Function persistPiPressureFromMessageEnd

packages/pi-plugin/src/index.ts:252–328

(args: {
	db: ContextDatabase;
	sessionId: string;
	message: unknown;
	piContextWindow: number;
	piTokens?: number;
	notifyIssue?: (message: string) => unknown | Promise<unknown>;
})

}

export async function persistPiPressureFromMessageEnd(args: {
	db: ContextDatabase;
	sessionId: string;
	message: unknown;
	piContextWindow: number;
	piTokens?: number;
	notifyIssue?: (message: string) => unknown | Promise<unknown>;
}): Promise<void> {
	const { provider, model } = getPiMessageModel(args.message);
	const effectiveContextLimit = resolvePiPressureContextLimit({
		db: args.db,
		sessionId: args.sessionId,
		piContextWindow: args.piContextWindow,
	});
	const usage = extractAssistantUsage(args.message);
	const pressure = computePiPressure(usage, effectiveContextLimit);
	const msg =
		args.message && typeof args.message === "object"
			? (args.message as { errorMessage?: unknown })
			: undefined;
	const messageHadOverflowError =
		typeof msg?.errorMessage === "string" &&
		detectOverflow(msg.errorMessage).isOverflow;
	const updates: Partial<{
		lastResponseTime: number;
		lastContextPercentage: number;
		lastInputTokens: number;
		observedSafeInputTokens: number;
		cacheAlertSent: boolean;
	}> = { lastResponseTime: Date.now() };

	if (pressure) {
		const percentage = pressure.percentage;
		const contextLimit = effectiveContextLimit;
		const meta = getOrCreateSessionMeta(args.db, args.sessionId);
		const observedSafeInputTokens = meta.observedSafeInputTokens ?? 0;
		if (
			percentage > 100 &&
			observedSafeInputTokens > 0 &&
			pressure.inputTokens <= observedSafeInputTokens * 2
		) {
			// Pi resolves the window from its own runtime, not a cache we could
			// reload — so a >100% reading with a known-good safe baseline means
			// Pi's reported contextWindow is genuinely wrong. There's nothing to
			// re-fetch; surface the alert (overflow detection still captures a
			// real lower cap separately).
			if (!meta.cacheAlertSent) {
				updates.cacheAlertSent = true;
				const safeTokens = Math.max(
					observedSafeInputTokens,
					pressure.inputTokens,
				);
				const modelLabel =
					provider && model ? `${provider}/${model}` : "the active model";
				await args.notifyIssue?.(
					`⚠️ Magic Context: Pi reports a context limit of ${formatTokens(contextLimit)} tokens for ${modelLabel} but you've successfully sent ${formatTokens(safeTokens)} tokens in this session — the reported limit looks wrong. Restart Pi if you suspect this is incorrect.`,
				);
			}