fix(opencode-go): route GPT-5.6+ models through Responses API

packages/types/src/__tests__/opencode-go.test.ts

@@ -5,6 +5,7 @@ import {
 	OPENCODE_GO_DEFAULT_TEMPERATURE,
 	OPENCODE_GO_ANTHROPIC_FORMAT_MODELS,
 	OPENCODE_GO_RESPONSES_FORMAT_MODELS,
+	OPENCODE_GO_RESPONSES_FORMAT_REGEX,
 	isOpencodeGoAnthropicFormatModel,
 	isOpencodeGoResponsesFormatModel,
 	getOpencodeGoModelInfo,
@@ -37,12 +38,7 @@ describe("opencode-go registry", () => {
 		"omen-alpha",
 		"grok-4.5",
 	]
-	const responsesFormatModels = [
-		"gpt-5.6-luna",
-		"grok-4.6",
-		"muse-spark-1.3-contributor",
-		"muse-spark-1.2-contributor",
-	]
+	const responsesFormatModels = ["grok-4.6", "muse-spark-1.3-contributor", "muse-spark-1.2-contributor"]
 
 	describe("isOpencodeGoAnthropicFormatModel", () => {
 		it("classifies Qwen and MiniMax models as Anthropic-format", () => {
@@ -142,6 +138,23 @@ describe("opencode-go registry", () => {
 	})
 
 	describe("OPENCODE_GO_RESPONSES_FORMAT_MODELS", () => {
+		it("classifies later numeric GPT models through the Responses regex", () => {
+			expect(OPENCODE_GO_RESPONSES_FORMAT_REGEX.some((regex) => regex.test("gpt-5.6-luna"))).toBe(true)
+			expect(OPENCODE_GO_RESPONSES_FORMAT_REGEX.some((regex) => regex.test("gpt-6-luna"))).toBe(true)
+			expect(isOpencodeGoResponsesFormatModel("gpt-6-luna")).toBe(true)
+			expect(isOpencodeGoResponsesFormatModel("gpt-7-luna")).toBe(true)
+			expect(isOpencodeGoResponsesFormatModel("gpt-8")).toBe(true)
+			expect(isOpencodeGoResponsesFormatModel("gpt-6.1-mini")).toBe(true)
+			expect(isOpencodeGoResponsesFormatModel("gpt-5.10-foo")).toBe(true)
+			expect(isOpencodeGoResponsesFormatModel("gpt-10-foo")).toBe(true)
+			expect(isOpencodeGoResponsesFormatModel("GPT-6-Luna")).toBe(true)
+			expect(isOpencodeGoResponsesFormatModel("gpt-5.5-pro")).toBe(false)
+			expect(isOpencodeGoResponsesFormatModel("gpt-5")).toBe(false)
+			expect(isOpencodeGoResponsesFormatModel("gpt-4.1")).toBe(false)
+			expect(isOpencodeGoResponsesFormatModel("gpt-6o")).toBe(false)
+			expect(isOpencodeGoResponsesFormatModel("gpt-oss-20b")).toBe(false)
+		})
+
 		it("contains exactly the Responses-only models", () => {
 			expect([...OPENCODE_GO_RESPONSES_FORMAT_MODELS].sort()).toEqual([...responsesFormatModels].sort())
 		})
@@ -189,6 +202,20 @@ describe("opencode-go registry", () => {
 			})
 		})
 
+		it("curates gpt-6-luna with its Go pricing and capabilities", () => {
+			expect(getOpencodeGoModelInfo("gpt-6-luna")).toMatchObject({
+				maxTokens: 128_000,
+				supportsMaxTokens: true,
+				contextWindow: 1_050_000,
+				supportsImages: true,
+				supportsPromptCache: true,
+				inputPrice: 0.1,
+				outputPrice: 0.5,
+				cacheWritesPrice: 0.13,
+				cacheReadsPrice: 0.01,
+			})
+		})
+
 		it("is disjoint from the Anthropic-format set", () => {
 			for (const id of OPENCODE_GO_RESPONSES_FORMAT_MODELS) {
 				expect(OPENCODE_GO_ANTHROPIC_FORMAT_MODELS.has(id)).toBe(false)

packages/types/src/providers/opencode-go.ts

@@ -658,6 +658,27 @@
 		description:
 			"Muse Spark 1.2 Contributor is Meta's multimodal coding model with a 1M context window. Available via the Opencode Go plan.",
 	},
+	"gpt-6-luna": {
+		maxTokens: 128_000,
+		contextWindow: 1_050_000,
+		supportsImages: true,
+		supportsPromptCache: true,
+		supportsMaxTokens: true,
+		supportsReasoningEffort: ["none", "low", "medium", "high", "xhigh", "max"],
+		reasoningEffort: "medium",
+		inputPrice: 0.1,
+		outputPrice: 0.5,
+		cacheWritesPrice: 0.13,
+		cacheReadsPrice: 0.01,
+		longContextPricing: {
+			thresholdTokens: 272_000,
+			inputPriceMultiplier: 2,
+			outputPriceMultiplier: 1.5,
+			cacheWritesPriceMultiplier: 2,
+			cacheReadsPriceMultiplier: 2,
+		},
+		description: "GPT-6 Luna via the OpenCode Go Responses API.",
+	},
 }
 
 /**
@@ -694,27 +715,21 @@
  * (`/v1/responses`), not the OpenAI-compatible Chat Completions endpoint
  * (`/v1/chat/completions`).
  *
- * The Go gateway maps every model to exactly one wire format. Responses-only
- * models are explicitly curated in `opencodeGoModels`: the gateway's
- * `/v1/chat/completions` adapter for these models can fail with an opaque HTTP 500
- * (`{"type":"error","error":{"type":"error","message":"Internal server error"}}`),
- * while `/v1/responses` succeeds (Zoo-Code-Org/Zoo-Code#1431).
- *
- * Drive routing from this set rather than from the model ID string so the
- * gateway's protocol contract stays explicit, testable, and easy to extend
- * when the next Responses-only model lands. Unknown model IDs default to the
- * OpenAI-compatible chat completions format.
+ * The Go gateway maps known non-GPT Responses models explicitly, while GPT-5.6
+ * and later numeric GPT generations are routed by pattern. The separate `gpt-oss`
+ * family remains on Chat Completions.
  */
 export const OPENCODE_GO_RESPONSES_FORMAT_MODELS = new Set<string>([
-	// --- OpenAI ---
-	"gpt-5.6-luna",
-	// --- xAI ---
 	"grok-4.6",
-	// --- Meta ---
 	"muse-spark-1.3-contributor",
 	"muse-spark-1.2-contributor",
 ])
 
+export const OPENCODE_GO_RESPONSES_FORMAT_REGEX: RegExp[] = [
+	// gpt-5.6 and above are routed by pattern for automatic discovery
+	/^gpt-(?:5\.(?:[6-9]|\d{2,})|[6-9]\d*(?:[.-]|$)|\d{2,}(?:[.-]|$))/i,
+]
+
 /**
  * Returns `true` when the given Go-plan model ID must be requested via the
  * Anthropic Messages format (`/v1/messages`) rather than the OpenAI-compatible
@@ -732,7 +747,10 @@
  * format, matching the gateway's default routing.
  */
 export function isOpencodeGoResponsesFormatModel(modelId: string): boolean {
-	return OPENCODE_GO_RESPONSES_FORMAT_MODELS.has(modelId)
+	return (
+		OPENCODE_GO_RESPONSES_FORMAT_MODELS.has(modelId) ||
+		OPENCODE_GO_RESPONSES_FORMAT_REGEX.some((regex) => regex.test(modelId))
+	)
 }
 
 /**

src/api/providers/__tests__/opencode-go.spec.ts

@@ -40,15 +40,17 @@ vitest.mock("../fetchers/modelCache", () => ({
 			"glm-5.1": { ...opencodeGoModels["glm-5.1"] },
 			// Anthropic-format model used to exercise the /v1/messages path.
 			"qwen3.7-max": { ...opencodeGoModels["qwen3.7-max"] },
-			// Responses-format model (Zoo-Code-Org/Zoo-Code#1431).
+			// Responses-format models (Zoo-Code-Org/Zoo-Code#1431 and #1979).
 			"gpt-5.6-luna": { ...opencodeGoModels["gpt-5.6-luna"] },
+			"gpt-6-luna": { ...opencodeGoModels["gpt-6-luna"] },
 		})
 	}),
 	refreshModels: vitest.fn().mockImplementation(function () {
 		return Promise.resolve({
 			"glm-5.1": { ...opencodeGoModels["glm-5.1"] },
 			"qwen3.7-max": { ...opencodeGoModels["qwen3.7-max"] },
 			"gpt-5.6-luna": { ...opencodeGoModels["gpt-5.6-luna"] },
+			"gpt-6-luna": { ...opencodeGoModels["gpt-6-luna"] },
 		})
 	}),
 	getModelsFromCache: vitest.fn().mockReturnValue(undefined),
@@ -1362,8 +1364,11 @@ describe("OpencodeGoHandler", () => {
 			}).rejects.toThrow("Opencode Go completion error: internal server error")
 		})
 
-		it("classifies documented Responses models as Responses-format and other models as not", () => {
+		it("classifies documented and numeric Responses models, excluding gpt-oss", () => {
 			expect(isOpencodeGoResponsesFormatModel("gpt-5.6-luna")).toBe(true)
+			expect(isOpencodeGoResponsesFormatModel("gpt-6-luna")).toBe(true)
+			expect(isOpencodeGoResponsesFormatModel("gpt-5.4-nano")).toBe(false)
+			expect(isOpencodeGoResponsesFormatModel("gpt-oss-20b")).toBe(false)
 			expect(isOpencodeGoResponsesFormatModel("grok-4.5")).toBe(false)
 			expect(isOpencodeGoResponsesFormatModel("grok-4.6")).toBe(true)
 			expect(isOpencodeGoResponsesFormatModel("muse-spark-1.3-contributor")).toBe(true)

src/api/providers/fetchers/__tests__/opencode-go.spec.ts

@@ -75,6 +75,22 @@ describe("Opencode Go Fetchers", () => {
 			})
 		})
 
+		it("uses Responses defaults for an uncurated numeric gpt model", async () => {
+			mockedAxios.get.mockResolvedValue({ data: { data: [{ id: "gpt-7-luna" }] } })
+
+			const models = await getOpencodeGoModels("k")
+
+			expect(models["gpt-7-luna"]).toMatchObject({
+				contextWindow: 1_050_000,
+				maxTokens: 128_000,
+				supportsMaxTokens: true,
+				supportsImages: true,
+				supportsPromptCache: true,
+				supportsReasoningEffort: ["none", "low", "medium", "high", "xhigh", "max"],
+				reasoningEffort: "medium",
+			})
+		})
+
 		it("falls back to default context/max tokens for an unknown model when metadata is absent", async () => {
 			mockedAxios.get.mockResolvedValue({ data: { data: [{ id: "some-unknown-model" }] } })
 
@@ -213,6 +229,7 @@ describe("Opencode Go Fetchers", () => {
 				"hy3",
 				"hy3-preview",
 				"gpt-5.6-luna",
+				"gpt-6-luna",
 				"grok-4.5",
 				"grok-4.6",
 				"muse-spark-1.3-contributor",
@@ -298,6 +315,40 @@ describe("Opencode Go Fetchers", () => {
 			expect(info.cacheReadsPrice).toBe(0.26)
 		})
 
+		it("uses Responses defaults when parsing an uncurated numeric gpt model", () => {
+			const info = parseOpencodeGoModel({ id: "gpt-7-foo" })
+			expect(info).toMatchObject({
+				contextWindow: 1_050_000,
+				maxTokens: 128_000,
+				supportsMaxTokens: true,
+				supportsImages: true,
+				supportsPromptCache: true,
+				supportsReasoningEffort: ["none", "low", "medium", "high", "xhigh", "max"],
+				reasoningEffort: "medium",
+			})
+		})
+
+		it("overrides Responses defaults with live metadata for an uncurated numeric gpt model", () => {
+			const info = parseOpencodeGoModel({
+				id: "gpt-7-foo",
+				context_length: 2_000_000,
+				max_output_tokens: 64_000,
+				supports_images: false,
+				description: "Live GPT model description",
+			})
+
+			expect(info).toMatchObject({
+				contextWindow: 2_000_000,
+				maxTokens: 64_000,
+				supportsImages: false,
+				description: "Live GPT model description",
+				supportsMaxTokens: true,
+				supportsPromptCache: true,
+				supportsReasoningEffort: ["none", "low", "medium", "high", "xhigh", "max"],
+				reasoningEffort: "medium",
+			})
+		})
+
 		it("falls back to defaults for an unknown model with no cache pricing", () => {
 			const info = parseOpencodeGoModel({ id: "x", context_window: 100000, max_tokens: 8000 })
 			expect(info.supportsPromptCache).toBe(false)

src/api/providers/fetchers/opencode-go.ts

@@ -2,7 +2,7 @@
 import { z } from "zod"
 
 import type { ModelInfo } from "@roo-code/types"
-import { opencodeGoDefaultModelInfo, getOpencodeGoModelInfo } from "@roo-code/types"
+import { getOpencodeGoModelInfo, isOpencodeGoResponsesFormatModel, opencodeGoDefaultModelInfo } from "@roo-code/types"
 
 import { throwIfAborted } from "../utils/abort-signal"
 
@@ -32,6 +32,19 @@
 	data: z.array(opencodeGoModelSchema),
 })
 
+// Capability defaults for uncurated Responses-format models. This lets newly
+// discovered GPT models route and expose their output-token control without
+// inventing model-specific pricing; prices remain available only for curated IDs.
+const opencodeGoResponsesModelDefaults: ModelInfo = {
+	maxTokens: 128_000,
+	contextWindow: 1_050_000,
+	supportsImages: true,
+	supportsPromptCache: true,
+	supportsMaxTokens: true,
+	supportsReasoningEffort: ["none", "low", "medium", "high", "xhigh", "max"],
+	reasoningEffort: "medium",
+}
+
 /**
  * Maps a raw Opencode Go model entry to the internal {@link ModelInfo} shape.
  *
@@ -47,8 +60,8 @@
  *      is curated, including its capabilities and pricing.
  *   2. Override static limits and image support with live `/models` values when
  *      present, keeping the gateway authoritative for volatile fields.
- *   3. Fall back to {@link opencodeGoDefaultModelInfo} for an unknown model,
- *      ensuring downstream consumers always receive a fully-populated object.
+ *   3. Use Responses-specific capability defaults for uncurated Responses
+ *      models; otherwise use {@link opencodeGoDefaultModelInfo} for unknowns.
  *
  * @param model - Validated model entry from the `/models` response.
  * @returns Normalised model metadata suitable for the model picker.
@@ -71,6 +84,16 @@
 		}
 	}
 
+	if (isOpencodeGoResponsesFormatModel(model.id)) {
+		return {
+			...opencodeGoResponsesModelDefaults,
+			...(liveContextWindow !== undefined && { contextWindow: liveContextWindow }),
+			...(liveMaxTokens !== undefined && { maxTokens: liveMaxTokens }),
+			...(liveSupportsImages !== undefined && { supportsImages: liveSupportsImages }),
+			description: model.description ?? model.name,
+		}
+	}
+
 	return {
 		maxTokens: liveMaxTokens ?? opencodeGoDefaultModelInfo.maxTokens,
 		contextWindow: liveContextWindow ?? opencodeGoDefaultModelInfo.contextWindow,

src/api/providers/opencode-go.ts

@@ -71,9 +71,9 @@ type OpencodeGoFormat = "anthropic" | "openai" | "responses"
  *   - Anthropic Messages (`/v1/messages`) — used by Qwen (qwen3.8-max,
  *     qwen3.7-max, qwen3.7-plus, qwen3.6-plus) and MiniMax (minimax-m3,
  *     minimax-m2.7, minimax-m2.5) models.
- *   - OpenAI Responses (`/v1/responses`) — used by gpt-5.6-luna, whose
- *     chat-completions adapter fails with an opaque HTTP 500
- *     (Zoo-Code-Org/Zoo-Code#1431).
+ *   - OpenAI Responses (`/v1/responses`) — used by numeric GPT models from
+ *     GPT-5.6 onward, such as `gpt-5.6-luna` and `gpt-6-luna`. Earlier numeric
+ *     GPT models and the separate `gpt-oss` family use chat completions.
  *
  * Sending an Anthropic-format model to the chat completions endpoint is
  * rejected with `401 Model <id> is not supported for format oa-compat`, so this
@@ -182,10 +182,10 @@ export class OpencodeGoHandler extends RouterProvider implements SingleCompletio
 	 * reasoning, partial tool calls, and token usage.
 	 *
 	 * Anthropic-format models (Qwen/MiniMax) are streamed via
-	 * {@link streamAnthropicMessage} against `/v1/messages`; Responses-format
-	 * models (gpt-5.6-luna) are streamed via {@link streamResponsesMessage}
-	 * against `/v1/responses`; all other models use the OpenAI-compatible chat
-	 * completions endpoint.
+	 * {@link streamAnthropicMessage} against `/v1/messages`; numeric GPT models
+	 * from GPT-5.6 onward, excluding `gpt-oss`, use
+	 * {@link streamResponsesMessage} against `/v1/responses`; all other models use
+	 * the OpenAI-compatible chat completions endpoint.
 	 *
 	 * For OpenAI-format models that require reasoning_content to be passed back
 	 * during multi-turn tool calls (`preserveReasoning`), messages are
@@ -292,7 +292,8 @@ export class OpencodeGoHandler extends RouterProvider implements SingleCompletio
 
 	/**
 	 * Streams an OpenAI Responses-format completion for Go models that only
-	 * accept the `/v1/responses` endpoint (currently gpt-5.6-luna).
+	 * accept the `/v1/responses` endpoint (numeric GPT model IDs from GPT-5.6
+	 * onward, excluding `gpt-oss`).
 	 *
 	 * Follows the focused xAI handler pattern: the conversation is converted
 	 * with the shared {@link convertToResponsesApiInput} transform, the system
@@ -690,8 +691,9 @@ export class OpencodeGoHandler extends RouterProvider implements SingleCompletio
 	 * Performs a non-streaming chat completion and returns the full response text.
 	 *
 	 * Anthropic-format models are completed via the `/v1/messages` endpoint;
-	 * Responses-format models via `/v1/responses`; all other
-	 * models use the OpenAI-compatible chat completions endpoint.
+	 * numeric GPT models from GPT-5.6 onward, excluding `gpt-oss`, use
+	 * `/v1/responses`; all other models use the OpenAI-compatible chat completions
+	 * endpoint.
 	 *
 	 * @param prompt - The user prompt to send as a single user message.
 	 * @returns The model's reply text, or an empty string if no content is returned.