@oh-my-pi/pi-catalog 16.1.13 → 16.1.15

package/CHANGELOG.md

@@ -2,6 +2,22 @@
 
 ## [Unreleased]
 
+## [16.1.14] - 2026-06-22
+
+### Added
+
+- Added Sakana AI provider support with Fugu model integration
+- Added Sakana AI/Fugu provider catalog entries with Fugu model discovery and Responses API metadata
+- Added support for "xhigh" reasoning tier across model configurations
+- Added configuration for new models GCP-5.4 Mini, GPT-5.5, and variants
+- Added `devin` variant collapse table to streamline model tiering
+
+### Changed
+
+- Updated reasoning label pattern to include "minimal" and "max" efforts
+- Simplified model identification logic for Devin-powered reasoning models
+- Refactored variant routing to consolidate and standardize tier definitions
+
 ## [16.1.13] - 2026-06-22
 
 ### Added

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

@@ -0,0 +1,9 @@
+import type { ModelSpec, ResolvedDevinCompat } from "../types";
+/**
+ * Resolve devin-agent (Codeium Cascade) compat. Cascade has no wire
+ * reasoning/effort field — effort is selected by routing to a sibling model id
+ * (the `thinking.effortRouting` baked by variant-collapse). So the thinking
+ * deriver must never fabricate an effort ladder from identity for these models;
+ * only explicit routed metadata counts.
+ */
+export declare function buildDevinCompat(_spec: ModelSpec<"devin-agent">): ResolvedDevinCompat;

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

@@ -266,6 +266,15 @@ export declare const CATALOG_PROVIDERS: readonly [{
         readonly label: "Qwen Portal";
         readonly oauthProvider: "qwen-portal";
     };
+}, {
+    readonly id: "sakana";
+    readonly defaultModel: "fugu";
+    readonly envVars: readonly ["SAKANA_API_KEY", "FUGU_API_KEY"];
+    readonly createModelManagerOptions: (config: ModelManagerConfig) => import("..").ModelManagerOptions<"openai-responses", unknown>;
+    readonly dynamicModelsAuthoritative: true;
+    readonly catalogDiscovery: {
+        readonly label: "Sakana AI";
+    };
 }, {
     readonly id: "synthetic";
     readonly defaultModel: "hf:zai-org/GLM-5.1";

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

@@ -313,6 +313,13 @@ export interface MoonshotModelManagerConfig {
     fetch?: FetchImpl;
 }
 export declare function moonshotModelManagerOptions(config?: MoonshotModelManagerConfig): ModelManagerOptions<"openai-completions">;
+export declare const SAKANA_FUGU_STATIC_MODELS: readonly ModelSpec<"openai-responses">[];
+export interface SakanaModelManagerConfig {
+    apiKey?: string;
+    baseUrl?: string;
+    fetch?: FetchImpl;
+}
+export declare function sakanaModelManagerOptions(config?: SakanaModelManagerConfig): ModelManagerOptions<"openai-responses">;
 export interface QwenPortalModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;

package/dist/types/types.d.ts

@@ -439,6 +439,7 @@ export interface ResolvedOpenAIResponsesCompat extends ResolvedOpenAISharedCompa
     supportsImageDetailOriginal: boolean;
     requiresJuiceZeroHack: boolean;
     supportsObfuscationOptOut: boolean;
+    streamIdleTimeoutMs?: number;
 }
 /**
  * OpenRouter is a pseudo API: runtime dispatch can use either Responses
@@ -456,10 +457,28 @@ export type ResolvedAnthropicCompat = Required<AnthropicCompat> & {
      */
     officialEndpoint: boolean;
 };
+/**
+ * Compatibility settings for the devin-agent (Codeium Cascade) API. Cascade
+ * selects reasoning effort only by routing to a sibling model id (the
+ * `thinking.effortRouting` baked by variant-collapse), never by a wire
+ * reasoning/effort field, so the model-thinking deriver must not invent an
+ * effort ladder from identity for these models.
+ */
+export interface DevinCompat {
+    /**
+     * Trust only explicit `thinking` metadata; never derive a thinking surface
+     * from model identity. A reasoning model with no explicit routed thinking
+     * resolves to `thinking: undefined` (`reasoning: true`, no controllable
+     * effort) instead of a fabricated minimal/low/medium/high ladder.
+     */
+    trustExplicitThinkingOnly?: boolean;
+}
+/** Fully-resolved devin-agent compat view. */
+export type ResolvedDevinCompat = Required<DevinCompat>;
 /** Sparse, user-authored compat overrides for a given API (models.json / config vocabulary). */
-export type CompatConfigOf<TApi extends Api> = TApi extends "openai-completions" | "openrouter" | "openai-responses" | "azure-openai-responses" | "openai-codex-responses" ? OpenAICompat : TApi extends "anthropic-messages" ? AnthropicCompat : undefined;
+export type CompatConfigOf<TApi extends Api> = TApi extends "openai-completions" | "openrouter" | "openai-responses" | "azure-openai-responses" | "openai-codex-responses" ? OpenAICompat : TApi extends "anthropic-messages" ? AnthropicCompat : TApi extends "devin-agent" ? DevinCompat : undefined;
 /** Resolved compat for a given API: complete record, materialized once by `buildModel`. */
-export type CompatOf<TApi extends Api> = TApi extends "openrouter" ? ResolvedOpenRouterCompat : TApi extends "openai-completions" ? ResolvedOpenAICompat : TApi extends "openai-responses" | "azure-openai-responses" | "openai-codex-responses" ? ResolvedOpenAIResponsesCompat : TApi extends "anthropic-messages" ? ResolvedAnthropicCompat : undefined;
+export type CompatOf<TApi extends Api> = TApi extends "openrouter" ? ResolvedOpenRouterCompat : TApi extends "openai-completions" ? ResolvedOpenAICompat : TApi extends "openai-responses" | "azure-openai-responses" | "openai-codex-responses" ? ResolvedOpenAIResponsesCompat : TApi extends "anthropic-messages" ? ResolvedAnthropicCompat : TApi extends "devin-agent" ? ResolvedDevinCompat : undefined;
 export interface Model<TApi extends Api = Api> {
     id: string;
     /**

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

@@ -57,6 +57,7 @@ export interface VariantCollapseTable {
 export declare const ANTIGRAVITY_VARIANT_COLLAPSE_TABLE: VariantCollapseTable;
 /** `google-gemini-cli` (cloudcode-pa): Gemini 3.x on the level transport (official CLI parity). */
 export declare const GEMINI_CLI_VARIANT_COLLAPSE_TABLE: VariantCollapseTable;
+export declare const DEVIN_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>>;
 /**

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "16.1.13",
+	"version": "16.1.15",
 	"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": "16.1.13",
+		"@oh-my-pi/pi-utils": "16.1.15",
 		"arktype": "^2.2.0",
 		"zod": "^4"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "16.1.13",
+		"@oh-my-pi/pi-ai": "16.1.15",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/build.ts

@@ -10,6 +10,7 @@
  * compat per request.
  */
 import { buildAnthropicCompat } from "./compat/anthropic";
+import { buildDevinCompat } from "./compat/devin";
 import { buildOpenAICompat, buildOpenAIResponsesCompat, buildOpenRouterCompat } from "./compat/openai";
 import { resolveModelThinking } from "./model-thinking";
 import type { Api, CompatOf, Model, ModelSpec } from "./types";
@@ -38,6 +39,8 @@ export function buildCompat(spec: ModelSpec<Api>): CompatOf<Api> {
 			return buildOpenAIResponsesCompat(spec as ModelSpec<"openai-responses">);
 		case "anthropic-messages":
 			return buildAnthropicCompat(spec as ModelSpec<"anthropic-messages">);
+		case "devin-agent":
+			return buildDevinCompat(spec as ModelSpec<"devin-agent">);
 		default:
 			return undefined;
 	}

package/src/compat/devin.ts

@@ -0,0 +1,12 @@
+import type { ModelSpec, ResolvedDevinCompat } from "../types";
+
+/**
+ * Resolve devin-agent (Codeium Cascade) compat. Cascade has no wire
+ * reasoning/effort field — effort is selected by routing to a sibling model id
+ * (the `thinking.effortRouting` baked by variant-collapse). So the thinking
+ * deriver must never fabricate an effort ladder from identity for these models;
+ * only explicit routed metadata counts.
+ */
+export function buildDevinCompat(_spec: ModelSpec<"devin-agent">): ResolvedDevinCompat {
+	return { trustExplicitThinkingOnly: true };
+}

package/src/compat/openai.ts

@@ -535,6 +535,7 @@ export function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): Resol
 		emptyLengthFinishIsContextError: spec.provider === "ollama",
 		usesOpenAIToolCallIdLimit: spec.provider === "openai",
 		promptCacheSessionHeader: spec.provider === "xai-oauth" ? "x-grok-conv-id" : undefined,
+		streamIdleTimeoutMs: spec.compat?.streamIdleTimeoutMs,
 	};
 	applyCompatOverrides(compat, spec.compat);
 	if (spec.compat?.reasoningDisableMode === undefined) {

package/src/discovery/devin.ts

@@ -17,9 +17,8 @@ const DEFAULT_CONTEXT_WINDOW = 200_000;
 const DEFAULT_MAX_TOKENS = 64_000;
 
 /** Best-effort match for labels whose wording implies a thinking / reasoning-effort variant. */
-const REASONING_LABEL_PATTERN = /think|thinking|high|medium|low|xhigh|reasoning/i;
+const REASONING_LABEL_PATTERN = /think|thinking|minimal|high|medium|low|xhigh|max|reasoning/i;
 const NO_REASONING_LABEL_PATTERN = /\bno thinking\b/i;
-
 function supportsDevinThinking(config: ClientModelConfig): boolean {
 	if (NO_REASONING_LABEL_PATTERN.test(config.label)) return false;
 	return config.modelInfo?.modelFeatures?.supportsThinking === true || REASONING_LABEL_PATTERN.test(config.label);

package/src/model-thinking.ts

@@ -35,6 +35,7 @@ import type {
 	CompatOf,
 	Model,
 	ModelSpec,
+	ResolvedDevinCompat,
 	ResolvedOpenAICompat,
 	ResolvedOpenAIResponsesCompat,
 	ThinkingConfig,
@@ -61,6 +62,10 @@ const GPT_5_1_CODEX_MINI_EFFORTS: readonly Effort[] = [Effort.Medium, Effort.Hig
 const LOW_MEDIUM_HIGH_REASONING_EFFORTS: readonly Effort[] = [Effort.Low, Effort.Medium, Effort.High];
 const GLM_52_HIGH_MAX_REASONING_EFFORTS: readonly Effort[] = [Effort.High, Effort.XHigh];
 
+const FUGU_REASONING_EFFORTS: readonly Effort[] = [Effort.High, Effort.XHigh];
+const FUGU_REASONING_EFFORT_MAP: Readonly<EffortMap> = {
+	[Effort.XHigh]: "max",
+};
 type EffortMap = Partial<Record<Effort, string>>;
 
 const GROQ_QWEN3_32B_REASONING_EFFORT_MAP: Readonly<EffortMap> = {
@@ -153,6 +158,10 @@ export function resolveModelThinking<TApi extends Api>(
 	if (spec.thinking && Array.isArray(spec.thinking.efforts) && spec.thinking.efforts.length > 0) {
 		return fillThinkingWireDefaults(spec, compat, spec.thinking);
 	}
+	// Cascade selects effort only by routing to a sibling model id, so a Devin
+	// model with no explicit routed thinking has no controllable surface —
+	// never fabricate an effort ladder from identity.
+	if ((compat as ResolvedDevinCompat | undefined)?.trustExplicitThinkingOnly === true) return undefined;
 	// Empty/malformed explicit metadata is treated as absent — infer instead.
 	return deriveThinking(spec, compat);
 }
@@ -301,6 +310,9 @@ function getModelDefinedEfforts<TApi extends Api>(
 			return GLM_52_HIGH_MAX_REASONING_EFFORTS;
 		}
 	}
+	if (isSakanaFuguReasoningModel(spec)) {
+		return FUGU_REASONING_EFFORTS;
+	}
 	return isOpenAICompatReasoningApi(spec.api) &&
 		(isMinimaxM2FamilyModelId(spec.id) ||
 			isOpenAIGptOssModelId(spec.id) ||
@@ -377,6 +389,9 @@ function inferDetectedEffortMap<TApi extends Api>(
 	if (isOllamaCloudGlm52ReasoningEffortModel(spec)) {
 		return GLM_52_XHIGH_MAX_EFFORT_MAP;
 	}
+	if (isSakanaFuguReasoningModel(spec)) {
+		return FUGU_REASONING_EFFORT_MAP;
+	}
 	if (!isOpenAICompatReasoningApi(spec.api)) {
 		return undefined;
 	}
@@ -400,6 +415,10 @@ function inferDetectedEffortMap<TApi extends Api>(
 	return map;
 }
 
+function isSakanaFuguReasoningModel<TApi extends Api>(spec: ModelSpec<TApi>): boolean {
+	return spec.provider === "sakana" && /^fugu(?:$|-)/i.test(spec.id);
+}
+
 function isDeepseekReasoningModel<TApi extends Api>(spec: ModelSpec<TApi>): boolean {
 	if (!spec.reasoning) return false;
 	const lowerId = spec.id.toLowerCase();