@oh-my-pi/pi-catalog 18.12.2 → 18.12.3

package/CHANGELOG.md

@@ -2,6 +2,16 @@
 
 ## [Unreleased]
 
+## [18.12.3] - 2026-10-19
+
+### Changed
+
+- MiniMax `MiniMax-M3.1-Flash` (the launch id) now gets the always-on `low`–`max` thinking the bundled `MiniMax-M3.1-Flash-Preview` rows have, and omits sampling params (`temperature`, `top_p`, `top_k`) it rejects ([#16315](https://github.com/can1357/oh-my-pi/pull/16315) by [@SaladDay](https://github.com/SaladDay))
+
+### Fixed
+
+- Gemma 4 models on Google AI Studio (`google/gemma-4-31b-it`, `google/gemma-4-26b-a4b-it`, …) no longer fail every request with `Thinking budget is not supported for this model`; thinking is now toggled with `thinkingLevel` (on = `high`, off = `minimal`) ([#16350](https://github.com/can1357/oh-my-pi/issues/16350))
+
 ## [18.12.2] - 2026-10-18
 
 ### Changed

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "18.12.2",
+	"version": "18.12.3",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"keywords": [
 		"ai",
@@ -93,12 +93,12 @@
 		"gen:proto": "bun scripts/generate-protocols.ts"
 	},
 	"dependencies": {
-		"@oh-my-pi/omptype": "18.12.2",
-		"@oh-my-pi/pi-utils": "18.12.2"
+		"@oh-my-pi/omptype": "18.12.3",
+		"@oh-my-pi/pi-utils": "18.12.3"
 	},
 	"devDependencies": {
 		"@bgotink/kdl": "0.4.0",
-		"@oh-my-pi/pi-ai": "18.12.2",
+		"@oh-my-pi/pi-ai": "18.12.3",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/rules/classes/gemma.kdl

@@ -11,4 +11,15 @@ class "gemma" {
 	on "google" "vercel-ai-gateway" {
 		thinking-mode "budget"
 	}
+	// Gemma 4 on the Gemini API toggles thinking only through `thinkingLevel`
+	// (`HIGH` on, `MINIMAL` off) and rejects any `thinkingBudget` with a 400
+	// (https://ai.google.dev/gemma/docs/core/gemma_on_gemini_api#thinking).
+	// Off must be sent explicitly as `MINIMAL` rather than omitting the config.
+	on "google" {
+		revision ">=4" {
+			thinking-mode "google-level"
+			thinking-efforts "high"
+			thinking-suppress-when-off #true
+		}
+	}
 }

package/src/compat/rules/classes/minimax.kdl

@@ -81,15 +81,17 @@ class "minimax" {
 			thinking-efforts "low" "medium" "high"
 		}
 	}
-	// MiniMax-M3.1-Flash-Preview always thinks (`thinking.type: "disabled"` and
-	// effort `none` 400 with error 2013) and tunes depth through
-	// `output_config.effort` from `low` to `max` (the default when omitted).
-	// The identity map replaces the M3 family's adaptive-tag-only map so each
-	// level reaches the wire. Like the M2 family, thinking cannot be turned off:
-	// `thinking-requires-effort` hides the off level and clamps forced-off
-	// requests to the lowest effort instead of sending `disabled`.
+	// MiniMax-M3.1-Flash (and its bundled `-Preview` id) always thinks
+	// (`thinking.type: "disabled"` and effort `none` 400 with error 2013) and
+	// tunes depth through `output_config.effort` from `low` to `max` (the
+	// default when omitted). The identity map replaces the M3 family's
+	// adaptive-tag-only map so each level reaches the wire. Like the M2 family,
+	// thinking cannot be turned off: `thinking-requires-effort` hides the off
+	// level and clamps forced-off requests to the lowest effort instead of
+	// sending `disabled`. The launch id also rejects sampling params
+	// (`temperature`, `top_p`, `top_k`).
 	on "minimax" "minimax-cn" "minimax-code" "minimax-code-cn" {
-		models "MiniMax-M3.1-Flash-Preview" {
+		models "MiniMax-M3.1-Flash" "MiniMax-M3.1-Flash-Preview" {
 			thinking-requires-effort #true
 			thinking-efforts "low" "medium" "high" "xhigh" "max"
 			thinking-effort-map {
@@ -98,6 +100,9 @@ class "minimax" {
 				high "high"
 			}
 		}
+		models "MiniMax-M3.1-Flash" {
+			supports-sampling-params #false
+		}
 	}
 	// Token Plan rows ride the Anthropic route; these OpenAI-only wire axes
 	// serve the legacy lightning rows that stay on the OpenAI-compatible one.