feat(openai-native): always send instructions with Responses API; remove developer/system role injection from input

- Always set top-level `instructions` for every Responses API call (not just when using `previous_response_id`)
- Remove embedding the system prompt as a developer role message in `input` for Responses API
- Keep SDK request builder gating: only include `temperature` when the model supports it
- Update provider tests to assert `instructions` presence and absence of developer/system role in input

feat(ui): gate TemperatureControl by model capability

- Only render the Advanced Settings temperature control when `selectedModelInfo.supportsTemperature !== false`
- Avoids surfacing a no-op UI for models that ignore temperature (e.g., o-series / certain GPT‑5)
- All related UI tests pass

Rationale:
- Responses API treats roles in `input` as free-form; the only standardized way to set system behavior is the top-level `instructions` field.
- UI should reflect model capabilities to prevent user confusion; backend already omits temperature when unsupported, now the UI matches that behavior.

Author: hannesrudolph

src/api/providers/__tests__/openai-native.spec.ts

@@ -219,12 +219,9 @@ describe("OpenAiNativeHandler", () => {
 			const body1 = (mockFetch.mock.calls[0][1] as any).body as string
 			const parsedBody = JSON.parse(body1)
 			expect(parsedBody.model).toBe("gpt-5-2025-08-07")
-			// Now using structured format with content arrays
+			expect(parsedBody.instructions).toBe("You are a helpful assistant.")
+			// Now using structured format with content arrays (no system prompt in input; it's provided via `instructions`)
 			expect(parsedBody.input).toEqual([
-				{
-					role: "developer",
-					content: [{ type: "input_text", text: "You are a helpful assistant." }],
-				},
 				{
 					role: "user",
 					content: [{ type: "input_text", text: "Hello!" }],
@@ -872,11 +869,8 @@ describe("OpenAiNativeHandler", () => {
 
 			// Verify first request sends full conversation in structured format
 			let firstCallBody = JSON.parse(mockFetch.mock.calls[0][1].body)
+			expect(firstCallBody.instructions).toBe(systemPrompt)
 			expect(firstCallBody.input).toEqual([
-				{
-					role: "developer",
-					content: [{ type: "input_text", text: systemPrompt }],
-				},
 				{
 					role: "user",
 					content: [{ type: "input_text", text: "Hello" }],
@@ -1224,11 +1218,8 @@ describe("GPT-5 streaming event coverage (additional)", () => {
 			const requestBody = JSON.parse(mockFetch.mock.calls[0][1].body)
 			expect(requestBody).toMatchObject({
 				model: "codex-mini-latest",
+				instructions: "You are a helpful coding assistant.",
 				input: [
-					{
-						role: "developer",
-						content: [{ type: "input_text", text: "You are a helpful coding assistant." }],
-					},
 					{
 						role: "user",
 						content: [{ type: "input_text", text: "Write a hello world function" }],
@@ -1321,13 +1312,10 @@ describe("GPT-5 streaming event coverage (additional)", () => {
 				chunks.push(chunk)
 			}
 
-			// Verify the request body includes full conversation in structured format
+			// Verify the request body includes full conversation in structured format (without embedding system prompt)
 			const requestBody = JSON.parse(mockFetch.mock.calls[0][1].body)
+			expect(requestBody.instructions).toBe("You are a helpful assistant.")
 			expect(requestBody.input).toEqual([
-				{
-					role: "developer",
-					content: [{ type: "input_text", text: "You are a helpful assistant." }],
-				},
 				{
 					role: "user",
 					content: [{ type: "input_text", text: "First question" }],

src/api/providers/openai-native.ts

@@ -214,9 +214,10 @@ export class OpenAiNativeHandler extends BaseProvider implements SingleCompletio
 			input: formattedInput,
 			stream: true,
 			store: metadata?.store !== false, // Default to true unless explicitly set to false
-			// Always include instructions (system prompt) when using previous_response_id
-			// This ensures the system prompt stays up-to-date even if it changes (e.g., mode switch)
-			...(requestPreviousResponseId && { instructions: systemPrompt }),
+			// Always include instructions (system prompt) for Responses API.
+			// Unlike Chat Completions, system/developer roles in input have no special semantics here.
+			// The official way to set system behavior is the top-level `instructions` field.
+			instructions: systemPrompt,
 			...(reasoningEffort && {
 				reasoning: {
 					effort: reasoningEffort,
@@ -316,11 +317,8 @@ export class OpenAiNativeHandler extends BaseProvider implements SingleCompletio
 		// This supports both text and images
 		const formattedMessages: any[] = []
 
-		// Add system prompt as developer message
-		formattedMessages.push({
-			role: "developer",
-			content: [{ type: "input_text", text: systemPrompt }],
-		})
+		// Do NOT embed the system prompt as a developer message in the Responses API input.
+		// The Responses API treats roles as free-form; use the top-level `instructions` field instead.
 
 		// Process each message
 		for (const message of messages) {

src/core/task/Task.ts

@@ -579,6 +579,18 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 
 	public async overwriteClineMessages(newMessages: ClineMessage[]) {
 		this.clineMessages = newMessages
+
+		// If deletion or history truncation leaves a condense_context as the last message,
+		// ensure the next API call suppresses previous_response_id so the condensed context is respected.
+		try {
+			const last = this.clineMessages.at(-1)
+			if (last && last.type === "say" && last.say === "condense_context") {
+				this.skipPrevResponseIdOnce = true
+			}
+		} catch {
+			// non-fatal
+		}
+
 		restoreTodoListForTask(this)
 		await this.saveClineMessages()
 	}
@@ -2633,4 +2645,4 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 	public get taskAsk(): ClineMessage | undefined {
 		return this.idleAsk || this.resumableAsk || this.interactiveAsk
 	}
-}
+}

src/core/task/__tests__/Task.spec.ts

@@ -1615,6 +1615,69 @@ describe("Cline", () => {
 		})
 	})
 
+	describe("Conversation continuity after condense and deletion", () => {
+		it("should set suppressPreviousResponseId when last message is condense_context", async () => {
+			// Arrange: create task
+			const task = new Task({
+				provider: mockProvider,
+				apiConfiguration: mockApiConfig,
+				task: "initial task",
+				startTask: false,
+			})
+
+			// Ensure provider state returns required fields for attemptApiRequest
+			mockProvider.getState = vi.fn().mockResolvedValue({
+				apiConfiguration: mockApiConfig,
+			})
+
+			// Simulate deletion that leaves a condense_context as the last message
+			const condenseMsg = {
+				ts: Date.now(),
+				type: "say" as const,
+				say: "condense_context" as const,
+				contextCondense: {
+					summary: "summarized",
+					cost: 0.001,
+					prevContextTokens: 1200,
+					newContextTokens: 400,
+				},
+			}
+			await task.overwriteClineMessages([condenseMsg])
+
+			// Spy and return a minimal successful stream to exercise attemptApiRequest
+			const mockStream = {
+				async *[Symbol.asyncIterator]() {
+					yield { type: "text", text: "ok" }
+				},
+				async next() {
+					return { done: true, value: { type: "text", text: "ok" } }
+				},
+				async return() {
+					return { done: true, value: undefined }
+				},
+				async throw(e: any) {
+					throw e
+				},
+				[Symbol.asyncDispose]: async () => {},
+			} as AsyncGenerator<ApiStreamChunk>
+
+			const createMessageSpy = vi.spyOn(task.api, "createMessage").mockReturnValue(mockStream)
+
+			// Act: initiate an API request
+			const iterator = task.attemptApiRequest(0)
+			await iterator.next() // read first chunk to ensure call happened
+
+			// Assert: metadata includes suppressPreviousResponseId set to true
+			expect(createMessageSpy).toHaveBeenCalled()
+			const callArgs = createMessageSpy.mock.calls[0]
+			// Args: [systemPrompt, cleanConversationHistory, metadata]
+			const metadata = callArgs?.[2]
+			expect(metadata?.suppressPreviousResponseId).toBe(true)
+
+			// The skip flag should be reset after the call
+			expect((task as any).skipPrevResponseIdOnce).toBe(false)
+		})
+	})
 	describe("abortTask", () => {
 		it("should set abort flag and emit TaskAborted event", async () => {
 			const task = new Task({