@oh-my-pi/pi-agent-core 17.2.1 → 17.2.2

package/CHANGELOG.md

@@ -2,6 +2,13 @@
 
 ## [Unreleased]
 
+## [17.2.2] - 2026-07-31
+
+### Fixed
+
+- Fixed an issue where response-only usage records were incorrectly treated as authoritative context anchors, while ensuring prompt and total-only provider telemetry remains preserved.
+- Fixed context compaction summaries growing excessively with large context windows by capping the summary output budget to 16,384 tokens, ensuring conversations are properly compressed rather than duplicated.
+
 ## [17.2.0] - 2026-07-30
 
 ### Fixed

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

@@ -52,18 +52,30 @@ export interface CompactionSettings {
 }
 /** Reserve applied when {@link CompactionSettings.reserveTokens} is unset. */
 export declare const DEFAULT_RESERVE_TOKENS = 16384;
+/**
+ * Hard ceiling on a generated compaction summary.
+ *
+ * The summary budget is `floor(0.8 * reserveTokens)`, and the effective reserve is
+ * at least 15% of the declared context window, so a 1M-token window authorizes a
+ * ~120k-token summary. At that size the model copies rather than compresses, and
+ * output is the slowest and most expensive token class. Capping absolutely keeps
+ * the compression ratio improving with window size instead of degrading. The value
+ * mirrors {@link DEFAULT_RESERVE_TOKENS} so this adds no new tuning constant.
+ */
+export declare const MAX_SUMMARY_TOKENS = 16384;
 export declare const DEFAULT_COMPACTION_SETTINGS: CompactionSettings;
 /** Whether a compaction candidate preserves provider-native transport under the effective settings. */
 export declare function shouldUseProviderNativeCompaction(model: Model, settings: Pick<CompactionSettings, "remoteEnabled" | "remoteStreamingV2Enabled">): boolean;
 /**
  * Calculate total context tokens from usage.
- * Uses the native totalTokens field when available, falls back to computing from components.
- * Provider-side orchestration tokens are billable but never replay into the
- * conversation prefix, so they are excluded from context sizing to keep
- * auto-compaction and context-promotion thresholds honest.
+ * Prefers an explicit provider-reported context occupancy when available.
+ * Otherwise uses totalTokens and falls back to computing from billable
+ * components. Provider-side orchestration tokens are billable but never replay
+ * into the conversation prefix, so they are excluded from context sizing.
  */
 export declare function calculateContextTokens(usage: Usage): number;
 export declare function calculatePromptTokens(usage: Usage): number;
+export declare function hasContextTokenUsage(usage: Usage): boolean;
 /**
  * Find the last non-aborted assistant message usage from session entries.
  */

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

@@ -65,6 +65,7 @@ export interface OpenAiRemoteCompactionResponse extends OpenAiRemoteCompactionPr
 export interface RemoteCompactionRequest {
     systemPrompt: string;
     prompt: string;
+    maxTokens?: number;
 }
 export interface RemoteCompactionResponse {
     summary: string;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-agent-core",
-	"version": "17.2.1",
+	"version": "17.2.2",
 	"description": "General-purpose agent with transport abstraction, state management, and attachment support",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -35,12 +35,12 @@
 		"fmt": "biome format --write ."
 	},
 	"dependencies": {
-		"@oh-my-pi/pi-ai": "17.2.1",
-		"@oh-my-pi/pi-catalog": "17.2.1",
-		"@oh-my-pi/pi-natives": "17.2.1",
-		"@oh-my-pi/pi-utils": "17.2.1",
-		"@oh-my-pi/pi-wire": "17.2.1",
-		"@oh-my-pi/snapcompact": "17.2.1",
+		"@oh-my-pi/pi-ai": "17.2.2",
+		"@oh-my-pi/pi-catalog": "17.2.2",
+		"@oh-my-pi/pi-natives": "17.2.2",
+		"@oh-my-pi/pi-utils": "17.2.2",
+		"@oh-my-pi/pi-wire": "17.2.2",
+		"@oh-my-pi/snapcompact": "17.2.2",
 		"@opentelemetry/api": "^1.9.1"
 	},
 	"devDependencies": {

package/src/compaction/compaction.ts

@@ -187,6 +187,18 @@ export interface CompactionSettings {
 /** Reserve applied when {@link CompactionSettings.reserveTokens} is unset. */
 export const DEFAULT_RESERVE_TOKENS = 16384;
 
+/**
+ * Hard ceiling on a generated compaction summary.
+ *
+ * The summary budget is `floor(0.8 * reserveTokens)`, and the effective reserve is
+ * at least 15% of the declared context window, so a 1M-token window authorizes a
+ * ~120k-token summary. At that size the model copies rather than compresses, and
+ * output is the slowest and most expensive token class. Capping absolutely keeps
+ * the compression ratio improving with window size instead of degrading. The value
+ * mirrors {@link DEFAULT_RESERVE_TOKENS} so this adds no new tuning constant.
+ */
+export const MAX_SUMMARY_TOKENS = DEFAULT_RESERVE_TOKENS;
+
 // reserveTokens is deliberately absent: an unset reserve is what marks it as
 // defaulted, which resolveBudgetReserveTokens needs to distinguish "user never
 // chose a reserve" from "user explicitly configured the default value".
@@ -221,12 +233,15 @@ export function shouldUseProviderNativeCompaction(
 
 /**
  * Calculate total context tokens from usage.
- * Uses the native totalTokens field when available, falls back to computing from components.
- * Provider-side orchestration tokens are billable but never replay into the
- * conversation prefix, so they are excluded from context sizing to keep
- * auto-compaction and context-promotion thresholds honest.
+ * Prefers an explicit provider-reported context occupancy when available.
+ * Otherwise uses totalTokens and falls back to computing from billable
+ * components. Provider-side orchestration tokens are billable but never replay
+ * into the conversation prefix, so they are excluded from context sizing.
  */
 export function calculateContextTokens(usage: Usage): number {
+	if (usage.contextTokens !== undefined) {
+		return Math.max(0, usage.contextTokens);
+	}
 	const orchestration = usage.orchestration;
 	const orchestrationTotal = orchestration
 		? (orchestration.input ?? 0) + (orchestration.output ?? 0) + (orchestration.cacheRead ?? 0)
@@ -236,6 +251,9 @@ export function calculateContextTokens(usage: Usage): number {
 }
 
 export function calculatePromptTokens(usage: Usage): number {
+	if (usage.contextTokens !== undefined) {
+		return Math.max(0, usage.contextTokens);
+	}
 	const promptTokens = usage.input + usage.cacheRead + usage.cacheWrite;
 	if (promptTokens > 0) {
 		return promptTokens;
@@ -243,6 +261,14 @@ export function calculatePromptTokens(usage: Usage): number {
 	return calculateContextTokens(usage);
 }
 
+export function hasContextTokenUsage(usage: Usage): boolean {
+	return (
+		(usage.contextTokens ?? 0) > 0 ||
+		usage.input + usage.cacheRead + usage.cacheWrite > 0 ||
+		calculateContextTokens(usage) > usage.output
+	);
+}
+
 /**
  * Get usage from an assistant message if available.
  * Skips aborted and error messages as they don't have valid usage data.
@@ -842,7 +868,7 @@ export async function generateSummary(
 	previousSummary?: string,
 	options?: SummaryOptions,
 ): Promise<string> {
-	const maxTokens = Math.floor(0.8 * reserveTokens);
+	const maxTokens = Math.min(Math.floor(0.8 * reserveTokens), MAX_SUMMARY_TOKENS);
 
 	// Use update prompt if we have a previous summary, otherwise initial prompt
 	let basePrompt = previousSummary ? UPDATE_SUMMARIZATION_PROMPT : SUMMARIZATION_PROMPT;
@@ -881,7 +907,7 @@ export async function generateSummary(
 			key =>
 				requestRemoteCompaction(
 					endpoint,
-					{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, prompt: promptText },
+					{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, prompt: promptText, maxTokens },
 					signal,
 					{ fetch: options.fetch, model, apiKey: key },
 				),
@@ -1086,7 +1112,7 @@ async function generateShortSummary(
 			key =>
 				requestRemoteCompaction(
 					endpoint,
-					{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, prompt: promptText },
+					{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, prompt: promptText, maxTokens },
 					signal,
 					{ fetch: options?.fetch, model, apiKey: key },
 				),
@@ -1657,7 +1683,7 @@ async function generateTurnPrefixSummary(
 	signal?: AbortSignal,
 	options?: SummaryOptions,
 ): Promise<string> {
-	const maxTokens = Math.floor(0.5 * reserveTokens); // Smaller budget for turn prefix
+	const maxTokens = Math.min(Math.floor(0.5 * reserveTokens), MAX_SUMMARY_TOKENS); // Smaller budget for turn prefix
 
 	const llmMessages = (options?.convertToLlm ?? defaultConvertToLlm)(messages);
 	const conversationText = serializeConversationForSummary(llmMessages, preferredDialect(model.id));

package/src/compaction/openai.ts

@@ -239,6 +239,7 @@ export interface OpenAiRemoteCompactionResponse extends OpenAiRemoteCompactionPr
 export interface RemoteCompactionRequest {
 	systemPrompt: string;
 	prompt: string;
+	maxTokens?: number;
 }
 
 export interface RemoteCompactionResponse {
@@ -928,8 +929,9 @@ export async function requestRemoteCompaction(
 					{ role: "user", content: request.prompt },
 				],
 				stream: false,
+				max_tokens: request.maxTokens,
 			}
-		: { systemPrompt: request.systemPrompt, prompt: request.prompt };
+		: { systemPrompt: request.systemPrompt, prompt: request.prompt, maxTokens: request.maxTokens };
 
 	const response = await (opts?.fetch ?? fetch)(endpoint, {
 		method: "POST",