@oh-my-pi/pi-agent-core 18.18.0 → 18.18.2

package/CHANGELOG.md

@@ -2,6 +2,18 @@
 
 ## [Unreleased]
 
+## [18.18.2] - 2026-11-08
+
+### Added
+
+- Added `canReplayRemoteCompactionFrom()` to determine whether a model can replay provider-native compaction produced by another model.
+
+### Fixed
+
+- Fixed GenAI message-content capture configured in omp's own `.env` files so it continues to apply when those values are no longer exposed through the process environment.
+- Fixed compaction preparation for remote OpenAI compaction that the active model cannot replay, ensuring the content is expanded and remains available for compaction.
+- Fixed generic remote compaction requests to `/chat/completions` so gateway sessions retain the correct per-request model attribution.
+
 ## [18.17.0] - 2026-11-06
 
 ### Added

package/dist/types/agent.d.ts

@@ -460,6 +460,8 @@ export declare class Agent {
      *  recompute their own snapshot from `peekSteeringQueue()`/`peekFollowUpQueue()`
      *  (or a higher-level view) on notification. */
     onQueueChange(listener: () => void): () => void;
+    /** Notify a session when this agent parks or resumes at a safe process-pause boundary. */
+    onPauseChange(listener: (paused: boolean) => void): () => void;
     /** Register an independently removable hook that runs before queued messages are consumed. */
     addBeforeQueuedMessageDequeueHook(hook: (signal?: AbortSignal) => Promise<void> | void): () => void;
     /** Register an independently removable hook that runs immediately before each model call, with the

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

@@ -205,6 +205,8 @@ export declare const AUTO_HANDOFF_THRESHOLD_FOCUS: string;
 export interface SummaryOptions {
     promptOverride?: string;
     extraContext?: string[];
+    /** Additional summary context fetched only when a compaction request consumes it. */
+    loadExtraContext?: () => Promise<string[] | undefined>;
     remoteEndpoint?: string;
     /** Stable system-prompt segments from the live turn, preserved for provider cache reuse. */
     remoteSystemPrompt?: string[];
@@ -264,6 +266,8 @@ export interface SummaryOptions {
     tools?: Tool[];
     /** Optional fetch implementation threaded into remote compaction calls. */
     fetch?: FetchImpl;
+    /** Inspect or replace a generic remote-compaction payload before it is sent. */
+    onPayload?: SimpleStreamOptions["onPayload"];
     /**
      * Optional completion transport override for host-level request wrappers
      * (e.g. the coding-agent provider-concurrency limiter). When provided,
@@ -378,6 +382,13 @@ export interface CompactionPreparation {
  * Responses-family replay. Local summaries have no provider restriction.
  */
 export declare function canReplayRemoteCompaction(preserveData: Record<string, unknown> | undefined, activeModel: Model): boolean;
+/**
+ * Whether `reader` can replay the provider-native compaction `compactor` would
+ * write. Only the OpenAI Responses lanes (the model's own API or its
+ * `remoteCompaction.api` override) store a provider-bound payload; the
+ * Anthropic lane and local methods leave a summary any model reads.
+ */
+export declare function canReplayRemoteCompactionFrom(compactor: Model, reader: Model): boolean;
 /**
  * Whether compaction preparation may reuse a native boundary instead of
  * re-expanding its original messages. This is deliberately stricter than normal

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

@@ -14,7 +14,7 @@
  *   summarization endpoints that accept `{ systemPrompt, prompt }` and reply
  *   with `{ summary, shortSummary? }`.
  */
-import type { CodexCompactionContext, FetchImpl, Message, Model, ProviderSessionState } from "@oh-my-pi/pi-ai/types";
+import type { CodexCompactionContext, FetchImpl, Message, Model, ProviderSessionState, SimpleStreamOptions } from "@oh-my-pi/pi-ai/types";
 import { Tokenizer } from "../tokenizer.js";
 export * from "./compaction-v2-streaming.js";
 export declare const OPENAI_REMOTE_COMPACTION_PRESERVE_KEY = "openaiRemoteCompaction";
@@ -153,14 +153,17 @@ export declare function requestOpenAiRemoteCompaction(model: Model, apiKey: stri
  * HTTP 400 `"'messages' is required"`, compaction silently fell back to
  * local summarization, and context grew unbounded.
  *
- * When `context.model` is provided the chat-completions body is tagged with
- * that model's wire id (llama.cpp requires the field) and `context.apiKey` is
- * forwarded as `Authorization: Bearer`. Callers wrap this in `withAuth` so
- * 401s force-refresh through the standard credential rotation policy.
+ * On chat-completions endpoints the model's request headers are resolved for
+ * each call, its wire id is sent (llama.cpp requires it), and `opts.apiKey` is
+ * forwarded as `Authorization: Bearer`. Custom summarizers receive no model
+ * credentials. `opts.onPayload` can inspect or replace either request body.
+ * Callers wrap this in `withAuth` so 401s force-refresh through the standard
+ * credential rotation policy.
  */
 export declare function requestRemoteCompaction(endpoint: string, request: RemoteCompactionRequest, signal?: AbortSignal, opts?: {
     fetch?: FetchImpl;
     timeoutMs?: number;
     model?: Model;
     apiKey?: string;
+    onPayload?: SimpleStreamOptions["onPayload"];
 }): Promise<RemoteCompactionResponse>;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-agent-core",
-	"version": "18.18.0",
+	"version": "18.18.2",
 	"description": "General-purpose agent with transport abstraction, state management, and attachment support",
 	"homepage": "https://omp.sh",
 	"author": {
@@ -38,16 +38,16 @@
 		"fmt": "oxfmt --no-error-on-unmatched-pattern 'src/**/*.{ts,tsx}' '{test,bench,examples,scripts}/**/*.ts' '*.ts'"
 	},
 	"dependencies": {
-		"@oh-my-pi/pi-ai": "18.18.0",
-		"@oh-my-pi/pi-catalog": "18.18.0",
-		"@oh-my-pi/pi-natives": "18.18.0",
-		"@oh-my-pi/pi-utils": "18.18.0",
-		"@oh-my-pi/pi-wire": "18.18.0",
-		"@oh-my-pi/snapcompact": "18.18.0",
+		"@oh-my-pi/pi-ai": "18.18.2",
+		"@oh-my-pi/pi-catalog": "18.18.2",
+		"@oh-my-pi/pi-natives": "18.18.2",
+		"@oh-my-pi/pi-utils": "18.18.2",
+		"@oh-my-pi/pi-wire": "18.18.2",
+		"@oh-my-pi/snapcompact": "18.18.2",
 		"@opentelemetry/api": "^1.9.1"
 	},
 	"devDependencies": {
-		"@oh-my-pi/omptype": "18.18.0",
+		"@oh-my-pi/omptype": "18.18.2",
 		"@opentelemetry/context-async-hooks": "2.9.0",
 		"@opentelemetry/sdk-trace-base": "2.9.0",
 		"@types/bun": "^1.3.14"

package/src/agent.ts

@@ -423,6 +423,7 @@ export class Agent {
 	#tokenizer = new Tokenizer(this.#state.model);
 	#listeners = new Set<(e: AgentEvent) => void>();
 	#queueListeners = new Set<() => void>();
+	#pauseListeners = new Set<(paused: boolean) => void>();
 	#abortController?: AbortController;
 	#convertToLlm: (messages: AgentMessage[]) => Message[] | Promise<Message[]>;
 	#transformContext?: (messages: AgentMessage[], signal?: AbortSignal) => Promise<AgentMessage[]>;
@@ -1014,6 +1015,12 @@ export class Agent {
 		return () => this.#queueListeners.delete(listener);
 	}
 
+	/** Notify a session when this agent parks or resumes at a safe process-pause boundary. */
+	onPauseChange(listener: (paused: boolean) => void): () => void {
+		this.#pauseListeners.add(listener);
+		return () => this.#pauseListeners.delete(listener);
+	}
+
 	/** Register an independently removable hook that runs before queued messages are consumed. */
 	addBeforeQueuedMessageDequeueHook(hook: (signal?: AbortSignal) => Promise<void> | void): () => void {
 		const registration = (signal?: AbortSignal) => hook(signal);
@@ -1378,6 +1385,19 @@ export class Agent {
 		}
 	}
 
+	/** Report a safe-boundary park/resume transition; a throwing observer never disrupts the loop. */
+	#emitPauseChanged(paused: boolean): void {
+		for (const listener of this.#pauseListeners) {
+			try {
+				listener(paused);
+			} catch (err) {
+				logger.warn("Agent pause listener threw", {
+					error: err instanceof Error ? err.message : String(err),
+				});
+			}
+		}
+	}
+
 	replaceQueues(steering: AgentMessage[], followUp: AgentMessage[]) {
 		this.#steeringQueue = steering.slice();
 		this.#followUpQueue = followUp.slice();
@@ -2078,7 +2098,9 @@ export class Agent {
 			onHarmonyLeak: this.#onHarmonyLeak,
 			onTurnEnd: (messages, signal, context) => this.#onTurnEnd?.(messages, signal, context),
 			onPauseChange: paused => {
+				if (this.#isPaused === paused) return;
 				this.#isPaused = paused;
+				this.#emitPauseChanged(paused);
 			},
 			getToolChoice,
 			softToolRequirementState: this.#softToolRequirementState,

package/src/compaction/compaction.ts

@@ -691,6 +691,8 @@ function shouldRetryHandoffWithAutoToolChoice(response: AssistantMessage): boole
 export interface SummaryOptions {
 	promptOverride?: string;
 	extraContext?: string[];
+	/** Additional summary context fetched only when a compaction request consumes it. */
+	loadExtraContext?: () => Promise<string[] | undefined>;
 	remoteEndpoint?: string;
 	/** Stable system-prompt segments from the live turn, preserved for provider cache reuse. */
 	remoteSystemPrompt?: string[];
@@ -755,6 +757,8 @@ export interface SummaryOptions {
 	tools?: Tool[];
 	/** Optional fetch implementation threaded into remote compaction calls. */
 	fetch?: FetchImpl;
+	/** Inspect or replace a generic remote-compaction payload before it is sent. */
+	onPayload?: SimpleStreamOptions["onPayload"];
 	/**
 	 * Optional completion transport override for host-level request wrappers
 	 * (e.g. the coding-agent provider-concurrency limiter). When provided,
@@ -1038,7 +1042,7 @@ async function summarizeConversationWindow(
 					endpoint,
 					{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, prompt: promptText, maxTokens },
 					signal,
-					{ fetch: options.fetch, model, apiKey: key },
+					{ fetch: options.fetch, model, apiKey: key, onPayload: options.onPayload },
 				),
 			{ signal, missingKeyMessage: "Remote compaction credentials unavailable" },
 		);
@@ -1269,7 +1273,7 @@ async function generateShortSummary(
 						endpoint,
 						{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, prompt: promptText, maxTokens },
 						signal,
-						{ fetch: options?.fetch, model, apiKey: key },
+						{ fetch: options?.fetch, model, apiKey: key, onPayload: options?.onPayload },
 					),
 				{ signal, missingKeyMessage: "Remote compaction credentials unavailable" },
 			);
@@ -1398,6 +1402,19 @@ export function canReplayRemoteCompaction(
 	return !remote || (remote.provider === activeModel.provider && isOpenAiRemoteCompactionApi(activeModel.api));
 }
 
+/**
+ * Whether `reader` can replay the provider-native compaction `compactor` would
+ * write. Only the OpenAI Responses lanes (the model's own API or its
+ * `remoteCompaction.api` override) store a provider-bound payload; the
+ * Anthropic lane and local methods leave a summary any model reads.
+ */
+export function canReplayRemoteCompactionFrom(compactor: Model, reader: Model): boolean {
+	return (
+		!isOpenAiRemoteCompactionApi(compactor.remoteCompaction?.api ?? compactor.api) ||
+		(compactor.provider === reader.provider && isOpenAiRemoteCompactionApi(reader.api))
+	);
+}
+
 /**
  * Whether compaction preparation may reuse a native boundary instead of
  * re-expanding its original messages. This is deliberately stricter than normal
@@ -1461,7 +1478,12 @@ export function prepareCompaction(
 ): CompactionPreparation | undefined {
 	const lastEntry = pathEntries[pathEntries.length - 1];
 	// A speculative native record may leave uncovered messages before the record.
-	if (lastEntry?.type === "compaction" && !lastEntry.providerReplayThroughEntryId) {
+	// A trailing boundary this preparation re-expands covers nothing yet.
+	if (
+		lastEntry?.type === "compaction" &&
+		!lastEntry.providerReplayThroughEntryId &&
+		(!activeModel || remotePreserveReusable(lastEntry.preserveData, activeModel, settings))
+	) {
 		return undefined;
 	}
 
@@ -1776,11 +1798,20 @@ export async function compact(
 		codexCompaction: options?.codexCompaction,
 		tools: options?.tools,
 		fetch: options?.fetch,
+		onPayload: options?.onPayload,
 		completeImpl: options?.completeImpl,
 		// The caller's opt-out must reach every summarization oneshot, otherwise
 		// an outer retry loop multiplies with the inner one (see SummaryOptions).
 		oneshotRetry: options?.oneshotRetry,
 	};
+	let extraContextLoaded = false;
+	const ensureExtraContext = async (): Promise<void> => {
+		if (extraContextLoaded || !options?.loadExtraContext) return;
+		const context = await options.loadExtraContext();
+		signal?.throwIfAborted();
+		if (context?.length) summaryOptions.extraContext = [...(summaryOptions.extraContext ?? []), ...context];
+		extraContextLoaded = true;
+	};
 
 	// An archive keeps the reasoning policy of the model it was rendered for;
 	// Claude refuses inputs that reproduce reasoning (issue #6093), so a
@@ -2214,6 +2245,7 @@ export async function compact(
 			].filter(files => files.after >= filesDueBefore);
 			if (carried.length > 0) nativeRetainedFiles = carried;
 		}
+		await ensureExtraContext();
 		try {
 			// The live turn's own system prompt, wire tools and transformed
 			// history when the host can build them: kept thinking remains valid
@@ -2274,6 +2306,7 @@ export async function compact(
 	if (!usedRemoteCompaction && nativeCompactionError !== undefined && !summaryOptions.remoteEndpoint) {
 		throw new NativeCompactionError(nativeCompactionError);
 	}
+	if (!usedRemoteCompaction && nativeSummary === undefined) await ensureExtraContext();
 
 	// Generate summaries (can be parallel if both needed) and merge into one
 	let summary: string;
@@ -2332,11 +2365,7 @@ export async function compact(
 
 	const shortSummary = usedRemoteCompaction
 		? "Remote compaction"
-		: await generateShortSummary(recentMessages, summary, model, reserveTokens, apiKey, signal, {
-				...summaryOptions,
-				extraContext: options?.extraContext,
-				thinkingLevel: options?.thinkingLevel,
-			});
+		: await generateShortSummary(recentMessages, summary, model, reserveTokens, apiKey, signal, summaryOptions);
 
 	// Compute file lists and append to summary
 	const { readFiles, modifiedFiles } = snapcompact.computeFileLists(fileOps);

package/src/compaction/openai.ts

@@ -39,6 +39,7 @@ import type {
 	Message,
 	Model,
 	ProviderSessionState,
+	SimpleStreamOptions,
 } from "@oh-my-pi/pi-ai/types";
 import {
 	dropMalformedOpenAIResponsesToolCalls,
@@ -59,7 +60,16 @@ import {
 	OPENAI_HEADER_VALUES,
 	OPENAI_HEADERS,
 } from "@oh-my-pi/pi-catalog/wire/codex";
-import { $env, isRecord, logger, prompt, ptree, stringifyJson, structuredCloneJSON } from "@oh-my-pi/pi-utils";
+import {
+	$env,
+	isRecord,
+	logger,
+	prompt,
+	ptree,
+	stringifyJson,
+	structuredCloneJSON,
+	untilAborted,
+} from "@oh-my-pi/pi-utils";
 import { dataUrlImageSize } from "../image-tokens";
 import { Tokenizer } from "../tokenizer";
 import { appendAzureApiVersion, resolveAzureOpenAiBaseUrl } from "./azure-openai-endpoint";
@@ -1058,16 +1068,24 @@ export async function requestOpenAiRemoteCompaction(
  * HTTP 400 `"'messages' is required"`, compaction silently fell back to
  * local summarization, and context grew unbounded.
  *
- * When `context.model` is provided the chat-completions body is tagged with
- * that model's wire id (llama.cpp requires the field) and `context.apiKey` is
- * forwarded as `Authorization: Bearer`. Callers wrap this in `withAuth` so
- * 401s force-refresh through the standard credential rotation policy.
+ * On chat-completions endpoints the model's request headers are resolved for
+ * each call, its wire id is sent (llama.cpp requires it), and `opts.apiKey` is
+ * forwarded as `Authorization: Bearer`. Custom summarizers receive no model
+ * credentials. `opts.onPayload` can inspect or replace either request body.
+ * Callers wrap this in `withAuth` so 401s force-refresh through the standard
+ * credential rotation policy.
  */
 export async function requestRemoteCompaction(
 	endpoint: string,
 	request: RemoteCompactionRequest,
 	signal?: AbortSignal,
-	opts?: { fetch?: FetchImpl; timeoutMs?: number; model?: Model; apiKey?: string },
+	opts?: {
+		fetch?: FetchImpl;
+		timeoutMs?: number;
+		model?: Model;
+		apiKey?: string;
+		onPayload?: SimpleStreamOptions["onPayload"];
+	},
 ): Promise<RemoteCompactionResponse> {
 	let endpointPath = endpoint;
 	try {
@@ -1076,13 +1094,22 @@ export async function requestRemoteCompaction(
 		// Keep the raw endpoint for relative/custom fetch implementations.
 	}
 	const isChatCompletions = /\/chat\/completions\/?$/.test(endpointPath);
-	const headers: Record<string, string> = { "content-type": "application/json" };
+	const headers = new Headers({ "content-type": "application/json" });
 	if (isChatCompletions) {
-		if (opts?.apiKey) headers.Authorization = `Bearer ${opts.apiKey}`;
-		if (opts?.model?.headers) Object.assign(headers, opts.model.headers);
+		if (opts?.apiKey) headers.set("Authorization", `Bearer ${opts.apiKey}`);
+		const model = opts?.model;
+		const resolveHeaders = model?.resolveHeaders;
+		const configuredHeaders = resolveHeaders
+			? await untilAborted(signal, () => resolveHeaders(signal))
+			: model?.headers;
+		signal?.throwIfAborted();
+		// `set` replaces case-insensitively, so a configured `authorization` overrides the default bearer.
+		if (configuredHeaders) {
+			for (const name in configuredHeaders) headers.set(name, configuredHeaders[name]!);
+		}
 	}
 
-	const body: Record<string, unknown> = isChatCompletions
+	let body: unknown = isChatCompletions
 		? {
 				model: opts?.model ? resolveOpenAiCompactModel(opts.model) : undefined,
 				messages: [
@@ -1093,6 +1120,8 @@ export async function requestRemoteCompaction(
 				max_tokens: request.maxTokens,
 			}
 		: { systemPrompt: request.systemPrompt, prompt: request.prompt, maxTokens: request.maxTokens };
+	const replacement = await opts?.onPayload?.(body, opts.model, signal);
+	if (replacement !== undefined) body = replacement;
 
 	const response = await (opts?.fetch ?? fetch)(endpoint, {
 		method: "POST",

package/src/telemetry.ts

@@ -39,6 +39,7 @@ import {
 	type ToolChoice,
 	type Usage,
 } from "@oh-my-pi/pi-ai";
+import { $otelEnv } from "@oh-my-pi/pi-utils";
 import {
 	type Attributes,
 	type AttributeValue,
@@ -453,7 +454,7 @@ export function resolveTelemetry(
 let contentCaptureEnvCache: ResolvedTelemetryContentCapture | undefined;
 function readContentCaptureEnv(): ResolvedTelemetryContentCapture {
 	if (contentCaptureEnvCache !== undefined) return contentCaptureEnvCache;
-	const raw = process.env[CONTENT_CAPTURE_ENV];
+	const raw = $otelEnv(CONTENT_CAPTURE_ENV);
 	if (!raw) {
 		contentCaptureEnvCache = "none";
 		return "none";

package/src/tokenizer.ts

@@ -3,16 +3,16 @@ import { renderDemotedThinking } from "@oh-my-pi/pi-ai/dialect/demotion";
 import { isFactoryDroidGeminiWire, keepsFactoryGeminiReasoning } from "@oh-my-pi/pi-ai/providers/factory-droid/gemini";
 import { ollamaChatDropsOwnThinking } from "@oh-my-pi/pi-ai/providers/ollama";
 import { completionsDropsOwnThinking, isLiteLLMThinkingSignature } from "@oh-my-pi/pi-ai/providers/openai-completions";
-import { isColdReplayableResponsesTurn } from "@oh-my-pi/pi-ai/providers/openai-shared";
+import {
+	getNativeResponsesHistoryItems,
+	getReplayableNativeResponsesHistoryItems,
+	replaysNativeResponsesTurn,
+} from "@oh-my-pi/pi-ai/providers/openai-shared";
 import {
 	dropsForeignReasoningPayload,
 	isSameModelTurn,
 	shouldDropTruncatedThinkingOnlyAssistant,
 } from "@oh-my-pi/pi-ai/providers/transform-messages";
-import {
-	getOpenAIResponsesHistoryItems,
-	sanitizeOpenAIResponsesAssistantHistoryItemsForReplay,
-} from "@oh-my-pi/pi-ai/utils";
 import { resolveThinkingReplayTokens, type ThinkingReplayTokens } from "@oh-my-pi/pi-catalog/compat/thinking-replay";
 import type { ModelTokenizer } from "@oh-my-pi/pi-catalog/types";
 import * as natives from "@oh-my-pi/pi-natives";
@@ -595,15 +595,19 @@ export class Tokenizer {
 	 * its reasoning, so those send no reasoning and no signature (0).
 	 */
 	#replayedReasoningTokens(message: AssistantMessage): number | null | undefined {
-		const reader = this.#reasoningReader;
-		if (!reader || message.api !== reader.api || message.model !== reader.id) return undefined;
-		const items = getOpenAIResponsesHistoryItems(message.providerPayload, reader.provider, message.provider);
-		if (!items) return undefined;
-		if (!items.some(item => item.type === "reasoning") || shouldDropTruncatedThinkingOnlyAssistant(message)) return 0;
-		const replay = sanitizeOpenAIResponsesAssistantHistoryItemsForReplay(items);
-		if (!replay?.some(item => item.type === "reasoning")) return 0;
-		// Cold, only plaintext reasoning bound to no server state replays natively.
-		if (!this.nativeHistoryReplay && !isColdReplayableResponsesTurn(replay)) return 0;
+		// A payload this reader does not own is not native history at all: the turn
+		// is rebuilt from its blocks, and its signatures are priced below.
+		if (!getNativeResponsesHistoryItems(message, this.#reasoningReader)) return undefined;
+		// The request decides whether the payload goes on the wire; `/shake` reads
+		// that same decision, so a kept turn and a billed turn always agree.
+		const replay = getReplayableNativeResponsesHistoryItems(message, this.#reasoningReader);
+		if (!replay || !replaysNativeResponsesTurn(replay, this.nativeHistoryReplay)) return 0;
+		// Then the amount: nothing reasoning-shaped survives to price, or the turn is
+		// dropped whole as truncated thinking, so it sends no reasoning and no
+		// signature either. An answer-only payload replays with neither.
+		if (!replay.some(item => item.type === "reasoning") || shouldDropTruncatedThinkingOnlyAssistant(message)) {
+			return 0;
+		}
 		return message.usage?.reasoningTokens ?? null;
 	}