@oh-my-pi/pi-catalog 18.0.7 → 18.0.8

package/CHANGELOG.md

@@ -2,6 +2,13 @@
 
 ## [Unreleased]
 
+## [18.0.8] - 2026-08-27
+
+### Fixed
+
+- Fixed the thinking control mode for OpenAI models served over Bedrock Converse (`global.openai.gpt-5.6-luna`, `-sol`, `-terra`), which are now classified as `effort` rather than `budget` so requests use OpenAI's reasoning schema.
+- Fixed LiteLLM discovery leaking a colliding bundled model's provider-specific transport onto custom endpoints: a discovered alias (e.g. `kimi-k3`) matching a bundled Fireworks model no longer inherits that model's wire-id transform, which had caused requests to POST a model id the endpoint never advertised and return HTTP 400 ([#9938](https://github.com/can1357/oh-my-pi/issues/9938)).
+
 ## [18.0.7] - 2026-08-26
 
 ### Added

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "18.0.7",
+	"version": "18.0.8",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"homepage": "https://omp.sh",
 	"author": "Stencil Labs, Inc.",
@@ -34,11 +34,11 @@
 		"gen:proto": "bun scripts/generate-protocols.ts"
 	},
 	"dependencies": {
-		"@oh-my-pi/omptype": "18.0.7",
-		"@oh-my-pi/pi-utils": "18.0.7"
+		"@oh-my-pi/omptype": "18.0.8",
+		"@oh-my-pi/pi-utils": "18.0.8"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "18.0.7",
+		"@oh-my-pi/pi-ai": "18.0.8",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/model-thinking.ts

@@ -777,6 +777,14 @@ function inferThinkingControlMode<TApi extends Api>(
 					return "anthropic-budget-effort";
 				}
 			}
+			// Bedrock serves the GPT-5.x models through OpenAI's own request
+			// schema, which rejects Anthropic's budget block outright:
+			// `unknown_parameter: 'thinking'`. It takes `reasoning.effort`
+			// instead. gpt-oss parses as `unknown` (no `gpt-<digits>`), so it
+			// keeps the budget path it ships with today.
+			if (parsedModel.family === "openai") {
+				return "effort";
+			}
 			return "budget";
 
 		default:

package/src/models.json

@@ -28688,7 +28688,7 @@
 			"contextWindow": 1050000,
 			"maxTokens": 128000,
 			"thinking": {
-				"mode": "budget",
+				"mode": "effort",
 				"efforts": [
 					"low",
 					"medium",
@@ -28727,7 +28727,7 @@
 			"contextWindow": 1050000,
 			"maxTokens": 128000,
 			"thinking": {
-				"mode": "budget",
+				"mode": "effort",
 				"efforts": [
 					"low",
 					"medium",
@@ -28766,7 +28766,7 @@
 			"contextWindow": 1050000,
 			"maxTokens": 128000,
 			"thinking": {
-				"mode": "budget",
+				"mode": "effort",
 				"efforts": [
 					"low",
 					"medium",

package/src/provider-models/cache-provider-id.ts

@@ -60,7 +60,10 @@ export function resolveModelCacheProviderId(providerId: string, options: ModelCa
 			return "cursor:default-effort-v4";
 		case "litellm": {
 			const baseUrl = options.baseUrl ?? getDefaultModelDiscoveryBaseUrl(providerId)!;
-			return `litellm:rich-v7:${Bun.hash(baseUrl).toString(36)}`;
+			// rich-v8 invalidates rows whose `compatConfig` retained a colliding
+			// bundled model's provider-specific transport (e.g. Fireworks
+			// `wireModelIdMode`) before that leak was fixed (issue #9938).
+			return `litellm:rich-v8:${Bun.hash(baseUrl).toString(36)}`;
 		}
 		case "opencode-go":
 		case "opencode-zen": {

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

@@ -5397,12 +5397,29 @@ function mapLiteLLMRichEntry<TApi extends Api>(
 							["tools", "tool_choice", "functions", "function_call"].includes(param),
 						)
 					: reference?.supportsTools;
+	// Enrich from the bundled reference with provider-INDEPENDENT reasoning
+	// hints only. The reference is resolved against the global bundled catalog,
+	// so a custom endpoint exposing an alias that collides with a bundled model
+	// (a LiteLLM proxy serving `kimi-k3`, which matches Fireworks' bundled
+	// `kimi-k3`) must not inherit that provider's transport compat. Spreading
+	// the resolved `reference.compat` wholesale leaked `wireModelIdMode`,
+	// `toolSchemaFlavor`, `thinkingFormat`, etc. across the provider boundary —
+	// rewriting the wire id to `accounts/fireworks/models/kimi-k3` for a
+	// non-Fireworks endpoint (issue #9938). `buildModel` re-derives every
+	// transport field from the discovered provider and model id, so only the
+	// effort vocabulary flows through here. Mirrors `discoverOpenAIModelsList`.
+	const referenceCompat = reference?.compat as OpenAICompat | undefined;
 	const compat: OpenAICompat = {
-		...(reference?.compat ?? {}),
 		supportsStore: false,
 		supportsDeveloperRole: false,
 		...(supportedOpenAIParams !== undefined
 			? { supportsReasoningEffort: supportedOpenAIParams.includes("reasoning_effort") }
+			: referenceCompat?.supportsReasoningEffort !== undefined
+				? { supportsReasoningEffort: referenceCompat.supportsReasoningEffort }
+				: {}),
+		...(referenceCompat?.reasoningEffortMap ? { reasoningEffortMap: referenceCompat.reasoningEffortMap } : {}),
+		...(referenceCompat?.omitReasoningEffort !== undefined
+			? { omitReasoningEffort: referenceCompat.omitReasoningEffort }
 			: {}),
 	};
 	return {
@@ -5628,12 +5645,14 @@ export function litellmModelManagerOptions(config?: LiteLLMModelManagerConfig):
 	const baseUrl = config?.baseUrl ?? getDefaultModelDiscoveryBaseUrl("litellm")!;
 	return {
 		providerId: "litellm",
-		// rich-v7 invalidates rows cached before discovery continued past endpoints
-		// that omitted cache pricing. Earlier versions added bundled reference fallback,
-		// moved OpenAI models to Responses, continued past incomplete vision and API
-		// metadata, stripped reseller usage suffixes, filtered placeholder rows, and
-		// mapped rich pricing. Bump the version whenever these mappers change, or warm
-		// authoritative caches keep serving pre-change rows for the full TTL.
+		// rich-v8 invalidates rows whose `compatConfig` retained a colliding
+		// bundled model's provider-specific transport (e.g. Fireworks
+		// `wireModelIdMode`) before that leak was fixed. Earlier versions added
+		// bundled reference fallback, moved OpenAI models to Responses, continued
+		// past incomplete vision/API metadata and endpoints omitting cache
+		// pricing, stripped reseller usage suffixes, filtered placeholder rows,
+		// and mapped rich pricing. Bump the version whenever these mappers change,
+		// or warm authoritative caches keep serving pre-change rows for the full TTL.
 		cacheProviderId: resolveModelCacheProviderId("litellm", { baseUrl }),
 		// 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