@oh-my-pi/pi-catalog 16.2.13 → 16.3.0

package/CHANGELOG.md

@@ -2,6 +2,21 @@
 
 ## [Unreleased]
 
+## [16.3.0] - 2026-07-02
+
+### Breaking Changes
+
+- Renamed the `requiresJuiceZeroHack` compatibility flag to `requiresReasoningSuppressionPrompt` (affecting `OpenAICompat` and `ResolvedOpenAIResponsesCompat`) and removed the unused `"juice-zero-developer-message"` member from `OpenAIReasoningDisableMode`.
+
+### Fixed
+
+- Fixed stream markup healing pattern misfires by disabling the healer on the official OpenAI endpoint.
+- Updated the Xiaomi provider's default model to the supported `mimo-v2.5` model.
+- Fixed model discovery probes (including Ollama and metadata fetches) failing behind private-CA gateways by ensuring they honor the `NODE_EXTRA_CA_CERTS` environment variable.
+- Fixed CoreWeave Serverless Inference project-header detection to ensure blank OpenAI-Project overrides do not block the `COREWEAVE_PROJECT` fallback.
+- Fixed LiteLLM MiniMax M3 discovery to remove reseller-only display suffixes and invalidated the model cache to clear stale suffixes immediately.
+- Fixed ZenMux's `anthropic-messages` proxy being misclassified as a non-signing reasoning endpoint (`replayUnsignedThinking: true`), matching the GitHub Copilot fix (#2851). ZenMux's `zenmux.ai/api/anthropic` route 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 `400 messages.1.content.0: Invalid signature in thinking` on Claude Sonnet 5 and other reasoning models. ([#4192](https://github.com/can1357/oh-my-pi/issues/4192))
+
 ## [16.2.13] - 2026-07-01
 
 ### Added
@@ -184,7 +199,7 @@
 
 ### Fixed
 
-- Fixed Claude 4.6 routing on the `google-antigravity` (and `google-gemini-cli`) Cloud Code Assist providers, whose backend exposes the models asymmetrically: `claude-sonnet-4-6` has no `-thinking` twin and `claude-opus-4-6` has only the `-thinking` twin. The shared `thinkingPair` family was routing thinking efforts on `claude-sonnet-4-6` to a non-existent `claude-sonnet-4-6-thinking` wire id (404 `Requested entity was not found`); replaced both 4.6 entries with bespoke single-wire families that declare the dead ids as `retiredMembers` so `reconcileRetiredRouting` re-points stale bundled-catalog and SQLite-cache rows away from the 404 wire id. Refreshed the bundled `models.json` Sonnet 4.6 entry whose stored `effortRouting` still targeted the dead `-thinking` id. Added `claude-sonnet-4-6` and `claude-opus-4-6-thinking` entries to `ANTIGRAVITY_MODEL_WIRE_PROFILES` capped at the backend's 64000-output-token limit (over-cap requests 400'd with `Request contains an invalid argument`); `modelEnum` is now optional on `AntigravityModelWireProfile` since the Claude wire ids are accepted without a captured `labels.model_enum`. ([#3067](https://github.com/can1357/oh-my-pi/issues/3067))
+- Fixed Claude 4.6 routing on the `google-antigravity` (and `google-gemini-cli`) Cloud Code Assist providers, whose backend exposes the models asymmetrically: `claude-sonnet-4-6` has no `-thinking` twin and `claude-opus-4-6` has only the `-thinking` twin. The shared `thinkingPair` family was routing thinking efforts on `claude-sonnet-4-6` to a non-existent `claude-sonnet-4-6-thinking` wire id (404 `Requested entity was not found`); replaced both 4.6 entries with bespoke single-wire families so every effort and off resolve to the live wire id. Added `claude-sonnet-4-6` and `claude-opus-4-6-thinking` entries to `ANTIGRAVITY_MODEL_WIRE_PROFILES` capped at the backend's 64000-output-token limit (over-cap requests 400'd with `Request contains an invalid argument`); `modelEnum` is now optional on `AntigravityModelWireProfile` since the Claude wire ids are accepted without a captured `labels.model_enum`. ([#3067](https://github.com/can1357/oh-my-pi/issues/3067))
 
 ## [16.1.3] - 2026-06-19
 

package/dist/types/compat/openai.d.ts

@@ -20,8 +20,8 @@ interface OpenAIResponsesSpecLike {
  * as well as URL — bundled `azure` models carry no baseUrl (the deployment host
  * is per-resource, resolved at runtime) — while OpenAI/Copilot developer-role
  * and prompt-cache detection stay URL-keyed, as the historical call sites were.
- * The GPT-5 juice-zero hack keys on the model name, matching the historical
- * request-time check.
+ * The GPT-5 reasoning-suppression prompt keys on the model name, matching the
+ * historical request-time check.
  */
 export declare function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): ResolvedOpenAIResponsesCompat;
 export declare function buildOpenRouterCompat(spec: ModelSpec<"openrouter">): ResolvedOpenRouterCompat;

package/dist/types/hosts.d.ts

@@ -115,6 +115,11 @@ export declare const KNOWN_HOSTS: {
         readonly providers: readonly ["opencode-go", "opencode-zen"];
         readonly urlMarkers: readonly ["opencode.ai"];
     };
+    /** ZenMux's Anthropic-compatible proxy (`zenmux.ai/api/anthropic`) forwards to signature-enforcing Anthropic. */
+    readonly zenmux: {
+        readonly providers: readonly ["zenmux"];
+        readonly urlMarkers: readonly ["zenmux.ai"];
+    };
     readonly chutes: {
         readonly urlMarkers: readonly ["chutes.ai"];
     };

package/dist/types/provider-models/descriptors.d.ts

@@ -373,7 +373,7 @@ export declare const CATALOG_PROVIDERS: readonly [{
     };
 }, {
     readonly id: "xiaomi";
-    readonly defaultModel: "mimo-v2-flash";
+    readonly defaultModel: "mimo-v2.5";
     readonly envVars: readonly ["XIAOMI_API_KEY"];
     readonly createModelManagerOptions: (config: ModelManagerConfig) => import("..").ModelManagerOptions<"openai-completions", unknown>;
     readonly catalogDiscovery: {

package/dist/types/types.d.ts

@@ -1,4 +1,5 @@
 import type { Effort } from "./effort";
+export type { FetchImpl } from "@oh-my-pi/pi-utils";
 export type { KnownProvider } from "./provider-models/descriptors";
 export type KnownApi = "openai-completions" | "openai-responses" | "openrouter" | "openai-codex-responses" | "azure-openai-responses" | "anthropic-messages" | "bedrock-converse-stream" | "google-generative-ai" | "google-gemini-cli" | "google-vertex" | "ollama-chat" | "cursor-agent" | "gitlab-duo-agent" | "devin-agent";
 export type Api = KnownApi | (string & {});
@@ -63,14 +64,6 @@ export type Provider = string;
 export type ThinkingBudgets = {
     [key in Effort]?: number;
 };
-/**
- * `fetch`-compatible function. Accepts any callable matching the standard
- * fetch signature; `preconnect` is optional because non-Bun runtimes (browsers,
- * test mocks) won't expose it.
- */
-export type FetchImpl = ((input: string | URL | Request, init?: RequestInit) => Promise<Response>) & {
-    preconnect?: typeof globalThis.fetch.preconnect;
-};
 export interface Usage {
     /** Non-cached input tokens (matches the bucket the provider bills as new input). */
     input: number;
@@ -118,7 +111,7 @@ export interface Usage {
     };
 }
 export type OpenAIReasoningFormat = "openai" | "openrouter" | "zai" | "qwen" | "qwen-chat-template";
-export type OpenAIReasoningDisableMode = "omit" | "lowest-effort" | "openrouter-enabled-false" | "zai-thinking-disabled" | "qwen-enable-thinking-false" | "qwen-template-false" | "juice-zero-developer-message";
+export type OpenAIReasoningDisableMode = "omit" | "lowest-effort" | "openrouter-enabled-false" | "zai-thinking-disabled" | "qwen-enable-thinking-false" | "qwen-template-false";
 export type OpenAIStreamMarkupHealingPattern = "kimi" | "dsml" | "thinking";
 /**
  * Compatibility settings for openai-completions API.
@@ -294,13 +287,13 @@ export interface OpenAICompat {
     /** 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
-     * cannot disable it via request params (Responses APIs only; see
+     * Append a trailing no-reasoning developer item when the caller did not
+     * request reasoning, suppressing default reasoning on models that cannot
+     * disable it via request params (Responses APIs only; see
      * https://community.openai.com/t/need-reasoning-false-option-for-gpt-5/1351588/7).
-     * Default: auto-detected (GPT-5-family model names).
+     * The prompt must not look like an execution or tool budget. Default: auto-detected (GPT-5-family model names).
      */
-    requiresJuiceZeroHack?: boolean;
+    requiresReasoningSuppressionPrompt?: boolean;
     /** Whether streamed reasoning deltas for the same field may repeat the full cumulative text snapshot. Default: false. */
     reasoningDeltasMayBeCumulative?: boolean;
     /** Strip leaked DeepSeek chat-template special tokens from visible content deltas. Default: auto-detected. */
@@ -470,7 +463,7 @@ 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" | "supportsNamedToolChoice" | "reasoningContentField" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "replayReasoningContent" | "qwenPreserveThinking" | "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">> & {
+export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "reasoningEffortMap" | "supportsReasoningParams" | "thinkingFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "supportsToolChoice" | "supportsForcedToolChoice" | "supportsNamedToolChoice" | "reasoningContentField" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "replayReasoningContent" | "qwenPreserveThinking" | "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" | "requiresReasoningSuppressionPrompt" | "enableGeminiThinkingLoopGuard" | "whenThinking">> & {
     vercelGatewayRouting?: OpenAICompat["vercelGatewayRouting"];
     extraBody?: OpenAICompat["extraBody"];
     cacheControlFormat?: OpenAICompat["cacheControlFormat"];
@@ -489,7 +482,7 @@ export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompa
     supportsLongPromptCacheRetention: boolean;
     strictResponsesPairing: boolean;
     supportsImageDetailOriginal: boolean;
-    requiresJuiceZeroHack: boolean;
+    requiresReasoningSuppressionPrompt: boolean;
     supportsObfuscationOptOut: boolean;
     streamIdleTimeoutMs?: number;
 }

package/dist/types/utils.d.ts

@@ -1,4 +1,10 @@
+import { type FetchImpl } from "@oh-my-pi/pi-utils";
 export { isRecord } from "@oh-my-pi/pi-utils";
+/**
+ * Fetch implementation for catalog discovery probes: the caller's override
+ * when given, otherwise global fetch wrapped for `NODE_EXTRA_CA_CERTS`.
+ */
+export declare function discoveryFetch(override?: FetchImpl): FetchImpl;
 export declare function toNumber(value: unknown): number | undefined;
 export declare function toPositiveNumber(value: unknown, fallback: number): number;
 export declare function toPositiveNumber(value: unknown, fallback: number | null): number | null;

package/dist/types/wire/coreweave.d.ts

@@ -9,3 +9,4 @@ export interface CoreWeaveProjectEnv {
 export declare function resolveCoreWeaveProject(env: CoreWeaveProjectEnv): string | undefined;
 export declare function coreWeaveProjectHeaders(env: CoreWeaveProjectEnv): Record<string, string> | undefined;
 export declare function hasCoreWeaveProjectHeader(headers: Record<string, string>): boolean;
+export declare function removeBlankCoreWeaveProjectHeaders(headers: Record<string, string>): void;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "16.2.13",
+	"version": "16.3.0",
 	"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": "16.2.13",
+		"@oh-my-pi/pi-utils": "16.3.0",
 		"arktype": "^2.2.0",
 		"zod": "^4"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "16.2.13",
+		"@oh-my-pi/pi-ai": "16.3.0",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/anthropic.ts

@@ -42,12 +42,16 @@ export function buildAnthropicCompat(spec: ModelSpec<"anthropic-messages">): Res
 	const official = isOfficialAnthropicApiUrl(baseUrl);
 	// Z.AI's Anthropic-compatible proxy lives at `api.z.ai/api/anthropic`.
 	const isZai = modelMatchesHost(spec, "zai");
-	// GitHub Copilot's Anthropic-compatible proxy (api.githubcopilot.com/v1/messages)
+	// GitHub Copilot's `anthropic-messages` proxy (api.githubcopilot.com/v1/messages)
 	// rejects the per-tool `eager_input_streaming` field with
 	// `tools.0.custom.eager_input_streaming: Extra inputs are not permitted` and
 	// doesn't whitelist the `fine-grained-tool-streaming-2025-05-14` beta either
 	// (issue #2558), so eager tool-input streaming is unavailable on this host.
 	const isCopilot = modelMatchesHost(spec, "githubCopilot");
+	// ZenMux's `anthropic-messages` route (zenmux.ai/api/anthropic) forwards to
+	// signature-enforcing Anthropic — same failure class as GitHub Copilot #2851
+	// (issue #4192).
+	const isZenmux = modelMatchesHost(spec, "zenmux");
 	const requiresThinkingEnabled = modelMatchesHost(spec, "moonshotNative") && matchesKimiK27CodeFamily(spec);
 	const compat: ResolvedAnthropicCompat = {
 		officialEndpoint: official,
@@ -77,15 +81,18 @@ export function buildAnthropicCompat(spec: ModelSpec<"anthropic-messages">): Res
 		// arguments (#2005). Known non-signing hosts (Z.AI, DeepSeek) are also
 		// preserved for compatibility.
 		//
-		// 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.
+		// GitHub Copilot's `anthropic-messages` proxy and ZenMux's Anthropic route
+		// are excluded: both forward to signature-enforcing Anthropic and return
+		// full thinking signatures, so they are SIGNING endpoints. 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 (issues
+		// #2851, #4192). Treating them like official Anthropic degrades such
+		// blocks to text instead, which the API accepts.
 		replayUnsignedThinking:
-			!isCopilot && (isZai || modelMatchesHost(spec, "deepseekFamily") || (spec.reasoning && !official)),
+			!isCopilot &&
+			!isZenmux &&
+			(isZai || modelMatchesHost(spec, "deepseekFamily") || (spec.reasoning && !official)),
 		escapeBuiltinToolNames: modelMatchesHost(spec, "umans"),
 	};
 	applyCompatOverrides(compat, spec.compat);

package/src/compat/openai.ts

@@ -117,20 +117,39 @@ function resolveReasoningDisableMode(
 /**
  * Pick the leaked-markup healer for an OpenAI-compatible visible-text stream.
  * Kimi chat-template tokens and DeepSeek DSML envelopes need their dedicated
- * tool-call grammars; every other model defaults to `"thinking"`. All patterns
- * run the generic thinking healer, so leaked reasoning idioms (e.g. a Gemini
- * ` ```thinking ` fence on OpenRouter) are always recovered from `delta.content`.
+ * tool-call grammars. Every other OpenAI-compatible model defaults to
+ * `"thinking"` so leaked reasoning idioms (e.g. a Gemini ` ```thinking ` fence
+ * on OpenRouter) are recovered from `delta.content` — **except** the official
+ * OpenAI endpoint (`provider: "openai"` + `api.openai.com`), which returns
+ * structured reasoning and never leaks, so it heals nothing (returns
+ * `undefined`) to avoid misfiring on legitimate fenced content.
  */
-function detectStreamMarkupHealingPattern(provider: string, modelId: string): OpenAIStreamMarkupHealingPattern {
+function detectStreamMarkupHealingPattern(
+	provider: string,
+	modelId: string,
+	baseUrl: string,
+): OpenAIStreamMarkupHealingPattern | undefined {
 	if (provider === "kimi-code" || provider === "moonshot" || /kimi[-/_.]?k2/i.test(modelId)) {
 		return "kimi";
 	}
 	if (isDeepseekModelIdOrName(modelId) && DSML_HEALING_PROVIDERS.has(provider)) {
 		return "dsml";
 	}
+	if (isOfficialOpenAIEndpoint(provider, baseUrl)) return undefined;
 	return "thinking";
 }
 
+/** Strict official-OpenAI check: provider id `openai` and an `api.openai.com` host (missing baseUrl defaults there). */
+function isOfficialOpenAIEndpoint(provider: string, baseUrl: string): boolean {
+	if (provider !== "openai") return false;
+	if (!baseUrl) return true;
+	try {
+		return new URL(baseUrl).hostname === "api.openai.com";
+	} catch {
+		return false;
+	}
+}
+
 /**
  * OpenCode's gateways (https://opencode.ai/zen|go) gate `reasoning_content`
  * on the request's thinking state for every model they front (Kimi K2.x,
@@ -517,7 +536,7 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
 		streamIdleTimeoutMs,
 		stripDeepseekSpecialTokens:
 			isDeepseekModelIdOrName(spec.id) && (provider === "nvidia" || provider === "deepseek"),
-		streamMarkupHealingPattern: detectStreamMarkupHealingPattern(provider, spec.id),
+		streamMarkupHealingPattern: detectStreamMarkupHealingPattern(provider, spec.id, baseUrl),
 		reasoningDeltasMayBeCumulative:
 			MINIMAX_PROVIDER_OR_ID_PATTERN.test(provider) || MINIMAX_PROVIDER_OR_ID_PATTERN.test(spec.id),
 		emptyLengthFinishIsContextError: provider === "ollama",
@@ -574,8 +593,8 @@ interface OpenAIResponsesSpecLike {
  * as well as URL — bundled `azure` models carry no baseUrl (the deployment host
  * is per-resource, resolved at runtime) — while OpenAI/Copilot developer-role
  * and prompt-cache detection stay URL-keyed, as the historical call sites were.
- * The GPT-5 juice-zero hack keys on the model name, matching the historical
- * request-time check.
+ * The GPT-5 reasoning-suppression prompt keys on the model name, matching the
+ * historical request-time check.
  */
 export function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): ResolvedOpenAIResponsesCompat {
 	const baseUrl = spec.baseUrl ?? "";
@@ -607,7 +626,7 @@ export function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): Resol
 		// 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"),
+		requiresReasoningSuppressionPrompt: spec.name.toLowerCase().startsWith("gpt-5"),
 		reasoningEffortMap: {},
 		supportsReasoningParams: true,
 		thinkingFormat,
@@ -646,7 +665,7 @@ export function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): Resol
 		supportsObfuscationOptOut: isOpenAIUrl || spec.provider === "openai",
 		stripDeepseekSpecialTokens:
 			Boolean(id) && isDeepseekModelIdOrName(id) && (spec.provider === "nvidia" || spec.provider === "deepseek"),
-		streamMarkupHealingPattern: id ? detectStreamMarkupHealingPattern(spec.provider, id) : undefined,
+		streamMarkupHealingPattern: id ? detectStreamMarkupHealingPattern(spec.provider, id, baseUrl) : undefined,
 		reasoningDeltasMayBeCumulative:
 			MINIMAX_PROVIDER_OR_ID_PATTERN.test(spec.provider) || (id ? MINIMAX_PROVIDER_OR_ID_PATTERN.test(id) : false),
 		emptyLengthFinishIsContextError: spec.provider === "ollama",
@@ -674,7 +693,7 @@ function pickResponsesOnly(compat: ResolvedOpenAIResponsesCompat): ResponsesOnly
 		supportsLongPromptCacheRetention: compat.supportsLongPromptCacheRetention,
 		strictResponsesPairing: compat.strictResponsesPairing,
 		supportsImageDetailOriginal: compat.supportsImageDetailOriginal,
-		requiresJuiceZeroHack: compat.requiresJuiceZeroHack,
+		requiresReasoningSuppressionPrompt: compat.requiresReasoningSuppressionPrompt,
 		supportsObfuscationOptOut: compat.supportsObfuscationOptOut,
 	} satisfies ResponsesOnlyCompat;
 }

package/src/discovery/antigravity.ts

@@ -1,6 +1,6 @@
 import { type } from "arktype";
 import type { ModelSpec } from "../types";
-import { toPositiveNumber } from "../utils";
+import { discoveryFetch, toPositiveNumber } from "../utils";
 import {
 	ANTIGRAVITY_VARIANT_COLLAPSE_TABLE,
 	collapseEffortVariants,
@@ -161,7 +161,7 @@ export interface FetchAntigravityDiscoveryModelsOptions {
 export async function fetchAntigravityDiscoveryModels(
 	options: FetchAntigravityDiscoveryModelsOptions,
 ): Promise<ModelSpec<"google-gemini-cli">[] | null> {
-	const fetcher = options.fetcher ?? fetch;
+	const fetcher = discoveryFetch(options.fetcher);
 	const endpoints = options.endpoint
 		? [trimTrailingSlashes(options.endpoint)]
 		: DEFAULT_ANTIGRAVITY_DISCOVERY_ENDPOINTS.map(trimTrailingSlashes);

package/src/discovery/codex.ts

@@ -1,6 +1,7 @@
+import type { FetchImpl } from "@oh-my-pi/pi-utils";
 import { type } from "arktype";
 import type { ModelSpec } from "../types";
-import { isRecord } from "../utils";
+import { discoveryFetch, isRecord } from "../utils";
 import { CODEX_BASE_URL, OPENAI_HEADER_VALUES, OPENAI_HEADERS } from "../wire/codex";
 
 const DEFAULT_MODEL_LIST_PATHS = ["/codex/models", "/models"] as const;
@@ -81,7 +82,7 @@ export interface CodexModelDiscoveryResult {
  * Returns `{ models: [] }` when a route succeeds but yields no usable models.
  */
 export async function fetchCodexModels(options: CodexModelDiscoveryOptions): Promise<CodexModelDiscoveryResult | null> {
-	const fetchFn = options.fetchFn ?? fetch;
+	const fetchFn = discoveryFetch(options.fetchFn);
 	const baseUrl = normalizeBaseUrl(options.baseUrl);
 	const paths = normalizePaths(options.paths);
 	const headers = buildCodexHeaders(options);
@@ -168,7 +169,7 @@ function buildCodexHeaders(options: CodexModelDiscoveryOptions): Headers {
 
 async function resolveCodexClientVersion(
 	clientVersion: string | undefined,
-	fetchFn: typeof fetch,
+	fetchFn: FetchImpl,
 	signal: AbortSignal | undefined,
 ): Promise<string> {
 	const normalizedClientVersion = normalizeClientVersion(clientVersion);

package/src/discovery/devin.ts

@@ -1,6 +1,7 @@
 import { gunzipSync } from "node:zlib";
 import { create, fromBinary, toBinary } from "@bufbuild/protobuf";
 import type { FetchImpl, ModelSpec } from "../types";
+import { discoveryFetch } from "../utils";
 import {
 	GetCliModelConfigsRequestSchema,
 	GetCliModelConfigsResponseSchema,
@@ -76,7 +77,7 @@ export async function fetchDevinModels(
 			accept: "*/*",
 		};
 
-		const fetchImpl = options.fetch ?? fetch;
+		const fetchImpl = discoveryFetch(options.fetch);
 		const response = await fetchImpl(requestUrl, { method: "POST", headers, body, signal });
 		if (!response.ok) {
 			return null;

package/src/discovery/gemini.ts

@@ -2,6 +2,7 @@ import { type } from "arktype";
 import { getBundledModels } from "../models";
 import { toModelSpec } from "../provider-models/bundled-references";
 import type { FetchImpl, Model, ModelSpec } from "../types";
+import { discoveryFetch } from "../utils";
 
 const GOOGLE_GENERATIVE_AI_BASE_URL = "https://generativelanguage.googleapis.com/v1beta";
 const DEFAULT_PAGE_SIZE = 100;
@@ -77,7 +78,7 @@ export async function fetchGeminiModels(
 		return null;
 	}
 
-	const fetchImpl = options.fetch ?? fetch;
+	const fetchImpl = discoveryFetch(options.fetch);
 	const baseUrl = normalizeBaseUrl(options.baseUrl);
 	const pageSize = normalizePositiveInt(options.pageSize, DEFAULT_PAGE_SIZE);
 	const maxPages = normalizePositiveInt(options.maxPages, DEFAULT_MAX_PAGES);

package/src/discovery/gitlab-duo-workflow.ts

@@ -2,7 +2,7 @@ import * as fs from "node:fs/promises";
 import * as path from "node:path";
 import { z } from "zod/v4";
 import type { FetchImpl, ModelSpec } from "../types";
-import { isRecord } from "../utils";
+import { discoveryFetch, isRecord } from "../utils";
 
 const GITLAB_DEFAULT_BASE_URL = "https://gitlab.com";
 const GRAPHQL_PATH = "/api/graphql";
@@ -340,7 +340,7 @@ async function fetchNamespaceOverrideCandidate(
 	if (!restNamespaceId) {
 		return null;
 	}
-	const fetchImpl = config.fetch ?? fetch;
+	const fetchImpl = discoveryFetch(config.fetch);
 	let response: Response;
 	try {
 		response = await fetchImpl(`${baseUrl}${GROUPS_PATH}/${encodeURIComponent(restNamespaceId)}`, {
@@ -406,7 +406,7 @@ async function fetchProjectRootNamespaceViaRest(
 	baseUrl: string,
 	projectIdOrPath: string,
 ): Promise<GitLabDuoWorkflowRestProject | null> {
-	const fetchImpl = config.fetch ?? fetch;
+	const fetchImpl = discoveryFetch(config.fetch);
 	let response: Response;
 	try {
 		response = await fetchImpl(`${baseUrl}${PROJECTS_PATH}/${encodeURIComponent(projectIdOrPath)}`, {
@@ -449,7 +449,7 @@ async function fetchTopLevelGroupNamespaceCandidates(
 	config: GitLabDuoWorkflowDiscoveryConfig,
 	baseUrl: string,
 ): Promise<GitLabDuoWorkflowCandidate[]> {
-	const fetchImpl = config.fetch ?? fetch;
+	const fetchImpl = discoveryFetch(config.fetch);
 	const candidates: (GitLabDuoWorkflowCandidate & { preferred: boolean })[] = [];
 	// GitLab paginates `/groups`; a token can belong to more than one page of top-level
 	// groups, and a usable Duo namespace may live on a later page. Follow the keyset/
@@ -518,7 +518,7 @@ async function postGraphQL(
 	query: string,
 	variables: Record<string, string>,
 ): Promise<unknown | null> {
-	const fetchImpl = config.fetch ?? fetch;
+	const fetchImpl = discoveryFetch(config.fetch);
 	let response: Response;
 	try {
 		response = await fetchImpl(`${baseUrl}${GRAPHQL_PATH}`, {

package/src/discovery/openai-compatible.ts

@@ -1,5 +1,6 @@
 import { type } from "arktype";
 import type { Api, FetchImpl, ModelSpec, Provider } from "../types";
+import { discoveryFetch } from "../utils";
 
 const MODELS_PATH = "/models";
 
@@ -137,7 +138,7 @@ export async function fetchOpenAICompatibleModels<TApi extends Api>(
 		requestHeaders.Authorization = `Bearer ${options.apiKey}`;
 	}
 
-	const fetchImpl = options.fetch ?? globalThis.fetch;
+	const fetchImpl = discoveryFetch(options.fetch);
 	const fetchPayload = async (signal?: AbortSignal): Promise<unknown | null> => {
 		let response: Response;
 		try {

package/src/hosts.ts

@@ -59,6 +59,8 @@ export const KNOWN_HOSTS = {
 	nvidia: { providers: ["nvidia"], urlMarkers: ["integrate.api.nvidia.com"] },
 	moonshotNative: { providers: ["moonshot", "kimi-code"], urlMarkers: ["api.moonshot.ai", "api.kimi.com"] },
 	opencode: { providers: ["opencode-go", "opencode-zen"], urlMarkers: ["opencode.ai"] },
+	/** ZenMux's Anthropic-compatible proxy (`zenmux.ai/api/anthropic`) forwards to signature-enforcing Anthropic. */
+	zenmux: { providers: ["zenmux"], urlMarkers: ["zenmux.ai"] },
 	chutes: { urlMarkers: ["chutes.ai"] },
 } as const satisfies Record<string, HostClassSpec>;
 

package/src/provider-models/descriptors.ts

@@ -415,7 +415,7 @@ export const CATALOG_PROVIDERS = [
 	},
 	{
 		id: "xiaomi",
-		defaultModel: "mimo-v2-flash",
+		defaultModel: "mimo-v2.5",
 		envVars: ["XIAOMI_API_KEY"],
 		createModelManagerOptions: (config: ModelManagerConfig) => xiaomiModelManagerOptions(config),
 		catalogDiscovery: { label: "Xiaomi" },

package/src/provider-models/ollama.ts

@@ -3,6 +3,7 @@ import { Effort } from "../effort";
 import { isGlm52ReasoningEffortModelId } from "../identity/family";
 import type { ModelManagerOptions } from "../model-manager";
 import type { FetchImpl, ThinkingConfig } from "../types";
+import { discoveryFetch } from "../utils";
 import { createBundledReferenceMap, createReferenceResolver } from "./bundled-references";
 
 export interface OllamaCloudModelManagerConfig {
@@ -75,7 +76,7 @@ async function fetchShowMetadata(
 	baseUrl: string,
 	apiKey: string,
 	model: string,
-	fetchImpl: FetchImpl = fetch,
+	fetchImpl: FetchImpl = discoveryFetch(),
 ): Promise<OllamaShowResponse | undefined> {
 	const response = await fetchImpl(`${baseUrl}/api/show`, {
 		method: "POST",
@@ -107,7 +108,7 @@ export function ollamaCloudModelManagerOptions(
 			const response = await fetchWithRetry(`${baseUrl}/api/tags`, {
 				method: "GET",
 				headers: createCloudHeaders(apiKey),
-				fetch: config?.fetch,
+				fetch: discoveryFetch(config?.fetch),
 				defaultDelayMs: OLLAMA_RETRY_DELAYS_MS,
 			});
 			if (!response.ok) {

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

@@ -9,7 +9,7 @@ import { isGlmVisionModelId, isGrokReasoningEffortCapable, isReasoningGlmModelId
 import type { ModelManagerOptions } from "../model-manager";
 import { getBundledModels } from "../models";
 import type { Api, FetchImpl, Model, ModelSpec, OpenAICompat, Provider, ThinkingConfig } from "../types";
-import { isAnthropicOAuthToken, isRecord, toBoolean, toNumber, toPositiveNumber } from "../utils";
+import { discoveryFetch, isAnthropicOAuthToken, isRecord, toBoolean, toNumber, toPositiveNumber } from "../utils";
 import { coreWeaveProjectHeaders } from "../wire/coreweave";
 import {
 	COPILOT_API_HEADERS,
@@ -81,7 +81,7 @@ function toInputCapabilities(value: unknown): ("text" | "image")[] {
 	return supportsImage ? ["text", "image"] : ["text"];
 }
 
-async function fetchModelsDevPayload(fetchImpl: FetchImpl = fetch): Promise<unknown> {
+async function fetchModelsDevPayload(fetchImpl: FetchImpl = discoveryFetch()): Promise<unknown> {
 	const response = await fetchImpl(MODELS_DEV_URL, {
 		method: "GET",
 		headers: { Accept: "application/json" },
@@ -270,7 +270,7 @@ function toOllamaNativeBaseUrl(baseUrl: string): string {
 async function fetchOllamaNativeModels(
 	baseUrl: string,
 	resolveMetadata: (modelId: string) => Promise<OllamaResolvedMetadata>,
-	fetchImpl: FetchImpl = fetch,
+	fetchImpl: FetchImpl = discoveryFetch(),
 ): Promise<ModelSpec<"openai-responses">[] | null> {
 	const nativeBaseUrl = toOllamaNativeBaseUrl(baseUrl);
 	let response: Response;
@@ -376,7 +376,7 @@ function getOllamaThinkingConfig(capabilities: string[] | undefined): ThinkingCo
 async function fetchOllamaShowMetadata(
 	nativeBaseUrl: string,
 	modelId: string,
-	fetchImpl: FetchImpl = fetch,
+	fetchImpl: FetchImpl = discoveryFetch(),
 ): Promise<OllamaShowMetadata | undefined> {
 	try {
 		const response = await fetchImpl(`${nativeBaseUrl}/api/show`, {
@@ -743,7 +743,7 @@ async function fetchUmansModelsInfo(options: {
 	if (options.apiKey) {
 		requestHeaders["x-api-key"] = options.apiKey;
 	}
-	const fetchImpl = options.fetch ?? fetch;
+	const fetchImpl = discoveryFetch(options.fetch);
 	let payload: unknown;
 	try {
 		const response = await fetchImpl(`${discoveryBaseUrl}${UMANS_MODELS_INFO_PATH}`, {
@@ -1541,7 +1541,7 @@ async function fetchFireworksServerlessModels(options: {
 }): Promise<ModelSpec<"openai-completions">[] | null> {
 	const listUrl = toFireworksControlPlaneModelsUrl(options.baseUrl, FIREWORKS_CONTROL_PLANE_ACCOUNT);
 	if (!listUrl) return null;
-	const fetchImpl = options.fetch ?? fetch;
+	const fetchImpl = discoveryFetch(options.fetch);
 	const collected = new Map<string, ModelSpec<"openai-completions">>();
 	let pageToken = "";
 	for (let page = 0; page < FIREWORKS_CONTROL_PLANE_MAX_PAGES; page++) {
@@ -2926,6 +2926,33 @@ function normalizeLiteLLMRuntimeBaseUrl(baseUrl: string): string {
 	return trimmed.endsWith("/") ? trimmed.slice(0, -1) : trimmed;
 }
 
+const LITELLM_RESELLER_USAGE_SUFFIX = /\s+\(\d+(?:\.\d+)?[x×] usage\)$/i;
+
+function stripLiteLLMResellerUsageSuffix(name: string): string {
+	const cleaned = name.replace(LITELLM_RESELLER_USAGE_SUFFIX, "").trim();
+	return cleaned.length > 0 ? cleaned : name;
+}
+
+function toLiteLLMDisplayName(modelName: string | undefined, referenceName: string | undefined, id: string): string {
+	const cleanedModelName = modelName ? stripLiteLLMResellerUsageSuffix(modelName) : undefined;
+	if (cleanedModelName && cleanedModelName !== id) {
+		return cleanedModelName;
+	}
+	return referenceName ? stripLiteLLMResellerUsageSuffix(referenceName) : id;
+}
+
+function mapLiteLLMOpenAICompatibleModel<TApi extends Api>(
+	entry: OpenAICompatibleModelRecord,
+	defaults: ModelSpec<TApi>,
+	reference: ModelSpec<TApi> | undefined,
+): ModelSpec<TApi> {
+	const model = mapWithBundledReference(entry, defaults, reference);
+	return {
+		...model,
+		name: stripLiteLLMResellerUsageSuffix(model.name),
+	};
+}
+
 function toNonEmptyString(value: unknown): string | undefined {
 	if (typeof value !== "string") {
 		return undefined;
@@ -3025,7 +3052,7 @@ function mapLiteLLMRichEntry<TApi extends Api>(
 	};
 	return {
 		id,
-		name: modelName && modelName !== id ? modelName : (reference?.name ?? id),
+		name: toLiteLLMDisplayName(modelName, reference?.name, id),
 		api: options.api,
 		provider: options.provider,
 		baseUrl: runtimeBaseUrl,
@@ -3052,7 +3079,7 @@ async function fetchLiteLLMRichEndpoint<TApi extends Api>(
 	runtimeBaseUrl: string,
 	signal?: AbortSignal,
 ): Promise<ModelSpec<TApi>[] | null> {
-	const fetchImpl = options.fetch ?? globalThis.fetch;
+	const fetchImpl = discoveryFetch(options.fetch);
 	const requestHeaders: Record<string, string> = {
 		Accept: "application/json",
 		...options.headers,
@@ -3126,7 +3153,11 @@ export function litellmModelManagerOptions(
 	const baseUrl = config?.baseUrl ?? Bun.env.LITELLM_BASE_URL ?? "http://localhost:4000/v1";
 	return {
 		providerId: "litellm",
-		cacheProviderId: `litellm:rich-v1:${Bun.hash(baseUrl).toString(36)}`,
+		// rich-v2 invalidates rows cached before reseller usage-suffix stripping
+		// (stale display names like `MiniMax-M3 (3x usage)`); bump the version
+		// whenever the mappers below change, or warm authoritative caches keep
+		// serving pre-change rows for the full TTL.
+		cacheProviderId: `litellm:rich-v2:${Bun.hash(baseUrl).toString(36)}`,
 		// litellm is a local-only proxy and is never bundled in models.json (that
 		// would leak the machine's localhost catalog). Prefer the proxy's richer
 		// management metadata, then fall back to /v1/models and enrich bare ids
@@ -3151,7 +3182,8 @@ export function litellmModelManagerOptions(
 				provider: "litellm",
 				baseUrl,
 				apiKey,
-				mapModel: (entry, defaults) => mapWithBundledReference(entry, defaults, resolveReference(defaults.id)),
+				mapModel: (entry, defaults) =>
+					mapLiteLLMOpenAICompatibleModel(entry, defaults, resolveReference(defaults.id)),
 				fetch: config?.fetch,
 			});
 		},

package/src/types.ts

@@ -1,5 +1,8 @@
 import type { Effort } from "./effort";
 
+// Re-exported from @oh-my-pi/pi-utils so the whole workspace shares one
+// `fetch`-compatible signature (tls-fetch's wrappers produce/accept it).
+export type { FetchImpl } from "@oh-my-pi/pi-utils";
 export type { KnownProvider } from "./provider-models/descriptors";
 
 export type KnownApi =
@@ -89,15 +92,6 @@ export type Provider = string;
 /** Token budgets for each thinking level (token-based providers only) */
 export type ThinkingBudgets = { [key in Effort]?: number };
 
-/**
- * `fetch`-compatible function. Accepts any callable matching the standard
- * fetch signature; `preconnect` is optional because non-Bun runtimes (browsers,
- * test mocks) won't expose it.
- */
-export type FetchImpl = ((input: string | URL | Request, init?: RequestInit) => Promise<Response>) & {
-	preconnect?: typeof globalThis.fetch.preconnect;
-};
-
 export interface Usage {
 	/** Non-cached input tokens (matches the bucket the provider bills as new input). */
 	input: number;
@@ -153,8 +147,7 @@ export type OpenAIReasoningDisableMode =
 	| "openrouter-enabled-false"
 	| "zai-thinking-disabled"
 	| "qwen-enable-thinking-false"
-	| "qwen-template-false"
-	| "juice-zero-developer-message";
+	| "qwen-template-false";
 
 export type OpenAIStreamMarkupHealingPattern = "kimi" | "dsml" | "thinking";
 
@@ -332,13 +325,13 @@ export interface OpenAICompat {
 	/** 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
-	 * cannot disable it via request params (Responses APIs only; see
+	 * Append a trailing no-reasoning developer item when the caller did not
+	 * request reasoning, suppressing default reasoning on models that cannot
+	 * disable it via request params (Responses APIs only; see
 	 * https://community.openai.com/t/need-reasoning-false-option-for-gpt-5/1351588/7).
-	 * Default: auto-detected (GPT-5-family model names).
+	 * The prompt must not look like an execution or tool budget. Default: auto-detected (GPT-5-family model names).
 	 */
-	requiresJuiceZeroHack?: boolean;
+	requiresReasoningSuppressionPrompt?: boolean;
 	/** Whether streamed reasoning deltas for the same field may repeat the full cumulative text snapshot. Default: false. */
 	reasoningDeltasMayBeCumulative?: boolean;
 	/** Strip leaked DeepSeek chat-template special tokens from visible content deltas. Default: auto-detected. */
@@ -564,7 +557,7 @@ export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat &
 			| "thinkingKeep"
 			| "strictResponsesPairing"
 			| "supportsImageDetailOriginal"
-			| "requiresJuiceZeroHack"
+			| "requiresReasoningSuppressionPrompt"
 			| "enableGeminiThinkingLoopGuard"
 			| "whenThinking"
 		>
@@ -588,7 +581,7 @@ export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompa
 	supportsLongPromptCacheRetention: boolean;
 	strictResponsesPairing: boolean;
 	supportsImageDetailOriginal: boolean;
-	requiresJuiceZeroHack: boolean;
+	requiresReasoningSuppressionPrompt: boolean;
 	supportsObfuscationOptOut: boolean;
 	streamIdleTimeoutMs?: number;
 }

package/src/utils.ts

@@ -1,5 +1,15 @@
+import { type FetchImpl, wrapFetchForExtraCa } from "@oh-my-pi/pi-utils";
+
 export { isRecord } from "@oh-my-pi/pi-utils";
 
+/**
+ * Fetch implementation for catalog discovery probes: the caller's override
+ * when given, otherwise global fetch wrapped for `NODE_EXTRA_CA_CERTS`.
+ */
+export function discoveryFetch(override?: FetchImpl): FetchImpl {
+	return override ?? wrapFetchForExtraCa(fetch);
+}
+
 export function toNumber(value: unknown): number | undefined {
 	if (typeof value === "number" && Number.isFinite(value)) {
 		return value;

package/src/wire/coreweave.ts

@@ -38,5 +38,14 @@ export function coreWeaveProjectHeaders(env: CoreWeaveProjectEnv): Record<string
 
 export function hasCoreWeaveProjectHeader(headers: Record<string, string>): boolean {
 	const normalized = COREWEAVE_PROJECT_HEADER.toLowerCase();
-	return Object.keys(headers).some(header => header.toLowerCase() === normalized);
+	return Object.entries(headers).some(([header, value]) => header.toLowerCase() === normalized && value.trim() !== "");
+}
+
+export function removeBlankCoreWeaveProjectHeaders(headers: Record<string, string>): void {
+	const normalized = COREWEAVE_PROJECT_HEADER.toLowerCase();
+	for (const [header, value] of Object.entries(headers)) {
+		if (header.toLowerCase() === normalized && value.trim() === "") {
+			delete headers[header];
+		}
+	}
 }