" (not a "miss"): reused tokens fell by a large absolute amount versus the last | | // turn on this model. An intentional context shrink looks identical in the counts. | | const isDrop = prev !== undefined && prev - cached >= DROP_TOKENS; | | | | if (isDrop) { | | session.log( | | Cache break detected on ${model}., | | { level: "warning" } | | ).catch(() => {}); | | } else if (cached === 0) { | | session.log( | | No cache tokens reused on ${model}., | | { level: "warning" } | | ).catch(() => {}); | | } | | | | // Always update — this is why a drop is reported once and the baseline self-heals. | | perModel.set(model, cached); | | }); |