jeopi-catalog 16.4.2 → 16.4.3

package/CHANGELOG.md

@@ -2,6 +2,16 @@
 
 ## [Unreleased]
 
+## [16.4.3] - 2026-07-22
+
+### Fixed
+
+- Fixed SAP AI Core Claude ids in version-first order (`anthropic--claude-4.8-opus`) parsing as unknown, restoring Anthropic adaptive thinking metadata and capability gates. Ported from oh-my-pi (upstream `c1480b29e`).
+- Fixed GitHub Copilot Business and Enterprise discovery downgrading vision-capable models to text-only unconditionally; the catalog now honors upstream's `supports.vision` flag on non-personal endpoints (only an explicit `false`, or an omitted flag, falls back to text-only) instead of ignoring it outright. Ported from oh-my-pi (upstream `e58d2c460`, `4bae9a42a`, `b0f22caf8`).
+- Fixed stale cached model limits overriding updated static catalog limits after a static catalog fingerprint mismatch: same-id cache rows are now sanitized (`contextWindow`/`maxTokens` reset to `null`, deferring to the fresh static entry) instead of passed through verbatim, while cache-only ids (no static counterpart) keep their cached limits. Ported from oh-my-pi (upstream `b7aa046ed`).
+- Fixed Cursor discovery to preserve `GetUsableModels` max-mode metadata for premium models and invalidate stale pre-max-mode cache rows (bumped the Cursor cache namespace so old cache entries written before `cursorMaxMode` was tracked can never satisfy a fresh-cache read). Ported from oh-my-pi (upstream `358811115`, `f98ef2e1e`).
+- Fixed Z.ai GLM-5.2 on the `anthropic-messages` coding endpoint (`api.z.ai/api/anthropic`) deriving generic `budget` mode with five synthetic effort tiers instead of the wire-exact `anthropic-budget-effort` mode with the two-tier high/max scale Z.ai's Anthropic proxy actually serves (same as Umans). The reasoning-effort predicate now matches both `umans` and `zai` providers on `anthropic-messages`. Ported from oh-my-pi (upstream `f34034fa7` by @oldschoola; adapted to jeopi's `Effort` enum, which has no `Max` tier — jeopi exposes `[High, XHigh]` with `effortMap: { xhigh: "max" }` rather than upstream's literal `[high, max]` efforts, matching the existing Umans/Ollama Cloud GLM-5.2 convention. The bundled `models.json` `zai/glm-5.2` entry's stale `thinking` blob was intentionally left as-is per repo convention — `buildModel()` always re-derives `thinking` from this policy at runtime via `resolveModelThinking`, so the stored JSON field is inert; a full live `gen:models` regeneration was out of scope for this targeted fix).
+
 ## [16.2.29] - 2026-07-10
 
 ### Fixed

package/dist/types/types.d.ts

@@ -572,6 +572,8 @@ export interface Model<TApi extends Api = Api> {
     supportsTools?: boolean;
     /** GitLab Duo Workflow root namespace selected during catalog discovery. */
     gitlabDuoWorkflowRootNamespaceId?: string;
+    /** Cursor `max_mode` request flag returned by `GetUsableModels` for premium models that require max mode. */
+    cursorMaxMode?: boolean;
     cost: {
         input: number;
         output: number;

package/dist/types/wire/github-copilot.d.ts

@@ -28,13 +28,7 @@ export type ParsedGitHubCopilotApiKey = {
     apiEndpoint?: string;
 };
 export declare function isPublicGitHubHost(host: string): boolean;
-/**
- * Canonical personal-Copilot API host. The business
- * (`api.business.githubcopilot.com`) and enterprise (`copilot-api.{domain}`)
- * endpoints respond with HTTP 400 "vision is not supported" on image inputs,
- * so catalog discovery and capability gates MUST honour the upstream's
- * `supports.vision` flag only for this exact base URL.
- */
+/** Canonical personal-Copilot API host. */
 export declare const PERSONAL_GITHUB_COPILOT_BASE_URL: "https://api.githubcopilot.com";
 /** `true` when the resolved base URL is the canonical personal-Copilot host. */
 export declare function isPersonalGitHubCopilotBaseUrl(baseUrl: string | undefined): boolean;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "jeopi-catalog",
-	"version": "16.4.2",
+	"version": "16.4.3",
 	"description": "Model catalog for jeopi: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"homepage": "https://github.com/akillness/jeopi",
 	"author": "Can Boluk",
@@ -34,12 +34,12 @@
 	},
 	"dependencies": {
 		"@bufbuild/protobuf": "^2.12.0",
-		"jeopi-utils": "16.4.2",
+		"jeopi-utils": "16.4.3",
 		"arktype": "^2.2.0",
 		"zod": "^4"
 	},
 	"devDependencies": {
-		"jeopi-ai": "16.4.2",
+		"jeopi-ai": "16.4.3",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/discovery/cursor.ts

@@ -28,6 +28,7 @@ const CursorModelDetailsSchema = type({
 	displayModelId: OptionalDisplayNameSchema.default(undefined),
 	aliases: CursorAliasesSchema.default(() => []),
 	"thinkingDetails?": "unknown",
+	maxMode: "boolean = false",
 });
 
 const CursorModelsInnerSchema = type("unknown[]");
@@ -283,6 +284,7 @@ function normalizeCursorModel(
 			name,
 			baseUrl: baseUrlOverride ?? reference.baseUrl,
 			reasoning,
+			cursorMaxMode: details.maxMode,
 		};
 	}
 	return {
@@ -296,6 +298,7 @@ function normalizeCursorModel(
 		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
 		contextWindow: DEFAULT_CONTEXT_WINDOW,
 		maxTokens: DEFAULT_MAX_TOKENS,
+		cursorMaxMode: details.maxMode,
 	};
 }
 

package/src/identity/classify.ts

@@ -106,15 +106,20 @@ export const parseGeminiModel = parser((modelId): GeminiModel | null => {
 });
 
 export const parseAnthropicModel = parser((modelId): AnthropicModel | null => {
-	const match = /claude-(opus|sonnet|fable|mythos)-(\d{1,2}(?:[.-]\d{1,2}){0,2})\b/.exec(modelId);
-	if (!match) {
+	const kindFirst = /claude-(opus|sonnet|fable|mythos)-(\d{1,2}(?:[.-]\d{1,2}){0,2})\b/.exec(modelId);
+	const versionFirst = kindFirst
+		? null
+		: /claude-(\d{1,2}(?:[.-]\d{1,2}){0,2})-(opus|sonnet|fable|mythos)\b/.exec(modelId);
+	const kind = kindFirst?.[1] ?? versionFirst?.[2];
+	const versionInput = kindFirst?.[2] ?? versionFirst?.[1];
+	if (!kind || !versionInput) {
 		return null;
 	}
-	const version = parseSemVer(match[2]);
+	const version = parseSemVer(versionInput);
 	if (!version) {
 		return null;
 	}
-	return { family: "anthropic", kind: match[1] as AnthropicKind, version };
+	return { family: "anthropic", kind: kind as AnthropicKind, version };
 });
 
 export const parseOpenAIModel = parser((modelId): OpenAIModel | null => {

package/src/model-manager.ts

@@ -152,8 +152,9 @@ export async function resolveProviderModels<TApi extends Api = Api, TModelsDevPa
 	const dynamicFetchSucceeded = fetchedDynamicModels !== null;
 	const cacheModels = dynamicFetchSucceeded
 		? []
-		: dropCachedModelIdsOnStaticMismatch(
+		: prepareCacheModelsForStaticMismatch(
 				normalizeModelList<TApi>(cache?.models ?? []),
+				staticModels,
 				cacheFingerprintMatches,
 				options.dropCachedModelIdsOnStaticMismatch,
 			);
@@ -188,8 +189,9 @@ export async function resolveProviderModels<TApi extends Api = Api, TModelsDevPa
 				collapseBuiltModelVariants(
 					mergeDynamicModels(
 						mergeModelSources(staticModels, modelsDevModels),
-						dropCachedModelIdsOnStaticMismatch(
+						prepareCacheModelsForStaticMismatch(
 							normalizeModelList<TApi>(latestCache?.models ?? cache?.models ?? []),
+							staticModels,
 							cacheFingerprintMatches,
 							options.dropCachedModelIdsOnStaticMismatch,
 						),
@@ -260,16 +262,29 @@ function shouldFetchRemoteSources(
 	return false;
 }
 
-function dropCachedModelIdsOnStaticMismatch<TApi extends Api>(
+function prepareCacheModelsForStaticMismatch<TApi extends Api>(
 	models: readonly Model<TApi>[],
+	staticModels: 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];
+	if (models.length === 0) {
+		return [];
+	}
+	if (cacheFingerprintMatches) {
+		return [...models];
+	}
+
+	const droppedIds = ids && ids.length > 0 ? new Set(ids) : undefined;
+	const staticIds = staticModels.length > 0 ? new Set(staticModels.map(model => model.id)) : undefined;
+	const sanitizedModels: Model<TApi>[] = [];
+	for (const model of models) {
+		if (droppedIds?.has(model.id)) {
+			continue;
+		}
+		sanitizedModels.push(staticIds?.has(model.id) ? { ...model, contextWindow: null, maxTokens: null } : model);
 	}
-	const droppedIds = new Set(ids);
-	return models.filter(model => !droppedIds.has(model.id));
+	return sanitizedModels;
 }
 
 function mergeModelSources<TApi extends Api>(...sources: readonly (readonly Model<TApi>[])[]): Model<TApi>[] {
@@ -351,11 +366,14 @@ function fingerprintStatic<TApi extends Api>(
 function mergeDynamicModel<TApi extends Api>(existingModel: Model<TApi>, dynamicModel: Model<TApi>): Model<TApi> {
 	// When discovery resolves the same model id to a different endpoint (e.g.
 	// a GitHub Copilot business/enterprise host), the bundled reference's
-	// capabilities are pinned to the canonical host and no longer apply —
-	// honour the dynamic value alone. Same-endpoint merges still OR-upgrade so
-	// a discovery that omits the capability flag doesn't drop bundled vision.
+	// capabilities are pinned to another endpoint and no longer apply. Copilot
+	// dynamic discovery also pre-applies the correct image fallback for omitted
+	// `supports.vision`, so its explicit `false` must not be OR-upgraded by the
+	// canonical bundled model.
 	const endpointChanged = existingModel.baseUrl !== dynamicModel.baseUrl;
-	const supportsImage = endpointChanged
+	const dynamicInputAuthoritative =
+		endpointChanged || (existingModel.provider === "github-copilot" && dynamicModel.provider === "github-copilot");
+	const supportsImage = dynamicInputAuthoritative
 		? dynamicModel.input.includes("image")
 		: existingModel.input.includes("image") || dynamicModel.input.includes("image");
 	// Re-build from spec stage: sparse compat comes from `compatConfig` (the

package/src/model-thinking.ts

@@ -306,7 +306,7 @@ function getModelDefinedEfforts<TApi extends Api>(
 		if (isZaiThinkingFormat(compat) || isOpenRouterThinkingFormat(compat)) {
 			return DEFAULT_REASONING_EFFORTS_WITH_XHIGH;
 		}
-		if (isUmansGlm52ReasoningEffortModel(spec) || isOllamaCloudGlm52ReasoningEffortModel(spec)) {
+		if (isAnthropicMessagesGlm52ReasoningEffortModel(spec) || isOllamaCloudGlm52ReasoningEffortModel(spec)) {
 			return GLM_52_HIGH_MAX_REASONING_EFFORTS;
 		}
 	}
@@ -325,8 +325,12 @@ function isOllamaCloudGlm52ReasoningEffortModel<TApi extends Api>(spec: ModelSpe
 	return spec.api === "ollama-chat" && spec.provider === "ollama-cloud" && isGlm52ReasoningEffortModelId(spec.id);
 }
 
-function isUmansGlm52ReasoningEffortModel<TApi extends Api>(spec: ModelSpec<TApi>): boolean {
-	return spec.api === "anthropic-messages" && spec.provider === "umans" && isGlm52ReasoningEffortModelId(spec.id);
+function isAnthropicMessagesGlm52ReasoningEffortModel<TApi extends Api>(spec: ModelSpec<TApi>): boolean {
+	return (
+		spec.api === "anthropic-messages" &&
+		(spec.provider === "umans" || spec.provider === "zai") &&
+		isGlm52ReasoningEffortModelId(spec.id)
+	);
 }
 
 function isMinimaxReasoningModelOnAnthropicEndpoint<TApi extends Api>(spec: ModelSpec<TApi>): boolean {
@@ -382,7 +386,8 @@ function inferDetectedEffortMap<TApi extends Api>(
 	//     `xhigh` 400s — collapse minimal->none, low/medium/high->high, xhigh->max.
 	//   - OpenRouter: `max` 400s and `xhigh` IS its max tier, so it passes `xhigh`
 	//     through literally (no map; the tier is exposed via getModelDefinedEfforts).
-	//   - Umans and Ollama Cloud expose only high/max on their GLM-5.2 routes.
+	//   - Umans, Ollama Cloud, and Z.ai's anthropic-messages proxy expose only
+	//     high/max on their GLM-5.2 routes.
 	//   - Other openai-compat hosts (Fireworks, resellers) keep their distinct
 	//     lower tiers and host quirks (e.g. Fireworks rejects `minimal`, so
 	//     `minimal->none` stays) and only remap the top `xhigh` UI tier onto the
@@ -391,7 +396,7 @@ function inferDetectedEffortMap<TApi extends Api>(
 	if (isGlm52 && isZaiThinkingFormat(compat)) {
 		return ZAI_GLM_52_REASONING_EFFORT_MAP;
 	}
-	if (isUmansGlm52ReasoningEffortModel(spec) || isOllamaCloudGlm52ReasoningEffortModel(spec)) {
+	if (isAnthropicMessagesGlm52ReasoningEffortModel(spec) || isOllamaCloudGlm52ReasoningEffortModel(spec)) {
 		return GLM_52_XHIGH_MAX_EFFORT_MAP;
 	}
 	if (isSakanaFuguReasoningModel(spec)) {
@@ -581,7 +586,7 @@ function inferThinkingControlMode<TApi extends Api>(
 			if (isMinimaxReasoningModelOnAnthropicEndpoint(spec)) {
 				return "anthropic-adaptive";
 			}
-			if (isUmansGlm52ReasoningEffortModel(spec)) {
+			if (isAnthropicMessagesGlm52ReasoningEffortModel(spec)) {
 				return "anthropic-budget-effort";
 			}
 			if (parsedModel.family === "anthropic") {

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

@@ -3510,16 +3510,13 @@ export function githubCopilotModelManagerOptions(config?: GithubCopilotModelMana
 								? entry.name
 								: (reference?.name ?? defaults.name);
 						const api = inferCopilotApi(defaults.id);
-						// `supports.vision` reports the model's intrinsic capability, but
-						// the business/enterprise endpoints respond `400 vision is not
-						// supported` on image inputs. Only honour the flag for the
-						// canonical personal-Copilot host.
 						const supportsVision = extractCopilotSupportsVision(entry);
-						const input: ModelSpec<Api>["input"] = isPersonalGitHubCopilotBaseUrl(baseUrl)
-							? supportsVision
+						const input: ModelSpec<Api>["input"] =
+							supportsVision === true
 								? ["text", "image"]
-								: (reference?.input ?? defaults.input)
-							: ["text"];
+								: supportsVision === false || !isPersonalGitHubCopilotBaseUrl(baseUrl)
+									? ["text"]
+									: (reference?.input ?? defaults.input);
 						// With COPILOT_API_HEADERS the served window is the long-context
 						// ceiling; the default tier ends at token_prices.default.context_max
 						// prompt tokens. Cap the base entry to the default tier — the long

package/src/provider-models/special.ts

@@ -42,10 +42,13 @@ export interface CursorModelManagerConfig {
 	clientVersion?: string;
 }
 
+const CURSOR_CACHE_PROVIDER_ID = "cursor:max-mode-v2";
+
 export function cursorModelManagerOptions(config: CursorModelManagerConfig = {}): ModelManagerOptions<"cursor-agent"> {
 	const { apiKey, baseUrl, clientVersion } = config;
 	return {
 		providerId: "cursor",
+		cacheProviderId: CURSOR_CACHE_PROVIDER_ID,
 		...(apiKey
 			? {
 					fetchDynamicModels: async () => {

package/src/types.ts

@@ -701,6 +701,8 @@ export interface Model<TApi extends Api = Api> {
 	supportsTools?: boolean;
 	/** GitLab Duo Workflow root namespace selected during catalog discovery. */
 	gitlabDuoWorkflowRootNamespaceId?: string;
+	/** Cursor `max_mode` request flag returned by `GetUsableModels` for premium models that require max mode. */
+	cursorMaxMode?: boolean;
 	cost: {
 		input: number; // $/million tokens
 		output: number; // $/million tokens

package/src/wire/github-copilot.ts

@@ -45,13 +45,7 @@ export function isPublicGitHubHost(host: string): boolean {
 	return PUBLIC_GITHUB_HOSTS.has(host.trim().toLowerCase());
 }
 
-/**
- * Canonical personal-Copilot API host. The business
- * (`api.business.githubcopilot.com`) and enterprise (`copilot-api.{domain}`)
- * endpoints respond with HTTP 400 "vision is not supported" on image inputs,
- * so catalog discovery and capability gates MUST honour the upstream's
- * `supports.vision` flag only for this exact base URL.
- */
+/** Canonical personal-Copilot API host. */
 export const PERSONAL_GITHUB_COPILOT_BASE_URL = "https://api.githubcopilot.com" as const;
 
 /** `true` when the resolved base URL is the canonical personal-Copilot host. */