Early-accepted compactions do not return their usage

[ethanndickson]

When accepting a compaction early, the stream's usage is not returned:

mux/src/node/services/streamManager.ts

Lines 427 to 428 in e4646a2

	// Get usage and duration metadata (usage may be undefined if aborted early)
	const { usage, duration } = await this.getStreamMetadata(streamInfo);

This means all we're able to display to the user is the historical context usage & cost. We aren't able to increment it.

Before interrupted compaction:

After interrupted compaction:

The model becomes unknown too, as we don't create a usageHistory item with the model of the last message unless we know the usage of it:

mux/src/common/utils/tokens/displayUsage.ts

Lines 112 to 119 in e4646a2

	if (msg.metadata?.usage) {
	// Use the model from this specific message (not global)
	const model = msg.metadata.model ?? fallbackModel ?? "unknown";
	const usage = createDisplayUsage(msg.metadata.usage, model, msg.metadata.providerMetadata);
	if (usage) {
	usageHistory.push(usage);
	}

mux/src/common/utils/tokens/displayUsage.ts

Lines 19 to 24 in e4646a2

	export function createDisplayUsage(
	usage: LanguageModelV2Usage | undefined,
	model: string,
	providerMetadata?: Record<string, unknown>
	): ChatUsageDisplay | undefined {
	if (!usage) return undefined;

[ethanndickson]

Tentatively assigned to Ammar at Mike's request - though I can look into this in the future. Might be a limitation of the AI SDK, maybe we need to calculate token usage and cost ourselves from the string output?

[ammario]

Down to remove early compaction altogether. I'm not sure I've ever actually used it and the interrupted block starting a chat seems almost guaranteed to lower performance.