fix(ai-chat): accumulate output tokens across tool-use steps in a turn

Fold each finished round-trip's streamed total into a base so the thinking
indicator's token count stays monotonic across a multi-step turn instead of
dropping back to near-zero when the next step's message begins.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

frontend/src/lib/components/copilot/chat/AIChatManager.svelte.ts

@@ -273,9 +273,15 @@ export class AIChatManager {
 	currentReply = $state<string>('')
 	currentReasoning = $state<string>('')
 	currentReasoningActive = $state<boolean>(false)
-	// Live output-token count for the in-flight assistant message; drives the typing
-	// indicator's progress readout. Only populated by providers that stream usage.
+	// Live output-token count for the current turn; drives the typing indicator's
+	// progress readout. Only populated by providers that stream usage.
 	currentOutputTokens = $state<number>(0)
+	// A turn can span several API round-trips (one per tool-use step). Each round-trip's
+	// streamed usage.output_tokens is a running total scoped to that one message, so it
+	// restarts from near-zero every step. To keep the readout monotonic across the whole
+	// turn, fold each finished step's total into a base and add the in-flight step on top.
+	#outputTokensBase = 0
+	#outputTokensStep = 0
 	displayMessages = $state<DisplayMessage[]>([])
 	messages = $state<ChatCompletionMessageParam[]>([])
 	/** Provider-reported context size of the last committed turn (prompt +
@@ -1641,6 +1647,8 @@ export class AIChatManager {
 			this.currentReasoning = ''
 			this.currentReasoningActive = false
 			this.currentOutputTokens = 0
+			this.#outputTokensBase = 0
+			this.#outputTokensStep = 0
 
 			// Compaction trigger. Without a known context window there is no limit
 			// to enforce, so compaction stays off rather than guessing one.
@@ -1696,7 +1704,14 @@ export class AIChatManager {
 					onNewToken: (token) => (this.currentReply += token),
 					onReasoningDelta: (token) => (this.currentReasoning += token),
 					onReasoningStart: () => (this.currentReasoningActive = true),
-					onOutputTokens: (n) => (this.currentOutputTokens = n),
+					onOutputTokens: (n) => {
+						// A drop means a new step's message started; bank the prior step's total.
+						if (n < this.#outputTokensStep) {
+							this.#outputTokensBase += this.#outputTokensStep
+						}
+						this.#outputTokensStep = n
+						this.currentOutputTokens = this.#outputTokensBase + n
+					},
 					onMessageEnd: () => {
 						// Keep the streamed text for the abort/error paths. Non-empty only:
 						// parsers flush (and reset) when a tool call starts after text, and