@gajae-code/ai 0.6.2 → 0.6.4

package/CHANGELOG.md

@@ -2,6 +2,12 @@
 
 ## [Unreleased]
 
+## [0.6.4] - 2026-06-20
+
+### Fixed
+
+- Fixed argument mis-attribution in the OpenAI-compatible Responses API streaming decoder when a single response emits multiple tool-call items. The decoder buffered streamed argument deltas against a single most-recent item/block slot, so interleaved or back-to-back `function_call`/`custom_tool_call` argument deltas could be applied to the wrong item — finalizing a tool call with another call's arguments (e.g. one tool's payload landing on a different tool's schema and tripping validation). Streamed deltas now accumulate against a per-item buffer keyed on stable item identity (`item_id` primary; positional `output_index` only when finite), each block records its content index at registration time, and finalization writes onto the same block stored in the message content while reading only the matching item's buffer. Single-tool-call streams, reasoning/text streaming, and the Chat Completions path are unchanged.
+
 ## [0.6.2] - 2026-06-19
 ### Added
 

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

@@ -228,6 +228,8 @@ export interface ModelsDevProviderDescriptor {
      * Can return null to skip the model, or an array to emit multiple models.
      */
     transformModel?: (model: Model<Api>, modelId: string, raw: ModelsDevModel) => Model<Api> | Model<Api>[] | null;
+    /** Optional static rows appended after mapped models. Used only for official provider catalogs missing from models.dev. */
+    appendModels?: readonly Model<Api>[];
     /**
      * Optional: override the API type per-model.
      * Called with (modelId, raw). Return the API type to use.

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@gajae-code/ai",
-	"version": "0.6.2",
+	"version": "0.6.4",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"homepage": "https://gaebal-gajae.dev",
 	"author": "Yeachan-Heo",
@@ -43,7 +43,7 @@
 	"dependencies": {
 		"@anthropic-ai/sdk": "^0.94.0",
 		"@bufbuild/protobuf": "^2.12.0",
-		"@gajae-code/utils": "0.6.2",
+		"@gajae-code/utils": "0.6.4",
 		"openai": "^6.36.0",
 		"partial-json": "^0.1.7",
 		"zod": "4.4.3"

package/src/model-thinking.ts

@@ -409,7 +409,7 @@ function applyGeneratedModelPolicy(model: ApiModel<Api>): void {
 	// 512K. The stale 512K survives generate-models (provider-scoped models bypass the
 	// models.dev refresh in applyGlobalModelsDevFallback), tripping auto-compaction /
 	// context-cap thresholds 2x early on MiniMax sessions. Pin to the true 1M.
-	if (model.id === "minimax-m3") {
+	if (model.provider !== "opencode-go" && model.id === "minimax-m3") {
 		model.contextWindow = 1_000_000;
 	}
 }