@gajae-code/agent-core 0.4.2 → 0.4.4

package/CHANGELOG.md

@@ -2,6 +2,12 @@
 
 ## [Unreleased]
 
+## [0.4.3] - 2026-06-10
+
+### Fixed
+
+- Separated a model's total context window from its safe input/prompt-packing budget in the compaction threshold. `effectiveReserveTokens`, `resolveThresholdTokens`, and `shouldCompact` now accept an optional `maxOutputTokens` and reserve at least that completion budget, so a large-output model (e.g. 400K context / 128K max output) caps input near 272K instead of 340K and cannot overflow the total window with reserved output ([#442](https://github.com/Yeachan-Heo/gajae-code/issues/442)).
+
 ## [0.4.0] - 2026-06-06
 
 ### Changed

package/dist/types/agent.d.ts

@@ -346,6 +346,12 @@ export declare class Agent {
      * Used by dequeue keybinding.
      */
     popLastFollowUp(): AgentMessage | undefined;
+    /** Remove queued steering+follow-up messages matching `predicate`, preserving order of the rest. */
+    removeQueuedMessages(predicate: (message: AgentMessage) => boolean): {
+        steering: number;
+        followUp: number;
+        total: number;
+    };
     clearMessages(): void;
     abort(): void;
     /**

package/dist/types/compaction/compaction.d.ts

@@ -50,14 +50,23 @@ export declare function calculatePromptTokens(usage: Usage): number;
  */
 export declare function getLastAssistantUsage(entries: SessionEntry[]): Usage | undefined;
 /**
- * Effective reserve: at least 15% of context window or the configured floor, whichever is larger.
+ * Effective reserve: the largest of 15% of the context window, the configured floor,
+ * and the model's reserved completion budget (`maxOutputTokens`).
+ *
+ * Reserving `maxOutputTokens` keeps the safe input/prompt-packing budget below the
+ * *total* context window for models whose completion reservation exceeds the 15%
+ * floor (e.g. a 400K-context model with 128K max output reserves 128K, not 60K, so
+ * input is capped near 272K instead of 340K).
  */
-export declare function effectiveReserveTokens(contextWindow: number, settings: CompactionSettings): number;
+export declare function effectiveReserveTokens(contextWindow: number, settings: CompactionSettings, maxOutputTokens?: number): number;
 /**
  * Check if compaction should trigger based on context usage.
+ *
+ * `maxOutputTokens` is the model's reserved completion budget; it is excluded from
+ * the safe input budget so prompt + reserved output cannot exceed the total window.
  */
-export declare function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean;
-export declare function resolveThresholdTokens(contextWindow: number, settings: CompactionSettings): number;
+export declare function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings, maxOutputTokens?: number): boolean;
+export declare function resolveThresholdTokens(contextWindow: number, settings: CompactionSettings, maxOutputTokens?: number): number;
 /**
  * Estimate token count for a message using cl100k_base via the native
  * tokenizer. This is not Anthropic's first-party tokenizer (Anthropic doesn't

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@gajae-code/agent-core",
-	"version": "0.4.2",
+	"version": "0.4.4",
 	"description": "General-purpose agent with transport abstraction, state management, and attachment support",
 	"homepage": "https://gaebal-gajae.dev",
 	"author": "Yeachan-Heo",
@@ -35,9 +35,9 @@
 		"fmt": "biome format --write ."
 	},
 	"dependencies": {
-		"@gajae-code/ai": "0.4.2",
-		"@gajae-code/natives": "0.4.2",
-		"@gajae-code/utils": "0.4.2",
+		"@gajae-code/ai": "0.4.4",
+		"@gajae-code/natives": "0.4.4",
+		"@gajae-code/utils": "0.4.4",
 		"@opentelemetry/api": "^1.9.0"
 	},
 	"devDependencies": {

package/src/agent.ts

@@ -943,6 +943,21 @@ export class Agent {
 		return this.#followUpQueue.pop();
 	}
 
+	/** Remove queued steering+follow-up messages matching `predicate`, preserving order of the rest. */
+	removeQueuedMessages(predicate: (message: AgentMessage) => boolean): {
+		steering: number;
+		followUp: number;
+		total: number;
+	} {
+		const beforeSteering = this.#steeringQueue.length;
+		const beforeFollowUp = this.#followUpQueue.length;
+		this.#steeringQueue = this.#steeringQueue.filter(m => !predicate(m));
+		this.#followUpQueue = this.#followUpQueue.filter(m => !predicate(m));
+		const steering = beforeSteering - this.#steeringQueue.length;
+		const followUp = beforeFollowUp - this.#followUpQueue.length;
+		return { steering, followUp, total: steering + followUp };
+	}
+
 	clearMessages() {
 		this.#state.messages = [];
 	}

package/src/compaction/compaction.ts

@@ -203,22 +203,44 @@ export function getLastAssistantUsage(entries: SessionEntry[]): Usage | undefine
 }
 
 /**
- * Effective reserve: at least 15% of context window or the configured floor, whichever is larger.
+ * Effective reserve: the largest of 15% of the context window, the configured floor,
+ * and the model's reserved completion budget (`maxOutputTokens`).
+ *
+ * Reserving `maxOutputTokens` keeps the safe input/prompt-packing budget below the
+ * *total* context window for models whose completion reservation exceeds the 15%
+ * floor (e.g. a 400K-context model with 128K max output reserves 128K, not 60K, so
+ * input is capped near 272K instead of 340K).
  */
-export function effectiveReserveTokens(contextWindow: number, settings: CompactionSettings): number {
-	return Math.max(Math.floor(contextWindow * 0.15), settings.reserveTokens);
+export function effectiveReserveTokens(
+	contextWindow: number,
+	settings: CompactionSettings,
+	maxOutputTokens = 0,
+): number {
+	return Math.max(Math.floor(contextWindow * 0.15), settings.reserveTokens, Math.max(0, maxOutputTokens));
 }
 
 /**
  * Check if compaction should trigger based on context usage.
+ *
+ * `maxOutputTokens` is the model's reserved completion budget; it is excluded from
+ * the safe input budget so prompt + reserved output cannot exceed the total window.
  */
-export function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean {
+export function shouldCompact(
+	contextTokens: number,
+	contextWindow: number,
+	settings: CompactionSettings,
+	maxOutputTokens = 0,
+): boolean {
 	if (!settings.enabled || settings.strategy === "off" || contextWindow <= 0) return false;
-	const thresholdTokens = resolveThresholdTokens(contextWindow, settings);
+	const thresholdTokens = resolveThresholdTokens(contextWindow, settings, maxOutputTokens);
 	return contextTokens > thresholdTokens;
 }
 
-export function resolveThresholdTokens(contextWindow: number, settings: CompactionSettings): number {
+export function resolveThresholdTokens(
+	contextWindow: number,
+	settings: CompactionSettings,
+	maxOutputTokens = 0,
+): number {
 	// Fixed token limit takes priority over percentage
 	const thresholdTokens = settings.thresholdTokens;
 	if (typeof thresholdTokens === "number" && Number.isFinite(thresholdTokens) && thresholdTokens > 0) {
@@ -229,7 +251,7 @@ export function resolveThresholdTokens(contextWindow: number, settings: Compacti
 	// Percentage-based threshold
 	const thresholdPercent = settings.thresholdPercent;
 	if (typeof thresholdPercent !== "number" || !Number.isFinite(thresholdPercent) || thresholdPercent <= 0) {
-		return contextWindow - effectiveReserveTokens(contextWindow, settings);
+		return contextWindow - effectiveReserveTokens(contextWindow, settings, maxOutputTokens);
 	}
 	const clampedThresholdPercent = Math.min(99, Math.max(1, thresholdPercent));
 	return Math.floor(contextWindow * (clampedThresholdPercent / 100));