@oh-my-pi/pi-catalog 16.1.9 → 16.1.11

package/CHANGELOG.md

@@ -2,6 +2,12 @@
 
 ## [Unreleased]
 
+## [16.1.11] - 2026-06-21
+
+### Fixed
+
+- Fixed Umans `umans-glm-5.1` / `umans-glm-5.2` advertising native image input. The `models/info` endpoint reports `supports_vision: "via-handoff"` for the GLM models, meaning vision routes through a separate handoff pre-analysis step instead of accepting raw image blocks; `umansSupportsVision` treated any non-empty string as native vision support, so image prompts went directly to GLM and were rejected with `400 This model does not support image inputs`. The helper now requires `supports_vision === true`, the bundled GLM 5.1/5.2 rows are corrected to text-only, and stale mismatched Umans cache rows for those ids are dropped so the vision-handoff path runs even before a successful refresh. ([#3184](https://github.com/can1357/oh-my-pi/issues/3184))
+
 ## [16.1.9] - 2026-06-21
 
 ### Fixed

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

@@ -28,6 +28,8 @@ export interface ModelManagerOptions<TApi extends Api = Api, TModelsDevPayload =
     cacheTtlMs?: number;
     /** When true, a successful dynamic fetch is the complete provider catalog and prunes static-only models. */
     dynamicModelsAuthoritative?: boolean;
+    /** Cached model ids to ignore when the cache was written against a different static catalog fingerprint. */
+    dropCachedModelIdsOnStaticMismatch?: readonly string[];
     /** Optional dynamic endpoint fetcher. */
     fetchDynamicModels?: () => Promise<readonly ModelSpec<TApi>[] | null>;
     /** Optional models.dev fallback hook. */

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "16.1.9",
+	"version": "16.1.11",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -34,12 +34,12 @@
 	},
 	"dependencies": {
 		"@bufbuild/protobuf": "^2.12.0",
-		"@oh-my-pi/pi-utils": "16.1.9",
+		"@oh-my-pi/pi-utils": "16.1.11",
 		"arktype": "^2.2.0",
 		"zod": "^4"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "16.1.9",
+		"@oh-my-pi/pi-ai": "16.1.11",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/openai.ts

@@ -39,6 +39,10 @@ const GLM_CODING_PLAN_STREAM_IDLE_TIMEOUT_MS = 600_000;
 const DEEPSEEK_REASONING_STREAM_IDLE_TIMEOUT_MS = 300_000;
 /** Kimi K2.6 can spend several minutes reasoning before the first visible token. */
 const KIMI_K26_REASONING_STREAM_IDLE_TIMEOUT_MS = 300_000;
+/** Xiaomi MiMo Pro on api.xiaomimimo.com can stall ~2min before the first event (issue #1770). */
+const XIAOMI_MIMO_STREAM_IDLE_TIMEOUT_MS = 300_000;
+/** Alibaba Coding Plan (coding-intl.dashscope) qwen models idle before the first event (issue #1770). */
+const ALIBABA_CODING_PLAN_STREAM_IDLE_TIMEOUT_MS = 600_000;
 const MINIMAX_PROVIDER_OR_ID_PATTERN = /minimax/i;
 const DSML_HEALING_PROVIDERS = new Set([
 	"ollama",
@@ -287,11 +291,15 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
 	const streamIdleTimeoutMs =
 		GLM_CODING_PLAN_MODEL_PATTERN.test(spec.id) && (isZai || isZhipu)
 			? GLM_CODING_PLAN_STREAM_IDLE_TIMEOUT_MS
-			: spec.reasoning && isKimiK26ModelId(spec.id)
-				? KIMI_K26_REASONING_STREAM_IDLE_TIMEOUT_MS
-				: spec.reasoning && isDirectDeepseekApi
-					? DEEPSEEK_REASONING_STREAM_IDLE_TIMEOUT_MS
-					: undefined;
+			: provider === "alibaba-coding-plan"
+				? ALIBABA_CODING_PLAN_STREAM_IDLE_TIMEOUT_MS
+				: isXiaomiMimo
+					? XIAOMI_MIMO_STREAM_IDLE_TIMEOUT_MS
+					: spec.reasoning && isKimiK26ModelId(spec.id)
+						? KIMI_K26_REASONING_STREAM_IDLE_TIMEOUT_MS
+						: spec.reasoning && isDirectDeepseekApi
+							? DEEPSEEK_REASONING_STREAM_IDLE_TIMEOUT_MS
+							: undefined;
 
 	// Fireworks "Fast" variants (`<id>-fast`) are served from the router
 	// namespace (`accounts/fireworks/routers/<id>-fast`), like Fire Pass, rather

package/src/model-manager.ts

@@ -39,6 +39,8 @@ export interface ModelManagerOptions<TApi extends Api = Api, TModelsDevPayload =
 	cacheTtlMs?: number;
 	/** When true, a successful dynamic fetch is the complete provider catalog and prunes static-only models. */
 	dynamicModelsAuthoritative?: boolean;
+	/** Cached model ids to ignore when the cache was written against a different static catalog fingerprint. */
+	dropCachedModelIdsOnStaticMismatch?: readonly string[];
 	/** Optional dynamic endpoint fetcher. */
 	fetchDynamicModels?: () => Promise<readonly ModelSpec<TApi>[] | null>;
 	/** Optional models.dev fallback hook. */
@@ -148,7 +150,13 @@ export async function resolveProviderModels<TApi extends Api = Api, TModelsDevPa
 	const shouldUseFreshCacheAsAuthoritative =
 		strategy === "online-if-uncached" && hasUsableFreshCache && hasAuthoritativeCache;
 	const dynamicFetchSucceeded = fetchedDynamicModels !== null;
-	const cacheModels = dynamicFetchSucceeded ? [] : normalizeModelList<TApi>(cache?.models ?? []);
+	const cacheModels = dynamicFetchSucceeded
+		? []
+		: dropCachedModelIdsOnStaticMismatch(
+				normalizeModelList<TApi>(cache?.models ?? []),
+				cacheFingerprintMatches,
+				options.dropCachedModelIdsOnStaticMismatch,
+			);
 	const dynamicModels = fetchedDynamicModels ?? [];
 	const mergedWithCache = mergeDynamicModels(mergeModelSources(staticModels, modelsDevModels), cacheModels);
 	const mergedModels = mergeDynamicModels(mergedWithCache, dynamicModels);
@@ -180,7 +188,11 @@ export async function resolveProviderModels<TApi extends Api = Api, TModelsDevPa
 				collapseBuiltModelVariants(
 					mergeDynamicModels(
 						mergeModelSources(staticModels, modelsDevModels),
-						normalizeModelList<TApi>(latestCache?.models ?? cache?.models ?? []),
+						dropCachedModelIdsOnStaticMismatch(
+							normalizeModelList<TApi>(latestCache?.models ?? cache?.models ?? []),
+							cacheFingerprintMatches,
+							options.dropCachedModelIdsOnStaticMismatch,
+						),
 					),
 				),
 				false,
@@ -248,6 +260,18 @@ function shouldFetchRemoteSources(
 	return false;
 }
 
+function dropCachedModelIdsOnStaticMismatch<TApi extends Api>(
+	models: readonly Model<TApi>[],
+	cacheFingerprintMatches: boolean,
+	ids: readonly string[] | undefined,
+): Model<TApi>[] {
+	if (cacheFingerprintMatches || ids === undefined || ids.length === 0 || models.length === 0) {
+		return models.length === 0 ? [] : [...models];
+	}
+	const droppedIds = new Set(ids);
+	return models.filter(model => !droppedIds.has(model.id));
+}
+
 function mergeModelSources<TApi extends Api>(...sources: readonly (readonly Model<TApi>[])[]): Model<TApi>[] {
 	// Strip out empty/missing sources up front. The hot path is `(static, [])`
 	// (modelsDev disabled / failed) — a single non-empty source means we can

package/src/models.json

@@ -69284,8 +69284,7 @@
 			"baseUrl": "https://api.code.umans.ai",
 			"reasoning": true,
 			"input": [
-				"text",
-				"image"
+				"text"
 			],
 			"cost": {
 				"input": 0,
@@ -69327,8 +69326,7 @@
 				]
 			},
 			"input": [
-				"text",
-				"image"
+				"text"
 			],
 			"cost": {
 				"input": 0,

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

@@ -568,6 +568,7 @@ const UMANS_REASONING_EFFORT_BY_LEVEL: Record<string, Effort> = {
 	xhigh: Effort.XHigh,
 };
 const UMANS_DEFAULT_REASONING_EFFORTS = [Effort.Minimal, Effort.Low, Effort.Medium, Effort.High, Effort.XHigh] as const;
+const UMANS_VIA_HANDOFF_MODEL_IDS = ["umans-glm-5.1", "umans-glm-5.2"] as const;
 
 export interface UmansModelManagerConfig {
 	apiKey?: string;
@@ -586,8 +587,18 @@ function normalizeUmansBaseUrl(baseUrl: string | undefined): string {
 	return normalized.endsWith("/v1") ? normalized.slice(0, -3) : normalized;
 }
 
+/**
+ * Umans `models/info` reports `supports_vision: true` for natively
+ * vision-capable models and a non-empty string sentinel (e.g.
+ * `"via-handoff"`) for models that route image inputs through a vision
+ * handoff pre-analysis step instead of accepting raw image blocks. Only
+ * `true` means the model accepts image content directly; sentinel values
+ * MUST map to text-only so the agent's vision-handoff path runs instead
+ * of triggering an upstream HTTP 400 (`This model does not support image
+ * inputs`).
+ */
 function umansSupportsVision(value: unknown): boolean {
-	return value === true || (typeof value === "string" && value.length > 0);
+	return value === true;
 }
 
 function umansReasoningSupported(value: unknown): boolean {
@@ -704,6 +715,7 @@ export function umansModelManagerOptions(config?: UmansModelManagerConfig): Mode
 	return {
 		providerId: "umans",
 		dynamicModelsAuthoritative: true,
+		dropCachedModelIdsOnStaticMismatch: UMANS_VIA_HANDOFF_MODEL_IDS,
 		fetchDynamicModels: () => fetchUmansModelsInfo({ baseUrl, apiKey, fetch: config?.fetch, references }),
 	};
 }