fix: handle both Anthropic and OpenAI token semantics correctly

The previous fix only handled Anthropic-style APIs. This update ensures
correct context token calculation when switching between OpenAI and
Anthropic providers in the same task.

Key changes:
- Store tokensIn based on API protocol:
  * Anthropic: inputTokens (base input, no cache)
  * OpenAI: totalInputTokens (total input, includes cache)
- Store apiProtocol in the message for protocol-aware calculations
- getApiMetrics now uses tokensIn directly since it represents total
  input tokens for both protocols

This handles the provider switching scenario correctly.

src/core/task/Task.ts

@@ -1920,17 +1920,20 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 									cacheReadTokens,
 								)
 
-					// For tokensIn, store the base input tokens (not including cache tokens)
-					// This ensures context length calculation works correctly in getApiMetrics
-					// where it adds tokensIn + tokensOut. The cache tokens are stored separately
-					// in cacheWrites and cacheReads for display purposes.
+					// Store tokensIn based on the API protocol:
+					// - For Anthropic: inputTokens is base input (without cache), so use inputTokens
+					// - For OpenAI: inputTokens already includes cache, so use totalInputTokens
+					// This ensures getApiMetrics can correctly calculate context tokens by checking apiProtocol
+					const tokensInValue = apiProtocol === "anthropic" ? inputTokens : costResult.totalInputTokens
+
 					this.clineMessages[lastApiReqIndex].text = JSON.stringify({
 						...existingData,
-						tokensIn: inputTokens,
+						tokensIn: tokensInValue,
 						tokensOut: costResult.totalOutputTokens,
 						cacheWrites: cacheWriteTokens,
 						cacheReads: cacheReadTokens,
 						cost: totalCost ?? costResult.totalCost,
+						apiProtocol,
 						cancelReason,
 						streamingFailedMessage,
 					} satisfies ClineApiReqInfo)

src/shared/getApiMetrics.ts

@@ -80,12 +80,11 @@ export function getApiMetrics(messages: ClineMessage[]) {
 		if (message.type === "say" && message.say === "api_req_started" && message.text) {
 			try {
 				const parsedText: ParsedApiReqStartedTextType = JSON.parse(message.text)
-				const { tokensIn, tokensOut } = parsedText
+				const { tokensIn, tokensOut, apiProtocol } = parsedText
 
-				// tokensIn contains base input tokens (without cache tokens).
-				// This works correctly for both Anthropic and OpenAI protocols because:
-				// - Cache tokens are stored separately in cacheWrites/cacheReads
-				// - Context length = base input + output tokens
+				// Context tokens = total tokens sent to the model
+				// For both Anthropic and OpenAI, tokensIn now stores the total input tokens
+				// (including cache tokens), so we just add tokensIn + tokensOut
 				result.contextTokens = (tokensIn || 0) + (tokensOut || 0)
 			} catch (error) {
 				console.error("Error parsing JSON:", error)