@gajae-code/ai 0.4.1 → 0.4.3

package/CHANGELOG.md

@@ -2,6 +2,12 @@
 
 ## [Unreleased]
 
+## [0.4.2] - 2026-06-09
+
+### Fixed
+
+- Treated `gpt-5.5` as a 400K-context model wherever context caps / auto-promote thresholds are resolved, so a ~272K session is no longer considered over-cap and no longer demotes to `gpt-5.4`. Pinned the OpenAI Codex `gpt-5.5` context window to 400K and removed its `gpt-5.4` promotion target (the smaller window made it a demotion) ([#428](https://github.com/Yeachan-Heo/gajae-code/issues/428)).
+
 ## [0.4.0] - 2026-06-06
 
 ### Added

package/dist/types/model-thinking.d.ts

@@ -40,7 +40,9 @@ export declare function applyGeneratedModelPolicies(models: ApiModel<Api>[]): vo
  * When a model's context is exhausted, the agent can promote to a sibling
  * model with a larger context window on the same provider:
  * - `OpenAI code backend-spark` variants promote to `gpt-5.5`.
- * - `gpt-5.5` (270K input) promotes to `gpt-5.4` (1M input).
+ *
+ * `gpt-5.5` itself is a 400K-context model and is not demoted to `gpt-5.4`
+ * (which has a smaller window), so it has no promotion target.
  */
 export declare function linkOpenAIPromotionTargets(models: ApiModel<Api>[]): void;
 /**

package/dist/types/types.d.ts

@@ -717,6 +717,14 @@ export interface Model<TApi extends Api = any> {
     baseUrl: string;
     reasoning: boolean;
     input: ("text" | "image")[];
+    /**
+     * Output modalities the model can produce. Defaults to text-only when
+     * unset. A model that lists `"image"` advertises image-generation support
+     * (e.g. an OpenAI-compatible `gpt-image` model behind a proxy), which the
+     * `generate_image` tool uses to route requests without first-party
+     * provider/id heuristics.
+     */
+    output?: ("text" | "image")[];
     cost: {
         input: number;
         output: number;

package/dist/types/utils/http-inspector.d.ts

@@ -13,6 +13,10 @@ export type CapturedHttpErrorResponse = {
     bodyText?: string;
     bodyJson?: unknown;
 };
+/** Whether `message` (from a 400 response) signals an unavailable/unknown model. */
+export declare function isModelUnavailableError(message: string, error: unknown): boolean;
+/** Actionable guidance for selecting an available model/provider. */
+export declare function formatModelUnavailableGuidance(dump: RawHttpRequestDump | undefined): string;
 export declare function appendRawHttpRequestDumpFor400(message: string, error: unknown, dump: RawHttpRequestDump | undefined): Promise<string>;
 export declare function finalizeErrorMessage(error: unknown, rawRequestDump: RawHttpRequestDump | undefined, capturedErrorResponse?: CapturedHttpErrorResponse): Promise<string>;
 export declare function withHttpStatus(error: unknown, status: number): Error;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@gajae-code/ai",
-	"version": "0.4.1",
+	"version": "0.4.3",
 	"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": "0.4.1",
+		"@gajae-code/utils": "0.4.3",
 		"openai": "^6.36.0",
 		"partial-json": "^0.1.7",
 		"zod": "4.4.3"

package/src/model-thinking.ts

@@ -195,7 +195,9 @@ export function applyGeneratedModelPolicies(models: ApiModel<Api>[]): void {
  * When a model's context is exhausted, the agent can promote to a sibling
  * model with a larger context window on the same provider:
  * - `OpenAI code backend-spark` variants promote to `gpt-5.5`.
- * - `gpt-5.5` (270K input) promotes to `gpt-5.4` (1M input).
+ *
+ * `gpt-5.5` itself is a 400K-context model and is not demoted to `gpt-5.4`
+ * (which has a smaller window), so it has no promotion target.
  */
 export function linkOpenAIPromotionTargets(models: ApiModel<Api>[]): void {
 	for (const candidate of models) {
@@ -204,8 +206,6 @@ export function linkOpenAIPromotionTargets(models: ApiModel<Api>[]): void {
 		let targetId: string | undefined;
 		if (parsedCandidate.variant === "codex-spark") {
 			targetId = "gpt-5.5";
-		} else if (parsedCandidate.variant === "base" && semverEqual(parsedCandidate.version, "5.5")) {
-			targetId = "gpt-5.4";
 		} else {
 			continue;
 		}
@@ -430,7 +430,22 @@ function inferGeneratedApplyPatchToolType(
 	return undefined;
 }
 
+function applyGpt55ContextWindow(model: ApiModel<Api>, parsedModel: OpenAIModel): boolean {
+	// gpt-5.5 is a 400K-context model. OpenAI code backend discovery can omit the
+	// context window, falling back to the 272K default, which incorrectly trips
+	// context-cap / auto-promote thresholds (a ~272K session would look over-cap
+	// and demote to gpt-5.4). Pin gpt-5.5 to its true 400K window.
+	if (parsedModel.variant === "base" && semverEqual(parsedModel.version, "5.5")) {
+		model.contextWindow = 400000;
+		return true;
+	}
+	return false;
+}
+
 function applyOpenAICatalogPolicy(model: ApiModel<Api>, parsedModel: OpenAIModel): void {
+	if (applyGpt55ContextWindow(model, parsedModel)) {
+		return;
+	}
 	// OpenAI code backend models: 400K figure includes output budget; input window is 272K.
 	if (parsedModel.variant.startsWith("codex") && parsedModel.variant !== "codex-spark") {
 		model.contextWindow = 272000;

package/src/models.json

@@ -9853,7 +9853,8 @@
 			"baseUrl": "https://api.kilo.ai/api/gateway",
 			"reasoning": false,
 			"input": [
-				"text"
+				"text",
+				"image"
 			],
 			"cost": {
 				"input": 0,
@@ -9872,7 +9873,8 @@
 			"baseUrl": "https://api.kilo.ai/api/gateway",
 			"reasoning": false,
 			"input": [
-				"text"
+				"text",
+				"image"
 			],
 			"cost": {
 				"input": 0,
@@ -16768,7 +16770,8 @@
 			"baseUrl": "https://api.kilo.ai/api/gateway",
 			"reasoning": false,
 			"input": [
-				"text"
+				"text",
+				"image"
 			],
 			"cost": {
 				"input": 0,
@@ -34660,7 +34663,7 @@
 		},
 		"minimax-m3": {
 			"id": "minimax-m3",
-			"name": "MiniMax M3",
+			"name": "MiniMax-M3",
 			"api": "anthropic-messages",
 			"provider": "minimax",
 			"baseUrl": "https://api.minimax.io/anthropic",
@@ -34855,7 +34858,7 @@
 		},
 		"minimax-m3": {
 			"id": "minimax-m3",
-			"name": "MiniMax M3",
+			"name": "MiniMax-M3",
 			"api": "anthropic-messages",
 			"provider": "minimax-cn",
 			"baseUrl": "https://api.minimaxi.com/anthropic",
@@ -35122,7 +35125,7 @@
 		},
 		"minimax-m3": {
 			"id": "minimax-m3",
-			"name": "MiniMax M3",
+			"name": "MiniMax-M3",
 			"api": "openai-completions",
 			"provider": "minimax-code",
 			"baseUrl": "https://api.minimax.io/v1",
@@ -35395,7 +35398,7 @@
 		},
 		"minimax-m3": {
 			"id": "minimax-m3",
-			"name": "MiniMax M3",
+			"name": "MiniMax-M3",
 			"api": "openai-completions",
 			"provider": "minimax-code-cn",
 			"baseUrl": "https://api.minimaxi.com/v1",
@@ -53264,7 +53267,7 @@
 				"cacheRead": 0.5,
 				"cacheWrite": 0
 			},
-			"contextWindow": 272000,
+			"contextWindow": 400000,
 			"maxTokens": 128000,
 			"preferWebsockets": true,
 			"priority": 9,
@@ -53274,8 +53277,7 @@
 				"maxLevel": "xhigh",
 				"defaultLevel": "xhigh"
 			},
-			"applyPatchToolType": "freeform",
-			"contextPromotionTarget": "openai-codex/gpt-5.4"
+			"applyPatchToolType": "freeform"
 		}
 	},
 	"opencode": {
@@ -63602,7 +63604,8 @@
 			"baseUrl": "https://api.venice.ai/api/v1",
 			"reasoning": false,
 			"input": [
-				"text"
+				"text",
+				"image"
 			],
 			"cost": {
 				"input": 0,

package/src/types.ts

@@ -854,6 +854,14 @@ export interface Model<TApi extends Api = any> {
 	baseUrl: string;
 	reasoning: boolean;
 	input: ("text" | "image")[];
+	/**
+	 * Output modalities the model can produce. Defaults to text-only when
+	 * unset. A model that lists `"image"` advertises image-generation support
+	 * (e.g. an OpenAI-compatible `gpt-image` model behind a proxy), which the
+	 * `generate_image` tool uses to route requests without first-party
+	 * provider/id heuristics.
+	 */
+	output?: ("text" | "image")[];
 	cost: {
 		input: number; // $/million tokens
 		output: number; // $/million tokens

package/src/utils/http-inspector.ts

@@ -1,5 +1,5 @@
 import * as path from "node:path";
-import { extractHttpStatusFromError, getLogsDir } from "@gajae-code/utils";
+import { APP_NAME, extractHttpStatusFromError, getLogsDir } from "@gajae-code/utils";
 import { isCopilotTransientModelError } from "./retry.js";
 import { formatErrorMessageWithRetryAfter } from "./retry-after.js";
 
@@ -26,6 +26,46 @@ type ErrorWithStatus = {
 
 const SENSITIVE_HEADERS = ["authorization", "x-api-key", "api-key", "cookie", "set-cookie", "proxy-authorization"];
 
+/**
+ * Privacy note appended next to a saved raw HTTP request dump. The dump is
+ * sanitized (secrets/thinking redacted) but can still contain prompt content
+ * and request metadata, so we explicitly discourage pasting it into public
+ * channels (issue #438).
+ */
+const RAW_HTTP_REQUEST_PRIVACY_NOTE =
+	"note: this local file is for your own debugging and may contain prompt content or request metadata — review it before sharing and do not paste its contents into public channels (issues, Discord, etc.).";
+
+/**
+ * Patterns that indicate the configured model is unavailable on the provider
+ * (e.g. OpenAI's "The requested model '...' does not exist."). Used to surface
+ * actionable model/provider guidance instead of only a raw 400 + log path.
+ */
+const MODEL_UNAVAILABLE_PATTERNS: readonly RegExp[] = [
+	/\bmodel\b[^\n]*\bdoes not exist\b/i,
+	/\bdoes not exist\b[^\n]*\bmodel\b/i,
+	/\bmodel\b[^\n]*\b(not found|unavailable|not supported|no access|does not have access)\b/i,
+	/\b(unknown|unsupported|invalid)\s+model\b/i,
+];
+
+/** Whether `message` (from a 400 response) signals an unavailable/unknown model. */
+export function isModelUnavailableError(message: string, error: unknown): boolean {
+	if (extractHttpStatusFromError(error) !== 400) return false;
+	return MODEL_UNAVAILABLE_PATTERNS.some(pattern => pattern.test(message));
+}
+
+/** Actionable guidance for selecting an available model/provider. */
+export function formatModelUnavailableGuidance(dump: RawHttpRequestDump | undefined): string {
+	const modelPart = dump?.model ? ` '${dump.model}'` : "";
+	const providerPart = dump?.provider ? ` on provider '${dump.provider}'` : "";
+	return [
+		`The configured model${modelPart}${providerPart} is not available on this account/provider.`,
+		"Pick an available model before retrying:",
+		`  • List available models:  ${APP_NAME} --list-models`,
+		`  • Run with a model:        ${APP_NAME} --model <model>`,
+		`  • Configure a provider:    ${APP_NAME} setup provider`,
+	].join("\n");
+}
+
 export async function appendRawHttpRequestDumpFor400(
 	message: string,
 	error: unknown,
@@ -41,7 +81,7 @@ export async function appendRawHttpRequestDumpFor400(
 
 	try {
 		await Bun.write(filePath, `${JSON.stringify(sanitizedDump, null, 2)}\n`);
-		return `${message}\nraw-http-request=${filePath}`;
+		return `${message}\nraw-http-request=${filePath}\n${RAW_HTTP_REQUEST_PRIVACY_NOTE}`;
 	} catch (writeError) {
 		const writeMessage = writeError instanceof Error ? writeError.message : String(writeError);
 		return `${message}\nraw-http-request-save-failed=${writeMessage}`;
@@ -62,6 +102,9 @@ export async function finalizeErrorMessage(
 			message = `${message}\n${capturedMessage}`;
 		}
 	}
+	if (isModelUnavailableError(message, error)) {
+		message = `${message}\n\n${formatModelUnavailableGuidance(rawRequestDump)}`;
+	}
 	return appendRawHttpRequestDumpFor400(message, error, rawRequestDump);
 }
 

package/src/utils/idle-iterator.ts

@@ -157,10 +157,14 @@ export async function* iterateWithIdleTimeout<T>(
 			activeTimeoutMs = firstItemTimeoutMs;
 		} else if (options.idleTimeoutMs !== undefined && options.idleTimeoutMs > 0) {
 			activeTimeoutMs = options.idleTimeoutMs - (Date.now() - lastProgressAt);
-			if (activeTimeoutMs <= 0) {
-				options.onIdle?.();
-				closeIterator();
-				throw new Error(options.errorMessage);
+			// The idle deadline may already have elapsed because the *consumer*
+			// was slow, not because the provider stalled — and the next item may
+			// already be buffered and ready to deliver. Clamp to 0 instead of
+			// throwing eagerly so the next() race below still gets a chance to
+			// win (it settles on a microtask, ahead of the 0ms timer). Only a
+			// genuinely hung iterator loses that race and surfaces as a stall.
+			if (activeTimeoutMs < 0) {
+				activeTimeoutMs = 0;
 			}
 		}
 
@@ -177,7 +181,10 @@ export async function* iterateWithIdleTimeout<T>(
 
 		let timer: NodeJS.Timeout | undefined;
 		let resolveTimeout: ((value: { kind: "timeout" }) => void) | undefined;
-		const enforceTimeout = !noTimeoutEnforced && activeTimeoutMs !== undefined && activeTimeoutMs > 0;
+		const enforceTimeout =
+			!noTimeoutEnforced &&
+			activeTimeoutMs !== undefined &&
+			(awaitingFirstItem ? activeTimeoutMs > 0 : activeTimeoutMs >= 0);
 		if (enforceTimeout) {
 			const { promise, resolve } = Promise.withResolvers<{ kind: "timeout" }>();
 			resolveTimeout = resolve;