fix(core): restore compaction input budget contract

sdk/packages/core/src/extensions/context/compaction.test.ts

@@ -90,7 +90,6 @@ function runForcedBasicCompaction(
 				info: { id: "mock-model", maxInputTokens: targetTokens },
 			},
 			maxInputTokens: targetTokens,
-			usableBudgetTokens: targetTokens,
 			triggerTokens: targetTokens,
 			thresholdRatio: 1,
 			utilizationRatio: 2,
@@ -266,7 +265,6 @@ describe("createContextCompactionPrepareTurn", () => {
 					info: { id: "mock-model", maxInputTokens: 100_000 },
 				},
 				maxInputTokens: 100_000,
-				usableBudgetTokens: 100_000,
 				triggerTokens: 100_000,
 				thresholdRatio: 1,
 				utilizationRatio: 0.1,
@@ -490,7 +488,6 @@ describe("createContextCompactionPrepareTurn", () => {
 					info: { id: "mock-model", maxInputTokens: 1_000 },
 				},
 				maxInputTokens: 1_000,
-				usableBudgetTokens: 1_000,
 				triggerTokens: 900,
 				targetTokens: 100,
 				thresholdRatio: 0.9,
@@ -1416,9 +1413,9 @@ describe("createContextCompactionPrepareTurn", () => {
 
 		expect(compact).toHaveBeenCalledTimes(1);
 		const context = compact.mock.calls[0]?.[0];
-		expect(context?.maxInputTokens).toBe(400_000);
+		expect(context?.maxInputTokens).toBe(272_000);
 		expect(context?.triggerTokens).toBe(244_800);
-		expect(context?.thresholdRatio).toBe(244_800 / 400_000);
+		expect(context?.thresholdRatio).toBe(0.9);
 		expect(result?.messages).toEqual([
 			{ role: "user", content: "Compacted by derived input budget" },
 		]);
@@ -1466,10 +1463,9 @@ describe("createContextCompactionPrepareTurn", () => {
 
 		expect(compact).toHaveBeenCalledTimes(1);
 		const context = compact.mock.calls[0]?.[0];
-		expect(context?.maxInputTokens).toBe(40_960);
-		expect(context?.usableBudgetTokens).toBe(24_576);
+		expect(context?.maxInputTokens).toBe(24_576);
 		expect(context?.triggerTokens).toBe(22_118);
-		expect(context?.thresholdRatio).toBe(22_118 / 40_960);
+		expect(context?.thresholdRatio).toBe(22_118 / 24_576);
 	});
 
 	it("caps shared-context output reserve at half the context window", async () => {
@@ -1514,7 +1510,7 @@ describe("createContextCompactionPrepareTurn", () => {
 
 		expect(compact).toHaveBeenCalledTimes(1);
 		const context = compact.mock.calls[0]?.[0];
-		expect(context?.maxInputTokens).toBe(40_960);
+		expect(context?.maxInputTokens).toBe(20_480);
 		expect(context?.triggerTokens).toBe(18_432);
 	});
 
@@ -1561,7 +1557,7 @@ describe("createContextCompactionPrepareTurn", () => {
 		expect(compact).toHaveBeenCalledTimes(1);
 		const context = compact.mock.calls[0]?.[0];
 		expect(context?.triggerTokens).toBe(19_660);
-		expect(context?.thresholdRatio).toBe(19_660 / 40_960);
+		expect(context?.thresholdRatio).toBe(19_660 / 24_576);
 	});
 
 	it("applies explicit reserve tokens to the usable budget", async () => {
@@ -1607,7 +1603,7 @@ describe("createContextCompactionPrepareTurn", () => {
 		expect(compact).toHaveBeenCalledTimes(1);
 		const context = compact.mock.calls[0]?.[0];
 		expect(context?.triggerTokens).toBe(20_480);
-		expect(context?.thresholdRatio).toBe(0.5);
+		expect(context?.thresholdRatio).toBe(20_480 / 24_576);
 	});
 
 	it("uses the lower split input budget when it is below context-derived input budget", async () => {
@@ -2013,7 +2009,6 @@ describe("createContextCompactionPrepareTurn", () => {
 					info: { id: "mock-model", maxInputTokens: 100 },
 				},
 				maxInputTokens: 100,
-				usableBudgetTokens: 100,
 				triggerTokens: 100,
 				thresholdRatio: 1,
 				utilizationRatio: 0.1,

sdk/packages/core/src/extensions/context/compaction.ts

@@ -100,7 +100,6 @@ interface CompactionBudget {
 	outputReserveTokens: number;
 	triggerTokens: number;
 	thresholdRatio: number;
-	usableThresholdRatio: number;
 	ceilingSource: CompactionBudgetSource;
 	reserveCapped: boolean;
 }
@@ -166,8 +165,6 @@ function resolveCompactionBudget(input: {
 		outputReserveTokens,
 		triggerTokens,
 		thresholdRatio:
-			ceilingTokens > 0 ? triggerTokens / ceilingTokens : 0,
-		usableThresholdRatio:
 			usableBudgetTokens > 0 ? triggerTokens / usableBudgetTokens : 0,
 		ceilingSource,
 		reserveCapped:
@@ -371,7 +368,7 @@ export function createContextCompactionPrepareTurn(
 			contextWindow: context.model.info?.contextWindow,
 			modelMaxTokens: context.model.info?.maxTokens,
 		});
-		const maxInputTokens = compactionBudget.ceilingTokens;
+		const maxInputTokens = compactionBudget.usableBudgetTokens;
 		const shouldCompact = inputTokens > compactionBudget.triggerTokens;
 		config.logger?.debug("Context compaction diagnostics", {
 			mode,
@@ -381,13 +378,12 @@ export function createContextCompactionPrepareTurn(
 			modelId: config.modelId,
 			inputTokens,
 			maxInputTokens,
-			usableBudgetTokens: compactionBudget.usableBudgetTokens,
+			ceilingTokens: compactionBudget.ceilingTokens,
 			outputReserveTokens: compactionBudget.outputReserveTokens,
 			ceilingSource: compactionBudget.ceilingSource,
 			reserveCapped: compactionBudget.reserveCapped,
 			triggerTokens: compactionBudget.triggerTokens,
 			thresholdRatio: compactionBudget.thresholdRatio,
-			usableThresholdRatio: compactionBudget.usableThresholdRatio,
 			shouldCompact,
 			messageCount: context.messages.length,
 			apiMessageCount: context.apiMessages.length,
@@ -425,7 +421,6 @@ export function createContextCompactionPrepareTurn(
 			messages: context.messages,
 			model: context.model,
 			maxInputTokens,
-			usableBudgetTokens: compactionBudget.usableBudgetTokens,
 			triggerTokens: targetState.triggerTokens,
 			targetTokens,
 			thresholdRatio: targetState.thresholdRatio,
@@ -442,7 +437,7 @@ export function createContextCompactionPrepareTurn(
 				iteration: context.iteration,
 				triggerTokens: targetState.triggerTokens,
 				maxInputTokens,
-				usableBudgetTokens: compactionBudget.usableBudgetTokens,
+				ceilingTokens: compactionBudget.ceilingTokens,
 				outputReserveTokens: compactionBudget.outputReserveTokens,
 				ceilingSource: compactionBudget.ceilingSource,
 			},
@@ -486,7 +481,7 @@ export function createContextCompactionPrepareTurn(
 				severity: "info",
 				strategy: strategy,
 				maxInputTokens,
-				usableBudgetTokens: compactionBudget.usableBudgetTokens,
+				ceilingTokens: compactionBudget.ceilingTokens,
 				outputReserveTokens: compactionBudget.outputReserveTokens,
 				ceilingSource: compactionBudget.ceilingSource,
 				reserveCapped: compactionBudget.reserveCapped,

sdk/packages/core/src/types/config.ts

@@ -67,20 +67,15 @@ export interface CoreCompactionContext {
 		info?: ModelInfo;
 	};
 	/**
-	 * Full input ceiling for the model (context window or explicit override),
-	 * before any output-token reserve is subtracted.
+	 * Usable input budget after reserving any shared context window space needed
+	 * for model output.
 	 */
 	maxInputTokens: number;
-	/**
-	 * Tokens the input may actually occupy: `maxInputTokens` minus the output
-	 * reserve held back on shared-context models. Compact below this value.
-	 */
-	usableBudgetTokens: number;
 	triggerTokens: number;
 	targetTokens?: number;
 	/**
-	 * Effective trigger point as a fraction of `maxInputTokens`
-	 * (`triggerTokens / maxInputTokens`), not the raw configured ratio.
+	 * Effective trigger point as a fraction of the usable input budget
+	 * (`triggerTokens / maxInputTokens`).
 	 */
 	thresholdRatio: number;
 	utilizationRatio: number;