feat(opencode-go): native model params, anthropic-format routing, and context-token fix (#652)

* opencode go

* Update default to GLM 5.2

* Use Andthropic messages for minmax and qwen

* fix context calculation with minmax

* Update coverage

* add opencode test

* fix typescript error

* fix(opencode-go): address PR #652 review feedback

- Type streamAnthropicMessage's info param as ModelInfo and drop the
  force-cast so calculateApiCostAnthropic can no longer silently return
  /bin/sh when pricing fields are absent.
- Wrap pre-stream Anthropic-format errors (401/429/network) with the
  'Opencode Go completion error:' prefix for consistency with completePrompt.
- Clarify the registry doc: supportsPromptCache controls client-side
  cache_control injection (Anthropic path) only; OA-compat models price
  server-side cached_tokens via cacheReadsPrice regardless of the flag
  (MiMo stays false, matching the dedicated mimo provider).
- Add cacheWritesPrice (0.375) to minimax-m3 so its cache writes are
  billed, matching M2.5/M2.7.
- Add supportsMaxTokens to DeepSeek V4 models so the max-output slider
  is exposed like GLM.
- Strengthen the streaming cost test to assert totalCost > 0, and add
  registry invariants for MiniMax cache-write pricing and DeepSeek
  supportsMaxTokens plus a streaming error-wrapping test.

Author: navedmerchant

.changeset/add-glm-5-2-support.md

@@ -3,3 +3,5 @@
 ---
 
 Add GLM-5.2 support with High/Max `reasoning_effort` tiers. The default effort is High (deep reasoning stays opt-in), Max is selected only when the user explicitly picks it, and the parameter is omitted entirely when reasoning is disabled.
+
+Also refines the Opencode Go provider per review: bill MiniMax M3 cache writes (`cacheWritesPrice`), expose the max-output slider for DeepSeek V4 models (`supportsMaxTokens`), wrap pre-stream Anthropic-format errors with the provider prefix, and type the Anthropic streaming path's model info as `ModelInfo` so cost calculation can no longer silently return `$0`.

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

@@ -0,0 +1,144 @@
+import {
+	opencodeGoDefaultModelId,
+	opencodeGoDefaultModelInfo,
+	opencodeGoModels,
+	OPENCODE_GO_DEFAULT_TEMPERATURE,
+	OPENCODE_GO_ANTHROPIC_FORMAT_MODELS,
+	isOpencodeGoAnthropicFormatModel,
+	getOpencodeGoModelInfo,
+} from "../providers/opencode-go.js"
+
+describe("opencode-go registry", () => {
+	const anthropicFormatModels = [
+		"qwen3.7-max",
+		"qwen3.7-plus",
+		"qwen3.6-plus",
+		"minimax-m3",
+		"minimax-m2.7",
+		"minimax-m2.5",
+	]
+	const openaiFormatModels = [
+		"glm-5",
+		"glm-5.1",
+		"glm-5.2",
+		"kimi-k2.5",
+		"kimi-k2.6",
+		"mimo-v2.5",
+		"mimo-v2.5-pro",
+		"deepseek-v4-pro",
+		"deepseek-v4-flash",
+	]
+
+	describe("isOpencodeGoAnthropicFormatModel", () => {
+		it("classifies Qwen and MiniMax models as Anthropic-format", () => {
+			for (const id of anthropicFormatModels) {
+				expect(isOpencodeGoAnthropicFormatModel(id)).toBe(true)
+			}
+		})
+
+		it("classifies GLM/Kimi/MiMo/DeepSeek models as OpenAI-compatible", () => {
+			for (const id of openaiFormatModels) {
+				expect(isOpencodeGoAnthropicFormatModel(id)).toBe(false)
+			}
+		})
+
+		it("defaults unknown model IDs to the OpenAI-compatible format", () => {
+			expect(isOpencodeGoAnthropicFormatModel("some-future-model")).toBe(false)
+			expect(isOpencodeGoAnthropicFormatModel("")).toBe(false)
+		})
+	})
+
+	describe("getOpencodeGoModelInfo", () => {
+		it("returns the native ModelInfo for a curated model", () => {
+			const info = getOpencodeGoModelInfo("qwen3.7-max")
+			expect(info).toBeDefined()
+			expect(info?.maxTokens).toBe(65_536)
+			expect(info?.contextWindow).toBe(1_000_000)
+			expect(info?.supportsPromptCache).toBe(true)
+		})
+
+		it("returns undefined for an unknown model ID", () => {
+			expect(getOpencodeGoModelInfo("not-a-real-go-model")).toBeUndefined()
+		})
+	})
+
+	describe("OPENCODE_GO_ANTHROPIC_FORMAT_MODELS", () => {
+		it("contains exactly the Qwen and MiniMax models", () => {
+			expect([...OPENCODE_GO_ANTHROPIC_FORMAT_MODELS].sort()).toEqual([...anthropicFormatModels].sort())
+		})
+
+		// The PR description calls out that the format-routing set must stay in
+		// sync with the Go model table — every routed model must have a native
+		// registry entry so capability flags and pricing resolve correctly.
+		it("every Anthropic-format model has a native registry entry", () => {
+			for (const id of OPENCODE_GO_ANTHROPIC_FORMAT_MODELS) {
+				expect(opencodeGoModels[id]).toBeDefined()
+			}
+		})
+	})
+
+	describe("opencodeGoModels registry invariants", () => {
+		it("every entry has a positive maxTokens and contextWindow", () => {
+			for (const [id, info] of Object.entries(opencodeGoModels)) {
+				expect(info.maxTokens).toBeGreaterThan(0)
+				expect(info.contextWindow).toBeGreaterThan(0)
+				// Sanity: max output must not exceed the context window.
+				expect(info.maxTokens).toBeLessThanOrEqual(info.contextWindow)
+				void id
+			}
+		})
+
+		it("every entry declares supportsImages", () => {
+			for (const info of Object.values(opencodeGoModels)) {
+				expect(typeof info.supportsImages).toBe("boolean")
+			}
+		})
+
+		it("models with an array supportsReasoningEffort expose a non-empty allow-list", () => {
+			for (const info of Object.values(opencodeGoModels)) {
+				if (Array.isArray(info.supportsReasoningEffort)) {
+					expect(info.supportsReasoningEffort.length).toBeGreaterThan(0)
+				}
+			}
+		})
+
+		it("every Anthropic-format model with prompt-cache injection declares a cacheWritesPrice", () => {
+			// MiniMax/Qwen route through /v1/messages with client-side
+			// cache_control breakpoints, so cache_creation_input_tokens are
+			// reported and billed — each must carry a cacheWritesPrice or the
+			// write cost is silently reported as $0.
+			for (const id of OPENCODE_GO_ANTHROPIC_FORMAT_MODELS) {
+				const info = getOpencodeGoModelInfo(id)
+				expect(info).toBeDefined()
+				if (info?.supportsPromptCache) {
+					expect(info.cacheWritesPrice).toBeDefined()
+					expect(info.cacheReadsPrice).toBeDefined()
+				}
+			}
+		})
+
+		it("DeepSeek entries expose supportsMaxTokens so the max-output slider is available", () => {
+			expect(getOpencodeGoModelInfo("deepseek-v4-pro")?.supportsMaxTokens).toBe(true)
+			expect(getOpencodeGoModelInfo("deepseek-v4-flash")?.supportsMaxTokens).toBe(true)
+		})
+	})
+
+	describe("defaults", () => {
+		it("the default model id is a curated OpenAI-compatible model", () => {
+			expect(opencodeGoDefaultModelId).toBe("glm-5.2")
+			expect(opencodeGoModels[opencodeGoDefaultModelId]).toBeDefined()
+			expect(isOpencodeGoAnthropicFormatModel(opencodeGoDefaultModelId)).toBe(false)
+		})
+
+		it("exposes a fully-populated default ModelInfo fallback", () => {
+			expect(opencodeGoDefaultModelInfo.maxTokens).toBeGreaterThan(0)
+			expect(opencodeGoDefaultModelInfo.contextWindow).toBeGreaterThan(0)
+			expect(opencodeGoDefaultModelInfo.supportsPromptCache).toBe(false)
+			expect(opencodeGoDefaultModelInfo.description).toBeTruthy()
+		})
+
+		it("exposes a deterministic default temperature", () => {
+			expect(OPENCODE_GO_DEFAULT_TEMPERATURE).toBe(0)
+		})
+	})
+})

packages/types/src/__tests__/provider-settings.test.ts

@@ -53,6 +53,32 @@ describe("getApiProtocol", () => {
 		})
 	})
 
+	describe("Opencode Go provider", () => {
+		it("should return 'anthropic' for opencode-go Anthropic-format models (Qwen/MiniMax)", () => {
+			expect(getApiProtocol("opencode-go", "qwen3.7-max")).toBe("anthropic")
+			expect(getApiProtocol("opencode-go", "qwen3.7-plus")).toBe("anthropic")
+			expect(getApiProtocol("opencode-go", "qwen3.6-plus")).toBe("anthropic")
+			expect(getApiProtocol("opencode-go", "minimax-m3")).toBe("anthropic")
+			expect(getApiProtocol("opencode-go", "minimax-m2.7")).toBe("anthropic")
+			expect(getApiProtocol("opencode-go", "minimax-m2.5")).toBe("anthropic")
+		})
+
+		it("should return 'openai' for opencode-go OpenAI-format models (GLM/DeepSeek/etc.)", () => {
+			expect(getApiProtocol("opencode-go", "glm-5.2")).toBe("openai")
+			expect(getApiProtocol("opencode-go", "deepseek-v4-pro")).toBe("openai")
+			expect(getApiProtocol("opencode-go", "kimi-k2.5")).toBe("openai")
+			expect(getApiProtocol("opencode-go", "mimo-v2.5")).toBe("openai")
+		})
+
+		it("should return 'openai' for opencode-go without a model", () => {
+			expect(getApiProtocol("opencode-go")).toBe("openai")
+		})
+
+		it("should return 'openai' for opencode-go with an unknown model id", () => {
+			expect(getApiProtocol("opencode-go", "some-future-model")).toBe("openai")
+		})
+	})
+
 	describe("Other providers", () => {
 		it("should return 'openai' for non-anthropic providers regardless of model", () => {
 			expect(getApiProtocol("openrouter", "claude-3-opus")).toBe("openai")

packages/types/src/provider-settings.ts

@@ -21,6 +21,7 @@ import {
 	internationalZAiModels,
 	minimaxModels,
 	mimoModels,
+	isOpencodeGoAnthropicFormatModel,
 } from "./providers/index.js"
 
 /**
@@ -595,6 +596,17 @@ export const getApiProtocol = (provider: ProviderName | undefined, modelId?: str
 		return "anthropic"
 	}
 
+	// Opencode Go routes a subset of its models (Qwen, MiniMax) through the
+	// Anthropic Messages wire format (`/v1/messages`), which reports usage in
+	// Anthropic style: `input_tokens` excludes cache tokens, with separate
+	// `cache_creation_input_tokens` / `cache_read_input_tokens` fields. These
+	// models must use the anthropic protocol so token/cost aggregation adds the
+	// cache tokens back into the input total — otherwise the cached prefix is
+	// dropped from `contextTokens`, undercounting context-window usage.
+	if (provider && provider === "opencode-go" && modelId && isOpencodeGoAnthropicFormatModel(modelId)) {
+		return "anthropic"
+	}
+
 	return "openai"
 }