fix(coding-agent): retarget OpenAI automatic defaults

packages/coding-agent/CHANGELOG.md

@@ -16,6 +16,8 @@
 
 ### Changed
 
+- Automatic startup selection for the built-in OpenAI and Codex providers now defaults to GPT-5.6 Sol instead of GPT-5.5. Explicitly saved GPT-5.5 selections remain supported.
+
 ### Removed
 
 ## [2026.8.22] - 2026-08-22

packages/coding-agent/src/core/changes.md

@@ -1,5 +1,23 @@
 # changes
 
+## 2026-08-22 - Retarget OpenAI automatic defaults to GPT-5.6 Sol
+
+### What changed
+
+- `packages/coding-agent/src/core/model-resolver.ts`: retargeted the `openai` and `openai-codex` provider defaults from `gpt-5.5` to `gpt-5.6-sol` while retaining GPT-5.5 in catalogs and explicit settings resolution.
+
+### Why
+
+- Automatic startup recommendation should follow the current recommended GPT-5.6 Sol model; saved GPT-5.5 selections remain explicitly selectable.
+
+### Why an extension could not handle it
+
+- `defaultModelPerProvider` is consumed by core initial-model resolution before extension recommendations are applied.
+
+### Expected merge conflict zones
+
+- LOW: the OpenAI provider entries in `packages/coding-agent/src/core/model-resolver.ts`.
+
 ## 2026-08-22 - emit agent_idle after settlement-deferred turns resolve
 
 ### What changed

packages/coding-agent/src/core/model-resolver.ts

@@ -39,9 +39,9 @@ export const defaultModelPerProvider: Record<KnownProvider, string> = {
 	"amazon-bedrock": "us.anthropic.claude-opus-4-6-v1",
 	"ant-ling": "Ring-2.6-1T",
 	anthropic: "claude-opus-4-8",
-	openai: "gpt-5.5",
+	openai: "gpt-5.6-sol",
 	"azure-openai-responses": "gpt-5.4",
-	"openai-codex": "gpt-5.5",
+	"openai-codex": "gpt-5.6-sol",
 	ollama: "qwen3.5:397b",
 	// Cursor ships no models until its chat protocol is ported; "auto" matches
 	// the Cursor agent's native model auto-selection once models exist.

packages/coding-agent/test/model-resolver.test.ts

@@ -697,8 +697,8 @@ describe("resolveCliModel", () => {
 
 describe("default model selection", () => {
 	test("openai defaults track current models", () => {
-		expect(defaultModelPerProvider.openai).toBe("gpt-5.5");
-		expect(defaultModelPerProvider["openai-codex"]).toBe("gpt-5.5");
+		expect(defaultModelPerProvider.openai).toBe("gpt-5.6-sol");
+		expect(defaultModelPerProvider["openai-codex"]).toBe("gpt-5.6-sol");
 	});
 
 	test("zai, minimax, cerebras, and ant-ling defaults track current models", () => {
@@ -931,8 +931,8 @@ describe("default model selection", () => {
 		};
 		const custom: Model<"anthropic-messages"> = {
 			...openAiDefault,
-			id: "custom-model",
-			provider: "custom",
+			id: "grok-4.5",
+			provider: "xai",
 		};
 		const runtime = {
 			getModels: () => [openAiDefault, custom],
@@ -957,8 +957,8 @@ describe("default model selection", () => {
 		const settings = await findInitialModel({
 			scopedModels: [],
 			isContinuing: false,
-			defaultProvider: "custom",
-			defaultModelId: "custom-model",
+			defaultProvider: "xai",
+			defaultModelId: "grok-4.5",
 			modelRuntime: runtime,
 		});
 		const providerDefault = await findInitialModel({
@@ -968,7 +968,7 @@ describe("default model selection", () => {
 		});
 		const firstAvailableRuntime = {
 			...runtime,
-			getAvailable: async () => [custom],
+			getAvailable: async () => [openAiDefault],
 		} as unknown as Parameters<typeof findInitialModel>[0]["modelRuntime"];
 		const firstAvailable = await findInitialModel({
 			scopedModels: [],

packages/coding-agent/test/provider-default-model-selection.test.ts

@@ -0,0 +1,35 @@
+import { getModels } from "@earendil-works/pi-ai/compat";
+import { describe, expect, test } from "vitest";
+import { findInitialModel } from "../src/core/model-resolver.ts";
+
+describe("OpenAI provider defaults", () => {
+	test("prefers GPT-5.6 Sol automatically while preserving explicit GPT-5.5", async () => {
+		const openAiModels = getModels("openai");
+		const codexModels = getModels("openai-codex");
+		const availableModels = [
+			openAiModels.find((model) => model.id === "gpt-5.6-sol"),
+			codexModels.find((model) => model.id === "gpt-5.6-sol"),
+			codexModels.find((model) => model.id === "gpt-5.5"),
+		].filter((model) => model !== undefined);
+		const runtime = {
+			getAvailableSnapshot: () => availableModels,
+			getModel: (provider: string, modelId: string) =>
+				availableModels.find((model) => model.provider === provider && model.id === modelId),
+			hasConfiguredAuth: () => true,
+		} as unknown as Parameters<typeof findInitialModel>[0]["modelRuntime"];
+
+		const automatic = await findInitialModel({ scopedModels: [], isContinuing: false, modelRuntime: runtime });
+		const explicit = await findInitialModel({
+			scopedModels: [],
+			isContinuing: false,
+			defaultProvider: "openai-codex",
+			defaultModelId: "gpt-5.5",
+			modelRuntime: runtime,
+		});
+
+		expect(automatic.model?.id).toBe("gpt-5.6-sol");
+		expect(automatic.provenance).toBe("provider-default");
+		expect(explicit.model?.id).toBe("gpt-5.5");
+		expect(explicit.provenance).toBe("settings");
+	});
+});

packages/coding-agent/test/suite/interactive-mode-scoped-settings.test.ts

@@ -117,7 +117,7 @@ describe("InteractiveMode scoped-setting caller compatibility", () => {
 
 	it("keeps post-auth default model selection on the global-setting setter", async () => {
 		// Given: authentication completes while the session still has the unknown placeholder model.
-		const defaultModel = { provider: "openai", id: "gpt-5.5" };
+		const defaultModel = { provider: "openai", id: "gpt-5.6-sol" };
 		const setModel = vi.fn(async () => undefined);
 		const setSessionModel = vi.fn(async () => undefined);
 		const fakeThis = {