@gajae-code/ai 0.6.1 → 0.6.2

package/CHANGELOG.md

@@ -2,6 +2,17 @@
 
 ## [Unreleased]
 
+## [0.6.2] - 2026-06-19
+### Added
+
+- Added opt-in `compat.sendSessionHeaders` for the `openai-completions` provider. When enabled (default off), the agent session id is forwarded as vendor-neutral `session_id` and `x-session-id` request headers to any OpenAI-compatible endpoint, letting relays/proxies do session-affinity routing and reuse a server-side prompt cache keyed by the session. Previously only the `openai-responses` provider injected session headers, and only against a first-party OpenAI base URL. Injection runs after the caller's `headers`/`extraHeaders` are merged and before `requestTransform`, and uses `??=` so any header the caller already set always wins; it is skipped entirely when the flag is off or no session id is available, leaving existing provider behavior byte-identical. The first-party `openai-responses` gating is unchanged.
+
+### Fixed
+
+- Prevented OpenAI Codex Responses `invalid_function_parameters` / tool-schema validation error events from being treated as retryable `server_error`s, so malformed request schemas fail fast instead of burning the full retry budget.
+- Read LM Studio `/v1/models` nested metadata such as `meta.n_ctx`, `meta.n_ctx_train`, and `details.max_tokens` when normalizing dynamically discovered GGUF-backed local models.
+- Corrected the bundled `openai-codex/gpt-5.5` context window from an overstated 400K back to its true 272K (272,000-token) window, so context-cap / auto-compaction thresholds no longer let gpt-5.5 sessions overrun the model's real limit before compaction (#873).
+
 ## [0.6.1] - 2026-06-18
 
 ### Fixed

package/README.md

@@ -778,6 +778,7 @@ The `openai-completions` API is implemented by many providers with minor differe
 interface OpenAICompat {
 	supportsStore?: boolean; // Whether provider supports the `store` field (default: true)
 	supportsDeveloperRole?: boolean; // Whether provider supports `developer` role vs `system` (default: true)
+	sendSessionHeaders?: boolean; // Forward the session id as `session_id`/`x-session-id` headers for relay session-affinity & prompt-cache reuse (default: false)
 	supportsReasoningEffort?: boolean; // Whether provider supports `reasoning_effort` (default: true)
 	maxTokensField?: "max_completion_tokens" | "max_tokens"; // Which field name to use (default: max_completion_tokens)
 	extraBody?: Record<string, unknown>; // Extra request-body fields for custom proxy routing or provider-specific options

package/dist/types/types.d.ts

@@ -606,6 +606,17 @@ export interface OpenAICompat extends ToolChoiceCompat {
     supportsStore?: boolean;
     /** Whether the provider supports the `developer` role (vs `system`). Default: auto-detected from URL. */
     supportsDeveloperRole?: boolean;
+    /**
+     * Whether to forward the agent session id as vendor-neutral session-identity
+     * headers (`session_id`, `x-session-id`) on every chat-completions request.
+     * Off by default. Opt in for OpenAI-compatible proxies/relays that route on
+     * session affinity or reuse a server-side prompt cache keyed by session.
+     * First-party OpenAI does not need this (it has its own gated injection in
+     * the openai-responses provider). Headers are only added when a non-empty
+     * session id is available and are never allowed to overwrite a header the
+     * caller already set via `headers`/`requestTransform`.
+     */
+    sendSessionHeaders?: boolean;
     /**
      * Whether the provider's chat-completions endpoint accepts multiple
      * leading `system`/`developer` messages. When false, ordered system

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@gajae-code/ai",
-	"version": "0.6.1",
+	"version": "0.6.2",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"homepage": "https://gaebal-gajae.dev",
 	"author": "Yeachan-Heo",
@@ -43,7 +43,7 @@
 	"dependencies": {
 		"@anthropic-ai/sdk": "^0.94.0",
 		"@bufbuild/protobuf": "^2.12.0",
-		"@gajae-code/utils": "0.6.1",
+		"@gajae-code/utils": "0.6.2",
 		"openai": "^6.36.0",
 		"partial-json": "^0.1.7",
 		"zod": "4.4.3"

package/src/models.json

@@ -53332,7 +53332,7 @@
 				"cacheRead": 0.5,
 				"cacheWrite": 0
 			},
-			"contextWindow": 400000,
+			"contextWindow": 272000,
 			"maxTokens": 128000,
 			"preferWebsockets": true,
 			"priority": 9,

package/src/provider-models/openai-compat.ts

@@ -176,6 +176,57 @@ function mapWithBundledReference<TApi extends Api>(
 	};
 }
 
+function getNestedModelValue(entry: OpenAICompatibleModelRecord, path: readonly string[]): unknown {
+	let current: unknown = entry;
+	for (const segment of path) {
+		if (!isRecord(current)) {
+			return undefined;
+		}
+		current = current[segment];
+	}
+	return current;
+}
+
+function firstPositiveModelNumber(fallback: number, ...candidates: readonly unknown[]): number {
+	for (const candidate of candidates) {
+		const value = toNumber(candidate);
+		if (value !== undefined && value > 0) {
+			return value;
+		}
+	}
+	return fallback;
+}
+
+function mapLmStudioModel(
+	entry: OpenAICompatibleModelRecord,
+	defaults: Model<"openai-completions">,
+	reference: Model<"openai-completions"> | undefined,
+): Model<"openai-completions"> {
+	const model = mapWithBundledReference(entry, defaults, reference);
+	return {
+		...model,
+		contextWindow: firstPositiveModelNumber(
+			model.contextWindow,
+			entry.context_length,
+			entry.max_context_length,
+			getNestedModelValue(entry, ["meta", "n_ctx"]),
+			getNestedModelValue(entry, ["details", "context_length"]),
+			getNestedModelValue(entry, ["details", "n_ctx"]),
+			getNestedModelValue(entry, ["meta", "n_ctx_train"]),
+		),
+		maxTokens: firstPositiveModelNumber(
+			model.maxTokens,
+			entry.max_completion_tokens,
+			entry.max_tokens,
+			entry.max_output_tokens,
+			getNestedModelValue(entry, ["details", "max_completion_tokens"]),
+			getNestedModelValue(entry, ["details", "max_tokens"]),
+			getNestedModelValue(entry, ["meta", "max_completion_tokens"]),
+			getNestedModelValue(entry, ["meta", "max_tokens"]),
+		),
+	};
+}
+
 function normalizeAnthropicBaseUrl(baseUrl: string | undefined, fallback: string): string {
 	const value = baseUrl?.trim();
 	if (!value) {
@@ -1215,7 +1266,7 @@ export function lmStudioModelManagerOptions(
 				apiKey,
 				mapModel: (entry, defaults) => {
 					const reference = references.get(defaults.id);
-					return mapWithBundledReference(entry, defaults, reference);
+					return mapLmStudioModel(entry, defaults, reference);
 				},
 			}),
 	};

package/src/providers/openai-codex-responses.ts

@@ -104,6 +104,14 @@ const CODEX_WEBSOCKET_RETRY_BUDGET = CODEX_MAX_RETRIES;
 const CODEX_WEBSOCKET_TRANSPORT_ERROR_PREFIX = "Codex websocket transport error";
 const CODEX_PREVIOUS_RESPONSE_STALE_CODES = new Set(["previous_response_not_found", "codex_previous_response_stale"]);
 const CODEX_RETRYABLE_EVENT_CODES = new Set(["model_error", "server_error", "internal_error"]);
+const CODEX_NON_RETRYABLE_EVENT_CODES = new Set([
+	"invalid_function_parameters",
+	"invalid_request_error",
+	"invalid_schema",
+	"invalid_tool_schema",
+]);
+const CODEX_NON_RETRYABLE_EVENT_MESSAGE =
+	/invalid[_ -]function[_ -]parameters|invalid schema for function|invalid[_ -]tool[_ -]schema|schema must have type ["']?object["']?/i;
 const CODEX_RETRYABLE_EVENT_MESSAGE =
 	/processing your request|retry your request|temporar(?:y|ily)|overloaded|service.?unavailable|internal error|server error/i;
 const CODEX_PROVIDER_SESSION_STATE_KEY = "openai-codex-responses";
@@ -2709,11 +2717,17 @@ class CodexProviderStreamError extends Error {
 }
 
 function isRetryableCodexFailureEvent(rawEvent: Record<string, unknown>): boolean {
-	const code = getCodexEventErrorCode(rawEvent);
-	if (code && CODEX_RETRYABLE_EVENT_CODES.has(code.toLowerCase())) {
+	const code = getCodexEventErrorCode(rawEvent).toLowerCase();
+	const message = getCodexEventErrorMessage(rawEvent);
+	if (
+		(code && CODEX_NON_RETRYABLE_EVENT_CODES.has(code)) ||
+		(!!message && CODEX_NON_RETRYABLE_EVENT_MESSAGE.test(message))
+	) {
+		return false;
+	}
+	if (code && CODEX_RETRYABLE_EVENT_CODES.has(code)) {
 		return true;
 	}
-	const message = getCodexEventErrorMessage(rawEvent);
 	return !!message && CODEX_RETRYABLE_EVENT_MESSAGE.test(message);
 }
 

package/src/providers/openai-completions-compat.ts

@@ -189,6 +189,7 @@ export function detectOpenAICompat(model: Model<"openai-completions">, resolvedB
 	return {
 		supportsStore: !isNonStandard,
 		supportsDeveloperRole: !isNonStandard,
+		sendSessionHeaders: false,
 		supportsMultipleSystemMessages: supportsMultipleSystemMessagesDefault,
 		supportsReasoningEffort: !isGrok && !isZai,
 		reasoningEffortMap,
@@ -254,6 +255,7 @@ export function resolveOpenAICompat(
 	return {
 		supportsStore: model.compat.supportsStore ?? detected.supportsStore,
 		supportsDeveloperRole: model.compat.supportsDeveloperRole ?? detected.supportsDeveloperRole,
+		sendSessionHeaders: model.compat.sendSessionHeaders ?? detected.sendSessionHeaders,
 		supportsMultipleSystemMessages:
 			model.compat.supportsMultipleSystemMessages ?? detected.supportsMultipleSystemMessages,
 		supportsReasoningEffort: model.compat.supportsReasoningEffort ?? detected.supportsReasoningEffort,

package/src/providers/openai-completions.ts

@@ -453,6 +453,7 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions"> = (
 				options?.streamFirstEventTimeoutMs,
 				options?.authCredentialType,
 				options?.requestMaxRetries,
+				options?.sessionId,
 			);
 			const premiumRequestsTotal = copilotPremiumRequests;
 			getCapturedErrorResponse = captureErrorResponse;
@@ -943,6 +944,7 @@ async function createClient(
 	streamFirstEventTimeoutOverride?: number,
 	authCredentialType?: OpenAICompletionsOptions["authCredentialType"],
 	requestMaxRetries?: number,
+	sessionId?: string,
 ): Promise<{
 	client: OpenAI;
 	copilotPremiumRequests: number | undefined;
@@ -981,6 +983,14 @@ async function createClient(
 		headers["X-OpenRouter-Cache-TTL"] = "3600";
 	}
 	Object.assign(headers, extraHeaders);
+	if (sessionId && resolveOpenAICompat(model).sendSessionHeaders) {
+		// Forward the agent session id as vendor-neutral session-identity headers so
+		// OpenAI-compatible proxies/relays can do session-affinity routing and reuse a
+		// server-side prompt cache. Opt-in via `compat.sendSessionHeaders`; never
+		// overwrite a header the caller already set (model.headers / requestTransform).
+		headers.session_id ??= sessionId;
+		headers["x-session-id"] ??= sessionId;
+	}
 	if (model.provider === "kimi-code") {
 		headers = { ...getKimiCommonHeaders(), ...headers };
 	}

package/src/types.ts

@@ -742,6 +742,17 @@ export interface OpenAICompat extends ToolChoiceCompat {
 	supportsStore?: boolean;
 	/** Whether the provider supports the `developer` role (vs `system`). Default: auto-detected from URL. */
 	supportsDeveloperRole?: boolean;
+	/**
+	 * Whether to forward the agent session id as vendor-neutral session-identity
+	 * headers (`session_id`, `x-session-id`) on every chat-completions request.
+	 * Off by default. Opt in for OpenAI-compatible proxies/relays that route on
+	 * session affinity or reuse a server-side prompt cache keyed by session.
+	 * First-party OpenAI does not need this (it has its own gated injection in
+	 * the openai-responses provider). Headers are only added when a non-empty
+	 * session id is available and are never allowed to overwrite a header the
+	 * caller already set via `headers`/`requestTransform`.
+	 */
+	sendSessionHeaders?: boolean;
 	/**
 	 * Whether the provider's chat-completions endpoint accepts multiple
 	 * leading `system`/`developer` messages. When false, ordered system