fix: correct truncation token display accounting

Author: roomote

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

@@ -56,6 +56,22 @@ describe("Context Management", () => {
 			TelemetryService.createInstance([])
 		}
 	})
+
+	const countEffectiveHistoryTokens = async (messages: ApiMessage[], systemPrompt: string) => {
+		const effectiveMessages = condenseModule.getEffectiveApiHistory(messages)
+		let total = await estimateTokenCount([{ type: "text", text: systemPrompt }], mockApiHandler)
+
+		for (const message of effectiveMessages) {
+			if (Array.isArray(message.content)) {
+				total += await estimateTokenCount(message.content, mockApiHandler)
+			} else if (typeof message.content === "string") {
+				total += await estimateTokenCount([{ type: "text", text: message.content }], mockApiHandler)
+			}
+		}
+
+		return total
+	}
+
 	/**
 	 * Tests for the truncateConversation function
 	 */
@@ -1700,5 +1716,46 @@ describe("Context Management", () => {
 			// With system prompt included, we expect roughly 50% of the messages remaining
 			expect(result.newContextTokensAfterTruncation).toBeGreaterThan(0)
 		})
+
+		it("should count only the effective API history after truncation when prior condenses exist", async () => {
+			const modelInfo = createModelInfo(100000, 30000)
+			const totalTokens = 70001
+			const condenseId = "prior-condense"
+			const hiddenContent = "hidden historical content ".repeat(4000)
+
+			const messages: ApiMessage[] = [
+				{ role: "user", content: hiddenContent, condenseParent: condenseId },
+				{ role: "assistant", content: hiddenContent, condenseParent: condenseId },
+				{ role: "user", content: hiddenContent, condenseParent: condenseId },
+				{ role: "user", content: "Earlier summary", isSummary: true, condenseId },
+				{ role: "assistant", content: "Recent visible reply" },
+				{ role: "user", content: "Recent visible follow-up" },
+				{ role: "assistant", content: "Most recent visible reply" },
+				{ role: "user", content: "" },
+			]
+
+			const systemPrompt = "System prompt for truncation recount"
+
+			const result = await manageContext({
+				messages,
+				totalTokens,
+				contextWindow: modelInfo.contextWindow,
+				maxTokens: modelInfo.maxTokens,
+				apiHandler: mockApiHandler,
+				autoCondenseContext: false,
+				autoCondenseContextPercent: 100,
+				systemPrompt,
+				taskId,
+				profileThresholds: {},
+				currentProfileId: "default",
+			})
+
+			expect(result.truncationId).toBeDefined()
+			expect(result.newContextTokensAfterTruncation).toBeDefined()
+
+			const expectedEffectiveTokens = await countEffectiveHistoryTokens(result.messages, systemPrompt)
+			expect(result.newContextTokensAfterTruncation).toBe(expectedEffectiveTokens)
+			expect(result.newContextTokensAfterTruncation).toBeLessThan(result.prevContextTokens)
+		})
 	})
 })

src/core/context-management/index.ts

@@ -4,7 +4,13 @@ import crypto from "crypto"
 import { TelemetryService } from "@roo-code/telemetry"
 
 import { ApiHandler, ApiHandlerCreateMessageMetadata } from "../../api"
-import { MAX_CONDENSE_THRESHOLD, MIN_CONDENSE_THRESHOLD, summarizeConversation, SummarizeResponse } from "../condense"
+import {
+	MAX_CONDENSE_THRESHOLD,
+	MIN_CONDENSE_THRESHOLD,
+	getEffectiveApiHistory,
+	summarizeConversation,
+	SummarizeResponse,
+} from "../condense"
 import { ApiMessage } from "../task-persistence/apiMessages"
 import { ANTHROPIC_DEFAULT_MAX_TOKENS } from "@roo-code/types"
 import { RooIgnoreController } from "../ignore/RooIgnoreController"
@@ -334,11 +340,10 @@ export async function manageContext({
 	if (prevContextTokens > allowedTokens) {
 		const truncationResult = truncateConversation(messages, 0.5, taskId)
 
-		// Calculate new context tokens after truncation by counting non-truncated messages
-		// Messages with truncationParent are hidden, so we count only those without it
-		const effectiveMessages = truncationResult.messages.filter(
-			(msg) => !msg.truncationParent && !msg.isTruncationMarker,
-		)
+		// Calculate new context tokens after truncation from the effective API history.
+		// This keeps the post-truncation recount aligned with the same filtered history
+		// we actually send to the provider, including prior condense/truncation layers.
+		const effectiveMessages = getEffectiveApiHistory(truncationResult.messages)
 
 		// Include system prompt tokens so this value matches what we send to the API.
 		// Note: `prevContextTokens` is computed locally here (totalTokens + lastMessageTokens).