@oh-my-pi/pi-catalog 15.10.11 → 15.11.0

package/CHANGELOG.md

@@ -2,6 +2,26 @@
 
 ## [Unreleased]
 
+## [15.11.0] - 2026-06-10
+
+### Fixed
+
+- Fixed `buildModel` so malformed explicit thinking metadata without `efforts` is treated as sparse input and inferred instead of crashing during model resolution ([#2251](https://github.com/can1357/oh-my-pi/issues/2251)).
+
+## [15.10.12] - 2026-06-10
+
+### Added
+
+- Added `grok-composer-2.5-fast` (Cursor "Composer 2.5 Fast") to the xAI Grok OAuth (SuperGrok) catalog: non-reasoning, text-only, 200K context.
+
+### Changed
+
+- Set every xAI Grok OAuth (SuperGrok) curated model's max output tokens to mirror its context window (`grok-build`, `grok-4.3`, `grok-4.20-0309-{reasoning,non-reasoning}`, `grok-4.20-multi-agent-0309`, `grok-composer-2.5-fast`), replacing the `8888` `UNK_MAX_TOKENS` placeholder (and a stale `30000` on three grok-4.x entries). xAI's OAuth `/v1/models` reports no per-request output limit, so the curated catalog now owns `maxTokens` like `contextWindow`, deterministic on both the static-seed and online-overlay paths; the `openai-responses` wire still clamps the actual request to `OPENAI_MAX_OUTPUT_TOKENS` (64k).
+
+### Fixed
+
+- Excluded zero-cost `xai-oauth` subscription entries from the model reference indexes (`buildModelReferenceIndex`, `createReferenceResolver`), so their zero pricing and context-window-sized `maxTokens` cannot outrank paid/public Grok references when resolving custom-provider model identities.
+
 ## [15.10.11] - 2026-06-10
 
 ### Added

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

@@ -13,6 +13,7 @@ export interface ModelReferenceIndex {
     exact: Map<string, Model<Api>>;
     suffixAlias: Map<string, Model<Api>>;
 }
+export declare function isZeroCostXaiOAuthReference(candidate: Model<Api>): boolean;
 /**
  * Build a reference index from a model catalog (typically the bundled models).
  * Pure: callers are responsible for memoizing the result.

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

@@ -151,15 +151,15 @@ export declare const CATALOG_PROVIDERS: readonly [{
     readonly allowUnauthenticated: true;
 }, {
     readonly id: "minimax";
-    readonly defaultModel: "MiniMax-M2.5";
+    readonly defaultModel: "MiniMax-M3";
     readonly envVars: readonly ["MINIMAX_API_KEY"];
 }, {
     readonly id: "minimax-code";
-    readonly defaultModel: "MiniMax-M2.5";
+    readonly defaultModel: "MiniMax-M3";
     readonly envVars: readonly ["MINIMAX_CODE_API_KEY"];
 }, {
     readonly id: "minimax-code-cn";
-    readonly defaultModel: "MiniMax-M2.5";
+    readonly defaultModel: "MiniMax-M3";
     readonly envVars: readonly ["MINIMAX_CODE_CN_API_KEY"];
 }, {
     readonly id: "mistral";

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

@@ -118,8 +118,9 @@ export declare const XAI_OAUTH_CURATED_MODELS: readonly XAICuratedModel[];
  *
  * `reasoning` defaults to `true` for the Grok-4.x family; the explicit
  * `grok-4.20-0309-non-reasoning` entry opts out via `XAICuratedModel.reasoning`.
- * `maxTokens` uses `UNK_MAX_TOKENS` so id-keyed overlays from a successful
- * dynamic fetch merge cleanly. Mirrors
+ * `maxTokens` mirrors each model's `contextWindow` (the OAuth surface reports
+ * no per-request output limit); the openai-responses wire still clamps the
+ * actual request to OPENAI_MAX_OUTPUT_TOKENS. Mirrors
  * `hermes-agent/hermes_cli/models.py:_XAI_STATIC_FALLBACK`.
  */
 export declare function buildXaiOAuthStaticSeed(baseUrl?: string): ModelSpec<"openai-responses">[];

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "15.10.11",
+	"version": "15.11.0",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -34,11 +34,11 @@
 	},
 	"dependencies": {
 		"@bufbuild/protobuf": "^2.12.0",
-		"@oh-my-pi/pi-utils": "15.10.11",
+		"@oh-my-pi/pi-utils": "15.11.0",
 		"zod": "4.4.3"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "15.10.11",
+		"@oh-my-pi/pi-ai": "15.11.0",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/identity/reference.ts

@@ -17,7 +17,18 @@ export interface ModelReferenceIndex {
 	suffixAlias: Map<string, Model<Api>>;
 }
 
-// Custom provider entries often front a known upstream model through a local proxy.
+// xai-oauth subscription entries carry zero public pricing and inflated maxTokens;
+// keep them provider-local so they cannot outrank paid/public Grok references.
+export function isZeroCostXaiOAuthReference(candidate: Model<Api>): boolean {
+	return (
+		candidate.provider === "xai-oauth" &&
+		candidate.cost.input === 0 &&
+		candidate.cost.output === 0 &&
+		candidate.cost.cacheRead === 0 &&
+		candidate.cost.cacheWrite === 0
+	);
+}
+
 // Prefer the reference with the largest limits and complete cache pricing, then
 // first-party OpenAI entries.
 function shouldReplaceReference(existing: Model<Api> | undefined, candidate: Model<Api>): boolean {
@@ -47,6 +58,9 @@ function normalizeReferenceKey(value: string): string {
 export function buildModelReferenceIndex(models: Iterable<Model<Api>>): ModelReferenceIndex {
 	const exact = new Map<string, Model<Api>>();
 	for (const candidate of models) {
+		if (isZeroCostXaiOAuthReference(candidate)) {
+			continue;
+		}
 		const key = normalizeReferenceKey(candidate.id);
 		if (shouldReplaceReference(exact.get(key), candidate)) {
 			exact.set(key, candidate);

package/src/model-thinking.ts

@@ -97,7 +97,7 @@ export function resolveModelThinking<TApi extends Api>(
 ): ThinkingConfig | undefined {
 	if (!spec.reasoning) return undefined;
 	if (omitsWireReasoningEffort(spec.api, compat)) return undefined;
-	if (spec.thinking && spec.thinking.efforts.length > 0) {
+	if (spec.thinking && Array.isArray(spec.thinking.efforts) && spec.thinking.efforts.length > 0) {
 		return fillThinkingWireDefaults(spec, spec.thinking);
 	}
 	// Empty/malformed explicit metadata is treated as absent — infer instead.

package/src/models.json

@@ -69968,7 +69968,7 @@
 				"cacheWrite": 0
 			},
 			"contextWindow": 2000000,
-			"maxTokens": 30000,
+			"maxTokens": 2000000,
 			"compat": {
 				"reasoningEffortMap": {
 					"minimal": "low"
@@ -69993,7 +69993,7 @@
 				"cacheWrite": 0
 			},
 			"contextWindow": 2000000,
-			"maxTokens": 30000,
+			"maxTokens": 2000000,
 			"compat": {
 				"reasoningEffortMap": {
 					"minimal": "low"
@@ -70018,7 +70018,7 @@
 				"cacheWrite": 0
 			},
 			"contextWindow": 2000000,
-			"maxTokens": 8888,
+			"maxTokens": 2000000,
 			"compat": {
 				"reasoningEffortMap": {
 					"minimal": "low"
@@ -70053,7 +70053,7 @@
 				"cacheWrite": 0
 			},
 			"contextWindow": 1000000,
-			"maxTokens": 30000,
+			"maxTokens": 1000000,
 			"compat": {
 				"reasoningEffortMap": {
 					"minimal": "low"
@@ -70087,13 +70087,37 @@
 				"cacheWrite": 0
 			},
 			"contextWindow": 512000,
-			"maxTokens": 8888,
+			"maxTokens": 512000,
 			"compat": {
 				"reasoningEffortMap": {
 					"minimal": "low"
 				},
 				"supportsReasoningEffort": false
 			}
+		},
+		"grok-composer-2.5-fast": {
+			"id": "grok-composer-2.5-fast",
+			"name": "Grok Composer 2.5 Fast",
+			"api": "openai-responses",
+			"provider": "xai-oauth",
+			"baseUrl": "https://api.x.ai/v1",
+			"reasoning": false,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 200000,
+			"maxTokens": 200000,
+			"compat": {
+				"reasoningEffortMap": {
+					"minimal": "low"
+				}
+			}
 		}
 	},
 	"xiaomi": {

package/src/provider-models/bundled-references.ts

@@ -1,3 +1,4 @@
+import { isZeroCostXaiOAuthReference } from "../identity/reference";
 import { getBundledModels, getBundledProviders } from "../models";
 import type { Api, Model, ModelSpec } from "../types";
 
@@ -29,6 +30,9 @@ export function createReferenceResolver<TApi extends Api>(
 	for (const provider of getBundledProviders()) {
 		for (const model of getBundledModels(provider as Parameters<typeof getBundledModels>[0])) {
 			const candidate = model as Model<Api>;
+			if (isZeroCostXaiOAuthReference(candidate)) {
+				continue;
+			}
 			const existing = globalRefs.get(candidate.id);
 			if (!existing) {
 				globalRefs.set(candidate.id, candidate);

package/src/provider-models/descriptors.ts

@@ -191,17 +191,17 @@ export const CATALOG_PROVIDERS = [
 	},
 	{
 		id: "minimax",
-		defaultModel: "MiniMax-M2.5",
+		defaultModel: "MiniMax-M3",
 		envVars: ["MINIMAX_API_KEY"],
 	},
 	{
 		id: "minimax-code",
-		defaultModel: "MiniMax-M2.5",
+		defaultModel: "MiniMax-M3",
 		envVars: ["MINIMAX_CODE_API_KEY"],
 	},
 	{
 		id: "minimax-code-cn",
-		defaultModel: "MiniMax-M2.5",
+		defaultModel: "MiniMax-M3",
 		envVars: ["MINIMAX_CODE_CN_API_KEY"],
 	},
 	{

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

@@ -739,6 +739,17 @@ export const XAI_OAUTH_CURATED_MODELS: readonly XAICuratedModel[] = [
 		reasoning: false,
 		input: ["text", "image"],
 	},
+	// Cursor's "Composer 2.5 Fast" exposed via SuperGrok: non-reasoning,
+	// text-only, 200K context (mirrors Cursor's composer-* catalog entries).
+	// Off the GROK_EFFORT_CAPABLE_PREFIXES allowlist, so the wire side already
+	// sets omitReasoningEffort=true; reasoning:false also hides the effort dial.
+	{
+		id: "grok-composer-2.5-fast",
+		contextWindow: 200_000,
+		name: "Grok Composer 2.5 Fast",
+		reasoning: false,
+		input: ["text"],
+	},
 ] as const;
 
 // xAI /v1/models returns chat, image, voice, and STT entries. Tool surfaces
@@ -754,6 +765,13 @@ const XAI_NON_CHAT_PREFIXES = ["grok-imagine-", "grok-stt-", "grok-voice-"] as c
 // at request time, downstream of the omitReasoningEffort gate in xai-responses.ts.
 const XAI_REASONING_EFFORT_MAP = { minimal: "low" } as const;
 
+// xai-oauth's /v1/models exposes no per-request output limit on the OAuth
+// (Grok Build / SuperGrok) surface, so the curated catalog owns `maxTokens`
+// like it owns `contextWindow`: each entry mirrors its context window. The
+// openai-responses wire clamps the actual request to
+// min(requested, model.maxTokens, OPENAI_MAX_OUTPUT_TOKENS=64000), so this is
+// just "no model-specific sub-cap below 64k", not an unbounded output budget.
+
 // Single source of truth for curated → Model fan-in. Used by the static-seed
 // and the dynamic overlay/inject paths (applyXAIOAuthCuration) so curated
 // reasoning/effort flags survive an online refresh (xAI's /v1/models lacks
@@ -776,6 +794,7 @@ function mergeCuratedIntoModel(
 	return {
 		...base,
 		contextWindow: curated.contextWindow,
+		maxTokens: curated.contextWindow,
 		name: curated.name ?? base.name,
 		reasoning: curated.reasoning ?? true,
 		input: curated.input ?? base.input,
@@ -850,8 +869,9 @@ function applyXAIOAuthCuration(dynamic: readonly ModelSpec<"openai-responses">[]
  *
  * `reasoning` defaults to `true` for the Grok-4.x family; the explicit
  * `grok-4.20-0309-non-reasoning` entry opts out via `XAICuratedModel.reasoning`.
- * `maxTokens` uses `UNK_MAX_TOKENS` so id-keyed overlays from a successful
- * dynamic fetch merge cleanly. Mirrors
+ * `maxTokens` mirrors each model's `contextWindow` (the OAuth surface reports
+ * no per-request output limit); the openai-responses wire still clamps the
+ * actual request to OPENAI_MAX_OUTPUT_TOKENS. Mirrors
  * `hermes-agent/hermes_cli/models.py:_XAI_STATIC_FALLBACK`.
  */
 export function buildXaiOAuthStaticSeed(baseUrl?: string): ModelSpec<"openai-responses">[] {
@@ -871,7 +891,7 @@ export function buildXaiOAuthStaticSeed(baseUrl?: string): ModelSpec<"openai-res
 			input: ["text"],
 			cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
 			contextWindow: curated.contextWindow,
-			maxTokens: UNK_MAX_TOKENS,
+			maxTokens: curated.contextWindow,
 			compat: { reasoningEffortMap: XAI_REASONING_EFFORT_MAP },
 		};
 		return mergeCuratedIntoModel(base, curated);