@sayknow-cli/agent-core 0.2.7 → 0.3.1

package/CHANGELOG.md

@@ -2,6 +2,12 @@
 
 ## [Unreleased]
 
+## [0.7.2] - 2026-06-24
+
+### Fixed
+
+- Reserved headroom when trimming OpenAI remote compaction input so `/responses/compact` requests stay below the model context window instead of filling the entire window.
+
 ## [0.6.2] - 2026-06-19
 
 ### Changed

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

@@ -43,6 +43,7 @@ export declare function getPreservedOpenAiRemoteCompactionData(preserveData: Rec
 export declare function withOpenAiRemoteCompactionPreserveData(preserveData: Record<string, unknown> | undefined, remoteCompaction: OpenAiRemoteCompactionPreserveData | undefined): Record<string, unknown> | undefined;
 export declare function estimateOpenAiCompactInputTokens(input: Array<Record<string, unknown>>, instructions: string): number;
 export declare function trimOpenAiCompactInput(input: Array<Record<string, unknown>>, contextWindow: number, instructions: string): Array<Record<string, unknown>>;
+export declare function resolveOpenAiCompactInputBudget(contextWindow: number, maxOutputTokens?: number): number;
 /**
  * Build the OpenAI Responses-API native history array from LLM messages.
  *

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@sayknow-cli/agent-core",
-	"version": "0.2.7",
+	"version": "0.3.1",
 	"description": "General-purpose agent with transport abstraction, state management, and attachment support",
 	"homepage": "https://github.com/jaybeyond/Sayknow_CLI",
 	"author": "jaybeyond",
@@ -35,9 +35,9 @@
 		"fmt": "biome format --write ."
 	},
 	"dependencies": {
-		"@sayknow-cli/ai": "0.2.7",
-		"@sayknow-cli/natives": "0.2.7",
-		"@sayknow-cli/utils": "0.2.7",
+		"@sayknow-cli/ai": "0.3.1",
+		"@sayknow-cli/natives": "0.3.1",
+		"@sayknow-cli/utils": "0.3.1",
 		"@opentelemetry/api": "^1.9.0"
 	},
 	"devDependencies": {

package/src/agent-loop.ts

@@ -1084,6 +1084,17 @@ async function executeToolCalls(
 
 		await runInActiveSpan(toolSpan, async () => {
 			try {
+				if (toolCall.incompleteArguments) {
+					// The provider flagged this call's argument JSON as truncated
+					// (the model hit its output-token limit mid-call). Executing the
+					// best-effort partial parse would run the tool on wrong input, so
+					// reject with a retryable, actionable error instead.
+					throw new Error(
+						`Tool call "${toolCall.name}" was cut off before its arguments finished streaming ` +
+							`(the response hit its output token limit). The partial arguments cannot be executed. ` +
+							`Re-issue the call with complete arguments, splitting the work into smaller steps if needed.`,
+					);
+				}
 				if (!tool) throw new Error(`Tool ${toolCall.name} not found`);
 
 				let effectiveArgs: Record<string, unknown>;

package/src/compaction/openai.ts

@@ -237,6 +237,12 @@ export function trimOpenAiCompactInput(
 	return trimmed;
 }
 
+export function resolveOpenAiCompactInputBudget(contextWindow: number, maxOutputTokens = 0): number {
+	if (contextWindow <= 0) return 0;
+	const reservedTokens = Math.max(Math.floor(contextWindow * 0.15), maxOutputTokens, 1);
+	return Math.max(1, contextWindow - reservedTokens);
+}
+
 function collectKnownOpenAiCallIds(items: Array<Record<string, unknown>>): Set<string> {
 	const knownCallIds = new Set<string>();
 	for (const item of items) {
@@ -473,7 +479,11 @@ export async function requestOpenAiRemoteCompaction(
 	const endpoint = resolveOpenAiCompactEndpoint(model, options?.authCredentialType);
 	const request: OpenAiRemoteCompactionRequest = {
 		model: model.id,
-		input: trimOpenAiCompactInput(compactInput, model.contextWindow, instructions),
+		input: trimOpenAiCompactInput(
+			compactInput,
+			resolveOpenAiCompactInputBudget(model.contextWindow, model.maxTokens),
+			instructions,
+		),
 		instructions,
 	};
 	const headers: Record<string, string> = {