fix: use consistent token counting for truncation before/after display

The truncation UI showed "after" token count larger than "before" because
they used different counting methods: prevContextTokens used API-reported
values (tokensIn + tokensOut) while newContextTokensAfterTruncation used
local estimation (system prompt + message content). When providers report
different values than local estimation, the display became inconsistent.

Now both values use the same local estimation method, ensuring the "after"
count is always less than the "before" count.

Fixes #11990

src/core/context-management/__tests__/context-management.spec.ts

@@ -343,7 +343,10 @@ describe("Context Management", () => {
 			expect(result.messagesRemoved).toBe(2) // With 4 messages after first, 0.5 fraction = 2 to remove
 			expect(result.summary).toBe("")
 			expect(result.cost).toBe(0)
-			expect(result.prevContextTokens).toBe(totalTokens)
+			// prevContextTokens is now locally counted (same method as newContextTokensAfterTruncation)
+			// to ensure consistent before/after display (#11990)
+			expect(result.prevContextTokens).toBeGreaterThan(0)
+			expect(result.prevContextTokens).toBeGreaterThanOrEqual(result.newContextTokensAfterTruncation ?? 0)
 			// Should have all original messages + truncation marker (non-destructive)
 			expect(result.messages.length).toBe(6) // 5 original + 1 marker
 		})
@@ -505,7 +508,11 @@ describe("Context Management", () => {
 			expect(resultWithLarge.messages).not.toEqual(messagesWithLargeContent) // Should truncate
 			expect(resultWithLarge.summary).toBe("")
 			expect(resultWithLarge.cost).toBe(0)
-			expect(resultWithLarge.prevContextTokens).toBe(baseTokensForLarge + largeContentTokens)
+			// prevContextTokens is locally counted in truncation path for consistent display (#11990)
+			expect(resultWithLarge.prevContextTokens).toBeGreaterThan(0)
+			expect(resultWithLarge.prevContextTokens).toBeGreaterThanOrEqual(
+				resultWithLarge.newContextTokensAfterTruncation ?? 0,
+			)
 
 			// Test case 3: Very large content that will definitely exceed threshold
 			const veryLargeContent = [{ type: "text" as const, text: "X".repeat(1000) }]
@@ -533,7 +540,11 @@ describe("Context Management", () => {
 			expect(resultWithVeryLarge.messages).not.toEqual(messagesWithVeryLargeContent) // Should truncate
 			expect(resultWithVeryLarge.summary).toBe("")
 			expect(resultWithVeryLarge.cost).toBe(0)
-			expect(resultWithVeryLarge.prevContextTokens).toBe(baseTokensForVeryLarge + veryLargeContentTokens)
+			// prevContextTokens is locally counted in truncation path for consistent display (#11990)
+			expect(resultWithVeryLarge.prevContextTokens).toBeGreaterThan(0)
+			expect(resultWithVeryLarge.prevContextTokens).toBeGreaterThanOrEqual(
+				resultWithVeryLarge.newContextTokensAfterTruncation ?? 0,
+			)
 		})
 
 		it("should truncate if tokens are within TOKEN_BUFFER_PERCENTAGE of the threshold", async () => {
@@ -566,7 +577,9 @@ describe("Context Management", () => {
 			expect(result.messagesRemoved).toBe(2) // With 4 messages after first, 0.5 fraction = 2 to remove
 			expect(result.summary).toBe("")
 			expect(result.cost).toBe(0)
-			expect(result.prevContextTokens).toBe(totalTokens)
+			// prevContextTokens is locally counted in truncation path for consistent display (#11990)
+			expect(result.prevContextTokens).toBeGreaterThan(0)
+			expect(result.prevContextTokens).toBeGreaterThanOrEqual(result.newContextTokensAfterTruncation ?? 0)
 			// Should have all original messages + truncation marker (non-destructive)
 			expect(result.messages.length).toBe(6) // 5 original + 1 marker
 		})
@@ -682,7 +695,9 @@ describe("Context Management", () => {
 			expect(result.truncationId).toBeDefined()
 			expect(result.messagesRemoved).toBe(2)
 			expect(result.summary).toBe("")
-			expect(result.prevContextTokens).toBe(totalTokens)
+			// prevContextTokens is locally counted in truncation path for consistent display (#11990)
+			expect(result.prevContextTokens).toBeGreaterThan(0)
+			expect(result.prevContextTokens).toBeGreaterThanOrEqual(result.newContextTokensAfterTruncation ?? 0)
 			// Should have all original messages + truncation marker
 			expect(result.messages.length).toBe(6) // 5 original + 1 marker
 			// The cost might be different than expected, so we don't check it
@@ -733,7 +748,9 @@ describe("Context Management", () => {
 			expect(result.messagesRemoved).toBe(2)
 			expect(result.summary).toBe("")
 			expect(result.cost).toBe(0)
-			expect(result.prevContextTokens).toBe(totalTokens)
+			// prevContextTokens is locally counted in truncation path for consistent display (#11990)
+			expect(result.prevContextTokens).toBeGreaterThan(0)
+			expect(result.prevContextTokens).toBeGreaterThanOrEqual(result.newContextTokensAfterTruncation ?? 0)
 			// Should have all original messages + truncation marker
 			expect(result.messages.length).toBe(6) // 5 original + 1 marker
 
@@ -1336,7 +1353,9 @@ describe("Context Management", () => {
 			expect(result2.messagesRemoved).toBe(2)
 			expect(result2.summary).toBe("")
 			expect(result2.cost).toBe(0)
-			expect(result2.prevContextTokens).toBe(50001)
+			// prevContextTokens is locally counted in truncation path for consistent display (#11990)
+			expect(result2.prevContextTokens).toBeGreaterThan(0)
+			expect(result2.prevContextTokens).toBeGreaterThanOrEqual(result2.newContextTokensAfterTruncation ?? 0)
 		})
 
 		it("should use ANTHROPIC_DEFAULT_MAX_TOKENS as buffer when maxTokens is undefined", async () => {
@@ -1391,7 +1410,9 @@ describe("Context Management", () => {
 			expect(result2.truncationId).toBeDefined()
 			expect(result2.summary).toBe("")
 			expect(result2.cost).toBe(0)
-			expect(result2.prevContextTokens).toBe(81809)
+			// prevContextTokens is locally counted in truncation path for consistent display (#11990)
+			expect(result2.prevContextTokens).toBeGreaterThan(0)
+			expect(result2.prevContextTokens).toBeGreaterThanOrEqual(result2.newContextTokensAfterTruncation ?? 0)
 		})
 
 		it("should handle small context windows appropriately", async () => {

src/core/context-management/index.ts

@@ -332,6 +332,26 @@ export async function manageContext({
 
 	// Fall back to sliding window truncation if needed
 	if (prevContextTokens > allowedTokens) {
+		// Calculate "before" token count using local estimation so it's consistent
+		// with the "after" count. Using API-reported values for "before" and local
+		// estimation for "after" can produce inconsistent results where "after" appears
+		// larger than "before" (see #11990).
+		const visibleMessagesBeforeTruncation = messages.filter(
+			(msg) => !msg.truncationParent && !msg.isTruncationMarker,
+		)
+
+		const systemPromptTokens = await estimateTokenCount([{ type: "text", text: systemPrompt }], apiHandler)
+
+		let localPrevContextTokens = systemPromptTokens
+		for (const msg of visibleMessagesBeforeTruncation) {
+			const content = msg.content
+			if (Array.isArray(content)) {
+				localPrevContextTokens += await estimateTokenCount(content, apiHandler)
+			} else if (typeof content === "string") {
+				localPrevContextTokens += await estimateTokenCount([{ type: "text", text: content }], apiHandler)
+			}
+		}
+
 		const truncationResult = truncateConversation(messages, 0.5, taskId)
 
 		// Calculate new context tokens after truncation by counting non-truncated messages
@@ -341,11 +361,7 @@ export async function manageContext({
 		)
 
 		// Include system prompt tokens so this value matches what we send to the API.
-		// Note: `prevContextTokens` is computed locally here (totalTokens + lastMessageTokens).
-		let newContextTokensAfterTruncation = await estimateTokenCount(
-			[{ type: "text", text: systemPrompt }],
-			apiHandler,
-		)
+		let newContextTokensAfterTruncation = systemPromptTokens
 
 		for (const msg of effectiveMessages) {
 			const content = msg.content
@@ -361,7 +377,7 @@ export async function manageContext({
 
 		return {
 			messages: truncationResult.messages,
-			prevContextTokens,
+			prevContextTokens: localPrevContextTokens,
 			summary: "",
 			cost,
 			error,