@oh-my-pi/pi-catalog 16.0.7 → 16.0.9

package/CHANGELOG.md

@@ -2,6 +2,23 @@
 
 ## [Unreleased]
 
+## [16.0.9] - 2026-06-18
+
+### Fixed
+
+- Fixed GitHub Copilot's `anthropic-messages` proxy being misclassified as a non-signing reasoning endpoint (`replayUnsignedThinking: true`). It forwards to signature-enforcing Anthropic, so replaying a stripped/unsigned historical `thinking` block as `signature: ""` — most visibly an end_turn-bound checkpoint/branch-return turn whose signature the transform must strip — caused a `400 Invalid signature` that corrupted the session and re-tripped on every full history re-send (e.g. after toggling MCP servers). Copilot now degrades such blocks to text like the official API. ([#2851](https://github.com/can1357/oh-my-pi/issues/2851))
+- Added a `supportsImageDetailOriginal` compat flag that resolves to `false` for GitHub Copilot, whose Responses endpoint rejects the `detail: "original"` image hint with a 400, and `true` for every other host. ([#2822](https://github.com/can1357/oh-my-pi/issues/2822))
+
+## [16.0.8] - 2026-06-18
+
+### Changed
+
+- Refactored model family ID predicates and capability checkers to use a shared, uniform process-lifetime `memo` utility to eliminate caching boilerplate.
+
+### Fixed
+
+- Fixed LM Studio dynamic discovery to use native `/api/v0/models` metadata so VLM models advertise image input. ([#2945](https://github.com/can1357/oh-my-pi/issues/2945))
+
 ## [16.0.7] - 2026-06-18
 
 ### Fixed

package/dist/types/identity/classify.d.ts

@@ -41,7 +41,6 @@ export interface UnknownModel {
     id: string;
 }
 export type ParsedModel = GeminiModel | AnthropicModel | OpenAIModel | UnknownModel;
-/** Strip a provider namespace prefix (`openai/gpt-5.4` → `gpt-5.4`). */
 export declare function bareModelId(modelId: string): string;
 export declare function parseKnownModel(modelId: string): ParsedModel;
 export declare const parseGeminiModel: (modelId: string) => GeminiModel | null;

package/dist/types/identity/family.d.ts

@@ -7,27 +7,27 @@
  * here.
  */
 /** Kimi family ids in any namespace form (`moonshotai/kimi-*`, `kimi-k2.6`, `vendor/kimi.x`). */
-export declare function isKimiModelId(modelId: string): boolean;
+export declare const isKimiModelId: (modelId: string) => boolean;
 /** Kimi K2.6 specifically, including router ids that spell the version `k2p6`. */
-export declare function isKimiK26ModelId(modelId: string): boolean;
+export declare const isKimiK26ModelId: (modelId: string) => boolean;
 /** Claude ids in any namespace form (`claude-*`, `vendor/claude.x`). */
-export declare function isClaudeModelId(modelId: string): boolean;
+export declare const isClaudeModelId: (modelId: string) => boolean;
 /** `anthropic/`-namespaced ids (aggregator catalogs like OpenRouter). */
-export declare function isAnthropicNamespacedModelId(modelId: string): boolean;
+export declare const isAnthropicNamespacedModelId: (modelId: string) => boolean;
 /** Qwen family ids (substring match — Qwen SKUs have no stable prefix shape). */
-export declare function isQwenModelId(modelId: string): boolean;
+export declare const isQwenModelId: (modelId: string) => boolean;
 /** Gemma open-weights family (`gemma-3-27b-it`, `google/gemma-4-E2B-it`, `gemma2-9b`). */
-export declare function isGemmaModelId(modelId: string): boolean;
+export declare const isGemmaModelId: (modelId: string) => boolean;
 /** DeepSeek family by id or display name (proxies often rename the id but keep the name). */
-export declare function isDeepseekModelIdOrName(value: string): boolean;
+export declare const isDeepseekModelIdOrName: (modelId: string) => boolean;
 /** Xiaomi MiMo family by id or display name. */
-export declare function isMimoModelIdOrName(value: string): boolean;
+export declare const isMimoModelIdOrName: (modelId: string) => boolean;
 /**
  * Grok SKUs that expose the wire `reasoning.effort` dial. Other Grok reasoners
  * (e.g. `grok-build`, `grok-4.20-0309-reasoning`) think natively but reject the
  * param, so callers must omit reasoning effort for them.
  */
-export declare function isGrokReasoningEffortCapable(modelId: string): boolean;
+export declare const isGrokReasoningEffortCapable: (modelId: string) => boolean;
 /**
  * MiniMax M2-generation family (M2, M2.1, M2.5, M2.7, including `-highspeed`/
  * `-lightning`/`-her`/`-turbo` variants, dotless aliases like `minimax-m21`,
@@ -37,18 +37,18 @@ export declare function isGrokReasoningEffortCapable(modelId: string): boolean;
  * `minimal` to `none` (Fireworks) or expects the full 5-tier scale must
  * clamp instead. Excludes M1, M3, MiniMax-Text-01, music, hailuo, voice ids.
  */
-export declare function isMinimaxM2FamilyModelId(modelId: string): boolean;
+export declare const isMinimaxM2FamilyModelId: (modelId: string) => boolean;
 /** MiniMax M3 family ids in bundled/default and aggregator namespace forms. */
-export declare function isMinimaxM3FamilyModelId(modelId: string): boolean;
+export declare const isMinimaxM3FamilyModelId: (modelId: string) => boolean;
 /**
  * OpenAI gpt-oss family (`gpt-oss-20b`, `gpt-oss-120b`, `gpt-oss:120b`,
  * `vendor/gpt-oss-…`). The Harmony reasoning format only accepts
  * `low|medium|high` for `reasoning_effort` and rejects `minimal`, `xhigh`,
  * and `none`.
  */
-export declare function isOpenAIGptOssModelId(modelId: string): boolean;
+export declare const isOpenAIGptOssModelId: (modelId: string) => boolean;
 /** OpenAI model ids (gpt-*, o1-*, o3-*, o4-*, or prefixed with openai/). */
-export declare function isOpenAIModelId(modelId: string): boolean;
+export declare const isOpenAIModelId: (modelId: string) => boolean;
 /**
  * Reasoning-capable GLM coding SKUs: glm-4.5 and up on the base / `-air` /
  * `-turbo` lines. Excludes the vision (`…v`) shape, the non-reasoning
@@ -56,11 +56,11 @@ export declare function isOpenAIModelId(modelId: string): boolean;
  * keeps newly-bumped integers (`glm-5.3`, `glm-6`, …) covered without a per-id
  * allowlist.
  */
-export declare function isReasoningGlmModelId(modelId: string): boolean;
+export declare const isReasoningGlmModelId: (modelId: string) => boolean;
 /** GLM-5.2+ coding SKUs accept `reasoning_effort` in addition to binary thinking. */
-export declare function isGlm52ReasoningEffortModelId(modelId: string): boolean;
+export declare const isGlm52ReasoningEffortModelId: (modelId: string) => boolean;
 /** GLM vision SKUs — the `v` that attaches to the version (`glm-4v`, `glm-4.5v`). */
-export declare function isGlmVisionModelId(modelId: string): boolean;
+export declare const isGlmVisionModelId: (modelId: string) => boolean;
 /**
  * Coarse vendor-lineage token for "are two models the same family?" checks
  * (e.g. picking a cross-family reviewer). All Claude point releases share a token,
@@ -72,7 +72,7 @@ export declare function isGlmVisionModelId(modelId: string): boolean;
  * Vendor-only by design: a model's kind/variant (opus vs sonnet, codex vs base) is
  * collapsed onto the single vendor token; use {@link parseKnownModel} for finer breakdowns.
  */
-export declare function modelFamilyToken(modelId: string): string;
+export declare const modelFamilyToken: (modelId: string) => string;
 /**
  * Adaptive thinking `display` is supported starting with Claude Opus 4.7 and
  * the Claude Fable/Mythos 5 generation. Older adaptive-thinking models
@@ -80,13 +80,13 @@ export declare function modelFamilyToken(modelId: string): string;
  * dashed version forms both match while bare dated ids
  * (`claude-opus-4-20250514` = Opus 4.0) stay excluded.
  */
-export declare function supportsAdaptiveThinkingDisplay(modelId: string): boolean;
+export declare const supportsAdaptiveThinkingDisplay: (modelId: string) => boolean;
 /**
  * Returns true for Anthropic models with Opus 4.7+/Fable/Mythos API restrictions:
  * - Sampling parameters (temperature/top_p/top_k) return 400 error
  * - Thinking content is omitted by default (needs display: "summarized")
  */
-export declare function hasOpus47ApiRestrictions(modelId: string): boolean;
+export declare const hasOpus47ApiRestrictions: (modelId: string) => boolean;
 /**
  * Mid-conversation `role: "system"` messages (system instructions appended at
  * non-first positions in the `messages` array) are supported starting with
@@ -94,8 +94,8 @@ export declare function hasOpus47ApiRestrictions(modelId: string): boolean;
  * models reject the role.
  * @see https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages
  */
-export declare function supportsMidConversationSystemMessages(modelId: string): boolean;
-export declare function isAnthropicFableOrMythosModel(modelId: string): boolean;
+export declare const supportsMidConversationSystemMessages: (modelId: string) => boolean;
+export declare const isAnthropicFableOrMythosModel: (modelId: string) => boolean;
 /** Thinking-variant token location inside a model id. */
 export interface ThinkingVariantToken {
     index: number;
@@ -115,4 +115,4 @@ export declare function findThinkingVariantToken(modelId: string): ThinkingVaria
  * token exists or nothing would remain. Callers MUST verify the result names
  * a live model.
  */
-export declare function stripThinkingVariantToken(modelId: string): string | undefined;
+export declare const stripThinkingVariantToken: (modelId: string) => string | undefined;

package/dist/types/provider-models/openai-compat.d.ts

@@ -265,6 +265,18 @@ export interface KimiCodeModelManagerConfig {
     fetch?: FetchImpl;
 }
 export declare function kimiCodeModelManagerOptions(config?: KimiCodeModelManagerConfig): ModelManagerOptions<"openai-completions">;
+/** Native LM Studio metadata keyed by model id from `/api/v0/models`. */
+export interface LmStudioNativeModelMetadata {
+    input: ("text" | "image")[];
+    contextWindow?: number;
+}
+/** Options for LM Studio's optional native metadata probe. */
+export interface LmStudioNativeModelMetadataOptions {
+    headers?: Record<string, string>;
+    signal?: AbortSignal;
+}
+/** Fetches LM Studio native model metadata used to mark VLM models as image-capable. */
+export declare function fetchLmStudioNativeModelMetadata(baseUrl: string, fetchImpl?: FetchImpl, options?: LmStudioNativeModelMetadataOptions): Promise<Map<string, LmStudioNativeModelMetadata> | null>;
 export interface LmStudioModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;

package/dist/types/types.d.ts

@@ -247,6 +247,8 @@ export interface OpenAICompat {
     alwaysSendMaxTokens?: boolean;
     /** Whether Responses-API tool-call/result history must be strictly paired. Default: auto-detected (Azure OpenAI, GitHub Copilot). */
     strictResponsesPairing?: boolean;
+    /** Whether the Responses API accepts the `detail: "original"` image hint. Default: auto-detected (false for GitHub Copilot, which rejects it with a 400). */
+    supportsImageDetailOriginal?: boolean;
     /**
      * Append a trailing `# Juice: 0 !important` developer item when the caller
      * did not request reasoning, suppressing default reasoning on models that
@@ -416,7 +418,7 @@ export interface ResolvedOpenAISharedCompat {
  * `buildModel`; request handlers read fields and never detect, resolve, or
  * allocate.
  */
-export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "reasoningEffortMap" | "supportsReasoningParams" | "thinkingFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "supportsToolChoice" | "supportsForcedToolChoice" | "reasoningContentField" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamIdleTimeoutMs" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "requiresJuiceZeroHack" | "enableGeminiThinkingLoopGuard" | "whenThinking">> & {
+export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "reasoningEffortMap" | "supportsReasoningParams" | "thinkingFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "supportsToolChoice" | "supportsForcedToolChoice" | "reasoningContentField" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamIdleTimeoutMs" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "supportsImageDetailOriginal" | "requiresJuiceZeroHack" | "enableGeminiThinkingLoopGuard" | "whenThinking">> & {
     vercelGatewayRouting?: OpenAICompat["vercelGatewayRouting"];
     extraBody?: OpenAICompat["extraBody"];
     cacheControlFormat?: OpenAICompat["cacheControlFormat"];
@@ -434,6 +436,7 @@ export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<Op
 export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompat {
     supportsLongPromptCacheRetention: boolean;
     strictResponsesPairing: boolean;
+    supportsImageDetailOriginal: boolean;
     requiresJuiceZeroHack: boolean;
     supportsObfuscationOptOut: boolean;
 }

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "16.0.7",
+	"version": "16.0.9",
 	"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": "16.0.7",
+		"@oh-my-pi/pi-utils": "16.0.9",
 		"arktype": "^2.2.0",
 		"zod": "^4"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "16.0.7",
+		"@oh-my-pi/pi-ai": "16.0.9",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/anthropic.ts

@@ -66,7 +66,16 @@ export function buildAnthropicCompat(spec: ModelSpec<"anthropic-messages">): Res
 		// loses the reasoning chain and can destabilize the next tool-call
 		// arguments (#2005). Known non-signing hosts (Z.AI, DeepSeek) are also
 		// preserved for compatibility.
-		replayUnsignedThinking: isZai || modelMatchesHost(spec, "deepseekFamily") || (spec.reasoning && !official),
+		//
+		// GitHub Copilot's `anthropic-messages` proxy is excluded: it forwards to
+		// signature-enforcing Anthropic and returns full thinking signatures, so it
+		// is a SIGNING endpoint. Replaying a stripped/unsigned thinking block as
+		// `signature: ""` there 400s the whole request ("Invalid signature") — most
+		// visibly when a checkpoint/branch-return turn's end_turn-bound signature is
+		// stripped on replay (issue #2851). Treating it like official Anthropic
+		// degrades such blocks to text instead, which the API accepts.
+		replayUnsignedThinking:
+			!isCopilot && (isZai || modelMatchesHost(spec, "deepseekFamily") || (spec.reasoning && !official)),
 		escapeBuiltinToolNames: modelMatchesHost(spec, "umans"),
 	};
 	applyCompatOverrides(compat, spec.compat);

package/src/compat/openai.ts

@@ -459,6 +459,12 @@ export function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): Resol
 		// Azure OpenAI and GitHub Copilot Responses paths require tool results
 		// to strictly match prior tool calls when building Responses inputs.
 		strictResponsesPairing: isAzure || spec.provider === "github-copilot",
+		// GitHub Copilot's Responses endpoint rejects the `detail: "original"`
+		// image hint with a 400; every other host preserves native-resolution
+		// frames (snapcompact relies on `original`). Detect Copilot by provider id
+		// or base-URL host (mirroring the Anthropic compat builder) so a model
+		// pointed at the Copilot host under a different provider id still clamps.
+		supportsImageDetailOriginal: !modelMatchesHost({ provider: spec.provider, baseUrl }, "githubCopilot"),
 		requiresJuiceZeroHack: spec.name.toLowerCase().startsWith("gpt-5"),
 		reasoningEffortMap: {},
 		supportsReasoningParams: true,
@@ -514,6 +520,7 @@ function pickResponsesOnly(compat: ResolvedOpenAIResponsesCompat): ResponsesOnly
 	return {
 		supportsLongPromptCacheRetention: compat.supportsLongPromptCacheRetention,
 		strictResponsesPairing: compat.strictResponsesPairing,
+		supportsImageDetailOriginal: compat.supportsImageDetailOriginal,
 		requiresJuiceZeroHack: compat.requiresJuiceZeroHack,
 		supportsObfuscationOptOut: compat.supportsObfuscationOptOut,
 	} satisfies ResponsesOnlyCompat;

package/src/hosts.ts

@@ -16,7 +16,7 @@ interface HostClassSpec {
 	readonly providers?: readonly string[];
 	/** Provider-id prefixes that imply this host class (e.g. `xiaomi-token-plan-`). */
 	readonly providerPrefixes?: readonly string[];
-	/** Case-insensitive substrings matched against the base URL. */
+	/** Lowercase ASCII substrings matched case-insensitively against the base URL. */
 	readonly urlMarkers: readonly string[];
 	// Strict hostname matching is intentionally not modeled here: the one
 	// auth-sensitive consumer (Anthropic official-endpoint) parses the URL
@@ -68,9 +68,8 @@ export type KnownHost = keyof typeof KNOWN_HOSTS;
 export function hostMatchesUrl(baseUrl: string | undefined, host: KnownHost): boolean {
 	if (!baseUrl) return false;
 	const spec: HostClassSpec = KNOWN_HOSTS[host];
-	const normalized = baseUrl.toLowerCase();
 	for (const marker of spec.urlMarkers) {
-		if (normalized.includes(marker)) return true;
+		if (includesAsciiCaseInsensitive(baseUrl, marker)) return true;
 	}
 	return false;
 }
@@ -91,6 +90,19 @@ export function modelMatchesHost(model: { provider: string; baseUrl: string }, h
 	return hostMatchesUrl(model.baseUrl, host);
 }
 
+function includesAsciiCaseInsensitive(value: string, lowerNeedle: string): boolean {
+	const needleLength = lowerNeedle.length;
+	const end = value.length - needleLength;
+	for (let start = 0; start <= end; start++) {
+		let offset = 0;
+		for (; offset < needleLength; offset++) {
+			if ((value.charCodeAt(start + offset) | 0x20) !== lowerNeedle.charCodeAt(offset)) break;
+		}
+		if (offset === needleLength) return true;
+	}
+	return false;
+}
+
 // --- Endpoint-shape predicates (URL path/verb shapes, not vendor hosts) ---
 
 /** Vertex AI express-mode OpenAI-compatible endpoint (`…/endpoints/openapi`). */

package/src/identity/classify.ts

@@ -51,9 +51,15 @@ export interface UnknownModel {
 export type ParsedModel = GeminiModel | AnthropicModel | OpenAIModel | UnknownModel;
 
 /** Strip a provider namespace prefix (`openai/gpt-5.4` → `gpt-5.4`). */
+// Cache keyed by model id (a bounded set of bundled/aggregator ids), so no eviction is needed.
+const bareModelIdCache = new Map<string, string>();
 export function bareModelId(modelId: string): string {
+	const cached = bareModelIdCache.get(modelId);
+	if (cached !== undefined) return cached;
 	const p = modelId.lastIndexOf("/");
-	return p !== -1 ? modelId.slice(p + 1) : modelId;
+	const result = p !== -1 ? modelId.slice(p + 1) : modelId;
+	bareModelIdCache.set(modelId, result);
+	return result;
 }
 
 export function parseKnownModel(modelId: string): ParsedModel {

package/src/identity/family.ts

@@ -16,45 +16,58 @@ import {
 	semverGte,
 } from "./classify";
 
+/** Bounded process-lifetime cache memo helper. */
+function memo<T>(fn: (modelId: string) => T): (modelId: string) => T {
+	const cache = new Map<string, T>();
+	return (modelId: string) => {
+		if (cache.has(modelId)) {
+			return cache.get(modelId) as T;
+		}
+		const result = fn(modelId);
+		cache.set(modelId, result);
+		return result;
+	};
+}
+
 /** Kimi family ids in any namespace form (`moonshotai/kimi-*`, `kimi-k2.6`, `vendor/kimi.x`). */
-export function isKimiModelId(modelId: string): boolean {
+export const isKimiModelId = memo((modelId: string): boolean => {
 	return modelId.includes("moonshotai/kimi") || /(^|\/)kimi[-.]/i.test(modelId);
-}
+});
 
 /** Kimi K2.6 specifically, including router ids that spell the version `k2p6`. */
-export function isKimiK26ModelId(modelId: string): boolean {
+export const isKimiK26ModelId = memo((modelId: string): boolean => {
 	return /(^|\/)kimi-k2(?:\.6|p6)(?:[-:]|$)/i.test(modelId);
-}
+});
 
 /** Claude ids in any namespace form (`claude-*`, `vendor/claude.x`). */
-export function isClaudeModelId(modelId: string): boolean {
+export const isClaudeModelId = memo((modelId: string): boolean => {
 	return /(^|\/)claude[-.]/i.test(modelId);
-}
+});
 
 /** `anthropic/`-namespaced ids (aggregator catalogs like OpenRouter). */
-export function isAnthropicNamespacedModelId(modelId: string): boolean {
+export const isAnthropicNamespacedModelId = memo((modelId: string): boolean => {
 	return /(^|\/)anthropic\//i.test(modelId);
-}
+});
 
 /** Qwen family ids (substring match — Qwen SKUs have no stable prefix shape). */
-export function isQwenModelId(modelId: string): boolean {
+export const isQwenModelId = memo((modelId: string): boolean => {
 	return modelId.toLowerCase().includes("qwen");
-}
+});
 
 /** Gemma open-weights family (`gemma-3-27b-it`, `google/gemma-4-E2B-it`, `gemma2-9b`). */
-export function isGemmaModelId(modelId: string): boolean {
+export const isGemmaModelId = memo((modelId: string): boolean => {
 	return /(^|\/)gemma[-.]?\d/i.test(modelId);
-}
+});
 
 /** DeepSeek family by id or display name (proxies often rename the id but keep the name). */
-export function isDeepseekModelIdOrName(value: string): boolean {
+export const isDeepseekModelIdOrName = memo((value: string): boolean => {
 	return value.toLowerCase().includes("deepseek");
-}
+});
 
 /** Xiaomi MiMo family by id or display name. */
-export function isMimoModelIdOrName(value: string): boolean {
+export const isMimoModelIdOrName = memo((value: string): boolean => {
 	return value.toLowerCase().includes("mimo");
-}
+});
 
 const GROK_EFFORT_CAPABLE_PREFIXES = ["grok-3-mini", "grok-4.20-multi-agent", "grok-4.3"] as const;
 
@@ -63,11 +76,11 @@ const GROK_EFFORT_CAPABLE_PREFIXES = ["grok-3-mini", "grok-4.20-multi-agent", "g
  * (e.g. `grok-build`, `grok-4.20-0309-reasoning`) think natively but reject the
  * param, so callers must omit reasoning effort for them.
  */
-export function isGrokReasoningEffortCapable(modelId: string): boolean {
+export const isGrokReasoningEffortCapable = memo((modelId: string): boolean => {
 	const bare = bareModelId(modelId).trim().toLowerCase();
 	if (!bare) return false;
 	return GROK_EFFORT_CAPABLE_PREFIXES.some(prefix => bare.startsWith(prefix));
-}
+});
 
 /**
  * MiniMax M2-generation family (M2, M2.1, M2.5, M2.7, including `-highspeed`/
@@ -78,20 +91,20 @@ export function isGrokReasoningEffortCapable(modelId: string): boolean {
  * `minimal` to `none` (Fireworks) or expects the full 5-tier scale must
  * clamp instead. Excludes M1, M3, MiniMax-Text-01, music, hailuo, voice ids.
  */
-export function isMinimaxM2FamilyModelId(modelId: string): boolean {
+export const isMinimaxM2FamilyModelId = memo((modelId: string): boolean => {
 	const lower = modelId.toLowerCase();
 	if (!lower.includes("minimax")) return false;
 	// Boundary-delimited `m2` token followed by zero or more digits (dotless
 	// variants like `m21`/`m25`/`m27`) and an optional dotted minor version.
 	return /(?:^|[/.-])m2\d*(?:[.-]\d+)?(?:[-.:_]|$)/i.test(lower);
-}
+});
 
 /** MiniMax M3 family ids in bundled/default and aggregator namespace forms. */
-export function isMinimaxM3FamilyModelId(modelId: string): boolean {
+export const isMinimaxM3FamilyModelId = memo((modelId: string): boolean => {
 	const lower = modelId.toLowerCase();
 	if (!lower.includes("minimax")) return false;
 	return /(?:^|[/._-])(?:minimax[/._-])?m3(?:[-.:_]|$)/i.test(lower);
-}
+});
 
 /**
  * OpenAI gpt-oss family (`gpt-oss-20b`, `gpt-oss-120b`, `gpt-oss:120b`,
@@ -99,14 +112,14 @@ export function isMinimaxM3FamilyModelId(modelId: string): boolean {
  * `low|medium|high` for `reasoning_effort` and rejects `minimal`, `xhigh`,
  * and `none`.
  */
-export function isOpenAIGptOssModelId(modelId: string): boolean {
+export const isOpenAIGptOssModelId = memo((modelId: string): boolean => {
 	return /(^|\/)gpt-oss[-:]/i.test(modelId);
-}
+});
 
 /** OpenAI model ids (gpt-*, o1-*, o3-*, o4-*, or prefixed with openai/). */
-export function isOpenAIModelId(modelId: string): boolean {
+export const isOpenAIModelId = memo((modelId: string): boolean => {
 	return /(^|\/)(gpt|o1|o3|o4)[-.]/i.test(modelId) || modelId.toLowerCase().includes("openai/");
-}
+});
 
 /**
  * Reasoning-capable GLM coding SKUs: glm-4.5 and up on the base / `-air` /
@@ -115,7 +128,7 @@ export function isOpenAIModelId(modelId: string): boolean {
  * keeps newly-bumped integers (`glm-5.3`, `glm-6`, …) covered without a per-id
  * allowlist.
  */
-export function isReasoningGlmModelId(modelId: string): boolean {
+export const isReasoningGlmModelId = memo((modelId: string): boolean => {
 	const glm = parseGlmModel(bareModelId(modelId));
 	if (!glm || glm.vision) {
 		return false;
@@ -124,9 +137,10 @@ export function isReasoningGlmModelId(modelId: string): boolean {
 		return false;
 	}
 	return semverGte(glm.version, "4.5");
-}
+});
+
 /** GLM-5.2+ coding SKUs accept `reasoning_effort` in addition to binary thinking. */
-export function isGlm52ReasoningEffortModelId(modelId: string): boolean {
+export const isGlm52ReasoningEffortModelId = memo((modelId: string): boolean => {
 	const glm = parseGlmModel(bareModelId(modelId));
 	if (!glm || glm.vision) {
 		return false;
@@ -135,12 +149,13 @@ export function isGlm52ReasoningEffortModelId(modelId: string): boolean {
 		return false;
 	}
 	return semverGte(glm.version, "5.2");
-}
+});
 
 /** GLM vision SKUs — the `v` that attaches to the version (`glm-4v`, `glm-4.5v`). */
-export function isGlmVisionModelId(modelId: string): boolean {
+export const isGlmVisionModelId = memo((modelId: string): boolean => {
 	return parseGlmModel(bareModelId(modelId))?.vision === true;
-}
+});
+
 /**
  * Coarse vendor-lineage token for "are two models the same family?" checks
  * (e.g. picking a cross-family reviewer). All Claude point releases share a token,
@@ -152,7 +167,7 @@ export function isGlmVisionModelId(modelId: string): boolean {
  * Vendor-only by design: a model's kind/variant (opus vs sonnet, codex vs base) is
  * collapsed onto the single vendor token; use {@link parseKnownModel} for finer breakdowns.
  */
-export function modelFamilyToken(modelId: string): string {
+export const modelFamilyToken = memo((modelId: string): string => {
 	const parsed = parseKnownModel(modelId);
 	if (parsed.family !== "unknown") return parsed.family;
 	if (isClaudeModelId(modelId) || isAnthropicNamespacedModelId(modelId)) return "anthropic";
@@ -166,7 +181,7 @@ export function modelFamilyToken(modelId: string): string {
 	if (isGemmaModelId(modelId)) return "gemma";
 	if (parseGlmModel(bareModelId(modelId))) return "glm";
 	return "";
-}
+});
 
 /**
  * Adaptive thinking `display` is supported starting with Claude Opus 4.7 and
@@ -175,23 +190,23 @@ export function modelFamilyToken(modelId: string): string {
  * dashed version forms both match while bare dated ids
  * (`claude-opus-4-20250514` = Opus 4.0) stay excluded.
  */
-export function supportsAdaptiveThinkingDisplay(modelId: string): boolean {
+export const supportsAdaptiveThinkingDisplay = memo((modelId: string): boolean => {
 	const parsed = parseAnthropicModel(bareModelId(modelId));
 	if (!parsed) return false;
 	if (isFableOrMythos(parsed.kind)) return semverGte(parsed.version, "5");
 	return parsed.kind === "opus" && semverGte(parsed.version, "4.7");
-}
+});
 
 /**
  * Returns true for Anthropic models with Opus 4.7+/Fable/Mythos API restrictions:
  * - Sampling parameters (temperature/top_p/top_k) return 400 error
  * - Thinking content is omitted by default (needs display: "summarized")
  */
-export function hasOpus47ApiRestrictions(modelId: string): boolean {
+export const hasOpus47ApiRestrictions = memo((modelId: string): boolean => {
 	const parsed = parseAnthropicModel(bareModelId(modelId));
 	if (!parsed) return false;
 	return (parsed.kind === "opus" && semverGte(parsed.version, "4.7")) || isFableOrMythos(parsed.kind);
-}
+});
 
 /**
  * Mid-conversation `role: "system"` messages (system instructions appended at
@@ -200,16 +215,16 @@ export function hasOpus47ApiRestrictions(modelId: string): boolean {
  * models reject the role.
  * @see https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages
  */
-export function supportsMidConversationSystemMessages(modelId: string): boolean {
+export const supportsMidConversationSystemMessages = memo((modelId: string): boolean => {
 	const parsed = parseAnthropicModel(bareModelId(modelId));
 	if (!parsed) return false;
 	return (parsed.kind === "opus" && semverGte(parsed.version, "4.8")) || isFableOrMythos(parsed.kind);
-}
+});
 
-export function isAnthropicFableOrMythosModel(modelId: string): boolean {
+export const isAnthropicFableOrMythosModel = memo((modelId: string): boolean => {
 	const parsed = parseAnthropicModel(bareModelId(modelId));
 	return parsed !== null && isFableOrMythos(parsed.kind);
-}
+});
 
 /** Thinking-variant token location inside a model id. */
 export interface ThinkingVariantToken {
@@ -245,9 +260,9 @@ export function findThinkingVariantToken(modelId: string): ThinkingVariantToken
  * token exists or nothing would remain. Callers MUST verify the result names
  * a live model.
  */
-export function stripThinkingVariantToken(modelId: string): string | undefined {
+export const stripThinkingVariantToken = memo((modelId: string): string | undefined => {
 	const token = findThinkingVariantToken(modelId);
 	if (!token) return undefined;
 	const stripped = modelId.slice(0, token.index) + modelId.slice(token.index + token.length);
 	return stripped.length > 0 ? stripped : undefined;
-}
+});

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

@@ -2184,6 +2184,86 @@ export function kimiCodeModelManagerOptions(
 // 12.5. LM Studio
 // ---------------------------------------------------------------------------
 
+/** Native LM Studio metadata keyed by model id from `/api/v0/models`. */
+export interface LmStudioNativeModelMetadata {
+	input: ("text" | "image")[];
+	contextWindow?: number;
+}
+
+/** Options for LM Studio's optional native metadata probe. */
+export interface LmStudioNativeModelMetadataOptions {
+	headers?: Record<string, string>;
+	signal?: AbortSignal;
+}
+
+const LM_STUDIO_NATIVE_METADATA_TIMEOUT_MS = 250;
+
+function toLmStudioNativeBaseUrl(baseUrl: string): string {
+	const trimmed = baseUrl.trim();
+	const normalized = trimmed.endsWith("/") ? trimmed.slice(0, -1) : trimmed;
+	return normalized.endsWith("/v1") ? normalized.slice(0, -3) : normalized;
+}
+
+function getLmStudioCapabilityNames(value: unknown): string[] {
+	if (!Array.isArray(value)) {
+		return [];
+	}
+	return value.flatMap(item => (typeof item === "string" ? [item.toLowerCase()] : []));
+}
+
+function getLmStudioNativeInput(entry: Record<string, unknown>): ("text" | "image")[] {
+	const modelType = typeof entry.type === "string" ? entry.type.toLowerCase() : "";
+	const capabilities = getLmStudioCapabilityNames(entry.capabilities);
+	const supportsImage = modelType === "vlm" || capabilities.includes("vision") || capabilities.includes("image");
+	return supportsImage ? ["text", "image"] : ["text"];
+}
+
+function getLmStudioNativeContextWindow(entry: Record<string, unknown>): number | undefined {
+	return (
+		toPositiveNumber(entry.max_context_length, null) ??
+		toPositiveNumber(entry.context_length, null) ??
+		toPositiveNumber(entry.max_model_len, null) ??
+		undefined
+	);
+}
+
+/** Fetches LM Studio native model metadata used to mark VLM models as image-capable. */
+export async function fetchLmStudioNativeModelMetadata(
+	baseUrl: string,
+	fetchImpl: FetchImpl = fetch,
+	options?: LmStudioNativeModelMetadataOptions,
+): Promise<Map<string, LmStudioNativeModelMetadata> | null> {
+	const nativeBaseUrl = toLmStudioNativeBaseUrl(baseUrl);
+	try {
+		const response = await fetchImpl(`${nativeBaseUrl}/api/v0/models`, {
+			method: "GET",
+			headers: { Accept: "application/json", ...(options?.headers ?? {}) },
+			signal: options?.signal ?? AbortSignal.timeout(LM_STUDIO_NATIVE_METADATA_TIMEOUT_MS),
+		});
+		if (!response.ok) {
+			return null;
+		}
+		const payload = await response.json();
+		if (!isRecord(payload) || !Array.isArray(payload.data)) {
+			return null;
+		}
+		const metadata = new Map<string, LmStudioNativeModelMetadata>();
+		for (const entry of payload.data) {
+			if (!isRecord(entry) || typeof entry.id !== "string" || entry.id.length === 0) {
+				continue;
+			}
+			const contextWindow = getLmStudioNativeContextWindow(entry);
+			metadata.set(entry.id, {
+				input: getLmStudioNativeInput(entry),
+				...(contextWindow === undefined ? {} : { contextWindow }),
+			});
+		}
+		return metadata;
+	} catch {
+		return null;
+	}
+}
+
 export interface LmStudioModelManagerConfig {
 	apiKey?: string;
 	baseUrl?: string;
@@ -2198,8 +2278,11 @@ export function lmStudioModelManagerOptions(
 	const references = createBundledReferenceMap<"openai-completions">("lm-studio" as any);
 	return {
 		providerId: "lm-studio",
-		fetchDynamicModels: () =>
-			fetchOpenAICompatibleModels({
+		fetchDynamicModels: async () => {
+			const nativeMetadataPromise = fetchLmStudioNativeModelMetadata(baseUrl, config?.fetch, {
+				headers: apiKey ? { Authorization: `Bearer ${apiKey}` } : undefined,
+			});
+			const models = await fetchOpenAICompatibleModels({
 				api: "openai-completions",
 				provider: "lm-studio",
 				baseUrl,
@@ -2209,7 +2292,26 @@ export function lmStudioModelManagerOptions(
 					return mapWithBundledReference(entry, defaults, reference);
 				},
 				fetch: config?.fetch,
-			}),
+			});
+			if (!models) {
+				return models;
+			}
+			const nativeMetadata = await nativeMetadataPromise;
+			if (!nativeMetadata) {
+				return models;
+			}
+			return models.map(model => {
+				const metadata = nativeMetadata.get(model.id);
+				if (!metadata) {
+					return model;
+				}
+				return {
+					...model,
+					input: metadata.input,
+					contextWindow: metadata.contextWindow ?? model.contextWindow,
+				};
+			});
+		},
 	};
 }
 

package/src/types.ts

@@ -283,6 +283,8 @@ export interface OpenAICompat {
 	alwaysSendMaxTokens?: boolean;
 	/** Whether Responses-API tool-call/result history must be strictly paired. Default: auto-detected (Azure OpenAI, GitHub Copilot). */
 	strictResponsesPairing?: boolean;
+	/** Whether the Responses API accepts the `detail: "original"` image hint. Default: auto-detected (false for GitHub Copilot, which rejects it with a 400). */
+	supportsImageDetailOriginal?: boolean;
 	/**
 	 * Append a trailing `# Juice: 0 !important` developer item when the caller
 	 * did not request reasoning, suppressing default reasoning on models that
@@ -504,6 +506,7 @@ export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat &
 			| "cacheControlFormat"
 			| "thinkingKeep"
 			| "strictResponsesPairing"
+			| "supportsImageDetailOriginal"
 			| "requiresJuiceZeroHack"
 			| "enableGeminiThinkingLoopGuard"
 			| "whenThinking"
@@ -527,6 +530,7 @@ export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat &
 export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompat {
 	supportsLongPromptCacheRetention: boolean;
 	strictResponsesPairing: boolean;
+	supportsImageDetailOriginal: boolean;
 	requiresJuiceZeroHack: boolean;
 	supportsObfuscationOptOut: boolean;
 }