fix(deepseek): force tool_choice=required for thinking models + e2e regression suite

apps/vscode-e2e/.env.local.sample

@@ -1 +1,2 @@
 OPENROUTER_API_KEY=sk-or-v1-...
+DEEPSEEK_API_KEY=sk-...

apps/vscode-e2e/AGENTS.md

@@ -82,10 +82,42 @@ Record mode uses **record-on-miss**: if an existing fixture already matches a re
 
 If the LLM calls a tool first (e.g. `read_file`) and then calls `attempt_completion` after seeing the result, you need two fixtures:
 
-- **Turn 1**: match on the task prompt → respond with the tool call
-- **Turn 2**: match on a stable part of the tool _result_ → respond with `attempt_completion`
+- **Turn 1**: match on the task prompt (with `sequenceIndex: 0` so it fires only once) → respond with the tool call, giving the tool call a unique `id`
+- **Turn 2**: match on `toolCallId` → respond with `attempt_completion`
+
+Using `toolCallId` (the `id` of the tool call emitted in turn 1) is the recommended approach for turn-2 matching. It is:
+
+- **Precise**: fires only when that exact tool call's result is in the conversation
+- **Cross-test safe**: each test's tool call ids are unique, so accumulated match counts from previous tests can't interfere
+- **Stateless**: no `sequenceIndex` needed on turn-2 fixtures — if the task makes extra API calls they'll keep getting the same `attempt_completion`
+
+Example:
+
+```json
+{
+	"fixtures": [
+		{
+			"match": {
+				"userMessage": "my-e2e-tag:my-test",
+				"sequenceIndex": 0
+			},
+			"response": {
+				"toolCalls": [{ "name": "read_file", "arguments": "{\"path\":\"marker.txt\"}", "id": "call_my_read" }]
+			}
+		},
+		{
+			"match": { "toolCallId": "call_my_read" },
+			"response": {
+				"toolCalls": [
+					{ "name": "attempt_completion", "arguments": "{\"result\":\"MY_MARKER\"}", "id": "call_my_done" }
+				]
+			}
+		}
+	]
+}
+```
 
-The tool result is provided by the extension (not the mock), so its content is deterministic if test files have stable names. Use a stable substring from the tool result as the turn-2 match string.
+The `model` field can be added to either match when a test targets a specific model.
 
 ## 404 errors in logs are expected
 
@@ -118,6 +150,23 @@ ZAI_API_KEY=<key> TEST_FILE=zai.test pnpm --filter @roo-code/vscode-e2e test:ci
 ```
 
 When adding a new test to this suite, add a matching fixture to the `installZAiFetchInterceptor` call in `suiteSetup`. Use a short unique prefix (e.g. `"zai-glm-e2e-mytest:"`) that won't appear in `<environment_details>`.
+
+### DeepSeek V4 (`suite/providers/deepseek-v4.test.ts`)
+
+DeepSeek exposes `deepSeekBaseUrl`, so the suite redirects the OpenAI-compatible DeepSeek client through aimock with `deepSeekBaseUrl: ${AIMOCK_URL}/v1`. The test still installs a lightweight fetch capture for request-shape assertions, but responses should come from aimock fixtures or aimock record mode.
+
+Record DeepSeek fixtures with the targeted file filter so aimock proxies OpenAI-compatible traffic to `https://api.deepseek.com`:
+
+```sh
+DEEPSEEK_API_KEY=<key> TEST_FILE=deepseek-v4.test pnpm --filter @roo-code/vscode-e2e test:record
+```
+
+After converting the generated `openai-*.json` files into stable named fixtures, verify in mock mode:
+
+```sh
+USE_MOCK=true TEST_FILE=deepseek-v4.test pnpm --filter @roo-code/vscode-e2e test:run
+```
+
 ## Tests that use a non-default provider
 
 If your test calls `api.setConfiguration({ apiProvider: "anthropic", ... })`, point aimock at the

apps/vscode-e2e/fixtures/deepseek-v4.json

@@ -0,0 +1,128 @@
+{
+	"fixtures": [
+		{
+			"match": {
+				"model": "deepseek-v4-flash",
+				"userMessage": "deepseek-v4-e2e:deepseek-v4-flash:reasoning-on",
+				"sequenceIndex": 0
+			},
+			"response": {
+				"toolCalls": [
+					{
+						"name": "read_file",
+						"arguments": "{\"path\":\"deepseek-v4-e2e-deepseek-v4-flash-reasoning-on.txt\"}",
+						"id": "call_dsv4_flash_on_read"
+					}
+				]
+			}
+		},
+		{
+			"match": {
+				"model": "deepseek-v4-flash",
+				"toolCallId": "call_dsv4_flash_on_read"
+			},
+			"response": {
+				"toolCalls": [
+					{
+						"name": "attempt_completion",
+						"arguments": "{\"result\":\"DEEPSEEK_V4_MARKER_deepseek_v4_flash_reasoning_on\"}",
+						"id": "call_dsv4_flash_on_done"
+					}
+				]
+			}
+		},
+		{
+			"match": {
+				"model": "deepseek-v4-flash",
+				"userMessage": "deepseek-v4-e2e:deepseek-v4-flash:reasoning-off",
+				"sequenceIndex": 0
+			},
+			"response": {
+				"toolCalls": [
+					{
+						"name": "read_file",
+						"arguments": "{\"path\":\"deepseek-v4-e2e-deepseek-v4-flash-reasoning-off.txt\"}",
+						"id": "call_dsv4_flash_off_read"
+					}
+				]
+			}
+		},
+		{
+			"match": {
+				"model": "deepseek-v4-flash",
+				"toolCallId": "call_dsv4_flash_off_read"
+			},
+			"response": {
+				"toolCalls": [
+					{
+						"name": "attempt_completion",
+						"arguments": "{\"result\":\"DEEPSEEK_V4_MARKER_deepseek_v4_flash_reasoning_off\"}",
+						"id": "call_dsv4_flash_off_done"
+					}
+				]
+			}
+		},
+		{
+			"match": {
+				"model": "deepseek-v4-pro",
+				"userMessage": "deepseek-v4-e2e:deepseek-v4-pro:reasoning-on",
+				"sequenceIndex": 0
+			},
+			"response": {
+				"toolCalls": [
+					{
+						"name": "read_file",
+						"arguments": "{\"path\":\"deepseek-v4-e2e-deepseek-v4-pro-reasoning-on.txt\"}",
+						"id": "call_dsv4_pro_on_read"
+					}
+				]
+			}
+		},
+		{
+			"match": {
+				"model": "deepseek-v4-pro",
+				"toolCallId": "call_dsv4_pro_on_read"
+			},
+			"response": {
+				"toolCalls": [
+					{
+						"name": "attempt_completion",
+						"arguments": "{\"result\":\"DEEPSEEK_V4_MARKER_deepseek_v4_pro_reasoning_on\"}",
+						"id": "call_dsv4_pro_on_done"
+					}
+				]
+			}
+		},
+		{
+			"match": {
+				"model": "deepseek-v4-pro",
+				"userMessage": "deepseek-v4-e2e:deepseek-v4-pro:reasoning-off",
+				"sequenceIndex": 0
+			},
+			"response": {
+				"toolCalls": [
+					{
+						"name": "read_file",
+						"arguments": "{\"path\":\"deepseek-v4-e2e-deepseek-v4-pro-reasoning-off.txt\"}",
+						"id": "call_dsv4_pro_off_read"
+					}
+				]
+			}
+		},
+		{
+			"match": {
+				"model": "deepseek-v4-pro",
+				"toolCallId": "call_dsv4_pro_off_read"
+			},
+			"response": {
+				"toolCalls": [
+					{
+						"name": "attempt_completion",
+						"arguments": "{\"result\":\"DEEPSEEK_V4_MARKER_deepseek_v4_pro_reasoning_off\"}",
+						"id": "call_dsv4_pro_off_done"
+					}
+				]
+			}
+		}
+	]
+}

apps/vscode-e2e/package.json

@@ -5,6 +5,7 @@
 		"lint": "eslint src --ext=ts --max-warnings=0",
 		"check-types": "tsc -p tsconfig.esm.json --noEmit",
 		"format": "prettier --write src",
+		"test:deepseek-v4": "TEST_FILE=deepseek-v4.test pnpm test:ci",
 		"test:ci": "pnpm -w bundle && pnpm --filter @roo-code/vscode-webview build && pnpm test:run",
 		"test:ci:mock": "pnpm -w bundle && pnpm --filter @roo-code/vscode-webview build && USE_MOCK=true pnpm test:run",
 		"test:record": "AIMOCK_RECORD=true pnpm test:ci",

apps/vscode-e2e/src/runTest.ts

@@ -7,8 +7,15 @@ import { LLMock } from "@copilotkit/aimock"
 
 async function main() {
 	const isRecord = process.env.AIMOCK_RECORD === "true"
+	const testGrep = process.argv.find((arg, i) => process.argv[i - 1] === "--grep") || process.env.TEST_GREP
+	const testFile = process.argv.find((arg, i) => process.argv[i - 1] === "--file") || process.env.TEST_FILE
+	const isDeepSeekTest = testFile?.includes("deepseek-v4") === true
 
-	if (isRecord && !process.env.OPENROUTER_API_KEY) {
+	if (isRecord && isDeepSeekTest && !process.env.DEEPSEEK_API_KEY) {
+		throw new Error("AIMOCK_RECORD=true requires DEEPSEEK_API_KEY to record DeepSeek fixtures")
+	}
+
+	if (isRecord && !isDeepSeekTest && !process.env.OPENROUTER_API_KEY) {
 		throw new Error("AIMOCK_RECORD=true requires OPENROUTER_API_KEY to record fixtures")
 	}
 
@@ -43,7 +50,7 @@ async function main() {
 						// Use /api (not /api/v1) — aimock appends the request path (/v1/chat/completions)
 						// so including /v1 here would produce a doubled /v1/v1 upstream URL.
 						providers: {
-							openai: "https://openrouter.ai/api",
+							openai: isDeepSeekTest ? "https://api.deepseek.com" : "https://openrouter.ai/api",
 							// aimock forwards the x-api-key header from the Anthropic SDK to the real API.
 							anthropic: "https://api.anthropic.com",
 						},
@@ -84,8 +91,6 @@ async function main() {
 		// - npm run test:e2e -- --grep "write-to-file"
 		// - TEST_GREP="apply-diff" npm run test:e2e
 		// - TEST_FILE="task.test.js" npm run test:e2e
-		const testGrep = process.argv.find((arg, i) => process.argv[i - 1] === "--grep") || process.env.TEST_GREP
-		const testFile = process.argv.find((arg, i) => process.argv[i - 1] === "--file") || process.env.TEST_FILE
 
 		// Pass test filters and mock URL as environment variables to the test runner
 		const extensionTestsEnv = {

apps/vscode-e2e/src/suite/index.ts

@@ -23,7 +23,11 @@ export async function run() {
 		apiProvider: "openrouter" as const,
 		// In record mode, forward the real key so aimock can proxy it to OpenRouter.
 		// In replay mode, "mock-key" is sufficient — aimock never contacts the real API.
-		openRouterApiKey: aimockUrl && !isRecord ? "mock-key" : process.env.OPENROUTER_API_KEY!,
+		openRouterApiKey: aimockUrl
+			? isRecord
+				? (process.env.OPENROUTER_API_KEY ?? "mock-key")
+				: "mock-key"
+			: process.env.OPENROUTER_API_KEY!,
 		openRouterModelId: "openai/gpt-4.1",
 		...(aimockUrl && { openRouterBaseUrl: `${aimockUrl}/v1` }),
 	})