fix: recover bounded context and implement engineering spec

Author: ponythewhite

package-lock.json

@@ -49,10 +49,10 @@
 	"engines": {
 		"node": "^22.12.0 || >=23.3.0"
 	},
-	"version": "1.0.7",
+	"version": "1.0.8",
 	"dependencies": {
 		"get-east-asian-width": "^1.6.0",
-		"@ponythewhite/base-context": "1.0.7"
+		"@ponythewhite/base-context": "1.0.8"
 	},
 	"overrides": {
 		"rimraf": "6.1.2",

package.json

@@ -1,5 +1,9 @@
 # Changelog
 
+## [1.0.8] - 2026-09-18
+
+- Allowed one deterministic context reprepare before the existing single request-recovery attempt, without restarting the invocation or completed tools.
+
 ## [1.0.7] - 2026-09-18
 
 - Updated lockstep packaging for the Base Context 1.0.7 session-resume fix.

packages/agent/CHANGELOG.md

@@ -1,6 +1,6 @@
 {
 	"name": "@ponythewhite/base-context-agent",
-	"version": "1.0.7",
+	"version": "1.0.8",
 	"description": "Synerise base-context agent runtime with context management and native recovery",
 	"type": "module",
 	"main": "./dist/index.js",
@@ -19,7 +19,7 @@
 		"prepublishOnly": "npm --prefix ../.. run build:source"
 	},
 	"dependencies": {
-		"@ponythewhite/base-context-ai": "1.0.7",
+		"@ponythewhite/base-context-ai": "1.0.8",
 		"typebox": "^1.3.9"
 	},
 	"keywords": [

packages/agent/package.json

@@ -597,6 +597,7 @@ async function streamAssistantResponse(
 	emit: AgentEventSink,
 	streamFn?: StreamFn,
 	allowRecovery = true,
+	allowReprepare = true,
 ): Promise<AssistantMessage> {
 	const build: { projection: AgentContextBuildResult } = { projection: undefined };
 	let addedPartial = false;
@@ -767,18 +768,22 @@ async function streamAssistantResponse(
 			isLocalRequestPreparationError(outcome.error) &&
 			config.recoverRequestPreparation
 		) {
-			let recovered: boolean;
+			let recovered: boolean | "reprepare";
 			try {
 				recovered = await config.recoverRequestPreparation(outcome.error, signal);
 			} catch (error) {
 				// Preserve the real owner error, including a post-ACK compaction failure.
 				markLocalRequestPreparationError(error);
 				throw error;
 			}
-			if (recovered) {
-				// Rebuild/adopt the committed owner view. Keep this invocation's output collector,
-				// completed tools and emitted messages; only this unsent request gets one reprepare.
-				return streamAssistantResponse(context, config, signal, emit, streamFn, false);
+			if (recovered === "reprepare" && allowReprepare) {
+				// A bounded deterministic change (for example, rebasing a TaskFrame) must
+				// not consume the one recovery available if the new request still cannot fit.
+				return streamAssistantResponse(context, config, signal, emit, streamFn, true, false);
+			}
+			if (recovered === true) {
+				// Keep this invocation's output collector, completed tools and emitted messages.
+				return streamAssistantResponse(context, config, signal, emit, streamFn, false, false);
 			}
 		}
 		throw outcome.error;

packages/agent/src/agent-loop.ts

@@ -210,8 +210,8 @@ export interface AgentLoopConfig extends SimpleStreamOptions {
 	/** Internal state mirror for adopted working sets. Awaited and excluded from provider options. */
 	onContextAdopted?: (messages: AgentMessage[]) => void | Promise<void>;
 
-	/** Native owner recovery after an unsent request's projection has been released. */
-	recoverRequestPreparation?: (error: unknown, signal?: AbortSignal) => Promise<boolean>;
+	/** Native owner recovery after release. One deterministic reprepare may precede one recovery. */
+	recoverRequestPreparation?: (error: unknown, signal?: AbortSignal) => Promise<boolean | "reprepare">;
 
 	/** Retry a settled provider failure within this invocation and its output allowance. */
 	recoverProviderFailure?: (message: AssistantMessage, signal?: AbortSignal) => Promise<boolean>;

packages/agent/src/types.ts

@@ -1,5 +1,16 @@
 # Changelog
 
+## [1.0.8] - 2026-09-18
+
+- Fixed Codex socket reuse after authentication or endpoint changes without interrupting active requests.
+- Fixed Anthropic tool schemas losing root constraints and definitions or sharing mutable schema data with callers.
+- Fixed Codex retry waits retaining abort listeners and ignoring the provider retry delay.
+- Fixed duplicate fallback item IDs when a Responses assistant message contains multiple unsigned text blocks.
+- Kept generated Responses item IDs stable for source-backed requests, with a fresh context epoch after restarting the session owner.
+- Respected disabled Codex prompt-cache affinity without changing account routing or connection reuse.
+- Added opt-in OpenAI token estimation and near-limit Responses input counting while preserving output reserves and control-call accounting.
+- Exposed conservative plaintext sizing through the existing request token meter.
+
 ## [1.0.7] - 2026-09-18
 
 - Updated lockstep packaging for the Base Context 1.0.7 session-resume fix.

packages/ai/CHANGELOG.md

@@ -1,6 +1,6 @@
 {
 	"name": "@ponythewhite/base-context-ai",
-	"version": "1.0.7",
+	"version": "1.0.8",
 	"description": "Synerise base-context unified LLM API and provider configuration",
 	"type": "module",
 	"main": "./dist/index.js",
@@ -80,6 +80,7 @@
 		"@mistralai/mistralai": "^2.2.0",
 		"typebox": "^1.3.9",
 		"chalk": "^5.6.2",
+		"js-tiktoken": "1.0.21",
 		"openai": "6.47.0",
 		"partial-json": "^0.1.7",
 		"proxy-agent": "^6.5.0",

packages/ai/package.json

@@ -1279,17 +1279,11 @@ function convertTools(
 	if (!tools) return [];
 
 	return tools.map((tool, index) => {
-		const schema = tool.parameters as { properties?: unknown; required?: string[] };
-
 		return {
 			name: isOAuthToken ? toClaudeCodeName(tool.name) : tool.name,
 			description: tool.description,
 			...(supportsEagerToolInputStreaming ? { eager_input_streaming: true } : {}),
-			input_schema: {
-				type: "object",
-				properties: schema.properties ?? {},
-				required: schema.required ?? [],
-			},
+			input_schema: structuredClone(tool.parameters) as Anthropic.Messages.Tool.InputSchema,
 			...(cacheControl && index === tools.length - 1 ? { cache_control: cacheControl } : {}),
 		};
 	});

packages/ai/src/providers/anthropic.ts

@@ -33,6 +33,7 @@ import type {
 	StreamOptions,
 	Usage,
 } from "../types.js";
+import { abortableDelay } from "../utils/abortable-delay.js";
 import {
 	appendAssistantMessageDiagnostic,
 	createAssistantMessageDiagnostic,
@@ -128,20 +129,6 @@ function codexTransportError(error: unknown): Error {
 	);
 }
 
-function sleep(ms: number, signal?: AbortSignal): Promise<void> {
-	return new Promise((resolve, reject) => {
-		if (signal?.aborted) {
-			reject(new Error("Request was aborted"));
-			return;
-		}
-		const timeout = setTimeout(resolve, ms);
-		signal?.addEventListener("abort", () => {
-			clearTimeout(timeout);
-			reject(new Error("Request was aborted"));
-		});
-	});
-}
-
 export const streamOpenAICodexResponses: StreamFunction<"openai-codex-responses", OpenAICodexResponsesOptions> = (
 	model: Model<"openai-codex-responses">,
 	context: Context,
@@ -383,8 +370,14 @@ export const streamOpenAICodexResponses: StreamFunction<"openai-codex-responses"
 				}
 				if (!(attempt < MAX_RETRIES && isRetryableError(lastError))) throw lastError;
 				await attempts.settle("failed");
-				const delayMs = BASE_DELAY_MS * 2 ** attempt;
-				await sleep(delayMs, options?.signal);
+				const retryAfter = response?.headers.get("retry-after");
+				const retryAfterMs = retryAfter
+					? /^\d+(\.\d+)?$/.test(retryAfter)
+						? Number(retryAfter) * 1000
+						: Date.parse(retryAfter) - Date.now()
+					: NaN;
+				const delayMs = Number.isFinite(retryAfterMs) ? Math.max(0, retryAfterMs) : BASE_DELAY_MS * 2 ** attempt;
+				await abortableDelay(delayMs, options?.signal);
 			}
 
 			if (!response?.ok) {
@@ -453,6 +446,7 @@ function buildRequestBody(
 		includeSystemPrompt: false,
 		onProjection,
 		pendingPublicMessageGroups: options?.attempts?.pendingPublicMessageGroups,
+		responsesMessageIds: options?.attempts?.responsesMessageIds,
 	});
 
 	const body: RequestBody = {
@@ -463,7 +457,7 @@ function buildRequestBody(
 		input: messages,
 		text: { verbosity: options?.textVerbosity || "low" },
 		include: ["reasoning.encrypted_content"],
-		prompt_cache_key: options?.sessionId,
+		prompt_cache_key: options?.cacheRetention === "none" ? undefined : options?.sessionId,
 		tool_choice: "auto",
 		parallel_tool_calls: true,
 	};
@@ -702,6 +696,13 @@ async function* parseSSE(response: Response): AsyncGenerator<Record<string, unkn
 
 const OPENAI_BETA_RESPONSES_WEBSOCKETS = "responses_websockets=2026-02-06";
 const SESSION_WEBSOCKET_CACHE_TTL_MS = 5 * 60 * 1000;
+const WEBSOCKET_AUTH_HEADERS = [
+	"authorization",
+	"chatgpt-account-id",
+	"openai-organization",
+	"openai-project",
+	"x-api-key",
+] as const;
 
 type WebSocketEventType = "open" | "message" | "error" | "close";
 type WebSocketListener = (event: unknown) => void;
@@ -723,6 +724,7 @@ interface CachedWebSocketContinuationState {
 interface CachedWebSocketConnection {
 	socket: WebSocketLike;
 	url: string;
+	headers: Headers;
 	busy: boolean;
 	idleTimer?: ReturnType<typeof setTimeout>;
 	continuation?: CachedWebSocketContinuationState;
@@ -886,7 +888,9 @@ function scheduleSessionWebSocketExpiry(sessionId: string, entry: CachedWebSocke
 	entry.idleTimer = setTimeout(() => {
 		if (entry.busy) return;
 		closeWebSocketSilently(entry.socket, 1000, "idle_timeout");
-		websocketSessionCache.delete(sessionId);
+		if (websocketSessionCache.get(sessionId) === entry) {
+			websocketSessionCache.delete(sessionId);
+		}
 	}, SESSION_WEBSOCKET_CACHE_TTL_MS);
 }
 
@@ -1005,16 +1009,20 @@ async function acquireWebSocket(
 			clearTimeout(cached.idleTimer);
 			cached.idleTimer = undefined;
 		}
-		if (!cached.busy && isWebSocketReusable(cached.socket)) {
+		const matchesConnection =
+			cached.url === url && WEBSOCKET_AUTH_HEADERS.every((name) => cached.headers.get(name) === headers.get(name));
+		if (!cached.busy && matchesConnection && isWebSocketReusable(cached.socket)) {
 			cached.busy = true;
 			return {
 				socket: cached.socket,
 				entry: cached,
 				reused: true,
 				release: ({ keep } = {}) => {
-					if (!keep || !isWebSocketReusable(cached.socket)) {
+					if (!keep || !isWebSocketReusable(cached.socket) || websocketSessionCache.get(sessionId) !== cached) {
 						closeWebSocketSilently(cached.socket);
-						websocketSessionCache.delete(sessionId);
+						if (websocketSessionCache.get(sessionId) === cached) {
+							websocketSessionCache.delete(sessionId);
+						}
 						return;
 					}
 					cached.busy = false;
@@ -1032,21 +1040,22 @@ async function acquireWebSocket(
 				},
 			};
 		}
-		if (!isWebSocketReusable(cached.socket)) {
-			closeWebSocketSilently(cached.socket);
+		cached.continuation = undefined;
+		closeWebSocketSilently(cached.socket);
+		if (websocketSessionCache.get(sessionId) === cached) {
 			websocketSessionCache.delete(sessionId);
 		}
 	}
 
 	const socket = await connectWebSocket(url, headers, signal, timeoutMs);
-	const entry: CachedWebSocketConnection = { socket, url, busy: true };
+	const entry: CachedWebSocketConnection = { socket, url, headers: new Headers(headers), busy: true };
 	websocketSessionCache.set(sessionId, entry);
 	return {
 		socket,
 		entry,
 		reused: false,
 		release: ({ keep } = {}) => {
-			if (!keep || !isWebSocketReusable(entry.socket)) {
+			if (!keep || !isWebSocketReusable(entry.socket) || websocketSessionCache.get(sessionId) !== entry) {
 				closeWebSocketSilently(entry.socket);
 				if (entry.idleTimer) clearTimeout(entry.idleTimer);
 				if (websocketSessionCache.get(sessionId) === entry) {