@oh-my-pi/pi-catalog 16.0.4 → 16.0.5

package/CHANGELOG.md

@@ -2,6 +2,30 @@
 
 ## [Unreleased]
 
+## [16.0.5] - 2026-06-17
+
+### Added
+
+- Added `enableGeminiThinkingLoopGuard` to OpenAI compatibility options to allow explicit opt-in or opt-out of the Gemini thinking-loop guard for OpenAI-compatible model aliases
+- Added `LITELLM_BASE_URL` as the LiteLLM provider discovery base URL fallback, with discovery caches scoped by the resolved proxy URL and explicit provider `baseUrl` config kept at higher precedence. ([#2726](https://github.com/can1357/oh-my-pi/issues/2726))
+- Added `ThinkingConfig.effortBudgets` (per-effort thinking-budget contract baked into collapsed variants) and `ANTIGRAVITY_MODEL_WIRE_PROFILES` (`maxOutputTokens` + `model_enum` per Antigravity wire id) to mirror the captured Antigravity Cloud Code Assist client request shape.
+
+### Changed
+
+- Defaulted `enableGeminiThinkingLoopGuard` from Gemini family detection for both OpenAI completions and responses compatibility specs so Gemini models now enable the thinking-loop guard automatically
+- Updated the default Gemini CLI user-agent version fallback to 0.46.0.
+- Changed the Antigravity (`google-antigravity`, daily-cloudcode-pa) gemini-3.x collapse families to the `budget` thinking transport with the client's per-tier `thinkingBudget` (3.5 Flash low/medium/high = 1000/4000/10000, 3.1 Pro low/high = 1001/10001) and corrected 3.5 Flash effort→wire routing (medium → `gemini-3.5-flash-low`, high → `gemini-3-flash-agent`). Split the shared CCA collapse table so `google-gemini-cli` (cloudcode-pa) keeps the `google-level` `thinkingLevel` transport for official Gemini CLI parity. Stale collapsed snapshots (bundled catalog, recycled `gemini-3-flash` alias) self-heal from the hand table at collapse time, and the model cache schema is bumped to v7 to invalidate pre-budget Antigravity rows.
+- Changed the Antigravity user-agent to the `antigravity/hub/<version>` format (default `2.1.4`) to match the captured client.
+
+### Fixed
+
+- Fixed `off` effort routing for `claude-opus-4-5` and `claude-opus-4-6` to use their base model IDs when thinking is disabled
+- Fixed `gemini-2.5-flash` effort routing so all non-off effort levels resolve to `gemini-2.5-flash-thinking`
+- Fixed shared variant alias provider resolution so `resolveBareVariantAlias` reports all matching providers when model aliases are present in both CCA collapse tables
+- Routed google-antigravity default baseUrl to the stable primary daily endpoint in the catalog generator and all fallback snapshots, resolving connection drops on heavy queries.
+- Fixed MiniMax M3 dialect selection so MiniMax-family OpenAI-compatible models use the MiniMax tool-call dialect instead of generic XML. ([#2759](https://github.com/can1357/oh-my-pi/issues/2759))
+- Fixed GitHub Copilot dynamic discovery to honor plan-specific API endpoints stored in structured OAuth credentials. ([#2876](https://github.com/can1357/oh-my-pi/issues/2876))
+
 ## [16.0.4] - 2026-06-17
 
 ### Fixed

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

@@ -5,6 +5,7 @@ import type { ModelSpec, OpenAICompat, ResolvedOpenAICompat, ResolvedOpenAIRespo
  */
 export declare function buildOpenAICompat(spec: ModelSpec<"openai-completions">): ResolvedOpenAICompat;
 interface OpenAIResponsesSpecLike {
+    id?: string;
     provider: string;
     name: string;
     baseUrl: string;

package/dist/types/discovery/antigravity.d.ts

@@ -1,4 +1,7 @@
 import type { ModelSpec } from "../types";
+import { type VariantCollapseTable } from "../variant-collapse";
+export declare const ANTIGRAVITY_PRIMARY_ENDPOINT = "https://daily-cloudcode-pa.googleapis.com";
+export declare const ANTIGRAVITY_SANDBOX_ENDPOINT = "https://daily-cloudcode-pa.sandbox.googleapis.com";
 /**
  * Raw model metadata returned by Antigravity's `fetchAvailableModels` endpoint.
  */
@@ -51,6 +54,12 @@ export interface FetchAntigravityDiscoveryModelsOptions {
     signal?: AbortSignal;
     /** Optional fetch implementation override for tests. */
     fetcher?: typeof fetch;
+    /**
+     * Hand collapse table to apply to the discovered list. Defaults to the
+     * Antigravity (budget-transport) table; `googleGeminiCli` passes the
+     * level-transport table so cloudcode-pa keeps `thinkingLevel`.
+     */
+    collapseTable?: VariantCollapseTable;
 }
 /**
  * Fetches discoverable Antigravity models and normalizes them into canonical model entries.

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

@@ -1,3 +1,3 @@
-export type Dialect = "glm" | "hermes" | "kimi" | "xml" | "anthropic" | "deepseek" | "harmony" | "pi" | "qwen3" | "gemini" | "gemma";
+export type Dialect = "glm" | "hermes" | "kimi" | "xml" | "anthropic" | "deepseek" | "harmony" | "pi" | "qwen3" | "gemini" | "gemma" | "minimax";
 export declare const FALLBACK_DIALECT: Dialect;
 export declare function preferredDialect(modelId: string): Dialect;

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

@@ -32,6 +32,8 @@ export declare function isMimoModelIdOrName(value: string): boolean;
  * clamp instead. Excludes M1, M3, MiniMax-Text-01, music, hailuo, voice ids.
  */
 export declare function isMinimaxM2FamilyModelId(modelId: string): boolean;
+/** MiniMax M3 family ids in bundled/default and aggregator namespace forms. */
+export declare function 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

package/dist/types/types.d.ts

@@ -33,6 +33,14 @@ export interface ThinkingConfig {
      * thinking is disabled. Missing keys fall back to `requestModelId ?? id`.
      */
     effortRouting?: Readonly<Partial<Record<Effort | "off", string>>>;
+    /**
+     * Per-effort thinking budget in tokens, baked at build time for collapsed
+     * variants whose upstream expects an explicit `thinkingBudget` instead of a
+     * value derived from the generic ladder (Antigravity Cloud Code Assist
+     * gemini-3.x). Request mapping prefers caller `thinkingBudgets`, then this
+     * map, then the provider default ladder. Only meaningful for `mode: "budget"`.
+     */
+    effortBudgets?: Readonly<Partial<Record<Effort, number>>>;
     /**
      * When true, a thinking-off request MUST explicitly suppress thinking on
      * the wire (google-level: `thinkingLevel: "MINIMAL"` + `includeThoughts:
@@ -137,6 +145,13 @@ export interface OpenAICompat {
     reasoningEffortMap?: Partial<Record<Effort, string>>;
     /** Whether the provider supports `stream_options: { include_usage: true }` for token usage in streaming responses. Default: true. */
     supportsUsageInStreaming?: boolean;
+    /**
+     * Enable the Gemini thinking-loop guard (pi-ai stream layer) for this model.
+     * Defaults to true when the model id classifies as the gemini family. Set
+     * explicitly to cover an opaque OpenAI-compat proxy alias (e.g. `my-model`)
+     * that routes to Gemini, or to false to opt a gemini-family id out.
+     */
+    enableGeminiThinkingLoopGuard?: boolean;
     /** Which field to use for max tokens. Default: auto-detected from URL. */
     maxTokensField?: "max_completion_tokens" | "max_tokens";
     /** Whether tool results require the `name` field. Default: auto-detected from URL. */
@@ -322,7 +337,7 @@ type ResolvedToolStrictMode = NonNullable<OpenAICompat["toolStrictMode"]> | "mix
  * `buildModel`; request handlers read fields and never detect, resolve, or
  * allocate.
  */
-export type ResolvedOpenAICompat = Required<Omit<OpenAICompat, "openRouterRouting" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "streamIdleTimeoutMs" | "supportsLongPromptCacheRetention" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "requiresJuiceZeroHack" | "whenThinking">> & {
+export type ResolvedOpenAICompat = Required<Omit<OpenAICompat, "openRouterRouting" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "streamIdleTimeoutMs" | "supportsLongPromptCacheRetention" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "requiresJuiceZeroHack" | "enableGeminiThinkingLoopGuard" | "whenThinking">> & {
     openRouterRouting?: OpenAICompat["openRouterRouting"];
     vercelGatewayRouting?: OpenAICompat["vercelGatewayRouting"];
     extraBody?: OpenAICompat["extraBody"];
@@ -334,6 +349,8 @@ export type ResolvedOpenAICompat = Required<Omit<OpenAICompat, "openRouterRoutin
     isOpenRouterHost: boolean;
     /** The model sits behind Vercel AI Gateway. */
     isVercelGatewayHost: boolean;
+    /** See {@link OpenAICompat.enableGeminiThinkingLoopGuard}. Set by the builder from the family classifier. */
+    enableGeminiThinkingLoopGuard?: boolean;
     /** Complete alternate view for thinking-engaged requests; swap pointers, never spread. */
     whenThinking?: ResolvedOpenAICompat;
 };
@@ -346,6 +363,8 @@ export interface ResolvedOpenAIResponsesCompat {
     strictResponsesPairing: boolean;
     requiresJuiceZeroHack: boolean;
     reasoningEffortMap: Partial<Record<Effort, string>>;
+    /** See {@link OpenAICompat.enableGeminiThinkingLoopGuard}. */
+    enableGeminiThinkingLoopGuard?: boolean;
 }
 /** Fully-resolved anthropic-messages compat view (same contract as `ResolvedOpenAICompat`). */
 export type ResolvedAnthropicCompat = Required<AnthropicCompat> & {

package/dist/types/variant-collapse.d.ts

@@ -46,12 +46,11 @@ export interface EffortVariantFamily {
 export interface VariantCollapseTable {
     families: readonly EffortVariantFamily[];
 }
-/**
- * Shared by `google-antigravity` and `google-gemini-cli` — both serve the
- * Antigravity discovery list (`fetchAntigravityDiscoveryModels`).
- */
+/** `google-antigravity` (daily-cloudcode-pa): Gemini 3.x on the budget transport. */
 export declare const ANTIGRAVITY_VARIANT_COLLAPSE_TABLE: VariantCollapseTable;
-/** Provider id → hand collapse table. Both CCA providers share one table. */
+/** `google-gemini-cli` (cloudcode-pa): Gemini 3.x on the level transport (official CLI parity). */
+export declare const GEMINI_CLI_VARIANT_COLLAPSE_TABLE: VariantCollapseTable;
+/** Provider id → hand collapse table. The CCA providers diverge on thinking transport. */
 export declare const VARIANT_COLLAPSE_TABLES: Readonly<Record<string, VariantCollapseTable>>;
 /**
  * The global automatic rule: derive an `X` + `X-thinking` family for every

package/dist/types/wire/gemini-headers.d.ts

@@ -9,7 +9,6 @@ export declare const getGeminiCliHeaders: (modelId?: string) => {
     "Client-Metadata": string;
 };
 export declare const ANTIGRAVITY_SYSTEM_INSTRUCTION: string;
-export declare const ANTIGRAVITY_NO_PREAMBLE_INSTRUCTION = "CRITICAL: NEVER output rule checks, formatting guidelines, constraint checklists (e.g. \"No emdashes\"), or your thinking/personality preambles in the final response. Output only the final response.";
 /**
  * Antigravity / Cloud Code Assist user agent. Lives in its own file so discovery
  * and usage code can read it without pulling the heavy google-gemini-cli provider
@@ -17,3 +16,19 @@ export declare const ANTIGRAVITY_NO_PREAMBLE_INSTRUCTION = "CRITICAL: NEVER outp
  * parse graph.
  */
 export declare let getAntigravityUserAgent: () => string;
+/**
+ * Per-wire-id Antigravity Cloud Code Assist request constants, captured from the
+ * real `antigravity/hub` client against `daily-cloudcode-pa`. `modelEnum` is the
+ * opaque `labels.model_enum` token the client tags each request with;
+ * `maxOutputTokens` is the fixed `generationConfig.maxOutputTokens` it sends
+ * regardless of the thinking budget. Keyed by the routed upstream wire id
+ * (post effort-routing), not the collapsed logical id. Checkpoint-only ids
+ * (e.g. `gemini-3.1-flash-lite`) are intentionally absent — this provider only
+ * emits agent requests.
+ */
+export interface AntigravityModelWireProfile {
+    modelEnum: string;
+    maxOutputTokens: number;
+}
+export declare const ANTIGRAVITY_MODEL_WIRE_PROFILES: Readonly<Record<string, AntigravityModelWireProfile>>;
+export declare function getAntigravityModelWireProfile(wireModelId: string): AntigravityModelWireProfile | undefined;

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

@@ -25,9 +25,11 @@ export declare const COPILOT_API_HEADERS: {
 export type ParsedGitHubCopilotApiKey = {
     accessToken: string;
     enterpriseUrl?: string;
+    apiEndpoint?: string;
 };
 export declare function isPublicGitHubHost(host: string): boolean;
 export declare function normalizeGitHubCopilotEnterpriseDomain(input: string | undefined): string | undefined;
+export declare function normalizeGitHubCopilotApiEndpoint(input: string | undefined): string | undefined;
 export declare function parseGitHubCopilotApiKey(apiKeyRaw: string): ParsedGitHubCopilotApiKey;
 export declare function normalizeDomain(input: string): string | null;
 export declare function getGitHubCopilotBaseUrl(enterpriseDomain?: string): string;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "16.0.4",
+	"version": "16.0.5",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -34,11 +34,11 @@
 	},
 	"dependencies": {
 		"@bufbuild/protobuf": "^2.12.0",
-		"@oh-my-pi/pi-utils": "16.0.4",
+		"@oh-my-pi/pi-utils": "16.0.5",
 		"zod": "^4"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "16.0.4",
+		"@oh-my-pi/pi-ai": "16.0.5",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/openai.ts

@@ -17,6 +17,7 @@ import {
 	isKimiModelId,
 	isMimoModelIdOrName,
 	isQwenModelId,
+	modelFamilyToken,
 } from "../identity/family";
 import type { ModelSpec, OpenAICompat, ResolvedOpenAICompat, ResolvedOpenAIResponsesCompat } from "../types";
 import { applyCompatOverrides } from "./apply";
@@ -211,6 +212,10 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
 		supportsReasoningParams: provider !== "github-copilot",
 		reasoningEffortMap: {},
 		supportsUsageInStreaming: !isCerebras,
+		// pi-ai's thinking-loop guard is gemini-only; default the flag from the
+		// family classifier so OpenAI-compat proxies serving Gemini are covered.
+		// An opaque alias can opt in via `compat.enableGeminiThinkingLoopGuard`.
+		enableGeminiThinkingLoopGuard: modelFamilyToken(spec.id) === "gemini",
 		// Kimi (including via OpenRouter and Fireworks router-form IDs such as
 		// `accounts/fireworks/routers/kimi-*`) calculates TPM rate limits based on
 		// max_tokens, not actual output. The official Kimi K2 model guidance
@@ -295,6 +300,7 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
 }
 
 interface OpenAIResponsesSpecLike {
+	id?: string;
 	provider: string;
 	name: string;
 	baseUrl: string;
@@ -329,6 +335,7 @@ export function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): Resol
 		strictResponsesPairing: isAzure || spec.provider === "github-copilot",
 		requiresJuiceZeroHack: spec.name.toLowerCase().startsWith("gpt-5"),
 		reasoningEffortMap: {},
+		enableGeminiThinkingLoopGuard: modelFamilyToken(spec.id ?? "") === "gemini",
 	};
 	applyCompatOverrides(compat, spec.compat);
 	return compat;

package/src/discovery/antigravity.ts

@@ -1,13 +1,16 @@
 import { z } from "zod/v4";
 import type { ModelSpec } from "../types";
 import { toPositiveNumber } from "../utils";
-import { ANTIGRAVITY_VARIANT_COLLAPSE_TABLE, collapseEffortVariants } from "../variant-collapse";
+import {
+	ANTIGRAVITY_VARIANT_COLLAPSE_TABLE,
+	collapseEffortVariants,
+	type VariantCollapseTable,
+} from "../variant-collapse";
 import { getAntigravityUserAgent } from "../wire/gemini-headers";
 
-const DEFAULT_ANTIGRAVITY_DISCOVERY_ENDPOINTS = [
-	"https://daily-cloudcode-pa.googleapis.com",
-	"https://daily-cloudcode-pa.sandbox.googleapis.com",
-] as const;
+export const ANTIGRAVITY_PRIMARY_ENDPOINT = "https://daily-cloudcode-pa.googleapis.com";
+export const ANTIGRAVITY_SANDBOX_ENDPOINT = "https://daily-cloudcode-pa.sandbox.googleapis.com";
+const DEFAULT_ANTIGRAVITY_DISCOVERY_ENDPOINTS = [ANTIGRAVITY_PRIMARY_ENDPOINT, ANTIGRAVITY_SANDBOX_ENDPOINT] as const;
 const FETCH_AVAILABLE_MODELS_PATH = "/v1internal:fetchAvailableModels";
 
 const DEFAULT_CONTEXT_WINDOW = 200_000;
@@ -157,6 +160,12 @@ export interface FetchAntigravityDiscoveryModelsOptions {
 	signal?: AbortSignal;
 	/** Optional fetch implementation override for tests. */
 	fetcher?: typeof fetch;
+	/**
+	 * Hand collapse table to apply to the discovered list. Defaults to the
+	 * Antigravity (budget-transport) table; `googleGeminiCli` passes the
+	 * level-transport table so cloudcode-pa keeps `thinkingLevel`.
+	 */
+	collapseTable?: VariantCollapseTable;
 }
 
 /**
@@ -239,7 +248,7 @@ export async function fetchAntigravityDiscoveryModels(
 		// Collapse effort-tier variants at the source so runtime discovery,
 		// the gemini-cli re-provision, and the catalog generator all see
 		// logical ids only.
-		const collapsed = collapseEffortVariants(models, ANTIGRAVITY_VARIANT_COLLAPSE_TABLE);
+		const collapsed = collapseEffortVariants(models, options.collapseTable ?? ANTIGRAVITY_VARIANT_COLLAPSE_TABLE);
 		collapsed.sort((a, b) => a.name.localeCompare(b.name) || a.id.localeCompare(b.id));
 		return collapsed;
 	}

package/src/identity/dialect.ts

@@ -11,7 +11,8 @@ export type Dialect =
 	| "pi"
 	| "qwen3"
 	| "gemini"
-	| "gemma";
+	| "gemma"
+	| "minimax";
 
 export const FALLBACK_DIALECT: Dialect = "xml";
 
@@ -31,6 +32,8 @@ export function preferredDialect(modelId: string): Dialect {
 			return "qwen3";
 		case "deepseek":
 			return "deepseek";
+		case "minimax":
+			return "minimax";
 		case "openai":
 		case "gpt-oss":
 			return "harmony";

package/src/identity/family.ts

@@ -73,6 +73,13 @@ export function isMinimaxM2FamilyModelId(modelId: string): boolean {
 	return /(?:^|[/.-])m2\d*(?:[.-]\d+)?(?:[-.:_]|$)/i.test(lower);
 }
 
+/** MiniMax M3 family ids in bundled/default and aggregator namespace forms. */
+export function isMinimaxM3FamilyModelId(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`,
  * `vendor/gpt-oss-…`). The Harmony reasoning format only accepts
@@ -139,7 +146,7 @@ export function modelFamilyToken(modelId: string): string {
 	if (isOpenAIModelId(modelId)) return "openai";
 	if (isKimiModelId(modelId)) return "kimi";
 	if (isQwenModelId(modelId)) return "qwen";
-	if (isMinimaxM2FamilyModelId(modelId)) return "minimax";
+	if (isMinimaxM2FamilyModelId(modelId) || isMinimaxM3FamilyModelId(modelId)) return "minimax";
 	if (isOpenAIGptOssModelId(modelId)) return "gpt-oss";
 	if (isDeepseekModelIdOrName(modelId)) return "deepseek";
 	if (isMimoModelIdOrName(modelId)) return "mimo";

package/src/model-cache.ts

@@ -7,12 +7,14 @@ import { getModelDbPath } from "@oh-my-pi/pi-utils";
 import type { Api, Model, ModelSpec } from "./types";
 
 // Rows persist ModelSpec JSON (sparse `compat`, never the resolved record);
-// the model manager rebuilds via `buildModel` on load. v6 invalidates rows
-// that may contain the retired unknown-limit sentinels (222222/8888); v5
-// invalidated rows predating effort-tier variant collapsing (raw
-// `-low`/`-high`/`-thinking` member ids); v4 dropped the pre-efforts
-// ThinkingConfig shape.
-const CACHE_SCHEMA_VERSION = 6;
+// the model manager rebuilds via `buildModel` on load. v7 invalidates rows
+// predating the Antigravity Gemini budget-mode migration (cached specs still
+// carrying `thinking.mode: "google-level"` and the old 3.5-flash effort
+// routing); v6 invalidates rows that may contain the retired unknown-limit
+// sentinels (222222/8888); v5 invalidated rows predating effort-tier variant
+// collapsing (raw `-low`/`-high`/`-thinking` member ids); v4 dropped the
+// pre-efforts ThinkingConfig shape.
+const CACHE_SCHEMA_VERSION = 7;
 
 interface CacheRow {
 	provider_id: string;

package/src/models.json

@@ -17726,7 +17726,7 @@
 			"name": "Claude Opus 4.5",
 			"api": "google-gemini-cli",
 			"provider": "google-antigravity",
-			"baseUrl": "https://daily-cloudcode-pa.sandbox.googleapis.com",
+			"baseUrl": "https://daily-cloudcode-pa.googleapis.com",
 			"reasoning": true,
 			"input": [
 				"text",
@@ -17749,6 +17749,7 @@
 					"high"
 				],
 				"effortRouting": {
+					"off": "claude-opus-4-5",
 					"minimal": "claude-opus-4-5-thinking",
 					"low": "claude-opus-4-5-thinking",
 					"medium": "claude-opus-4-5-thinking",
@@ -17762,7 +17763,7 @@
 			"name": "Claude Opus 4.6",
 			"api": "google-gemini-cli",
 			"provider": "google-antigravity",
-			"baseUrl": "https://daily-cloudcode-pa.sandbox.googleapis.com",
+			"baseUrl": "https://daily-cloudcode-pa.googleapis.com",
 			"reasoning": true,
 			"input": [
 				"text",
@@ -17785,6 +17786,7 @@
 					"high"
 				],
 				"effortRouting": {
+					"off": "claude-opus-4-6",
 					"minimal": "claude-opus-4-6-thinking",
 					"low": "claude-opus-4-6-thinking",
 					"medium": "claude-opus-4-6-thinking",
@@ -17798,7 +17800,7 @@
 			"name": "Claude Sonnet 4.5",
 			"api": "google-gemini-cli",
 			"provider": "google-antigravity",
-			"baseUrl": "https://daily-cloudcode-pa.sandbox.googleapis.com",
+			"baseUrl": "https://daily-cloudcode-pa.googleapis.com",
 			"reasoning": true,
 			"input": [
 				"text",
@@ -17835,7 +17837,7 @@
 			"name": "Claude Sonnet 4.6",
 			"api": "google-gemini-cli",
 			"provider": "google-antigravity",
-			"baseUrl": "https://daily-cloudcode-pa.sandbox.googleapis.com",
+			"baseUrl": "https://daily-cloudcode-pa.googleapis.com",
 			"reasoning": true,
 			"input": [
 				"text",
@@ -17872,7 +17874,7 @@
 			"name": "Gemini 2.5 Flash",
 			"api": "google-gemini-cli",
 			"provider": "google-antigravity",
-			"baseUrl": "https://daily-cloudcode-pa.sandbox.googleapis.com",
+			"baseUrl": "https://daily-cloudcode-pa.googleapis.com",
 			"reasoning": true,
 			"input": [
 				"text",
@@ -17908,7 +17910,7 @@
 			"name": "Gemini 2.5 Pro",
 			"api": "google-gemini-cli",
 			"provider": "google-antigravity",
-			"baseUrl": "https://daily-cloudcode-pa.sandbox.googleapis.com",
+			"baseUrl": "https://daily-cloudcode-pa.googleapis.com",
 			"reasoning": true,
 			"input": [
 				"text",
@@ -17938,7 +17940,7 @@
 			"name": "Gemini 3 Flash",
 			"api": "google-gemini-cli",
 			"provider": "google-antigravity",
-			"baseUrl": "https://daily-cloudcode-pa.sandbox.googleapis.com",
+			"baseUrl": "https://daily-cloudcode-pa.googleapis.com",
 			"reasoning": true,
 			"input": [
 				"text",
@@ -17953,22 +17955,36 @@
 			"contextWindow": 1048576,
 			"maxTokens": 65536,
 			"thinking": {
-				"mode": "google-level",
+				"mode": "budget",
 				"efforts": [
 					"minimal",
 					"low",
 					"medium",
 					"high"
 				],
-				"requiresEffort": true
-			}
+				"effortBudgets": {
+					"minimal": 1000,
+					"low": 1000,
+					"medium": 4000,
+					"high": 10000
+				},
+				"effortRouting": {
+					"off": "gemini-3.5-flash-extra-low",
+					"minimal": "gemini-3.5-flash-extra-low",
+					"low": "gemini-3.5-flash-extra-low",
+					"medium": "gemini-3.5-flash-low",
+					"high": "gemini-3-flash-agent"
+				},
+				"suppressWhenOff": true
+			},
+			"requestModelId": "gemini-3.5-flash-extra-low"
 		},
 		"gemini-3-pro": {
 			"id": "gemini-3-pro",
 			"name": "Gemini 3 Pro",
 			"api": "google-gemini-cli",
 			"provider": "google-antigravity",
-			"baseUrl": "https://daily-cloudcode-pa.sandbox.googleapis.com",
+			"baseUrl": "https://daily-cloudcode-pa.googleapis.com",
 			"reasoning": true,
 			"input": [
 				"text",
@@ -18002,7 +18018,7 @@
 			"name": "Gemini 3.1 Pro Preview",
 			"api": "google-gemini-cli",
 			"provider": "google-antigravity",
-			"baseUrl": "https://daily-cloudcode-pa.sandbox.googleapis.com",
+			"baseUrl": "https://daily-cloudcode-pa.googleapis.com",
 			"reasoning": true,
 			"input": [
 				"text",
@@ -18017,11 +18033,15 @@
 			"contextWindow": 1048576,
 			"maxTokens": 65535,
 			"thinking": {
-				"mode": "google-level",
+				"mode": "budget",
 				"efforts": [
 					"low",
 					"high"
 				],
+				"effortBudgets": {
+					"low": 1001,
+					"high": 10001
+				},
 				"effortRouting": {
 					"off": "gemini-3.1-pro-low",
 					"low": "gemini-3.1-pro-low",
@@ -18036,7 +18056,7 @@
 			"name": "GPT OSS 120B",
 			"api": "google-gemini-cli",
 			"provider": "google-antigravity",
-			"baseUrl": "https://daily-cloudcode-pa.sandbox.googleapis.com",
+			"baseUrl": "https://daily-cloudcode-pa.googleapis.com",
 			"reasoning": true,
 			"input": [
 				"text"
@@ -18110,7 +18130,11 @@
 					"high"
 				],
 				"effortRouting": {
-					"off": "gemini-2.5-flash"
+					"off": "gemini-2.5-flash",
+					"minimal": "gemini-2.5-flash-thinking",
+					"low": "gemini-2.5-flash-thinking",
+					"medium": "gemini-2.5-flash-thinking",
+					"high": "gemini-2.5-flash-thinking"
 				}
 			}
 		},
@@ -69307,7 +69331,7 @@
 				"cacheRead": 0,
 				"cacheWrite": 0
 			},
-			"contextWindow": null,
+			"contextWindow": 1000000,
 			"maxTokens": null,
 			"compat": {
 				"supportsUsageInStreaming": false

package/src/provider-models/google.ts

@@ -2,6 +2,7 @@ import { fetchAntigravityDiscoveryModels } from "../discovery/antigravity";
 import { fetchGeminiModels } from "../discovery/gemini";
 import type { ModelManagerOptions } from "../model-manager";
 import type { FetchImpl } from "../types";
+import { GEMINI_CLI_VARIANT_COLLAPSE_TABLE } from "../variant-collapse";
 
 export interface GoogleModelManagerConfig {
 	apiKey?: string;
@@ -89,6 +90,7 @@ export function googleGeminiCliModelManagerOptions(
 							token,
 							endpoint,
 							fetcher: toDiscoveryFetch(config?.fetch),
+							collapseTable: GEMINI_CLI_VARIANT_COLLAPSE_TABLE,
 						});
 						if (models === null) {
 							return null;

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

@@ -2522,9 +2522,10 @@ export function litellmModelManagerOptions(
 	config?: LiteLLMModelManagerConfig,
 ): ModelManagerOptions<"openai-completions"> {
 	const apiKey = config?.apiKey;
-	const baseUrl = config?.baseUrl ?? "http://localhost:4000/v1";
+	const baseUrl = config?.baseUrl ?? Bun.env.LITELLM_BASE_URL ?? "http://localhost:4000/v1";
 	return {
 		providerId: "litellm",
+		cacheProviderId: `litellm:${Bun.hash(baseUrl).toString(36)}`,
 		// litellm is a local-only proxy whose /v1/models returns bare ids with no
 		// metadata, and it is never bundled in models.json (that would leak the
 		// machine's localhost catalog). It proxies known upstream models, so we
@@ -2807,9 +2808,11 @@ export function githubCopilotModelManagerOptions(config?: GithubCopilotModelMana
 	const parsedApiKey = rawApiKey ? parseGitHubCopilotApiKey(rawApiKey) : undefined;
 	const apiKey = parsedApiKey?.accessToken;
 	const baseUrl =
-		parsedApiKey?.enterpriseUrl && configuredBaseUrl.includes("githubcopilot.com")
-			? getGitHubCopilotBaseUrl(parsedApiKey.enterpriseUrl)
-			: configuredBaseUrl;
+		parsedApiKey?.apiEndpoint && configuredBaseUrl.includes("githubcopilot.com")
+			? parsedApiKey.apiEndpoint
+			: parsedApiKey?.enterpriseUrl && configuredBaseUrl.includes("githubcopilot.com")
+				? getGitHubCopilotBaseUrl(parsedApiKey.enterpriseUrl)
+				: configuredBaseUrl;
 	const providerRefs = createBundledReferenceMap<Api>("github-copilot");
 	const resolveReference = createReferenceResolver(providerRefs);
 	return {

package/src/types.ts

@@ -53,6 +53,14 @@ export interface ThinkingConfig {
 	 * thinking is disabled. Missing keys fall back to `requestModelId ?? id`.
 	 */
 	effortRouting?: Readonly<Partial<Record<Effort | "off", string>>>;
+	/**
+	 * Per-effort thinking budget in tokens, baked at build time for collapsed
+	 * variants whose upstream expects an explicit `thinkingBudget` instead of a
+	 * value derived from the generic ladder (Antigravity Cloud Code Assist
+	 * gemini-3.x). Request mapping prefers caller `thinkingBudgets`, then this
+	 * map, then the provider default ladder. Only meaningful for `mode: "budget"`.
+	 */
+	effortBudgets?: Readonly<Partial<Record<Effort, number>>>;
 	/**
 	 * When true, a thinking-off request MUST explicitly suppress thinking on
 	 * the wire (google-level: `thinkingLevel: "MINIMAL"` + `includeThoughts:
@@ -162,6 +170,13 @@ export interface OpenAICompat {
 	reasoningEffortMap?: Partial<Record<Effort, string>>;
 	/** Whether the provider supports `stream_options: { include_usage: true }` for token usage in streaming responses. Default: true. */
 	supportsUsageInStreaming?: boolean;
+	/**
+	 * Enable the Gemini thinking-loop guard (pi-ai stream layer) for this model.
+	 * Defaults to true when the model id classifies as the gemini family. Set
+	 * explicitly to cover an opaque OpenAI-compat proxy alias (e.g. `my-model`)
+	 * that routes to Gemini, or to false to opt a gemini-family id out.
+	 */
+	enableGeminiThinkingLoopGuard?: boolean;
 	/** Which field to use for max tokens. Default: auto-detected from URL. */
 	maxTokensField?: "max_completion_tokens" | "max_tokens";
 	/** Whether tool results require the `name` field. Default: auto-detected from URL. */
@@ -365,6 +380,7 @@ export type ResolvedOpenAICompat = Required<
 		| "thinkingKeep"
 		| "strictResponsesPairing"
 		| "requiresJuiceZeroHack"
+		| "enableGeminiThinkingLoopGuard"
 		| "whenThinking"
 	>
 > & {
@@ -379,6 +395,8 @@ export type ResolvedOpenAICompat = Required<
 	isOpenRouterHost: boolean;
 	/** The model sits behind Vercel AI Gateway. */
 	isVercelGatewayHost: boolean;
+	/** See {@link OpenAICompat.enableGeminiThinkingLoopGuard}. Set by the builder from the family classifier. */
+	enableGeminiThinkingLoopGuard?: boolean;
 	/** Complete alternate view for thinking-engaged requests; swap pointers, never spread. */
 	whenThinking?: ResolvedOpenAICompat;
 };
@@ -392,6 +410,8 @@ export interface ResolvedOpenAIResponsesCompat {
 	strictResponsesPairing: boolean;
 	requiresJuiceZeroHack: boolean;
 	reasoningEffortMap: Partial<Record<Effort, string>>;
+	/** See {@link OpenAICompat.enableGeminiThinkingLoopGuard}. */
+	enableGeminiThinkingLoopGuard?: boolean;
 }
 
 /** Fully-resolved anthropic-messages compat view (same contract as `ResolvedOpenAICompat`). */

package/src/variant-collapse.ts

@@ -107,80 +107,132 @@ const GEMINI_3_FLASH_FAMILY_EFFORTS: readonly Effort[] = [Effort.Minimal, Effort
 const GEMINI_3_PRO_FAMILY_EFFORTS: readonly Effort[] = [Effort.Low, Effort.High];
 
 /**
- * Shared by `google-antigravity` and `google-gemini-cli` — both serve the
- * Antigravity discovery list (`fetchAntigravityDiscoveryModels`).
+ * Antigravity Cloud Code Assist sends an explicit `thinkingBudget` per tier
+ * (verified against captured `daily-cloudcode-pa` requests). Flash uses round
+ * budgets; Pro offsets every budget by +1. Minimal mirrors Low (the Antigravity
+ * UI exposes Low/Medium/High only) so the effort stays selectable.
  */
-export const ANTIGRAVITY_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
-	families: [
-		{
-			id: "gemini-3.5-flash",
-			name: "Gemini 3.5 Flash",
-			members: ["gemini-3.5-flash-extra-low", "gemini-3.5-flash-low", "gemini-3-flash-agent"],
-			routing: {
-				off: "gemini-3.5-flash-extra-low",
-				[Effort.Minimal]: "gemini-3-flash-agent",
-				[Effort.Low]: "gemini-3.5-flash-extra-low",
-				[Effort.Medium]: "gemini-3.5-flash-extra-low",
-				[Effort.High]: "gemini-3.5-flash-low",
-			},
-			thinking: { mode: "google-level", efforts: GEMINI_3_FLASH_FAMILY_EFFORTS },
-			suppressWhenOff: true,
-			// Retired bare id; the alias only fires when no live model holds it
-			// (exact match wins in every resolver).
-			extraAliases: ["gemini-3-flash"],
-		},
-		{
-			id: "gemini-3.1-pro",
-			name: "Gemini 3.1 Pro",
-			// High routes to `gemini-pro-agent` — the upstream `gemini-3.1-pro-high`
-			// deployment returns INVALID_ARGUMENT on every streamGenerateContent
-			// request (both CCA endpoints) while discovery still lists it;
-			// `gemini-pro-agent` is the same model ("Gemini 3.1 Pro (High)", same
-			// thinking budget/caps) and accepts the identical request body.
-			// `gemini-3.1-pro-high` stays a member so the dead raw id is consumed.
-			members: ["gemini-3.1-pro-low", "gemini-pro-agent", "gemini-3.1-pro-high"],
-			retiredMembers: ["gemini-3.1-pro-high"],
-			routing: {
-				off: "gemini-3.1-pro-low",
-				[Effort.Low]: "gemini-3.1-pro-low",
-				[Effort.High]: "gemini-pro-agent",
-			},
-			thinking: { mode: "google-level", efforts: GEMINI_3_PRO_FAMILY_EFFORTS },
-			suppressWhenOff: true,
-		},
-		{
-			// Legacy static family — covers stale snapshots and caches.
-			id: "gemini-3-pro",
-			name: "Gemini 3 Pro",
-			members: ["gemini-3-pro-low", "gemini-3-pro-high"],
-			routing: {
-				off: "gemini-3-pro-low",
-				[Effort.Low]: "gemini-3-pro-low",
-				[Effort.High]: "gemini-3-pro-high",
-			},
-			thinking: { mode: "google-level", efforts: GEMINI_3_PRO_FAMILY_EFFORTS },
-			suppressWhenOff: true,
+const GEMINI_3_FLASH_FAMILY_BUDGETS: Readonly<Partial<Record<Effort, number>>> = {
+	[Effort.Minimal]: 1000,
+	[Effort.Low]: 1000,
+	[Effort.Medium]: 4000,
+	[Effort.High]: 10000,
+};
+const GEMINI_3_PRO_FAMILY_BUDGETS: Readonly<Partial<Record<Effort, number>>> = {
+	[Effort.Low]: 1001,
+	[Effort.High]: 10001,
+};
+
+/**
+ * The two Cloud Code Assist providers share the same Antigravity discovery list
+ * but disagree on the thinking transport: `google-antigravity` (daily-cloudcode-pa)
+ * sends an explicit `thinkingBudget` (verified against captured requests), while
+ * `google-gemini-cli` (cloudcode-pa) follows the official Gemini CLI and uses
+ * `thinkingLevel`. The Gemini 3.x families therefore differ only in thinking
+ * transport (and, for Flash, the per-tier wire-id routing); everything else is
+ * shared verbatim.
+ */
+function geminiFlashFamily(mode: "budget" | "google-level"): EffortVariantFamily {
+	const budget = mode === "budget";
+	return {
+		id: "gemini-3.5-flash",
+		name: "Gemini 3.5 Flash",
+		members: ["gemini-3.5-flash-extra-low", "gemini-3.5-flash-low", "gemini-3-flash-agent"],
+		routing: budget
+			? {
+					off: "gemini-3.5-flash-extra-low",
+					[Effort.Minimal]: "gemini-3.5-flash-extra-low",
+					[Effort.Low]: "gemini-3.5-flash-extra-low",
+					[Effort.Medium]: "gemini-3.5-flash-low",
+					[Effort.High]: "gemini-3-flash-agent",
+				}
+			: {
+					off: "gemini-3.5-flash-extra-low",
+					[Effort.Minimal]: "gemini-3-flash-agent",
+					[Effort.Low]: "gemini-3.5-flash-extra-low",
+					[Effort.Medium]: "gemini-3.5-flash-extra-low",
+					[Effort.High]: "gemini-3.5-flash-low",
+				},
+		thinking: budget
+			? { mode: "budget", efforts: GEMINI_3_FLASH_FAMILY_EFFORTS, effortBudgets: GEMINI_3_FLASH_FAMILY_BUDGETS }
+			: { mode: "google-level", efforts: GEMINI_3_FLASH_FAMILY_EFFORTS },
+		suppressWhenOff: true,
+		// Retired bare id; the alias only fires when no live model holds it
+		// (exact match wins in every resolver).
+		extraAliases: ["gemini-3-flash"],
+	};
+}
+
+function geminiProFamily(mode: "budget" | "google-level"): EffortVariantFamily {
+	const budget = mode === "budget";
+	return {
+		id: "gemini-3.1-pro",
+		name: "Gemini 3.1 Pro",
+		// High routes to `gemini-pro-agent` — the upstream `gemini-3.1-pro-high`
+		// deployment returns INVALID_ARGUMENT on every streamGenerateContent
+		// request (both CCA endpoints) while discovery still lists it;
+		// `gemini-pro-agent` is the same model ("Gemini 3.1 Pro (High)", same
+		// thinking budget/caps) and accepts the identical request body.
+		// `gemini-3.1-pro-high` stays a member so the dead raw id is consumed.
+		members: ["gemini-3.1-pro-low", "gemini-pro-agent", "gemini-3.1-pro-high"],
+		retiredMembers: ["gemini-3.1-pro-high"],
+		routing: {
+			off: "gemini-3.1-pro-low",
+			[Effort.Low]: "gemini-3.1-pro-low",
+			[Effort.High]: "gemini-pro-agent",
 		},
-		{
-			// Rename-only collapse: every effort and off fall back to the wire id.
-			id: "gpt-oss-120b",
-			name: "GPT-OSS 120B",
-			members: ["gpt-oss-120b-medium"],
-			routing: {},
-			thinking: { mode: "budget", efforts: [Effort.Minimal, Effort.Low, Effort.Medium, Effort.High] },
+		thinking: budget
+			? { mode: "budget", efforts: GEMINI_3_PRO_FAMILY_EFFORTS, effortBudgets: GEMINI_3_PRO_FAMILY_BUDGETS }
+			: { mode: "google-level", efforts: GEMINI_3_PRO_FAMILY_EFFORTS },
+		suppressWhenOff: true,
+	};
+}
+
+/** CCA families shared verbatim by both providers (transport-agnostic). */
+const SHARED_CCA_FAMILIES: readonly EffortVariantFamily[] = [
+	{
+		// Legacy static family — covers stale snapshots and caches. Stale ids are
+		// unverified against the budget-mode CCA contract; keep them on level.
+		id: "gemini-3-pro",
+		name: "Gemini 3 Pro",
+		members: ["gemini-3-pro-low", "gemini-3-pro-high"],
+		routing: {
+			off: "gemini-3-pro-low",
+			[Effort.Low]: "gemini-3-pro-low",
+			[Effort.High]: "gemini-3-pro-high",
 		},
-		thinkingPair("claude-sonnet-4-6", "Claude Sonnet 4.6"),
-		thinkingPair("claude-opus-4-6", "Claude Opus 4.6"),
-		thinkingPair("claude-sonnet-4-5", "Claude Sonnet 4.5"),
-		thinkingPair("claude-opus-4-5", "Claude Opus 4.5"),
-		thinkingPair("gemini-2.5-flash", "Gemini 2.5 Flash"),
-	],
+		thinking: { mode: "google-level", efforts: GEMINI_3_PRO_FAMILY_EFFORTS },
+		suppressWhenOff: true,
+	},
+	{
+		// Rename-only collapse: every effort and off fall back to the wire id.
+		id: "gpt-oss-120b",
+		name: "GPT-OSS 120B",
+		members: ["gpt-oss-120b-medium"],
+		routing: {},
+		thinking: { mode: "budget", efforts: [Effort.Minimal, Effort.Low, Effort.Medium, Effort.High] },
+	},
+	thinkingPair("claude-sonnet-4-6", "Claude Sonnet 4.6"),
+	thinkingPair("claude-opus-4-6", "Claude Opus 4.6"),
+	thinkingPair("claude-sonnet-4-5", "Claude Sonnet 4.5"),
+	thinkingPair("claude-opus-4-5", "Claude Opus 4.5"),
+	thinkingPair("gemini-2.5-flash", "Gemini 2.5 Flash"),
+];
+
+/** `google-antigravity` (daily-cloudcode-pa): Gemini 3.x on the budget transport. */
+export const ANTIGRAVITY_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
+	families: [geminiFlashFamily("budget"), geminiProFamily("budget"), ...SHARED_CCA_FAMILIES],
 };
 
-/** Provider id → hand collapse table. Both CCA providers share one table. */
+/** `google-gemini-cli` (cloudcode-pa): Gemini 3.x on the level transport (official CLI parity). */
+export const GEMINI_CLI_VARIANT_COLLAPSE_TABLE: VariantCollapseTable = {
+	families: [geminiFlashFamily("google-level"), geminiProFamily("google-level"), ...SHARED_CCA_FAMILIES],
+};
+
+/** Provider id → hand collapse table. The CCA providers diverge on thinking transport. */
 export const VARIANT_COLLAPSE_TABLES: Readonly<Record<string, VariantCollapseTable>> = {
 	"google-antigravity": ANTIGRAVITY_VARIANT_COLLAPSE_TABLE,
-	"google-gemini-cli": ANTIGRAVITY_VARIANT_COLLAPSE_TABLE,
+	"google-gemini-cli": GEMINI_CLI_VARIANT_COLLAPSE_TABLE,
 };
 
 /**
@@ -360,6 +412,47 @@ function reconcileRetiredRouting<TSpec extends VariantSpecLike>(
 	return next;
 }
 
+/**
+ * Refresh a collapsed snapshot's thinking surface in place. Bundled catalog and
+ * prev-generation snapshots freeze a family's transport, budgets, and routing;
+ * discovery emits the canonical id but the exact-id merge never overwrites a
+ * stale `family.id` row (e.g. `gemini-3.1-pro`) nor a recycled `extraAliases`
+ * row (e.g. `gemini-3-flash`). This re-applies the hand-table family's thinking,
+ * routing, and default wire id while keeping the spec id (load-bearing for exact
+ * selectors and bundled lookups). Returns `spec` by reference when unchanged.
+ */
+function refreshCollapsedThinking<TSpec extends VariantSpecLike>(
+	spec: TSpec,
+	family: EffortVariantFamily,
+	retired: ReadonlySet<string> | undefined,
+): TSpec {
+	// Scope snapshot self-heal to families carrying a curated per-effort budget
+	// contract (Antigravity gemini-3.x). Their routing targets are all verified
+	// live, so rebuilding routing here is safe; families without `effortBudgets`
+	// (derived `X`/`X-thinking` pairs, claude pairs) keep their presence-filtered
+	// snapshot routing untouched.
+	if (!spec.reasoning || family.thinking.effortBudgets === undefined) return spec;
+	const routing: Partial<Record<Effort | "off", string>> = {};
+	let hasRouting = false;
+	for (const effortKey in family.routing) {
+		const target = family.routing[effortKey as Effort | "off"];
+		if (target !== undefined && !retired?.has(target)) {
+			routing[effortKey as Effort | "off"] = target;
+			hasRouting = true;
+		}
+	}
+	const thinking: ThinkingConfig = { ...family.thinking };
+	if (hasRouting) thinking.effortRouting = routing;
+	if (family.suppressWhenOff) thinking.suppressWhenOff = true;
+	const offTarget = family.routing.off;
+	const requestModelId =
+		offTarget !== undefined && !retired?.has(offTarget) && offTarget !== spec.id ? offTarget : spec.requestModelId;
+	if (Bun.deepEquals(thinking, spec.thinking) && requestModelId === spec.requestModelId) {
+		return spec;
+	}
+	return { ...spec, thinking, ...(requestModelId !== undefined ? { requestModelId } : {}) };
+}
+
 /**
  * Collapse every family in `table` found in `specs`. Non-member specs pass
  * through verbatim (by reference), order preserved; the collapsed spec
@@ -394,11 +487,17 @@ export function collapseEffortVariants<TSpec extends VariantSpecLike>(
 				: existing;
 		const rawPresent = family.members.filter(id => byId.has(id) && !(id === family.id && existingCollapsed));
 		if (rawPresent.length === 0) {
-			// Inert (no members) or already collapsed (pass-through) — idempotence.
-			// A stale collapsed entry still gets retired routing re-pointed.
-			if (reconciled !== undefined && reconciled !== existing) {
+			// Inert (no members) or already collapsed (pass-through). A stale
+			// family.id-keyed snapshot is refreshed in place from the current
+			// hand-table family (transport/budgets/routing); retired targets drop.
+			// Recycled extraAliases rows are healed in a later pass.
+			const refreshed =
+				existing !== undefined && existingCollapsed
+					? refreshCollapsedThinking(reconciled ?? existing, family, retired)
+					: reconciled;
+			if (refreshed !== undefined && refreshed !== existing) {
 				familyIdBySpecId.set(family.id, family.id);
-				replacement.set(family.id, reconciled);
+				replacement.set(family.id, refreshed);
 			}
 			continue;
 		}
@@ -464,6 +563,27 @@ export function collapseEffortVariants<TSpec extends VariantSpecLike>(
 		replacement.set(family.id, collapsed);
 	}
 
+	// Refresh stale alias-keyed snapshots in place (recycled bare ids). Runs even
+	// when the canonical family.id row is also present, since the exact-id merge
+	// keeps the stale alias row alongside the discovered canonical one.
+	for (const family of table.families) {
+		if (family.extraAliases === undefined) continue;
+		const retired =
+			family.retiredMembers !== undefined && family.retiredMembers.length > 0
+				? new Set(family.retiredMembers)
+				: undefined;
+		for (const alias of family.extraAliases) {
+			if (alias === family.id || familyIdBySpecId.has(alias)) continue;
+			const aliasSpec = byId.get(alias);
+			if (aliasSpec === undefined) continue;
+			const refreshed = refreshCollapsedThinking(aliasSpec, family, retired);
+			if (refreshed !== aliasSpec) {
+				familyIdBySpecId.set(alias, alias);
+				replacement.set(alias, refreshed);
+			}
+		}
+	}
+
 	if (replacement.size === 0) return [...specs];
 
 	const emitted = new Set<string>();
@@ -602,7 +722,13 @@ export function resolveBareVariantAlias(modelId: string): BareVariantAliasHit |
 		if (hit === undefined) continue;
 		const providers: Provider[] = [];
 		for (const candidate in VARIANT_COLLAPSE_TABLES) {
-			if (VARIANT_COLLAPSE_TABLES[candidate] === table) providers.push(candidate);
+			// Match by resolved alias target, not table identity: the CCA providers
+			// now hold distinct table objects that still share these aliases.
+			if (
+				getAliasIndex(VARIANT_COLLAPSE_TABLES[candidate] as VariantCollapseTable).forward.get(normalized) === hit
+			) {
+				providers.push(candidate);
+			}
 		}
 		return { id: hit, providers };
 	}

package/src/wire/gemini-headers.ts

@@ -4,7 +4,7 @@
  * GeminiCLI/VERSION/MODEL (PLATFORM; ARCH; SURFACE)
  */
 export function getGeminiCliUserAgent(modelId = "gemini-3.1-pro-preview"): string {
-	const version = process.env.PI_AI_GEMINI_CLI_VERSION || "0.35.3";
+	const version = process.env.PI_AI_GEMINI_CLI_VERSION || "0.46.0";
 	const platform = process.platform === "win32" ? "win32" : process.platform;
 	const arch = process.arch === "x64" ? "x64" : process.arch;
 	return `GeminiCLI/${version}/${modelId} (${platform}; ${arch}; terminal)`;
@@ -20,8 +20,6 @@ export const ANTIGRAVITY_SYSTEM_INSTRUCTION =
 	"You are pair programming with a USER to solve their coding task. The task may require creating a new codebase, modifying or debugging an existing codebase, or simply answering a question." +
 	"**Absolute paths only**" +
 	"**Proactiveness**";
-export const ANTIGRAVITY_NO_PREAMBLE_INSTRUCTION =
-	'CRITICAL: NEVER output rule checks, formatting guidelines, constraint checklists (e.g. "No emdashes"), or your thinking/personality preambles in the final response. Output only the final response.';
 /**
  * Antigravity / Cloud Code Assist user agent. Lives in its own file so discovery
  * and usage code can read it without pulling the heavy google-gemini-cli provider
@@ -29,7 +27,7 @@ export const ANTIGRAVITY_NO_PREAMBLE_INSTRUCTION =
  * parse graph.
  */
 export let getAntigravityUserAgent = () => {
-	const DEFAULT_ANTIGRAVITY_VERSION = "1.104.0";
+	const DEFAULT_ANTIGRAVITY_VERSION = "2.1.4";
 	const version = process.env.PI_AI_ANTIGRAVITY_VERSION || DEFAULT_ANTIGRAVITY_VERSION;
 	// Map Node.js platform/arch to Antigravity's expected format.
 	// Verified against Antigravity source: _qn() and wqn() in main.js.
@@ -37,7 +35,32 @@ export let getAntigravityUserAgent = () => {
 	// process.arch:     x64→amd64, ia32→386, others pass through (arm64)
 	const os = process.platform === "win32" ? "windows" : process.platform;
 	const arch = process.arch === "x64" ? "amd64" : process.arch === "ia32" ? "386" : process.arch;
-	const userAgent = `antigravity/${version} ${os}/${arch}`;
+	const userAgent = `antigravity/hub/${version} ${os}/${arch}`;
 	getAntigravityUserAgent = () => userAgent;
 	return userAgent;
 };
+
+/**
+ * Per-wire-id Antigravity Cloud Code Assist request constants, captured from the
+ * real `antigravity/hub` client against `daily-cloudcode-pa`. `modelEnum` is the
+ * opaque `labels.model_enum` token the client tags each request with;
+ * `maxOutputTokens` is the fixed `generationConfig.maxOutputTokens` it sends
+ * regardless of the thinking budget. Keyed by the routed upstream wire id
+ * (post effort-routing), not the collapsed logical id. Checkpoint-only ids
+ * (e.g. `gemini-3.1-flash-lite`) are intentionally absent — this provider only
+ * emits agent requests.
+ */
+export interface AntigravityModelWireProfile {
+	modelEnum: string;
+	maxOutputTokens: number;
+}
+export const ANTIGRAVITY_MODEL_WIRE_PROFILES: Readonly<Record<string, AntigravityModelWireProfile>> = {
+	"gemini-3.5-flash-extra-low": { modelEnum: "MODEL_PLACEHOLDER_M187", maxOutputTokens: 65536 },
+	"gemini-3.5-flash-low": { modelEnum: "MODEL_PLACEHOLDER_M20", maxOutputTokens: 65536 },
+	"gemini-3-flash-agent": { modelEnum: "MODEL_PLACEHOLDER_M132", maxOutputTokens: 65536 },
+	"gemini-3.1-pro-low": { modelEnum: "MODEL_PLACEHOLDER_M36", maxOutputTokens: 65535 },
+	"gemini-pro-agent": { modelEnum: "MODEL_PLACEHOLDER_M16", maxOutputTokens: 65535 },
+};
+export function getAntigravityModelWireProfile(wireModelId: string): AntigravityModelWireProfile | undefined {
+	return ANTIGRAVITY_MODEL_WIRE_PROFILES[wireModelId];
+}

package/src/wire/github-copilot.ts

@@ -30,11 +30,13 @@ export const COPILOT_API_HEADERS = {
 type GitHubCopilotApiKeyPayload = {
 	token?: unknown;
 	enterpriseUrl?: unknown;
+	apiEndpoint?: unknown;
 };
 
 export type ParsedGitHubCopilotApiKey = {
 	accessToken: string;
 	enterpriseUrl?: string;
+	apiEndpoint?: string;
 };
 
 const PUBLIC_GITHUB_HOSTS = new Set(["api.github.com", "github.com", "www.github.com"]);
@@ -51,6 +53,18 @@ export function normalizeGitHubCopilotEnterpriseDomain(input: string | undefined
 	return normalized;
 }
 
+export function normalizeGitHubCopilotApiEndpoint(input: string | undefined): string | undefined {
+	const trimmed = input?.trim();
+	if (!trimmed?.startsWith("https://")) return undefined;
+	try {
+		const url = new URL(trimmed);
+		if (url.protocol !== "https:" || !url.hostname) return undefined;
+		return trimmed.replace(/\/+$/, "");
+	} catch {
+		return undefined;
+	}
+}
+
 export function parseGitHubCopilotApiKey(apiKeyRaw: string): ParsedGitHubCopilotApiKey {
 	try {
 		const parsed = JSON.parse(apiKeyRaw) as GitHubCopilotApiKeyPayload;
@@ -61,6 +75,10 @@ export function parseGitHubCopilotApiKey(apiKeyRaw: string): ParsedGitHubCopilot
 					typeof parsed.enterpriseUrl === "string"
 						? normalizeGitHubCopilotEnterpriseDomain(parsed.enterpriseUrl)
 						: undefined,
+				apiEndpoint:
+					typeof parsed.apiEndpoint === "string"
+						? normalizeGitHubCopilotApiEndpoint(parsed.apiEndpoint)
+						: undefined,
 			};
 		}
 	} catch {}